net: don't mix device locking in dev_close_many() calls

Lockdep found the following dependency:

  &dev_instance_lock_key#3 -->
     &rdev->wiphy.mtx -->
        &net->xdp.lock -->
	   &xs->mutex -->
	      &dev_instance_lock_key#3

The first dependency is the problem. wiphy mutex should be outside
the instance locks. The problem happens in notifiers (as always)
for CLOSE. We only hold the instance lock for ops locked devices
during CLOSE, and WiFi netdevs are not ops locked. Unfortunately,
when we dev_close_many() during netns dismantle we may be holding
the instance lock of _another_ netdev when issuing a CLOSE for
a WiFi device.

Lockdep's "Possible unsafe locking scenario" only prints 3 locks
and we have 4, plus I think we'd need 3 CPUs, like this:

       CPU0                 CPU1              CPU2
       ----                 ----              ----
  lock(&xs->mutex);
                       lock(&dev_instance_lock_key#3);
                                         lock(&rdev->wiphy.mtx);
                                         lock(&net->xdp.lock);
                                         lock(&xs->mutex);
                       lock(&rdev->wiphy.mtx);
  lock(&dev_instance_lock_key#3);

Tho, I don't think that's possible as CPU1 and CPU2 would
be under rtnl_lock. Even if we have per-netns rtnl_lock and
wiphy can span network namespaces - CPU0 and CPU1 must be
in the same netns to see dev_instance_lock, so CPU0 can't
be installing a socket as CPU1 is tearing the netns down.

Regardless, our expected lock ordering is that wiphy lock
is taken before instance locks, so let's fix this.

Go over the ops locked and non-locked devices separately.
Note that calling dev_close_many() on an empty list is perfectly
fine. All processing (including RCU syncs) are conditional
on the list not being empty, already.

Fixes: 7e4d784 ("net: hold netdev instance lock during rtnetlink operations")
Reported-by: [email protected]
Acked-by: Stanislav Fomichev <[email protected]>
Link: https://patch.msgid.link/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>

Author: kuba-moo

net/core/dev.c

@@ -11932,15 +11932,24 @@ void unregister_netdevice_many_notify(struct list_head *head,
 		BUG_ON(dev->reg_state != NETREG_REGISTERED);
 	}
 
-	/* If device is running, close it first. */
+	/* If device is running, close it first. Start with ops locked... */
 	list_for_each_entry(dev, head, unreg_list) {
-		list_add_tail(&dev->close_list, &close_head);
-		netdev_lock_ops(dev);
+		if (netdev_need_ops_lock(dev)) {
+			list_add_tail(&dev->close_list, &close_head);
+			netdev_lock(dev);
+		}
+	}
+	dev_close_many(&close_head, true);
+	/* ... now unlock them and go over the rest. */
+	list_for_each_entry(dev, head, unreg_list) {
+		if (netdev_need_ops_lock(dev))
+			netdev_unlock(dev);
+		else
+			list_add_tail(&dev->close_list, &close_head);
 	}
 	dev_close_many(&close_head, true);
 
 	list_for_each_entry(dev, head, unreg_list) {
-		netdev_unlock_ops(dev);
 		/* And unlink it from device chain. */
 		unlist_netdevice(dev);
 		netdev_lock(dev);