On Fri, May 03, 2019 at 05:37:48PM +0200, Oleg Nesterov wrote:

(And if we change this code to use wait_event(xchg(readers_block) == 0) we
 can remove rw_sem altogether).

That patch you just saw and didn't look at did just that.

The main problem is that this is sub-optimal. We can have a lot of readers
sleeping in __down_read() when percpu_down_write() succeeds, then after
percpu_down_write_non_owner() does up_write() they all will be woken just
to hang in readers_block(). Plus the new readers will need to pass the
lock-check-unlock-schedule path.

Yes, that's gone. Still, write side locking on percpu-rwsem _should_ be
relatively rare and is certainly not a fast path.

Peter, just in case... I see another patch from you but I need to run away
till Monday.

n/p, enjoy the weekend!