Re: [PATCH v2] make buffer_locked provide an acquire semantics

[Linus Torvalds]

On Sun, Jul 31, 2022 at 1:39 PM Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote:
>
> Do you think that wait_event also needs a read memory barrier?

Not really, no.

__wait_event() uses prepare_to_wait(), and it uses set_current_state()
very much so that the process state setting is serialized with the
test afterwards.

And the only race wait_event should worry about is exactly the "go to
sleep" vs "make the event true and then wake up" race, so that it
doesn't wait forever.

There is no guarantee of memory ordering wrt anything else, and I
don't think there is any need for such a guarantee.

If somebody wants that guarantee, they should probably make the
condition for wait_event() to be a "load_acquire()" or similar. Those
cases do exist.

                       Linus