Re: [PATCH 4/4] drm/ttm: Optimize reservation slightly

[Maarten Lankhorst <maarten.lankhorst@canonical.com>]

Hey,

Op 05-11-12 14:31, Thomas Hellstrom schreef:
> Reservation locking currently always takes place under the LRU spinlock.
> Hence, strictly there is no need for an atomic_cmpxchg call; we can use
> atomic_read followed by atomic_write since nobody else will ever reserve
> without the lru spinlock held.
> At least on Intel this should remove a locked bus cycle on successful
> reserve.
>
> Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
>
Is that really a good thing to submit when I am busy killing lru lock around reserve? :-)

-	while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
+	while (unlikely(atomic_xchg(&bo->reserved, 1) != 0)) {

Works without lru lock too!

In fact mutexes are done in a similar way[1], except with some more magic, and unlocked state is 1, not 0.
However I do think that to get that right (saves a irq disable in unlock path, and less wakeups in contended
case), I should really just post the mutex extension patches for reservations and ride the flames. It's
getting too close to real mutexes so I really want it to be a mutex in that case. So lets convert it.. Soon! :-)

~Maarten

[1] See linux/include/asm-generic/mutex-xchg.h and linux/include/asm-generic/mutex-dec.h for how
archs generally implement mutex fastpaths.