rbd: define flags field, use it for exists flag
authorAlex Elder <elder@inktank.com>
Mon, 14 Jan 2013 18:43:31 +0000 (12:43 -0600)
committerSage Weil <sage@inktank.com>
Thu, 14 Feb 2013 02:29:10 +0000 (18:29 -0800)
Define a new rbd device flags field, manipulated using bit
operations.  Replace the use of the current "exists" flag with a bit
in this new "flags" field.  Add a little commentary about the
"exists" flag, which does not need to be manipulated atomically.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
drivers/block/rbd.c

index fd9656b5fdb97ba976e8881c6e8dfa6cee8a2b0e..8c90a39c2a91144704933975108fcdcb27207aa9 100644 (file)
@@ -264,7 +264,7 @@ struct rbd_device {
        spinlock_t              lock;           /* queue lock */
 
        struct rbd_image_header header;
-       atomic_t                exists;
+       unsigned long           flags;
        struct rbd_spec         *spec;
 
        char                    *header_name;
@@ -292,6 +292,12 @@ struct rbd_device {
        unsigned long           open_count;
 };
 
+/* Flag bits for rbd_dev->flags */
+
+enum rbd_dev_flags {
+       RBD_DEV_FLAG_EXISTS,    /* mapped snapshot has not been deleted */
+};
+
 static DEFINE_MUTEX(ctl_mutex);          /* Serialize open/close/setup/teardown */
 
 static LIST_HEAD(rbd_dev_list);    /* devices */
@@ -782,7 +788,8 @@ static int rbd_dev_set_mapping(struct rbd_device *rbd_dev)
                        goto done;
                rbd_dev->mapping.read_only = true;
        }
-       atomic_set(&rbd_dev->exists, 1);
+       set_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
+
 done:
        return ret;
 }
@@ -1877,9 +1884,14 @@ static void rbd_request_fn(struct request_queue *q)
                        rbd_assert(rbd_dev->spec->snap_id == CEPH_NOSNAP);
                }
 
-               /* Quit early if the snapshot has disappeared */
-
-               if (!atomic_read(&rbd_dev->exists)) {
+               /*
+                * Quit early if the mapped snapshot no longer
+                * exists.  It's still possible the snapshot will
+                * have disappeared by the time our request arrives
+                * at the osd, but there's no sense in sending it if
+                * we already know.
+                */
+               if (!test_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags)) {
                        dout("request for non-existent snapshot");
                        rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
                        result = -ENXIO;
@@ -2571,7 +2583,7 @@ struct rbd_device *rbd_dev_create(struct rbd_client *rbdc,
                return NULL;
 
        spin_lock_init(&rbd_dev->lock);
-       atomic_set(&rbd_dev->exists, 0);
+       rbd_dev->flags = 0;
        INIT_LIST_HEAD(&rbd_dev->node);
        INIT_LIST_HEAD(&rbd_dev->snaps);
        init_rwsem(&rbd_dev->header_rwsem);
@@ -3200,10 +3212,17 @@ static int rbd_dev_snaps_update(struct rbd_device *rbd_dev)
                if (snap_id == CEPH_NOSNAP || (snap && snap->id > snap_id)) {
                        struct list_head *next = links->next;
 
-                       /* Existing snapshot not in the new snap context */
-
+                       /*
+                        * A previously-existing snapshot is not in
+                        * the new snap context.
+                        *
+                        * If the now missing snapshot is the one the
+                        * image is mapped to, clear its exists flag
+                        * so we can avoid sending any more requests
+                        * to it.
+                        */
                        if (rbd_dev->spec->snap_id == snap->id)
-                               atomic_set(&rbd_dev->exists, 0);
+                               clear_bit(RBD_DEV_FLAG_EXISTS, &rbd_dev->flags);
                        rbd_remove_snap_dev(snap);
                        dout("%ssnap id %llu has been removed\n",
                                rbd_dev->spec->snap_id == snap->id ?