Re: [rfc] optimise unlock_page

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

> Not good enough, I'm afraid.  It looks like Ben's right and you need
> a count - and counts in the page struct are a lot harder to add than
> page flags.
> 
> I've now played around with the hangs on my three 4CPU machines
> (all of them in io_schedule below __lock_page, waiting on pages
> which were neither PG_locked nor PG_waiters when I looked).
> 
> Seeing Ben's mail, I thought the answer would be just to remove
> the "_exclusive" from your three prepare_to_wait_exclusive()s.
> That helped, but it didn't eliminate the hangs.

There might be a way ... by having the flags manipulation always
atomically deal with PG_locked and PG_waiters together. This is possible
but we would need even more weirdo bitops abstractions from the arch I'm
afraid... unless we start using atomic_* rather that bitops in order to
manipulate multiple bits at a time.

Ben.