Skip to content

R
raspberrypi-kernel

openeuler / raspberrypi-kernel

通知 13
Star 1
Fork 0
R
raspberrypi-kernel
切换分支/标签
  1. 23 8月, 2009 1 次提交
  3. 09 6月, 2009 1 次提交
  5. 18 5月, 2009 1 次提交
    • E
      bnx2: bnx2_tx_int() optimizations · d62fda08
      由 Eric Dumazet 提交于 
      When using bnx2 in a high transmit load, bnx2_tx_int() cost is pretty high.

There are two reasons.

One is an expensive call to bnx2_get_hw_tx_cons(bnapi) for each freed skb

One is cpu stalls when accessing skb_is_gso(skb) / skb_shinfo(skb)->nr_frags
because of two cache line misses.
(One to get skb->end/head to compute skb_shinfo(skb),
 one to get is_gso/nr_frags)

This patch :

1) avoids calling bnx2_get_hw_tx_cons(bnapi) too many times.

2) makes bnx2_start_xmit() cache is_gso & nr_frags into sw_tx_bd descriptor.
   This uses a litle bit more ram (256 longs per device on x86), but helps a lot.

3) uses a prefetch(&skb->end) to speedup dev_kfree_skb(), bringing
  cache line that will be needed in skb_release_data()

result is 5 % bandwidth increase in benchmarks, involving UDP or TCP receive
 & transmits, when a cpu is dedicated to ksoftirqd for bnx2.

bnx2_tx_int going from 3.33 % cpu to 0.5 % cpu in oprofile

Note : skb_dma_unmap() still very expensive but this is for another patch,
not related to bnx2 (2.9 % of cpu, while it does nothing on x86_32)
Signed-off-by: NEric Dumazet <dada1@cosmosbay.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      d62fda08
  7. 05 4月, 2009 1 次提交
    • M
      bnx2: Use request_firmware() · 57579f76
      由 Michael Chan 提交于 
      Based on original patch by Ben Hutchings <ben@decadent.org.uk> and
Bastian Blank <waldi@debian.org>, with the following main changes:

Separated the mips firmware and rv2p firmware into different files
to make it easier to update them separately.

Added some code to fixup the rv2p code with run-time information
such as PAGE_SIZE.

Update version to 2.0.0.
Signed-off-by: NMichael Chan <mchan@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      57579f76
  9. 13 2月, 2009 1 次提交
  11. 17 12月, 2008 2 次提交
  13. 03 12月, 2008 1 次提交
    • M
      bnx2: Add workaround to handle missed MSI. · efba0180
      由 Michael Chan 提交于 
      The bnx2 chips do not support per MSI vector masking.  On 5706/5708, new MSI
address/data are stored only when the MSI enable bit is toggled.  As a result,
SMP affinity no longer works in the latest kernel.  A more serious problem is
that the driver will no longer receive interrupts when the MSI receiving CPU
goes offline.

The workaround in this patch only addresses the problem of CPU going offline.
When that happens, the driver's timer function will detect that it is making
no forward progress on pending interrupt events and will recover from it.

Eric Dumazet reported the problem.

We also found that if an interrupt is internally asserted while MSI and INTA
are disabled, the chip will end up in the same state after MSI is re-enabled.
The same workaround is needed for this problem. 
Signed-off-by: NMichael Chan <mchan@broadcom.com>
Tested-by: NEric Dumazet <dada1@cosmosbay.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      efba0180
  15. 29 11月, 2008 1 次提交
  17. 13 11月, 2008 2 次提交
  19. 10 10月, 2008 1 次提交
  21. 19 9月, 2008 3 次提交
  23. 19 7月, 2008 1 次提交
  25. 15 7月, 2008 2 次提交
  27. 20 6月, 2008 4 次提交
  29. 12 6月, 2008 1 次提交
    • D
      net: Eliminate flush_scheduled_work() calls while RTNL is held. · 4bb073c0
      由 David S. Miller 提交于 
      If the RTNL is held when we invoke flush_scheduled_work() we could
deadlock.  One such case is linkwatch, it is a work struct which tries
to grab the RTNL semaphore.

The most common case are net driver ->stop() methods.  The
simplest conversion is to instead use cancel_{delayed_}work_sync()
explicitly on the various work struct the driver uses.

This is an OK transformation because these work structs are doing
things like resetting the chip, restarting link negotiation, and so
forth.  And if we're bringing down the device, we're about to turn the
chip off and reset it anways.  So if we cancel a pending work event,
that's fine here.

Some drivers were working around this deadlock by using a msleep()
polling loop of some sort, and those cases are converted to instead
use cancel_{delayed_}work_sync() as well.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      4bb073c0
  31. 17 5月, 2008 2 次提交
  33. 24 2月, 2008 1 次提交
  35. 01 2月, 2008 5 次提交
  37. 29 1月, 2008 9 次提交