Re: [Bugme-new] [Bug 38102] New: BUG kmalloc-2048: Poison overwritten

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 05 juillet 2011 à 07:56 +0400, Alexey Zaytsev a écrit :
> On Tue, Jul 5, 2011 at 07:44, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >
> > I dont care about duplicate acks at this point.
> >
> > Thats a separate issue (TCP layer)
> >
> 
> Maybe some tx packets are just sent out more then once? Or a single
> packet is sent out instead of some other packets?
> The delays between two dups is short, and they come in bursts, up to a
> few hundreds of duplicate packets at a time.
> 

Thats a completely different problem. SSH is very expensive for your
receiver (your dump1 file has small packets (560 bytes)), and it cannot
cope with the stress.

You're filling the b44 rx ring, and b44 driver has no choice to zap 200
packets at once. This sure is a problem for tcp, as it stalls the thing.

You could avoid this by doing this at b44 machine (the receiver)

echo "4096 32768 87380" >/proc/sys/net/ipv4/tcp_rmem

So that sender wont be able to push so many packets

> > Do you still have memory scribbles ?
> Yes.

OK

> 
> >
> > I wonder if the problem is not coming from the "fast recovery" added in
> > commit 32737e934a952c (PATCH: b44 Handle RX FIFO overflow better
> > (simplified))
> >
> 
> I've tested back to 2.6.27. I did not test all releases of course, so
> maybe this was fixed, and then broken again.