On Tue, 31 Dec 2024 at 03:14, Manfred Spraul <manfred@xxxxxxxxxxxxxxxx> wrote:
>
> Should we add the missing memory barriers and switch to
> wait_queue_active() in front of all wakeup calls?
If we *really* want to optimize this, we could even get rid of the
memory barrier at least on x86, because
(a) mutex_unlock() is a full memory barrier on x86 (it involves a
locked cmpxchg)
(b) the condition is always set inside the locked region
(c) the wakeup is after releasing the lock
but this is architecture-specific (ie "mutex_unlock()" is not
*guaranteed* to be a memory barrier (ie on other architectures it
might be only a release barrier).
We have "smp_mb__after_atomic()" and "smp_mb__after_spinlock()", but
we don't have a "smp_mb__after_mutex_unlock()".
So we'd have to add a new helper or config option.
Anyway, I'm perfectly happy to get these optimizations, but because of
historical trouble in this area, I want any patches to be very clearly
documented.
Oleg's patch to only wake up writers when readers have actually opened
up a slot may not make any actual difference (because readers in
*practice* always do big reads), but I like it because it feels
obviously correct and doesn't have any locking or memory ordering
subtleties (and actually makes the logic more logical and
straightforward).
Linus
