rbd: only get snap context for write requests
authorAlex Elder <elder@inktank.com>
Fri, 9 Nov 2012 14:43:15 +0000 (08:43 -0600)
committerAlex Elder <elder@inktank.com>
Thu, 17 Jan 2013 21:52:00 +0000 (15:52 -0600)
Right now we get the snapshot context for an rbd image (under
protection of the header semaphore) for every request processed.

There's no need to get the snap context if we're doing a read,
so avoid doing so in that case.

Note that we no longer need to hold the header semaphore to
check the rbd_dev's existence flag.

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

index 2846536..0fbe9ad 100644 (file)
@@ -1331,7 +1331,7 @@ static int rbd_do_op(struct request *rq,
        } else {
                opcode = CEPH_OSD_OP_READ;
                flags = CEPH_OSD_FLAG_READ;
-               snapc = NULL;
+               rbd_assert(!snapc);
                snapid = rbd_dev->spec->snap_id;
                payload_len = 0;
        }
@@ -1662,22 +1662,25 @@ static void rbd_rq_fn(struct request_queue *q)
                }
                spin_unlock_irq(q->queue_lock);
 
-               /* Stop writes to a read-only device */
-
-               result = -EROFS;
-               if (read_only && rq_data_dir(rq) == WRITE)
-                       goto out_end_request;
-
-               /* Grab a reference to the snapshot context */
-
-               down_read(&rbd_dev->header_rwsem);
-               if (atomic_read(&rbd_dev->exists)) {
+               /* Write requests need a reference to the snapshot context */
+
+               if (rq_data_dir(rq) == WRITE) {
+                       result = -EROFS;
+                       if (read_only) /* Can't write to a read-only device */
+                               goto out_end_request;
+
+                       /*
+                        * Note that each osd request will take its
+                        * own reference to the snapshot context
+                        * supplied.  The reference we take here
+                        * just guarantees the one we provide stays
+                        * valid.
+                        */
+                       down_read(&rbd_dev->header_rwsem);
                        snapc = ceph_get_snap_context(rbd_dev->header.snapc);
+                       up_read(&rbd_dev->header_rwsem);
                        rbd_assert(snapc != NULL);
-               }
-               up_read(&rbd_dev->header_rwsem);
-
-               if (!snapc) {
+               } else if (!atomic_read(&rbd_dev->exists)) {
                        rbd_assert(rbd_dev->spec->snap_id != CEPH_NOSNAP);
                        dout("request for non-existent snapshot");
                        result = -ENXIO;
@@ -1689,7 +1692,8 @@ static void rbd_rq_fn(struct request_queue *q)
                                blk_rq_pos(rq) * SECTOR_SIZE,
                                size, rq->bio);
 out_end_request:
-               ceph_put_snap_context(snapc);
+               if (snapc)
+                       ceph_put_snap_context(snapc);
                spin_lock_irq(q->queue_lock);
                if (!size || result < 0)
                        __blk_end_request_all(rq, result);