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

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

Le mardi 05 juillet 2011 à 02:29 +0400, Alexey Zaytsev a écrit :
> On Tue, Jul 5, 2011 at 00:25, Alexey Zaytsev <alexey.zaytsev@gmail.com> wrote:
> > On Mon, Jul 4, 2011 at 19:12, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >> Le lundi 04 juillet 2011 à 16:43 +0200, Michael Büsch a écrit :
> >>> On Mon, 4 Jul 2011 16:27:26 +0200
> >>> Michael Büsch <m@bues.ch> wrote:
> >>> > We do this in b43, which has exactly the same DMA engine.
> >>>
> >>> (Ok, it turns out we don't do this in b43 (We only do it on the TX side).
> >>>  But that's a bug. We should do a wmb() on the RX side before advancing the
> >>>  descriptor ring pointer.)
> >>
> >> I am wondering what happens if RX ring is set to 64, and we receive
> >> exactly 64 buffers in one round, B44_DMARX_PTR wont change at all ?
> >>
> >> Alexey, could you try this patch please ?
> >
> > Sorry, did not help.
> >
> 
> Ran a few rounds of tcpdump. Seeing a significant number or duplicate
> ACKs from the problematic machine. Not seeing them when testing
> between this machine and an other linux box. Or the illumos machine
> and the other linux box.
> 
> Dumps are available here:
> 
> http://zaytsev.su/tmp/caps/
> 
> dump1-3 - between the problematic machine an the illumos box,
> collected on illumos side. All show dups.
> dump5 - between an other linux box and the illumos machine, no dups.
> Collcted on the illumos side.
> dump-linux - between 2 linux machines, collected on the
> non-problematic side. No dups, no corruptions.
> 
> 192.168.0.33 - the problematic machine.
> 192.168.0.72 - the illumos machine.
> 192.168.0.122 - an other linux machine.

??

I dont care about duplicate acks at this point.

Thats a separate issue (TCP layer)

Do you still have memory scribbles ?

I wonder if the problem is not coming from the "fast recovery" added in
commit 32737e934a952c (PATCH: b44 Handle RX FIFO overflow better
(simplified))

Maybe we should do instead a fast dequeue of packets (recycling them
instead of pushing them to upper stack) in case too many packets are
ready to be delivered, and always make sure NIC has a reserve of
available buffers for DMA accesses, before it can assert ISTAT_RFO