linux-kernel - Re: [PATCH net-next] devlink: Require devlink lock during device reload

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 8 Nov 2021 21:58:36 +0200
From:   Leon Romanovsky <leon@...nel.org>
To:     Jakub Kicinski <kuba@...nel.org>
Cc:     Ido Schimmel <idosch@...sch.org>, Jiri Pirko <jiri@...nulli.us>,
        "David S . Miller" <davem@...emloft.net>,
        Jiri Pirko <jiri@...dia.com>, linux-kernel@...r.kernel.org,
        netdev@...r.kernel.org, edwin.peer@...adcom.com
Subject: Re: [PATCH net-next] devlink: Require devlink lock during device
 reload

On Mon, Nov 08, 2021 at 10:46:08AM -0800, Jakub Kicinski wrote:
> On Mon, 8 Nov 2021 20:24:37 +0200 Leon Romanovsky wrote:
> > > I prefer my version. I think I asked you to show how the changes make
> > > drivers simpler, which you failed to do.  
> > 
> > Why did I fail? My version requires **zero** changes to the drivers.
> > Everything works without them changing anything. You can't ask for more.
> 
> For the last time.
> 
> "Your version" does require driver changes, but for better or worse
> we have already committed them to the tree. All the re-ordering to make
> sure devlink is registered last and more work is done at alloc,
> remember?

It fixed access to devlink before driver is ready. Also it fixed devlink
reload of simple drivers (without net namespaces support). So yes, at
least for now, we have a workaround to devlink reload bugs. We rmmod
mlx5_ib before reload and after. Everything thanks to reordering.

> 
> The goal is to make the upstream drivers simpler. You failed to show
> how your code does that.
> 
> Maybe you don't see the benefit because upstream simplifications are
> hard to depend on in out-of-tree drivers?

I don't care about out-of-tree drivers, mlx5 is fully upstream.

> 
> > > I already told you how this is going to go, don't expect me to comment
> > > too much.
> > >   
> > > > However for net namespace aware drivers it still stays DOA.
> > > > 
> > > > As you can see, devlink reload holds pernet_ops_rwsem, which drivers should
> > > > take too in order to unregister_netdevice_notifier.
> > > > 
> > > > So for me, the difference between netdevsim and real device (mlx5) is
> > > > too huge to really invest time into netdevsim-centric API, because it
> > > > won't solve any of real world problems.  
> > > 
> > > Did we not already go over this? Sorry, it feels like you're repeating
> > > arguments which I replied to before. This is exhausting.  
> > 
> > I don't enjoy it either.
> > 
> > > nfp will benefit from the simplified locking as well, and so will bnxt,
> > > although I'm not sure the maintainers will opt for using devlink framework
> > > due to the downstream requirements.  
> > 
> > Exactly why devlink should be fixed first.
> 
> If by "fixed first" you mean it needs 5 locks to be added and to remove
> any guarantees on sub-object lifetime then no thanks.

How do you plan to fix pernet_ops_rwsem lock? By exposing devlink state
to the drivers? By providing unlocked version of unregister_netdevice_notifier?

This simple scenario has deadlocks:
sudo ip netns add n1
sudo devlink dev reload pci/0000:00:09.0 netns n1
sudo ip netns del n1

https://lore.kernel.org/netdev/20211108104608.378c106e@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com/T/#m94b5c173f134c7d19daf455e3f6bad5ba6afd90d