Re: [rfc] optimise unlock_page

On Wed, May 09, 2007 at 12:41:24AM +0200, Nick Piggin wrote:
On Wed, May 09, 2007 at 07:30:27AM +1000, Benjamin Herrenschmidt wrote:
On Tue, 2007-05-08 at 13:40 +0200, Nick Piggin wrote:
This patch trades a page flag for a significant improvement in the unlock_page
fastpath. Various problems in the previous version were spotted by Hugh and
Ben (and fixed in this one).

Comments?

--

Speed up unlock_page by introducing a new page flag to signal that there are
page waitqueue waiters for PG_locked. This means a memory barrier and a random
waitqueue hash cacheline load can be avoided in the fastpath when there is no
contention.

I'm not 100% familiar with the exclusive vs. non exclusive wait thingy
but wake_up_page() does __wake_up_bit() which calls __wake_up() with
nr_exclusive set to 1. Doesn't that mean that only one waiter will be
woken up ?

If that's the case, then we lose because we'll have clear PG_waiters but
only wake up one of them.

Waking them all would fix it but at the risk of causing other
problems... Maybe PG_waiters need to actually be a counter but if that
is the case, then it complicates things even more.

Any smart idea ?

It will wake up 1 exclusive waiter, but no limit on non exclusive waiters.
Hmm, but it won't wake up waiters behind the exclusive guy... maybe the
wake up code can check whether the waitqueue is still active after the
wakeup, and set PG_waiters again in that case?

Hm, I don't know if we can do that without a race either...

OTOH, waking all non exclusive waiters may not be a really bad idea.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/