rbd: don't hold spinlock during messenger flush

A recent change made changes to the rbd_client_list be protected by
a spinlock.  Unfortunately in rbd_put_client(), the lock is taken
before possibly dropping the last reference to an rbd_client, and on
the last reference that eventually calls flush_workqueue() which can
sleep.

The problem was flagged by a debug spinlock warning:
    BUG: spinlock wrong CPU on CPU#3, rbd/27814

The solution is to move the spinlock acquisition and release inside
rbd_client_release(), which is the spot where it's really needed for
protecting the removal of the rbd_client from the client list.

Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>

drivers/block/rbd.c

@@ -450,7 +450,9 @@ static void rbd_client_release(struct kref *kref)
 	struct rbd_client *rbdc = container_of(kref, struct rbd_client, kref);
 
 	dout("rbd_release_client %p\n", rbdc);
+	spin_lock(&rbd_client_list_lock);
 	list_del(&rbdc->node);
+	spin_unlock(&rbd_client_list_lock);
 
 	ceph_destroy_client(rbdc->client);
 	kfree(rbdc->rbd_opts);
@@ -463,9 +465,7 @@ static void rbd_client_release(struct kref *kref)
  */
 static void rbd_put_client(struct rbd_device *rbd_dev)
 {
-	spin_lock(&rbd_client_list_lock);
 	kref_put(&rbd_dev->rbd_client->kref, rbd_client_release);
-	spin_unlock(&rbd_client_list_lock);
 	rbd_dev->rbd_client = NULL;
 }