dm mpath: fix racey management of PG initialization
authorMike Snitzer <snitzer@redhat.com>
Mon, 24 Aug 2020 18:19:55 +0000 (14:19 -0400)
committerMike Snitzer <snitzer@redhat.com>
Tue, 1 Sep 2020 20:41:56 +0000 (16:41 -0400)
Commit 935fcc56abc3 ("dm mpath: only flush workqueue when needed")
changed flush_multipath_work() to avoid needless workqueue
flushing (of a multipath global workqueue). But that change didn't
realize the surrounding flush_multipath_work() code should also only
run if 'pg_init_in_progress' is set.

Fix this by only doing all of flush_multipath_work()'s PG init related
work if 'pg_init_in_progress' is set.

Otherwise multipath_wait_for_pg_init_completion() will run
unconditionally but the preceeding flush_workqueue(kmpath_handlerd)
may not. This could lead to deadlock (though only if kmpath_handlerd
never runs a corresponding work to decrement 'pg_init_in_progress').

It could also be, though highly unlikely, that the kmpath_handlerd
work that does PG init completes before 'pg_init_in_progress' is set,
and then an intervening DM table reload's multipath_postsuspend()
triggers flush_multipath_work().

Fixes: 935fcc56abc3 ("dm mpath: only flush workqueue when needed")
Cc: stable@vger.kernel.org
Reported-by: Ben Marzinski <bmarzins@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-mpath.c

index e3283d3..de4da82 100644 (file)
@@ -1287,17 +1287,25 @@ static void multipath_wait_for_pg_init_completion(struct multipath *m)
 static void flush_multipath_work(struct multipath *m)
 {
        if (m->hw_handler_name) {
-               set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
-               smp_mb__after_atomic();
+               unsigned long flags;
+
+               if (!atomic_read(&m->pg_init_in_progress))
+                       goto skip;
+
+               spin_lock_irqsave(&m->lock, flags);
+               if (atomic_read(&m->pg_init_in_progress) &&
+                   !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
+                       spin_unlock_irqrestore(&m->lock, flags);
 
-               if (atomic_read(&m->pg_init_in_progress))
                        flush_workqueue(kmpath_handlerd);
-               multipath_wait_for_pg_init_completion(m);
+                       multipath_wait_for_pg_init_completion(m);
 
-               clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
-               smp_mb__after_atomic();
+                       spin_lock_irqsave(&m->lock, flags);
+                       clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
+               }
+               spin_unlock_irqrestore(&m->lock, flags);
        }
-
+skip:
        if (m->queue_mode == DM_TYPE_BIO_BASED)
                flush_work(&m->process_queued_bios);
        flush_work(&m->trigger_event);