Re: [PATCH net V3 1/3] bonding: implement xdo_dev_state_free and call it after deletion

[Hangbin Liu <liuhangbin@gmail.com>]

On Thu, Aug 15, 2024 at 01:23:05AM +0000, Jianbo Liu wrote:
> On Wed, 2024-08-14 at 11:11 +0800, Hangbin Liu wrote:
> > On Wed, Aug 14, 2024 at 02:03:58AM +0000, Jianbo Liu wrote:
> > > On Tue, 2024-08-13 at 07:14 -0700, Jakub Kicinski wrote:
> > > > On Tue, 13 Aug 2024 02:58:12 +0000 Jianbo Liu wrote:
> > > > > > > +       rcu_read_lock();
> > > > > > > +       bond = netdev_priv(bond_dev);
> > > > > > > +       slave = rcu_dereference(bond->curr_active_slave);
> > > > > > > +       real_dev = slave ? slave->dev : NULL;
> > > > > > > +       rcu_read_unlock();  
> > > > > > 
> > > > > > What's holding onto real_dev once you drop the rcu lock
> > > > > > here?  
> > > > > 
> > > > > I think it should be xfrm state (and bond device).
> > > > 
> > > > Please explain it in the commit message in more certain terms.
> > > 
> > > Sorry, I don't understand. The real_dev is saved in xs-
> > > >xso.real_dev,
> > > and also bond's slave. It's straightforward. What else do I need to
> > > explain?
> > 
> > I think Jakub means you need to make sure the real_dev is not freed
> > during
> > xfrmdev_ops. See bond_ipsec_add_sa(). You unlock it too early and
> > later
> > xfrmdev_ops is not protected.
> 
> This RCU lock is to protect the reading of curr_active_slave, which is
> pointing to a big stuct - slave struct, so there is no error to get
> real_dev from slave->dev.

It's not about getting real_dev from slave->dev. As Jakub said, What's holding
on real_dev once you drop the rcu lock?

Thanks
Hangbin