Re: "new style" netdevice allocation patch for TUN driver (2.4.18 kernel)

["Maksim (Max) Krasnyanskiy" <maxk@qualcomm.com>]

>    From: "Maksim (Max) Krasnyanskiy" <maxk@qualcomm.com>
>    Date: Tue, 06 Aug 2002 10:07:49 -0700
>
>    Dave, how about this
>
>    --- net/core/dev.c.orig Mon Aug  5 21:48:54 2002
>    +++ net/core/dev.c      Mon Aug  5 21:54:01 2002
>    @@ -2577,6 +2577,11 @@
>
>First, the call-chain notifiers are probably not safe
>to run without rtnl_lock held.
Good point.

>Second, why not just fix the bug instead of applying band-aids
>to device unregistry?  I know it's nice in that it allows you
>to configure devices some more, but it doesn't make the real
>problem go away.
I completely agree. However sleeping and holding a lock that you
don't have to hold is a bug on it's own :).
Things like sockets drop the lock before calling schedule() and acquire
it on wakeup. I think we should do the same in unregister_netdevice().

How about this:

--- dev.c.orig  Tue Aug  6 00:58:46 2002
+++ dev.c       Tue Aug  6 01:00:00 2002
@@ -2584,9 +2584,12 @@
                         notifier_call_chain(&netdev_chain, 
NETDEV_UNREGISTER, dev);
                 }

+               rtnl_unlock();
                 current->state = TASK_INTERRUPTIBLE;
                 schedule_timeout(HZ/4);
                 current->state = TASK_RUNNING;
+               rtnl_lock();
+
                 if ((jiffies - warning_time) > 10*HZ) {
                         printk(KERN_EMERG "unregister_netdevice: waiting 
for %s to "
                                         "become free. Usage count = %d\n",
-----

Max