Re: [PATCH v3 1/1] workqueue: ignore dead tasks in a workqueue sleep hook

[Roman Penyaev]

On Tue, Oct 25, 2016 at 2:56 PM, Oleg Nesterov <oleg@xxxxxxxxxx> 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.

--
Roman