Newer devices have more PAPD channel groups.
Signed-off-by: Eran Harary <eran.harary@intel.com>
Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
When a TDLS station is being drained its Tx queues are still in use. Don't
allocate them to a different station in the meantime.
Signed-off-by: Arik Nemtsov <arikx.nemtsov@intel.com>
Reviewed-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
Due to issues with Miracast adapters MIMO reception disable
use of MIMO when for low latency P2P traffic.
Signed-off-by: Eyal Shapira <eyalx.shapira@intel.com>
Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
Once the FW supports autonomous decision between STBC/BFER/SISO
we no longer set the STBC bit and ANT_AB in the rate table.
However the FW rate in the tx response will have the STBC
or BFER bit set and the antennas set to ANT_AB in case these
were chosen by it. This will cause us to discard any such
response as unmatching the current LQ table and thus break
the rs search cycle completely.
Fix this by relaxing the rate matching in case we're working
with the new API and STBC/BFER are used.
Signed-off-by: Eyal Shapira <eyalx.shapira@intel.com>
Reviewed-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
The check to avoid the shared antenna was passed the wrong
antenna parameter. It should have checked whether the antenna of
the next column we're considering is allowed and instead it was
passed the current antenna.
This could lead to a wrong choice of the next column in the rs
algorithm and non optimal performance.
Fixes: commit 219fb66b49 ("iwlwifi: mvm: rs - don't use the shared antenna when BT load is high")
Signed-off-by: Eyal Shapira <eyalx.shapira@intel.com>
Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
Integrate the module into the kernel configuration
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Integrate the assembler modules into the kernel crypto
framework. Take care to avoid long intervals of disabled
preemption.
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The assembler block cipher module that controls the core
AES functions.
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Key generation for big endian core routines.
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The assembler AES encryption and decryption core routines.
Implemented & optimized for big endian. Nevertheless they
work on little endian too.
For most efficient reuse in (higher level) block cipher
routines they are implemented as "fast" call modules without
any stack handling or register saving. The caller must
take care of that part.
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
4K AES tables for big endian. To reduce the possiblity of
timing attacks, the size has been cut to 8KB + 256 bytes
in contrast to 16KB in the generic implementation. That
is not perfect but at least a good tradeoff for CPU limited
router devices.
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Define some register aliases for better readability.
Signed-off-by: Markus Stockhausen <stockhausen@collogia.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Simplify the driver's probe function and error handling by using the
device managed allocators, while at it, drop the redundant "out of
memory" messages since these are already printed by the allocator.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Now that these definitions have been moved to
drivers/char/hw_random/bcm63xx-rng.c where they belong to make the
driver standalone, we can safely remove these definitions from
bcm63xx_regs.h.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
arch/mips/include/asm/mach-bcm63xx/bcm63xx_regs.h contains the register
definitions for this random number generator block, incorporate these
register definitions directly into the bcm63xx-rng driver so we do not
rely on this header to be provided.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
bcm_{readl,writel} macros expand to __raw_{readl,writel}, use these
directly such that we do not rely on the platform to provide these for
us. As a result, we no longer use bcm63xx_io.h, so remove that inclusion
too.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
For received packet stream, the offset of 'RX_SEQ_START' locates after
the offset of 'RX_NUMBER_MIDI', although current macro and proc output
includes wrong offsets.
Fortunately, this bug doesn't affect streaming functionality because
these macro is not used.
This commit fixes these wrong macro and outputs.
Signed-off-by: Takashi Sakamoto <o-takashi@sakamocchi.jp>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
The "usual" path is:
- rt_mutex_slowlock()
- set_current_state()
- task_blocks_on_rt_mutex() (ret 0)
- __rt_mutex_slowlock()
- sleep or not but do return with __set_current_state(TASK_RUNNING)
- back to caller.
