Re: Regression associated with commit c8628155ece3 - "tcp: reduce out_of_order memory use"

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

On Thu, 2012-08-23 at 15:57 -0500, Larry Finger wrote:
> On 08/22/2012 11:03 PM, Eric Dumazet wrote:
> >
> > Changing the allocation size removes the problem ? thats really strange.
> >
> > If you try different sizes in the 9100-30720 range, can you pinpoint the
> > failure threshold ?
> 
> The allocation size change did not fix the problem. It turned out that 10 tries 
> from a secure web page were not enough to trigger this intermittent problem that 
> particular test.
> 
> Based on DaveM's comment that skb->truesize could be wrong, I tried setting 
> truesize after every netdev_alloc_skb() call. Of course, that had no effect. I 
> then found https://lkml.org/lkml/2010/11/19/505I, which clearly states why this 
> need not be done.
> 
> What skb modifications require that truesize be adjusted? The driver never 
> resets skb->len or skb->data_len for any buffers, other than setting skb->len to 
> zero.

skb->truesize is adjusted when a frag is added to one skb, or when
skb->head is re-allocated.

Are you sure you dont have another problem, because as I said commit
c8628155ece3  had a bug, so a bisect is not very useful.

How many reloads are needed to trigger the bug, do you have a script to
reproduce it ?

Could it be a PMTU problem ? (check
http://git.kernel.org/?p=linux/kernel/git/davem/net.git;a=commit;h=9b04f350057863d1fad1ba071e09362a1da3503e )