If we hack the virtio_net driver to always allocate full-sized (64k+)
skbuffs, the driver slows down (lguest numbers):
Time to receive 1GB (small buffers): 10.85 seconds
Time to receive 1GB (64k+ buffers): 24.75 seconds
Of course, large buffers use up more space in the ring, so we increase
that from 128 to 2048:
Time to receive 1GB (64k+ buffers, 2k ring): 16.61 seconds
If we recycle pages rather than using alloc_page/free_page:
Time to receive 1GB (64k+ buffers, 2k ring, recycle pages): 10.81 seconds
This demonstrates that with efficient allocation, we don't need to
have a separate "small buffer" queue.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Finally this patch lets virtio_net receive GSO packets in addition
to sending them. This can definitely be optimised for the non-GSO
case. For comparison the Xen approach stores one page in each skb
and uses subsequent skb's pages to construct an SG skb instead of
preallocating the maximum amount of pages per skb.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> (added feature bits)
This patch adds some basic ethtool operations to virtio_net so
I could test SG without GSO (which was really useful because TSO
turned out to be buggy :)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> (remove MTU setting)
On Mon, 2008-05-26 at 17:42 +1000, Rusty Russell wrote:
> If we fail to transmit a packet, we assume the queue is full and put
> the skb into last_xmit_skb. However, if more space frees up before we
> xmit it, we loop, and the result can be transmitting the same skb twice.
>
> Fix is simple: set skb to NULL if we've used it in some way, and check
> before sending.
...
> diff -r 564237b31993 drivers/net/virtio_net.c
> --- a/drivers/net/virtio_net.c Mon May 19 12:22:00 2008 +1000
> +++ b/drivers/net/virtio_net.c Mon May 19 12:24:58 2008 +1000
> @@ -287,21 +287,25 @@ again:
> free_old_xmit_skbs(vi);
>
> /* If we has a buffer left over from last time, send it now. */
> - if (vi->last_xmit_skb) {
> + if (unlikely(vi->last_xmit_skb)) {
> if (xmit_skb(vi, vi->last_xmit_skb) != 0) {
> /* Drop this skb: we only queue one. */
> vi->dev->stats.tx_dropped++;
> kfree_skb(skb);
> + skb = NULL;
> goto stop_queue;
> }
> vi->last_xmit_skb = NULL;
With this, may drop an skb and then later in the function discover that
we could have sent it after all. Poor wee skb :)
How about the incremental patch below?
Cheers,
Mark.
Subject: [PATCH] virtio_net: Delay dropping tx skbs
Currently we drop the skb in start_xmit() if we have a
queued buffer and fail to transmit it.
However, if we delay dropping it until we've stopped the
queue and enabled the tx notification callback, then there
is a chance space might become available for it.
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband:
MAINTAINERS: Remove Glenn Streiff from NetEffect entry
mlx4_core: Improve error message when not enough UAR pages are available
IB/mlx4: Add support for memory management extensions and local DMA L_Key
IB/mthca: Keep free count for MTT buddy allocator
mlx4_core: Keep free count for MTT buddy allocator
mlx4_code: Add missing FW status return code
IB/mlx4: Rename struct mlx4_lso_seg to mlx4_wqe_lso_seg
mlx4_core: Add module parameter to enable QoS support
RDMA/iwcm: Remove IB_ACCESS_LOCAL_WRITE from remote QP attributes
IPoIB: Include err code in trace message for ib_sa_path_rec_get() failures
IB/sa_query: Check if sm_ah is NULL in ib_sa_remove_one()
IB/ehca: Release mutex in error path of alloc_small_queue_page()
IB/ehca: Use default value for Local CA ACK Delay if FW returns 0
IB/ehca: Filter PATH_MIG events if QP was never armed
IB/iser: Add support for RDMA_CM_EVENT_ADDR_CHANGE event
RDMA/cma: Add RDMA_CM_EVENT_TIMEWAIT_EXIT event
RDMA/cma: Add RDMA_CM_EVENT_ADDR_CHANGE event
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/gerg/m68knommu:
m68knommu: put ColdFire head code into .text.head section
m68knommu: remove last use of CONFIG_FADS and CONFIG_RPXCLASSIC
m68knommu: remove RPXCLASSIC from the m68k tree
m68knommu: fec: remove FADS
m68knommu: MCF5307 PIT GENERIC_CLOCKEVENTS support
m68knommu: add read_barrier_depends() and irqs_disabled_flags()
m68knommu: add byteswap assembly opcode for ISA A+
m68knommu: add ffs and __ffs plattform which support ISA A+ or ISA C
m68knommu: add sched_clock() for the DMA timer
m68knommu: complete generic time
m68knommu: move code within time.c
m68knommu: m68knommu: add old stack trace method
m68knommu: Add Coldfire DMA Timer support
m68knommu: defconfig for M5407C3 board
m68knommu: defconfig for M5307C3 board
m68knommu: defconfig for M5275EVB board
m68knommu: defconfig for M5249EVB board
m68knommu: change to a configs directory for board configurations
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6:
pkt_sched: sch_sfq: dump a real number of flows
atm: [fore200e] use MODULE_FIRMWARE() and other suggested cleanups
netfilter: make security table depend on NETFILTER_ADVANCED
tcp: Clear probes_out more aggressively in tcp_ack().
