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

[Jeff Layton]

On Tue, 2020-03-10 at 08:42 +1100, NeilBrown wrote:
> On Mon, Mar 09 2020, Jeff Layton wrote:
> 
> > On Mon, 2020-03-09 at 13:22 -0400, Jeff Layton wrote:
> > > On Mon, 2020-03-09 at 08:52 -0700, Linus Torvalds wrote:
> > > > On Mon, Mar 9, 2020 at 7:36 AM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > > > > On Sun, 2020-03-08 at 22:03 +0800, kernel test robot wrote:
> > > > > > FYI, we noticed a -96.6% regression of will-it-scale.per_process_ops due to commit:
> > > > > 
> > > > > This is not completely unexpected as we're banging on the global
> > > > > blocked_lock_lock now for every unlock. This test just thrashes file
> > > > > locks and unlocks without doing anything in between, so the workload
> > > > > looks pretty artificial [1].
> > > > > 
> > > > > It would be nice to avoid the global lock in this codepath, but it
> > > > > doesn't look simple to do. I'll keep thinking about it, but for now I'm
> > > > > inclined to ignore this result unless we see a problem in more realistic
> > > > > workloads.
> > > > 
> > > > That is a _huge_ regression, though.
> > > > 
> > > > What about something like the attached? Wouldn't that work? And make
> > > > the code actually match the old comment about wow "fl_blocker" being
> > > > NULL being special.
> > > > 
> > > > The old code seemed to not know about things like memory ordering either.
> > > > 
> > > > Patch is entirely untested, but aims to have that "smp_store_release()
> > > > means I'm done and not going to touch it any more", making that
> > > > smp_load_acquire() test hopefully be valid as per the comment..
> > > 
> > > Yeah, something along those lines maybe. I don't think we can use
> > > fl_blocker that way though, as the wait_event_interruptible is waiting
> > > on it to go to NULL, and the wake_up happens before fl_blocker is
> > > cleared.
> > > 
> > > Maybe we need to mix in some sort of FL_BLOCK_ACTIVE flag and use that
> > > instead of testing for !fl_blocker to see whether we can avoid the
> > > blocked_lock_lock?
> > >   
> > 
> > How about something like this instead? (untested other than for
> > compilation)
> > 
> > Basically, this just switches the waiters over to wait for
> > fl_blocked_member to go empty. That still happens before the wakeup, so
> > it should be ok to wait on that.
> > 
> > I think we can also eliminate the lockless list_empty check in
> > locks_delete_block, as the fl_blocker check should be sufficient now.
> > -- 
> > Jeff Layton <jlayton@xxxxxxxxxx>
> > From c179d779c9b72838ed9996a65d686d86679d1639 Mon Sep 17 00:00:00 2001
> > From: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> > Date: Mon, 9 Mar 2020 14:35:43 -0400
> > Subject: [PATCH] locks: reinstate locks_delete_lock optimization
> > 
> > ...by using smp_load_acquire and smp_store_release to close the race
> > window.
> > 
> > [ jlayton: wait on the fl_blocked_requests list to go empty instead of
> > 	   the fl_blocker pointer to clear. Remove the list_empty check
> > 	   from locks_delete_lock shortcut. ]
> 
> Why do you think it is OK to remove that list_empty check?  I don't
> think it is.  There might be locked requests that need to be woken up.
> 
> As the problem here is a use-after-free due to a race, one option would
> be to use rcu_free() on the file_lock, and hold rcu_read_lock() around
> test/use.
> 
> Another option is to use a different lock.  The fl_wait contains a
> spinlock, and we have wake_up_locked() which is provided for exactly
> these sorts of situations where the wake_up call can race with a thread
> waking up.
> 
> So my compile-tested-only proposal is below.
> I can probably a proper change-log entry if you think the patch is a
> good way to go.
> 
> NeilBrown
> 
> 
> diff --git a/fs/locks.c b/fs/locks.c
> index 426b55d333d5..8aa04d5ac8b3 100644
> --- a/fs/locks.c
> +++ b/fs/locks.c
> @@ -735,11 +735,13 @@ static void __locks_wake_up_blocks(struct file_lock *blocker)
>  
>  		waiter = list_first_entry(&blocker->fl_blocked_requests,
>  					  struct file_lock, fl_blocked_member);
> +		spin_lock(&waiter->fl_wait.lock);
>  		__locks_delete_block(waiter);
>  		if (waiter->fl_lmops && waiter->fl_lmops->lm_notify)
>  			waiter->fl_lmops->lm_notify(waiter);
>  		else
> -			wake_up(&waiter->fl_wait);
> +			wake_up_locked(&waiter->fl_wait);
> +		spin_unlock(&waiter->fl_wait.lock);
>  	}
>  }
>  
> @@ -753,6 +755,31 @@ int locks_delete_block(struct file_lock *waiter)
>  {
>  	int status = -ENOENT;
>  
> +	/*
> +	 * If fl_blocker is NULL, it won't be set again as this thread
> +	 * "owns" the lock and is the only one that might try to claim
> +	 * the lock.  So it is safe to test fl_blocker locklessly.
> +	 * Also if fl_blocker is NULL, this waiter is not listed on
> +	 * fl_blocked_requests for some lock, so no other request can
> +	 * be added to the list of fl_blocked_requests for this
> +	 * request.  So if fl_blocker is NULL, it is safe to
> +	 * locklessly check if fl_blocked_requests is empty.  If both
> +	 * of these checks succeed, there is no need to take the lock.
> +	 * However, some other thread might have only *just* set
> +	 * fl_blocker to NULL and it about to send a wakeup on
> +	 * fl_wait, so we mustn't return too soon or we might free waiter
> +	 * before that wakeup can be sent.  So take the fl_wait.lock
> +	 * to serialize with the wakeup in __locks_wake_up_blocks().
> +	 */
> +	if (waiter->fl_blocker == NULL) {
> +		spin_lock(&waiter->fl_wait.lock);
> +		if (waiter->fl_blocker == NULL &&
> +		    list_empty(&waiter->fl_blocked_requests)) {
> +			spin_unlock(&waiter->fl_wait.lock);
> +			return status;
> +		}
> +		spin_unlock(&waiter->fl_wait.lock);
> +	}
>  	spin_lock(&blocked_lock_lock);
>  	if (waiter->fl_blocker)
>  		status = 0;
> 

Looks good on a cursory check, and I'm inclined to go with this since
it's less fiddly for people to backport.

One other difference to note -- we are holding the fl_wait lock when
calling lm_notify, but I don't think it will matter to any of the
existing lm_notify functions.

If you want to clean up the changelog and resend that would be great.

Thanks,
-- 
Jeff Layton <jlayton@xxxxxxxxxx>