Re: [PATCH v4 2/6] perf: Enqueue SIGTRAP always via task_work.

[Frederic Weisbecker <frederic@kernel.org>]

+Cc Oleg.

Le Thu, Nov 07, 2024 at 03:46:17PM +0100, Sebastian Andrzej Siewior a écrit :
> On 2024-10-30 16:46:22 [+0100], Frederic Weisbecker wrote:
> > This needs more thoughts. We must make sure that the parent is put _after_
> > the child because it's dereferenced on release, for example:
> …
> > put_event()
> >    free_event()
> >       irq_work_sync(&event->pending_irq);
> >       ====> IRQ or irq_workd
> >       perf_event_wakeup()
> >          ring_buffer_wakeup()
> > 	    event = event->parent;
> > 	    rcu_dereference(event->rb);
> > 
> > And now after this patch it's possible that this happens after
> > the parent has been released.
> > 
> > We could put the parent from the child's free_event() but some
> > places (inherit_event()) may call free_event() on a child without
> > having held a reference to the parent.
> > 
> > Also note that with this patch the task may receive late irrelevant
> > signals after the event is removed. It's probably not that bad but
> > still... This could be a concern for exec(), is there a missing
> > task_work_run() there before flush_signal_handlers()?
> 
> So if this causes so much pain, what about taking only one item at a
> item? The following passes the test, too:
> 
> diff --git a/kernel/task_work.c b/kernel/task_work.c
> index c969f1f26be58..fc796ffddfc74 100644
> --- a/kernel/task_work.c
> +++ b/kernel/task_work.c
> @@ -206,7 +206,7 @@ bool task_work_cancel(struct task_struct *task, struct callback_head *cb)
>  void task_work_run(void)
>  {
>  	struct task_struct *task = current;
> -	struct callback_head *work, *head, *next;
> +	struct callback_head *work, *head;
>  
>  	for (;;) {
>  		/*
> @@ -214,17 +214,7 @@ void task_work_run(void)
>  		 * work_exited unless the list is empty.
>  		 */
>  		work = READ_ONCE(task->task_works);
> -		do {
> -			head = NULL;
> -			if (!work) {
> -				if (task->flags & PF_EXITING)
> -					head = &work_exited;
> -				else
> -					break;
> -			}
> -		} while (!try_cmpxchg(&task->task_works, &work, head));
> -
> -		if (!work)
> +		if (!work && !(task->flags & PF_EXITING))
>  			break;
>  		/*
>  		 * Synchronize with task_work_cancel_match(). It can not remove
> @@ -232,13 +222,24 @@ void task_work_run(void)
>  		 * But it can remove another entry from the ->next list.
>  		 */
>  		raw_spin_lock_irq(&task->pi_lock);
> +		do {
> +			head = NULL;
> +			if (work) {
> +				head = READ_ONCE(work->next);
> +			} else {
> +				if (task->flags & PF_EXITING)
> +					head = &work_exited;
> +				else
> +					break;
> +			}
> +		} while (!try_cmpxchg(&task->task_works, &work, head));
>  		raw_spin_unlock_irq(&task->pi_lock);

And having more than one task work should be sufficiently rare
that we don't care about doing the locking + cmpxchg() for each
of them pending.

I like it!

Thanks.

>  
> -		do {
> -			next = work->next;
> -			work->func(work);
> -			work = next;
> +		if (!work)
> +			break;
> +		work->func(work);
> +
> +		if (head)
>  			cond_resched();
> -		} while (work);
>  	}
>  }
> 
> Sebastian