e1000e: fix e1000_netpoll(), remove extraneous e1000_clean_tx_irq() call
net: Update entry in af_family_clock_key_strings
netdev: Remove warning from __netif_schedule().
sky2: don't stop queue on shutdown
On 32-bit architectures PAGE_ALIGN() truncates 64-bit values to the 32-bit
boundary. For example:
u64 val = PAGE_ALIGN(size);
always returns a value < 4GB even if size is greater than 4GB.
The problem resides in PAGE_MASK definition (from include/asm-x86/page.h for
example):
#define PAGE_SHIFT 12
#define PAGE_SIZE (_AC(1,UL) << PAGE_SHIFT)
#define PAGE_MASK (~(PAGE_SIZE-1))
...
#define PAGE_ALIGN(addr) (((addr)+PAGE_SIZE-1)&PAGE_MASK)
The "~" is performed on a 32-bit value, so everything in "and" with
PAGE_MASK greater than 4GB will be truncated to the 32-bit boundary.
Using the ALIGN() macro seems to be the right way, because it uses
typeof(addr) for the mask.
Also move the PAGE_ALIGN() definitions out of include/asm-*/page.h in
include/linux/mm.h.
See also lkml discussion: http://lkml.org/lkml/2008/6/11/237
[akpm@linux-foundation.org: fix drivers/media/video/uvc/uvc_queue.c]
[akpm@linux-foundation.org: fix v850]
[akpm@linux-foundation.org: fix powerpc]
[akpm@linux-foundation.org: fix arm]
[akpm@linux-foundation.org: fix mips]
[akpm@linux-foundation.org: fix drivers/media/video/pvrusb2/pvrusb2-dvb.c]
[akpm@linux-foundation.org: fix drivers/mtd/maps/uclinux.c]
[akpm@linux-foundation.org: fix powerpc]
Signed-off-by: Andrea Righi <righi.andrea@gmail.com>
Cc: <linux-arch@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Although mv643xx_eth has no hardware support for inserting a vlan
tag by twiddling some bits in the TX descriptor, it does support
hardware TX checksumming on packets where the IP header starts {a
limited set of values other than 14} bytes into the packet.
This patch sets mv643xx_eth's ->vlan_features to NETIF_F_SG |
NETIF_F_IP_CSUM, which prevents the stack from checksumming vlan'ed
packets in software, and if vlan tags are present on a transmitted
packet, notifies the hardware of this fact by toggling the right
bits in the TX descriptor.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
When there is a link status change (link or phy status interrupt),
print a message notifying the user of the new link status.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
The mv643xx_eth hardware has a provision for polling the PHY's
MII management registers to obtain the (R)(G)MII interface speed
(10/100/1000) and duplex (half/full) and pause (off/symmetric)
settings to use to talk to the PHY.
The driver currently does not make use of this feature. Instead,
whenever there is a link status change event, it reads the current
link parameters from the PHY, and programs those parameters into
the mv643xx_eth MAC by hand.
This patch switches the mv643xx_eth driver to letting the MAC
auto-determine the (R)(G)MII link parameters by PHY polling, if there
is a PHY present. For PHYless ports (when e.g. the (R)(G)MII
interface is connected to a hardware switch), we keep hardcoding the
MII interface parameters.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Instead of hardcoding MII register addresses and values, use the
symbolic constants defined in linux/mii.h.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
The mv643xx_eth driver is limiting DMA bursts to 32 bytes, while
using the largest burst size (128 bytes) gives a couple percentage
points performance improvement in throughput tests, and the docs
say that the 128 byte default should not need to be changed, so
use 128 byte bursts instead.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
The recommended sequence for waiting for the transmit path to clear
after disabling all of the transmit queues is to wait for the
TX_FIFO_EMPTY bit in the Port Status register to become set as well
as the TX_IN_PROGRESS bit to clear.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
The maximum receive packet size field in the Port Serial Control
register controls at what size received packets are flagged
overlength in the receive descriptor, but it doesn't prevent
overlength packets from being DMAd to memory and signaled to the
host like other received packets.
mv643xx_eth does not support receiving jumbo frames in 10/100 mode,
but setting the packet threshold to larger than 1522 bytes in 10/100
mode won't cause breakage by itself.
