Re: [PATCH net] nfp: fix rcu_read_lock/unlock while rcu_derefrencing

[Louis Peens <louis.peens@corigine.com>]

On Tue, May 09, 2023 at 07:40:13PM -0700, Jakub Kicinski wrote:
> On Tue,  9 May 2023 08:06:32 +0200 Louis Peens wrote:
> > +static inline
> > +struct net_device *nfp_app_dev_get_locked(struct nfp_app *app, u32 id,
> 
> _locked() in what way? RCU functions typically use an _rcu suffix, no?
We were discussing the naming during internal review, for some reason didn't
think about using _rcu, will update if there is a v2.
> 
> > +					  bool *redir_egress)
> > +{
> > +	struct net_device *dev;
> > +
> > +	if (unlikely(!app || !app->type->dev_get))
> > +		return NULL;
> > +
> > +	rcu_read_lock();
> > +	dev = app->type->dev_get(app, id, redir_egress);
> > +	rcu_read_unlock();
> > +
> > +	return dev;
> 
> this looks very suspicious, RCU takes care primarily of the lifetime of
> objects, in this case dev. Returning it after dropping the lock seems
> wrong.
> 
> If the context is safe maybe it's a better idea to change the 
> condition in rcu_dereference_check() to include rcu_read_lock_bh_held()?
Thanks, will take a closer look at this.
> -- 
> pw-bot: cr
>