linux/net/core/dev.c

9025 lines
224 KiB

include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 17:04:11 +09:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
hlist: drop the node parameter from iterators I'm not sure why, but the hlist for each entry iterators were conceived list_for_each_entry(pos, head, member) The hlist ones were greedy and wanted an extra parameter: hlist_for_each_entry(tpos, pos, head, member) Why did they need an extra pos parameter? I'm not quite sure. Not only they don't really need it, it also prevents the iterator from looking exactly like the list iterator, which is unfortunate. Besides the semantic patch, there was some manual work required: - Fix up the actual hlist iterators in linux/list.h - Fix up the declaration of other iterators based on the hlist ones. - A very small amount of places were using the 'node' parameter, this was modified to use 'obj->member' instead. - Coccinelle didn't handle the hlist_for_each_entry_safe iterator properly, so those had to be fixed up manually. The semantic patch which is mostly the work of Peter Senna Tschudin is here: @@ iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host; type T; expression a,c,d,e; identifier b; statement S; @@ -T b; <+... when != b ( hlist_for_each_entry(a, - b, c, d) S | hlist_for_each_entry_continue(a, - b, c) S | hlist_for_each_entry_from(a, - b, c) S | hlist_for_each_entry_rcu(a, - b, c, d) S | hlist_for_each_entry_rcu_bh(a, - b, c, d) S | hlist_for_each_entry_continue_rcu_bh(a, - b, c) S | for_each_busy_worker(a, c, - b, d) S | ax25_uid_for_each(a, - b, c) S | ax25_for_each(a, - b, c) S | inet_bind_bucket_for_each(a, - b, c) S | sctp_for_each_hentry(a, - b, c) S | sk_for_each(a, - b, c) S | sk_for_each_rcu(a, - b, c) S | sk_for_each_from -(a, b) +(a) S + sk_for_each_from(a) S | sk_for_each_safe(a, - b, c, d) S | sk_for_each_bound(a, - b, c) S | hlist_for_each_entry_safe(a, - b, c, d, e) S | hlist_for_each_entry_continue_rcu(a, - b, c) S | nr_neigh_for_each(a, - b, c) S | nr_neigh_for_each_safe(a, - b, c, d) S | nr_node_for_each(a, - b, c) S | nr_node_for_each_safe(a, - b, c, d) S | - for_each_gfn_sp(a, c, d, b) S + for_each_gfn_sp(a, c, d) S | - for_each_gfn_indirect_valid_sp(a, c, d, b) S + for_each_gfn_indirect_valid_sp(a, c, d) S | for_each_host(a, - b, c) S | for_each_host_safe(a, - b, c, d) S | for_each_mesh_entry(a, - b, c, d) S ) ...+> [akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c] [akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c] [akpm@linux-foundation.org: checkpatch fixes] [akpm@linux-foundation.org: fix warnings] [akpm@linux-foudnation.org: redo intrusive kvm changes] Tested-by: Peter Senna Tschudin <peter.senna@gmail.com> Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Gleb Natapov <gleb@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-27 17:06:00 -08:00
hlist: drop the node parameter from iterators I'm not sure why, but the hlist for each entry iterators were conceived list_for_each_entry(pos, head, member) The hlist ones were greedy and wanted an extra parameter: hlist_for_each_entry(tpos, pos, head, member) Why did they need an extra pos parameter? I'm not quite sure. Not only they don't really need it, it also prevents the iterator from looking exactly like the list iterator, which is unfortunate. Besides the semantic patch, there was some manual work required: - Fix up the actual hlist iterators in linux/list.h - Fix up the declaration of other iterators based on the hlist ones. - A very small amount of places were using the 'node' parameter, this was modified to use 'obj->member' instead. - Coccinelle didn't handle the hlist_for_each_entry_safe iterator properly, so those had to be fixed up manually. The semantic patch which is mostly the work of Peter Senna Tschudin is here: @@ iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host; type T; expression a,c,d,e; identifier b; statement S; @@ -T b; <+... when != b ( hlist_for_each_entry(a, - b, c, d) S | hlist_for_each_entry_continue(a, - b, c) S | hlist_for_each_entry_from(a, - b, c) S | hlist_for_each_entry_rcu(a, - b, c, d) S | hlist_for_each_entry_rcu_bh(a, - b, c, d) S | hlist_for_each_entry_continue_rcu_bh(a, - b, c) S | for_each_busy_worker(a, c, - b, d) S | ax25_uid_for_each(a, - b, c) S | ax25_for_each(a, - b, c) S | inet_bind_bucket_for_each(a, - b, c) S | sctp_for_each_hentry(a, - b, c) S | sk_for_each(a, - b, c) S | sk_for_each_rcu(a, - b, c) S | sk_for_each_from -(a, b) +(a) S + sk_for_each_from(a) S | sk_for_each_safe(a, - b, c, d) S | sk_for_each_bound(a, - b, c) S | hlist_for_each_entry_safe(a, - b, c, d, e) S | hlist_for_each_entry_continue_rcu(a, - b, c) S | nr_neigh_for_each(a, - b, c) S | nr_neigh_for_each_safe(a, - b, c, d) S | nr_node_for_each(a, - b, c) S | nr_node_for_each_safe(a, - b, c, d) S | - for_each_gfn_sp(a, c, d, b) S + for_each_gfn_sp(a, c, d) S | - for_each_gfn_indirect_valid_sp(a, c, d, b) S + for_each_gfn_indirect_valid_sp(a, c, d) S | for_each_host(a, - b, c) S | for_each_host_safe(a, - b, c, d) S | for_each_mesh_entry(a, - b, c, d) S ) ...+> [akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c] [akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c] [akpm@linux-foundation.org: checkpatch fixes] [akpm@linux-foundation.org: fix warnings] [akpm@linux-foudnation.org: redo intrusive kvm changes] Tested-by: Peter Senna Tschudin <peter.senna@gmail.com> Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Gleb Natapov <gleb@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-27 17:06:00 -08:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
hlist: drop the node parameter from iterators I'm not sure why, but the hlist for each entry iterators were conceived list_for_each_entry(pos, head, member) The hlist ones were greedy and wanted an extra parameter: hlist_for_each_entry(tpos, pos, head, member) Why did they need an extra pos parameter? I'm not quite sure. Not only they don't really need it, it also prevents the iterator from looking exactly like the list iterator, which is unfortunate. Besides the semantic patch, there was some manual work required: - Fix up the actual hlist iterators in linux/list.h - Fix up the declaration of other iterators based on the hlist ones. - A very small amount of places were using the 'node' parameter, this was modified to use 'obj->member' instead. - Coccinelle didn't handle the hlist_for_each_entry_safe iterator properly, so those had to be fixed up manually. The semantic patch which is mostly the work of Peter Senna Tschudin is here: @@ iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host; type T; expression a,c,d,e; identifier b; statement S; @@ -T b; <+... when != b ( hlist_for_each_entry(a, - b, c, d) S | hlist_for_each_entry_continue(a, - b, c) S | hlist_for_each_entry_from(a, - b, c) S | hlist_for_each_entry_rcu(a, - b, c, d) S | hlist_for_each_entry_rcu_bh(a, - b, c, d) S | hlist_for_each_entry_continue_rcu_bh(a, - b, c) S | for_each_busy_worker(a, c, - b, d) S | ax25_uid_for_each(a, - b, c) S | ax25_for_each(a, - b, c) S | inet_bind_bucket_for_each(a, - b, c) S | sctp_for_each_hentry(a, - b, c) S | sk_for_each(a, - b, c) S | sk_for_each_rcu(a, - b, c) S | sk_for_each_from -(a, b) +(a) S + sk_for_each_from(a) S | sk_for_each_safe(a, - b, c, d) S | sk_for_each_bound(a, - b, c) S | hlist_for_each_entry_safe(a, - b, c, d, e) S | hlist_for_each_entry_continue_rcu(a, - b, c) S | nr_neigh_for_each(a, - b, c) S | nr_neigh_for_each_safe(a, - b, c, d) S | nr_node_for_each(a, - b, c) S | nr_node_for_each_safe(a, - b, c, d) S | - for_each_gfn_sp(a, c, d, b) S + for_each_gfn_sp(a, c, d) S | - for_each_gfn_indirect_valid_sp(a, c, d, b) S + for_each_gfn_indirect_valid_sp(a, c, d) S | for_each_host(a, - b, c) S | for_each_host_safe(a, - b, c, d) S | for_each_mesh_entry(a, - b, c, d) S ) ...+> [akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c] [akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c] [akpm@linux-foundation.org: checkpatch fixes] [akpm@linux-foundation.org: fix warnings] [akpm@linux-foudnation.org: redo intrusive kvm changes] Tested-by: Peter Senna Tschudin <peter.senna@gmail.com> Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Gleb Natapov <gleb@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-27 17:06:00 -08:00
hlist: drop the node parameter from iterators I'm not sure why, but the hlist for each entry iterators were conceived list_for_each_entry(pos, head, member) The hlist ones were greedy and wanted an extra parameter: hlist_for_each_entry(tpos, pos, head, member) Why did they need an extra pos parameter? I'm not quite sure. Not only they don't really need it, it also prevents the iterator from looking exactly like the list iterator, which is unfortunate. Besides the semantic patch, there was some manual work required: - Fix up the actual hlist iterators in linux/list.h - Fix up the declaration of other iterators based on the hlist ones. - A very small amount of places were using the 'node' parameter, this was modified to use 'obj->member' instead. - Coccinelle didn't handle the hlist_for_each_entry_safe iterator properly, so those had to be fixed up manually. The semantic patch which is mostly the work of Peter Senna Tschudin is here: @@ iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host; type T; expression a,c,d,e; identifier b; statement S; @@ -T b; <+... when != b ( hlist_for_each_entry(a, - b, c, d) S | hlist_for_each_entry_continue(a, - b, c) S | hlist_for_each_entry_from(a, - b, c) S | hlist_for_each_entry_rcu(a, - b, c, d) S | hlist_for_each_entry_rcu_bh(a, - b, c, d) S | hlist_for_each_entry_continue_rcu_bh(a, - b, c) S | for_each_busy_worker(a, c, - b, d) S | ax25_uid_for_each(a, - b, c) S | ax25_for_each(a, - b, c) S | inet_bind_bucket_for_each(a, - b, c) S | sctp_for_each_hentry(a, - b, c) S | sk_for_each(a, - b, c) S | sk_for_each_rcu(a, - b, c) S | sk_for_each_from -(a, b) +(a) S + sk_for_each_from(a) S | sk_for_each_safe(a, - b, c, d) S | sk_for_each_bound(a, - b, c) S | hlist_for_each_entry_safe(a, - b, c, d, e) S | hlist_for_each_entry_continue_rcu(a, - b, c) S | nr_neigh_for_each(a, - b, c) S | nr_neigh_for_each_safe(a, - b, c, d) S | nr_node_for_each(a, - b, c) S | nr_node_for_each_safe(a, - b, c, d) S | - for_each_gfn_sp(a, c, d, b) S + for_each_gfn_sp(a, c, d) S | - for_each_gfn_indirect_valid_sp(a, c, d, b) S + for_each_gfn_indirect_valid_sp(a, c, d) S | for_each_host(a, - b, c) S | for_each_host_safe(a, - b, c, d) S | for_each_mesh_entry(a, - b, c, d) S ) ...+> [akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c] [akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c] [akpm@linux-foundation.org: checkpatch fixes] [akpm@linux-foundation.org: fix warnings] [akpm@linux-foudnation.org: redo intrusive kvm changes] Tested-by: Peter Senna Tschudin <peter.senna@gmail.com> Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Gleb Natapov <gleb@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-27 17:06:00 -08:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
net: fix rtnl notification in atomic context commit 991fb3f74c "dev: always advertise rx_flags changes via netlink" introduced rtnl notification from __dev_set_promiscuity(), which can be called in atomic context. Steps to reproduce: ip tuntap add dev tap1 mode tap ifconfig tap1 up tcpdump -nei tap1 & ip tuntap del dev tap1 mode tap [ 271.627994] device tap1 left promiscuous mode [ 271.639897] BUG: sleeping function called from invalid context at mm/slub.c:940 [ 271.664491] in_atomic(): 1, irqs_disabled(): 0, pid: 3394, name: ip [ 271.677525] INFO: lockdep is turned off. [ 271.690503] CPU: 0 PID: 3394 Comm: ip Tainted: G W 3.12.0-rc3+ #73 [ 271.703996] Hardware name: System manufacturer System Product Name/P8Z77 WS, BIOS 3007 07/26/2012 [ 271.731254] ffffffff81a58506 ffff8807f0d57a58 ffffffff817544e5 ffff88082fa0f428 [ 271.760261] ffff8808071f5f40 ffff8807f0d57a88 ffffffff8108bad1 ffffffff81110ff8 [ 271.790683] 0000000000000010 00000000000000d0 00000000000000d0 ffff8807f0d57af8 [ 271.822332] Call Trace: [ 271.838234] [<ffffffff817544e5>] dump_stack+0x55/0x76 [ 271.854446] [<ffffffff8108bad1>] __might_sleep+0x181/0x240 [ 271.870836] [<ffffffff81110ff8>] ? rcu_irq_exit+0x68/0xb0 [ 271.887076] [<ffffffff811a80be>] kmem_cache_alloc_node+0x4e/0x2a0 [ 271.903368] [<ffffffff810b4ddc>] ? vprintk_emit+0x1dc/0x5a0 [ 271.919716] [<ffffffff81614d67>] ? __alloc_skb+0x57/0x2a0 [ 271.936088] [<ffffffff810b4de0>] ? vprintk_emit+0x1e0/0x5a0 [ 271.952504] [<ffffffff81614d67>] __alloc_skb+0x57/0x2a0 [ 271.968902] [<ffffffff8163a0b2>] rtmsg_ifinfo+0x52/0x100 [ 271.985302] [<ffffffff8162ac6d>] __dev_notify_flags+0xad/0xc0 [ 272.001642] [<ffffffff8162ad0c>] __dev_set_promiscuity+0x8c/0x1c0 [ 272.017917] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.033961] [<ffffffff8162b109>] dev_set_promiscuity+0x29/0x50 [ 272.049855] [<ffffffff8172e937>] packet_dev_mc+0x87/0xc0 [ 272.065494] [<ffffffff81732052>] packet_notifier+0x1b2/0x380 [ 272.080915] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.096009] [<ffffffff81761c66>] notifier_call_chain+0x66/0x150 [ 272.110803] [<ffffffff8108503e>] __raw_notifier_call_chain+0xe/0x10 [ 272.125468] [<ffffffff81085056>] raw_notifier_call_chain+0x16/0x20 [ 272.139984] [<ffffffff81620190>] call_netdevice_notifiers_info+0x40/0x70 [ 272.154523] [<ffffffff816201d6>] call_netdevice_notifiers+0x16/0x20 [ 272.168552] [<ffffffff816224c5>] rollback_registered_many+0x145/0x240 [ 272.182263] [<ffffffff81622641>] rollback_registered+0x31/0x40 [ 272.195369] [<ffffffff816229c8>] unregister_netdevice_queue+0x58/0x90 [ 272.208230] [<ffffffff81547ca0>] __tun_detach+0x140/0x340 [ 272.220686] [<ffffffff81547ed6>] tun_chr_close+0x36/0x60 Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-10-23 16:02:42 -07:00
netpoll: protect napi_poll and poll_controller during dev_[open|close] Ivan Vercera was recently backporting commit 9c13cb8bb477a83b9a3c9e5a5478a4e21294a760 to a RHEL kernel, and I noticed that, while this patch protects the tg3 driver from having its ndo_poll_controller routine called during device initalization, it does nothing for the driver during shutdown. I.e. it would be entirely possible to have the ndo_poll_controller method (or subsequently the ndo_poll) routine called for a driver in the netpoll path on CPU A while in parallel on CPU B, the ndo_close or ndo_open routine could be called. Given that the two latter routines tend to initizlize and free many data structures that the former two rely on, the result can easily be data corruption or various other crashes. Furthermore, it seems that this is potentially a problem with all net drivers that support netpoll, and so this should ideally be fixed in a common path. As Ben H Pointed out to me, we can't preform dev_open/dev_close in atomic context, so I've come up with this solution. We can use a mutex to sleep in open/close paths and just do a mutex_trylock in the napi poll path and abandon the poll attempt if we're locked, as we'll just retry the poll on the next send anyway. I've tested this here by flooding netconsole with messages on a system whos nic driver I modfied to periodically return NETDEV_TX_BUSY, so that the netpoll tx workqueue would be forced to send frames and poll the device. While this was going on I rapidly ifdown/up'ed the interface and watched for any problems. I've not found any. Signed-off-by: Neil Horman <nhorman@tuxdriver.com> CC: Ivan Vecera <ivecera@redhat.com> CC: "David S. Miller" <davem@davemloft.net> CC: Ben Hutchings <bhutchings@solarflare.com> CC: Francois Romieu <romieu@fr.zoreil.com> CC: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-02-05 08:05:43 +00:00
netpoll: protect napi_poll and poll_controller during dev_[open|close] Ivan Vercera was recently backporting commit 9c13cb8bb477a83b9a3c9e5a5478a4e21294a760 to a RHEL kernel, and I noticed that, while this patch protects the tg3 driver from having its ndo_poll_controller routine called during device initalization, it does nothing for the driver during shutdown. I.e. it would be entirely possible to have the ndo_poll_controller method (or subsequently the ndo_poll) routine called for a driver in the netpoll path on CPU A while in parallel on CPU B, the ndo_close or ndo_open routine could be called. Given that the two latter routines tend to initizlize and free many data structures that the former two rely on, the result can easily be data corruption or various other crashes. Furthermore, it seems that this is potentially a problem with all net drivers that support netpoll, and so this should ideally be fixed in a common path. As Ben H Pointed out to me, we can't preform dev_open/dev_close in atomic context, so I've come up with this solution. We can use a mutex to sleep in open/close paths and just do a mutex_trylock in the napi poll path and abandon the poll attempt if we're locked, as we'll just retry the poll on the next send anyway. I've tested this here by flooding netconsole with messages on a system whos nic driver I modfied to periodically return NETDEV_TX_BUSY, so that the netpoll tx workqueue would be forced to send frames and poll the device. While this was going on I rapidly ifdown/up'ed the interface and watched for any problems. I've not found any. Signed-off-by: Neil Horman <nhorman@tuxdriver.com> CC: Ivan Vecera <ivecera@redhat.com> CC: "David S. Miller" <davem@davemloft.net> CC: Ben Hutchings <bhutchings@solarflare.com> CC: Francois Romieu <romieu@fr.zoreil.com> CC: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-02-05 08:05:43 +00:00
netpoll: protect napi_poll and poll_controller during dev_[open|close] Ivan Vercera was recently backporting commit 9c13cb8bb477a83b9a3c9e5a5478a4e21294a760 to a RHEL kernel, and I noticed that, while this patch protects the tg3 driver from having its ndo_poll_controller routine called during device initalization, it does nothing for the driver during shutdown. I.e. it would be entirely possible to have the ndo_poll_controller method (or subsequently the ndo_poll) routine called for a driver in the netpoll path on CPU A while in parallel on CPU B, the ndo_close or ndo_open routine could be called. Given that the two latter routines tend to initizlize and free many data structures that the former two rely on, the result can easily be data corruption or various other crashes. Furthermore, it seems that this is potentially a problem with all net drivers that support netpoll, and so this should ideally be fixed in a common path. As Ben H Pointed out to me, we can't preform dev_open/dev_close in atomic context, so I've come up with this solution. We can use a mutex to sleep in open/close paths and just do a mutex_trylock in the napi poll path and abandon the poll attempt if we're locked, as we'll just retry the poll on the next send anyway. I've tested this here by flooding netconsole with messages on a system whos nic driver I modfied to periodically return NETDEV_TX_BUSY, so that the netpoll tx workqueue would be forced to send frames and poll the device. While this was going on I rapidly ifdown/up'ed the interface and watched for any problems. I've not found any. Signed-off-by: Neil Horman <nhorman@tuxdriver.com> CC: Ivan Vecera <ivecera@redhat.com> CC: "David S. Miller" <davem@davemloft.net> CC: Ben Hutchings <bhutchings@solarflare.com> CC: Francois Romieu <romieu@fr.zoreil.com> CC: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-02-05 08:05:43 +00:00
net: fix rtnl notification in atomic context commit 991fb3f74c "dev: always advertise rx_flags changes via netlink" introduced rtnl notification from __dev_set_promiscuity(), which can be called in atomic context. Steps to reproduce: ip tuntap add dev tap1 mode tap ifconfig tap1 up tcpdump -nei tap1 & ip tuntap del dev tap1 mode tap [ 271.627994] device tap1 left promiscuous mode [ 271.639897] BUG: sleeping function called from invalid context at mm/slub.c:940 [ 271.664491] in_atomic(): 1, irqs_disabled(): 0, pid: 3394, name: ip [ 271.677525] INFO: lockdep is turned off. [ 271.690503] CPU: 0 PID: 3394 Comm: ip Tainted: G W 3.12.0-rc3+ #73 [ 271.703996] Hardware name: System manufacturer System Product Name/P8Z77 WS, BIOS 3007 07/26/2012 [ 271.731254] ffffffff81a58506 ffff8807f0d57a58 ffffffff817544e5 ffff88082fa0f428 [ 271.760261] ffff8808071f5f40 ffff8807f0d57a88 ffffffff8108bad1 ffffffff81110ff8 [ 271.790683] 0000000000000010 00000000000000d0 00000000000000d0 ffff8807f0d57af8 [ 271.822332] Call Trace: [ 271.838234] [<ffffffff817544e5>] dump_stack+0x55/0x76 [ 271.854446] [<ffffffff8108bad1>] __might_sleep+0x181/0x240 [ 271.870836] [<ffffffff81110ff8>] ? rcu_irq_exit+0x68/0xb0 [ 271.887076] [<ffffffff811a80be>] kmem_cache_alloc_node+0x4e/0x2a0 [ 271.903368] [<ffffffff810b4ddc>] ? vprintk_emit+0x1dc/0x5a0 [ 271.919716] [<ffffffff81614d67>] ? __alloc_skb+0x57/0x2a0 [ 271.936088] [<ffffffff810b4de0>] ? vprintk_emit+0x1e0/0x5a0 [ 271.952504] [<ffffffff81614d67>] __alloc_skb+0x57/0x2a0 [ 271.968902] [<ffffffff8163a0b2>] rtmsg_ifinfo+0x52/0x100 [ 271.985302] [<ffffffff8162ac6d>] __dev_notify_flags+0xad/0xc0 [ 272.001642] [<ffffffff8162ad0c>] __dev_set_promiscuity+0x8c/0x1c0 [ 272.017917] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.033961] [<ffffffff8162b109>] dev_set_promiscuity+0x29/0x50 [ 272.049855] [<ffffffff8172e937>] packet_dev_mc+0x87/0xc0 [ 272.065494] [<ffffffff81732052>] packet_notifier+0x1b2/0x380 [ 272.080915] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.096009] [<ffffffff81761c66>] notifier_call_chain+0x66/0x150 [ 272.110803] [<ffffffff8108503e>] __raw_notifier_call_chain+0xe/0x10 [ 272.125468] [<ffffffff81085056>] raw_notifier_call_chain+0x16/0x20 [ 272.139984] [<ffffffff81620190>] call_netdevice_notifiers_info+0x40/0x70 [ 272.154523] [<ffffffff816201d6>] call_netdevice_notifiers+0x16/0x20 [ 272.168552] [<ffffffff816224c5>] rollback_registered_many+0x145/0x240 [ 272.182263] [<ffffffff81622641>] rollback_registered+0x31/0x40 [ 272.195369] [<ffffffff816229c8>] unregister_netdevice_queue+0x58/0x90 [ 272.208230] [<ffffffff81547ca0>] __tun_detach+0x140/0x340 [ 272.220686] [<ffffffff81547ed6>] tun_chr_close+0x36/0x60 Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-10-23 16:02:42 -07:00
net: fix rtnl notification in atomic context commit 991fb3f74c "dev: always advertise rx_flags changes via netlink" introduced rtnl notification from __dev_set_promiscuity(), which can be called in atomic context. Steps to reproduce: ip tuntap add dev tap1 mode tap ifconfig tap1 up tcpdump -nei tap1 & ip tuntap del dev tap1 mode tap [ 271.627994] device tap1 left promiscuous mode [ 271.639897] BUG: sleeping function called from invalid context at mm/slub.c:940 [ 271.664491] in_atomic(): 1, irqs_disabled(): 0, pid: 3394, name: ip [ 271.677525] INFO: lockdep is turned off. [ 271.690503] CPU: 0 PID: 3394 Comm: ip Tainted: G W 3.12.0-rc3+ #73 [ 271.703996] Hardware name: System manufacturer System Product Name/P8Z77 WS, BIOS 3007 07/26/2012 [ 271.731254] ffffffff81a58506 ffff8807f0d57a58 ffffffff817544e5 ffff88082fa0f428 [ 271.760261] ffff8808071f5f40 ffff8807f0d57a88 ffffffff8108bad1 ffffffff81110ff8 [ 271.790683] 0000000000000010 00000000000000d0 00000000000000d0 ffff8807f0d57af8 [ 271.822332] Call Trace: [ 271.838234] [<ffffffff817544e5>] dump_stack+0x55/0x76 [ 271.854446] [<ffffffff8108bad1>] __might_sleep+0x181/0x240 [ 271.870836] [<ffffffff81110ff8>] ? rcu_irq_exit+0x68/0xb0 [ 271.887076] [<ffffffff811a80be>] kmem_cache_alloc_node+0x4e/0x2a0 [ 271.903368] [<ffffffff810b4ddc>] ? vprintk_emit+0x1dc/0x5a0 [ 271.919716] [<ffffffff81614d67>] ? __alloc_skb+0x57/0x2a0 [ 271.936088] [<ffffffff810b4de0>] ? vprintk_emit+0x1e0/0x5a0 [ 271.952504] [<ffffffff81614d67>] __alloc_skb+0x57/0x2a0 [ 271.968902] [<ffffffff8163a0b2>] rtmsg_ifinfo+0x52/0x100 [ 271.985302] [<ffffffff8162ac6d>] __dev_notify_flags+0xad/0xc0 [ 272.001642] [<ffffffff8162ad0c>] __dev_set_promiscuity+0x8c/0x1c0 [ 272.017917] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.033961] [<ffffffff8162b109>] dev_set_promiscuity+0x29/0x50 [ 272.049855] [<ffffffff8172e937>] packet_dev_mc+0x87/0xc0 [ 272.065494] [<ffffffff81732052>] packet_notifier+0x1b2/0x380 [ 272.080915] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.096009] [<ffffffff81761c66>] notifier_call_chain+0x66/0x150 [ 272.110803] [<ffffffff8108503e>] __raw_notifier_call_chain+0xe/0x10 [ 272.125468] [<ffffffff81085056>] raw_notifier_call_chain+0x16/0x20 [ 272.139984] [<ffffffff81620190>] call_netdevice_notifiers_info+0x40/0x70 [ 272.154523] [<ffffffff816201d6>] call_netdevice_notifiers+0x16/0x20 [ 272.168552] [<ffffffff816224c5>] rollback_registered_many+0x145/0x240 [ 272.182263] [<ffffffff81622641>] rollback_registered+0x31/0x40 [ 272.195369] [<ffffffff816229c8>] unregister_netdevice_queue+0x58/0x90 [ 272.208230] [<ffffffff81547ca0>] __tun_detach+0x140/0x340 [ 272.220686] [<ffffffff81547ed6>] tun_chr_close+0x36/0x60 Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-10-23 16:02:42 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
net: dont call jump_label_dec from irq context Igor Maravic reported an error caused by jump_label_dec() being called from IRQ context : BUG: sleeping function called from invalid context at kernel/mutex.c:271 in_atomic(): 1, irqs_disabled(): 0, pid: 0, name: swapper 1 lock held by swapper/0: #0: (&n->timer){+.-...}, at: [<ffffffff8107ce90>] call_timer_fn+0x0/0x340 Pid: 0, comm: swapper Not tainted 3.2.0-rc2-net-next-mpls+ #1 Call Trace: <IRQ> [<ffffffff8104f417>] __might_sleep+0x137/0x1f0 [<ffffffff816b9a2f>] mutex_lock_nested+0x2f/0x370 [<ffffffff810a89fd>] ? trace_hardirqs_off+0xd/0x10 [<ffffffff8109a37f>] ? local_clock+0x6f/0x80 [<ffffffff810a90a5>] ? lock_release_holdtime.part.22+0x15/0x1a0 [<ffffffff81557929>] ? sock_def_write_space+0x59/0x160 [<ffffffff815e936e>] ? arp_error_report+0x3e/0x90 [<ffffffff810969cd>] atomic_dec_and_mutex_lock+0x5d/0x80 [<ffffffff8112fc1d>] jump_label_dec+0x1d/0x50 [<ffffffff81566525>] net_disable_timestamp+0x15/0x20 [<ffffffff81557a75>] sock_disable_timestamp+0x45/0x50 [<ffffffff81557b00>] __sk_free+0x80/0x200 [<ffffffff815578d0>] ? sk_send_sigurg+0x70/0x70 [<ffffffff815e936e>] ? arp_error_report+0x3e/0x90 [<ffffffff81557cba>] sock_wfree+0x3a/0x70 [<ffffffff8155c2b0>] skb_release_head_state+0x70/0x120 [<ffffffff8155c0b6>] __kfree_skb+0x16/0x30 [<ffffffff8155c119>] kfree_skb+0x49/0x170 [<ffffffff815e936e>] arp_error_report+0x3e/0x90 [<ffffffff81575bd9>] neigh_invalidate+0x89/0xc0 [<ffffffff81578dbe>] neigh_timer_handler+0x9e/0x2a0 [<ffffffff81578d20>] ? neigh_update+0x640/0x640 [<ffffffff81073558>] __do_softirq+0xc8/0x3a0 Since jump_label_{inc|dec} must be called from process context only, we must defer jump_label_dec() if net_disable_timestamp() is called from interrupt context. Reported-by: Igor Maravic <igorm@etf.rs> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-28 11:16:50 +00:00
net: use a work queue to defer net_disable_timestamp() work Dmitry reported a warning [1] showing that we were calling net_disable_timestamp() -> static_key_slow_dec() from a non process context. Grabbing a mutex while holding a spinlock or rcu_read_lock() is not allowed. As Cong suggested, we now use a work queue. It is possible netstamp_clear() exits while netstamp_needed_deferred is not zero, but it is probably not worth trying to do better than that. netstamp_needed_deferred atomic tracks the exact number of deferred decrements. [1] [ INFO: suspicious RCU usage. ] 4.10.0-rc5+ #192 Not tainted ------------------------------- ./include/linux/rcupdate.h:561 Illegal context switch in RCU read-side critical section! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 0 2 locks held by syz-executor14/23111: #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] lock_sock include/net/sock.h:1454 [inline] #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] rawv6_sendmsg+0x1e65/0x3ec0 net/ipv6/raw.c:919 #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] nf_hook include/linux/netfilter.h:201 [inline] #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] __ip6_local_out+0x258/0x840 net/ipv6/output_core.c:160 stack backtrace: CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 lockdep_rcu_suspicious+0x139/0x180 kernel/locking/lockdep.c:4452 rcu_preempt_sleep_check include/linux/rcupdate.h:560 [inline] ___might_sleep+0x560/0x650 kernel/sched/core.c:7748 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 RSP: 002b:00007f6f46fceb58 EFLAGS: 00000292 ORIG_RAX: 0000000000000014 RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 0000000000445559 RDX: 0000000000000001 RSI: 0000000020f1eff0 RDI: 0000000000000005 RBP: 00000000006e19c0 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000292 R12: 0000000000700000 R13: 0000000020f59000 R14: 0000000000000015 R15: 0000000000020400 BUG: sleeping function called from invalid context at kernel/locking/mutex.c:752 in_atomic(): 1, irqs_disabled(): 0, pid: 23111, name: syz-executor14 INFO: lockdep is turned off. CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 ___might_sleep+0x47e/0x650 kernel/sched/core.c:7780 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 Fixes: b90e5794c5bd ("net: dont call jump_label_dec from irq context") Suggested-by: Cong Wang <xiyou.wangcong@gmail.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-02 10:31:35 -08:00
net: use a work queue to defer net_disable_timestamp() work Dmitry reported a warning [1] showing that we were calling net_disable_timestamp() -> static_key_slow_dec() from a non process context. Grabbing a mutex while holding a spinlock or rcu_read_lock() is not allowed. As Cong suggested, we now use a work queue. It is possible netstamp_clear() exits while netstamp_needed_deferred is not zero, but it is probably not worth trying to do better than that. netstamp_needed_deferred atomic tracks the exact number of deferred decrements. [1] [ INFO: suspicious RCU usage. ] 4.10.0-rc5+ #192 Not tainted ------------------------------- ./include/linux/rcupdate.h:561 Illegal context switch in RCU read-side critical section! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 0 2 locks held by syz-executor14/23111: #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] lock_sock include/net/sock.h:1454 [inline] #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] rawv6_sendmsg+0x1e65/0x3ec0 net/ipv6/raw.c:919 #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] nf_hook include/linux/netfilter.h:201 [inline] #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] __ip6_local_out+0x258/0x840 net/ipv6/output_core.c:160 stack backtrace: CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 lockdep_rcu_suspicious+0x139/0x180 kernel/locking/lockdep.c:4452 rcu_preempt_sleep_check include/linux/rcupdate.h:560 [inline] ___might_sleep+0x560/0x650 kernel/sched/core.c:7748 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 RSP: 002b:00007f6f46fceb58 EFLAGS: 00000292 ORIG_RAX: 0000000000000014 RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 0000000000445559 RDX: 0000000000000001 RSI: 0000000020f1eff0 RDI: 0000000000000005 RBP: 00000000006e19c0 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000292 R12: 0000000000700000 R13: 0000000020f59000 R14: 0000000000000015 R15: 0000000000020400 BUG: sleeping function called from invalid context at kernel/locking/mutex.c:752 in_atomic(): 1, irqs_disabled(): 0, pid: 23111, name: syz-executor14 INFO: lockdep is turned off. CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 ___might_sleep+0x47e/0x650 kernel/sched/core.c:7780 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 Fixes: b90e5794c5bd ("net: dont call jump_label_dec from irq context") Suggested-by: Cong Wang <xiyou.wangcong@gmail.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-02 10:31:35 -08:00
net: use a work queue to defer net_disable_timestamp() work Dmitry reported a warning [1] showing that we were calling net_disable_timestamp() -> static_key_slow_dec() from a non process context. Grabbing a mutex while holding a spinlock or rcu_read_lock() is not allowed. As Cong suggested, we now use a work queue. It is possible netstamp_clear() exits while netstamp_needed_deferred is not zero, but it is probably not worth trying to do better than that. netstamp_needed_deferred atomic tracks the exact number of deferred decrements. [1] [ INFO: suspicious RCU usage. ] 4.10.0-rc5+ #192 Not tainted ------------------------------- ./include/linux/rcupdate.h:561 Illegal context switch in RCU read-side critical section! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 0 2 locks held by syz-executor14/23111: #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] lock_sock include/net/sock.h:1454 [inline] #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] rawv6_sendmsg+0x1e65/0x3ec0 net/ipv6/raw.c:919 #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] nf_hook include/linux/netfilter.h:201 [inline] #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] __ip6_local_out+0x258/0x840 net/ipv6/output_core.c:160 stack backtrace: CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 lockdep_rcu_suspicious+0x139/0x180 kernel/locking/lockdep.c:4452 rcu_preempt_sleep_check include/linux/rcupdate.h:560 [inline] ___might_sleep+0x560/0x650 kernel/sched/core.c:7748 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 RSP: 002b:00007f6f46fceb58 EFLAGS: 00000292 ORIG_RAX: 0000000000000014 RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 0000000000445559 RDX: 0000000000000001 RSI: 0000000020f1eff0 RDI: 0000000000000005 RBP: 00000000006e19c0 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000292 R12: 0000000000700000 R13: 0000000020f59000 R14: 0000000000000015 R15: 0000000000020400 BUG: sleeping function called from invalid context at kernel/locking/mutex.c:752 in_atomic(): 1, irqs_disabled(): 0, pid: 23111, name: syz-executor14 INFO: lockdep is turned off. CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 ___might_sleep+0x47e/0x650 kernel/sched/core.c:7780 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 Fixes: b90e5794c5bd ("net: dont call jump_label_dec from irq context") Suggested-by: Cong Wang <xiyou.wangcong@gmail.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-02 10:31:35 -08:00
