Re: Kernel oops when clearing bgp neighbor info with TCP MD5SUM enabled

[Oleg Nesterov <oleg@redhat.com>]

Hi Anirban,

On 10/18, Anirban Sinha wrote:
>
> I have a question for you. The queue_work() routine which is called from
> schedule_work() does a put_cpu() which in turn does a enable_preempt(). Is
> this an attempt to trigger the scheduler?

No. please note that queue_work() does get_cpu() + put_cpu() to protect
against cpu_down() in between.

This can trigger the scheduler of course, but everything should be OK.

> One of the side affects of
> this enable_preempt() is the crash that we see below. What is happening
> is that  a timer callback routine, in  this case inet_twdr_hangman(),
> tries a bunch of cleanup until a threshold is reached. If further cleanups
> needs to be done beyond the threshold, it queues a work function. Now when
> the timer callback is run in __run_timers(), the routine grabs the value
> of preempt_count before and after the callback function call. If the two
> counts do not match, it calls BUG() (line 1037 in kernel/timer.c).

Yes, but I can't see how queue_work() can be involved, it doesn't change
->preempt_count. Note again it does put after get.

> Is is
> it illegal to schedule a work function from within a timer callback?

Yes sure.

I'd suppose that this unbalance comes from inet_twdr_hangman() pathes.

Could you verify this?

Oleg.