Re: Revert "gro: Fix legacy path napi_complete crash",

[David Miller]

From: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
Date: Thu, 26 Mar 2009 10:41:29 +0800

> On Wed, Mar 25, 2009 at 11:54:56PM +0100, Jarek Poplawski wrote:
> > 
> > Of course it's too late for verifying this now, but (for the future)
> > I think, this scenario could be considered:
> > 
> > process_backlog()			netif_rx()
> > 
> > if (!skb)
> > local_irq_enable()
> > 					if (queue.qlen) //NO
> > 					napi_schedule() //NOTHING
> > 					__skb_queue_tail() //qlen > 0
> > napi_complete()
> > ...					...
> > 					Every next netif_rx() sees
> > 					qlen > 0, so napi is never
> > 					scheduled again.
> > 
> > Then, something like this might work...
> 
> Yes this is why my original patch that started all this is broken.
> However, this doesn't apply to my patch that open-codes __napi_complete.

There is still a difference compared to your fix Herbert.  Jarek's
patch flushes GRO first before the unlink.

I still believe that's critical, although like you I can't pinpoint
why.

I know that GRO ought to be disabled here, but what if for some reason
it isn't? :-)

Adam Richter has successfully tested Jarek's variant, and if Ingo's
tests show that it makes his problem go away too then I'm definitely
going with Jarek's patch.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/