net: dont call jump_label_dec from irq context Igor Maravic reported an error caused by jump_label_dec() being called from IRQ context : BUG: sleeping function called from invalid context at kernel/mutex.c:271 in_atomic(): 1, irqs_disabled(): 0, pid: 0, name: swapper 1 lock held by swapper/0: #0: (&n->timer){+.-...}, at: [<ffffffff8107ce90>] call_timer_fn+0x0/0x340 Pid: 0, comm: swapper Not tainted 3.2.0-rc2-net-next-mpls+ #1 Call Trace: <IRQ> [<ffffffff8104f417>] __might_sleep+0x137/0x1f0 [<ffffffff816b9a2f>] mutex_lock_nested+0x2f/0x370 [<ffffffff810a89fd>] ? trace_hardirqs_off+0xd/0x10 [<ffffffff8109a37f>] ? local_clock+0x6f/0x80 [<ffffffff810a90a5>] ? lock_release_holdtime.part.22+0x15/0x1a0 [<ffffffff81557929>] ? sock_def_write_space+0x59/0x160 [<ffffffff815e936e>] ? arp_error_report+0x3e/0x90 [<ffffffff810969cd>] atomic_dec_and_mutex_lock+0x5d/0x80 [<ffffffff8112fc1d>] jump_label_dec+0x1d/0x50 [<ffffffff81566525>] net_disable_timestamp+0x15/0x20 [<ffffffff81557a75>] sock_disable_timestamp+0x45/0x50 [<ffffffff81557b00>] __sk_free+0x80/0x200 [<ffffffff815578d0>] ? sk_send_sigurg+0x70/0x70 [<ffffffff815e936e>] ? arp_error_report+0x3e/0x90 [<ffffffff81557cba>] sock_wfree+0x3a/0x70 [<ffffffff8155c2b0>] skb_release_head_state+0x70/0x120 [<ffffffff8155c0b6>] __kfree_skb+0x16/0x30 [<ffffffff8155c119>] kfree_skb+0x49/0x170 [<ffffffff815e936e>] arp_error_report+0x3e/0x90 [<ffffffff81575bd9>] neigh_invalidate+0x89/0xc0 [<ffffffff81578dbe>] neigh_timer_handler+0x9e/0x2a0 [<ffffffff81578d20>] ? neigh_update+0x640/0x640 [<ffffffff81073558>] __do_softirq+0xc8/0x3a0 Since jump_label_{inc|dec} must be called from process context only, we must defer jump_label_dec() if net_disable_timestamp() is called from interrupt context. Reported-by: Igor Maravic <igorm@etf.rs> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-28 11:16:50 +00:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
net: dont call jump_label_dec from irq context Igor Maravic reported an error caused by jump_label_dec() being called from IRQ context : BUG: sleeping function called from invalid context at kernel/mutex.c:271 in_atomic(): 1, irqs_disabled(): 0, pid: 0, name: swapper 1 lock held by swapper/0: #0: (&n->timer){+.-...}, at: [<ffffffff8107ce90>] call_timer_fn+0x0/0x340 Pid: 0, comm: swapper Not tainted 3.2.0-rc2-net-next-mpls+ #1 Call Trace: <IRQ> [<ffffffff8104f417>] __might_sleep+0x137/0x1f0 [<ffffffff816b9a2f>] mutex_lock_nested+0x2f/0x370 [<ffffffff810a89fd>] ? trace_hardirqs_off+0xd/0x10 [<ffffffff8109a37f>] ? local_clock+0x6f/0x80 [<ffffffff810a90a5>] ? lock_release_holdtime.part.22+0x15/0x1a0 [<ffffffff81557929>] ? sock_def_write_space+0x59/0x160 [<ffffffff815e936e>] ? arp_error_report+0x3e/0x90 [<ffffffff810969cd>] atomic_dec_and_mutex_lock+0x5d/0x80 [<ffffffff8112fc1d>] jump_label_dec+0x1d/0x50 [<ffffffff81566525>] net_disable_timestamp+0x15/0x20 [<ffffffff81557a75>] sock_disable_timestamp+0x45/0x50 [<ffffffff81557b00>] __sk_free+0x80/0x200 [<ffffffff815578d0>] ? sk_send_sigurg+0x70/0x70 [<ffffffff815e936e>] ? arp_error_report+0x3e/0x90 [<ffffffff81557cba>] sock_wfree+0x3a/0x70 [<ffffffff8155c2b0>] skb_release_head_state+0x70/0x120 [<ffffffff8155c0b6>] __kfree_skb+0x16/0x30 [<ffffffff8155c119>] kfree_skb+0x49/0x170 [<ffffffff815e936e>] arp_error_report+0x3e/0x90 [<ffffffff81575bd9>] neigh_invalidate+0x89/0xc0 [<ffffffff81578dbe>] neigh_timer_handler+0x9e/0x2a0 [<ffffffff81578d20>] ? neigh_update+0x640/0x640 [<ffffffff81073558>] __do_softirq+0xc8/0x3a0 Since jump_label_{inc|dec} must be called from process context only, we must defer jump_label_dec() if net_disable_timestamp() is called from interrupt context. Reported-by: Igor Maravic <igorm@etf.rs> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-11-28 11:16:50 +00:00
net: use a work queue to defer net_disable_timestamp() work Dmitry reported a warning [1] showing that we were calling net_disable_timestamp() -> static_key_slow_dec() from a non process context. Grabbing a mutex while holding a spinlock or rcu_read_lock() is not allowed. As Cong suggested, we now use a work queue. It is possible netstamp_clear() exits while netstamp_needed_deferred is not zero, but it is probably not worth trying to do better than that. netstamp_needed_deferred atomic tracks the exact number of deferred decrements. [1] [ INFO: suspicious RCU usage. ] 4.10.0-rc5+ #192 Not tainted ------------------------------- ./include/linux/rcupdate.h:561 Illegal context switch in RCU read-side critical section! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 0 2 locks held by syz-executor14/23111: #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] lock_sock include/net/sock.h:1454 [inline] #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] rawv6_sendmsg+0x1e65/0x3ec0 net/ipv6/raw.c:919 #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] nf_hook include/linux/netfilter.h:201 [inline] #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] __ip6_local_out+0x258/0x840 net/ipv6/output_core.c:160 stack backtrace: CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 lockdep_rcu_suspicious+0x139/0x180 kernel/locking/lockdep.c:4452 rcu_preempt_sleep_check include/linux/rcupdate.h:560 [inline] ___might_sleep+0x560/0x650 kernel/sched/core.c:7748 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 RSP: 002b:00007f6f46fceb58 EFLAGS: 00000292 ORIG_RAX: 0000000000000014 RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 0000000000445559 RDX: 0000000000000001 RSI: 0000000020f1eff0 RDI: 0000000000000005 RBP: 00000000006e19c0 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000292 R12: 0000000000700000 R13: 0000000020f59000 R14: 0000000000000015 R15: 0000000000020400 BUG: sleeping function called from invalid context at kernel/locking/mutex.c:752 in_atomic(): 1, irqs_disabled(): 0, pid: 23111, name: syz-executor14 INFO: lockdep is turned off. CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 ___might_sleep+0x47e/0x650 kernel/sched/core.c:7780 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 Fixes: b90e5794c5bd ("net: dont call jump_label_dec from irq context") Suggested-by: Cong Wang <xiyou.wangcong@gmail.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-02 10:31:35 -08:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
net: use a work queue to defer net_disable_timestamp() work Dmitry reported a warning [1] showing that we were calling net_disable_timestamp() -> static_key_slow_dec() from a non process context. Grabbing a mutex while holding a spinlock or rcu_read_lock() is not allowed. As Cong suggested, we now use a work queue. It is possible netstamp_clear() exits while netstamp_needed_deferred is not zero, but it is probably not worth trying to do better than that. netstamp_needed_deferred atomic tracks the exact number of deferred decrements. [1] [ INFO: suspicious RCU usage. ] 4.10.0-rc5+ #192 Not tainted ------------------------------- ./include/linux/rcupdate.h:561 Illegal context switch in RCU read-side critical section! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 0 2 locks held by syz-executor14/23111: #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] lock_sock include/net/sock.h:1454 [inline] #0: (sk_lock-AF_INET6){+.+.+.}, at: [<ffffffff83a35c35>] rawv6_sendmsg+0x1e65/0x3ec0 net/ipv6/raw.c:919 #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] nf_hook include/linux/netfilter.h:201 [inline] #1: (rcu_read_lock){......}, at: [<ffffffff83ae2678>] __ip6_local_out+0x258/0x840 net/ipv6/output_core.c:160 stack backtrace: CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 lockdep_rcu_suspicious+0x139/0x180 kernel/locking/lockdep.c:4452 rcu_preempt_sleep_check include/linux/rcupdate.h:560 [inline] ___might_sleep+0x560/0x650 kernel/sched/core.c:7748 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 RSP: 002b:00007f6f46fceb58 EFLAGS: 00000292 ORIG_RAX: 0000000000000014 RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 0000000000445559 RDX: 0000000000000001 RSI: 0000000020f1eff0 RDI: 0000000000000005 RBP: 00000000006e19c0 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000292 R12: 0000000000700000 R13: 0000000020f59000 R14: 0000000000000015 R15: 0000000000020400 BUG: sleeping function called from invalid context at kernel/locking/mutex.c:752 in_atomic(): 1, irqs_disabled(): 0, pid: 23111, name: syz-executor14 INFO: lockdep is turned off. CPU: 2 PID: 23111 Comm: syz-executor14 Not tainted 4.10.0-rc5+ #192 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:15 [inline] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 ___might_sleep+0x47e/0x650 kernel/sched/core.c:7780 __might_sleep+0x95/0x1a0 kernel/sched/core.c:7739 mutex_lock_nested+0x24f/0x1730 kernel/locking/mutex.c:752 atomic_dec_and_mutex_lock+0x119/0x160 kernel/locking/mutex.c:1060 __static_key_slow_dec+0x7a/0x1e0 kernel/jump_label.c:149 static_key_slow_dec+0x51/0x90 kernel/jump_label.c:174 net_disable_timestamp+0x3b/0x50 net/core/dev.c:1728 sock_disable_timestamp+0x98/0xc0 net/core/sock.c:403 __sk_destruct+0x27d/0x6b0 net/core/sock.c:1441 sk_destruct+0x47/0x80 net/core/sock.c:1460 __sk_free+0x57/0x230 net/core/sock.c:1468 sock_wfree+0xae/0x120 net/core/sock.c:1645 skb_release_head_state+0xfc/0x200 net/core/skbuff.c:655 skb_release_all+0x15/0x60 net/core/skbuff.c:668 __kfree_skb+0x15/0x20 net/core/skbuff.c:684 kfree_skb+0x16e/0x4c0 net/core/skbuff.c:705 inet_frag_destroy+0x121/0x290 net/ipv4/inet_fragment.c:304 inet_frag_put include/net/inet_frag.h:133 [inline] nf_ct_frag6_gather+0x1106/0x3840 net/ipv6/netfilter/nf_conntrack_reasm.c:617 ipv6_defrag+0x1be/0x2b0 net/ipv6/netfilter/nf_defrag_ipv6_hooks.c:68 nf_hook_entry_hookfn include/linux/netfilter.h:102 [inline] nf_hook_slow+0xc3/0x290 net/netfilter/core.c:310 nf_hook include/linux/netfilter.h:212 [inline] __ip6_local_out+0x489/0x840 net/ipv6/output_core.c:160 ip6_local_out+0x2d/0x170 net/ipv6/output_core.c:170 ip6_send_skb+0xa1/0x340 net/ipv6/ip6_output.c:1722 ip6_push_pending_frames+0xb3/0xe0 net/ipv6/ip6_output.c:1742 rawv6_push_pending_frames net/ipv6/raw.c:613 [inline] rawv6_sendmsg+0x2d1a/0x3ec0 net/ipv6/raw.c:927 inet_sendmsg+0x164/0x5b0 net/ipv4/af_inet.c:744 sock_sendmsg_nosec net/socket.c:635 [inline] sock_sendmsg+0xca/0x110 net/socket.c:645 sock_write_iter+0x326/0x600 net/socket.c:848 do_iter_readv_writev+0x2e3/0x5b0 fs/read_write.c:695 do_readv_writev+0x42c/0x9b0 fs/read_write.c:872 vfs_writev+0x87/0xc0 fs/read_write.c:911 do_writev+0x110/0x2c0 fs/read_write.c:944 SYSC_writev fs/read_write.c:1017 [inline] SyS_writev+0x27/0x30 fs/read_write.c:1014 entry_SYSCALL_64_fastpath+0x1f/0xc2 RIP: 0033:0x445559 Fixes: b90e5794c5bd ("net: dont call jump_label_dec from irq context") Suggested-by: Cong Wang <xiyou.wangcong@gmail.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-02 10:31:35 -08:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 15:57:29 -08:00
net: accept UFO datagrams from tuntap and packet Tuntap and similar devices can inject GSO packets. Accept type VIRTIO_NET_HDR_GSO_UDP, even though not generating UFO natively. Processes are expected to use feature negotiation such as TUNSETOFFLOAD to detect supported offload types and refrain from injecting other packets. This process breaks down with live migration: guest kernels do not renegotiate flags, so destination hosts need to expose all features that the source host does. Partially revert the UFO removal from 182e0b6b5846~1..d9d30adf5677. This patch introduces nearly(*) no new code to simplify verification. It brings back verbatim tuntap UFO negotiation, VIRTIO_NET_HDR_GSO_UDP insertion and software UFO segmentation. It does not reinstate protocol stack support, hardware offload (NETIF_F_UFO), SKB_GSO_UDP tunneling in SKB_GSO_SOFTWARE or reception of VIRTIO_NET_HDR_GSO_UDP packets in tuntap. To support SKB_GSO_UDP reappearing in the stack, also reinstate logic in act_csum and openvswitch. Achieve equivalence with v4.13 HEAD by squashing in commit 939912216fa8 ("net: skb_needs_check() removes CHECKSUM_UNNECESSARY check for tx.") and reverting commit 8d63bee643f1 ("net: avoid skb_warn_bad_offload false positives on UFO"). (*) To avoid having to bring back skb_shinfo(skb)->ip6_frag_id, ipv6_proxy_select_ident is changed to return a __be32 and this is assigned directly to the frag_hdr. Also, SKB_GSO_UDP is inserted at the end of the enum to minimize code churn. Tested Booted a v4.13 guest kernel with QEMU. On a host kernel before this patch `ethtool -k eth0` shows UFO disabled. After the patch, it is enabled, same as on a v4.13 host kernel. A UFO packet sent from the guest appears on the tap device: host: nc -l -p -u 8000 & tcpdump -n -i tap0 guest: dd if=/dev/zero of=payload.txt bs=1 count=2000 nc -u 192.16.1.1 8000 < payload.txt Direct tap to tap transmission of VIRTIO_NET_HDR_GSO_UDP succeeds, packets arriving fragmented: ./with_tap_pair.sh ./tap_send_ufo tap0 tap1 (from https://github.com/wdebruij/kerneltools/tree/master/tests) Changes v1 -> v2 - simplified set_offload change (review comment) - documented test procedure Link: http://lkml.kernel.org/r/<CAF=yD-LuUeDuL9YWPJD9ykOZ0QCjNeznPDr6whqZ9NGMNF12Mw@mail.gmail.com> Fixes: fb652fdfe837 ("macvlan/macvtap: Remove NETIF_F_UFO advertisement.") Reported-by: Michal Kubecek <mkubecek@suse.cz> Signed-off-by: Willem de Bruijn <willemb@google.com> Acked-by: Jason Wang <jasowang@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-11-21 10:22:25 -05:00
net: reduce skb_warn_bad_offload() noise Dmitry reported warnings occurring in __skb_gso_segment() [1] All SKB_GSO_DODGY producers can allow user space to feed packets that trigger the current check. We could prevent them from doing so, rejecting packets, but this might add regressions to existing programs. It turns out our SKB_GSO_DODGY handlers properly set up checksum information that is needed anyway when packets needs to be segmented. By checking again skb_needs_check() after skb_mac_gso_segment(), we should remove these pesky warnings, at a very minor cost. With help from Willem de Bruijn [1] WARNING: CPU: 1 PID: 6768 at net/core/dev.c:2439 skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 lo: caps=(0x000000a2803b7c69, 0x0000000000000000) len=138 data_len=0 gso_size=15883 gso_type=4 ip_summed=0 Kernel panic - not syncing: panic_on_warn set ... CPU: 1 PID: 6768 Comm: syz-executor1 Not tainted 4.9.0 #5 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 ffff8801c063ecd8 ffffffff82346bdf ffffffff00000001 1ffff100380c7d2e ffffed00380c7d26 0000000041b58ab3 ffffffff84b37e38 ffffffff823468f1 ffffffff84820740 ffffffff84f289c0 dffffc0000000000 ffff8801c063ee20 Call Trace: [<ffffffff82346bdf>] __dump_stack lib/dump_stack.c:15 [inline] [<ffffffff82346bdf>] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 [<ffffffff81827e34>] panic+0x1fb/0x412 kernel/panic.c:179 [<ffffffff8141f704>] __warn+0x1c4/0x1e0 kernel/panic.c:542 [<ffffffff8141f7e5>] warn_slowpath_fmt+0xc5/0x100 kernel/panic.c:565 [<ffffffff8356cbaf>] skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 [<ffffffff83585cd2>] __skb_gso_segment+0x482/0x780 net/core/dev.c:2706 [<ffffffff83586f19>] skb_gso_segment include/linux/netdevice.h:3985 [inline] [<ffffffff83586f19>] validate_xmit_skb+0x5c9/0xc20 net/core/dev.c:2969 [<ffffffff835892bb>] __dev_queue_xmit+0xe6b/0x1e70 net/core/dev.c:3383 [<ffffffff8358a2d7>] dev_queue_xmit+0x17/0x20 net/core/dev.c:3424 [<ffffffff83ad161d>] packet_snd net/packet/af_packet.c:2930 [inline] [<ffffffff83ad161d>] packet_sendmsg+0x32ed/0x4d30 net/packet/af_packet.c:2955 [<ffffffff834f0aaa>] sock_sendmsg_nosec net/socket.c:621 [inline] [<ffffffff834f0aaa>] sock_sendmsg+0xca/0x110 net/socket.c:631 [<ffffffff834f329a>] ___sys_sendmsg+0x8fa/0x9f0 net/socket.c:1954 [<ffffffff834f5e58>] __sys_sendmsg+0x138/0x300 net/socket.c:1988 [<ffffffff834f604d>] SYSC_sendmsg net/socket.c:1999 [inline] [<ffffffff834f604d>] SyS_sendmsg+0x2d/0x50 net/socket.c:1995 [<ffffffff84371941>] entry_SYSCALL_64_fastpath+0x1f/0xc2 Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Willem de Bruijn <willemb@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-01-31 10:20:32 -08:00
net: reduce skb_warn_bad_offload() noise Dmitry reported warnings occurring in __skb_gso_segment() [1] All SKB_GSO_DODGY producers can allow user space to feed packets that trigger the current check. We could prevent them from doing so, rejecting packets, but this might add regressions to existing programs. It turns out our SKB_GSO_DODGY handlers properly set up checksum information that is needed anyway when packets needs to be segmented. By checking again skb_needs_check() after skb_mac_gso_segment(), we should remove these pesky warnings, at a very minor cost. With help from Willem de Bruijn [1] WARNING: CPU: 1 PID: 6768 at net/core/dev.c:2439 skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 lo: caps=(0x000000a2803b7c69, 0x0000000000000000) len=138 data_len=0 gso_size=15883 gso_type=4 ip_summed=0 Kernel panic - not syncing: panic_on_warn set ... CPU: 1 PID: 6768 Comm: syz-executor1 Not tainted 4.9.0 #5 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 ffff8801c063ecd8 ffffffff82346bdf ffffffff00000001 1ffff100380c7d2e ffffed00380c7d26 0000000041b58ab3 ffffffff84b37e38 ffffffff823468f1 ffffffff84820740 ffffffff84f289c0 dffffc0000000000 ffff8801c063ee20 Call Trace: [<ffffffff82346bdf>] __dump_stack lib/dump_stack.c:15 [inline] [<ffffffff82346bdf>] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 [<ffffffff81827e34>] panic+0x1fb/0x412 kernel/panic.c:179 [<ffffffff8141f704>] __warn+0x1c4/0x1e0 kernel/panic.c:542 [<ffffffff8141f7e5>] warn_slowpath_fmt+0xc5/0x100 kernel/panic.c:565 [<ffffffff8356cbaf>] skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 [<ffffffff83585cd2>] __skb_gso_segment+0x482/0x780 net/core/dev.c:2706 [<ffffffff83586f19>] skb_gso_segment include/linux/netdevice.h:3985 [inline] [<ffffffff83586f19>] validate_xmit_skb+0x5c9/0xc20 net/core/dev.c:2969 [<ffffffff835892bb>] __dev_queue_xmit+0xe6b/0x1e70 net/core/dev.c:3383 [<ffffffff8358a2d7>] dev_queue_xmit+0x17/0x20 net/core/dev.c:3424 [<ffffffff83ad161d>] packet_snd net/packet/af_packet.c:2930 [inline] [<ffffffff83ad161d>] packet_sendmsg+0x32ed/0x4d30 net/packet/af_packet.c:2955 [<ffffffff834f0aaa>] sock_sendmsg_nosec net/socket.c:621 [inline] [<ffffffff834f0aaa>] sock_sendmsg+0xca/0x110 net/socket.c:631 [<ffffffff834f329a>] ___sys_sendmsg+0x8fa/0x9f0 net/socket.c:1954 [<ffffffff834f5e58>] __sys_sendmsg+0x138/0x300 net/socket.c:1988 [<ffffffff834f604d>] SYSC_sendmsg net/socket.c:1999 [inline] [<ffffffff834f604d>] SyS_sendmsg+0x2d/0x50 net/socket.c:1995 [<ffffffff84371941>] entry_SYSCALL_64_fastpath+0x1f/0xc2 Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Willem de Bruijn <willemb@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-01-31 10:20:32 -08:00
net: reduce skb_warn_bad_offload() noise Dmitry reported warnings occurring in __skb_gso_segment() [1] All SKB_GSO_DODGY producers can allow user space to feed packets that trigger the current check. We could prevent them from doing so, rejecting packets, but this might add regressions to existing programs. It turns out our SKB_GSO_DODGY handlers properly set up checksum information that is needed anyway when packets needs to be segmented. By checking again skb_needs_check() after skb_mac_gso_segment(), we should remove these pesky warnings, at a very minor cost. With help from Willem de Bruijn [1] WARNING: CPU: 1 PID: 6768 at net/core/dev.c:2439 skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 lo: caps=(0x000000a2803b7c69, 0x0000000000000000) len=138 data_len=0 gso_size=15883 gso_type=4 ip_summed=0 Kernel panic - not syncing: panic_on_warn set ... CPU: 1 PID: 6768 Comm: syz-executor1 Not tainted 4.9.0 #5 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 ffff8801c063ecd8 ffffffff82346bdf ffffffff00000001 1ffff100380c7d2e ffffed00380c7d26 0000000041b58ab3 ffffffff84b37e38 ffffffff823468f1 ffffffff84820740 ffffffff84f289c0 dffffc0000000000 ffff8801c063ee20 Call Trace: [<ffffffff82346bdf>] __dump_stack lib/dump_stack.c:15 [inline] [<ffffffff82346bdf>] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 [<ffffffff81827e34>] panic+0x1fb/0x412 kernel/panic.c:179 [<ffffffff8141f704>] __warn+0x1c4/0x1e0 kernel/panic.c:542 [<ffffffff8141f7e5>] warn_slowpath_fmt+0xc5/0x100 kernel/panic.c:565 [<ffffffff8356cbaf>] skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 [<ffffffff83585cd2>] __skb_gso_segment+0x482/0x780 net/core/dev.c:2706 [<ffffffff83586f19>] skb_gso_segment include/linux/netdevice.h:3985 [inline] [<ffffffff83586f19>] validate_xmit_skb+0x5c9/0xc20 net/core/dev.c:2969 [<ffffffff835892bb>] __dev_queue_xmit+0xe6b/0x1e70 net/core/dev.c:3383 [<ffffffff8358a2d7>] dev_queue_xmit+0x17/0x20 net/core/dev.c:3424 [<ffffffff83ad161d>] packet_snd net/packet/af_packet.c:2930 [inline] [<ffffffff83ad161d>] packet_sendmsg+0x32ed/0x4d30 net/packet/af_packet.c:2955 [<ffffffff834f0aaa>] sock_sendmsg_nosec net/socket.c:621 [inline] [<ffffffff834f0aaa>] sock_sendmsg+0xca/0x110 net/socket.c:631 [<ffffffff834f329a>] ___sys_sendmsg+0x8fa/0x9f0 net/socket.c:1954 [<ffffffff834f5e58>] __sys_sendmsg+0x138/0x300 net/socket.c:1988 [<ffffffff834f604d>] SYSC_sendmsg net/socket.c:1999 [inline] [<ffffffff834f604d>] SyS_sendmsg+0x2d/0x50 net/socket.c:1995 [<ffffffff84371941>] entry_SYSCALL_64_fastpath+0x1f/0xc2 Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Willem de Bruijn <willemb@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-01-31 10:20:32 -08:00
net: reduce skb_warn_bad_offload() noise Dmitry reported warnings occurring in __skb_gso_segment() [1] All SKB_GSO_DODGY producers can allow user space to feed packets that trigger the current check. We could prevent them from doing so, rejecting packets, but this might add regressions to existing programs. It turns out our SKB_GSO_DODGY handlers properly set up checksum information that is needed anyway when packets needs to be segmented. By checking again skb_needs_check() after skb_mac_gso_segment(), we should remove these pesky warnings, at a very minor cost. With help from Willem de Bruijn [1] WARNING: CPU: 1 PID: 6768 at net/core/dev.c:2439 skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 lo: caps=(0x000000a2803b7c69, 0x0000000000000000) len=138 data_len=0 gso_size=15883 gso_type=4 ip_summed=0 Kernel panic - not syncing: panic_on_warn set ... CPU: 1 PID: 6768 Comm: syz-executor1 Not tainted 4.9.0 #5 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 ffff8801c063ecd8 ffffffff82346bdf ffffffff00000001 1ffff100380c7d2e ffffed00380c7d26 0000000041b58ab3 ffffffff84b37e38 ffffffff823468f1 ffffffff84820740 ffffffff84f289c0 dffffc0000000000 ffff8801c063ee20 Call Trace: [<ffffffff82346bdf>] __dump_stack lib/dump_stack.c:15 [inline] [<ffffffff82346bdf>] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51 [<ffffffff81827e34>] panic+0x1fb/0x412 kernel/panic.c:179 [<ffffffff8141f704>] __warn+0x1c4/0x1e0 kernel/panic.c:542 [<ffffffff8141f7e5>] warn_slowpath_fmt+0xc5/0x100 kernel/panic.c:565 [<ffffffff8356cbaf>] skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434 [<ffffffff83585cd2>] __skb_gso_segment+0x482/0x780 net/core/dev.c:2706 [<ffffffff83586f19>] skb_gso_segment include/linux/netdevice.h:3985 [inline] [<ffffffff83586f19>] validate_xmit_skb+0x5c9/0xc20 net/core/dev.c:2969 [<ffffffff835892bb>] __dev_queue_xmit+0xe6b/0x1e70 net/core/dev.c:3383 [<ffffffff8358a2d7>] dev_queue_xmit+0x17/0x20 net/core/dev.c:3424 [<ffffffff83ad161d>] packet_snd net/packet/af_packet.c:2930 [inline] [<ffffffff83ad161d>] packet_sendmsg+0x32ed/0x4d30 net/packet/af_packet.c:2955 [<ffffffff834f0aaa>] sock_sendmsg_nosec net/socket.c:621 [inline] [<ffffffff834f0aaa>] sock_sendmsg+0xca/0x110 net/socket.c:631 [<ffffffff834f329a>] ___sys_sendmsg+0x8fa/0x9f0 net/socket.c:1954 [<ffffffff834f5e58>] __sys_sendmsg+0x138/0x300 net/socket.c:1988 [<ffffffff834f604d>] SYSC_sendmsg net/socket.c:1999 [inline] [<ffffffff834f604d>] SyS_sendmsg+0x2d/0x50 net/socket.c:1995 [<ffffffff84371941>] entry_SYSCALL_64_fastpath+0x1f/0xc2 Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Dmitry Vyukov <dvyukov@google.com> Cc: Willem de Bruijn <willemb@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-01-31 10:20:32 -08:00
net: do not disable sg for packets requiring no checksum A change in a series of VLAN-related changes appears to have inadvertently disabled the use of the scatter gather feature of network cards for transmission of non-IP ethernet protocols like ATA over Ethernet (AoE). Below is a reference to the commit that introduces a "harmonize_features" function that turns off scatter gather when the NIC does not support hardware checksumming for the ethernet protocol of an sk buff. commit f01a5236bd4b140198fbcc550f085e8361fd73fa Author: Jesse Gross <jesse@nicira.com> Date: Sun Jan 9 06:23:31 2011 +0000 net offloading: Generalize netif_get_vlan_features(). The can_checksum_protocol function is not equipped to consider a protocol that does not require checksumming. Calling it for a protocol that requires no checksum is inappropriate. The patch below has harmonize_features call can_checksum_protocol when the protocol needs a checksum, so that the network layer is not forced to perform unnecessary skb linearization on the transmission of AoE packets. Unnecessary linearization results in decreased performance and increased memory pressure, as reported here: http://www.spinics.net/lists/linux-mm/msg15184.html The problem has probably not been widely experienced yet, because only recently has the kernel.org-distributed aoe driver acquired the ability to use payloads of over a page in size, with the patchset recently included in the mm tree: https://lkml.org/lkml/2012/8/28/140 The coraid.com-distributed aoe driver already could use payloads of greater than a page in size, but its users generally do not use the newest kernels. Signed-off-by: Ed Cashin <ecashin@coraid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-19 15:49:00 +00:00
packet: on direct_xmit, limit tso and csum to supported devices When transmitting on a packet socket with PACKET_VNET_HDR and PACKET_QDISC_BYPASS, validate device support for features requested in vnet_hdr. Drop TSO packets sent to devices that do not support TSO or have the feature disabled. Note that the latter currently do process those packets correctly, regardless of not advertising the feature. Because of SKB_GSO_DODGY, it is not sufficient to test device features with netif_needs_gso. Full validate_xmit_skb is needed. Switch to software checksum for non-TSO packets that request checksum offload if that device feature is unsupported or disabled. Note that similar to the TSO case, device drivers may perform checksum offload correctly even when not advertising it. When switching to software checksum, packets hit skb_checksum_help, which has two BUG_ON checksum not in linear segment. Packet sockets always allocate at least up to csum_start + csum_off + 2 as linear. Tested by running github.com/wdebruij/kerneltools/psock_txring_vnet.c ethtool -K eth0 tso off tx on psock_txring_vnet -d $dst -s $src -i eth0 -l 2000 -n 1 -q -v psock_txring_vnet -d $dst -s $src -i eth0 -l 2000 -n 1 -q -v -N ethtool -K eth0 tx off psock_txring_vnet -d $dst -s $src -i eth0 -l 1000 -n 1 -q -v -G psock_txring_vnet -d $dst -s $src -i eth0 -l 1000 -n 1 -q -v -G -N v2: - add EXPORT_SYMBOL_GPL(validate_xmit_skb_list) Fixes: d346a3fae3ff ("packet: introduce PACKET_QDISC_BYPASS socket option") Signed-off-by: Willem de Bruijn <willemb@google.com> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-10-26 11:23:07 -04:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net: core: explicitly select a txq before doing l2 forwarding Currently, the tx queue were selected implicitly in ndo_dfwd_start_xmit(). The will cause several issues: - NETIF_F_LLTX were removed for macvlan, so txq lock were done for macvlan instead of lower device which misses the necessary txq synchronization for lower device such as txq stopping or frozen required by dev watchdog or control path. - dev_hard_start_xmit() was called with NULL txq which bypasses the net device watchdog. - dev_hard_start_xmit() does not check txq everywhere which will lead a crash when tso is disabled for lower device. Fix this by explicitly introducing a new param for .ndo_select_queue() for just selecting queues in the case of l2 forwarding offload. netdev_pick_tx() was also extended to accept this parameter and dev_queue_xmit_accel() was used to do l2 forwarding transmission. With this fixes, NETIF_F_LLTX could be preserved for macvlan and there's no need to check txq against NULL in dev_hard_start_xmit(). Also there's no need to keep a dedicated ndo_dfwd_start_xmit() and we can just reuse the code of dev_queue_xmit() to do the transmission. In the future, it was also required for macvtap l2 forwarding support since it provides a necessary synchronization method. Cc: John Fastabend <john.r.fastabend@intel.com> Cc: Neil Horman <nhorman@tuxdriver.com> Cc: e1000-devel@lists.sourceforge.net Signed-off-by: Jason Wang <jasowang@redhat.com> Acked-by: Neil Horman <nhorman@tuxdriver.com> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-10 16:18:26 +08:00
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
net: core: explicitly select a txq before doing l2 forwarding Currently, the tx queue were selected implicitly in ndo_dfwd_start_xmit(). The will cause several issues: - NETIF_F_LLTX were removed for macvlan, so txq lock were done for macvlan instead of lower device which misses the necessary txq synchronization for lower device such as txq stopping or frozen required by dev watchdog or control path. - dev_hard_start_xmit() was called with NULL txq which bypasses the net device watchdog. - dev_hard_start_xmit() does not check txq everywhere which will lead a crash when tso is disabled for lower device. Fix this by explicitly introducing a new param for .ndo_select_queue() for just selecting queues in the case of l2 forwarding offload. netdev_pick_tx() was also extended to accept this parameter and dev_queue_xmit_accel() was used to do l2 forwarding transmission. With this fixes, NETIF_F_LLTX could be preserved for macvlan and there's no need to check txq against NULL in dev_hard_start_xmit(). Also there's no need to keep a dedicated ndo_dfwd_start_xmit() and we can just reuse the code of dev_queue_xmit() to do the transmission. In the future, it was also required for macvtap l2 forwarding support since it provides a necessary synchronization method. Cc: John Fastabend <john.r.fastabend@intel.com> Cc: Neil Horman <nhorman@tuxdriver.com> Cc: e1000-devel@lists.sourceforge.net Signed-off-by: Jason Wang <jasowang@redhat.com> Acked-by: Neil Horman <nhorman@tuxdriver.com> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-10 16:18:26 +08:00
net: core: explicitly select a txq before doing l2 forwarding Currently, the tx queue were selected implicitly in ndo_dfwd_start_xmit(). The will cause several issues: - NETIF_F_LLTX were removed for macvlan, so txq lock were done for macvlan instead of lower device which misses the necessary txq synchronization for lower device such as txq stopping or frozen required by dev watchdog or control path. - dev_hard_start_xmit() was called with NULL txq which bypasses the net device watchdog. - dev_hard_start_xmit() does not check txq everywhere which will lead a crash when tso is disabled for lower device. Fix this by explicitly introducing a new param for .ndo_select_queue() for just selecting queues in the case of l2 forwarding offload. netdev_pick_tx() was also extended to accept this parameter and dev_queue_xmit_accel() was used to do l2 forwarding transmission. With this fixes, NETIF_F_LLTX could be preserved for macvlan and there's no need to check txq against NULL in dev_hard_start_xmit(). Also there's no need to keep a dedicated ndo_dfwd_start_xmit() and we can just reuse the code of dev_queue_xmit() to do the transmission. In the future, it was also required for macvtap l2 forwarding support since it provides a necessary synchronization method. Cc: John Fastabend <john.r.fastabend@intel.com> Cc: Neil Horman <nhorman@tuxdriver.com> Cc: e1000-devel@lists.sourceforge.net Signed-off-by: Jason Wang <jasowang@redhat.com> Acked-by: Neil Horman <nhorman@tuxdriver.com> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-10 16:18:26 +08:00
