[Patch net] ipv6: only call ip6_route_dev_notify() once for NETDEV_UNREGISTER

[Patch net] ipv6: only call ip6_route_dev_notify() once for NETDEV_UNREGISTER

[Cong Wang]

In commit 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
I assumed NETDEV_REGISTER and NETDEV_UNREGISTER are paired,
unfortunately, as reported by jeffy, netdev_wait_allrefs()
could rebroadcast NETDEV_UNREGISTER event until all refs are
gone.

We have to add an additional check to avoid this corner case.
For netdev_wait_allrefs() dev->reg_state is NETREG_UNREGISTERED,
for dev_change_net_namespace(), dev->reg_state is
NETREG_REGISTERED. So check for dev->reg_state != NETREG_UNREGISTERED.

Fixes: 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
Reported-by: jeffy <jeffy.chen@rock-chips.com>
Cc: David Ahern <dsahern@gmail.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
---
 net/ipv6/route.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index 7cebd95..322bd62 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -3722,7 +3722,11 @@ static int ip6_route_dev_notify(struct notifier_block *this,
 		net->ipv6.ip6_blk_hole_entry->dst.dev = dev;
 		net->ipv6.ip6_blk_hole_entry->rt6i_idev = in6_dev_get(dev);
 #endif
-	 } else if (event == NETDEV_UNREGISTER) {
+	 } else if (event == NETDEV_UNREGISTER &&
+		    dev->reg_state != NETREG_UNREGISTERED) {
+		/* NETDEV_UNREGISTER could be fired for multiple times by
+		 * netdev_wait_allrefs(). Make sure we only call this once.
+		 */
 		in6_dev_put(net->ipv6.ip6_null_entry->rt6i_idev);
 #ifdef CONFIG_IPV6_MULTIPLE_TABLES
 		in6_dev_put(net->ipv6.ip6_prohibit_entry->rt6i_idev);
-- 
2.5.5

Re: [Patch net] ipv6: only call ip6_route_dev_notify() once for NETDEV_UNREGISTER

[jeffy]

Hi Cong Wang,

i don't know much about net core, maybe i'm misreading the code...but

On 06/21/2017 02:42 AM, Cong Wang wrote:
> In commit 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
> I assumed NETDEV_REGISTER and NETDEV_UNREGISTER are paired,
> unfortunately, as reported by jeffy, netdev_wait_allrefs()
> could rebroadcast NETDEV_UNREGISTER event until all refs are
> gone.
>
> We have to add an additional check to avoid this corner case.
> For netdev_wait_allrefs() dev->reg_state is NETREG_UNREGISTERED,
> for dev_change_net_namespace(), dev->reg_state is
> NETREG_REGISTERED. So check for dev->reg_state != NETREG_UNREGISTERED.

i saw we are calling NETDEV_REGISTER in these cases:
1/ register_netdevice_notifier:
the paired unregister would be:
a) normal unregister: rollback_registered_many
b) the error path:
register_netdevice_notifier->rollback jump label

2/ register_netdevice:
the paired unregister would both be rollback_registered_many for 
normal/error cases

3/ dev_change_net_namespace:
the paired unregister is the one right before the register notify

i think we are handling all register notifies, but only unregister 
notify from rollback_registered_many now?


>
> Fixes: 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
> Reported-by: jeffy <jeffy.chen@rock-chips.com>
> Cc: David Ahern <dsahern@gmail.com>
> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
> ---
>   net/ipv6/route.c | 6 +++++-
>   1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> index 7cebd95..322bd62 100644
> --- a/net/ipv6/route.c
> +++ b/net/ipv6/route.c
> @@ -3722,7 +3722,11 @@ static int ip6_route_dev_notify(struct notifier_block *this,
>   		net->ipv6.ip6_blk_hole_entry->dst.dev = dev;
>   		net->ipv6.ip6_blk_hole_entry->rt6i_idev = in6_dev_get(dev);
>   #endif
> -	 } else if (event == NETDEV_UNREGISTER) {
> +	 } else if (event == NETDEV_UNREGISTER &&
> +		    dev->reg_state != NETREG_UNREGISTERED) {
> +		/* NETDEV_UNREGISTER could be fired for multiple times by
> +		 * netdev_wait_allrefs(). Make sure we only call this once.
> +		 */
>   		in6_dev_put(net->ipv6.ip6_null_entry->rt6i_idev);
>   #ifdef CONFIG_IPV6_MULTIPLE_TABLES
>   		in6_dev_put(net->ipv6.ip6_prohibit_entry->rt6i_idev);
>

