bluetooth: rfcomm_dev_state_change deadlock fix
authorDave Young <hidave.darkstar@gmail.com>
Mon, 2 Jun 2008 06:50:52 +0000 (23:50 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 3 Jun 2008 21:27:17 +0000 (14:27 -0700)
There's logic in __rfcomm_dlc_close:
rfcomm_dlc_lock(d);
d->state = BT_CLOSED;
d->state_changed(d, err);
rfcomm_dlc_unlock(d);

In rfcomm_dev_state_change, it's possible that rfcomm_dev_put try to
take the dlc lock, then we will deadlock.

Here fixed it by unlock dlc before rfcomm_dev_get in
rfcomm_dev_state_change.

why not unlock just before rfcomm_dev_put? it's because there's
another problem.  rfcomm_dev_get/rfcomm_dev_del will take
rfcomm_dev_lock, but in rfcomm_dev_add the lock order is :
rfcomm_dev_lock --> dlc lock

so I unlock dlc before the taken of rfcomm_dev_lock.

Actually it's a regression caused by commit
1905f6c736cb618e07eca0c96e60e3c024023428 ("bluetooth :
__rfcomm_dlc_close lock fix"), the dlc state_change could be two
callbacks : rfcomm_sk_state_change and rfcomm_dev_state_change. I
missed the rfcomm_sk_state_change that time.

Thanks Arjan van de Ven <arjan@linux.intel.com> for the effort in
commit 4c8411f8c115def968820a4df6658ccfd55d7f1a ("bluetooth: fix
locking bug in the rfcomm socket cleanup handling") but he missed the
rfcomm_dev_state_change lock issue.

Signed-off-by: Dave Young <hidave.darkstar@gmail.com>
Acked-by: Marcel Holtmann <marcel@holtmann.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/bluetooth/rfcomm/tty.c

index c3f749abb2d07a1f99274fc71873d86e02f7e9b4..c9191871c1e0921cf9b096a3fdcad79c28f30309 100644 (file)
@@ -566,11 +566,22 @@ static void rfcomm_dev_state_change(struct rfcomm_dlc *dlc, int err)
        if (dlc->state == BT_CLOSED) {
                if (!dev->tty) {
                        if (test_bit(RFCOMM_RELEASE_ONHUP, &dev->flags)) {
-                               if (rfcomm_dev_get(dev->id) == NULL)
+                               /* Drop DLC lock here to avoid deadlock
+                                * 1. rfcomm_dev_get will take rfcomm_dev_lock
+                                *    but in rfcomm_dev_add there's lock order:
+                                *    rfcomm_dev_lock -> dlc lock
+                                * 2. rfcomm_dev_put will deadlock if it's
+                                *    the last reference
+                                */
+                               rfcomm_dlc_unlock(dlc);
+                               if (rfcomm_dev_get(dev->id) == NULL) {
+                                       rfcomm_dlc_lock(dlc);
                                        return;
+                               }
 
                                rfcomm_dev_del(dev);
                                rfcomm_dev_put(dev);
+                               rfcomm_dlc_lock(dlc);
                        }
                } else
                        tty_hangup(dev->tty);