Re: circular locking, mirred, 2.6.24.2

[Jarek Poplawski <jarkao2@gmail.com>]

On Thu, Mar 06, 2008 at 08:59:55AM -0500, jamal wrote:
> On Thu, 2008-06-03 at 13:40 +0000, Jarek Poplawski wrote:
> 
> > I'm not sure this lockdep report is because of this, but there is
> > really a problem with lock order while using sch_ingress with
> > act_mirred: dev->queue_lock is taken after dev->ingress_lock, so
> > reversely to e.g. qdisc_lock_tree(). This shouldn't be a problem
> > when one of the devices is ifb yet.
> 
> Are there more details? Ingress of which netdevice is redirecting to
> egress of which netdevice?
> Sorry, I dont understand much about the internals of lockdep so i dont
> know what you are teaching it in the patch below...

Every netdevice after register_netdevice() has its queue_lock and
ingress_lock initialized with the same static lock_class_keys, so unless
changed later, these locks are treated by lockdep as 2 global locks. So,
taking such locks with different order should be reported. This really
happens in act_mirred, and I don't know yet, why this wasn't reported
earlier.

Of course, if there are two different devices this could be safe, but
not always (e.g. thread1 has dev_eth0->ingress_lock and wants
dev_eth1->queue_lock, thread2 has dev_eth1->ingress_lock, wants
dev_eth0->qdisc_lock, and thread3 has dev_eth0->qdisc_lock and wants
dev_eth0->ingress_lock). With ifb this shouldn't be the case, but
anyway we have to tell lockdep that ifb uses a different pair of locks.

My patch teaches lockdep about queue_lock, but after rethinking it
seems it's not enough, and I'll have to update this patch with
ingress_lock too.

Denys, I'll have to read your script yet, so you can wait with this
patching (unless you've started already - anyway this patch shouldn't
be harmful).

Thanks,
Jarek P.