[Patch net] ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf

[Patch net] ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf

[Cong Wang]

For each netns (except init_net), we initialize its null entry
in 3 places:

1) The template itself, as we use kmemdup()
2) Code around dst_init_metrics() in ip6_route_net_init()
3) ip6_route_dev_notify(), which is supposed to initialize it after
loopback registers

Unfortunately the last one still happens in a wrong order because
we expect to initialize net->ipv6.ip6_null_entry->rt6i_idev to
net->loopback_dev's idev, so we have to do that after we add
idev to it. However, this notifier has priority == 0 same as
ipv6_dev_notf, and ipv6_dev_notf is registered after
ip6_route_dev_notifier so it is called actually after
ip6_route_dev_notifier.

Fix it by specifying a smaller priority for ip6_route_dev_notifier.

Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
---
 net/ipv6/route.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index 2f11366..4dbf7e2 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -4024,7 +4024,7 @@ static struct pernet_operations ip6_route_net_late_ops = {
 
 static struct notifier_block ip6_route_dev_notifier = {
 	.notifier_call = ip6_route_dev_notify,
-	.priority = 0,
+	.priority = -10, /* Must be called after addrconf_notify!! */
 };
 
 void __init ip6_route_init_special_entries(void)
-- 
2.5.5

Re: [Patch net] ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf

[David Ahern]

On 5/3/17 11:07 PM, Cong Wang wrote:
> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> index 2f11366..4dbf7e2 100644
> --- a/net/ipv6/route.c
> +++ b/net/ipv6/route.c
> @@ -4024,7 +4024,7 @@ static struct pernet_operations ip6_route_net_late_ops = {
>  
>  static struct notifier_block ip6_route_dev_notifier = {
>  	.notifier_call = ip6_route_dev_notify,
> -	.priority = 0,
> +	.priority = -10, /* Must be called after addrconf_notify!! */
>  };
>  
>  void __init ip6_route_init_special_entries(void)
> 

That should be codified such that if someone changes the priority of the
addrconf notifier, it fixes up the route notifier as well. e.g.,

diff --git a/include/net/addrconf.h b/include/net/addrconf.h
index 1aeb25dd42a7..6c0ee3ccbe0f 100644
--- a/include/net/addrconf.h
+++ b/include/net/addrconf.h
@@ -20,6 +20,8 @@
 #define ADDRCONF_TIMER_FUZZ            (HZ / 4)
 #define ADDRCONF_TIMER_FUZZ_MAX                (HZ)
+#define ADDRCONF_NOTIFY_PRIORITY       0
+
 #include <linux/in.h>
 #include <linux/in6.h>

diff --git a/net/ipv6/addrconf.c b/net/ipv6/addrconf.c
index a2a370b71249..7a4e6b2996a1 100644
--- a/net/ipv6/addrconf.c
+++ b/net/ipv6/addrconf.c
@@ -3548,6 +3548,7 @@ static int addrconf_notify(struct notifier_block
*this, unsigned long event,
  */
 static struct notifier_block ipv6_dev_notf = {
        .notifier_call = addrconf_notify,
+       .priority      = ADDRCONF_NOTIFY_PRIORITY,
 };

 static void addrconf_type_change(struct net_device *dev, unsigned long
event)
diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index a1bf426c959b..e0d83f8bc01a 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -4024,7 +4024,7 @@ static struct pernet_operations
ip6_route_net_late_ops = {

 static struct notifier_block ip6_route_dev_notifier = {
        .notifier_call = ip6_route_dev_notify,
-       .priority = 0,
+       .priority = ADDRCONF_NOTIFY_PRIORITY - 10,
 };


And you marked this as a patch for net; I think this should go into
net-next.