net: core: explicitly select a txq before doing l2 forwarding Currently, the tx queue were selected implicitly in ndo_dfwd_start_xmit(). The will cause several issues: - NETIF_F_LLTX were removed for macvlan, so txq lock were done for macvlan instead of lower device which misses the necessary txq synchronization for lower device such as txq stopping or frozen required by dev watchdog or control path. - dev_hard_start_xmit() was called with NULL txq which bypasses the net device watchdog. - dev_hard_start_xmit() does not check txq everywhere which will lead a crash when tso is disabled for lower device. Fix this by explicitly introducing a new param for .ndo_select_queue() for just selecting queues in the case of l2 forwarding offload. netdev_pick_tx() was also extended to accept this parameter and dev_queue_xmit_accel() was used to do l2 forwarding transmission. With this fixes, NETIF_F_LLTX could be preserved for macvlan and there's no need to check txq against NULL in dev_hard_start_xmit(). Also there's no need to keep a dedicated ndo_dfwd_start_xmit() and we can just reuse the code of dev_queue_xmit() to do the transmission. In the future, it was also required for macvtap l2 forwarding support since it provides a necessary synchronization method. Cc: John Fastabend <john.r.fastabend@intel.com> Cc: Neil Horman <nhorman@tuxdriver.com> Cc: e1000-devel@lists.sourceforge.net Signed-off-by: Jason Wang <jasowang@redhat.com> Acked-by: Neil Horman <nhorman@tuxdriver.com> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-10 16:18:26 +08:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
locking/atomics: COCCINELLE/treewide: Convert trivial ACCESS_ONCE() patterns to READ_ONCE()/WRITE_ONCE() Please do not apply this to mainline directly, instead please re-run the coccinelle script shown below and apply its output. For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't harmful, and changing them results in churn. However, for some features, the read/write distinction is critical to correct operation. To distinguish these cases, separate read/write accessors must be used. This patch migrates (most) remaining ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following coccinelle script: ---- // Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and // WRITE_ONCE() // $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: tj@kernel.org Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 14:07:29 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
net: do not process device backlog during unregistration commit 381c759d9916 ("ipv4: Avoid crashing in ip_error") fixes a problem where processed packet comes from device with destroyed inetdev (dev->ip_ptr). This is not expected because inetdev_destroy is called in NETDEV_UNREGISTER phase and packets should not be processed after dev_close_many() and synchronize_net(). Above fix is still required because inetdev_destroy can be called for other reasons. But it shows the real problem: backlog can keep packets for long time and they do not hold reference to device. Such packets are then delivered to upper levels at the same time when device is unregistered. Calling flush_backlog after NETDEV_UNREGISTER_FINAL still accounts all packets from backlog but before that some packets continue to be delivered to upper levels long after the synchronize_net call which is supposed to wait the last ones. Also, as Eric pointed out, processed packets, mostly from other devices, can continue to add new packets to backlog. Fix the problem by moving flush_backlog early, after the device driver is stopped and before the synchronize_net() call. Then use netif_running check to make sure we do not add more packets to backlog. We have to do it in enqueue_to_backlog context when the local IRQ is disabled. As result, after the flush_backlog and synchronize_net sequence all packets should be accounted. Thanks to Eric W. Biederman for the test script and his valuable feedback! Reported-by: Vittorio Gambaletta <linuxbugs@vittgam.net> Fixes: 6e583ce5242f ("net: eliminate refcounting in backlog queue") Cc: Eric W. Biederman <ebiederm@xmission.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Signed-off-by: Julian Anastasov <ja@ssi.bg> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-07-09 09:59:09 +03:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
net: do not process device backlog during unregistration commit 381c759d9916 ("ipv4: Avoid crashing in ip_error") fixes a problem where processed packet comes from device with destroyed inetdev (dev->ip_ptr). This is not expected because inetdev_destroy is called in NETDEV_UNREGISTER phase and packets should not be processed after dev_close_many() and synchronize_net(). Above fix is still required because inetdev_destroy can be called for other reasons. But it shows the real problem: backlog can keep packets for long time and they do not hold reference to device. Such packets are then delivered to upper levels at the same time when device is unregistered. Calling flush_backlog after NETDEV_UNREGISTER_FINAL still accounts all packets from backlog but before that some packets continue to be delivered to upper levels long after the synchronize_net call which is supposed to wait the last ones. Also, as Eric pointed out, processed packets, mostly from other devices, can continue to add new packets to backlog. Fix the problem by moving flush_backlog early, after the device driver is stopped and before the synchronize_net() call. Then use netif_running check to make sure we do not add more packets to backlog. We have to do it in enqueue_to_backlog context when the local IRQ is disabled. As result, after the flush_backlog and synchronize_net sequence all packets should be accounted. Thanks to Eric W. Biederman for the test script and his valuable feedback! Reported-by: Vittorio Gambaletta <linuxbugs@vittgam.net> Fixes: 6e583ce5242f ("net: eliminate refcounting in backlog queue") Cc: Eric W. Biederman <ebiederm@xmission.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Signed-off-by: Julian Anastasov <ja@ssi.bg> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-07-09 09:59:09 +03:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
bpf: add meta pointer for direct access This work enables generic transfer of metadata from XDP into skb. The basic idea is that we can make use of the fact that the resulting skb must be linear and already comes with a larger headroom for supporting bpf_xdp_adjust_head(), which mangles xdp->data. Here, we base our work on a similar principle and introduce a small helper bpf_xdp_adjust_meta() for adjusting a new pointer called xdp->data_meta. Thus, the packet has a flexible and programmable room for meta data, followed by the actual packet data. struct xdp_buff is therefore laid out that we first point to data_hard_start, then data_meta directly prepended to data followed by data_end marking the end of packet. bpf_xdp_adjust_head() takes into account whether we have meta data already prepended and if so, memmove()s this along with the given offset provided there's enough room. xdp->data_meta is optional and programs are not required to use it. The rationale is that when we process the packet in XDP (e.g. as DoS filter), we can push further meta data along with it for the XDP_PASS case, and give the guarantee that a clsact ingress BPF program on the same device can pick this up for further post-processing. Since we work with skb there, we can also set skb->mark, skb->priority or other skb meta data out of BPF, thus having this scratch space generic and programmable allows for more flexibility than defining a direct 1:1 transfer of potentially new XDP members into skb (it's also more efficient as we don't need to initialize/handle each of such new members). The facility also works together with GRO aggregation. The scratch space at the head of the packet can be multiple of 4 byte up to 32 byte large. Drivers not yet supporting xdp->data_meta can simply be set up with xdp->data_meta as xdp->data + 1 as bpf_xdp_adjust_meta() will detect this and bail out, such that the subsequent match against xdp->data for later access is guaranteed to fail. The verifier treats xdp->data_meta/xdp->data the same way as we treat xdp->data/xdp->data_end pointer comparisons. The requirement for doing the compare against xdp->data is that it hasn't been modified from it's original address we got from ctx access. It may have a range marking already from prior successful xdp->data/xdp->data_end pointer comparisons though. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-25 02:25:51 +02:00
bpf: add meta pointer for direct access This work enables generic transfer of metadata from XDP into skb. The basic idea is that we can make use of the fact that the resulting skb must be linear and already comes with a larger headroom for supporting bpf_xdp_adjust_head(), which mangles xdp->data. Here, we base our work on a similar principle and introduce a small helper bpf_xdp_adjust_meta() for adjusting a new pointer called xdp->data_meta. Thus, the packet has a flexible and programmable room for meta data, followed by the actual packet data. struct xdp_buff is therefore laid out that we first point to data_hard_start, then data_meta directly prepended to data followed by data_end marking the end of packet. bpf_xdp_adjust_head() takes into account whether we have meta data already prepended and if so, memmove()s this along with the given offset provided there's enough room. xdp->data_meta is optional and programs are not required to use it. The rationale is that when we process the packet in XDP (e.g. as DoS filter), we can push further meta data along with it for the XDP_PASS case, and give the guarantee that a clsact ingress BPF program on the same device can pick this up for further post-processing. Since we work with skb there, we can also set skb->mark, skb->priority or other skb meta data out of BPF, thus having this scratch space generic and programmable allows for more flexibility than defining a direct 1:1 transfer of potentially new XDP members into skb (it's also more efficient as we don't need to initialize/handle each of such new members). The facility also works together with GRO aggregation. The scratch space at the head of the packet can be multiple of 4 byte up to 32 byte large. Drivers not yet supporting xdp->data_meta can simply be set up with xdp->data_meta as xdp->data + 1 as bpf_xdp_adjust_meta() will detect this and bail out, such that the subsequent match against xdp->data for later access is guaranteed to fail. The verifier treats xdp->data_meta/xdp->data the same way as we treat xdp->data/xdp->data_end pointer comparisons. The requirement for doing the compare against xdp->data is that it hasn't been modified from it's original address we got from ctx access. It may have a range marking already from prior successful xdp->data/xdp->data_end pointer comparisons though. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-25 02:25:51 +02:00
bpf: add meta pointer for direct access This work enables generic transfer of metadata from XDP into skb. The basic idea is that we can make use of the fact that the resulting skb must be linear and already comes with a larger headroom for supporting bpf_xdp_adjust_head(), which mangles xdp->data. Here, we base our work on a similar principle and introduce a small helper bpf_xdp_adjust_meta() for adjusting a new pointer called xdp->data_meta. Thus, the packet has a flexible and programmable room for meta data, followed by the actual packet data. struct xdp_buff is therefore laid out that we first point to data_hard_start, then data_meta directly prepended to data followed by data_end marking the end of packet. bpf_xdp_adjust_head() takes into account whether we have meta data already prepended and if so, memmove()s this along with the given offset provided there's enough room. xdp->data_meta is optional and programs are not required to use it. The rationale is that when we process the packet in XDP (e.g. as DoS filter), we can push further meta data along with it for the XDP_PASS case, and give the guarantee that a clsact ingress BPF program on the same device can pick this up for further post-processing. Since we work with skb there, we can also set skb->mark, skb->priority or other skb meta data out of BPF, thus having this scratch space generic and programmable allows for more flexibility than defining a direct 1:1 transfer of potentially new XDP members into skb (it's also more efficient as we don't need to initialize/handle each of such new members). The facility also works together with GRO aggregation. The scratch space at the head of the packet can be multiple of 4 byte up to 32 byte large. Drivers not yet supporting xdp->data_meta can simply be set up with xdp->data_meta as xdp->data + 1 as bpf_xdp_adjust_meta() will detect this and bail out, such that the subsequent match against xdp->data for later access is guaranteed to fail. The verifier treats xdp->data_meta/xdp->data the same way as we treat xdp->data/xdp->data_end pointer comparisons. The requirement for doing the compare against xdp->data is that it hasn't been modified from it's original address we got from ctx access. It may have a range marking already from prior successful xdp->data/xdp->data_end pointer comparisons though. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-25 02:25:51 +02:00
bpf: add meta pointer for direct access This work enables generic transfer of metadata from XDP into skb. The basic idea is that we can make use of the fact that the resulting skb must be linear and already comes with a larger headroom for supporting bpf_xdp_adjust_head(), which mangles xdp->data. Here, we base our work on a similar principle and introduce a small helper bpf_xdp_adjust_meta() for adjusting a new pointer called xdp->data_meta. Thus, the packet has a flexible and programmable room for meta data, followed by the actual packet data. struct xdp_buff is therefore laid out that we first point to data_hard_start, then data_meta directly prepended to data followed by data_end marking the end of packet. bpf_xdp_adjust_head() takes into account whether we have meta data already prepended and if so, memmove()s this along with the given offset provided there's enough room. xdp->data_meta is optional and programs are not required to use it. The rationale is that when we process the packet in XDP (e.g. as DoS filter), we can push further meta data along with it for the XDP_PASS case, and give the guarantee that a clsact ingress BPF program on the same device can pick this up for further post-processing. Since we work with skb there, we can also set skb->mark, skb->priority or other skb meta data out of BPF, thus having this scratch space generic and programmable allows for more flexibility than defining a direct 1:1 transfer of potentially new XDP members into skb (it's also more efficient as we don't need to initialize/handle each of such new members). The facility also works together with GRO aggregation. The scratch space at the head of the packet can be multiple of 4 byte up to 32 byte large. Drivers not yet supporting xdp->data_meta can simply be set up with xdp->data_meta as xdp->data + 1 as bpf_xdp_adjust_meta() will detect this and bail out, such that the subsequent match against xdp->data for later access is guaranteed to fail. The verifier treats xdp->data_meta/xdp->data the same way as we treat xdp->data/xdp->data_end pointer comparisons. The requirement for doing the compare against xdp->data is that it hasn't been modified from it's original address we got from ctx access. It may have a range marking already from prior successful xdp->data/xdp->data_end pointer comparisons though. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-25 02:25:51 +02:00
bpf: add meta pointer for direct access This work enables generic transfer of metadata from XDP into skb. The basic idea is that we can make use of the fact that the resulting skb must be linear and already comes with a larger headroom for supporting bpf_xdp_adjust_head(), which mangles xdp->data. Here, we base our work on a similar principle and introduce a small helper bpf_xdp_adjust_meta() for adjusting a new pointer called xdp->data_meta. Thus, the packet has a flexible and programmable room for meta data, followed by the actual packet data. struct xdp_buff is therefore laid out that we first point to data_hard_start, then data_meta directly prepended to data followed by data_end marking the end of packet. bpf_xdp_adjust_head() takes into account whether we have meta data already prepended and if so, memmove()s this along with the given offset provided there's enough room. xdp->data_meta is optional and programs are not required to use it. The rationale is that when we process the packet in XDP (e.g. as DoS filter), we can push further meta data along with it for the XDP_PASS case, and give the guarantee that a clsact ingress BPF program on the same device can pick this up for further post-processing. Since we work with skb there, we can also set skb->mark, skb->priority or other skb meta data out of BPF, thus having this scratch space generic and programmable allows for more flexibility than defining a direct 1:1 transfer of potentially new XDP members into skb (it's also more efficient as we don't need to initialize/handle each of such new members). The facility also works together with GRO aggregation. The scratch space at the head of the packet can be multiple of 4 byte up to 32 byte large. Drivers not yet supporting xdp->data_meta can simply be set up with xdp->data_meta as xdp->data + 1 as bpf_xdp_adjust_meta() will detect this and bail out, such that the subsequent match against xdp->data for later access is guaranteed to fail. The verifier treats xdp->data_meta/xdp->data the same way as we treat xdp->data/xdp->data_end pointer comparisons. The requirement for doing the compare against xdp->data is that it hasn't been modified from it's original address we got from ctx access. It may have a range marking already from prior successful xdp->data/xdp->data_end pointer comparisons though. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-25 02:25:51 +02:00
net: rcu lock and preempt disable missing around generic xdp do_xdp_generic must be called inside rcu critical section with preempt disabled to ensure BPF programs are valid and per-cpu variables used for redirect operations are consistent. This patch ensures this is true and fixes the splat below. The netif_receive_skb_internal() code path is now broken into two rcu critical sections. I decided it was better to limit the preempt_enable/disable block to just the xdp static key portion and the fallout is more rcu_read_lock/unlock calls. Seems like the best option to me. [ 607.596901] ============================= [ 607.596906] WARNING: suspicious RCU usage [ 607.596912] 4.13.0-rc4+ #570 Not tainted [ 607.596917] ----------------------------- [ 607.596923] net/core/dev.c:3948 suspicious rcu_dereference_check() usage! [ 607.596927] [ 607.596927] other info that might help us debug this: [ 607.596927] [ 607.596933] [ 607.596933] rcu_scheduler_active = 2, debug_locks = 1 [ 607.596938] 2 locks held by pool/14624: [ 607.596943] #0: (rcu_read_lock_bh){......}, at: [<ffffffff95445ffd>] ip_finish_output2+0x14d/0x890 [ 607.596973] #1: (rcu_read_lock_bh){......}, at: [<ffffffff953c8e3a>] __dev_queue_xmit+0x14a/0xfd0 [ 607.597000] [ 607.597000] stack backtrace: [ 607.597006] CPU: 5 PID: 14624 Comm: pool Not tainted 4.13.0-rc4+ #570 [ 607.597011] Hardware name: Dell Inc. Precision Tower 5810/0HHV7N, BIOS A17 03/01/2017 [ 607.597016] Call Trace: [ 607.597027] dump_stack+0x67/0x92 [ 607.597040] lockdep_rcu_suspicious+0xdd/0x110 [ 607.597054] do_xdp_generic+0x313/0xa50 [ 607.597068] ? time_hardirqs_on+0x5b/0x150 [ 607.597076] ? mark_held_locks+0x6b/0xc0 [ 607.597088] ? netdev_pick_tx+0x150/0x150 [ 607.597117] netif_rx_internal+0x205/0x3f0 [ 607.597127] ? do_xdp_generic+0xa50/0xa50 [ 607.597144] ? lock_downgrade+0x2b0/0x2b0 [ 607.597158] ? __lock_is_held+0x93/0x100 [ 607.597187] netif_rx+0x119/0x190 [ 607.597202] loopback_xmit+0xfd/0x1b0 [ 607.597214] dev_hard_start_xmit+0x127/0x4e0 Fixes: d445516966dc ("net: xdp: support xdp generic on virtual devices") Fixes: b5cdae3291f7 ("net: Generic XDP") Acked-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-08 14:00:30 -07:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
net: sched: further simplify handle_ing Ingress qdisc has no other purpose than calling into tc_classify() that executes attached classifier(s) and action(s). It has a 1:1 relationship to dev->ingress_queue. After having commit 087c1a601ad7 ("net: sched: run ingress qdisc without locks") removed the central ingress lock, one major contention point is gone. The extra indirection layers however, are not necessary for calling into ingress qdisc. pktgen calling locally into netif_receive_skb() with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps. We can redirect the private classifier list to the netdev directly, without changing any classifier API bits (!) and execute on that from handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed, ingress qdisc doesn't have a queue and thus dev_deactivate_queue() is also not applicable, ingress_cl_list provides similar behaviour. In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc. One next possible step is the removal of the dev's ingress (dummy) netdev_queue, and to only have the list member in the netdevice itself. Note, the filter chain is RCU protected and individual filter elements are being kfree'd by sched subsystem after RCU grace period. RCU read lock is being held by __netif_receive_skb_core(). Joint work with Alexei Starovoitov. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-09 22:51:32 +02:00
net: sched: further simplify handle_ing Ingress qdisc has no other purpose than calling into tc_classify() that executes attached classifier(s) and action(s). It has a 1:1 relationship to dev->ingress_queue. After having commit 087c1a601ad7 ("net: sched: run ingress qdisc without locks") removed the central ingress lock, one major contention point is gone. The extra indirection layers however, are not necessary for calling into ingress qdisc. pktgen calling locally into netif_receive_skb() with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps. We can redirect the private classifier list to the netdev directly, without changing any classifier API bits (!) and execute on that from handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed, ingress qdisc doesn't have a queue and thus dev_deactivate_queue() is also not applicable, ingress_cl_list provides similar behaviour. In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc. One next possible step is the removal of the dev's ingress (dummy) netdev_queue, and to only have the list member in the netdevice itself. Note, the filter chain is RCU protected and individual filter elements are being kfree'd by sched subsystem after RCU grace period. RCU read lock is being held by __netif_receive_skb_core(). Joint work with Alexei Starovoitov. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-09 22:51:32 +02:00
2015-05-14 00:36:28 +02:00
net: sched: further simplify handle_ing Ingress qdisc has no other purpose than calling into tc_classify() that executes attached classifier(s) and action(s). It has a 1:1 relationship to dev->ingress_queue. After having commit 087c1a601ad7 ("net: sched: run ingress qdisc without locks") removed the central ingress lock, one major contention point is gone. The extra indirection layers however, are not necessary for calling into ingress qdisc. pktgen calling locally into netif_receive_skb() with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps. We can redirect the private classifier list to the netdev directly, without changing any classifier API bits (!) and execute on that from handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed, ingress qdisc doesn't have a queue and thus dev_deactivate_queue() is also not applicable, ingress_cl_list provides similar behaviour. In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc. One next possible step is the removal of the dev's ingress (dummy) netdev_queue, and to only have the list member in the netdevice itself. Note, the filter chain is RCU protected and individual filter elements are being kfree'd by sched subsystem after RCU grace period. RCU read lock is being held by __netif_receive_skb_core(). Joint work with Alexei Starovoitov. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-09 22:51:32 +02:00
net: sched: further simplify handle_ing Ingress qdisc has no other purpose than calling into tc_classify() that executes attached classifier(s) and action(s). It has a 1:1 relationship to dev->ingress_queue. After having commit 087c1a601ad7 ("net: sched: run ingress qdisc without locks") removed the central ingress lock, one major contention point is gone. The extra indirection layers however, are not necessary for calling into ingress qdisc. pktgen calling locally into netif_receive_skb() with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps. We can redirect the private classifier list to the netdev directly, without changing any classifier API bits (!) and execute on that from handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed, ingress qdisc doesn't have a queue and thus dev_deactivate_queue() is also not applicable, ingress_cl_list provides similar behaviour. In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc. One next possible step is the removal of the dev's ingress (dummy) netdev_queue, and to only have the list member in the netdevice itself. Note, the filter chain is RCU protected and individual filter elements are being kfree'd by sched subsystem after RCU grace period. RCU read lock is being held by __netif_receive_skb_core(). Joint work with Alexei Starovoitov. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-09 22:51:32 +02:00
net: sched: further simplify handle_ing Ingress qdisc has no other purpose than calling into tc_classify() that executes attached classifier(s) and action(s). It has a 1:1 relationship to dev->ingress_queue. After having commit 087c1a601ad7 ("net: sched: run ingress qdisc without locks") removed the central ingress lock, one major contention point is gone. The extra indirection layers however, are not necessary for calling into ingress qdisc. pktgen calling locally into netif_receive_skb() with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps. We can redirect the private classifier list to the netdev directly, without changing any classifier API bits (!) and execute on that from handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed, ingress qdisc doesn't have a queue and thus dev_deactivate_queue() is also not applicable, ingress_cl_list provides similar behaviour. In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc. One next possible step is the removal of the dev's ingress (dummy) netdev_queue, and to only have the list member in the netdevice itself. Note, the filter chain is RCU protected and individual filter elements are being kfree'd by sched subsystem after RCU grace period. RCU read lock is being held by __netif_receive_skb_core(). Joint work with Alexei Starovoitov. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-09 22:51:32 +02:00
2015-09-15 23:05:43 -07:00
net: sched: further simplify handle_ing Ingress qdisc has no other purpose than calling into tc_classify() that executes attached classifier(s) and action(s). It has a 1:1 relationship to dev->ingress_queue. After having commit 087c1a601ad7 ("net: sched: run ingress qdisc without locks") removed the central ingress lock, one major contention point is gone. The extra indirection layers however, are not necessary for calling into ingress qdisc. pktgen calling locally into netif_receive_skb() with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps. We can redirect the private classifier list to the netdev directly, without changing any classifier API bits (!) and execute on that from handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed, ingress qdisc doesn't have a queue and thus dev_deactivate_queue() is also not applicable, ingress_cl_list provides similar behaviour. In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc. One next possible step is the removal of the dev's ingress (dummy) netdev_queue, and to only have the list member in the netdevice itself. Note, the filter chain is RCU protected and individual filter elements are being kfree'd by sched subsystem after RCU grace period. RCU read lock is being held by __netif_receive_skb_core(). Joint work with Alexei Starovoitov. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-09 22:51:32 +02:00
net: add a synchronize_net() in netdev_rx_handler_unregister() commit 35d48903e97819 (bonding: fix rx_handler locking) added a race in bonding driver, reported by Steven Rostedt who did a very good diagnosis : <quoting Steven> I'm currently debugging a crash in an old 3.0-rt kernel that one of our customers is seeing. The bug happens with a stress test that loads and unloads the bonding module in a loop (I don't know all the details as I'm not the one that is directly interacting with the customer). But the bug looks to be something that may still be present and possibly present in mainline too. It will just be much harder to trigger it in mainline. In -rt, interrupts are threads, and can schedule in and out just like any other thread. Note, mainline now supports interrupt threads so this may be easily reproducible in mainline as well. I don't have the ability to tell the customer to try mainline or other kernels, so my hands are somewhat tied to what I can do. But according to a core dump, I tracked down that the eth irq thread crashed in bond_handle_frame() here: slave = bond_slave_get_rcu(skb->dev); bond = slave->bond; <--- BUG the slave returned was NULL and accessing slave->bond caused a NULL pointer dereference. Looking at the code that unregisters the handler: void netdev_rx_handler_unregister(struct net_device *dev) { ASSERT_RTNL(); RCU_INIT_POINTER(dev->rx_handler, NULL); RCU_INIT_POINTER(dev->rx_handler_data, NULL); } Which is basically: dev->rx_handler = NULL; dev->rx_handler_data = NULL; And looking at __netif_receive_skb() we have: rx_handler = rcu_dereference(skb->dev->rx_handler); if (rx_handler) { if (pt_prev) { ret = deliver_skb(skb, pt_prev, orig_dev); pt_prev = NULL; } switch (rx_handler(&skb)) { My question to all of you is, what stops this interrupt from happening while the bonding module is unloading? What happens if the interrupt triggers and we have this: CPU0 CPU1 ---- ---- rx_handler = skb->dev->rx_handler netdev_rx_handler_unregister() { dev->rx_handler = NULL; dev->rx_handler_data = NULL; rx_handler() bond_handle_frame() { slave = skb->dev->rx_handler; bond = slave->bond; <-- NULL pointer dereference!!! What protection am I missing in the bond release handler that would prevent the above from happening? </quoting Steven> We can fix bug this in two ways. First is adding a test in bond_handle_frame() and others to check if rx_handler_data is NULL. A second way is adding a synchronize_net() in netdev_rx_handler_unregister() to make sure that a rcu protected reader has the guarantee to see a non NULL rx_handler_data. The second way is better as it avoids an extra test in fast path. Reported-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Jiri Pirko <jpirko@redhat.com> Cc: Paul E. McKenney <paulmck@us.ibm.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-29 03:01:22 +00:00
net: add a synchronize_net() in netdev_rx_handler_unregister() commit 35d48903e97819 (bonding: fix rx_handler locking) added a race in bonding driver, reported by Steven Rostedt who did a very good diagnosis : <quoting Steven> I'm currently debugging a crash in an old 3.0-rt kernel that one of our customers is seeing. The bug happens with a stress test that loads and unloads the bonding module in a loop (I don't know all the details as I'm not the one that is directly interacting with the customer). But the bug looks to be something that may still be present and possibly present in mainline too. It will just be much harder to trigger it in mainline. In -rt, interrupts are threads, and can schedule in and out just like any other thread. Note, mainline now supports interrupt threads so this may be easily reproducible in mainline as well. I don't have the ability to tell the customer to try mainline or other kernels, so my hands are somewhat tied to what I can do. But according to a core dump, I tracked down that the eth irq thread crashed in bond_handle_frame() here: slave = bond_slave_get_rcu(skb->dev); bond = slave->bond; <--- BUG the slave returned was NULL and accessing slave->bond caused a NULL pointer dereference. Looking at the code that unregisters the handler: void netdev_rx_handler_unregister(struct net_device *dev) { ASSERT_RTNL(); RCU_INIT_POINTER(dev->rx_handler, NULL); RCU_INIT_POINTER(dev->rx_handler_data, NULL); } Which is basically: dev->rx_handler = NULL; dev->rx_handler_data = NULL; And looking at __netif_receive_skb() we have: rx_handler = rcu_dereference(skb->dev->rx_handler); if (rx_handler) { if (pt_prev) { ret = deliver_skb(skb, pt_prev, orig_dev); pt_prev = NULL; } switch (rx_handler(&skb)) { My question to all of you is, what stops this interrupt from happening while the bonding module is unloading? What happens if the interrupt triggers and we have this: CPU0 CPU1 ---- ---- rx_handler = skb->dev->rx_handler netdev_rx_handler_unregister() { dev->rx_handler = NULL; dev->rx_handler_data = NULL; rx_handler() bond_handle_frame() { slave = skb->dev->rx_handler; bond = slave->bond; <-- NULL pointer dereference!!! What protection am I missing in the bond release handler that would prevent the above from happening? </quoting Steven> We can fix bug this in two ways. First is adding a test in bond_handle_frame() and others to check if rx_handler_data is NULL. A second way is adding a synchronize_net() in netdev_rx_handler_unregister() to make sure that a rcu protected reader has the guarantee to see a non NULL rx_handler_data. The second way is better as it avoids an extra test in fast path. Reported-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Jiri Pirko <jpirko@redhat.com> Cc: Paul E. McKenney <paulmck@us.ibm.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-29 03:01:22 +00:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