In the early error case where task_blocks_on_rt_mutex() return
-EDEADLK we never change the task's state back to RUNNING. I
assume this is intended. Without this change after ww_mutex
using rt_mutex the selftest passes but later I get plenty of:
| bad: scheduling from the idle thread!
backtraces.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Acked-by: Mike Galbraith <umgwanakikbuti@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: afffc6c180 ("locking/rtmutex: Optimize setting task running after being blocked")
Link: http://lkml.kernel.org/r/1425056229-22326-4-git-send-email-bigeasy@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We did a failed attempt in the past to only use rcu in rtnl dump
operations (commit e67f88dd12 "net: dont hold rtnl mutex during
netlink dump callbacks")
Now that dumps are holding RTNL anyway, there is no need to also
use rcu locking, as it forbids any scheduling ability, like
GFP_KERNEL allocations that controlling path should use instead
of GFP_ATOMIC whenever possible.
This should fix following splat Cong Wang reported :
[ INFO: suspicious RCU usage. ]
3.19.0+ #805 Tainted: G W
include/linux/rcupdate.h:538 Illegal context switch in RCU read-side critical section!
other info that might help us debug this:
rcu_scheduler_active = 1, debug_locks = 0
2 locks held by ip/771:
#0: (rtnl_mutex){+.+.+.}, at: [<ffffffff8182b8f4>] netlink_dump+0x21/0x26c
#1: (rcu_read_lock){......}, at: [<ffffffff817d785b>] rcu_read_lock+0x0/0x6e
stack backtrace:
CPU: 3 PID: 771 Comm: ip Tainted: G W 3.19.0+ #805
Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
0000000000000001 ffff8800d51e7718 ffffffff81a27457 0000000029e729e6
ffff8800d6108000 ffff8800d51e7748 ffffffff810b539b ffffffff820013dd
00000000000001c8 0000000000000000 ffff8800d7448088 ffff8800d51e7758
Call Trace:
[<ffffffff81a27457>] dump_stack+0x4c/0x65
[<ffffffff810b539b>] lockdep_rcu_suspicious+0x107/0x110
[<ffffffff8109796f>] rcu_preempt_sleep_check+0x45/0x47
[<ffffffff8109e457>] ___might_sleep+0x1d/0x1cb
[<ffffffff8109e67d>] __might_sleep+0x78/0x80
[<ffffffff814b9b1f>] idr_alloc+0x45/0xd1
[<ffffffff810cb7ab>] ? rcu_read_lock_held+0x3b/0x3d
[<ffffffff814b9f9d>] ? idr_for_each+0x53/0x101
[<ffffffff817c1383>] alloc_netid+0x61/0x69
[<ffffffff817c14c3>] __peernet2id+0x79/0x8d
[<ffffffff817c1ab7>] peernet2id+0x13/0x1f
[<ffffffff817d8673>] rtnl_fill_ifinfo+0xa8d/0xc20
[<ffffffff810b17d9>] ? __lock_is_held+0x39/0x52
[<ffffffff817d894f>] rtnl_dump_ifinfo+0x149/0x213
[<ffffffff8182b9c2>] netlink_dump+0xef/0x26c
[<ffffffff8182bcba>] netlink_recvmsg+0x17b/0x2c5
[<ffffffff817b0adc>] __sock_recvmsg+0x4e/0x59
[<ffffffff817b1b40>] sock_recvmsg+0x3f/0x51
[<ffffffff817b1f9a>] ___sys_recvmsg+0xf6/0x1d9
[<ffffffff8115dc67>] ? handle_pte_fault+0x6e1/0xd3d
[<ffffffff8100a3a0>] ? native_sched_clock+0x35/0x37
[<ffffffff8109f45b>] ? sched_clock_local+0x12/0x72
[<ffffffff8109f6ac>] ? sched_clock_cpu+0x9e/0xb7
[<ffffffff810cb7ab>] ? rcu_read_lock_held+0x3b/0x3d
[<ffffffff811abde8>] ? __fcheck_files+0x4c/0x58
[<ffffffff811ac556>] ? __fget_light+0x2d/0x52
[<ffffffff817b376f>] __sys_recvmsg+0x42/0x60
[<ffffffff817b379f>] SyS_recvmsg+0x12/0x1c
Signed-off-by: Eric Dumazet <edumazet@google.com>
Fixes: 0c7aecd4bd ("netns: add rtnl cmd to add and get peer netns ids")
Cc: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Reported-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 740c7f31c0 ("sh_eth: Ensure DMA engines are stopped before
freeing buffers") added a call to sh_eth_reset() to the
sh_eth_set_ringparam() and sh_eth_close() paths.
