net: make sure devices go through netdev_wait_all_refs
authorJakub Kicinski <kuba@kernel.org>
Wed, 6 Jan 2021 18:40:07 +0000 (10:40 -0800)
committerJakub Kicinski <kuba@kernel.org>
Sat, 9 Jan 2021 03:27:41 +0000 (19:27 -0800)
If register_netdevice() fails at the very last stage - the
notifier call - some subsystems may have already seen it and
grabbed a reference. struct net_device can't be freed right
away without calling netdev_wait_all_refs().

Now that we have a clean interface in form of dev->needs_free_netdev
and lenient free_netdev() we can undo what commit 93ee31f14f6f ("[NET]:
Fix free_netdev on register_netdev failure.") has done and complete
the unregistration path by bringing the net_set_todo() call back.

After registration fails user is still expected to explicitly
free the net_device, so make sure ->needs_free_netdev is cleared,
otherwise rolling back the registration will cause the old double
free for callers who release rtnl_lock before the free.

This also solves the problem of priv_destructor not being called
on notifier error.

net_set_todo() will be moved back into unregister_netdevice_queue()
in a follow up.

Reported-by: Hulk Robot <hulkci@huawei.com>
Reported-by: Yang Yingliang <yangyingliang@huawei.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net/core/dev.c

index adde93cbca9f0b1934f45383d23cc61f899e0dd1..0071a11a6dc3cb36087b1afc2c4d1ddc1191fe6f 100644 (file)
@@ -10077,17 +10077,11 @@ int register_netdevice(struct net_device *dev)
        ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
        ret = notifier_to_errno(ret);
        if (ret) {
+               /* Expect explicit free_netdev() on failure */
+               dev->needs_free_netdev = false;
                rollback_registered(dev);
-               rcu_barrier();
-
-               dev->reg_state = NETREG_UNREGISTERED;
-               /* We should put the kobject that hold in
-                * netdev_unregister_kobject(), otherwise
-                * the net device cannot be freed when
-                * driver calls free_netdev(), because the
-                * kobject is being hold.
-                */
-               kobject_put(&dev->dev.kobj);
+               net_set_todo(dev);
+               goto out;
        }
        /*
         *      Prevent userspace races by waiting until the network