dm raid: fix deadlock caused by premature md_stop_writes()
authorHeinz Mauelshagen <heinzm@redhat.com>
Sat, 2 Dec 2017 00:03:48 +0000 (01:03 +0100)
committerMike Snitzer <snitzer@redhat.com>
Fri, 8 Dec 2017 15:59:57 +0000 (10:59 -0500)
md_stop_writes() is called in raid_presuspend() causing deadlocks on
bios submitted afterwards -- which happens on loaded raid sets with
conversion requests.

Fix by moving md_stop_writes() to raid_postsuspend().  NOTE: when the
recovery's frozen (MD_RECOVERY_FROZEN), writes haven't been started (or
are already stopped) so don't stop them again.

Also remove superfluous readonly setting.

Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-raid.c

index 6319d846e0adb8c264f98676f0c2ddb8e85bdbb7..398314b6c31af76457c9cdfd585d1bbe051e6c09 100644 (file)
@@ -3613,24 +3613,19 @@ static void raid_io_hints(struct dm_target *ti, struct queue_limits *limits)
        blk_limits_io_opt(limits, chunk_size * mddev_data_stripes(rs));
 }
 
-static void raid_presuspend(struct dm_target *ti)
-{
-       struct raid_set *rs = ti->private;
-
-       md_stop_writes(&rs->md);
-}
-
 static void raid_postsuspend(struct dm_target *ti)
 {
        struct raid_set *rs = ti->private;
 
        if (!test_and_set_bit(RT_FLAG_RS_SUSPENDED, &rs->runtime_flags)) {
+               /* Writes have to be stopped before suspending to avoid deadlocks. */
+               if (!test_bit(MD_RECOVERY_FROZEN, &rs->md.recovery))
+                       md_stop_writes(&rs->md);
+
                mddev_lock_nointr(&rs->md);
                mddev_suspend(&rs->md);
                mddev_unlock(&rs->md);
        }
-
-       rs->md.ro = 1;
 }
 
 static void attempt_restore_of_faulty_devices(struct raid_set *rs)
@@ -3903,7 +3898,6 @@ static struct target_type raid_target = {
        .message = raid_message,
        .iterate_devices = raid_iterate_devices,
        .io_hints = raid_io_hints,
-       .presuspend = raid_presuspend,
        .postsuspend = raid_postsuspend,
        .preresume = raid_preresume,
        .resume = raid_resume,