[ofa-general] [PATCH] IPoIB: keep ib_wc[] on stack

[akepner at sgi.com]

On Wed, May 14, 2008 at 10:25:48AM +0300, Eli Cohen wrote:

> On Tue, 2008-05-13 at 18:21 -0700, akepner at sgi.com wrote:
> > We're getting panics like this one on big clusters:
> > 
> > skb_over_panic: text:ffffffff8821f32e len:160 put:100 head:ffff810372b0f000 data:ffff810372b0f01c tail:ffff810372b0f0bc end:ffff810372b0f080 dev:ib0
> 
> RX SKBs are large enough to contain 100 bytes... this looks like
> corruption. 

Exactly.

> Can you give more information on OS, kernel version, OFED
> version.

SUSE Linux Enterprise Server 10 SP1 (x86_64) - Kernel 2.6.16.46-0.12-smp
OFED 1.3 GA

> .....
> >From NAPI_HOWTO.txt, although the file has been removed but I think the
> statement is still valid:
> 
> -Guarantee: Only one CPU at any time can call dev->poll(); this is
> because only one CPU can pick the initial interrupt and hence the
> initial netif_rx_schedule(dev);
> 

Yes, you're correct. I missed the use of the __LINK_STATE_RX_SCHED 
bit in __netif_rx_schedule_prep()/netif_rx_complete() that serializes 
this. (Roland also pointed this out to me.)

-- 
Arthur