diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index 40e845fb97170..92b490aac93ee 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -1706,6 +1706,11 @@ static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
 	} else
 		r10_bio->master_bio = (struct bio *)first_r10bio;
 
+	/*
+	 * first select target devices under rcu_lock and
+	 * inc refcount on their rdev.  Record them by setting
+	 * bios[x] to bio
+	 */
 	rcu_read_lock();
 	for (disk = 0; disk < geo->raid_disks; disk++) {
 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
@@ -1737,9 +1742,6 @@ static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
 	for (disk = 0; disk < geo->raid_disks; disk++) {
 		sector_t dev_start, dev_end;
 		struct bio *mbio, *rbio = NULL;
-		struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
-		struct md_rdev *rrdev = rcu_dereference(
-			conf->mirrors[disk].replacement);
 
 		/*
 		 * Now start to calculate the start and end address for each disk.
@@ -1769,9 +1771,12 @@ static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
 
 		/*
 		 * It only handles discard bio which size is >= stripe size, so
-		 * dev_end > dev_start all the time
+		 * dev_end > dev_start all the time.
+		 * It doesn't need to use rcu lock to get rdev here. We already
+		 * add rdev->nr_pending in the first loop.
 		 */
 		if (r10_bio->devs[disk].bio) {
+			struct md_rdev *rdev = conf->mirrors[disk].rdev;
 			mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
 			mbio->bi_end_io = raid10_end_discard_request;
 			mbio->bi_private = r10_bio;
@@ -1784,6 +1789,7 @@ static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
 			bio_endio(mbio);
 		}
 		if (r10_bio->devs[disk].repl_bio) {
+			struct md_rdev *rrdev = conf->mirrors[disk].replacement;
 			rbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
 			rbio->bi_end_io = raid10_end_discard_request;
 			rbio->bi_private = r10_bio;