dm mpath: fix racey management of PG initialization

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>

1 files changed, 15 insertions, 7 deletions

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index e3283d35c7fd..de4da825ade6 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -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);