net: Always untag vlan-tagged traffic on input. Currently the functionality to untag traffic on input resides as part of the vlan module and is build only when VLAN support is enabled in the kernel. When VLAN is disabled, the function vlan_untag() turns into a stub and doesn't really untag the packets. This seems to create an interesting interaction between VMs supporting checksum offloading and some network drivers. There are some drivers that do not allow the user to change tx-vlan-offload feature of the driver. These drivers also seem to assume that any VLAN-tagged traffic they transmit will have the vlan information in the vlan_tci and not in the vlan header already in the skb. When transmitting skbs that already have tagged data with partial checksum set, the checksum doesn't appear to be updated correctly by the card thus resulting in a failure to establish TCP connections. The following is a packet trace taken on the receiver where a sender is a VM with a VLAN configued. The host VM is running on doest not have VLAN support and the outging interface on the host is tg3: 10:12:43.503055 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q (0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27243, offset 0, flags [DF], proto TCP (6), length 60) 10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect -> 0x48d9), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val 4294837885 ecr 0,nop,wscale 7], length 0 10:12:44.505556 52:54:00:ae:42:3f > 28:d2:44:7d:c2:de, ethertype 802.1Q (0x8100), length 78: vlan 100, p 0, ethertype IPv4, (tos 0x0, ttl 64, id 27244, offset 0, flags [DF], proto TCP (6), length 60) 10.0.100.1.58545 > 10.0.100.10.ircu-2: Flags [S], cksum 0xdc39 (incorrect -> 0x44ee), seq 1069378582, win 29200, options [mss 1460,sackOK,TS val 4294838888 ecr 0,nop,wscale 7], length 0 This connection finally times out. I've only access to the TG3 hardware in this configuration thus have only tested this with TG3 driver. There are a lot of other drivers that do not permit user changes to vlan acceleration features, and I don't know if they all suffere from a similar issue. The patch attempt to fix this another way. It moves the vlan header stipping code out of the vlan module and always builds it into the kernel network core. This way, even if vlan is not supported on a virtualizatoin host, the virtual machines running on top of such host will still work with VLANs enabled. CC: Patrick McHardy <kaber@trash.net> CC: Nithin Nayak Sujir <nsujir@broadcom.com> CC: Michael Chan <mchan@broadcom.com> CC: Jiri Pirko <jiri@resnulli.us> Signed-off-by: Vladislav Yasevich <vyasevic@redhat.com> Acked-by: Jiri Pirko <jiri@resnulli.us> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-08-08 14:42:13 -04:00
net, sched: add clsact qdisc This work adds a generalization of the ingress qdisc as a qdisc holding only classifiers. The clsact qdisc works on ingress, but also on egress. In both cases, it's execution happens without taking the qdisc lock, and the main difference for the egress part compared to prior version of [1] is that this can be applied with _any_ underlying real egress qdisc (also classless ones). Besides solving the use-case of [1], that is, allowing for more programmability on assigning skb->priority for the mqprio case that is supported by most popular 10G+ NICs, it also opens up a lot more flexibility for other tc applications. The main work on classification can already be done at clsact egress time if the use-case allows and state stored for later retrieval f.e. again in skb->priority with major/minors (which is checked by most classful qdiscs before consulting tc_classify()) and/or in other skb fields like skb->tc_index for some light-weight post-processing to get to the eventual classid in case of a classful qdisc. Another use case is that the clsact egress part allows to have a central egress counterpart to the ingress classifiers, so that classifiers can easily share state (e.g. in cls_bpf via eBPF maps) for ingress and egress. Currently, default setups like mq + pfifo_fast would require for this to use, for example, prio qdisc instead (to get a tc_classify() run) and to duplicate the egress classifier for each queue. With clsact, it allows for leaving the setup as is, it can additionally assign skb->priority to put the skb in one of pfifo_fast's bands and it can share state with maps. Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid) w/o the need to perform a skb_dst_force() to hold on to it any longer. In lwt case, we can also use this facility to setup dst metadata via cls_bpf (bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for that (case of IFF_NO_QUEUE devices, for example). The realization can be done without any changes to the scheduler core framework. All it takes is that we have two a-priori defined minors/child classes, where we can mux between ingress and egress classifier list (dev->ingress_cl_list and dev->egress_cl_list, latter stored close to dev->_tx to avoid extra cacheline miss for moderate loads). The egress part is a bit similar modelled to handle_ing() and patched to a noop in case the functionality is not used. Both handlers are now called sch_handle_ingress() and sch_handle_egress(), code sharing among the two doesn't seem practical as there are various minor differences in both paths, so that making them conditional in a single handler would rather slow things down. Full compatibility to ingress qdisc is provided as well. Since both piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist per netdevice, and thus ingress qdisc specific behaviour can be retained for user space. This means, either a user does 'tc qdisc add dev foo ingress' and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact' alternative, where both, ingress and egress classifier can be configured as in the below example. ingress qdisc supports attaching classifier to any minor number whereas clsact has two fixed minors for muxing between the lists, therefore to not break user space setups, they are better done as two separate qdiscs. I decided to extend the sch_ingress module with clsact functionality so that commonly used code can be reused, the module is being aliased with sch_clsact so that it can be auto-loaded properly. Alternative would have been to add a flag when initializing ingress to alter its behaviour plus aliasing to a different name (as it's more than just ingress). However, the first would end up, based on the flag, choosing the new/old behaviour by calling different function implementations to handle each anyway, the latter would require to register ingress qdisc once again under different alias. So, this really begs to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops by its own that share callbacks used by both. Example, adding qdisc: # tc qdisc add dev foo clsact # tc qdisc show dev foo qdisc mq 0: root qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1 qdisc clsact ffff: parent ffff:fff1 Adding filters (deleting, etc works analogous by specifying ingress/egress): # tc filter add dev foo ingress bpf da obj bar.o sec ingress # tc filter add dev foo egress bpf da obj bar.o sec egress # tc filter show dev foo ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action # tc filter show dev foo egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will show an empty list for clsact. Either using the parent names (ingress/egress) or specifying the full major/minor will then show the related filter lists. Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend. [1] http://patchwork.ozlabs.org/patch/512949/ Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-01-07 22:29:47 +01:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
net: allow vlan traffic to be received under bond The following configuration used to work as I expected. At least we could use the fcoe interfaces to do MPIO and the bond0 iface to do load balancing or failover. ---eth2.228-fcoe | eth2 -----| | |---- bond0 | eth3 -----| | ---eth3.228-fcoe This worked because of a change we added to allow inactive slaves to rx 'exact' matches. This functionality was kept intact with the rx_handler mechanism. However now the vlan interface attached to the active slave never receives traffic because the bonding rx_handler updates the skb->dev and goto's another_round. Previously, the vlan_do_receive() logic was called before the bonding rx_handler. Now by the time vlan_do_receive calls vlan_find_dev() the skb->dev is set to bond0 and it is clear no vlan is attached to this iface. The vlan lookup fails. This patch moves the VLAN check above the rx_handler. A VLAN tagged frame is now routed to the eth2.228-fcoe iface in the above schematic. Untagged frames continue to the bond0 as normal. This case also remains intact, eth2 --> bond0 --> vlan.228 Here the skb is VLAN tagged but the vlan lookup fails on eth2 causing the bonding rx_handler to be called. On the second pass the vlan lookup is on the bond0 iface and completes as expected. Putting a VLAN.228 on both the bond0 and eth2 device will result in eth2.228 receiving the skb. I don't think this is completely unexpected and was the result prior to the rx_handler result. Note, the same setup is also used for other storage traffic that MPIO is used with eg. iSCSI and similar setups can be contrived without storage protocols. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Acked-by: Jesse Gross <jesse@nicira.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Tested-by: Hans Schillstrom <hams.schillstrom@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-10-10 09:16:41 +00:00
net: allow vlan traffic to be received under bond The following configuration used to work as I expected. At least we could use the fcoe interfaces to do MPIO and the bond0 iface to do load balancing or failover. ---eth2.228-fcoe | eth2 -----| | |---- bond0 | eth3 -----| | ---eth3.228-fcoe This worked because of a change we added to allow inactive slaves to rx 'exact' matches. This functionality was kept intact with the rx_handler mechanism. However now the vlan interface attached to the active slave never receives traffic because the bonding rx_handler updates the skb->dev and goto's another_round. Previously, the vlan_do_receive() logic was called before the bonding rx_handler. Now by the time vlan_do_receive calls vlan_find_dev() the skb->dev is set to bond0 and it is clear no vlan is attached to this iface. The vlan lookup fails. This patch moves the VLAN check above the rx_handler. A VLAN tagged frame is now routed to the eth2.228-fcoe iface in the above schematic. Untagged frames continue to the bond0 as normal. This case also remains intact, eth2 --> bond0 --> vlan.228 Here the skb is VLAN tagged but the vlan lookup fails on eth2 causing the bonding rx_handler to be called. On the second pass the vlan lookup is on the bond0 iface and completes as expected. Putting a VLAN.228 on both the bond0 and eth2 device will result in eth2.228 receiving the skb. I don't think this is completely unexpected and was the result prior to the rx_handler result. Note, the same setup is also used for other storage traffic that MPIO is used with eg. iSCSI and similar setups can be contrived without storage protocols. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Acked-by: Jesse Gross <jesse@nicira.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Tested-by: Hans Schillstrom <hams.schillstrom@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-10-10 09:16:41 +00:00
net: allow vlan traffic to be received under bond The following configuration used to work as I expected. At least we could use the fcoe interfaces to do MPIO and the bond0 iface to do load balancing or failover. ---eth2.228-fcoe | eth2 -----| | |---- bond0 | eth3 -----| | ---eth3.228-fcoe This worked because of a change we added to allow inactive slaves to rx 'exact' matches. This functionality was kept intact with the rx_handler mechanism. However now the vlan interface attached to the active slave never receives traffic because the bonding rx_handler updates the skb->dev and goto's another_round. Previously, the vlan_do_receive() logic was called before the bonding rx_handler. Now by the time vlan_do_receive calls vlan_find_dev() the skb->dev is set to bond0 and it is clear no vlan is attached to this iface. The vlan lookup fails. This patch moves the VLAN check above the rx_handler. A VLAN tagged frame is now routed to the eth2.228-fcoe iface in the above schematic. Untagged frames continue to the bond0 as normal. This case also remains intact, eth2 --> bond0 --> vlan.228 Here the skb is VLAN tagged but the vlan lookup fails on eth2 causing the bonding rx_handler to be called. On the second pass the vlan lookup is on the bond0 iface and completes as expected. Putting a VLAN.228 on both the bond0 and eth2 device will result in eth2.228 receiving the skb. I don't think this is completely unexpected and was the result prior to the rx_handler result. Note, the same setup is also used for other storage traffic that MPIO is used with eg. iSCSI and similar setups can be contrived without storage protocols. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Acked-by: Jesse Gross <jesse@nicira.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Tested-by: Hans Schillstrom <hams.schillstrom@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-10-10 09:16:41 +00:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: rcu lock and preempt disable missing around generic xdp do_xdp_generic must be called inside rcu critical section with preempt disabled to ensure BPF programs are valid and per-cpu variables used for redirect operations are consistent. This patch ensures this is true and fixes the splat below. The netif_receive_skb_internal() code path is now broken into two rcu critical sections. I decided it was better to limit the preempt_enable/disable block to just the xdp static key portion and the fallout is more rcu_read_lock/unlock calls. Seems like the best option to me. [ 607.596901] ============================= [ 607.596906] WARNING: suspicious RCU usage [ 607.596912] 4.13.0-rc4+ #570 Not tainted [ 607.596917] ----------------------------- [ 607.596923] net/core/dev.c:3948 suspicious rcu_dereference_check() usage! [ 607.596927] [ 607.596927] other info that might help us debug this: [ 607.596927] [ 607.596933] [ 607.596933] rcu_scheduler_active = 2, debug_locks = 1 [ 607.596938] 2 locks held by pool/14624: [ 607.596943] #0: (rcu_read_lock_bh){......}, at: [<ffffffff95445ffd>] ip_finish_output2+0x14d/0x890 [ 607.596973] #1: (rcu_read_lock_bh){......}, at: [<ffffffff953c8e3a>] __dev_queue_xmit+0x14a/0xfd0 [ 607.597000] [ 607.597000] stack backtrace: [ 607.597006] CPU: 5 PID: 14624 Comm: pool Not tainted 4.13.0-rc4+ #570 [ 607.597011] Hardware name: Dell Inc. Precision Tower 5810/0HHV7N, BIOS A17 03/01/2017 [ 607.597016] Call Trace: [ 607.597027] dump_stack+0x67/0x92 [ 607.597040] lockdep_rcu_suspicious+0xdd/0x110 [ 607.597054] do_xdp_generic+0x313/0xa50 [ 607.597068] ? time_hardirqs_on+0x5b/0x150 [ 607.597076] ? mark_held_locks+0x6b/0xc0 [ 607.597088] ? netdev_pick_tx+0x150/0x150 [ 607.597117] netif_rx_internal+0x205/0x3f0 [ 607.597127] ? do_xdp_generic+0xa50/0xa50 [ 607.597144] ? lock_downgrade+0x2b0/0x2b0 [ 607.597158] ? __lock_is_held+0x93/0x100 [ 607.597187] netif_rx+0x119/0x190 [ 607.597202] loopback_xmit+0xfd/0x1b0 [ 607.597214] dev_hard_start_xmit+0x127/0x4e0 Fixes: d445516966dc ("net: xdp: support xdp generic on virtual devices") Fixes: b5cdae3291f7 ("net: Generic XDP") Acked-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-08 14:00:30 -07:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: rcu lock and preempt disable missing around generic xdp do_xdp_generic must be called inside rcu critical section with preempt disabled to ensure BPF programs are valid and per-cpu variables used for redirect operations are consistent. This patch ensures this is true and fixes the splat below. The netif_receive_skb_internal() code path is now broken into two rcu critical sections. I decided it was better to limit the preempt_enable/disable block to just the xdp static key portion and the fallout is more rcu_read_lock/unlock calls. Seems like the best option to me. [ 607.596901] ============================= [ 607.596906] WARNING: suspicious RCU usage [ 607.596912] 4.13.0-rc4+ #570 Not tainted [ 607.596917] ----------------------------- [ 607.596923] net/core/dev.c:3948 suspicious rcu_dereference_check() usage! [ 607.596927] [ 607.596927] other info that might help us debug this: [ 607.596927] [ 607.596933] [ 607.596933] rcu_scheduler_active = 2, debug_locks = 1 [ 607.596938] 2 locks held by pool/14624: [ 607.596943] #0: (rcu_read_lock_bh){......}, at: [<ffffffff95445ffd>] ip_finish_output2+0x14d/0x890 [ 607.596973] #1: (rcu_read_lock_bh){......}, at: [<ffffffff953c8e3a>] __dev_queue_xmit+0x14a/0xfd0 [ 607.597000] [ 607.597000] stack backtrace: [ 607.597006] CPU: 5 PID: 14624 Comm: pool Not tainted 4.13.0-rc4+ #570 [ 607.597011] Hardware name: Dell Inc. Precision Tower 5810/0HHV7N, BIOS A17 03/01/2017 [ 607.597016] Call Trace: [ 607.597027] dump_stack+0x67/0x92 [ 607.597040] lockdep_rcu_suspicious+0xdd/0x110 [ 607.597054] do_xdp_generic+0x313/0xa50 [ 607.597068] ? time_hardirqs_on+0x5b/0x150 [ 607.597076] ? mark_held_locks+0x6b/0xc0 [ 607.597088] ? netdev_pick_tx+0x150/0x150 [ 607.597117] netif_rx_internal+0x205/0x3f0 [ 607.597127] ? do_xdp_generic+0xa50/0xa50 [ 607.597144] ? lock_downgrade+0x2b0/0x2b0 [ 607.597158] ? __lock_is_held+0x93/0x100 [ 607.597187] netif_rx+0x119/0x190 [ 607.597202] loopback_xmit+0xfd/0x1b0 [ 607.597214] dev_hard_start_xmit+0x127/0x4e0 Fixes: d445516966dc ("net: xdp: support xdp generic on virtual devices") Fixes: b5cdae3291f7 ("net: Generic XDP") Acked-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-08 14:00:30 -07:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: rcu lock and preempt disable missing around generic xdp do_xdp_generic must be called inside rcu critical section with preempt disabled to ensure BPF programs are valid and per-cpu variables used for redirect operations are consistent. This patch ensures this is true and fixes the splat below. The netif_receive_skb_internal() code path is now broken into two rcu critical sections. I decided it was better to limit the preempt_enable/disable block to just the xdp static key portion and the fallout is more rcu_read_lock/unlock calls. Seems like the best option to me. [ 607.596901] ============================= [ 607.596906] WARNING: suspicious RCU usage [ 607.596912] 4.13.0-rc4+ #570 Not tainted [ 607.596917] ----------------------------- [ 607.596923] net/core/dev.c:3948 suspicious rcu_dereference_check() usage! [ 607.596927] [ 607.596927] other info that might help us debug this: [ 607.596927] [ 607.596933] [ 607.596933] rcu_scheduler_active = 2, debug_locks = 1 [ 607.596938] 2 locks held by pool/14624: [ 607.596943] #0: (rcu_read_lock_bh){......}, at: [<ffffffff95445ffd>] ip_finish_output2+0x14d/0x890 [ 607.596973] #1: (rcu_read_lock_bh){......}, at: [<ffffffff953c8e3a>] __dev_queue_xmit+0x14a/0xfd0 [ 607.597000] [ 607.597000] stack backtrace: [ 607.597006] CPU: 5 PID: 14624 Comm: pool Not tainted 4.13.0-rc4+ #570 [ 607.597011] Hardware name: Dell Inc. Precision Tower 5810/0HHV7N, BIOS A17 03/01/2017 [ 607.597016] Call Trace: [ 607.597027] dump_stack+0x67/0x92 [ 607.597040] lockdep_rcu_suspicious+0xdd/0x110 [ 607.597054] do_xdp_generic+0x313/0xa50 [ 607.597068] ? time_hardirqs_on+0x5b/0x150 [ 607.597076] ? mark_held_locks+0x6b/0xc0 [ 607.597088] ? netdev_pick_tx+0x150/0x150 [ 607.597117] netif_rx_internal+0x205/0x3f0 [ 607.597127] ? do_xdp_generic+0xa50/0xa50 [ 607.597144] ? lock_downgrade+0x2b0/0x2b0 [ 607.597158] ? __lock_is_held+0x93/0x100 [ 607.597187] netif_rx+0x119/0x190 [ 607.597202] loopback_xmit+0xfd/0x1b0 [ 607.597214] dev_hard_start_xmit+0x127/0x4e0 Fixes: d445516966dc ("net: xdp: support xdp generic on virtual devices") Fixes: b5cdae3291f7 ("net: Generic XDP") Acked-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: John Fastabend <john.fastabend@gmail.com> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-08 14:00:30 -07:00
static keys: Introduce 'struct static_key', static_key_true()/false() and static_key_slow_[inc|dec]() So here's a boot tested patch on top of Jason's series that does all the cleanups I talked about and turns jump labels into a more intuitive to use facility. It should also address the various misconceptions and confusions that surround jump labels. Typical usage scenarios: #include <linux/static_key.h> struct static_key key = STATIC_KEY_INIT_TRUE; if (static_key_false(&key)) do unlikely code else do likely code Or: if (static_key_true(&key)) do likely code else do unlikely code The static key is modified via: static_key_slow_inc(&key); ... static_key_slow_dec(&key); The 'slow' prefix makes it abundantly clear that this is an expensive operation. I've updated all in-kernel code to use this everywhere. Note that I (intentionally) have not pushed through the rename blindly through to the lowest levels: the actual jump-label patching arch facility should be named like that, so we want to decouple jump labels from the static-key facility a bit. On non-jump-label enabled architectures static keys default to likely()/unlikely() branches. Signed-off-by: Ingo Molnar <mingo@elte.hu> Acked-by: Jason Baron <jbaron@redhat.com> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: a.p.zijlstra@chello.nl Cc: mathieu.desnoyers@efficios.com Cc: davem@davemloft.net Cc: ddaney.cavm@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20120222085809.GA26397@elte.hu Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-02-24 08:31:31 +01:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
bpf: add meta pointer for direct access This work enables generic transfer of metadata from XDP into skb. The basic idea is that we can make use of the fact that the resulting skb must be linear and already comes with a larger headroom for supporting bpf_xdp_adjust_head(), which mangles xdp->data. Here, we base our work on a similar principle and introduce a small helper bpf_xdp_adjust_meta() for adjusting a new pointer called xdp->data_meta. Thus, the packet has a flexible and programmable room for meta data, followed by the actual packet data. struct xdp_buff is therefore laid out that we first point to data_hard_start, then data_meta directly prepended to data followed by data_end marking the end of packet. bpf_xdp_adjust_head() takes into account whether we have meta data already prepended and if so, memmove()s this along with the given offset provided there's enough room. xdp->data_meta is optional and programs are not required to use it. The rationale is that when we process the packet in XDP (e.g. as DoS filter), we can push further meta data along with it for the XDP_PASS case, and give the guarantee that a clsact ingress BPF program on the same device can pick this up for further post-processing. Since we work with skb there, we can also set skb->mark, skb->priority or other skb meta data out of BPF, thus having this scratch space generic and programmable allows for more flexibility than defining a direct 1:1 transfer of potentially new XDP members into skb (it's also more efficient as we don't need to initialize/handle each of such new members). The facility also works together with GRO aggregation. The scratch space at the head of the packet can be multiple of 4 byte up to 32 byte large. Drivers not yet supporting xdp->data_meta can simply be set up with xdp->data_meta as xdp->data + 1 as bpf_xdp_adjust_meta() will detect this and bail out, such that the subsequent match against xdp->data for later access is guaranteed to fail. The verifier treats xdp->data_meta/xdp->data the same way as we treat xdp->data/xdp->data_end pointer comparisons. The requirement for doing the compare against xdp->data is that it hasn't been modified from it's original address we got from ctx access. It may have a range marking already from prior successful xdp->data/xdp->data_end pointer comparisons though. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Acked-by: John Fastabend <john.fastabend@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-09-25 02:25:51 +02:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net-gre-gro: Add GRE support to the GRO stack This patch built on top of Commit 299603e8370a93dd5d8e8d800f0dff1ce2c53d36 ("net-gro: Prepare GRO stack for the upcoming tunneling support") to add the support of the standard GRE (RFC1701/RFC2784/RFC2890) to the GRO stack. It also serves as an example for supporting other encapsulation protocols in the GRO stack in the future. The patch supports version 0 and all the flags (key, csum, seq#) but will flush any pkt with the S (seq#) flag. This is because the S flag is not support by GSO, and a GRO pkt may end up in the forwarding path, thus requiring GSO support to break it up correctly. Currently the "packet_offload" structure only contains L3 (ETH_P_IP/ ETH_P_IPV6) GRO offload support so the encapped pkts are limited to IP pkts (i.e., w/o L2 hdr). But support for other protocol type can be easily added, so is the support for GRE variations like NVGRE. The patch also support csum offload. Specifically if the csum flag is on and the h/w is capable of checksumming the payload (CHECKSUM_COMPLETE), the code will take advantage of the csum computed by the h/w when validating the GRE csum. Note that commit 60769a5dcd8755715c7143b4571d5c44f01796f1 "ipv4: gre: add GRO capability" already introduces GRO capability to IPv4 GRE tunnels, using the gro_cells infrastructure. But GRO is done after GRE hdr has been removed (i.e., decapped). The following patch applies GRO when pkts first come in (before hitting the GRE tunnel code). There is some performance advantage for applying GRO as early as possible. Also this approach is transparent to other subsystem like Open vSwitch where GRE decap is handled outside of the IP stack hence making it harder for the gro_cells stuff to apply. On the other hand, some NICs are still not capable of hashing on the inner hdr of a GRE pkt (RSS). In that case the GRO processing of pkts from the same remote host will all happen on the same CPU and the performance may be suboptimal. I'm including some rough preliminary performance numbers below. Note that the performance will be highly dependent on traffic load, mix as usual. Moreover it also depends on NIC offload features hence the following is by no means a comprehesive study. Local testing and tuning will be needed to decide the best setting. All tests spawned 50 copies of netperf TCP_STREAM and ran for 30 secs. (super_netperf 50 -H 192.168.1.18 -l 30) An IP GRE tunnel with only the key flag on (e.g., ip tunnel add gre1 mode gre local 10.246.17.18 remote 10.246.17.17 ttl 255 key 123) is configured. The GRO support for pkts AFTER decap are controlled through the device feature of the GRE device (e.g., ethtool -K gre1 gro on/off). 1.1 ethtool -K gre1 gro off; ethtool -K eth0 gro off thruput: 9.16Gbps CPU utilization: 19% 1.2 ethtool -K gre1 gro on; ethtool -K eth0 gro off thruput: 5.9Gbps CPU utilization: 15% 1.3 ethtool -K gre1 gro off; ethtool -K eth0 gro on thruput: 9.26Gbps CPU utilization: 12-13% 1.4 ethtool -K gre1 gro on; ethtool -K eth0 gro on thruput: 9.26Gbps CPU utilization: 10% The following tests were performed on a different NIC that is capable of csum offload. I.e., the h/w is capable of computing IP payload csum (CHECKSUM_COMPLETE). 2.1 ethtool -K gre1 gro on (hence will use gro_cells) 2.1.1 ethtool -K eth0 gro off; csum offload disabled thruput: 8.53Gbps CPU utilization: 9% 2.1.2 ethtool -K eth0 gro off; csum offload enabled thruput: 8.97Gbps CPU utilization: 7-8% 2.1.3 ethtool -K eth0 gro on; csum offload disabled thruput: 8.83Gbps CPU utilization: 5-6% 2.1.4 ethtool -K eth0 gro on; csum offload enabled thruput: 8.98Gbps CPU utilization: 5% 2.2 ethtool -K gre1 gro off 2.2.1 ethtool -K eth0 gro off; csum offload disabled thruput: 5.93Gbps CPU utilization: 9% 2.2.2 ethtool -K eth0 gro off; csum offload enabled thruput: 5.62Gbps CPU utilization: 8% 2.2.3 ethtool -K eth0 gro on; csum offload disabled thruput: 7.69Gbps CPU utilization: 8% 2.2.4 ethtool -K eth0 gro on; csum offload enabled thruput: 8.96Gbps CPU utilization: 5-6% Signed-off-by: H.K. Jerry Chu <hkchu@google.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-07 10:23:19 -08:00
net-gre-gro: Add GRE support to the GRO stack This patch built on top of Commit 299603e8370a93dd5d8e8d800f0dff1ce2c53d36 ("net-gro: Prepare GRO stack for the upcoming tunneling support") to add the support of the standard GRE (RFC1701/RFC2784/RFC2890) to the GRO stack. It also serves as an example for supporting other encapsulation protocols in the GRO stack in the future. The patch supports version 0 and all the flags (key, csum, seq#) but will flush any pkt with the S (seq#) flag. This is because the S flag is not support by GSO, and a GRO pkt may end up in the forwarding path, thus requiring GSO support to break it up correctly. Currently the "packet_offload" structure only contains L3 (ETH_P_IP/ ETH_P_IPV6) GRO offload support so the encapped pkts are limited to IP pkts (i.e., w/o L2 hdr). But support for other protocol type can be easily added, so is the support for GRE variations like NVGRE. The patch also support csum offload. Specifically if the csum flag is on and the h/w is capable of checksumming the payload (CHECKSUM_COMPLETE), the code will take advantage of the csum computed by the h/w when validating the GRE csum. Note that commit 60769a5dcd8755715c7143b4571d5c44f01796f1 "ipv4: gre: add GRO capability" already introduces GRO capability to IPv4 GRE tunnels, using the gro_cells infrastructure. But GRO is done after GRE hdr has been removed (i.e., decapped). The following patch applies GRO when pkts first come in (before hitting the GRE tunnel code). There is some performance advantage for applying GRO as early as possible. Also this approach is transparent to other subsystem like Open vSwitch where GRE decap is handled outside of the IP stack hence making it harder for the gro_cells stuff to apply. On the other hand, some NICs are still not capable of hashing on the inner hdr of a GRE pkt (RSS). In that case the GRO processing of pkts from the same remote host will all happen on the same CPU and the performance may be suboptimal. I'm including some rough preliminary performance numbers below. Note that the performance will be highly dependent on traffic load, mix as usual. Moreover it also depends on NIC offload features hence the following is by no means a comprehesive study. Local testing and tuning will be needed to decide the best setting. All tests spawned 50 copies of netperf TCP_STREAM and ran for 30 secs. (super_netperf 50 -H 192.168.1.18 -l 30) An IP GRE tunnel with only the key flag on (e.g., ip tunnel add gre1 mode gre local 10.246.17.18 remote 10.246.17.17 ttl 255 key 123) is configured. The GRO support for pkts AFTER decap are controlled through the device feature of the GRE device (e.g., ethtool -K gre1 gro on/off). 1.1 ethtool -K gre1 gro off; ethtool -K eth0 gro off thruput: 9.16Gbps CPU utilization: 19% 1.2 ethtool -K gre1 gro on; ethtool -K eth0 gro off thruput: 5.9Gbps CPU utilization: 15% 1.3 ethtool -K gre1 gro off; ethtool -K eth0 gro on thruput: 9.26Gbps CPU utilization: 12-13% 1.4 ethtool -K gre1 gro on; ethtool -K eth0 gro on thruput: 9.26Gbps CPU utilization: 10% The following tests were performed on a different NIC that is capable of csum offload. I.e., the h/w is capable of computing IP payload csum (CHECKSUM_COMPLETE). 2.1 ethtool -K gre1 gro on (hence will use gro_cells) 2.1.1 ethtool -K eth0 gro off; csum offload disabled thruput: 8.53Gbps CPU utilization: 9% 2.1.2 ethtool -K eth0 gro off; csum offload enabled thruput: 8.97Gbps CPU utilization: 7-8% 2.1.3 ethtool -K eth0 gro on; csum offload disabled thruput: 8.83Gbps CPU utilization: 5-6% 2.1.4 ethtool -K eth0 gro on; csum offload enabled thruput: 8.98Gbps CPU utilization: 5% 2.2 ethtool -K gre1 gro off 2.2.1 ethtool -K eth0 gro off; csum offload disabled thruput: 5.93Gbps CPU utilization: 9% 2.2.2 ethtool -K eth0 gro off; csum offload enabled thruput: 5.62Gbps CPU utilization: 8% 2.2.3 ethtool -K eth0 gro on; csum offload disabled thruput: 7.69Gbps CPU utilization: 8% 2.2.4 ethtool -K eth0 gro on; csum offload enabled thruput: 8.96Gbps CPU utilization: 5-6% Signed-off-by: H.K. Jerry Chu <hkchu@google.com> Reviewed-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-01-07 10:23:19 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
