Re: Null pointer crash at find_idlest_group on db845c w/ linus/master

[Vincent Guittot]

On Wed, 4 Dec 2019 at 11:41, Valentin Schneider
<valentin.schneider@xxxxxxx> wrote:
>
> On 04/12/2019 10:09, Vincent Guittot wrote:
> > Now, we test that a group has at least one allowed CPU for the task so we
> > could skip the local group with the correct/wrong p->cpus_ptr
> >
> > The path is used for fork/exec ibut also for wakeup path for b.L when the task doesn't fit in the CPUs
> >
> > So we can probably imagine a scenario where we change task affinity while
> > sleeping. If the wakeup happens on a CPU that belongs to the group that is not
> > allowed, we can imagine that we skip the local_group
> >
>
> Shoot, I think you're right. If it is the local group that is NULL, then
> we most likely splat on:
>
>                 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
>                         return NULL;
>
> We don't splat before because we just use local_sgs, which is uninitialized
> but on the stack.
>
> Also; does it really have to involve an affinity "race"? AFAIU affinity
> could have been changed a while back, but the waking CPU isn't allowed
> so we skip the local_group (in simpler cases where each CPU is a group).

In fact, this will depend of the uninitialized values of local_sgs. I
have been able to reproduce the situation where we skip local group
but not to trigger the crash because the values already in the stack
don't trigger the misfit comparison.

I  wait for John feedback to confirm that this fix his problem and
will send a clean version of the patch
>
>