Re: [PATCH] slow_work_execute() needs mb() before test_bit(SLOW_WORK_PENDING)

[Oleg Nesterov <oleg@redhat.com>]

On 04/13, David Howells wrote:
>
> Oleg Nesterov <oleg@redhat.com> wrote:
>
> > clear_bit_unlock() implies release semantics, iow we have a one-way barrier
> > before clear_bit().
>
> That is correct.  The stuff before the clear_bit_unlock() must not leak out
> past it.  SLOW_WORK_EXECUTING is a lock around the execution of the work item.
>
> > But we need the mb() semantics after clear_bit(), before we test
> > SLOW_WORK_PENDING. Otherwise we can miss SLOW_WORK_ENQ_DEFERRED if we race
> > slow_work_enqueue().
>
> But if you move the mb() to the other side, and reduce to clear_bit() you
> break the above non-reentrancy guarantee.

Hmm. Can't understand... please see below.

> > However, given that both clear_bit() and set_bit() use the same word,
> > perhaps this is not possible.
>
> I would assume clear_bit() and set_bit() on different bits of the same word
> must interact as if they're on the same bit, with regard to atomiticity, but I
> don't know for certain that it is guaranteed.

Yes, this was my concern. Reading the docs, I can't see how it is possible
to "prove" that clear_bit(x, WORD) + test_bit(y, WORD) can't be "reversed".

> > But in that case I don't understand why do we need clear_bit_unlock(), not
> > just clear_bit(), and how "mb is not needeed" could be derived from
> > documentation.
>
> As mentioned above, SLOW_WORK_EXECUTING is a lock against multiple entrance to
> the execution function.

Yes I see.

But why do we need the barrier before clear_bit(SLOW_WORK_EXECUTING) ?
We do have one after test_and_clear_bit(SLOW_WORK_PENDING) above, and it
should be enough, no?

Oleg.