vlan: Centralize handling of hardware acceleration. Currently each driver that is capable of vlan hardware acceleration must be aware of the vlan groups that are configured and then pass the stripped tag to a specialized receive function. This is different from other types of hardware offload in that it places a significant amount of knowledge in the driver itself rather keeping it in the networking core. This makes vlan offloading function more similarly to other forms of offloading (such as checksum offloading or TSO) by doing the following: * On receive, stripped vlans are passed directly to the network core, without attempting to check for vlan groups or reconstructing the header if no group * vlans are made less special by folding the logic into the main receive routines * On transmit, the device layer will add the vlan header in software if the hardware doesn't support it, instead of spreading that logic out in upper layers, such as bonding. There are a number of advantages to this: * Fixes all bugs with drivers incorrectly dropping vlan headers at once. * Avoids having to disable VLAN acceleration when in promiscuous mode (good for bridging since it always puts devices in promiscuous mode). * Keeps VLAN tag separate until given to ultimate consumer, which avoids needing to do header reconstruction as in tg3 unless absolutely necessary. * Consolidates common code in core networking. Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-10-20 13:56:06 +00:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
net: solve a NAPI race While playing with mlx4 hardware timestamping of RX packets, I found that some packets were received by TCP stack with a ~200 ms delay... Since the timestamp was provided by the NIC, and my probe was added in tcp_v4_rcv() while in BH handler, I was confident it was not a sender issue, or a drop in the network. This would happen with a very low probability, but hurting RPC workloads. A NAPI driver normally arms the IRQ after the napi_complete_done(), after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab it. Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit while IRQ are not disabled, we might have later an IRQ firing and finding this bit set, right before napi_complete_done() clears it. This can happen with busy polling users, or if gro_flush_timeout is used. But some other uses of napi_schedule() in drivers can cause this as well. thread 1 thread 2 (could be on same cpu, or not) // busy polling or napi_watchdog() napi_schedule(); ... napi->poll() device polling: read 2 packets from ring buffer Additional 3rd packet is available. device hard irq // does nothing because NAPI_STATE_SCHED bit is owned by thread 1 napi_schedule(); napi_complete_done(napi, 2); rearm_irq(); Note that rearm_irq() will not force the device to send an additional IRQ for the packet it already signaled (3rd packet in my example) This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep() can set if it could not grab NAPI_STATE_SCHED Then napi_complete_done() properly reschedules the napi to make sure we do not miss something. Since we manipulate multiple bits at once, use cmpxchg() like in sk_busy_loop() to provide proper transactions. In v2, I changed napi_watchdog() to use a relaxed variant of napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point. In v3, I added more details in the changelog and clears NAPI_STATE_MISSED in busy_poll_stop() In v4, I added the ideas given by Alexander Duyck in v3 review Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Alexander Duyck <alexander.duyck@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-28 10:34:50 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: solve a NAPI race While playing with mlx4 hardware timestamping of RX packets, I found that some packets were received by TCP stack with a ~200 ms delay... Since the timestamp was provided by the NIC, and my probe was added in tcp_v4_rcv() while in BH handler, I was confident it was not a sender issue, or a drop in the network. This would happen with a very low probability, but hurting RPC workloads. A NAPI driver normally arms the IRQ after the napi_complete_done(), after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab it. Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit while IRQ are not disabled, we might have later an IRQ firing and finding this bit set, right before napi_complete_done() clears it. This can happen with busy polling users, or if gro_flush_timeout is used. But some other uses of napi_schedule() in drivers can cause this as well. thread 1 thread 2 (could be on same cpu, or not) // busy polling or napi_watchdog() napi_schedule(); ... napi->poll() device polling: read 2 packets from ring buffer Additional 3rd packet is available. device hard irq // does nothing because NAPI_STATE_SCHED bit is owned by thread 1 napi_schedule(); napi_complete_done(napi, 2); rearm_irq(); Note that rearm_irq() will not force the device to send an additional IRQ for the packet it already signaled (3rd packet in my example) This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep() can set if it could not grab NAPI_STATE_SCHED Then napi_complete_done() properly reschedules the napi to make sure we do not miss something. Since we manipulate multiple bits at once, use cmpxchg() like in sk_busy_loop() to provide proper transactions. In v2, I changed napi_watchdog() to use a relaxed variant of napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point. In v3, I added more details in the changelog and clears NAPI_STATE_MISSED in busy_poll_stop() In v4, I added the ideas given by Alexander Duyck in v3 review Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Alexander Duyck <alexander.duyck@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-28 10:34:50 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: solve a NAPI race While playing with mlx4 hardware timestamping of RX packets, I found that some packets were received by TCP stack with a ~200 ms delay... Since the timestamp was provided by the NIC, and my probe was added in tcp_v4_rcv() while in BH handler, I was confident it was not a sender issue, or a drop in the network. This would happen with a very low probability, but hurting RPC workloads. A NAPI driver normally arms the IRQ after the napi_complete_done(), after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab it. Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit while IRQ are not disabled, we might have later an IRQ firing and finding this bit set, right before napi_complete_done() clears it. This can happen with busy polling users, or if gro_flush_timeout is used. But some other uses of napi_schedule() in drivers can cause this as well. thread 1 thread 2 (could be on same cpu, or not) // busy polling or napi_watchdog() napi_schedule(); ... napi->poll() device polling: read 2 packets from ring buffer Additional 3rd packet is available. device hard irq // does nothing because NAPI_STATE_SCHED bit is owned by thread 1 napi_schedule(); napi_complete_done(napi, 2); rearm_irq(); Note that rearm_irq() will not force the device to send an additional IRQ for the packet it already signaled (3rd packet in my example) This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep() can set if it could not grab NAPI_STATE_SCHED Then napi_complete_done() properly reschedules the napi to make sure we do not miss something. Since we manipulate multiple bits at once, use cmpxchg() like in sk_busy_loop() to provide proper transactions. In v2, I changed napi_watchdog() to use a relaxed variant of napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point. In v3, I added more details in the changelog and clears NAPI_STATE_MISSED in busy_poll_stop() In v4, I added the ideas given by Alexander Duyck in v3 review Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Alexander Duyck <alexander.duyck@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-28 10:34:50 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: solve a NAPI race While playing with mlx4 hardware timestamping of RX packets, I found that some packets were received by TCP stack with a ~200 ms delay... Since the timestamp was provided by the NIC, and my probe was added in tcp_v4_rcv() while in BH handler, I was confident it was not a sender issue, or a drop in the network. This would happen with a very low probability, but hurting RPC workloads. A NAPI driver normally arms the IRQ after the napi_complete_done(), after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab it. Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit while IRQ are not disabled, we might have later an IRQ firing and finding this bit set, right before napi_complete_done() clears it. This can happen with busy polling users, or if gro_flush_timeout is used. But some other uses of napi_schedule() in drivers can cause this as well. thread 1 thread 2 (could be on same cpu, or not) // busy polling or napi_watchdog() napi_schedule(); ... napi->poll() device polling: read 2 packets from ring buffer Additional 3rd packet is available. device hard irq // does nothing because NAPI_STATE_SCHED bit is owned by thread 1 napi_schedule(); napi_complete_done(napi, 2); rearm_irq(); Note that rearm_irq() will not force the device to send an additional IRQ for the packet it already signaled (3rd packet in my example) This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep() can set if it could not grab NAPI_STATE_SCHED Then napi_complete_done() properly reschedules the napi to make sure we do not miss something. Since we manipulate multiple bits at once, use cmpxchg() like in sk_busy_loop() to provide proper transactions. In v2, I changed napi_watchdog() to use a relaxed variant of napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point. In v3, I added more details in the changelog and clears NAPI_STATE_MISSED in busy_poll_stop() In v4, I added the ideas given by Alexander Duyck in v3 review Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Alexander Duyck <alexander.duyck@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-28 10:34:50 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: allow BH servicing in sk_busy_loop() Instead of blocking BH in whole sk_busy_loop(), block them only around ->ndo_busy_poll() calls. This has many benefits. 1) allow tunneled traffic to use busy poll as well as native traffic. Tunnels handlers usually call netif_rx() and depend on net_rx_action() being run (from sofirq handler) 2) allow RFS/RPS being used (sending IPI to other cpus if needed) 3) use the 'lets burn cpu cycles' budget to do useful work (like TX completions, timers, RCU callbacks...) 4) reduce BH latencies, making busy poll a better citizen. Tested: Tested with SIT tunnel lpaa5:~# echo 0 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 37373.93 16384 87380 Now enable busy poll on both hosts lpaa5:~# echo 70 >/proc/sys/net/core/busy_read lpaa6:~# echo 70 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 58314.77 16384 87380 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-18 06:30:53 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: allow BH servicing in sk_busy_loop() Instead of blocking BH in whole sk_busy_loop(), block them only around ->ndo_busy_poll() calls. This has many benefits. 1) allow tunneled traffic to use busy poll as well as native traffic. Tunnels handlers usually call netif_rx() and depend on net_rx_action() being run (from sofirq handler) 2) allow RFS/RPS being used (sending IPI to other cpus if needed) 3) use the 'lets burn cpu cycles' budget to do useful work (like TX completions, timers, RCU callbacks...) 4) reduce BH latencies, making busy poll a better citizen. Tested: Tested with SIT tunnel lpaa5:~# echo 0 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 37373.93 16384 87380 Now enable busy poll on both hosts lpaa5:~# echo 70 >/proc/sys/net/core/busy_read lpaa6:~# echo 70 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 58314.77 16384 87380 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-18 06:30:53 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: allow BH servicing in sk_busy_loop() Instead of blocking BH in whole sk_busy_loop(), block them only around ->ndo_busy_poll() calls. This has many benefits. 1) allow tunneled traffic to use busy poll as well as native traffic. Tunnels handlers usually call netif_rx() and depend on net_rx_action() being run (from sofirq handler) 2) allow RFS/RPS being used (sending IPI to other cpus if needed) 3) use the 'lets burn cpu cycles' budget to do useful work (like TX completions, timers, RCU callbacks...) 4) reduce BH latencies, making busy poll a better citizen. Tested: Tested with SIT tunnel lpaa5:~# echo 0 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 37373.93 16384 87380 Now enable busy poll on both hosts lpaa5:~# echo 70 >/proc/sys/net/core/busy_read lpaa6:~# echo 70 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 58314.77 16384 87380 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-18 06:30:53 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
Merge branch 'locking-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip Pull locking updates from Ingo Molnar: "The tree got pretty big in this development cycle, but the net effect is pretty good: 115 files changed, 673 insertions(+), 1522 deletions(-) The main changes were: - Rework and generalize the mutex code to remove per arch mutex primitives. (Peter Zijlstra) - Add vCPU preemption support: add an interface to query the preemption status of vCPUs and use it in locking primitives - this optimizes paravirt performance. (Pan Xinhui, Juergen Gross, Christian Borntraeger) - Introduce cpu_relax_yield() and remov cpu_relax_lowlatency() to clean up and improve the s390 lock yielding machinery and its core kernel impact. (Christian Borntraeger) - Micro-optimize mutexes some more. (Waiman Long) - Reluctantly add the to-be-deprecated mutex_trylock_recursive() interface on a temporary basis, to give the DRM code more time to get rid of its locking hacks. Any other users will be NAK-ed on sight. (We turned off the deprecation warning for the time being to not pollute the build log.) (Peter Zijlstra) - Improve the rtmutex code a bit, in light of recent long lived bugs/races. (Thomas Gleixner) - Misc fixes, cleanups" * 'locking-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (36 commits) x86/paravirt: Fix bool return type for PVOP_CALL() x86/paravirt: Fix native_patch() locking/ww_mutex: Use relaxed atomics locking/rtmutex: Explain locking rules for rt_mutex_proxy_unlock()/init_proxy_locked() locking/rtmutex: Get rid of RT_MUTEX_OWNER_MASKALL x86/paravirt: Optimize native pv_lock_ops.vcpu_is_preempted() locking/mutex: Break out of expensive busy-loop on {mutex,rwsem}_spin_on_owner() when owner vCPU is preempted locking/osq: Break out of spin-wait busy waiting loop for a preempted vCPU in osq_lock() Documentation/virtual/kvm: Support the vCPU preemption check x86/xen: Support the vCPU preemption check x86/kvm: Support the vCPU preemption check x86/kvm: Support the vCPU preemption check kvm: Introduce kvm_write_guest_offset_cached() locking/core, x86/paravirt: Implement vcpu_is_preempted(cpu) for KVM and Xen guests locking/spinlocks, s390: Implement vcpu_is_preempted(cpu) locking/core, powerpc: Implement vcpu_is_preempted(cpu) sched/core: Introduce the vcpu_is_preempted(cpu) interface sched/wake_q: Rename WAKE_Q to DEFINE_WAKE_Q locking/core: Provide common cpu_relax_yield() definition locking/mutex: Don't mark mutex_trylock_recursive() as deprecated, temporarily ...
2016-12-12 10:48:02 -08:00
net: busy-poll: allow preemption in sk_busy_loop() After commit 4cd13c21b207 ("softirq: Let ksoftirqd do its job"), sk_busy_loop() needs a bit of care : softirqs might be delayed since we do not allow preemption yet. This patch adds preemptiom points in sk_busy_loop(), and makes sure no unnecessary cache line dirtying or atomic operations are done while looping. A new flag is added into napi->state : NAPI_STATE_IN_BUSY_POLL This prevents napi_complete_done() from clearing NAPIF_STATE_SCHED, so that sk_busy_loop() does not have to grab it again. Similarly, netpoll_poll_lock() is done one time. This gives about 10 to 20 % improvement in various busy polling tests, especially when many threads are busy polling in configurations with large number of NIC queues. This should allow experimenting with bigger delays without hurting overall latencies. Tested: On a 40Gb mlx4 NIC, 32 RX/TX queues. echo 70 >/proc/sys/net/core/busy_read for i in `seq 1 40`; do echo -n $i: ; ./super_netperf $i -H lpaa24 -t UDP_RR -- -N -n; done Before: After: 1: 90072 92819 2: 157289 184007 3: 235772 213504 4: 344074 357513 5: 394755 458267 6: 461151 487819 7: 549116 625963 8: 544423 716219 9: 720460 738446 10: 794686 837612 11: 915998 923960 12: 937507 925107 13: 1019677 971506 14: 1046831 1113650 15: 1114154 1148902 16: 1105221 1179263 17: 1266552 1299585 18: 1258454 1383817 19: 1341453 1312194 20: 1363557 1488487 21: 1387979 1501004 22: 1417552 1601683 23: 1550049 1642002 24: 1568876 1601915 25: 1560239 1683607 26: 1640207 1745211 27: 1706540 1723574 28: 1638518 1722036 29: 1734309 1757447 30: 1782007 1855436 31: 1724806 1888539 32: 1717716 1944297 33: 1778716 1869118 34: 1805738 1983466 35: 1815694 2020758 36: 1893059 2035632 37: 1843406 2034653 38: 1888830 2086580 39: 1972827 2143567 40: 1877729 2181851 Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Willem de Bruijn <willemb@google.com> Cc: Adam Belay <abelay@google.com> Cc: Tariq Toukan <tariqt@mellanox.com> Cc: Yuval Mintz <Yuval.Mintz@cavium.com> Cc: Ariel Elior <ariel.elior@cavium.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-15 10:15:11 -08:00
net: allow BH servicing in sk_busy_loop() Instead of blocking BH in whole sk_busy_loop(), block them only around ->ndo_busy_poll() calls. This has many benefits. 1) allow tunneled traffic to use busy poll as well as native traffic. Tunnels handlers usually call netif_rx() and depend on net_rx_action() being run (from sofirq handler) 2) allow RFS/RPS being used (sending IPI to other cpus if needed) 3) use the 'lets burn cpu cycles' budget to do useful work (like TX completions, timers, RCU callbacks...) 4) reduce BH latencies, making busy poll a better citizen. Tested: Tested with SIT tunnel lpaa5:~# echo 0 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 37373.93 16384 87380 Now enable busy poll on both hosts lpaa5:~# echo 70 >/proc/sys/net/core/busy_read lpaa6:~# echo 70 >/proc/sys/net/core/busy_read lpaa5:~# ./netperf -H 2002:af6:786::1 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from ::0 (::) port 0 AF_INET6 to 2002:af6:786::1 () port 0 AF_INET6 : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 58314.77 16384 87380 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-18 06:30:53 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: solve a NAPI race While playing with mlx4 hardware timestamping of RX packets, I found that some packets were received by TCP stack with a ~200 ms delay... Since the timestamp was provided by the NIC, and my probe was added in tcp_v4_rcv() while in BH handler, I was confident it was not a sender issue, or a drop in the network. This would happen with a very low probability, but hurting RPC workloads. A NAPI driver normally arms the IRQ after the napi_complete_done(), after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab it. Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit while IRQ are not disabled, we might have later an IRQ firing and finding this bit set, right before napi_complete_done() clears it. This can happen with busy polling users, or if gro_flush_timeout is used. But some other uses of napi_schedule() in drivers can cause this as well. thread 1 thread 2 (could be on same cpu, or not) // busy polling or napi_watchdog() napi_schedule(); ... napi->poll() device polling: read 2 packets from ring buffer Additional 3rd packet is available. device hard irq // does nothing because NAPI_STATE_SCHED bit is owned by thread 1 napi_schedule(); napi_complete_done(napi, 2); rearm_irq(); Note that rearm_irq() will not force the device to send an additional IRQ for the packet it already signaled (3rd packet in my example) This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep() can set if it could not grab NAPI_STATE_SCHED Then napi_complete_done() properly reschedules the napi to make sure we do not miss something. Since we manipulate multiple bits at once, use cmpxchg() like in sk_busy_loop() to provide proper transactions. In v2, I changed napi_watchdog() to use a relaxed variant of napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point. In v3, I added more details in the changelog and clears NAPI_STATE_MISSED in busy_poll_stop() In v4, I added the ideas given by Alexander Duyck in v3 review Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Alexander Duyck <alexander.duyck@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-02-28 10:34:50 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
netpoll: Close race condition between poll_one_napi and napi_disable Drivers might call napi_disable while not holding the napi instance poll_lock. In those instances, its possible for a race condition to exist between poll_one_napi and napi_disable. That is to say, poll_one_napi only tests the NAPI_STATE_SCHED bit to see if there is work to do during a poll, and as such the following may happen: CPU0 CPU1 ndo_tx_timeout napi_poll_dev napi_disable poll_one_napi test_and_set_bit (ret 0) test_bit (ret 1) reset adapter napi_poll_routine If the adapter gets a tx timeout without a napi instance scheduled, its possible for the adapter to think it has exclusive access to the hardware (as the napi instance is now scheduled via the napi_disable call), while the netpoll code thinks there is simply work to do. The result is parallel hardware access leading to corrupt data structures in the driver, and a crash. Additionaly, there is another, more critical race between netpoll and napi_disable. The disabled napi state is actually identical to the scheduled state for a given napi instance. The implication being that, if a napi instance is disabled, a netconsole instance would see the napi state of the device as having been scheduled, and poll it, likely while the driver was dong something requiring exclusive access. In the case above, its fairly clear that not having the rings in a state ready to be polled will cause any number of crashes. The fix should be pretty easy. netpoll uses its own bit to indicate that that the napi instance is in a state of being serviced by netpoll (NAPI_STATE_NPSVC). We can just gate disabling on that bit as well as the sched bit. That should prevent netpoll from conducting a napi poll if we convert its set bit to a test_and_set_bit operation to provide mutual exclusion Change notes: V2) Remove a trailing whtiespace Resubmit with proper subject prefix V3) Clean up spacing nits Signed-off-by: Neil Horman <nhorman@tuxdriver.com> CC: "David S. Miller" <davem@davemloft.net> CC: jmaxwell@redhat.com Tested-by: jmaxwell@redhat.com Signed-off-by: David S. Miller <davem@davemloft.net>
2015-09-23 14:57:58 -04:00
net: gro: add a per device gro flush timer Tuning coalescing parameters on NIC can be really hard. Servers can handle both bulk and RPC like traffic, with conflicting goals : bulk flows want as big GRO packets as possible, RPC want minimal latencies. To reach big GRO packets on 10Gbe NIC, one can use : ethtool -C eth0 rx-usecs 4 rx-frames 44 But this penalizes rpc sessions, with an increase of latencies, up to 50% in some cases, as NICs generally do not force an interrupt when a packet with TCP Push flag is received. Some NICs do not have an absolute timer, only a timer rearmed for every incoming packet. This patch uses a different strategy : Let GRO stack decides what do do, based on traffic pattern. Packets with Push flag wont be delayed. Packets without Push flag might be held in GRO engine, if we keep receiving data. This new mechanism is off by default, and shall be enabled by setting /sys/class/net/ethX/gro_flush_timeout to a value in nanosecond. To fully enable this mechanism, drivers should use napi_complete_done() instead of napi_complete(). Tested: Ran 200 netperf TCP_STREAM from A to B (10Gbe mlx4 link, 8 RX queues) Without this feature, we send back about 305,000 ACK per second. GRO aggregation ratio is low (811/305 = 2.65 segments per GRO packet) Setting a timer of 2000 nsec is enough to increase GRO packet sizes and reduce number of ACK packets. (811/19.2 = 42) Receiver performs less calls to upper stacks, less wakes up. This also reduces cpu usage on the sender, as it receives less ACK packets. Note that reducing number of wakes up increases cpu efficiency, but can decrease QPS, as applications wont have the chance to warmup cpu caches doing a partial read of RPC requests/answers if they fit in one skb. B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811269.80 305732.30 1199462.57 19705.72 0.00 0.00 0.50 B:~# echo 2000 >/sys/class/net/eth0/gro_flush_timeout B:~# sar -n DEV 1 10 | grep eth0 | tail -1 Average: eth0 811577.30 19230.80 1199916.51 1239.80 0.00 0.00 0.50 Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-11-06 21:09:44 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 16:41:36 -07:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
bonding: Fix stacked device detection in arp monitoring Prior to commit fbd929f2dce460456807a51e18d623db3db9f077 bonding: support QinQ for bond arp interval the arp monitoring code allowed for proper detection of devices stacked on top of vlans. Since the above commit, the code can still detect a device stacked on top of single vlan, but not a device stacked on top of Q-in-Q configuration. The search will only set the inner vlan tag if the route device is the vlan device. However, this is not always the case, as it is possible to extend the stacked configuration. With this patch it is possible to provision devices on top Q-in-Q vlan configuration that should be used as a source of ARP monitoring information. For example: ip link add link bond0 vlan10 type vlan proto 802.1q id 10 ip link add link vlan10 vlan100 type vlan proto 802.1q id 100 ip link add link vlan100 type macvlan Note: This patch limites the number of stacked VLANs to 2, just like before. The original, however had another issue in that if we had more then 2 levels of VLANs, we would end up generating incorrectly tagged traffic. This is no longer possible. Fixes: fbd929f2dce460456807a51e18d623db3db9f077 (bonding: support QinQ for bond arp interval) CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@redhat.com> CC: Andy Gospodarek <andy@greyhouse.net> CC: Ding Tianhong <dingtianhong@huawei.com> CC: Patric McHardy <kaber@trash.net> Signed-off-by: Vlad Yasevich <vyasevic@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-16 17:20:38 -04:00
net: make netdev_for_each_lower_dev safe for device removal When I used netdev_for_each_lower_dev in commit bad531623253 ("vrf: remove slave queue and private slave struct") I thought that it acts like netdev_for_each_lower_private and can be used to remove the current device from the list while walking, but unfortunately it acts more like netdev_for_each_lower_private_rcu and doesn't allow it. The difference is where the "iter" points to, right now it points to the current element and that makes it impossible to remove it. Change the logic to be similar to netdev_for_each_lower_private and make it point to the "next" element so we can safely delete the current one. VRF is the only such user right now, there's no change for the read-only users. Here's what can happen now: [98423.249858] general protection fault: 0000 [#1] SMP [98423.250175] Modules linked in: vrf bridge(O) stp llc nfsd auth_rpcgss oid_registry nfs_acl nfs lockd grace sunrpc crct10dif_pclmul crc32_pclmul crc32c_intel ghash_clmulni_intel jitterentropy_rng sha256_generic hmac drbg ppdev aesni_intel aes_x86_64 glue_helper lrw gf128mul ablk_helper cryptd evdev serio_raw pcspkr virtio_balloon parport_pc parport i2c_piix4 i2c_core virtio_console acpi_cpufreq button 9pnet_virtio 9p 9pnet fscache ipv6 autofs4 ext4 crc16 mbcache jbd2 sg virtio_blk virtio_net sr_mod cdrom e1000 ata_generic ehci_pci uhci_hcd ehci_hcd usbcore usb_common virtio_pci ata_piix libata floppy virtio_ring virtio scsi_mod [last unloaded: bridge] [98423.255040] CPU: 1 PID: 14173 Comm: ip Tainted: G O 4.5.0-rc2+ #81 [98423.255386] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.8.1-20150318_183358- 04/01/2014 [98423.255777] task: ffff8800547f5540 ti: ffff88003428c000 task.ti: ffff88003428c000 [98423.256123] RIP: 0010:[<ffffffff81514f3e>] [<ffffffff81514f3e>] netdev_lower_get_next+0x1e/0x30 [98423.256534] RSP: 0018:ffff88003428f940 EFLAGS: 00010207 [98423.256766] RAX: 0002000100000004 RBX: ffff880054ff9000 RCX: 0000000000000000 [98423.257039] RDX: ffff88003428f8b8 RSI: ffff88003428f950 RDI: ffff880054ff90c0 [98423.257287] RBP: ffff88003428f940 R08: 0000000000000000 R09: 0000000000000000 [98423.257537] R10: 0000000000000001 R11: 0000000000000000 R12: ffff88003428f9e0 [98423.257802] R13: ffff880054a5fd00 R14: ffff88003428f970 R15: 0000000000000001 [98423.258055] FS: 00007f3d76881700(0000) GS:ffff88005d000000(0000) knlGS:0000000000000000 [98423.258418] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [98423.258650] CR2: 00007ffe5951ffa8 CR3: 0000000052077000 CR4: 00000000000406e0 [98423.258902] Stack: [98423.259075] ffff88003428f960 ffffffffa0442636 0002000100000004 ffff880054ff9000 [98423.259647] ffff88003428f9b0 ffffffff81518205 ffff880054ff9000 ffff88003428f978 [98423.260208] ffff88003428f978 ffff88003428f9e0 ffff88003428f9e0 ffff880035b35f00 [98423.260739] Call Trace: [98423.260920] [<ffffffffa0442636>] vrf_dev_uninit+0x76/0xa0 [vrf] [98423.261156] [<ffffffff81518205>] rollback_registered_many+0x205/0x390 [98423.261401] [<ffffffff815183ec>] unregister_netdevice_many+0x1c/0x70 [98423.261641] [<ffffffff8153223c>] rtnl_delete_link+0x3c/0x50 [98423.271557] [<ffffffff815335bb>] rtnl_dellink+0xcb/0x1d0 [98423.271800] [<ffffffff811cd7da>] ? __inc_zone_state+0x4a/0x90 [98423.272049] [<ffffffff815337b4>] rtnetlink_rcv_msg+0x84/0x200 [98423.272279] [<ffffffff810cfe7d>] ? trace_hardirqs_on+0xd/0x10 [98423.272513] [<ffffffff8153370b>] ? rtnetlink_rcv+0x1b/0x40 [98423.272755] [<ffffffff81533730>] ? rtnetlink_rcv+0x40/0x40 [98423.272983] [<ffffffff8155d6e7>] netlink_rcv_skb+0x97/0xb0 [98423.273209] [<ffffffff8153371a>] rtnetlink_rcv+0x2a/0x40 [98423.273476] [<ffffffff8155ce8b>] netlink_unicast+0x11b/0x1a0 [98423.273710] [<ffffffff8155d2f1>] netlink_sendmsg+0x3e1/0x610 [98423.273947] [<ffffffff814fbc98>] sock_sendmsg+0x38/0x70 [98423.274175] [<ffffffff814fc253>] ___sys_sendmsg+0x2e3/0x2f0 [98423.274416] [<ffffffff810d841e>] ? do_raw_spin_unlock+0xbe/0x140 [98423.274658] [<ffffffff811e1bec>] ? handle_mm_fault+0x26c/0x2210 [98423.274894] [<ffffffff811e19cd>] ? handle_mm_fault+0x4d/0x2210 [98423.275130] [<ffffffff81269611>] ? __fget_light+0x91/0xb0 [98423.275365] [<ffffffff814fcd42>] __sys_sendmsg+0x42/0x80 [98423.275595] [<ffffffff814fcd92>] SyS_sendmsg+0x12/0x20 [98423.275827] [<ffffffff81611bb6>] entry_SYSCALL_64_fastpath+0x16/0x7a [98423.276073] Code: c3 31 c0 5d c3 0f 1f 84 00 00 00 00 00 66 66 66 66 90 48 8b 06 55 48 81 c7 c0 00 00 00 48 89 e5 48 8b 00 48 39 f8 74 09 48 89 06 <48> 8b 40 e8 5d c3 31 c0 5d c3 0f 1f 84 00 00 00 00 00 66 66 66 [98423.279639] RIP [<ffffffff81514f3e>] netdev_lower_get_next+0x1e/0x30 [98423.279920] RSP <ffff88003428f940> CC: David Ahern <dsa@cumulusnetworks.com> CC: David S. Miller <davem@davemloft.net> CC: Roopa Prabhu <roopa@cumulusnetworks.com> CC: Vlad Yasevich <vyasevic@redhat.com> Fixes: bad531623253 ("vrf: remove slave queue and private slave struct") Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> Reviewed-by: David Ahern <dsa@cumulusnetworks.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-17 18:00:31 +01:00
