ALSA: seq: Fix nested rwsem annotation for lockdep splat

commit 1f20f9ff57ca23b9f5502fca85ce3977e8496cb1 upstream.

syzkaller reported the lockdep splat due to the possible deadlock of
grp->list_mutex of each sequencer client object.  Actually this is
rather a false-positive report due to the missing nested lock
annotations.  The sequencer client may deliver the event directly to
another client which takes another own lock.

For addressing this issue, this patch replaces the simple down_read()
with down_read_nested().  As a lock subclass, the already existing
"hop" can be re-used, which indicates the depth of the call.

Reference: http://lkml.kernel.org/r/089e082686ac9b482e055c832617@google.com
Reported-by: syzbot <bot+7feb8de6b4d6bf810cf098bef942cc387e79d0ad@syzkaller.appspotmail.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>

diff --git a/sound/core/seq/seq_clientmgr.c b/sound/core/seq/seq_clientmgr.c
index 7b5a790..c789f91 100644 (file)
--- a/sound/core/seq/seq_clientmgr.c
+++ b/sound/core/seq/seq_clientmgr.c
@@ -676,7 +676,7 @@ static int deliver_to_subscribers(struct snd_seq_client *client,
        if (atomic)
                read_lock(&grp->list_lock);
        else
-               down_read(&grp->list_mutex);
+               down_read_nested(&grp->list_mutex, hop);
        list_for_each_entry(subs, &grp->list_head, src_list) {
                /* both ports ready? */
                if (atomic_read(&subs->ref_count) != 2)