Le samedi 02 juin 2012 12:56:45, vous avez écrit :

And partly because the patch I sent included its content in the commit
message as well. :o/

I noticed the repetition after trying to apply on 3.4, and dropped one. And 
only then realised it was really already applied.

If the inlined patch makes a difference, you should see it with 3.4.

It made a difference, when testing with netcat: without any change over 
vanilla 3.3.7, network trafic drops to 0 in a matter of seconds (up to around 
10s). With it, it stayed stable for 10 minutes, until I killed nc.
I reproduced this with 3.4 as well (no patch = bug, patch = no problem).
In both version without patch, I got the watchdog warning 10 minutes after 
traffic drop - though without the IO_PAGE_FAULT message.

I spent quite some time testing with nc in UDP mode first, and couldn't 
reproduce the issue (then I switched to TCP as said above). Does that make any 
sense ?
I also noticed the significant lag at bootup when eth0 is brought up is much 
reduced on patched kernel. Does that makes sense ?

FWIW, the commands I used were based on:
  nc -l -p 5555 < /dev/zero > /dev/null

With/without -u flag, and of course client-side equivalent command so the 
connection was used full-duplex at maximum speed: 450Mb/s in TCP, 800+Mb/s in 
UDP, each way. UDP was limited by CPU on one side (~450Mb/s upload from 
that box, 800Mb/s download, 100% cpu on it).
Values are as reported by nload & htop.
All tests were done in runlevel 2, with rsyslog manually started with its init 
script.

My life is a bit easier when you work somewhere in the main branch
(or in davem's -next but it is not relevant for regression fixes).

I'm not sure: does 3.4 tarball from kernel.org qualify as "main branch" ? 
Otherwise, which git repos & branch should I use ?

Regards,
-- 
Vincent Pelletier