Re: [locks] 6d390e4b5d: will-it-scale.per_process_ops -96.6% regression

[Linus Torvalds]

On Tue, Mar 10, 2020 at 3:07 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
>
> Given that, and the fact that Neil pointed out that yangerkun's latest
> patch would reintroduce the original race, I'm leaning back toward the
> patch Neil sent yesterday. It relies solely on spinlocks, and so doesn't
> have the subtle memory-ordering requirements of the others.

It has subtle locking changes, though.

It now calls the "->lm_notify()" callback with the wait queue spinlock held.

is that ok? It's not obvious. Those functions take other spinlocks,
and wake up other things. See for example nlmsvc_notify_blocked()..
Yes, it was called under the blocked_lock_lock spinlock before too,
but now there's an _additional_ spinlock, and it must not call
"wake_up(&waiter->fl_wait))" in the callback, for example, because it
already holds the lock on that wait queue.

Maybe that is never done. I don't know the callbacks.

I was really hoping that the simple memory ordering of using that
smp_store_release -> smp_load_acquire using fl_blocker would be
sufficient. That's a particularly simple and efficient ordering.

Oh well. If you want to go that spinlock way, it needs to document why
it's safe to do a callback under it.

                  Linus