md/raid10: close race that lose writes lost when replacement completes.

When a replacement operation completes there is a small window
when the original device is marked 'faulty' and the replacement
still looks like a replacement.  The faulty should be removed and
the replacement moved in place very quickly, bit it isn't instant.

So the code write out to the array must handle the possibility that
the only working device for some slot in the replacement - but it
doesn't.  If the primary device is faulty it just gives up.  This
can lead to corruption.

So make the code more robust: if either  the primary or the
replacement is present and working, write to them.  Only when
neither are present do we give up.

This bug has been present since replacement was introduced in
3.3, so it is suitable for any -stable kernel since then.
Reported-by: "George Spelvin" <linux@horizon.com>
Cc: stable@vger.kernel.org
Signed-off-by: NeilBrown <neilb@suse.de>

drivers/md/raid10.c

@@ -1334,18 +1334,21 @@ static void make_request(struct mddev *mddev, struct bio * bio)
 			blocked_rdev = rrdev;
 			break;
 		}
+		if (rdev && (test_bit(Faulty, &rdev->flags)
+			     || test_bit(Unmerged, &rdev->flags)))
+			rdev = NULL;
 		if (rrdev && (test_bit(Faulty, &rrdev->flags)
 			      || test_bit(Unmerged, &rrdev->flags)))
 			rrdev = NULL;
 
 		r10_bio->devs[i].bio = NULL;
 		r10_bio->devs[i].repl_bio = NULL;
-		if (!rdev || test_bit(Faulty, &rdev->flags) ||
-		    test_bit(Unmerged, &rdev->flags)) {
+
+		if (!rdev && !rrdev) {
 			set_bit(R10BIO_Degraded, &r10_bio->state);
 			continue;
 		}
-		if (test_bit(WriteErrorSeen, &rdev->flags)) {
+		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
 			sector_t first_bad;
 			sector_t dev_sector = r10_bio->devs[i].addr;
 			int bad_sectors;
@@ -1387,8 +1390,10 @@ static void make_request(struct mddev *mddev, struct bio * bio)
 					max_sectors = good_sectors;
 			}
 		}
+		if (rdev) {
 			r10_bio->devs[i].bio = bio;
 			atomic_inc(&rdev->nr_pending);
+		}
 		if (rrdev) {
 			r10_bio->devs[i].repl_bio = bio;
 			atomic_inc(&rrdev->nr_pending);
@@ -1444,9 +1449,8 @@ static void make_request(struct mddev *mddev, struct bio * bio)
 	for (i = 0; i < conf->copies; i++) {
 		struct bio *mbio;
 		int d = r10_bio->devs[i].devnum;
-		if (!r10_bio->devs[i].bio)
-			continue;
-
+		if (r10_bio->devs[i].bio) {
+			struct md_rdev *rdev = conf->mirrors[d].rdev;
 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 			md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
 				    max_sectors);
@@ -1454,17 +1458,19 @@ static void make_request(struct mddev *mddev, struct bio * bio)
 
 			mbio->bi_sector	= (r10_bio->devs[i].addr+
 					   choose_data_offset(r10_bio,
-						      conf->mirrors[d].rdev));
-		mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
+							      rdev));
+			mbio->bi_bdev = rdev->bdev;
 			mbio->bi_end_io	= raid10_end_write_request;
 			mbio->bi_rw = WRITE | do_sync | do_fua | do_discard;
 			mbio->bi_private = r10_bio;
 
 			atomic_inc(&r10_bio->remaining);
 
-		cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
+			cb = blk_check_plugged(raid10_unplug, mddev,
+					       sizeof(*plug));
 			if (cb)
-			plug = container_of(cb, struct raid10_plug_cb, cb);
+				plug = container_of(cb, struct raid10_plug_cb,
+						    cb);
 			else
 				plug = NULL;
 			spin_lock_irqsave(&conf->device_lock, flags);
@@ -1478,24 +1484,24 @@ static void make_request(struct mddev *mddev, struct bio * bio)
 			spin_unlock_irqrestore(&conf->device_lock, flags);
 			if (!plug)
 				md_wakeup_thread(mddev->thread);
+		}
 
-		if (!r10_bio->devs[i].repl_bio)
-			continue;
-
+		if (r10_bio->devs[i].repl_bio) {
+			struct md_rdev *rdev = conf->mirrors[d].replacement;
+			if (rdev == NULL) {
+				/* Replacement just got moved to main 'rdev' */
+				smp_mb();
+				rdev = conf->mirrors[d].rdev;
+			}
 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 			md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
 				    max_sectors);
 			r10_bio->devs[i].repl_bio = mbio;
 
-		/* We are actively writing to the original device
-		 * so it cannot disappear, so the replacement cannot
-		 * become NULL here
-		 */
 			mbio->bi_sector	= (r10_bio->devs[i].addr +
 					   choose_data_offset(
-					   r10_bio,
-					   conf->mirrors[d].replacement));
-		mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
+						   r10_bio, rdev));
+			mbio->bi_bdev = rdev->bdev;
 			mbio->bi_end_io	= raid10_end_write_request;
 			mbio->bi_rw = WRITE | do_sync | do_fua | do_discard;
 			mbio->bi_private = r10_bio;
@@ -1508,6 +1514,7 @@ static void make_request(struct mddev *mddev, struct bio * bio)
 			if (!mddev_check_plugged(mddev))
 				md_wakeup_thread(mddev->thread);
 		}
+	}
 
 	/* Don't remove the bias on 'remaining' (one_write_done) until
 	 * after checking if we need to go around again.