Re: Patch for lost wakeups

[Oleg Nesterov <oleg@redhat.com>]

On 08/09, Linus Torvalds wrote:
>
> I guess that instead of a "smp_wmb()", we could do another
> "smp_mb__before_spinlock()" thing, like we already allow for other
> architectures to do a weaker form of mb in case the spinlock is
> already a full mb. That would allow avoiding extra synchronization. Do
> a
>
>    #ifndef smp_mb__before_spinlock
>      #define smp_mb__before_spinlock() smp_wmb()
>    #endif
>
> in <linux/spinlock.h> to not force everybody to implement it. Because
> a wmb+acquire should be close enough to a full mb that nobody cares
> (ok, so reads could move into the critical region from outside, but by
> the time anybody has called "schedule()", I can't see it mattering, so
> "close enough").

Yes, this is what I tried to suggest. And of course we should turn that
wmb() in try_to_wake_up() into smp_mb__before_spinlock().

I event started the patch, but we already have smp_mb__after_lock(), so
it should be smp_mb__before_lock() for consistency and we need to turn
it to "define" too. Or change ARCH_HAS_SMP_MB_AFTER_LOCK, or add
ARCH_HAS_SMP_MB_BEFORE_LOCK.

Oleg.