net: make netdev_for_each_lower_dev safe for device removal When I used netdev_for_each_lower_dev in commit bad531623253 ("vrf: remove slave queue and private slave struct") I thought that it acts like netdev_for_each_lower_private and can be used to remove the current device from the list while walking, but unfortunately it acts more like netdev_for_each_lower_private_rcu and doesn't allow it. The difference is where the "iter" points to, right now it points to the current element and that makes it impossible to remove it. Change the logic to be similar to netdev_for_each_lower_private and make it point to the "next" element so we can safely delete the current one. VRF is the only such user right now, there's no change for the read-only users. Here's what can happen now: [98423.249858] general protection fault: 0000 [#1] SMP [98423.250175] Modules linked in: vrf bridge(O) stp llc nfsd auth_rpcgss oid_registry nfs_acl nfs lockd grace sunrpc crct10dif_pclmul crc32_pclmul crc32c_intel ghash_clmulni_intel jitterentropy_rng sha256_generic hmac drbg ppdev aesni_intel aes_x86_64 glue_helper lrw gf128mul ablk_helper cryptd evdev serio_raw pcspkr virtio_balloon parport_pc parport i2c_piix4 i2c_core virtio_console acpi_cpufreq button 9pnet_virtio 9p 9pnet fscache ipv6 autofs4 ext4 crc16 mbcache jbd2 sg virtio_blk virtio_net sr_mod cdrom e1000 ata_generic ehci_pci uhci_hcd ehci_hcd usbcore usb_common virtio_pci ata_piix libata floppy virtio_ring virtio scsi_mod [last unloaded: bridge] [98423.255040] CPU: 1 PID: 14173 Comm: ip Tainted: G O 4.5.0-rc2+ #81 [98423.255386] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.8.1-20150318_183358- 04/01/2014 [98423.255777] task: ffff8800547f5540 ti: ffff88003428c000 task.ti: ffff88003428c000 [98423.256123] RIP: 0010:[<ffffffff81514f3e>] [<ffffffff81514f3e>] netdev_lower_get_next+0x1e/0x30 [98423.256534] RSP: 0018:ffff88003428f940 EFLAGS: 00010207 [98423.256766] RAX: 0002000100000004 RBX: ffff880054ff9000 RCX: 0000000000000000 [98423.257039] RDX: ffff88003428f8b8 RSI: ffff88003428f950 RDI: ffff880054ff90c0 [98423.257287] RBP: ffff88003428f940 R08: 0000000000000000 R09: 0000000000000000 [98423.257537] R10: 0000000000000001 R11: 0000000000000000 R12: ffff88003428f9e0 [98423.257802] R13: ffff880054a5fd00 R14: ffff88003428f970 R15: 0000000000000001 [98423.258055] FS: 00007f3d76881700(0000) GS:ffff88005d000000(0000) knlGS:0000000000000000 [98423.258418] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [98423.258650] CR2: 00007ffe5951ffa8 CR3: 0000000052077000 CR4: 00000000000406e0 [98423.258902] Stack: [98423.259075] ffff88003428f960 ffffffffa0442636 0002000100000004 ffff880054ff9000 [98423.259647] ffff88003428f9b0 ffffffff81518205 ffff880054ff9000 ffff88003428f978 [98423.260208] ffff88003428f978 ffff88003428f9e0 ffff88003428f9e0 ffff880035b35f00 [98423.260739] Call Trace: [98423.260920] [<ffffffffa0442636>] vrf_dev_uninit+0x76/0xa0 [vrf] [98423.261156] [<ffffffff81518205>] rollback_registered_many+0x205/0x390 [98423.261401] [<ffffffff815183ec>] unregister_netdevice_many+0x1c/0x70 [98423.261641] [<ffffffff8153223c>] rtnl_delete_link+0x3c/0x50 [98423.271557] [<ffffffff815335bb>] rtnl_dellink+0xcb/0x1d0 [98423.271800] [<ffffffff811cd7da>] ? __inc_zone_state+0x4a/0x90 [98423.272049] [<ffffffff815337b4>] rtnetlink_rcv_msg+0x84/0x200 [98423.272279] [<ffffffff810cfe7d>] ? trace_hardirqs_on+0xd/0x10 [98423.272513] [<ffffffff8153370b>] ? rtnetlink_rcv+0x1b/0x40 [98423.272755] [<ffffffff81533730>] ? rtnetlink_rcv+0x40/0x40 [98423.272983] [<ffffffff8155d6e7>] netlink_rcv_skb+0x97/0xb0 [98423.273209] [<ffffffff8153371a>] rtnetlink_rcv+0x2a/0x40 [98423.273476] [<ffffffff8155ce8b>] netlink_unicast+0x11b/0x1a0 [98423.273710] [<ffffffff8155d2f1>] netlink_sendmsg+0x3e1/0x610 [98423.273947] [<ffffffff814fbc98>] sock_sendmsg+0x38/0x70 [98423.274175] [<ffffffff814fc253>] ___sys_sendmsg+0x2e3/0x2f0 [98423.274416] [<ffffffff810d841e>] ? do_raw_spin_unlock+0xbe/0x140 [98423.274658] [<ffffffff811e1bec>] ? handle_mm_fault+0x26c/0x2210 [98423.274894] [<ffffffff811e19cd>] ? handle_mm_fault+0x4d/0x2210 [98423.275130] [<ffffffff81269611>] ? __fget_light+0x91/0xb0 [98423.275365] [<ffffffff814fcd42>] __sys_sendmsg+0x42/0x80 [98423.275595] [<ffffffff814fcd92>] SyS_sendmsg+0x12/0x20 [98423.275827] [<ffffffff81611bb6>] entry_SYSCALL_64_fastpath+0x16/0x7a [98423.276073] Code: c3 31 c0 5d c3 0f 1f 84 00 00 00 00 00 66 66 66 66 90 48 8b 06 55 48 81 c7 c0 00 00 00 48 89 e5 48 8b 00 48 39 f8 74 09 48 89 06 <48> 8b 40 e8 5d c3 31 c0 5d c3 0f 1f 84 00 00 00 00 00 66 66 66 [98423.279639] RIP [<ffffffff81514f3e>] netdev_lower_get_next+0x1e/0x30 [98423.279920] RSP <ffff88003428f940> CC: David Ahern <dsa@cumulusnetworks.com> CC: David S. Miller <davem@davemloft.net> CC: Roopa Prabhu <roopa@cumulusnetworks.com> CC: Vlad Yasevich <vyasevic@redhat.com> Fixes: bad531623253 ("vrf: remove slave queue and private slave struct") Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> Reviewed-by: David Ahern <dsa@cumulusnetworks.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-17 18:00:31 +01:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net: add lower_dev_list to net_device and make a full mesh This patch adds lower_dev_list list_head to net_device, which is the same as upper_dev_list, only for lower devices, and begins to use it in the same way as the upper list. It also changes the way the whole adjacent device lists work - now they contain *all* of upper/lower devices, not only the first level. The first level devices are distinguished by the bool neighbour field in netdev_adjacent, also added by this patch. There are cases when a device can be added several times to the adjacent list, the simplest would be: /---- eth0.10 ---\ eth0- --- bond0 \---- eth0.20 ---/ where both bond0 and eth0 'see' each other in the adjacent lists two times. To avoid duplication of netdev_adjacent structures ref_nr is being kept as the number of times the device was added to the list. The 'full view' is achieved by adding, on link creation, all of the upper_dev's upper_dev_list devices as upper devices to all of the lower_dev's lower_dev_list devices (and to the lower_dev itself), and vice versa. On unlink they are removed using the same logic. I've tested it with thousands vlans/bonds/bridges, everything works ok and no observable lags even on a huge number of interfaces. Memory footprint for 128 devices interconnected with each other via both upper and lower (which is impossible, but for the comparison) lists would be: 128*128*2*sizeof(netdev_adjacent) = 1.5MB but in the real world we usualy have at most several devices with slaves and a lot of vlans, so the footprint will be much lower. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-08-28 23:25:05 +02:00
net: core: Always propagate flag changes to interfaces The following commit: b6c40d68ff6498b7f63ddf97cf0aa818d748dee7 net: only invoke dev->change_rx_flags when device is UP tried to fix a problem with VLAN devices and promiscuouse flag setting. The issue was that VLAN device was setting a flag on an interface that was down, thus resulting in bad promiscuity count. This commit blocked flag propagation to any device that is currently down. A later commit: deede2fabe24e00bd7e246eb81cd5767dc6fcfc7 vlan: Don't propagate flag changes on down interfaces fixed VLAN code to only propagate flags when the VLAN interface is up, thus fixing the same issue as above, only localized to VLAN. The problem we have now is that if we have create a complex stack involving multiple software devices like bridges, bonds, and vlans, then it is possible that the flags would not propagate properly to the physical devices. A simple examle of the scenario is the following: eth0----> bond0 ----> bridge0 ---> vlan50 If bond0 or eth0 happen to be down at the time bond0 is added to the bridge, then eth0 will never have promisc mode set which is currently required for operation as part of the bridge. As a result, packets with vlan50 will be dropped by the interface. The only 2 devices that implement the special flag handling are VLAN and DSA and they both have required code to prevent incorrect flag propagation. As a result we can remove the generic solution introduced in b6c40d68ff6498b7f63ddf97cf0aa818d748dee7 and leave it to the individual devices to decide whether they will block flag propagation or not. Reported-by: Stefan Priebe <s.priebe@profihost.ag> Suggested-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: Vlad Yasevich <vyasevic@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-11-19 20:47:15 -05:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-14 18:51:31 -07:00
net: fix rtnl notification in atomic context commit 991fb3f74c "dev: always advertise rx_flags changes via netlink" introduced rtnl notification from __dev_set_promiscuity(), which can be called in atomic context. Steps to reproduce: ip tuntap add dev tap1 mode tap ifconfig tap1 up tcpdump -nei tap1 & ip tuntap del dev tap1 mode tap [ 271.627994] device tap1 left promiscuous mode [ 271.639897] BUG: sleeping function called from invalid context at mm/slub.c:940 [ 271.664491] in_atomic(): 1, irqs_disabled(): 0, pid: 3394, name: ip [ 271.677525] INFO: lockdep is turned off. [ 271.690503] CPU: 0 PID: 3394 Comm: ip Tainted: G W 3.12.0-rc3+ #73 [ 271.703996] Hardware name: System manufacturer System Product Name/P8Z77 WS, BIOS 3007 07/26/2012 [ 271.731254] ffffffff81a58506 ffff8807f0d57a58 ffffffff817544e5 ffff88082fa0f428 [ 271.760261] ffff8808071f5f40 ffff8807f0d57a88 ffffffff8108bad1 ffffffff81110ff8 [ 271.790683] 0000000000000010 00000000000000d0 00000000000000d0 ffff8807f0d57af8 [ 271.822332] Call Trace: [ 271.838234] [<ffffffff817544e5>] dump_stack+0x55/0x76 [ 271.854446] [<ffffffff8108bad1>] __might_sleep+0x181/0x240 [ 271.870836] [<ffffffff81110ff8>] ? rcu_irq_exit+0x68/0xb0 [ 271.887076] [<ffffffff811a80be>] kmem_cache_alloc_node+0x4e/0x2a0 [ 271.903368] [<ffffffff810b4ddc>] ? vprintk_emit+0x1dc/0x5a0 [ 271.919716] [<ffffffff81614d67>] ? __alloc_skb+0x57/0x2a0 [ 271.936088] [<ffffffff810b4de0>] ? vprintk_emit+0x1e0/0x5a0 [ 271.952504] [<ffffffff81614d67>] __alloc_skb+0x57/0x2a0 [ 271.968902] [<ffffffff8163a0b2>] rtmsg_ifinfo+0x52/0x100 [ 271.985302] [<ffffffff8162ac6d>] __dev_notify_flags+0xad/0xc0 [ 272.001642] [<ffffffff8162ad0c>] __dev_set_promiscuity+0x8c/0x1c0 [ 272.017917] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.033961] [<ffffffff8162b109>] dev_set_promiscuity+0x29/0x50 [ 272.049855] [<ffffffff8172e937>] packet_dev_mc+0x87/0xc0 [ 272.065494] [<ffffffff81732052>] packet_notifier+0x1b2/0x380 [ 272.080915] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.096009] [<ffffffff81761c66>] notifier_call_chain+0x66/0x150 [ 272.110803] [<ffffffff8108503e>] __raw_notifier_call_chain+0xe/0x10 [ 272.125468] [<ffffffff81085056>] raw_notifier_call_chain+0x16/0x20 [ 272.139984] [<ffffffff81620190>] call_netdevice_notifiers_info+0x40/0x70 [ 272.154523] [<ffffffff816201d6>] call_netdevice_notifiers+0x16/0x20 [ 272.168552] [<ffffffff816224c5>] rollback_registered_many+0x145/0x240 [ 272.182263] [<ffffffff81622641>] rollback_registered+0x31/0x40 [ 272.195369] [<ffffffff816229c8>] unregister_netdevice_queue+0x58/0x90 [ 272.208230] [<ffffffff81547ca0>] __tun_detach+0x140/0x340 [ 272.220686] [<ffffffff81547ed6>] tun_chr_close+0x36/0x60 Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-10-23 16:02:42 -07:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
net: Generic XDP This provides a generic SKB based non-optimized XDP path which is used if either the driver lacks a specific XDP implementation, or the user requests it via a new IFLA_XDP_FLAGS value named XDP_FLAGS_SKB_MODE. It is arguable that perhaps I should have required something like this as part of the initial XDP feature merge. I believe this is critical for two reasons: 1) Accessibility. More people can play with XDP with less dependencies. Yes I know we have XDP support in virtio_net, but that just creates another depedency for learning how to use this facility. I wrote this to make life easier for the XDP newbies. 2) As a model for what the expected semantics are. If there is a pure generic core implementation, it serves as a semantic example for driver folks adding XDP support. One thing I have not tried to address here is the issue of XDP_PACKET_HEADROOM, thanks to Daniel for spotting that. It seems incredibly expensive to do a skb_cow(skb, XDP_PACKET_HEADROOM) or whatever even if the XDP program doesn't try to push headers at all. I think we really need the verifier to somehow propagate whether certain XDP helpers are used or not. v5: - Handle both negative and positive offset after running prog - Fix mac length in XDP_TX case (Alexei) - Use rcu_dereference_protected() in free_netdev (kbuild test robot) v4: - Fix MAC header adjustmnet before calling prog (David Ahern) - Disable LRO when generic XDP is installed (Michael Chan) - Bypass qdisc et al. on XDP_TX and record the event (Alexei) - Do not perform generic XDP on reinjected packets (DaveM) v3: - Make sure XDP program sees packet at MAC header, push back MAC header if we do XDP_TX. (Alexei) - Elide GRO when generic XDP is in use. (Alexei) - Add XDP_FLAG_SKB_MODE flag which the user can use to request generic XDP even if the driver has an XDP implementation. (Alexei) - Report whether SKB mode is in use in rtnl_xdp_fill() via XDP_FLAGS attribute. (Daniel) v2: - Add some "fall through" comments in switch statements based upon feedback from Andrew Lunn - Use RCU for generic xdp_prog, thanks to Johannes Berg. Tested-by: Andy Gospodarek <andy@greyhouse.net> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Tested-by: David Ahern <dsa@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-04-18 15:36:58 -04:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
xdp: refine xdp api with regards to generic xdp While working on the iproute2 generic XDP frontend, I noticed that as of right now it's possible to have native *and* generic XDP programs loaded both at the same time for the case when a driver supports native XDP. The intended model for generic XDP from b5cdae3291f7 ("net: Generic XDP") is, however, that only one out of the two can be present at once which is also indicated as such in the XDP netlink dump part. The main rationale for generic XDP is to ease accessibility (in case a driver does not yet have XDP support) and to generically provide a semantical model as an example for driver developers wanting to add XDP support. The generic XDP option for an XDP aware driver can still be useful for comparing and testing both implementations. However, it is not intended to have a second XDP processing stage or layer with exactly the same functionality of the first native stage. Only reason could be to have a partial fallback for future XDP features that are not supported yet in the native implementation and we probably also shouldn't strive for such fallback and instead encourage native feature support in the first place. Given there's currently no such fallback issue or use case, lets not go there yet if we don't need to. Therefore, change semantics for loading XDP and bail out if the user tries to load a generic XDP program when a native one is present and vice versa. Another alternative to bailing out would be to handle the transition from one flavor to another gracefully, but that would require to bring the device down, exchange both types of programs, and bring it up again in order to avoid a tiny window where a packet could hit both hooks. Given this complicates the logic for just a debugging feature in the native case, I went with the simpler variant. For the dump, remove IFLA_XDP_FLAGS that was added with b5cdae3291f7 and reuse IFLA_XDP_ATTACHED for indicating the mode. Dumping all or just a subset of flags that were used for loading the XDP prog is suboptimal in the long run since not all flags are useful for dumping and if we start to reuse the same flag definitions for load and dump, then we'll waste bit space. What we really just want is to dump the mode for now. Current IFLA_XDP_ATTACHED semantics are: nothing was installed (0), a program is running at the native driver layer (1). Thus, add a mode that says that a program is running at generic XDP layer (2). Applications will handle this fine in that older binaries will just indicate that something is attached at XDP layer, effectively this is similar to IFLA_XDP_FLAGS attr that we would have had modulo the redundancy. Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Alexei Starovoitov <ast@kernel.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-12 01:04:46 +02:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
rtnetlink: delay RTM_DELLINK notification until after ndo_uninit() The commit 56bfa7ee7c ("unregister_netdevice : move RTM_DELLINK to until after ndo_uninit") tried to do this ealier but while doing so it created a problem. Unfortunately the delayed rtmsg_ifinfo() also delayed call to fill_info(). So this translated into asking driver to remove private state and then query it's private state. This could have catastropic consequences. This change breaks the rtmsg_ifinfo() into two parts - one takes the precise snapshot of the device by called fill_info() before calling the ndo_uninit() and the second part sends the notification using collected snapshot. It was brought to notice when last link is deleted from an ipvlan device when it has free-ed the port and the subsequent .fill_info() call is trying to get the info from the port. kernel: [ 255.139429] ------------[ cut here ]------------ kernel: [ 255.139439] WARNING: CPU: 12 PID: 11173 at net/core/rtnetlink.c:2238 rtmsg_ifinfo+0x100/0x110() kernel: [ 255.139493] Modules linked in: ipvlan bonding w1_therm ds2482 wire cdc_acm ehci_pci ehci_hcd i2c_dev i2c_i801 i2c_core msr cpuid bnx2x ptp pps_core mdio libcrc32c kernel: [ 255.139513] CPU: 12 PID: 11173 Comm: ip Not tainted 3.18.0-smp-DEV #167 kernel: [ 255.139514] Hardware name: Intel RML,PCH/Ibis_QC_18, BIOS 1.0.10 05/15/2012 kernel: [ 255.139515] 0000000000000009 ffff880851b6b828 ffffffff815d87f4 00000000000000e0 kernel: [ 255.139516] 0000000000000000 ffff880851b6b868 ffffffff8109c29c 0000000000000000 kernel: [ 255.139518] 00000000ffffffa6 00000000000000d0 ffffffff81aaf580 0000000000000011 kernel: [ 255.139520] Call Trace: kernel: [ 255.139527] [<ffffffff815d87f4>] dump_stack+0x46/0x58 kernel: [ 255.139531] [<ffffffff8109c29c>] warn_slowpath_common+0x8c/0xc0 kernel: [ 255.139540] [<ffffffff8109c2ea>] warn_slowpath_null+0x1a/0x20 kernel: [ 255.139544] [<ffffffff8150d570>] rtmsg_ifinfo+0x100/0x110 kernel: [ 255.139547] [<ffffffff814f78b5>] rollback_registered_many+0x1d5/0x2d0 kernel: [ 255.139549] [<ffffffff814f79cf>] unregister_netdevice_many+0x1f/0xb0 kernel: [ 255.139551] [<ffffffff8150acab>] rtnl_dellink+0xbb/0x110 kernel: [ 255.139553] [<ffffffff8150da90>] rtnetlink_rcv_msg+0xa0/0x240 kernel: [ 255.139557] [<ffffffff81329283>] ? rhashtable_lookup_compare+0x43/0x80 kernel: [ 255.139558] [<ffffffff8150d9f0>] ? __rtnl_unlock+0x20/0x20 kernel: [ 255.139562] [<ffffffff8152cb11>] netlink_rcv_skb+0xb1/0xc0 kernel: [ 255.139563] [<ffffffff8150a495>] rtnetlink_rcv+0x25/0x40 kernel: [ 255.139565] [<ffffffff8152c398>] netlink_unicast+0x178/0x230 kernel: [ 255.139567] [<ffffffff8152c75f>] netlink_sendmsg+0x30f/0x420 kernel: [ 255.139571] [<ffffffff814e0b0c>] sock_sendmsg+0x9c/0xd0 kernel: [ 255.139575] [<ffffffff811d1d7f>] ? rw_copy_check_uvector+0x6f/0x130 kernel: [ 255.139577] [<ffffffff814e11c9>] ? copy_msghdr_from_user+0x139/0x1b0 kernel: [ 255.139578] [<ffffffff814e1774>] ___sys_sendmsg+0x304/0x310 kernel: [ 255.139581] [<ffffffff81198723>] ? handle_mm_fault+0xca3/0xde0 kernel: [ 255.139585] [<ffffffff811ebc4c>] ? destroy_inode+0x3c/0x70 kernel: [ 255.139589] [<ffffffff8108e6ec>] ? __do_page_fault+0x20c/0x500 kernel: [ 255.139597] [<ffffffff811e8336>] ? dput+0xb6/0x190 kernel: [ 255.139606] [<ffffffff811f05f6>] ? mntput+0x26/0x40 kernel: [ 255.139611] [<ffffffff811d2b94>] ? __fput+0x174/0x1e0 kernel: [ 255.139613] [<ffffffff814e2129>] __sys_sendmsg+0x49/0x90 kernel: [ 255.139615] [<ffffffff814e2182>] SyS_sendmsg+0x12/0x20 kernel: [ 255.139617] [<ffffffff815df092>] system_call_fastpath+0x12/0x17 kernel: [ 255.139619] ---[ end trace 5e6703e87d984f6b ]--- Signed-off-by: Mahesh Bandewar <maheshb@google.com> Reported-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp> Cc: Eric Dumazet <edumazet@google.com> Cc: Roopa Prabhu <roopa@cumulusnetworks.com> Cc: David S. Miller <davem@davemloft.net> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Thomas Graf <tgraf@suug.ch> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-03 13:46:24 -08:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
rtnetlink: delay RTM_DELLINK notification until after ndo_uninit() The commit 56bfa7ee7c ("unregister_netdevice : move RTM_DELLINK to until after ndo_uninit") tried to do this ealier but while doing so it created a problem. Unfortunately the delayed rtmsg_ifinfo() also delayed call to fill_info(). So this translated into asking driver to remove private state and then query it's private state. This could have catastropic consequences. This change breaks the rtmsg_ifinfo() into two parts - one takes the precise snapshot of the device by called fill_info() before calling the ndo_uninit() and the second part sends the notification using collected snapshot. It was brought to notice when last link is deleted from an ipvlan device when it has free-ed the port and the subsequent .fill_info() call is trying to get the info from the port. kernel: [ 255.139429] ------------[ cut here ]------------ kernel: [ 255.139439] WARNING: CPU: 12 PID: 11173 at net/core/rtnetlink.c:2238 rtmsg_ifinfo+0x100/0x110() kernel: [ 255.139493] Modules linked in: ipvlan bonding w1_therm ds2482 wire cdc_acm ehci_pci ehci_hcd i2c_dev i2c_i801 i2c_core msr cpuid bnx2x ptp pps_core mdio libcrc32c kernel: [ 255.139513] CPU: 12 PID: 11173 Comm: ip Not tainted 3.18.0-smp-DEV #167 kernel: [ 255.139514] Hardware name: Intel RML,PCH/Ibis_QC_18, BIOS 1.0.10 05/15/2012 kernel: [ 255.139515] 0000000000000009 ffff880851b6b828 ffffffff815d87f4 00000000000000e0 kernel: [ 255.139516] 0000000000000000 ffff880851b6b868 ffffffff8109c29c 0000000000000000 kernel: [ 255.139518] 00000000ffffffa6 00000000000000d0 ffffffff81aaf580 0000000000000011 kernel: [ 255.139520] Call Trace: kernel: [ 255.139527] [<ffffffff815d87f4>] dump_stack+0x46/0x58 kernel: [ 255.139531] [<ffffffff8109c29c>] warn_slowpath_common+0x8c/0xc0 kernel: [ 255.139540] [<ffffffff8109c2ea>] warn_slowpath_null+0x1a/0x20 kernel: [ 255.139544] [<ffffffff8150d570>] rtmsg_ifinfo+0x100/0x110 kernel: [ 255.139547] [<ffffffff814f78b5>] rollback_registered_many+0x1d5/0x2d0 kernel: [ 255.139549] [<ffffffff814f79cf>] unregister_netdevice_many+0x1f/0xb0 kernel: [ 255.139551] [<ffffffff8150acab>] rtnl_dellink+0xbb/0x110 kernel: [ 255.139553] [<ffffffff8150da90>] rtnetlink_rcv_msg+0xa0/0x240 kernel: [ 255.139557] [<ffffffff81329283>] ? rhashtable_lookup_compare+0x43/0x80 kernel: [ 255.139558] [<ffffffff8150d9f0>] ? __rtnl_unlock+0x20/0x20 kernel: [ 255.139562] [<ffffffff8152cb11>] netlink_rcv_skb+0xb1/0xc0 kernel: [ 255.139563] [<ffffffff8150a495>] rtnetlink_rcv+0x25/0x40 kernel: [ 255.139565] [<ffffffff8152c398>] netlink_unicast+0x178/0x230 kernel: [ 255.139567] [<ffffffff8152c75f>] netlink_sendmsg+0x30f/0x420 kernel: [ 255.139571] [<ffffffff814e0b0c>] sock_sendmsg+0x9c/0xd0 kernel: [ 255.139575] [<ffffffff811d1d7f>] ? rw_copy_check_uvector+0x6f/0x130 kernel: [ 255.139577] [<ffffffff814e11c9>] ? copy_msghdr_from_user+0x139/0x1b0 kernel: [ 255.139578] [<ffffffff814e1774>] ___sys_sendmsg+0x304/0x310 kernel: [ 255.139581] [<ffffffff81198723>] ? handle_mm_fault+0xca3/0xde0 kernel: [ 255.139585] [<ffffffff811ebc4c>] ? destroy_inode+0x3c/0x70 kernel: [ 255.139589] [<ffffffff8108e6ec>] ? __do_page_fault+0x20c/0x500 kernel: [ 255.139597] [<ffffffff811e8336>] ? dput+0xb6/0x190 kernel: [ 255.139606] [<ffffffff811f05f6>] ? mntput+0x26/0x40 kernel: [ 255.139611] [<ffffffff811d2b94>] ? __fput+0x174/0x1e0 kernel: [ 255.139613] [<ffffffff814e2129>] __sys_sendmsg+0x49/0x90 kernel: [ 255.139615] [<ffffffff814e2182>] SyS_sendmsg+0x12/0x20 kernel: [ 255.139617] [<ffffffff815df092>] system_call_fastpath+0x12/0x17 kernel: [ 255.139619] ---[ end trace 5e6703e87d984f6b ]--- Signed-off-by: Mahesh Bandewar <maheshb@google.com> Reported-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp> Cc: Eric Dumazet <edumazet@google.com> Cc: Roopa Prabhu <roopa@cumulusnetworks.com> Cc: David S. Miller <davem@davemloft.net> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Thomas Graf <tgraf@suug.ch> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-03 13:46:24 -08:00
rtnetlink: delay RTM_DELLINK notification until after ndo_uninit() The commit 56bfa7ee7c ("unregister_netdevice : move RTM_DELLINK to until after ndo_uninit") tried to do this ealier but while doing so it created a problem. Unfortunately the delayed rtmsg_ifinfo() also delayed call to fill_info(). So this translated into asking driver to remove private state and then query it's private state. This could have catastropic consequences. This change breaks the rtmsg_ifinfo() into two parts - one takes the precise snapshot of the device by called fill_info() before calling the ndo_uninit() and the second part sends the notification using collected snapshot. It was brought to notice when last link is deleted from an ipvlan device when it has free-ed the port and the subsequent .fill_info() call is trying to get the info from the port. kernel: [ 255.139429] ------------[ cut here ]------------ kernel: [ 255.139439] WARNING: CPU: 12 PID: 11173 at net/core/rtnetlink.c:2238 rtmsg_ifinfo+0x100/0x110() kernel: [ 255.139493] Modules linked in: ipvlan bonding w1_therm ds2482 wire cdc_acm ehci_pci ehci_hcd i2c_dev i2c_i801 i2c_core msr cpuid bnx2x ptp pps_core mdio libcrc32c kernel: [ 255.139513] CPU: 12 PID: 11173 Comm: ip Not tainted 3.18.0-smp-DEV #167 kernel: [ 255.139514] Hardware name: Intel RML,PCH/Ibis_QC_18, BIOS 1.0.10 05/15/2012 kernel: [ 255.139515] 0000000000000009 ffff880851b6b828 ffffffff815d87f4 00000000000000e0 kernel: [ 255.139516] 0000000000000000 ffff880851b6b868 ffffffff8109c29c 0000000000000000 kernel: [ 255.139518] 00000000ffffffa6 00000000000000d0 ffffffff81aaf580 0000000000000011 kernel: [ 255.139520] Call Trace: kernel: [ 255.139527] [<ffffffff815d87f4>] dump_stack+0x46/0x58 kernel: [ 255.139531] [<ffffffff8109c29c>] warn_slowpath_common+0x8c/0xc0 kernel: [ 255.139540] [<ffffffff8109c2ea>] warn_slowpath_null+0x1a/0x20 kernel: [ 255.139544] [<ffffffff8150d570>] rtmsg_ifinfo+0x100/0x110 kernel: [ 255.139547] [<ffffffff814f78b5>] rollback_registered_many+0x1d5/0x2d0 kernel: [ 255.139549] [<ffffffff814f79cf>] unregister_netdevice_many+0x1f/0xb0 kernel: [ 255.139551] [<ffffffff8150acab>] rtnl_dellink+0xbb/0x110 kernel: [ 255.139553] [<ffffffff8150da90>] rtnetlink_rcv_msg+0xa0/0x240 kernel: [ 255.139557] [<ffffffff81329283>] ? rhashtable_lookup_compare+0x43/0x80 kernel: [ 255.139558] [<ffffffff8150d9f0>] ? __rtnl_unlock+0x20/0x20 kernel: [ 255.139562] [<ffffffff8152cb11>] netlink_rcv_skb+0xb1/0xc0 kernel: [ 255.139563] [<ffffffff8150a495>] rtnetlink_rcv+0x25/0x40 kernel: [ 255.139565] [<ffffffff8152c398>] netlink_unicast+0x178/0x230 kernel: [ 255.139567] [<ffffffff8152c75f>] netlink_sendmsg+0x30f/0x420 kernel: [ 255.139571] [<ffffffff814e0b0c>] sock_sendmsg+0x9c/0xd0 kernel: [ 255.139575] [<ffffffff811d1d7f>] ? rw_copy_check_uvector+0x6f/0x130 kernel: [ 255.139577] [<ffffffff814e11c9>] ? copy_msghdr_from_user+0x139/0x1b0 kernel: [ 255.139578] [<ffffffff814e1774>] ___sys_sendmsg+0x304/0x310 kernel: [ 255.139581] [<ffffffff81198723>] ? handle_mm_fault+0xca3/0xde0 kernel: [ 255.139585] [<ffffffff811ebc4c>] ? destroy_inode+0x3c/0x70 kernel: [ 255.139589] [<ffffffff8108e6ec>] ? __do_page_fault+0x20c/0x500 kernel: [ 255.139597] [<ffffffff811e8336>] ? dput+0xb6/0x190 kernel: [ 255.139606] [<ffffffff811f05f6>] ? mntput+0x26/0x40 kernel: [ 255.139611] [<ffffffff811d2b94>] ? __fput+0x174/0x1e0 kernel: [ 255.139613] [<ffffffff814e2129>] __sys_sendmsg+0x49/0x90 kernel: [ 255.139615] [<ffffffff814e2182>] SyS_sendmsg+0x12/0x20 kernel: [ 255.139617] [<ffffffff815df092>] system_call_fastpath+0x12/0x17 kernel: [ 255.139619] ---[ end trace 5e6703e87d984f6b ]--- Signed-off-by: Mahesh Bandewar <maheshb@google.com> Reported-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp> Cc: Eric Dumazet <edumazet@google.com> Cc: Roopa Prabhu <roopa@cumulusnetworks.com> Cc: David S. Miller <davem@davemloft.net> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Thomas Graf <tgraf@suug.ch> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-03 13:46:24 -08:00
rtnetlink: delay RTM_DELLINK notification until after ndo_uninit() The commit 56bfa7ee7c ("unregister_netdevice : move RTM_DELLINK to until after ndo_uninit") tried to do this ealier but while doing so it created a problem. Unfortunately the delayed rtmsg_ifinfo() also delayed call to fill_info(). So this translated into asking driver to remove private state and then query it's private state. This could have catastropic consequences. This change breaks the rtmsg_ifinfo() into two parts - one takes the precise snapshot of the device by called fill_info() before calling the ndo_uninit() and the second part sends the notification using collected snapshot. It was brought to notice when last link is deleted from an ipvlan device when it has free-ed the port and the subsequent .fill_info() call is trying to get the info from the port. kernel: [ 255.139429] ------------[ cut here ]------------ kernel: [ 255.139439] WARNING: CPU: 12 PID: 11173 at net/core/rtnetlink.c:2238 rtmsg_ifinfo+0x100/0x110() kernel: [ 255.139493] Modules linked in: ipvlan bonding w1_therm ds2482 wire cdc_acm ehci_pci ehci_hcd i2c_dev i2c_i801 i2c_core msr cpuid bnx2x ptp pps_core mdio libcrc32c kernel: [ 255.139513] CPU: 12 PID: 11173 Comm: ip Not tainted 3.18.0-smp-DEV #167 kernel: [ 255.139514] Hardware name: Intel RML,PCH/Ibis_QC_18, BIOS 1.0.10 05/15/2012 kernel: [ 255.139515] 0000000000000009 ffff880851b6b828 ffffffff815d87f4 00000000000000e0 kernel: [ 255.139516] 0000000000000000 ffff880851b6b868 ffffffff8109c29c 0000000000000000 kernel: [ 255.139518] 00000000ffffffa6 00000000000000d0 ffffffff81aaf580 0000000000000011 kernel: [ 255.139520] Call Trace: kernel: [ 255.139527] [<ffffffff815d87f4>] dump_stack+0x46/0x58 kernel: [ 255.139531] [<ffffffff8109c29c>] warn_slowpath_common+0x8c/0xc0 kernel: [ 255.139540] [<ffffffff8109c2ea>] warn_slowpath_null+0x1a/0x20 kernel: [ 255.139544] [<ffffffff8150d570>] rtmsg_ifinfo+0x100/0x110 kernel: [ 255.139547] [<ffffffff814f78b5>] rollback_registered_many+0x1d5/0x2d0 kernel: [ 255.139549] [<ffffffff814f79cf>] unregister_netdevice_many+0x1f/0xb0 kernel: [ 255.139551] [<ffffffff8150acab>] rtnl_dellink+0xbb/0x110 kernel: [ 255.139553] [<ffffffff8150da90>] rtnetlink_rcv_msg+0xa0/0x240 kernel: [ 255.139557] [<ffffffff81329283>] ? rhashtable_lookup_compare+0x43/0x80 kernel: [ 255.139558] [<ffffffff8150d9f0>] ? __rtnl_unlock+0x20/0x20 kernel: [ 255.139562] [<ffffffff8152cb11>] netlink_rcv_skb+0xb1/0xc0 kernel: [ 255.139563] [<ffffffff8150a495>] rtnetlink_rcv+0x25/0x40 kernel: [ 255.139565] [<ffffffff8152c398>] netlink_unicast+0x178/0x230 kernel: [ 255.139567] [<ffffffff8152c75f>] netlink_sendmsg+0x30f/0x420 kernel: [ 255.139571] [<ffffffff814e0b0c>] sock_sendmsg+0x9c/0xd0 kernel: [ 255.139575] [<ffffffff811d1d7f>] ? rw_copy_check_uvector+0x6f/0x130 kernel: [ 255.139577] [<ffffffff814e11c9>] ? copy_msghdr_from_user+0x139/0x1b0 kernel: [ 255.139578] [<ffffffff814e1774>] ___sys_sendmsg+0x304/0x310 kernel: [ 255.139581] [<ffffffff81198723>] ? handle_mm_fault+0xca3/0xde0 kernel: [ 255.139585] [<ffffffff811ebc4c>] ? destroy_inode+0x3c/0x70 kernel: [ 255.139589] [<ffffffff8108e6ec>] ? __do_page_fault+0x20c/0x500 kernel: [ 255.139597] [<ffffffff811e8336>] ? dput+0xb6/0x190 kernel: [ 255.139606] [<ffffffff811f05f6>] ? mntput+0x26/0x40 kernel: [ 255.139611] [<ffffffff811d2b94>] ? __fput+0x174/0x1e0 kernel: [ 255.139613] [<ffffffff814e2129>] __sys_sendmsg+0x49/0x90 kernel: [ 255.139615] [<ffffffff814e2182>] SyS_sendmsg+0x12/0x20 kernel: [ 255.139617] [<ffffffff815df092>] system_call_fastpath+0x12/0x17 kernel: [ 255.139619] ---[ end trace 5e6703e87d984f6b ]--- Signed-off-by: Mahesh Bandewar <maheshb@google.com> Reported-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp> Cc: Eric Dumazet <edumazet@google.com> Cc: Roopa Prabhu <roopa@cumulusnetworks.com> Cc: David S. Miller <davem@davemloft.net> Acked-by: Eric Dumazet <edumazet@google.com> Acked-by: Thomas Graf <tgraf@suug.ch> Signed-off-by: David S. Miller <davem@davemloft.net>
