md/raid1: don't clear bitmap bit when bad-block-list write fails.
authorNeilBrown <neilb@suse.com>
Sat, 24 Oct 2015 05:02:16 +0000 (16:02 +1100)
committerNeilBrown <neilb@suse.com>
Sat, 24 Oct 2015 05:24:22 +0000 (16:24 +1100)
When a write fails and a bad-block-list is present, we can
update the bad-block-list instead of writing the data.  If
this succeeds then it is OK clear the relevant bitmap-bit as
no further 'sync' of the block is needed.

However if writing the bad-block-list fails then we need to
treat the write as failed and particularly must not clear
the bitmap bit.  Otherwise the device can be re-added (after
any hardware connection issues are resolved) and because the
relevant bit in the bitmap is clear, that block will not be
resynced.  This leads to data corruption.

We already delay the final bio_endio() on the write until
the bad-block-list is written so that when the write
returns: either that data is safe, the bad-block record is
safe, or the fact that the device is faulty is safe.
However we *don't* delay the clearing of the bitmap, so the
bitmap bit can be recorded as cleared before we know if the
bad-block-list was written safely.

So: delay that until the write really is safe.
i.e. move the call to close_write() until just before
calling bio_endio(), and recheck the 'is array degraded'
status before making that call.

This bug goes back to v3.1 when bad-block-lists were
introduced, though it only affects arrays created with
mdadm-3.3 or later as only those have bad-block lists.

Backports will require at least
Commit: 55ce74d4bfe1 ("md/raid1: ensure device failure recorded before write request returns.")
as well.  I'll send that to 'stable' separately.

Note that of the two tests of R1BIO_WriteError that this
patch adds, the first is certain to fail and the second is
certain to succeed.  However doing it this way makes the
patch more obviously correct.  I will tidy the code up in a
future merge window.

Reported-and-tested-by: Nate Dailey <nate.dailey@stratus.com>
Cc: Jes Sorensen <Jes.Sorensen@redhat.com>
Fixes: cd5ff9a16f08 ("md/raid1:  Handle write errors by updating badblock log.")
Signed-off-by: NeilBrown <neilb@suse.com>
drivers/md/raid1.c

index cfca6ed..d9d031e 100644 (file)
@@ -2258,15 +2258,16 @@ static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
                        rdev_dec_pending(conf->mirrors[m].rdev,
                                         conf->mddev);
                }
-       if (test_bit(R1BIO_WriteError, &r1_bio->state))
-               close_write(r1_bio);
        if (fail) {
                spin_lock_irq(&conf->device_lock);
                list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
                spin_unlock_irq(&conf->device_lock);
                md_wakeup_thread(conf->mddev->thread);
-       } else
+       } else {
+               if (test_bit(R1BIO_WriteError, &r1_bio->state))
+                       close_write(r1_bio);
                raid_end_bio_io(r1_bio);
+       }
 }
 
 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
@@ -2385,6 +2386,10 @@ static void raid1d(struct md_thread *thread)
                        r1_bio = list_first_entry(&tmp, struct r1bio,
                                                  retry_list);
                        list_del(&r1_bio->retry_list);
+                       if (mddev->degraded)
+                               set_bit(R1BIO_Degraded, &r1_bio->state);
+                       if (test_bit(R1BIO_WriteError, &r1_bio->state))
+                               close_write(r1_bio);
                        raid_end_bio_io(r1_bio);
                }
        }