Re: [Patch net] ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf

[David Ahern]

On 5/3/17 11:07 PM, Cong Wang wrote:
> For each netns (except init_net), we initialize its null entry
> in 3 places:
> 
> 1) The template itself, as we use kmemdup()
> 2) Code around dst_init_metrics() in ip6_route_net_init()
> 3) ip6_route_dev_notify(), which is supposed to initialize it after
> loopback registers
> 
> Unfortunately the last one still happens in a wrong order because
> we expect to initialize net->ipv6.ip6_null_entry->rt6i_idev to
> net->loopback_dev's idev, so we have to do that after we add
> idev to it. However, this notifier has priority == 0 same as
> ipv6_dev_notf, and ipv6_dev_notf is registered after
> ip6_route_dev_notifier so it is called actually after
> ip6_route_dev_notifier.
> 
> Fix it by specifying a smaller priority for ip6_route_dev_notifier.
> 
> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
> ---
>  net/ipv6/route.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> index 2f11366..4dbf7e2 100644
> --- a/net/ipv6/route.c
> +++ b/net/ipv6/route.c
> @@ -4024,7 +4024,7 @@ static struct pernet_operations ip6_route_net_late_ops = {
>  
>  static struct notifier_block ip6_route_dev_notifier = {
>  	.notifier_call = ip6_route_dev_notify,
> -	.priority = 0,
> +	.priority = -10, /* Must be called after addrconf_notify!! */
>  };
>  
>  void __init ip6_route_init_special_entries(void)
> 

And I see a refcnt problem with this change:

root@kenny-jessie2:~# unshare -n
root@kenny-jessie2:~# logout
root@kenny-jessie2:~# unshare -n

Message from syslogd@kenny-jessie2 at May  4 07:04:38 ...
 kernel:[   62.581552] unregister_netdevice: waiting for lo to become
free. Usage count = 1

Re: [Patch net] ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf

[Cong Wang]

On Thu, May 4, 2017 at 7:04 AM, David Ahern <dsahern@gmail.com> wrote:
> On 5/3/17 11:07 PM, Cong Wang wrote:
>> For each netns (except init_net), we initialize its null entry
>> in 3 places:
>>
>> 1) The template itself, as we use kmemdup()
>> 2) Code around dst_init_metrics() in ip6_route_net_init()
>> 3) ip6_route_dev_notify(), which is supposed to initialize it after
>> loopback registers
>>
>> Unfortunately the last one still happens in a wrong order because
>> we expect to initialize net->ipv6.ip6_null_entry->rt6i_idev to
>> net->loopback_dev's idev, so we have to do that after we add
>> idev to it. However, this notifier has priority == 0 same as
>> ipv6_dev_notf, and ipv6_dev_notf is registered after
>> ip6_route_dev_notifier so it is called actually after
>> ip6_route_dev_notifier.
>>
>> Fix it by specifying a smaller priority for ip6_route_dev_notifier.
>>
>> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
>> ---
>>  net/ipv6/route.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
>> index 2f11366..4dbf7e2 100644
>> --- a/net/ipv6/route.c
>> +++ b/net/ipv6/route.c
>> @@ -4024,7 +4024,7 @@ static struct pernet_operations ip6_route_net_late_ops = {
>>
>>  static struct notifier_block ip6_route_dev_notifier = {
>>       .notifier_call = ip6_route_dev_notify,
>> -     .priority = 0,
>> +     .priority = -10, /* Must be called after addrconf_notify!! */
>>  };
>>
>>  void __init ip6_route_init_special_entries(void)
>>
>
> And I see a refcnt problem with this change:
>
> root@kenny-jessie2:~# unshare -n
> root@kenny-jessie2:~# logout
> root@kenny-jessie2:~# unshare -n
>
> Message from syslogd@kenny-jessie2 at May  4 07:04:38 ...
>  kernel:[   62.581552] unregister_netdevice: waiting for lo to become
> free. Usage count = 1

Ah, looks like we need to put the refcnt for UNREGISTER too.

Will send v2 to include your ADDRCONF_NOTIFY_PRIORITY suggestion.