On 01/31/2017 01:31 AM, Bart Van Assche wrote:

On Wed, 2017-01-18 at 10:48 +0100, Hannes Reinecke wrote:

quoted

@@ -1488,26 +1487,13 @@ static unsigned long disk_events_poll_jiffies(struct gendisk *disk)
 void disk_block_events(struct gendisk *disk)
 {
        struct disk_events *ev = disk->ev;
-       unsigned long flags;
-       bool cancel;
 
        if (!ev)
                return;
 
-       /*
-        * Outer mutex ensures that the first blocker completes canceling
-        * the event work before further blockers are allowed to finish.
-        */
-       mutex_lock(&ev->block_mutex);
-
-       spin_lock_irqsave(&ev->lock, flags);
-       cancel = !ev->block++;
-       spin_unlock_irqrestore(&ev->lock, flags);
-
-       if (cancel)
+       if (atomic_inc_return(&ev->block) == 1)
                cancel_delayed_work_sync(&disk->ev->dwork);
 
-       mutex_unlock(&ev->block_mutex);
 }

Hello Hannes,

I have already encountered a few times a deadlock that was caused by the
event checking code so I agree with you that it would be a big step forward
if such deadlocks wouldn't occur anymore. However, this patch realizes a
change that has not been described in the patch description, namely that
disk_block_events() calls are no longer serialized. Are you sure it is safe
to drop the serialization of disk_block_events() calls?

Well, this whole synchronization stuff it a bit weird; I so totally fail
to see the rationale for it.
But anyway, once we've converted ev->block to atomics I _think_ the
mutex_lock can remain; will be checking.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		               zSeries & Storage
hare@suse.com			               +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 N�rnberg
GF: F. Imend�rffer, J. Smithard, D. Upmanyu, G. Norton
HRB 21284 (AG N�rnberg)