Re: [RFC PATCH net-next 17/22] nexthop: Replay nexthops when registering a notifier

[Ido Schimmel <idosch@idosch.org>]

On Tue, Sep 08, 2020 at 09:37:10AM -0600, David Ahern wrote:
> On 9/8/20 3:10 AM, Ido Schimmel wrote:
> > From: Ido Schimmel <idosch@nvidia.com>
> > 
> > When registering a new notifier to the nexthop notification chain,
> > replay all the existing nexthops to the new notifier so that it will
> > have a complete picture of the available nexthops.
> > 
> > Signed-off-by: Ido Schimmel <idosch@nvidia.com>
> > ---
> >  net/ipv4/nexthop.c | 54 ++++++++++++++++++++++++++++++++++++++++++++--
> >  1 file changed, 52 insertions(+), 2 deletions(-)
> > 
> > diff --git a/net/ipv4/nexthop.c b/net/ipv4/nexthop.c
> > index b40c343ca969..6505a0a28df2 100644
> > --- a/net/ipv4/nexthop.c
> > +++ b/net/ipv4/nexthop.c
> > @@ -156,6 +156,27 @@ static int call_nexthop_notifiers(struct net *net,
> >  	return notifier_to_errno(err);
> >  }
> >  
> > +static int call_nexthop_notifier(struct notifier_block *nb, struct net *net,
> > +				 enum nexthop_event_type event_type,
> > +				 struct nexthop *nh,
> > +				 struct netlink_ext_ack *extack)
> > +{
> > +	struct nh_notifier_info info = {
> > +		.net = net,
> > +		.extack = extack,
> > +	};
> > +	int err;
> > +
> > +	err = nh_notifier_info_init(&info, nh);
> > +	if (err)
> > +		return err;
> > +
> > +	err = nb->notifier_call(nb, event_type, &info);
> > +	nh_notifier_info_fini(&info);
> > +
> > +	return notifier_to_errno(err);
> > +}
> > +
> >  static unsigned int nh_dev_hashfn(unsigned int val)
> >  {
> >  	unsigned int mask = NH_DEV_HASHSIZE - 1;
> > @@ -2116,11 +2137,40 @@ static struct notifier_block nh_netdev_notifier = {
> >  	.notifier_call = nh_netdev_event,
> >  };
> >  
> > +static int nexthops_dump(struct net *net, struct notifier_block *nb,
> > +			 struct netlink_ext_ack *extack)
> > +{
> > +	struct rb_root *root = &net->nexthop.rb_root;
> > +	struct rb_node *node;
> > +	int err = 0;
> > +
> > +	for (node = rb_first(root); node; node = rb_next(node)) {
> > +		struct nexthop *nh;
> > +
> > +		nh = rb_entry(node, struct nexthop, rb_node);
> > +		err = call_nexthop_notifier(nb, net, NEXTHOP_EVENT_REPLACE, nh,
> > +					    extack);
> > +		if (err)
> > +			break;
> > +	}
> > +
> > +	return err;
> > +}
> > +
> >  int register_nexthop_notifier(struct net *net, struct notifier_block *nb,
> >  			      struct netlink_ext_ack *extack)
> >  {
> > -	return blocking_notifier_chain_register(&net->nexthop.notifier_chain,
> > -						nb);
> > +	int err;
> > +
> > +	rtnl_lock();
> > +	err = nexthops_dump(net, nb, extack);
> 
> can the unlock be moved here? register function below should not need it.

It can result in this unlikely race:

<t0> - rtnl_lock(); nexthops_dump(); rtnl_unlock()
<t1> - Nexthop is added / deleted
<t2> - blocking_notifier_chain_register()

It is possible to flip the order:

<t0> - blocking_notifier_chain_register()
<t1> - rtnl_lock(); nexthops_dump(); rtnl_unlock()

Worst case:

<t0> - blocking_notifier_chain_register()
<t1> - Nexthop is added / deleted
<t2> - rtnl_lock(); nexthops_dump(); rtnl_unlock()

Which is OK. If we get a delete notification for a nexthop we don't
know, we ignore it. If we get two replace notifications for the same
nexthop we just "update" it.

> 
> > +	if (err)
> > +		goto unlock;
> > +	err = blocking_notifier_chain_register(&net->nexthop.notifier_chain,
> > +					       nb);
> > +unlock:
> > +	rtnl_unlock();
> > +	return err;
> >  }
> >  EXPORT_SYMBOL(register_nexthop_notifier);
> >  
> > 
>