On 9/11/20 10:40 AM, Ido Schimmel wrote:
>>> @@ -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()
>
ok. Let's keep the order you have which I believe is consistent with FIB
notifiers.
