Re: em(4) watchdog timeout

Jack, others,

On Tue, Jul 25, 2006 at 11:19:22AM -0700, Jack Vogel wrote:
> On 7/25/06, Jeremie Le Hen <jeremie_at_le-hen.org> wrote:
> >I am rebuilding a fresher one right now.  According to Ian's post,
> >the problem is likely to remain.  What do you advice me to do to
> >track this down ?
> >
> >FYI, my -CURRENT kernel (as well userland) is patched with ProPolice.
> >I don't think this can lead to this kind of problem, the overhead is
> >really small, in an order of 3 percent.  Do you think such an
> >overhead in a time-critial path could trigger a watchdog timeout ?
> 
> Well, watchdogs ARE about timeouts :)
> 
> I know nothing about ProPolice but would suggest removing for a test
> and see if the problem goes away.

As a matter of fact ProPolice protects stack-based buffer overflows
by pushing an additional value in the stack during vulnerable functions'
prologue and checking it within the epilogue.  This is really a matter
of a few CPU cycles.

Nevertheless, I will try to reproduce the problem as-is and will also
recompile my kernel without ProPolice ASAP.


> As for the debug_info and stats data you sent, there was nothing that
> looked bad in it, but those are more of a dynamic tool, its when the
> problem occurs that this will possibly show why.... I know, it doesnt
> make it easy :(

It would be a pain to do it manually, I will try to write a small
script thates watch over dmesg and issues a few sysctl on debug_info
whenever it detects a watch dog timeout.

Thank you.
Regards,
-- 
Jeremie Le Hen
< jeremie at le-hen dot org >< ttz at chchile dot org >