However, setting the software reset bit(s) in the EDMR register resets
the MAC Address Registers to zero. Hence after kexec, the new kernel
doesn't detect a valid MAC address and assigns a random MAC address,
breaking DHCP.
Set the MAC address again after the reset in sh_eth_dev_exit() to fix
this.
Tested on r8a7740/armadillo (GETHER) and r8a7791/koelsch (FAST_RCAR).
Fixes: 740c7f31c0 ("sh_eth: Ensure DMA engines are stopped before freeing buffers")
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds bcmgenet_tx_poll for the tx_rings. This can reduce the
interrupt load and send xmit in network stack on time. This also
separated for the completion of tx_ring16 from bcmgenet_poll.
The bcmgenet_tx_reclaim of tx_ring[{0,1,2,3}] operative by an interrupt
is to be not more than a certain number TxBDs. It is caused by too
slowly reclaiming the transmitted skb. Therefore, performance
degradation of xmit after 605ad7f ("tcp: refine TSO autosizing").
Signed-off-by: Jaedon Shin <jaedon.shin@gmail.com>
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Brian reported crashes using IPv6 traffic with macvtap/veth combo.
I tracked the crashes in neigh_hh_output()
-> memcpy(skb->data - HH_DATA_MOD, hh->hh_data, HH_DATA_MOD);
Neighbour code assumes headroom to push Ethernet header is
at least 16 bytes.
It appears macvtap has only 14 bytes available on arches
where NET_IP_ALIGN is 0 (like x86)
Effect is a corruption of 2 bytes right before skb->head,
and possible crashes if accessing non existing memory.
This fix should also increase IPv4 performance, as paranoid code
in ip_finish_output2() wont have to call skb_realloc_headroom()
Reported-by: Brian Rak <brak@vultr.com>
Tested-by: Brian Rak <brak@vultr.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ursula Braun says:
====================
s390: network patches for net-next
here are some s390 related patches for net-next
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
remove Frank Blaschka as S390 NETWORK DRIVERS maintainer
Acked-by: Frank Blaschka <blaschka@linux.vnet.ibm.com>
Signed-off-by: Ursula Braun <ursula.braun@de.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adjusts two instances where we were using the (too big)
struct qeth_ipacmd_setadpparms size instead of the commands' actual
size. This didn't do any harm, but wasted a few bytes.
Signed-off-by: Stefan Raspl <raspl@linux.vnet.ibm.com>
Signed-off-by: Ursula Braun <ursula.braun@de.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
claw devices are outdated and no longer supported.
This patch removes the claw driver.
Signed-off-by: Ursula Braun <ursula.braun@de.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
four-way-handshake problem. The others are various small fixes
for problems all over, nothing really stands out.
-----BEGIN PGP SIGNATURE-----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=H3+7
-----END PGP SIGNATURE-----
Merge tag 'mac80211-for-davem-2015-02-27' of git://git.kernel.org/pub/scm/linux/kernel/git/jberg/mac80211
Johannes Berg says:
====================
A few patches have accumulated, among them the fix for Linus's
four-way-handshake problem. The others are various small fixes
for problems all over, nothing really stands out.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The PHY IRQ line is connected to external IRQ8, not IRQ0. Fix DT
accordingly.
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
Adds sound card using the ak4643 codec.
Signed-off-by: Ulrich Hecht <ulrich.hecht+renesas@gmail.com>
Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
Adds an entry for an R-Car Gen. 1 sound controller.
