Re: [PATCH net-next 2/9] net: hold instance lock during NETDEV_REGISTER/UP/UNREGISTER

[Stanislav Fomichev <stfomichev@gmail.com>]

On 03/26, Cosmin Ratiu wrote:
> On Wed, 2025-03-26 at 15:03 +0000, Cosmin Ratiu wrote:
> > On Tue, 2025-03-25 at 14:30 -0700, Stanislav Fomichev wrote:
> > > @@ -2072,8 +2087,8 @@ static void
> > > __move_netdevice_notifier_net(struct net *src_net,
> > >  					  struct net *dst_net,
> > >  					  struct notifier_block
> > > *nb)
> > >  {
> > > -	__unregister_netdevice_notifier_net(src_net, nb);
> > > -	__register_netdevice_notifier_net(dst_net, nb, true);
> > > +	__unregister_netdevice_notifier_net(src_net, nb, false);
> > > +	__register_netdevice_notifier_net(dst_net, nb, true,
> > > false);
> > >  }
> > 
> > I tested with your (and the rest of Jakub's) patches.
> > The problem with this approach is that when a netdev's net is
> > changed,
> > its lock will be acquired, but the notifiers for ALL netdevs in the
> > old
> > and the new namespace will be called, which will result in correct
> > behavior for that device and lockdep_assert_held failure for all
> > others.
> 
> But a thing I've learned many years ago about locking is that locks
> should protect data, not code. Shouldn't we avoid locking deep call
> hierarchies (like notifiers) with the instance lock and instead focus
> on 1) what fields need to be protected by the lock and 2) reduce
> critical section length for those fields.
> 
> That plus reference counting usually does the trick and should avoid
> these ugly deadlocks.

We want the operations to look atomic from the userspace if possible.
So the whole device is either moved or not, some other thread should
not be able to change, say, mtu mid-way.

And we do try to clarify what's specifically protected in terms of data:
https://web.git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/tree/include/linux/netdevice.h#n2494

But the notifiers are super tricky. There are years of natural growth
with the assumption of a single rtnl lock :-(