block: fix race between switching elevator and removing queues
authorMing Lei <ming.lei@redhat.com>
Thu, 12 Sep 2019 04:02:24 +0000 (12:02 +0800)
committerJens Axboe <axboe@kernel.dk>
Thu, 12 Sep 2019 13:13:22 +0000 (07:13 -0600)
cecf5d87ff20 ("block: split .sysfs_lock into two locks") starts to
release & actuire sysfs_lock again during switching elevator. So it
isn't enough to prevent switching elevator from happening by simply
clearing QUEUE_FLAG_REGISTERED with holding sysfs_lock, because
in-progress switch still can move on after re-acquiring the lock,
meantime the flag of QUEUE_FLAG_REGISTERED won't get checked.

Fixes this issue by checking 'q->elevator' directly & locklessly after
q->kobj is removed in blk_unregister_queue(), this way is safe because
q->elevator can't be changed at that time.

Fixes: cecf5d87ff20 ("block: split .sysfs_lock into two locks")
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-sysfs.c

index 1075134..3af7983 100644 (file)
@@ -1030,7 +1030,6 @@ EXPORT_SYMBOL_GPL(blk_register_queue);
 void blk_unregister_queue(struct gendisk *disk)
 {
        struct request_queue *q = disk->queue;
-       bool has_elevator;
 
        if (WARN_ON(!q))
                return;
@@ -1046,7 +1045,6 @@ void blk_unregister_queue(struct gendisk *disk)
         */
        mutex_lock(&q->sysfs_lock);
        blk_queue_flag_clear(QUEUE_FLAG_REGISTERED, q);
-       has_elevator = !!q->elevator;
        mutex_unlock(&q->sysfs_lock);
 
        mutex_lock(&q->sysfs_dir_lock);
@@ -1061,7 +1059,11 @@ void blk_unregister_queue(struct gendisk *disk)
        kobject_del(&q->kobj);
        blk_trace_remove_sysfs(disk_to_dev(disk));
 
-       if (has_elevator)
+       /*
+        * q->kobj has been removed, so it is safe to check if elevator
+        * exists without holding q->sysfs_lock.
+        */
+       if (q->elevator)
                elv_unregister_queue(q);
        mutex_unlock(&q->sysfs_dir_lock);