Signed-off-by: Ulrich Hecht <ulrich.hecht+renesas@gmail.com>
[horms: Removed undocumented "renesas,rcar_sound" compat string]
Signed-off-by: Simon Horman <horms+renesas@verge.net.au>
Modify cpuidle_enter_freeze() to do the sanity checks done by
cpuidle_select() to avoid crashing the suspend-to-idle code
path in case something is missing.
Fixes: 3810631332 (PM / sleep: Re-implement suspend-to-idle handling)
Original-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Disabling interrupts at the end of cpuidle_enter_freeze() is not
useful, because its caller, cpuidle_idle_call(), re-enables them
right away after invoking it.
To avoid that unnecessary back and forth dance with interrupts,
make cpuidle_enter_freeze() enable interrupts after calling
enter_freeze_proper() and drop the local_irq_disable() at its
end, so that all of the code paths in it end up with interrupts
enabled. Then, cpuidle_idle_call() will not need to re-enable
interrupts after calling cpuidle_enter_freeze() any more, because
the latter will return with interrupts enabled, in analogy with
cpuidle_enter().
Reported-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
tcp_fastopen_create_child() is static and should not be exported.
tcp4_gso_segment() and tcp6_gso_segment() should be static.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds support for reporting the actual and maximum combined channels
count of the hv_netvsc driver via 'ethtool --show-channels'.
This required adding 'max_chn' to 'struct netvsc_device', and assigning
it 'rsscap.num_recv_que' in 'rndis_filter_device_add'. Now we can access
the combined maximum channel count via 'struct netvsc_device' in the
ethtool callback.
Signed-off-by: Andrew Schwartzmeyer <andrew@schwartzmeyer.com>
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When using the wext compatibility code in cfg80211, part of the IEs
can be truncated if the passed user buffer is large enough for part
of the BSS but not large enough for all of the IEs. This can cause
an EAP network to show up as a PSK network.
Always return -E2BIG in this case to avoid truncating data.
Since this changes the control flow, use an on-stack variable for
a small buffer instead of allocating it.
Signed-off-by: James Minor <james.minor@ni.com>
[rework patch to error out immediately, use _check wrappers]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
These checked wrappers are necessary for the next patch, which
will use them to avoid sending out partial scan results.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Revert commit ad38bfc916 ("mac80211: Tx frame latency statistics")
(along with some follow-up fixes).
This code turned out not to be as useful in the current form as we
thought, and we've internally hacked it up more, but that's not
very suitable for upstream (for now), and we might just do that
with tracing instead.
Therefore, for now at least, remove this code. We might also need
to use the skb->tstamp field for the TCP performance issue, which
is more important than the debugging.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Both wpa_supplicant and mac80211 have and inactivity timer. By default
wpa_supplicant will be timed out in 5 minutes and mac80211's it is 30
minutes. If wpa_supplicant uses a longer timer than mac80211, it will
get unexpected disconnection by mac80211.
Using 0xffffffff instead as the configured value could solve this w/o
changing the code, but due to integer overflow in the expression used
this doesn't work. The expression is:
(current jiffies) > (frame Rx jiffies + NL80211_MESHCONF_PLINK_TIMEOUT * 250)
On 32bit system, the right side would overflow and be a very small
value if NL80211_MESHCONF_PLINK_TIMEOUT is sufficiently large,
causing unexpectedly early disconnections.
Instead allow disabling the inactivity timer to avoid this situation,
by passing the (previously invalid and useless) value 0.
Signed-off-by: Masashi Honma <masashi.honma@gmail.com>
[reword/rewrap commit log]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Due to the checks in get_hwsim_data_ref_from_addr, wmediumd
was only able to use the second mac address (those starting with
0x42). This is confusing and needlessly limiting, so allow any
configured address.
Signed-off-by: Bob Copeland <me@bobcopeland.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
When a fully converted cfg80211 driver needs cfg80211-wext for
userspace API purposes, the symbols need not be exported. When
other drivers (orinoco/hermes or ipw2200) are enabled, they do
need the symbols exported as they use them directly.
