On Fri, 23 Aug 2024 at 11:05, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> Looking around, we do seem to have a pattern of
>
> smp_store_release() -> wake_up_var()
>
> instead of a memory barrier. I don't think that actually works.
Hmm. It might not work for the wakeup race, but it might be a good
idea to do the store_release anyway, just for the actual user (ie then
the *use* of the variable may start with a "smp_load_acquire()", and
the release->acquire semantics means that everything that was done
before the release is visible after the acquire.
Of course, the smp_mb() will force that ordering too, but for a true
concurrent user that doesn't actually need waking up, it might see the
newly stores var value before the smp_mb() happens on the generating
side.
End result: those code paths may want *both* the smp_store_release()
and the smp_mb(), because they are ordering different accesses.
Just goes to show that this whole thing is more complicated than just
"wait for a value".
Linus
