Re: task_stat splat

[Oleg Nesterov <oleg@redhat.com>]

Damn, sorry for noise ;)

On 11/23, Oleg Nesterov wrote:
>
> On 11/23, Oleg Nesterov wrote:
> >
> > On 11/23, Oleg Nesterov wrote:
> > >
> > > On 11/23, Borislav Petkov wrote:
> > > >
> > > > where we end up with a zero PMD. RIP is corrupted too so we're somewhere
> > > > off in the fields.
> > >
> > > PMD = 0 is fine I guess, addr == 0 is not mapped.
> > >
> > > > Comment over thread_group_cputime() talks about dead tasks accounting
> > >
> > > This comment simply means that we also need to read the accumulated
> > > counters in tsk->signal.
> > >
> > > > which might be relevant as we're seeing not mapped page hierarchy so
> > > > something must have gone away recently but we try to look at it.
> > >
> > > This is called under ->siglock, we can't race with exit/etc. But this
> > > doesn't matter, it is not that we (say) get t == NULL or something like
> > > this.
> > >
> > > RIP == 0, and this looks "impossible", I do not see indirect function
> > > calls in this paths.
> >
> > Ah, I didn't notice you mentioned tip/master... so it looks as if
> > sched_class->update_curr is NULL?
>
> Perhaps this is migration thread? stop_sched_class doesn't have ->update_curr.

Yes, I think this can explain the problem, but

> could you try to cat /proc/pid-of-migration-thread/stat on your machine?

This won't trigger the crash unless it is running.

Oleg.