Hello!
I'm working on stacking some virtual interfaces, and ran into a problem with
locking.
Basically, I have an ethernet-like device, and on top of that I am putting
802.1Q vlans..and on top of that, some other vlan (macvlan).
In the event notifier for the vlan, I attempt to delete all vlans who's
under-lying device is going away. I do the same for the macvlan.
Currently, I am setting a deadlock warning because the macvlan code
is calling unregister_netdev, but the 'rmmod' method already has
a lock in there due to having earlier called unregister_netdev on the
.1q vlan.
So, the question is: Should I be calling unregister_netdev from the notifier
callback?
Thanks,
Ben
--
Ben Greear [off-list ref]
Candela Technologies Inc http://www.candelatech.com
From: Ben Greear <redacted>
Date: Tue, 11 Jul 2006 16:27:25 -0700
So, the question is: Should I be calling unregister_netdev from the
notifier callback?
Not really, as you saw it can be deadlock prone.
One idea is to run the unregister asynchronously via a workqueue.
That should be able to get you out of the deadlock.
David Miller wrote:
From: Ben Greear <redacted>
Date: Tue, 11 Jul 2006 16:27:25 -0700
quoted
So, the question is: Should I be calling unregister_netdev from the
notifier callback?
Not really, as you saw it can be deadlock prone.
One idea is to run the unregister asynchronously via a workqueue.
That should be able to get you out of the deadlock.
After more poking, I notice that 802.1q uses unregister_netdevice,
which does not grab rtnl. It seems that the notifiers are always called
with rtnl asserted. That would be a recursive call to unregister_netdevice,
however...any reason that won't work?
Ben
--
Ben Greear [off-list ref]
Candela Technologies Inc http://www.candelatech.com