2014-12-03 13:46:24 -08:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: fix for_each_netdev_feature As pointed out by Nikolay and further explained by Geert, the initial for_each_netdev_feature macro was broken, as feature would get set outside of the block of code it was intended to run in, thus only ever working for the first feature bit in the mask. While less pretty this way, this is tested and confirmed functional with multiple feature bits set in NETIF_F_UPPER_DISABLES. [root@dell-per730-01 ~]# ethtool -K bond0 lro off ... [ 242.761394] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 243.552178] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 244.353978] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 245.147420] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 [root@dell-per730-01 ~]# ethtool -K bond0 gro off ... [ 251.925645] bond0: Disabling feature 0x0000000000004000 on lower dev p5p2. [ 252.713693] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 253.499085] bond0: Disabling feature 0x0000000000004000 on lower dev p5p1. [ 254.290922] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 Fixes: fd867d51f ("net/core: generic support for disabling netdev features down stack") CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: Geert Uytterhoeven <geert@linux-m68k.org> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Acked-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-03 10:15:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: fix for_each_netdev_feature As pointed out by Nikolay and further explained by Geert, the initial for_each_netdev_feature macro was broken, as feature would get set outside of the block of code it was intended to run in, thus only ever working for the first feature bit in the mask. While less pretty this way, this is tested and confirmed functional with multiple feature bits set in NETIF_F_UPPER_DISABLES. [root@dell-per730-01 ~]# ethtool -K bond0 lro off ... [ 242.761394] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 243.552178] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 244.353978] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 245.147420] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 [root@dell-per730-01 ~]# ethtool -K bond0 gro off ... [ 251.925645] bond0: Disabling feature 0x0000000000004000 on lower dev p5p2. [ 252.713693] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 253.499085] bond0: Disabling feature 0x0000000000004000 on lower dev p5p1. [ 254.290922] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 Fixes: fd867d51f ("net/core: generic support for disabling netdev features down stack") CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: Geert Uytterhoeven <geert@linux-m68k.org> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Acked-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-03 10:15:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: fix for_each_netdev_feature As pointed out by Nikolay and further explained by Geert, the initial for_each_netdev_feature macro was broken, as feature would get set outside of the block of code it was intended to run in, thus only ever working for the first feature bit in the mask. While less pretty this way, this is tested and confirmed functional with multiple feature bits set in NETIF_F_UPPER_DISABLES. [root@dell-per730-01 ~]# ethtool -K bond0 lro off ... [ 242.761394] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 243.552178] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 244.353978] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 245.147420] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 [root@dell-per730-01 ~]# ethtool -K bond0 gro off ... [ 251.925645] bond0: Disabling feature 0x0000000000004000 on lower dev p5p2. [ 252.713693] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 253.499085] bond0: Disabling feature 0x0000000000004000 on lower dev p5p1. [ 254.290922] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 Fixes: fd867d51f ("net/core: generic support for disabling netdev features down stack") CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: Geert Uytterhoeven <geert@linux-m68k.org> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Acked-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-03 10:15:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: fix for_each_netdev_feature As pointed out by Nikolay and further explained by Geert, the initial for_each_netdev_feature macro was broken, as feature would get set outside of the block of code it was intended to run in, thus only ever working for the first feature bit in the mask. While less pretty this way, this is tested and confirmed functional with multiple feature bits set in NETIF_F_UPPER_DISABLES. [root@dell-per730-01 ~]# ethtool -K bond0 lro off ... [ 242.761394] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 243.552178] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 244.353978] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 245.147420] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 [root@dell-per730-01 ~]# ethtool -K bond0 gro off ... [ 251.925645] bond0: Disabling feature 0x0000000000004000 on lower dev p5p2. [ 252.713693] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 253.499085] bond0: Disabling feature 0x0000000000004000 on lower dev p5p1. [ 254.290922] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 Fixes: fd867d51f ("net/core: generic support for disabling netdev features down stack") CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: Geert Uytterhoeven <geert@linux-m68k.org> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Acked-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-03 10:15:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
net/core: generic support for disabling netdev features down stack There are some netdev features, which when disabled on an upper device, such as a bonding master or a bridge, must be disabled and cannot be re-enabled on underlying devices. This is a rework of an earlier more heavy-handed appraoch, which simply disables and prevents re-enabling of netdev features listed in a new define in include/net/netdev_features.h, NETIF_F_UPPER_DISABLES. Any upper device that disables a flag in that feature mask, the disabling will propagate down the stack, and any lower device that has any upper device with one of those flags disabled should not be able to enable said flag. Initially, only LRO is included for proof of concept, and because this code effectively does the same thing as dev_disable_lro(), though it will also activate from the ethtool path, which was one of the goals here. [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: on [root@dell-per730-01 ~]# ethtool -K bond0 lro off [root@dell-per730-01 ~]# ethtool -k bond0 |grep large large-receive-offload: off [root@dell-per730-01 ~]# ethtool -k p5p1 |grep large large-receive-offload: off dmesg dump: [ 1033.277986] bond0: Disabling feature 0x0000000000008000 on lower dev p5p2. [ 1034.067949] bnx2x 0000:06:00.1 p5p2: using MSI-X IRQs: sp 74 fp[0] 76 ... fp[7] 83 [ 1034.753612] bond0: Disabling feature 0x0000000000008000 on lower dev p5p1. [ 1035.591019] bnx2x 0000:06:00.0 p5p1: using MSI-X IRQs: sp 62 fp[0] 64 ... fp[7] 71 This has been successfully tested with bnx2x, qlcnic and netxen network cards as slaves in a bond interface. Turning LRO on or off on the master also turns it on or off on each of the slaves, new slaves are added with LRO in the same state as the master, and LRO can't be toggled on the slaves. Also, this should largely remove the need for dev_disable_lro(), and most, if not all, of its call sites can be replaced by simply making sure NETIF_F_LRO isn't included in the relevant device's feature flags. Note that this patch is driven by bug reports from users saying it was confusing that bonds and slaves had different settings for the same features, and while it won't be 100% in sync if a lower device doesn't support a feature like LRO, I think this is a good step in the right direction. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jay Vosburgh <j.vosburgh@gmail.com> CC: Veaceslav Falico <vfalico@gmail.com> CC: Andy Gospodarek <gospo@cumulusnetworks.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Nikolay Aleksandrov <razor@blackwall.org> CC: Michal Kubecek <mkubecek@suse.cz> CC: Alexander Duyck <alexander.duyck@gmail.com> CC: netdev@vger.kernel.org Signed-off-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-11-02 21:55:59 -05:00
mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic __GFP_REPEAT was designed to allow retry-but-eventually-fail semantic to the page allocator. This has been true but only for allocations requests larger than PAGE_ALLOC_COSTLY_ORDER. It has been always ignored for smaller sizes. This is a bit unfortunate because there is no way to express the same semantic for those requests and they are considered too important to fail so they might end up looping in the page allocator for ever, similarly to GFP_NOFAIL requests. Now that the whole tree has been cleaned up and accidental or misled usage of __GFP_REPEAT flag has been removed for !costly requests we can give the original flag a better name and more importantly a more useful semantic. Let's rename it to __GFP_RETRY_MAYFAIL which tells the user that the allocator would try really hard but there is no promise of a success. This will work independent of the order and overrides the default allocator behavior. Page allocator users have several levels of guarantee vs. cost options (take GFP_KERNEL as an example) - GFP_KERNEL & ~__GFP_RECLAIM - optimistic allocation without _any_ attempt to free memory at all. The most light weight mode which even doesn't kick the background reclaim. Should be used carefully because it might deplete the memory and the next user might hit the more aggressive reclaim - GFP_KERNEL & ~__GFP_DIRECT_RECLAIM (or GFP_NOWAIT)- optimistic allocation without any attempt to free memory from the current context but can wake kswapd to reclaim memory if the zone is below the low watermark. Can be used from either atomic contexts or when the request is a performance optimization and there is another fallback for a slow path. - (GFP_KERNEL|__GFP_HIGH) & ~__GFP_DIRECT_RECLAIM (aka GFP_ATOMIC) - non sleeping allocation with an expensive fallback so it can access some portion of memory reserves. Usually used from interrupt/bh context with an expensive slow path fallback. - GFP_KERNEL - both background and direct reclaim are allowed and the _default_ page allocator behavior is used. That means that !costly allocation requests are basically nofail but there is no guarantee of that behavior so failures have to be checked properly by callers (e.g. OOM killer victim is allowed to fail currently). - GFP_KERNEL | __GFP_NORETRY - overrides the default allocator behavior and all allocation requests fail early rather than cause disruptive reclaim (one round of reclaim in this implementation). The OOM killer is not invoked. - GFP_KERNEL | __GFP_RETRY_MAYFAIL - overrides the default allocator behavior and all allocation requests try really hard. The request will fail if the reclaim cannot make any progress. The OOM killer won't be triggered. - GFP_KERNEL | __GFP_NOFAIL - overrides the default allocator behavior and all allocation requests will loop endlessly until they succeed. This might be really dangerous especially for larger orders. Existing users of __GFP_REPEAT are changed to __GFP_RETRY_MAYFAIL because they already had their semantic. No new users are added. __alloc_pages_slowpath is changed to bail out for __GFP_RETRY_MAYFAIL if there is no progress and we have already passed the OOM point. This means that all the reclaim opportunities have been exhausted except the most disruptive one (the OOM killer) and a user defined fallback behavior is more sensible than keep retrying in the page allocator. [akpm@linux-foundation.org: fix arch/sparc/kernel/mdesc.c] [mhocko@suse.com: semantic fix] Link: http://lkml.kernel.org/r/20170626123847.GM11534@dhcp22.suse.cz [mhocko@kernel.org: address other thing spotted by Vlastimil] Link: http://lkml.kernel.org/r/20170626124233.GN11534@dhcp22.suse.cz Link: http://lkml.kernel.org/r/20170623085345.11304-3-mhocko@kernel.org Signed-off-by: Michal Hocko <mhocko@suse.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Alex Belits <alex.belits@cavium.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Christoph Hellwig <hch@infradead.org> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: David Daney <david.daney@cavium.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: NeilBrown <neilb@suse.com> Cc: Ralf Baechle <ralf@linux-mips.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-12 14:36:45 -07:00
mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic __GFP_REPEAT was designed to allow retry-but-eventually-fail semantic to the page allocator. This has been true but only for allocations requests larger than PAGE_ALLOC_COSTLY_ORDER. It has been always ignored for smaller sizes. This is a bit unfortunate because there is no way to express the same semantic for those requests and they are considered too important to fail so they might end up looping in the page allocator for ever, similarly to GFP_NOFAIL requests. Now that the whole tree has been cleaned up and accidental or misled usage of __GFP_REPEAT flag has been removed for !costly requests we can give the original flag a better name and more importantly a more useful semantic. Let's rename it to __GFP_RETRY_MAYFAIL which tells the user that the allocator would try really hard but there is no promise of a success. This will work independent of the order and overrides the default allocator behavior. Page allocator users have several levels of guarantee vs. cost options (take GFP_KERNEL as an example) - GFP_KERNEL & ~__GFP_RECLAIM - optimistic allocation without _any_ attempt to free memory at all. The most light weight mode which even doesn't kick the background reclaim. Should be used carefully because it might deplete the memory and the next user might hit the more aggressive reclaim - GFP_KERNEL & ~__GFP_DIRECT_RECLAIM (or GFP_NOWAIT)- optimistic allocation without any attempt to free memory from the current context but can wake kswapd to reclaim memory if the zone is below the low watermark. Can be used from either atomic contexts or when the request is a performance optimization and there is another fallback for a slow path. - (GFP_KERNEL|__GFP_HIGH) & ~__GFP_DIRECT_RECLAIM (aka GFP_ATOMIC) - non sleeping allocation with an expensive fallback so it can access some portion of memory reserves. Usually used from interrupt/bh context with an expensive slow path fallback. - GFP_KERNEL - both background and direct reclaim are allowed and the _default_ page allocator behavior is used. That means that !costly allocation requests are basically nofail but there is no guarantee of that behavior so failures have to be checked properly by callers (e.g. OOM killer victim is allowed to fail currently). - GFP_KERNEL | __GFP_NORETRY - overrides the default allocator behavior and all allocation requests fail early rather than cause disruptive reclaim (one round of reclaim in this implementation). The OOM killer is not invoked. - GFP_KERNEL | __GFP_RETRY_MAYFAIL - overrides the default allocator behavior and all allocation requests try really hard. The request will fail if the reclaim cannot make any progress. The OOM killer won't be triggered. - GFP_KERNEL | __GFP_NOFAIL - overrides the default allocator behavior and all allocation requests will loop endlessly until they succeed. This might be really dangerous especially for larger orders. Existing users of __GFP_REPEAT are changed to __GFP_RETRY_MAYFAIL because they already had their semantic. No new users are added. __alloc_pages_slowpath is changed to bail out for __GFP_RETRY_MAYFAIL if there is no progress and we have already passed the OOM point. This means that all the reclaim opportunities have been exhausted except the most disruptive one (the OOM killer) and a user defined fallback behavior is more sensible than keep retrying in the page allocator. [akpm@linux-foundation.org: fix arch/sparc/kernel/mdesc.c] [mhocko@suse.com: semantic fix] Link: http://lkml.kernel.org/r/20170626123847.GM11534@dhcp22.suse.cz [mhocko@kernel.org: address other thing spotted by Vlastimil] Link: http://lkml.kernel.org/r/20170626124233.GN11534@dhcp22.suse.cz Link: http://lkml.kernel.org/r/20170623085345.11304-3-mhocko@kernel.org Signed-off-by: Michal Hocko <mhocko@suse.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Alex Belits <alex.belits@cavium.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Christoph Hellwig <hch@infradead.org> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: David Daney <david.daney@cavium.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: NeilBrown <neilb@suse.com> Cc: Ralf Baechle <ralf@linux-mips.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-12 14:36:45 -07:00
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
2015-05-11 21:17:53 +02:00
2015-05-11 21:17:53 +02:00
net:set valid name before calling ndo_init() In commit 1c5cae815d19 (net: call dev_alloc_name from register_netdevice), a bug of bonding was involved, see example 1 and 2. In register_netdevice(), the name of net_device is not valid until dev_get_valid_name() is called. But dev->netdev_ops->ndo_init(that is bond_init) is called before dev_get_valid_name(), and it uses the invalid name of net_device. I think register_netdevice() should make sure that the name of net_device is valid before calling ndo_init(). example 1: modprobe bonding ls /proc/net/bonding/bond%d ps -eLf root 3398 2 3398 0 1 21:34 ? 00:00:00 [bond%d] example 2: modprobe bonding max_bonds=3 [ 170.100292] bonding: Ethernet Channel Bonding Driver: v3.7.1 (April 27, 2011) [ 170.101090] bonding: Warning: either miimon or arp_interval and arp_ip_target module parameters must be specified, otherwise bonding will not detect link failures! see bonding.txt for details. [ 170.102469] ------------[ cut here ]------------ [ 170.103150] WARNING: at /home/pwp/net-next-2.6/fs/proc/generic.c:586 proc_register+0x126/0x157() [ 170.104075] Hardware name: VirtualBox [ 170.105065] proc_dir_entry 'bonding/bond%d' already registered [ 170.105613] Modules linked in: bonding(+) sunrpc ipv6 uinput microcode ppdev parport_pc parport joydev e1000 pcspkr i2c_piix4 i2c_core [last unloaded: bonding] [ 170.108397] Pid: 3457, comm: modprobe Not tainted 2.6.39-rc2+ #14 [ 170.108935] Call Trace: [ 170.109382] [<c0438f3b>] warn_slowpath_common+0x6a/0x7f [ 170.109911] [<c051a42a>] ? proc_register+0x126/0x157 [ 170.110329] [<c0438fc3>] warn_slowpath_fmt+0x2b/0x2f [ 170.110846] [<c051a42a>] proc_register+0x126/0x157 [ 170.111870] [<c051a4dd>] proc_create_data+0x82/0x98 [ 170.112335] [<f94e6af6>] bond_create_proc_entry+0x3f/0x73 [bonding] [ 170.112905] [<f94dd806>] bond_init+0x77/0xa5 [bonding] [ 170.113319] [<c0721ac6>] register_netdevice+0x8c/0x1d3 [ 170.113848] [<f94e0e30>] bond_create+0x6c/0x90 [bonding] [ 170.114322] [<f94f4763>] bonding_init+0x763/0x7b1 [bonding] [ 170.114879] [<c0401240>] do_one_initcall+0x76/0x122 [ 170.115317] [<f94f4000>] ? 0xf94f3fff [ 170.115799] [<c0463f1e>] sys_init_module+0x1286/0x140d [ 170.116879] [<c07c6d9f>] sysenter_do_call+0x12/0x28 [ 170.117404] ---[ end trace 64e4fac3ae5fff1a ]--- [ 170.117924] bond%d: Warning: failed to register to debugfs [ 170.128728] ------------[ cut here ]------------ [ 170.129360] WARNING: at /home/pwp/net-next-2.6/fs/proc/generic.c:586 proc_register+0x126/0x157() [ 170.130323] Hardware name: VirtualBox [ 170.130797] proc_dir_entry 'bonding/bond%d' already registered [ 170.131315] Modules linked in: bonding(+) sunrpc ipv6 uinput microcode ppdev parport_pc parport joydev e1000 pcspkr i2c_piix4 i2c_core [last unloaded: bonding] [ 170.133731] Pid: 3457, comm: modprobe Tainted: G W 2.6.39-rc2+ #14 [ 170.134308] Call Trace: [ 170.134743] [<c0438f3b>] warn_slowpath_common+0x6a/0x7f [ 170.135305] [<c051a42a>] ? proc_register+0x126/0x157 [ 170.135820] [<c0438fc3>] warn_slowpath_fmt+0x2b/0x2f [ 170.137168] [<c051a42a>] proc_register+0x126/0x157 [ 170.137700] [<c051a4dd>] proc_create_data+0x82/0x98 [ 170.138174] [<f94e6af6>] bond_create_proc_entry+0x3f/0x73 [bonding] [ 170.138745] [<f94dd806>] bond_init+0x77/0xa5 [bonding] [ 170.139278] [<c0721ac6>] register_netdevice+0x8c/0x1d3 [ 170.139828] [<f94e0e30>] bond_create+0x6c/0x90 [bonding] [ 170.140361] [<f94f4763>] bonding_init+0x763/0x7b1 [bonding] [ 170.140927] [<c0401240>] do_one_initcall+0x76/0x122 [ 170.141494] [<f94f4000>] ? 0xf94f3fff [ 170.141975] [<c0463f1e>] sys_init_module+0x1286/0x140d [ 170.142463] [<c07c6d9f>] sysenter_do_call+0x12/0x28 [ 170.142974] ---[ end trace 64e4fac3ae5fff1b ]--- [ 170.144949] bond%d: Warning: failed to register to debugfs Signed-off-by: Weiping Pan <panweiping3@gmail.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-05-12 15:46:56 +00:00
net: Allow no-cache copy from user on transmit This patch uses __copy_from_user_nocache on transmit to bypass data cache for a performance improvement. skb_add_data_nocache and skb_copy_to_page_nocache can be called by sendmsg functions to use this feature, initial support is in tcp_sendmsg. This functionality is configurable per device using ethtool. Presumably, this feature would only be useful when the driver does not touch the data. The feature is turned on by default if a device indicates that it does some form of checksum offload; it is off by default for devices that do no checksum offload or indicate no checksum is necessary. For the former case copy-checksum is probably done anyway, in the latter case the device is likely loopback in which case the no cache copy is probably not beneficial. This patch was tested using 200 instances of netperf TCP_RR with 1400 byte request and one byte reply. Platform is 16 core AMD x86. No-cache copy disabled: 672703 tps, 97.13% utilization 50/90/99% latency:244.31 484.205 1028.41 No-cache copy enabled: 702113 tps, 96.16% utilization, 50/90/99% latency 238.56 467.56 956.955 Using 14000 byte request and response sizes demonstrate the effects more dramatically: No-cache copy disabled: 79571 tps, 34.34 %utlization 50/90/95% latency 1584.46 2319.59 5001.76 No-cache copy enabled: 83856 tps, 34.81% utilization 50/90/95% latency 2508.42 2622.62 2735.88 Note especially the effect on latency tail (95th percentile). This seems to provide a nice performance improvement and is consistent in the tests I ran. Presumably, this would provide the greatest benfits in the presence of an application workload stressing the cache and a lot of transmit data happening. Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-04-04 22:30:30 -07:00
MPLS: Add limited GSO support In the case where a non-MPLS packet is received and an MPLS stack is added it may well be the case that the original skb is GSO but the NIC used for transmit does not support GSO of MPLS packets. The aim of this code is to provide GSO in software for MPLS packets whose skbs are GSO. SKB Usage: When an implementation adds an MPLS stack to a non-MPLS packet it should do the following to skb metadata: * Set skb->inner_protocol to the old non-MPLS ethertype of the packet. skb->inner_protocol is added by this patch. * Set skb->protocol to the new MPLS ethertype of the packet. * Set skb->network_header to correspond to the end of the L3 header, including the MPLS label stack. I have posted a patch, "[PATCH v3.29] datapath: Add basic MPLS support to kernel" which adds MPLS support to the kernel datapath of Open vSwtich. That patch sets the above requirements in datapath/actions.c:push_mpls() and was used to exercise this code. The datapath patch is against the Open vSwtich tree but it is intended that it be added to the Open vSwtich code present in the mainline Linux kernel at some point. Features: I believe that the approach that I have taken is at least partially consistent with the handling of other protocols. Jesse, I understand that you have some ideas here. I am more than happy to change my implementation. This patch adds dev->mpls_features which may be used by devices to advertise features supported for MPLS packets. A new NETIF_F_MPLS_GSO feature is added for devices which support hardware MPLS GSO offload. Currently no devices support this and MPLS GSO always falls back to software. Alternate Implementation: One possible alternate implementation is to teach netif_skb_features() and skb_network_protocol() about MPLS, in a similar way to their understanding of VLANs. I believe this would avoid the need for net/mpls/mpls_gso.c and in particular the calls to __skb_push() and __skb_push() in mpls_gso_segment(). I have decided on the implementation in this patch as it should not introduce any overhead in the case where mpls_gso is not compiled into the kernel or inserted as a module. MPLS GSO suggested by Jesse Gross. Based in part on "v4 GRE: Add TCP segmentation offload for GRE" by Pravin B Shelar. Cc: Jesse Gross <jesse@nicira.com> Cc: Pravin B Shelar <pshelar@nicira.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-05-23 21:02:52 +00:00
net: Set device operstate at registration time The operstate of a device is initially IF_OPER_UNKNOWN and is updated asynchronously by linkwatch after each change of carrier state reported by the driver. The default carrier state of a net device is on, and this will never be changed on drivers that do not support carrier detection, thus the operstate remains IF_OPER_UNKNOWN. For devices that do support carrier detection, the driver must set the carrier state to off initially, then poll the hardware state when the device is opened. However, we must not activate linkwatch for a unregistered device, and commit b473001 ('net: Do not fire linkwatch events until the device is registered.') ensured that we don't. But this means that the operstate for many devices that support carrier detection remains IF_OPER_UNKNOWN when it should be IF_OPER_DOWN. The same issue exists with the dormant state. The proper initialisation sequence, avoiding a race with opening of the device, is: rtnl_lock(); rc = register_netdevice(dev); if (rc) goto out_unlock; netif_carrier_off(dev); /* or netif_dormant_on(dev) */ rtnl_unlock(); but it seems silly that this should have to be repeated in so many drivers. Further, the operstate seen immediately after opening the device may still be IF_OPER_UNKNOWN due to the asynchronous nature of linkwatch. Commit 22604c8 ('net: Fix for initial link state in 2.6.28') attempted to fix this by setting the operstate synchronously, but it was reverted as it could lead to deadlock. This initialises the operstate synchronously at registration time only. Signed-off-by: Ben Hutchings <bhutchings@solarflare.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-08-20 22:16:51 +01:00
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 15:38:18 -07:00
net: fix rtnl notification in atomic context commit 991fb3f74c "dev: always advertise rx_flags changes via netlink" introduced rtnl notification from __dev_set_promiscuity(), which can be called in atomic context. Steps to reproduce: ip tuntap add dev tap1 mode tap ifconfig tap1 up tcpdump -nei tap1 & ip tuntap del dev tap1 mode tap [ 271.627994] device tap1 left promiscuous mode [ 271.639897] BUG: sleeping function called from invalid context at mm/slub.c:940 [ 271.664491] in_atomic(): 1, irqs_disabled(): 0, pid: 3394, name: ip [ 271.677525] INFO: lockdep is turned off. [ 271.690503] CPU: 0 PID: 3394 Comm: ip Tainted: G W 3.12.0-rc3+ #73 [ 271.703996] Hardware name: System manufacturer System Product Name/P8Z77 WS, BIOS 3007 07/26/2012 [ 271.731254] ffffffff81a58506 ffff8807f0d57a58 ffffffff817544e5 ffff88082fa0f428 [ 271.760261] ffff8808071f5f40 ffff8807f0d57a88 ffffffff8108bad1 ffffffff81110ff8 [ 271.790683] 0000000000000010 00000000000000d0 00000000000000d0 ffff8807f0d57af8 [ 271.822332] Call Trace: [ 271.838234] [<ffffffff817544e5>] dump_stack+0x55/0x76 [ 271.854446] [<ffffffff8108bad1>] __might_sleep+0x181/0x240 [ 271.870836] [<ffffffff81110ff8>] ? rcu_irq_exit+0x68/0xb0 [ 271.887076] [<ffffffff811a80be>] kmem_cache_alloc_node+0x4e/0x2a0 [ 271.903368] [<ffffffff810b4ddc>] ? vprintk_emit+0x1dc/0x5a0 [ 271.919716] [<ffffffff81614d67>] ? __alloc_skb+0x57/0x2a0 [ 271.936088] [<ffffffff810b4de0>] ? vprintk_emit+0x1e0/0x5a0 [ 271.952504] [<ffffffff81614d67>] __alloc_skb+0x57/0x2a0 [ 271.968902] [<ffffffff8163a0b2>] rtmsg_ifinfo+0x52/0x100 [ 271.985302] [<ffffffff8162ac6d>] __dev_notify_flags+0xad/0xc0 [ 272.001642] [<ffffffff8162ad0c>] __dev_set_promiscuity+0x8c/0x1c0 [ 272.017917] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.033961] [<ffffffff8162b109>] dev_set_promiscuity+0x29/0x50 [ 272.049855] [<ffffffff8172e937>] packet_dev_mc+0x87/0xc0 [ 272.065494] [<ffffffff81732052>] packet_notifier+0x1b2/0x380 [ 272.080915] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.096009] [<ffffffff81761c66>] notifier_call_chain+0x66/0x150 [ 272.110803] [<ffffffff8108503e>] __raw_notifier_call_chain+0xe/0x10 [ 272.125468] [<ffffffff81085056>] raw_notifier_call_chain+0x16/0x20 [ 272.139984] [<ffffffff81620190>] call_netdevice_notifiers_info+0x40/0x70 [ 272.154523] [<ffffffff816201d6>] call_netdevice_notifiers+0x16/0x20 [ 272.168552] [<ffffffff816224c5>] rollback_registered_many+0x145/0x240 [ 272.182263] [<ffffffff81622641>] rollback_registered+0x31/0x40 [ 272.195369] [<ffffffff816229c8>] unregister_netdevice_queue+0x58/0x90 [ 272.208230] [<ffffffff81547ca0>] __tun_detach+0x140/0x340 [ 272.220686] [<ffffffff81547ed6>] tun_chr_close+0x36/0x60 Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-10-23 16:02:42 -07:00
net: Fix inconsistent teardown and release of private netdev state. Network devices can allocate reasources and private memory using netdev_ops->ndo_init(). However, the release of these resources can occur in one of two different places. Either netdev_ops->ndo_uninit() or netdev->destructor(). The decision of which operation frees the resources depends upon whether it is necessary for all netdev refs to be released before it is safe to perform the freeing. netdev_ops->ndo_uninit() presumably can occur right after the NETDEV_UNREGISTER notifier completes and the unicast and multicast address lists are flushed. netdev->destructor(), on the other hand, does not run until the netdev references all go away. Further complicating the situation is that netdev->destructor() almost universally does also a free_netdev(). This creates a problem for the logic in register_netdevice(). Because all callers of register_netdevice() manage the freeing of the netdev, and invoke free_netdev(dev) if register_netdevice() fails. If netdev_ops->ndo_init() succeeds, but something else fails inside of register_netdevice(), it does call ndo_ops->ndo_uninit(). But it is not able to invoke netdev->destructor(). This is because netdev->destructor() will do a free_netdev() and then the caller of register_netdevice() will do the same. However, this means that the resources that would normally be released by netdev->destructor() will not be. Over the years drivers have added local hacks to deal with this, by invoking their destructor parts by hand when register_netdevice() fails. Many drivers do not try to deal with this, and instead we have leaks. Let's close this hole by formalizing the distinction between what private things need to be freed up by netdev->destructor() and whether the driver needs unregister_netdevice() to perform the free_netdev(). netdev->priv_destructor() performs all actions to free up the private resources that used to be freed by netdev->destructor(), except for free_netdev(). netdev->needs_free_netdev is a boolean that indicates whether free_netdev() should be done at the end of unregister_netdevice(). Now, register_netdevice() can sanely release all resources after ndo_ops->ndo_init() succeeds, by invoking both ndo_ops->ndo_uninit() and netdev->priv_destructor(). And at the end of unregister_netdevice(), we invoke netdev->priv_destructor() and optionally call free_netdev(). Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-08 12:52:56 -04:00
linkwatch: linkwatch_forget_dev() to speedup device dismantle Herbert Xu a écrit : > On Tue, Nov 17, 2009 at 04:26:04AM -0800, David Miller wrote: >> Really, the link watch stuff is just due for a redesign. I don't >> think a simple hack is going to cut it this time, sorry Eric :-) > > I have no objections against any redesigns, but since the only > caller of linkwatch_forget_dev runs in process context with the > RTNL, it could also legally emit those events. Thanks guys, here an updated version then, before linkwatch surgery ? In this version, I force the event to be sent synchronously. [PATCH net-next-2.6] linkwatch: linkwatch_forget_dev() to speedup device dismantle time ip link del eth3.103 ; time ip link del eth3.104 ; time ip link del eth3.105 real 0m0.266s user 0m0.000s sys 0m0.001s real 0m0.770s user 0m0.000s sys 0m0.000s real 0m1.022s user 0m0.000s sys 0m0.000s One problem of current schem in vlan dismantle phase is the holding of device done by following chain : vlan_dev_stop() -> netif_carrier_off(dev) -> linkwatch_fire_event(dev) -> dev_hold() ... And __linkwatch_run_queue() runs up to one second later... A generic fix to this problem is to add a linkwatch_forget_dev() method to unlink the device from the list of watched devices. dev->link_watch_next becomes dev->link_watch_list (and use a bit more memory), to be able to unlink device in O(1). After patch : time ip link del eth3.103 ; time ip link del eth3.104 ; time ip link del eth3.105 real 0m0.024s user 0m0.000s sys 0m0.000s real 0m0.032s user 0m0.000s sys 0m0.001s real 0m0.033s user 0m0.000s sys 0m0.000s Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-11-17 05:59:21 +00:00
net: Fix inconsistent teardown and release of private netdev state. Network devices can allocate reasources and private memory using netdev_ops->ndo_init(). However, the release of these resources can occur in one of two different places. Either netdev_ops->ndo_uninit() or netdev->destructor(). The decision of which operation frees the resources depends upon whether it is necessary for all netdev refs to be released before it is safe to perform the freeing. netdev_ops->ndo_uninit() presumably can occur right after the NETDEV_UNREGISTER notifier completes and the unicast and multicast address lists are flushed. netdev->destructor(), on the other hand, does not run until the netdev references all go away. Further complicating the situation is that netdev->destructor() almost universally does also a free_netdev(). This creates a problem for the logic in register_netdevice(). Because all callers of register_netdevice() manage the freeing of the netdev, and invoke free_netdev(dev) if register_netdevice() fails. If netdev_ops->ndo_init() succeeds, but something else fails inside of register_netdevice(), it does call ndo_ops->ndo_uninit(). But it is not able to invoke netdev->destructor(). This is because netdev->destructor() will do a free_netdev() and then the caller of register_netdevice() will do the same. However, this means that the resources that would normally be released by netdev->destructor() will not be. Over the years drivers have added local hacks to deal with this, by invoking their destructor parts by hand when register_netdevice() fails. Many drivers do not try to deal with this, and instead we have leaks. Let's close this hole by formalizing the distinction between what private things need to be freed up by netdev->destructor() and whether the driver needs unregister_netdevice() to perform the free_netdev(). netdev->priv_destructor() performs all actions to free up the private resources that used to be freed by netdev->destructor(), except for free_netdev(). netdev->needs_free_netdev is a boolean that indicates whether free_netdev() should be done at the end of unregister_netdevice(). Now, register_netdevice() can sanely release all resources after ndo_ops->ndo_init() succeeds, by invoking both ndo_ops->ndo_uninit() and netdev->priv_destructor(). And at the end of unregister_netdevice(), we invoke netdev->priv_destructor() and optionally call free_netdev(). Signed-off-by: David S. Miller <davem@davemloft.net>