Make those drivers select a new CFG80211_WEXT_EXPORT Kconfig
symbol (instead of just CFG80211_WEXT) and export the functions
only if requested - this saves about 1/2k due to the size of
EXPORT_SYMBOL() itself.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
When changing AP SMPS, we need to look up all the stations
for this interface, so there's no reason to iterate over
hash chains rather than doing the simpler iteration over
the station list.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Since multicast addresses don't exist as stations, don't attempt
to look them up in the hashtable on TX.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
When applicable verify that the caller has permisson to the underlying
network namespace for a newly created network device.
Similary checks exist for the network namespace a network device will
be created in.
Fixes: 317f4810e4 ("rtnl: allow to create device with IFLA_LINK_NETNSID set")
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When applicable verify that the caller has permision to create a
network device in another network namespace. This check is already
present when moving a network device between network namespaces in
setlink so all that is needed is to duplicate that check in newlink.
This change almost backports cleanly, but there are context conflicts
as the code that follows was added in v4.0-rc1
Fixes: b51642f6d7 net: Enable a userns root rtnl calls that are safe for unprivilged users
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric Dumazet says:
====================
tcp: tso improvements
This patch serie reworks tcp_tso_should_defer() a bit
to get less bursts, and better ECN behavior.
We also removed tso_deferred field in tcp socket.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Another TCP issue is triggered by ECN.
Under pressure, receiver gets ECN marks, and send back ACK packets
with ECE TCP flag. Senders enter CA_CWR state.
In this state, tcp_tso_should_defer() is short cut :
if (icsk->icsk_ca_state != TCP_CA_Open)
goto send_now;
This means that about all ACK packets we receive are triggering
a partial send, and because cwnd is kept small, we can only send
a small amount of data for each incoming ACK,
which in return generate more ACK packets.
Allowing CA_Open and CA_CWR states to enable TSO defer in
tcp_tso_should_defer() brings performance back :
TSO autodefer has more chance to defer under pressure.
This patch increases TSO and LRO/GRO efficiency back to normal levels,
and does not impact overall ECN behavior.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With sysctl_tcp_min_tso_segs being 4, it is very possible
that tcp_tso_should_defer() decides not sending last 2 MSS
of initial window of 10 packets. This also applies if
autosizing decides to send X MSS per GSO packet, and cwnd
is not a multiple of X.
This patch implements an heuristic based on age of first
skb in write queue : If it was sent very recently (less than half srtt),
we can predict that no ACK packet will come in less than half rtt,
so deferring might cause an under utilization of our window.
This is visible on initial send (IW10) on web servers,
but more generally on some RPC, as the last part of the message
might need an extra RTT to get delivered.
Tested:
Ran following packetdrill test
// A simple server-side test that sends exactly an initial window (IW10)
// worth of packets.
`sysctl -e -q net.ipv4.tcp_min_tso_segs=4`
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0 bind(3, ..., ...) = 0
+0 listen(3, 1) = 0
+.1 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7>
+0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6>
+.1 < . 1:1(0) ack 1 win 257
+0 accept(3, ..., ...) = 4
+0 write(4, ..., 14600) = 14600
+0 > . 1:5841(5840) ack 1 win 457
+0 > . 5841:11681(5840) ack 1 win 457
// Following packet should be sent right now.
+0 > P. 11681:14601(2920) ack 1 win 457
+.1 < . 1:1(0) ack 14601 win 257
+0 close(4) = 0
+0 > F. 14601:14601(0) ack 1
+.1 < F. 1:1(0) ack 14602 win 257
+0 > . 14602:14602(0) ack 2
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TSO relies on ability to defer sending a small amount of packets.
Heuristic is to wait for future ACKS in hope to send more packets at once.
Current algorithm uses a per socket tso_deferred field as a pseudo timer.
This pseudo timer relies on future ACK, but there is no guarantee
we receive them in time.
Fix would be to use a real timer, but cost of such timer is probably too
expensive for typical cases.
This patch changes the logic to test the time of last transmit,
because we should not add bursts of more than 1ms for any given flow.
We've used this patch for about two years at Google, before FQ/pacing
as it would reduce a fair amount of bursts.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
