On 10/25, Roman Penyaev wrote:
>
> On Tue, Oct 25, 2016 at 2:56 PM, Oleg Nesterov <[email protected]> wrote:
> > On 10/25, Roman Pen wrote:
> >>
> >> struct task_struct *wq_worker_sleeping(struct task_struct *task)
> >> {
> >> - struct worker *worker = kthread_data(task), *to_wakeup = NULL;
> >> + struct worker *worker, *to_wakeup = NULL;
> >> struct worker_pool *pool;
> >>
> >> +
> >> + if (task->state == TASK_DEAD) {
> >> + /*
> >> + * Here we try to catch the following path before
> >> + * accessing NULL kthread->vfork_done ptr thru
> >> + * kthread_data():
> >> + *
> >> + * oops_end()
> >> + * do_exit()
> >> + * schedule()
> >> + *
> >> + * If panic_on_oops is not set and oops happens on
> >> + * a workqueue execution path, thread will be killed.
> >> + * That is definitly sad, but not to make the situation
> >> + * even worse we have to ignore dead tasks in order not
> >> + * to step on zeroed out members (e.g. t->vfork_done is
> >> + * already NULL on that path, since we were called by
> >> + * do_exit())).
> >> + */
> >> + return NULL;
> >> + }
> >
> > I still think that PF_EXITING check makes more sense than TASK_DEAD,
> > but I won't insist.
>
> Why? I probably do not see the corner cases, so, please, explain.
If nothing else the crashed worker can schedule() before do_task_dead(),
But mainly, to me PF_EXITING just looks better. TASK_DEAD is the very
special state, only sched/core.c should use it.
and... perhaps we can just add
void oops_end_exit(void)
{
current->flags &= ~PF_WQ_WORKER;
perhaps sonething else;
}
called by oops_end() before rewind_stack_do_exit() ?
Oleg.
