On 02/11/2015 09:24 AM, Oleg Nesterov wrote:
> I agree, and I have to admit I am not sure I fully understand why
> unlock uses the locked add. Except we need a barrier to avoid the race
> with the enter_slowpath() users, of course. Perhaps this is the only
> reason?
Right now it needs to be a locked operation to prevent read-reordering.
x86 memory ordering rules state that all writes are seen in a globally
consistent order, and are globally ordered wrt reads *on the same
addresses*, but reads to different addresses can be reordered wrt to writes.
So, if the unlocking add were not a locked operation:
__add(&lock->tickets.head, TICKET_LOCK_INC); /* not locked */
if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
__ticket_unlock_slowpath(lock, prev);
Then the read of lock->tickets.tail can be reordered before the unlock,
which introduces a race:
/* read reordered here */
if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG)) /* false */
/* ... */;
/* other CPU sets SLOWPATH and blocks */
__add(&lock->tickets.head, TICKET_LOCK_INC); /* not locked */
/* other CPU hung */
So it doesn't *have* to be a locked operation. This should also work:
__add(&lock->tickets.head, TICKET_LOCK_INC); /* not locked */
lfence(); /* prevent read
reordering */
if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
__ticket_unlock_slowpath(lock, prev);
but in practice a locked add is cheaper than an lfence (or at least was).
This *might* be OK, but I think it's on dubious ground:
__add(&lock->tickets.head, TICKET_LOCK_INC); /* not locked */
/* read overlaps write, and so is ordered */
if (unlikely(lock->head_tail & (TICKET_SLOWPATH_FLAG << TICKET_SHIFT))
__ticket_unlock_slowpath(lock, prev);
because I think Intel and AMD differed in interpretation about how
overlapping but different-sized reads & writes are ordered (or it simply
isn't architecturally defined).
If the slowpath flag is moved to head, then it would always have to be
locked anyway, because it needs to be atomic against other CPU's RMW
operations setting the flag.
J
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
