On Fri, 26 Jun 2009, Oleg Nesterov wrote:

And if we remove waitqueue_active() in xxx_update(), then lock/unlock is
not needed too.

If xxx_poll() takes q->lock first, it can safely miss the changes in ->status
and schedule(): xxx_update() will take q->lock, notice the sleeper and wake
it up (ok, it will set ->triggered but this doesn't matter).

If xxx_update() takes q->lock first, xxx_poll() must see the changes in
status after poll_wait()->unlock(&q->lock) (in fact, after lock, not unlock).

Sure. The snippet above was just to show what typically the code does, not 
a suggestion on how to solve the socket case.
But yeah, the problem in this case is the waitqueue_active() call. Without 
that, the wait queue lock/unlock in poll_wait() and the one in wake_up() 
guarantees the necessary barriers.
Some might argue the costs of the lock/unlock of q->lock, and wonder if 
MBs are a more efficient solution. This is something I'm not going into.
To me, it just looked not right having cross-matching MB in different 
subsystems.



- Davide