Re: [PATCH] for account_group_exec_runtime(), make sure ->signal can't be freed under rq->lock

[Oleg Nesterov <oleg@redhat.com>]

On 11/11, Ingo Molnar wrote:
>
> * Oleg Nesterov <oleg@redhat.com> wrote:
>
> > The patch is ugly, but I don't see the better fix for now. Needs the
> > review from Peter/Ingo.
>
> this is indeed too ugly,

Agreed. It was "unless we find another fix for 2.6.28".

> Regarding this teardown bug. Stupid question: why cannot the signal
> structure live as long as the last user is around? It's a tiny amount
> of RAM.

Well, release_task()->__exit_signal() clears/frees ->signal exactly
because it doesn't (must not) have users any longer. And we have the
code which checks ->signal != NULL to know if the task was already
released or not.

Now scheduler wants to play with ->signal. We can change the code so
that we don't actually free it until the task does the last schedule.
Say, we can free it __from put_task_struct(). But this means we need
another counter in signal_struct (signal_struct->count can't work).
And, until we change the code which checks ->signal != NULL, we need
another pointer in task_struct.


Perhaps this makes sense regardless of this bug, but I don't think
this is 2.6.28 material anyway.

Oleg.