If we really want to enforce maximum packet size on the receiving
end instead of on the sending end where it should be done, we can
always just add a length check to the software receive handler
instead of relying on the hardware to do the comparison for us.
What's more, changing the maximum packet size field requires
temporarily disabling the RX/TX paths. So once the link comes
up in 10/100 Mb/s mode or 1000 Mb/s mode, we'd have to disable it
again just to set the right maximum packet size field (1522 in
10/100 Mb/s mode or 9700 in 1000 Mb/s mode), just so that we can
offload one comparison operation to hardware that we might as well
do in software, assuming that we'd want to do it at all.
Contrary to what the documentation suggests, there is no harm in
just setting a 9700 byte maximum packet size in 10/100 mode, so use
the maximum maximum packet size for all modes.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
The mv643xx_eth driver allows doing transmit reclaim from within the
napi poll routine, but after doing reclaim, it would forget to check
the free transmit descriptor count and wake up the transmit queue if
the reclaim caused enough descriptors for a new packet to become
available. This would cause the netdev watchdog to occasionally kick
in during certain workloads with combined receive and transmit traffic.
Fix this by adding a wakeup check identical to the one in the
interrupt handler to the napi poll routine.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
When the ethernet link goes down while mv643xx_eth is transmitting
data, transmit DMA can stop before all queued transmit descriptors
have been processed. But even the descriptors that _have_ been
processed might not be properly marked as done before the transmit
DMA unit shuts down.
Then when the link comes up again, the hardware transmit pointer
might have advanced while not all previous packet descriptors have
been marked as transmitted, causing software transmit reclaim to
hang waiting for the hardware to finish transmitting a descriptor
that it has already skipped.
This patch forcibly reclaims all packets on the transmit ring on a
link down interrupt, and then resyncs the hardware transmit pointer to
what the software's idea of the first free descriptor is. Also, we
need to prevent re-waking the transmit queue if we get a 'transmit
done' interrupt at the same time as a 'link down' interrupt, which
this patch does as well.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
The previously merged TX hang erratum workaround ("mv643xx_eth:
work around TX hang hardware issue") assumes that TX_END interrupts
are delivered simultaneously with or after their corresponding TX
interrupts, but this is not always true in practise.
In particular, it appears that TX_END interrupts are issued as soon
as descriptor fetch returns an invalid descriptor, which may happen
before earlier descriptors have been fully transmitted and written
back to memory as being done.
This hardware behavior can lead to a situation where the current
driver code mistakenly assumes that the MAC has given up transmitting
before noticing the packets that it is in fact still currently working
on, causing the driver to re-kick the transmit queue, which will only
cause the MAC to re-fetch the invalid head descriptor, and generate
another TX_END interrupt, et cetera, until the packets in the pipe
finally finish transmitting and have their descriptors written back
to memory, which will then finally break the loop.
Fix this by having the erratum workaround not check the 'number of
unfinished descriptor', but instead, to compare the software's idea
of what the head descriptor pointer should be to the hardware's head
descriptor pointer (which is updated on the same conditions as the
TX_END interupt is generated on, i.e. possibly before all previous
descriptors have been transmitted and written back).
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
* 'devel' of master.kernel.org:/home/rmk/linux-2.6-arm: (85 commits)
[ARM] pxa: add base support for PXA930 Handheld Platform (aka SAAR)
[ARM] pxa: add base support for PXA930 Evaluation Board (aka TavorEVB)
[ARM] pxa: add base support for PXA930 (aka Tavor-P)
[ARM] Update mach-types
[ARM] pxa: make littleton to use the new smc91x platform data
[ARM] pxa: make zylonite to use the new smc91x platform data
[ARM] pxa: make mainstone to use the new smc91x platform data
[ARM] pxa: make lubbock to use new smc91x platform data
[NET] smc91x: prepare SMC_USE_PXA_DMA to be specified in platform data
[NET] smc91x: prepare for SMC_IO_SHIFT to be a platform configurable variable
[NET] smc91x: add SMC91X_NOWAIT flag to platform data
[NET] smc91x: favor the use of SMC91X_USE_* instead of SMC_CAN_USE_*
[NET] smc91x: remove "irq_flags" from "struct smc91x_platdata"
[ARM] 5146/1: pxa2xx: convert all boards to call pxa2xx_transceiver_mode helper
Support for LCD on e740 e750 e400 and e800 e-series PDAs
E-series UDC support
PXA UDC - allow use of inverted GPIO for pullup
Add e350 support
Fix broken e-series build
E-series GPIO / IRQ definitions.
...
Evgeniy Polyakov noticed that drivers/net/e1000e/netdev.c:e1000_netpoll()
was calling e1000_clean_tx_irq() without taking the TX lock.
