Re: [PATCH net v2 08/11] docs: net: document netdev notifier expectations

[Stanislav Fomichev <stfomichev@gmail.com>]

On 03/27, Jakub Kicinski wrote:
> On Thu, 27 Mar 2025 12:16:13 -0700 Jakub Kicinski wrote:
> > > +* ``NETDEV_REGISTER``
> > > +* ``NETDEV_UP``
> > > +* ``NETDEV_UNREGISTER``  
> > 
> > Can I ask the obvious question - anything specific that's hard in also
> > taking it in DOWN or just no time to investigate? Symmetry would be
> > great.

The latter: I added locks for DOWN, hit a dev_close somewhere, and
decided no to go down the rabbit hole.

> Looking at patch 4 maybe we should do the opposite. This was my
> original commit msg for locking UNREGISTER:
> 
>     net: make NETDEV_UNREGISTER and instance lock more consistent
>     
>     The NETDEV_UNREGISTER notifier gets called under the ops lock
>     when device changes namespace but not during real unregistration.
>     Take it consistently, XSK tries to poke at netdev queue state
>     from this notifier.
> 
> So if the only caller currently under the lock is netns change, and 
> we already split that to release the lock - maybe we can make
> UNREGISTER always unlocked instead?

That sounds very sensible, let me try it out and run the tests.
I'll have to drop the lock twice, once for NETDEV_UNREGISTER
and another time for move_netdevice_notifiers_dev_net, but since
the device is unlisted, nothing should touch it (in theory)?

netif_change_net_namespace is already the first thing that happens
in do_setlink, so I won't be converting it to dev_xxx (lmk if I
miss something here).

If it goes well, I'll also hack your xsk patch to grab the ops lock
in xsk_notifier:NETDEV_UNREGISTER, I think that's the only thing
that needs changes. The ugly bonding and teaming unlock/locks
will go away (which is a very nice side effect).