md/raid10: close race that lose writes lost when replacement completes.
authorNeilBrown <neilb@suse.de>
Thu, 22 Nov 2012 03:42:49 +0000 (14:42 +1100)
committerNeilBrown <neilb@suse.de>
Thu, 22 Nov 2012 04:12:36 +0000 (15:12 +1100)
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

index d1295aff41739eea048f0e43513dfba6ade4c8f1..ad032518fdc00981b344dc5c1c98880defd79095 100644 (file)
@@ -1334,18 +1334,21 @@ retry_write:
                        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 @@ retry_write:
                                        max_sectors = good_sectors;
                        }
                }
-               r10_bio->devs[i].bio = bio;
-               atomic_inc(&rdev->nr_pending);
+               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);
@@ -1444,69 +1449,71 @@ retry_write:
        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);
+                       r10_bio->devs[i].bio = mbio;
+
+                       mbio->bi_sector = (r10_bio->devs[i].addr+
+                                          choose_data_offset(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;
 
-               mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
-               md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
-                           max_sectors);
-               r10_bio->devs[i].bio = mbio;
+                       atomic_inc(&r10_bio->remaining);
 
-               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;
-               mbio->bi_end_io = raid10_end_write_request;
-               mbio->bi_rw = WRITE | do_sync | do_fua | do_discard;
-               mbio->bi_private = r10_bio;
+                       cb = blk_check_plugged(raid10_unplug, mddev,
+                                              sizeof(*plug));
+                       if (cb)
+                               plug = container_of(cb, struct raid10_plug_cb,
+                                                   cb);
+                       else
+                               plug = NULL;
+                       spin_lock_irqsave(&conf->device_lock, flags);
+                       if (plug) {
+                               bio_list_add(&plug->pending, mbio);
+                               plug->pending_cnt++;
+                       } else {
+                               bio_list_add(&conf->pending_bio_list, mbio);
+                               conf->pending_count++;
+                       }
+                       spin_unlock_irqrestore(&conf->device_lock, flags);
+                       if (!plug)
+                               md_wakeup_thread(mddev->thread);
+               }
 
-               atomic_inc(&r10_bio->remaining);
+               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;
+
+                       mbio->bi_sector = (r10_bio->devs[i].addr +
+                                          choose_data_offset(
+                                                  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;
 
-               cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
-               if (cb)
-                       plug = container_of(cb, struct raid10_plug_cb, cb);
-               else
-                       plug = NULL;
-               spin_lock_irqsave(&conf->device_lock, flags);
-               if (plug) {
-                       bio_list_add(&plug->pending, mbio);
-                       plug->pending_cnt++;
-               } else {
+                       atomic_inc(&r10_bio->remaining);
+                       spin_lock_irqsave(&conf->device_lock, flags);
                        bio_list_add(&conf->pending_bio_list, mbio);
                        conf->pending_count++;
+                       spin_unlock_irqrestore(&conf->device_lock, flags);
+                       if (!mddev_check_plugged(mddev))
+                               md_wakeup_thread(mddev->thread);
                }
-               spin_unlock_irqrestore(&conf->device_lock, flags);
-               if (!plug)
-                       md_wakeup_thread(mddev->thread);
-
-               if (!r10_bio->devs[i].repl_bio)
-                       continue;
-
-               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;
-               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);
-               spin_lock_irqsave(&conf->device_lock, flags);
-               bio_list_add(&conf->pending_bio_list, mbio);
-               conf->pending_count++;
-               spin_unlock_irqrestore(&conf->device_lock, flags);
-               if (!mddev_check_plugged(mddev))
-                       md_wakeup_thread(mddev->thread);
        }
 
        /* Don't remove the bias on 'remaining' (one_write_done) until