David Miller suggested to remove the call altogether: since in this
callpah there's periodic calls to ->poll() anyway which will do
e1000_clean_tx_irq() and will garbage-collect any finished TX ring
descriptors.
This fix solved the e1000e+netconsole crashes i've been seeing:
=============================================================================
BUG skbuff_head_cache: Poison overwritten
-----------------------------------------------------------------------------
INFO: 0xf658ae9c-0xf658ae9c. First byte 0x6a instead of 0x6b
INFO: Allocated in __alloc_skb+0x2c/0x110 age=0 cpu=0 pid=5098
INFO: Freed in __kfree_skb+0x31/0x80 age=0 cpu=1 pid=4440
INFO: Slab 0xc16cc140 objects=16 used=1 fp=0xf658ae00 flags=0x400000c3
INFO: Object 0xf658ae00 @offset=3584 fp=0xf658af00
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: David S. Miller <davem@davemloft.net>
If an mlx4 device with default FW (which gives a UAR BAR size of 8 MB)
is used in a system with 64 KB pages, then there are only 8192/64==128
UAR pages available. However, the first 128 UAR pages are reserved
for use with event queue doorbells, so no UAR pages are available to
do anything else with, which means that the driver cannot work.
The current driver fails with a fairly cryptic "Failed to allocate
driver access region, aborting" message in this situation. Fix the
driver to detect the problem earlier and print out a clearer
description of the problem and a suggestion of how to fix it (use a
new firmware image).
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add support for the following operations to mlx4 when device firmware
supports them:
- Send with invalidate and local invalidate send queue work requests;
- Allocate/free fast register MRs;
- Allocate/free fast register MR page lists;
- Fast register MR send queue work requests;
- Local DMA L_Key.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
They have never been used in this port of the driver. It is has only
ever been used on the ColdFire SoC ethernet core.
Signed-off-by: Greg Ungerer <gerg@uclinux.org>
This ifdefs are leftovers from the time as the driver was running
on a ppc.
Signed-off-by: Sebastian Siewior <bigeasy@linutronix.de>
Signed-off-by: Greg Ungerer <gerg@uclinux.org>
I found config FADS only in ppc/Kconfig. Bye bye relic.
Signed-off-by: Sebastian Siewior <bigeasy@linutronix.de>
Signed-off-by: Greg Ungerer <gerg@uclinux.org>
It is unnecessary, to stop queue and turn off carrier in shutdown
routine. With new netdev_queue this causes warnings.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6: (82 commits)
ipw2200: Call netif_*_queue() interfaces properly.
netxen: Needs to include linux/vmalloc.h
[netdrvr] atl1d: fix !CONFIG_PM build
r6040: rework init_one error handling
r6040: bump release number to 0.18
r6040: handle RX fifo full and no descriptor interrupts
r6040: change the default waiting time
r6040: use definitions for magic values in descriptor status
r6040: completely rework the RX path
r6040: call napi_disable when puting down the interface and set lp->dev accordingly.
mv643xx_eth: fix NETPOLL build
r6040: rework the RX buffers allocation routine
r6040: fix scheduling while atomic in r6040_tx_timeout
r6040: fix null pointer access and tx timeouts
r6040: prefix all functions with r6040
rndis_host: support WM6 devices as modems
at91_ether: use netstats in net_device structure
sfc: Create one RX queue and interrupt per CPU package by default
sfc: Use a separate workqueue for resets
sfc: I2C adapter initialisation fixes
...
netif_carrier_{on,off}() handles starting and stopping packet
flow into the driver. So there is no reason to invoke netif_stop_queue()
and netif_wake_queue() in response to link status events.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch reworks the error handling in r6040_init_one
in order not to leak resources and correcly unmap and release
PCI regions of the MAC. Also prefix printk's with the driver name
for clarity.
Signed-off-by: Florian Fainelli <florian.fainelli@telecomint.eu>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
This patch bumps the release of the r6040 driver. There has been
quite some versions of it out there, but this one is the one
people should report bugs against.
Signed-off-by: Florian Fainelli <florian.fainelli@telecomint.eu>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
This patch allows the MAC to handle the RX FIFO full
and no descriptor available interrupts. While we are at it
replace the TX interrupt with its corresponding definition.
Signed-off-by: Florian Fainelli <florian.fainelli@telecomint.eu>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
This patch changes the default waiting time of a packet, which
along with our previous r6040_rx path, was causing huge delays
with another host (160 to 230 ms).
Signed-off-by: Florian Fainelli <florian.fainelli@telecomint.eu>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Define all the descriptor status the MAC can set.
Signed-off-by: Florian Fainelli <florian.fainelli@telecomint.eu>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
