Re: [PATCH] slow_work_thread() should do the exclusive wait

[Oleg Nesterov]

(add Ingo)

On 04/23, David Howells wrote:
>
> Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> > I wonder if slow_work_cull_timeout() should have some sort of barrier,
> > so the write is suitably visible to the woken thread.  Bearing in mind
> > that the thread might _already_ have been woken by someone else?
>
> Perhaps the attached patch?
>
> David
> ---
> From: David Howells <dhowells@xxxxxxxxxx>
> Subject: [PATCH] slow_work_cull_timeout() should have a memory barrier
>
> slow_work_cull_timeout() should have a write memory barrier so that the setting
> of the cull flag is seen before the wakeup takes place.  This is required
> because wake_up() does not guarantee any memory barriership at all.
>
> Concomitant to this, slow_work_thread() should have a read memory barrier
> between its return from schedule() and its testing of slow_work_cull() as
> finish_wait() isn't a guaranteed barrier either.
>
> Signed-off-by: David Howells <dhowells@xxxxxxxxxx>
> ---
>
>  kernel/slow-work.c |    2 ++
>  1 files changed, 2 insertions(+), 0 deletions(-)
>
>
> diff --git a/kernel/slow-work.c b/kernel/slow-work.c
> index 521ed20..96e418d 100644
> --- a/kernel/slow-work.c
> +++ b/kernel/slow-work.c
> @@ -382,6 +382,7 @@ static int slow_work_thread(void *_data)
>  		finish_wait(&slow_work_thread_wq, &wait);
>
>  		try_to_freeze();
> +		smp_rmb();
>
>  		vsmax = vslow_work_proportion;
>  		vsmax *= atomic_read(&slow_work_thread_count);
> @@ -416,6 +417,7 @@ static int slow_work_thread(void *_data)
>  static void slow_work_cull_timeout(unsigned long data)
>  {
>  	slow_work_cull = true;
> +	smp_wmb();
>  	wake_up(&slow_work_thread_wq);
>  }

Confused. If we need this barrier, a lot of similar code is broken.

	slow_work_cull_timeout:

		slow_work_cull = true;
		wake_up(&slow_work_thread_wq);


	slow_work_thread:

		prepare_to_wait(&slow_work_thread_wq);
		if (!slow_work_cull)
			schedule();
		finish_wait(&slow_work_thread_wq);

		if (slow_work_cull)
			.....

Both wake_up() and prepare_to_wait() take the same wait_queue_head_t->lock,
and prepare_to_wait() does set_current_state() under this lock.

How can we miss the event? If wake_up() happens before prepare_to_wait(),
slow_work_thread() must see slow_work_cull = T, otherwise the subsequent
wake_up() must see the result of list_add() + set_current_state() and
wake up the sleeping thread.

Could you please clarify?

Oleg.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/