2017-05-08 12:52:56 -04:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
net: core: Fix slab-out-of-bounds in netdev_stats_to_stats64 commit 9256645af098 ("net/core: relax BUILD_BUG_ON in netdev_stats_to_stats64") made an attempt to read beyond the size of the source a possibility. Fix to only copy src size to dest. As dest might be bigger than src. ================================================================== BUG: KASAN: slab-out-of-bounds in netdev_stats_to_stats64+0xe/0x30 at addr ffff8801be248b20 Read of size 192 by task VBoxNetAdpCtl/6734 CPU: 1 PID: 6734 Comm: VBoxNetAdpCtl Tainted: G O 4.11.4prahal+intel+ #118 Hardware name: LENOVO 20CDCTO1WW/20CDCTO1WW, BIOS GQET52WW (1.32 ) 05/04/2017 Call Trace: dump_stack+0x63/0x86 kasan_object_err+0x1c/0x70 kasan_report+0x270/0x520 ? netdev_stats_to_stats64+0xe/0x30 ? sched_clock_cpu+0x1b/0x190 ? __module_address+0x3e/0x3b0 ? unwind_next_frame+0x1ea/0xb00 check_memory_region+0x13c/0x1a0 memcpy+0x23/0x50 netdev_stats_to_stats64+0xe/0x30 dev_get_stats+0x1b9/0x230 rtnl_fill_stats+0x44/0xc00 ? nla_put+0xc6/0x130 rtnl_fill_ifinfo+0xe9e/0x3700 ? rtnl_fill_vfinfo+0xde0/0xde0 ? sched_clock+0x9/0x10 ? sched_clock+0x9/0x10 ? sched_clock_local+0x120/0x130 ? __module_address+0x3e/0x3b0 ? unwind_next_frame+0x1ea/0xb00 ? sched_clock+0x9/0x10 ? sched_clock+0x9/0x10 ? sched_clock_cpu+0x1b/0x190 ? VBoxNetAdpLinuxIOCtlUnlocked+0x14b/0x280 [vboxnetadp] ? depot_save_stack+0x1d8/0x4a0 ? depot_save_stack+0x34f/0x4a0 ? depot_save_stack+0x34f/0x4a0 ? save_stack+0xb1/0xd0 ? save_stack_trace+0x16/0x20 ? save_stack+0x46/0xd0 ? kasan_slab_alloc+0x12/0x20 ? __kmalloc_node_track_caller+0x10d/0x350 ? __kmalloc_reserve.isra.36+0x2c/0xc0 ? __alloc_skb+0xd0/0x560 ? rtmsg_ifinfo_build_skb+0x61/0x120 ? rtmsg_ifinfo.part.25+0x16/0xb0 ? rtmsg_ifinfo+0x47/0x70 ? register_netdev+0x15/0x30 ? vboxNetAdpOsCreate+0xc0/0x1c0 [vboxnetadp] ? vboxNetAdpCreate+0x210/0x400 [vboxnetadp] ? VBoxNetAdpLinuxIOCtlUnlocked+0x14b/0x280 [vboxnetadp] ? do_vfs_ioctl+0x17f/0xff0 ? SyS_ioctl+0x74/0x80 ? do_syscall_64+0x182/0x390 ? __alloc_skb+0xd0/0x560 ? __alloc_skb+0xd0/0x560 ? save_stack_trace+0x16/0x20 ? init_object+0x64/0xa0 ? ___slab_alloc+0x1ae/0x5c0 ? ___slab_alloc+0x1ae/0x5c0 ? __alloc_skb+0xd0/0x560 ? sched_clock+0x9/0x10 ? kasan_unpoison_shadow+0x35/0x50 ? kasan_kmalloc+0xad/0xe0 ? __kmalloc_node_track_caller+0x246/0x350 ? __alloc_skb+0xd0/0x560 ? kasan_unpoison_shadow+0x35/0x50 ? memset+0x31/0x40 ? __alloc_skb+0x31f/0x560 ? napi_consume_skb+0x320/0x320 ? br_get_link_af_size_filtered+0xb7/0x120 [bridge] ? if_nlmsg_size+0x440/0x630 rtmsg_ifinfo_build_skb+0x83/0x120 rtmsg_ifinfo.part.25+0x16/0xb0 rtmsg_ifinfo+0x47/0x70 register_netdevice+0xa2b/0xe50 ? __kmalloc+0x171/0x2d0 ? netdev_change_features+0x80/0x80 register_netdev+0x15/0x30 vboxNetAdpOsCreate+0xc0/0x1c0 [vboxnetadp] vboxNetAdpCreate+0x210/0x400 [vboxnetadp] ? vboxNetAdpComposeMACAddress+0x1d0/0x1d0 [vboxnetadp] ? kasan_check_write+0x14/0x20 VBoxNetAdpLinuxIOCtlUnlocked+0x14b/0x280 [vboxnetadp] ? VBoxNetAdpLinuxOpen+0x20/0x20 [vboxnetadp] ? lock_acquire+0x11c/0x270 ? __audit_syscall_entry+0x2fb/0x660 do_vfs_ioctl+0x17f/0xff0 ? __audit_syscall_entry+0x2fb/0x660 ? ioctl_preallocate+0x1d0/0x1d0 ? __audit_syscall_entry+0x2fb/0x660 ? kmem_cache_free+0xb2/0x250 ? syscall_trace_enter+0x537/0xd00 ? exit_to_usermode_loop+0x100/0x100 SyS_ioctl+0x74/0x80 ? do_sys_open+0x350/0x350 ? do_vfs_ioctl+0xff0/0xff0 do_syscall_64+0x182/0x390 entry_SYSCALL64_slow_path+0x25/0x25 RIP: 0033:0x7f7e39a1ae07 RSP: 002b:00007ffc6f04c6d8 EFLAGS: 00000206 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 00007ffc6f04c730 RCX: 00007f7e39a1ae07 RDX: 00007ffc6f04c730 RSI: 00000000c0207601 RDI: 0000000000000007 RBP: 00007ffc6f04c700 R08: 00007ffc6f04c780 R09: 0000000000000008 R10: 0000000000000541 R11: 0000000000000206 R12: 0000000000000007 R13: 00000000c0207601 R14: 00007ffc6f04c730 R15: 0000000000000012 Object at ffff8801be248008, in cache kmalloc-4096 size: 4096 Allocated: PID = 6734 save_stack_trace+0x16/0x20 save_stack+0x46/0xd0 kasan_kmalloc+0xad/0xe0 __kmalloc+0x171/0x2d0 alloc_netdev_mqs+0x8a7/0xbe0 vboxNetAdpOsCreate+0x65/0x1c0 [vboxnetadp] vboxNetAdpCreate+0x210/0x400 [vboxnetadp] VBoxNetAdpLinuxIOCtlUnlocked+0x14b/0x280 [vboxnetadp] do_vfs_ioctl+0x17f/0xff0 SyS_ioctl+0x74/0x80 do_syscall_64+0x182/0x390 return_from_SYSCALL_64+0x0/0x6a Freed: PID = 5600 save_stack_trace+0x16/0x20 save_stack+0x46/0xd0 kasan_slab_free+0x73/0xc0 kfree+0xe4/0x220 kvfree+0x25/0x30 single_release+0x74/0xb0 __fput+0x265/0x6b0 ____fput+0x9/0x10 task_work_run+0xd5/0x150 exit_to_usermode_loop+0xe2/0x100 do_syscall_64+0x26c/0x390 return_from_SYSCALL_64+0x0/0x6a Memory state around the buggy address: ffff8801be248a80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ffff8801be248b00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 >ffff8801be248b80: 00 00 00 00 00 00 00 00 00 00 00 07 fc fc fc fc ^ ffff8801be248c00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ffff8801be248c80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ================================================================== Signed-off-by: Alban Browaeys <alban.browaeys@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-07-03 03:20:13 +02:00
mm, tree wide: replace __GFP_REPEAT by __GFP_RETRY_MAYFAIL with more useful semantic __GFP_REPEAT was designed to allow retry-but-eventually-fail semantic to the page allocator. This has been true but only for allocations requests larger than PAGE_ALLOC_COSTLY_ORDER. It has been always ignored for smaller sizes. This is a bit unfortunate because there is no way to express the same semantic for those requests and they are considered too important to fail so they might end up looping in the page allocator for ever, similarly to GFP_NOFAIL requests. Now that the whole tree has been cleaned up and accidental or misled usage of __GFP_REPEAT flag has been removed for !costly requests we can give the original flag a better name and more importantly a more useful semantic. Let's rename it to __GFP_RETRY_MAYFAIL which tells the user that the allocator would try really hard but there is no promise of a success. This will work independent of the order and overrides the default allocator behavior. Page allocator users have several levels of guarantee vs. cost options (take GFP_KERNEL as an example) - GFP_KERNEL & ~__GFP_RECLAIM - optimistic allocation without _any_ attempt to free memory at all. The most light weight mode which even doesn't kick the background reclaim. Should be used carefully because it might deplete the memory and the next user might hit the more aggressive reclaim - GFP_KERNEL & ~__GFP_DIRECT_RECLAIM (or GFP_NOWAIT)- optimistic allocation without any attempt to free memory from the current context but can wake kswapd to reclaim memory if the zone is below the low watermark. Can be used from either atomic contexts or when the request is a performance optimization and there is another fallback for a slow path. - (GFP_KERNEL|__GFP_HIGH) & ~__GFP_DIRECT_RECLAIM (aka GFP_ATOMIC) - non sleeping allocation with an expensive fallback so it can access some portion of memory reserves. Usually used from interrupt/bh context with an expensive slow path fallback. - GFP_KERNEL - both background and direct reclaim are allowed and the _default_ page allocator behavior is used. That means that !costly allocation requests are basically nofail but there is no guarantee of that behavior so failures have to be checked properly by callers (e.g. OOM killer victim is allowed to fail currently). - GFP_KERNEL | __GFP_NORETRY - overrides the default allocator behavior and all allocation requests fail early rather than cause disruptive reclaim (one round of reclaim in this implementation). The OOM killer is not invoked. - GFP_KERNEL | __GFP_RETRY_MAYFAIL - overrides the default allocator behavior and all allocation requests try really hard. The request will fail if the reclaim cannot make any progress. The OOM killer won't be triggered. - GFP_KERNEL | __GFP_NOFAIL - overrides the default allocator behavior and all allocation requests will loop endlessly until they succeed. This might be really dangerous especially for larger orders. Existing users of __GFP_REPEAT are changed to __GFP_RETRY_MAYFAIL because they already had their semantic. No new users are added. __alloc_pages_slowpath is changed to bail out for __GFP_RETRY_MAYFAIL if there is no progress and we have already passed the OOM point. This means that all the reclaim opportunities have been exhausted except the most disruptive one (the OOM killer) and a user defined fallback behavior is more sensible than keep retrying in the page allocator. [akpm@linux-foundation.org: fix arch/sparc/kernel/mdesc.c] [mhocko@suse.com: semantic fix] Link: http://lkml.kernel.org/r/20170626123847.GM11534@dhcp22.suse.cz [mhocko@kernel.org: address other thing spotted by Vlastimil] Link: http://lkml.kernel.org/r/20170626124233.GN11534@dhcp22.suse.cz Link: http://lkml.kernel.org/r/20170623085345.11304-3-mhocko@kernel.org Signed-off-by: Michal Hocko <mhocko@suse.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Alex Belits <alex.belits@cavium.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Christoph Hellwig <hch@infradead.org> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: David Daney <david.daney@cavium.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: NeilBrown <neilb@suse.com> Cc: Ralf Baechle <ralf@linux-mips.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-12 14:36:45 -07:00
[NET]: Add per-connection option to set max TSO frame size Update: My mailer ate one of Jarek's feedback mails... Fixed the parameter in netif_set_gso_max_size() to be u32, not u16. Fixed the whitespace issue due to a patch import botch. Changed the types from u32 to unsigned int to be more consistent with other variables in the area. Also brought the patch up to the latest net-2.6.26 tree. Update: Made gso_max_size container 32 bits, not 16. Moved the location of gso_max_size within netdev to be less hotpath. Made more consistent names between the sock and netdev layers, and added a define for the max GSO size. Update: Respun for net-2.6.26 tree. Update: changed max_gso_frame_size and sk_gso_max_size from signed to unsigned - thanks Stephen! This patch adds the ability for device drivers to control the size of the TSO frames being sent to them, per TCP connection. By setting the netdevice's gso_max_size value, the socket layer will set the GSO frame size based on that value. This will propogate into the TCP layer, and send TSO's of that size to the hardware. This can be desirable to help tune the bursty nature of TSO on a per-adapter basis, where one may have 1 GbE and 10 GbE devices coexisting in a system, one running multiqueue and the other not, etc. This can also be desirable for devices that cannot support full 64 KB TSO's, but still want to benefit from some level of segmentation offloading. Signed-off-by: Peter P Waskiewicz Jr <peter.p.waskiewicz.jr@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-03-21 03:43:19 -07:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
linkwatch: linkwatch_forget_dev() to speedup device dismantle Herbert Xu a écrit : > On Tue, Nov 17, 2009 at 04:26:04AM -0800, David Miller wrote: >> Really, the link watch stuff is just due for a redesign. I don't >> think a simple hack is going to cut it this time, sorry Eric :-) > > I have no objections against any redesigns, but since the only > caller of linkwatch_forget_dev runs in process context with the > RTNL, it could also legally emit those events. Thanks guys, here an updated version then, before linkwatch surgery ? In this version, I force the event to be sent synchronously. [PATCH net-next-2.6] linkwatch: linkwatch_forget_dev() to speedup device dismantle time ip link del eth3.103 ; time ip link del eth3.104 ; time ip link del eth3.105 real 0m0.266s user 0m0.000s sys 0m0.001s real 0m0.770s user 0m0.000s sys 0m0.000s real 0m1.022s user 0m0.000s sys 0m0.000s One problem of current schem in vlan dismantle phase is the holding of device done by following chain : vlan_dev_stop() -> netif_carrier_off(dev) -> linkwatch_fire_event(dev) -> dev_hold() ... And __linkwatch_run_queue() runs up to one second later... A generic fix to this problem is to add a linkwatch_forget_dev() method to unlink the device from the list of watched devices. dev->link_watch_next becomes dev->link_watch_list (and use a bit more memory), to be able to unlink device in O(1). After patch : time ip link del eth3.103 ; time ip link del eth3.104 ; time ip link del eth3.105 real 0m0.024s user 0m0.000s sys 0m0.000s real 0m0.032s user 0m0.000s sys 0m0.001s real 0m0.033s user 0m0.000s sys 0m0.000s Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-11-17 05:59:21 +00:00
net: add adj_list to save only neighbours Currently, we distinguish neighbours (first-level linked devices) from non-neighbours by the neighbour bool in the netdev_adjacent. This could be quite time-consuming in case we would like to traverse *only* through neighbours - cause we'd have to traverse through all devices and check for this flag, and in a (quite common) scenario where we have lots of vlans on top of bridge, which is on top of a bond - the bonding would have to go through all those vlans to get its upper neighbour linked devices. This situation is really unpleasant, cause there are already a lot of cases when a device with slaves needs to go through them in hot path. To fix this, introduce a new upper/lower device lists structure - adj_list, which contains only the neighbours. It works always in pair with the all_adj_list structure (renamed from upper/lower_dev_list), i.e. both of them contain the same links, only that all_adj_list contains also non-neighbour device links. It's really a small change visible, currently, only for __netdev_adjacent_dev_insert/remove(), and doesn't change the main linked logic at all. Also, add some comments a fix a name collision in netdev_for_each_upper_dev_rcu() and rework the naming by the following rules: netdev_(all_)(upper|lower)_* If "all_" is present, then we work with the whole list of upper/lower devices, otherwise - only with direct neighbours. Uninline functions - to get better stack traces. CC: "David S. Miller" <davem@davemloft.net> CC: Eric Dumazet <edumazet@google.com> CC: Jiri Pirko <jiri@resnulli.us> CC: Alexander Duyck <alexander.h.duyck@intel.com> CC: Cong Wang <amwang@redhat.com> Signed-off-by: Veaceslav Falico <vfalico@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-25 09:20:07 +02:00
net/qdisc: IFF_NO_QUEUE drivers should use consistent TX queue len The flag IFF_NO_QUEUE marks virtual device drivers that doesn't need a default qdisc attached, given they will be backed by physical device, that already have a qdisc attached for pushback. It is still supported to attach a qdisc to a IFF_NO_QUEUE device, as this can be useful for difference policy reasons (e.g. bandwidth limiting containers). For this to work, the tx_queue_len need to have a sane value, because some qdiscs inherit/copy the tx_queue_len (namely, pfifo, bfifo, gred, htb, plug and sfb). Commit a813104d9233 ("IFF_NO_QUEUE: Fix for drivers not calling ether_setup()") caught situations where some drivers didn't initialize tx_queue_len. The problem with the commit was choosing 1 as the fallback value. A qdisc queue length of 1 causes more harm than good, because it creates hard to debug situations for userspace. It gives userspace a false sense of a working config after attaching a qdisc. As low volume traffic (that doesn't activate the qdisc policy) works, like ping, while traffic that e.g. needs shaping cannot reach the configured policy levels, given the queue length is too small. This patch change the value to DEFAULT_TX_QUEUE_LEN, given other IFF_NO_QUEUE devices (that call ether_setup()) also use this value. Fixes: a813104d9233 ("IFF_NO_QUEUE: Fix for drivers not calling ether_setup()") Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-11-03 14:56:06 +01:00
netfilter: add netfilter ingress hook after handle_ing() under unique static key This patch adds the Netfilter ingress hook just after the existing tc ingress hook, that seems to be the consensus solution for this. Note that the Netfilter hook resides under the global static key that enables ingress filtering. Nonetheless, Netfilter still also has its own static key for minimal impact on the existing handle_ing(). * Without this patch: Result: OK: 6216490(c6216338+d152) usec, 100000000 (60byte,0frags) 16086246pps 7721Mb/sec (7721398080bps) errors: 100000000 42.46% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 25.92% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.81% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.62% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.70% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.34% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.44% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch: Result: OK: 6214833(c6214731+d101) usec, 100000000 (60byte,0frags) 16090536pps 7723Mb/sec (7723457280bps) errors: 100000000 41.23% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 26.57% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 7.72% kpktgend_0 [pktgen] [k] pktgen_thread_worker 5.55% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.78% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 2.06% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 1.43% kpktgend_0 [kernel.kallsyms] [k] __build_skb * Without this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9269001(c9268821+d179) usec, 100000000 (60byte,0frags) 10788648pps 5178Mb/sec (5178551040bps) errors: 100000000 40.99% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.50% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.77% kpktgend_0 [cls_u32] [k] u32_classify 5.62% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.18% kpktgend_0 [pktgen] [k] pktgen_thread_worker 3.23% kpktgend_0 [kernel.kallsyms] [k] tc_classify 2.97% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 1.83% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.50% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk 0.99% kpktgend_0 [kernel.kallsyms] [k] __build_skb * With this patch + tc ingress: tc filter add dev eth4 parent ffff: protocol ip prio 1 \ u32 match ip dst 4.3.2.1/32 Result: OK: 9308218(c9308091+d126) usec, 100000000 (60byte,0frags) 10743194pps 5156Mb/sec (5156733120bps) errors: 100000000 42.01% kpktgend_0 [kernel.kallsyms] [k] __netif_receive_skb_core 17.78% kpktgend_0 [kernel.kallsyms] [k] kfree_skb 11.70% kpktgend_0 [cls_u32] [k] u32_classify 5.46% kpktgend_0 [kernel.kallsyms] [k] tc_classify_compat 5.16% kpktgend_0 [pktgen] [k] pktgen_thread_worker 2.98% kpktgend_0 [kernel.kallsyms] [k] ip_rcv 2.84% kpktgend_0 [kernel.kallsyms] [k] tc_classify 1.96% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_internal 1.57% kpktgend_0 [kernel.kallsyms] [k] netif_receive_skb_sk Note that the results are very similar before and after. I can see gcc gets the code under the ingress static key out of the hot path. Then, on that cold branch, it generates the code to accomodate the netfilter ingress static key. My explanation for this is that this reduces the pressure on the instruction cache for non-users as the new code is out of the hot path, and it comes with minimal impact for tc ingress users. Using gcc version 4.8.4 on: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 [...] L1d cache: 16K L1i cache: 64K L2 cache: 2048K L3 cache: 8192K Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Acked-by: Alexei Starovoitov <ast@plumgrid.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-05-13 18:19:38 +02:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-15 23:38:52 -08:00
net: use synchronize_rcu_expedited() synchronize_rcu() is very slow in various situations (HZ=100, CONFIG_NO_HZ=y, CONFIG_PREEMPT=n) Extract from my (mostly idle) 8 core machine : synchronize_rcu() in 99985 us synchronize_rcu() in 79982 us synchronize_rcu() in 87612 us synchronize_rcu() in 79827 us synchronize_rcu() in 109860 us synchronize_rcu() in 98039 us synchronize_rcu() in 89841 us synchronize_rcu() in 79842 us synchronize_rcu() in 80151 us synchronize_rcu() in 119833 us synchronize_rcu() in 99858 us synchronize_rcu() in 73999 us synchronize_rcu() in 79855 us synchronize_rcu() in 79853 us When we hold RTNL mutex, we would like to spend some cpu cycles but not block too long other processes waiting for this mutex. We also want to setup/dismantle network features as fast as possible at boot/shutdown time. This patch makes synchronize_net() call the expedited version if RTNL is locked. synchronize_rcu_expedited() typical delay is about 20 us on my machine. synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 20 us synchronize_rcu_expedited() in 16 us synchronize_rcu_expedited() in 20 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> CC: Paul E. McKenney <paulmck@linux.vnet.ibm.com> CC: Ben Greear <greearb@candelatech.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-05-23 23:07:32 +00:00
net: dev_change_net_namespace: send a KOBJ_REMOVED/KOBJ_ADD When a new nic is created in namespace ns1, the kernel sends a KOBJ_ADD uevent to ns1. When the nic is moved to ns2, we only send a KOBJ_MOVE to ns2, and nothing to ns1. This patch changes that behavior so that when moving a nic from ns1 to ns2, we send a KOBJ_REMOVED to ns1 and KOBJ_ADD to ns2. (The KOBJ_MOVE is still sent to ns2). The effects of this can be seen when starting and stopping containers in an upstart based host. Lxc will create a pair of veth nics, the kernel sends KOBJ_ADD, and upstart starts network-instance jobs for each. When one nic is moved to the container, because no KOBJ_REMOVED event is received, the network-instance job for that veth never goes away. This was reported at https://bugs.launchpad.net/ubuntu/+source/lxc/+bug/1065589 With this patch the networ-instance jobs properly go away. The other oddness solved here is that if a nic is passed into a running upstart-based container, without this patch no network-instance job is started in the container. But when the container creates a new nic itself (ip link add new type veth) then network-interface jobs are created. With this patch, behavior comes in line with a regular host. v2: also send KOBJ_ADD to new netns. There will then be a _MOVE event from the device_rename() call, but that should be innocuous. Signed-off-by: Serge Hallyn <serge.hallyn@canonical.com> Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Acked-by: Daniel Lezcano <daniel.lezcano@free.fr> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-12-03 16:17:12 +00:00
net: dev_change_net_namespace: send a KOBJ_REMOVED/KOBJ_ADD When a new nic is created in namespace ns1, the kernel sends a KOBJ_ADD uevent to ns1. When the nic is moved to ns2, we only send a KOBJ_MOVE to ns2, and nothing to ns1. This patch changes that behavior so that when moving a nic from ns1 to ns2, we send a KOBJ_REMOVED to ns1 and KOBJ_ADD to ns2. (The KOBJ_MOVE is still sent to ns2). The effects of this can be seen when starting and stopping containers in an upstart based host. Lxc will create a pair of veth nics, the kernel sends KOBJ_ADD, and upstart starts network-instance jobs for each. When one nic is moved to the container, because no KOBJ_REMOVED event is received, the network-instance job for that veth never goes away. This was reported at https://bugs.launchpad.net/ubuntu/+source/lxc/+bug/1065589 With this patch the networ-instance jobs properly go away. The other oddness solved here is that if a nic is passed into a running upstart-based container, without this patch no network-instance job is started in the container. But when the container creates a new nic itself (ip link add new type veth) then network-interface jobs are created. With this patch, behavior comes in line with a regular host. v2: also send KOBJ_ADD to new netns. There will then be a _MOVE event from the device_rename() call, but that should be innocuous. Signed-off-by: Serge Hallyn <serge.hallyn@canonical.com> Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Acked-by: Daniel Lezcano <daniel.lezcano@free.fr> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-12-03 16:17:12 +00:00
net: dev_change_net_namespace: send a KOBJ_REMOVED/KOBJ_ADD When a new nic is created in namespace ns1, the kernel sends a KOBJ_ADD uevent to ns1. When the nic is moved to ns2, we only send a KOBJ_MOVE to ns2, and nothing to ns1. This patch changes that behavior so that when moving a nic from ns1 to ns2, we send a KOBJ_REMOVED to ns1 and KOBJ_ADD to ns2. (The KOBJ_MOVE is still sent to ns2). The effects of this can be seen when starting and stopping containers in an upstart based host. Lxc will create a pair of veth nics, the kernel sends KOBJ_ADD, and upstart starts network-instance jobs for each. When one nic is moved to the container, because no KOBJ_REMOVED event is received, the network-instance job for that veth never goes away. This was reported at https://bugs.launchpad.net/ubuntu/+source/lxc/+bug/1065589 With this patch the networ-instance jobs properly go away. The other oddness solved here is that if a nic is passed into a running upstart-based container, without this patch no network-instance job is started in the container. But when the container creates a new nic itself (ip link add new type veth) then network-interface jobs are created. With this patch, behavior comes in line with a regular host. v2: also send KOBJ_ADD to new netns. There will then be a _MOVE event from the device_rename() call, but that should be innocuous. Signed-off-by: Serge Hallyn <serge.hallyn@canonical.com> Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Acked-by: Daniel Lezcano <daniel.lezcano@free.fr> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-12-03 16:17:12 +00:00
net: dev_change_net_namespace: send a KOBJ_REMOVED/KOBJ_ADD When a new nic is created in namespace ns1, the kernel sends a KOBJ_ADD uevent to ns1. When the nic is moved to ns2, we only send a KOBJ_MOVE to ns2, and nothing to ns1. This patch changes that behavior so that when moving a nic from ns1 to ns2, we send a KOBJ_REMOVED to ns1 and KOBJ_ADD to ns2. (The KOBJ_MOVE is still sent to ns2). The effects of this can be seen when starting and stopping containers in an upstart based host. Lxc will create a pair of veth nics, the kernel sends KOBJ_ADD, and upstart starts network-instance jobs for each. When one nic is moved to the container, because no KOBJ_REMOVED event is received, the network-instance job for that veth never goes away. This was reported at https://bugs.launchpad.net/ubuntu/+source/lxc/+bug/1065589 With this patch the networ-instance jobs properly go away. The other oddness solved here is that if a nic is passed into a running upstart-based container, without this patch no network-instance job is started in the container. But when the container creates a new nic itself (ip link add new type veth) then network-interface jobs are created. With this patch, behavior comes in line with a regular host. v2: also send KOBJ_ADD to new netns. There will then be a _MOVE event from the device_rename() call, but that should be innocuous. Signed-off-by: Serge Hallyn <serge.hallyn@canonical.com> Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Acked-by: Daniel Lezcano <daniel.lezcano@free.fr> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-12-03 16:17:12 +00:00
net: fix rtnl notification in atomic context commit 991fb3f74c "dev: always advertise rx_flags changes via netlink" introduced rtnl notification from __dev_set_promiscuity(), which can be called in atomic context. Steps to reproduce: ip tuntap add dev tap1 mode tap ifconfig tap1 up tcpdump -nei tap1 & ip tuntap del dev tap1 mode tap [ 271.627994] device tap1 left promiscuous mode [ 271.639897] BUG: sleeping function called from invalid context at mm/slub.c:940 [ 271.664491] in_atomic(): 1, irqs_disabled(): 0, pid: 3394, name: ip [ 271.677525] INFO: lockdep is turned off. [ 271.690503] CPU: 0 PID: 3394 Comm: ip Tainted: G W 3.12.0-rc3+ #73 [ 271.703996] Hardware name: System manufacturer System Product Name/P8Z77 WS, BIOS 3007 07/26/2012 [ 271.731254] ffffffff81a58506 ffff8807f0d57a58 ffffffff817544e5 ffff88082fa0f428 [ 271.760261] ffff8808071f5f40 ffff8807f0d57a88 ffffffff8108bad1 ffffffff81110ff8 [ 271.790683] 0000000000000010 00000000000000d0 00000000000000d0 ffff8807f0d57af8 [ 271.822332] Call Trace: [ 271.838234] [<ffffffff817544e5>] dump_stack+0x55/0x76 [ 271.854446] [<ffffffff8108bad1>] __might_sleep+0x181/0x240 [ 271.870836] [<ffffffff81110ff8>] ? rcu_irq_exit+0x68/0xb0 [ 271.887076] [<ffffffff811a80be>] kmem_cache_alloc_node+0x4e/0x2a0 [ 271.903368] [<ffffffff810b4ddc>] ? vprintk_emit+0x1dc/0x5a0 [ 271.919716] [<ffffffff81614d67>] ? __alloc_skb+0x57/0x2a0 [ 271.936088] [<ffffffff810b4de0>] ? vprintk_emit+0x1e0/0x5a0 [ 271.952504] [<ffffffff81614d67>] __alloc_skb+0x57/0x2a0 [ 271.968902] [<ffffffff8163a0b2>] rtmsg_ifinfo+0x52/0x100 [ 271.985302] [<ffffffff8162ac6d>] __dev_notify_flags+0xad/0xc0 [ 272.001642] [<ffffffff8162ad0c>] __dev_set_promiscuity+0x8c/0x1c0 [ 272.017917] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.033961] [<ffffffff8162b109>] dev_set_promiscuity+0x29/0x50 [ 272.049855] [<ffffffff8172e937>] packet_dev_mc+0x87/0xc0 [ 272.065494] [<ffffffff81732052>] packet_notifier+0x1b2/0x380 [ 272.080915] [<ffffffff81731ea5>] ? packet_notifier+0x5/0x380 [ 272.096009] [<ffffffff81761c66>] notifier_call_chain+0x66/0x150 [ 272.110803] [<ffffffff8108503e>] __raw_notifier_call_chain+0xe/0x10 [ 272.125468] [<ffffffff81085056>] raw_notifier_call_chain+0x16/0x20 [ 272.139984] [<ffffffff81620190>] call_netdevice_notifiers_info+0x40/0x70 [ 272.154523] [<ffffffff816201d6>] call_netdevice_notifiers+0x16/0x20 [ 272.168552] [<ffffffff816224c5>] rollback_registered_many+0x145/0x240 [ 272.182263] [<ffffffff81622641>] rollback_registered+0x31/0x40 [ 272.195369] [<ffffffff816229c8>] unregister_netdevice_queue+0x58/0x90 [ 272.208230] [<ffffffff81547ca0>] __tun_detach+0x140/0x340 [ 272.220686] [<ffffffff81547ed6>] tun_chr_close+0x36/0x60 Signed-off-by: Alexei Starovoitov <ast@plumgrid.com> Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-10-23 16:02:42 -07:00
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 16:01:27 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
net: Delay default_device_exit_batch until no devices are unregistering v2 There is currently serialization network namespaces exiting and network devices exiting as the final part of netdev_run_todo does not happen under the rtnl_lock. This is compounded by the fact that the only list of devices unregistering in netdev_run_todo is local to the netdev_run_todo. This lack of serialization in extreme cases results in network devices unregistering in netdev_run_todo after the loopback device of their network namespace has been freed (making dst_ifdown unsafe), and after the their network namespace has exited (making the NETDEV_UNREGISTER, and NETDEV_UNREGISTER_FINAL callbacks unsafe). Add the missing serialization by a per network namespace count of how many network devices are unregistering and having a wait queue that is woken up whenever the count is decreased. The count and wait queue allow default_device_exit_batch to wait until all of the unregistration activity for a network namespace has finished before proceeding to unregister the loopback device and then allowing the network namespace to exit. Only a single global wait queue is used because there is a single global lock, and there is a single waiter, per network namespace wait queues would be a waste of resources. The per network namespace count of unregistering devices gives a progress guarantee because the number of network devices unregistering in an exiting network namespace must ultimately drop to zero (assuming network device unregistration completes). The basic logic remains the same as in v1. This patch is now half comment and half rtnl_lock_unregistering an expanded version of wait_event performs no extra work in the common case where no network devices are unregistering when we get to default_device_exit_batch. Reported-by: Francesco Ruggeri <fruggeri@aristanetworks.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-09-23 21:19:49 -07:00
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 11:56:21 -07:00
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00