Re: [Patch net] ipv6: only call ip6_route_dev_notify() once for NETDEV_UNREGISTER

[jeffy]

Hi Cong Wang,

oh, oops, i did misread.

also, Tested-by: Jeffy Chen <jeffy.chen@rock-chips.com>

On 06/21/2017 11:01 AM, jeffy wrote:
> Hi Cong Wang,
>
> i don't know much about net core, maybe i'm misreading the code...but
>
> On 06/21/2017 02:42 AM, Cong Wang wrote:
>> In commit 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after
>> ipv6_dev_notf")
>> I assumed NETDEV_REGISTER and NETDEV_UNREGISTER are paired,
>> unfortunately, as reported by jeffy, netdev_wait_allrefs()
>> could rebroadcast NETDEV_UNREGISTER event until all refs are
>> gone.
>>
>> We have to add an additional check to avoid this corner case.
>> For netdev_wait_allrefs() dev->reg_state is NETREG_UNREGISTERED,
>> for dev_change_net_namespace(), dev->reg_state is
>> NETREG_REGISTERED. So check for dev->reg_state != NETREG_UNREGISTERED.
>
> i saw we are calling NETDEV_REGISTER in these cases:
> 1/ register_netdevice_notifier:
> the paired unregister would be:
> a) normal unregister: rollback_registered_many
> b) the error path:
> register_netdevice_notifier->rollback jump label
>
> 2/ register_netdevice:
> the paired unregister would both be rollback_registered_many for
> normal/error cases
>
> 3/ dev_change_net_namespace:
> the paired unregister is the one right before the register notify
>
> i think we are handling all register notifies, but only unregister
> notify from rollback_registered_many now?
>
you are checking for NETREG_UNREGISTERED, so only filter out the 
unregistered after rollback_registered_many, so that indeed covers all 
cases :)
>
>>
>> Fixes: 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after
>> ipv6_dev_notf")
>> Reported-by: jeffy <jeffy.chen@rock-chips.com>
>> Cc: David Ahern <dsahern@gmail.com>
>> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
>> ---
>>   net/ipv6/route.c | 6 +++++-
>>   1 file changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
>> index 7cebd95..322bd62 100644
>> --- a/net/ipv6/route.c
>> +++ b/net/ipv6/route.c
>> @@ -3722,7 +3722,11 @@ static int ip6_route_dev_notify(struct
>> notifier_block *this,
>>           net->ipv6.ip6_blk_hole_entry->dst.dev = dev;
>>           net->ipv6.ip6_blk_hole_entry->rt6i_idev = in6_dev_get(dev);
>>   #endif
>> -     } else if (event == NETDEV_UNREGISTER) {
>> +     } else if (event == NETDEV_UNREGISTER &&
>> +            dev->reg_state != NETREG_UNREGISTERED) {
>> +        /* NETDEV_UNREGISTER could be fired for multiple times by
>> +         * netdev_wait_allrefs(). Make sure we only call this once.
>> +         */
>>           in6_dev_put(net->ipv6.ip6_null_entry->rt6i_idev);
>>   #ifdef CONFIG_IPV6_MULTIPLE_TABLES
>>           in6_dev_put(net->ipv6.ip6_prohibit_entry->rt6i_idev);
>>
>

Re: [Patch net] ipv6: only call ip6_route_dev_notify() once for NETDEV_UNREGISTER

[David Ahern]

On 6/20/17 2:42 PM, Cong Wang wrote:
> In commit 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
> I assumed NETDEV_REGISTER and NETDEV_UNREGISTER are paired,
> unfortunately, as reported by jeffy, netdev_wait_allrefs()
> could rebroadcast NETDEV_UNREGISTER event until all refs are
> gone.
> 
> We have to add an additional check to avoid this corner case.
> For netdev_wait_allrefs() dev->reg_state is NETREG_UNREGISTERED,
> for dev_change_net_namespace(), dev->reg_state is
> NETREG_REGISTERED. So check for dev->reg_state != NETREG_UNREGISTERED.
> 
> Fixes: 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
> Reported-by: jeffy <jeffy.chen@rock-chips.com>
> Cc: David Ahern <dsahern@gmail.com>
> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
> ---
>  net/ipv6/route.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)

Acked-by: David Ahern <dsahern@gmail.com>

Re: [Patch net] ipv6: only call ip6_route_dev_notify() once for NETDEV_UNREGISTER

[David Miller]

From: Cong Wang <xiyou.wangcong@gmail.com>
Date: Tue, 20 Jun 2017 11:42:27 -0700

> In commit 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
> I assumed NETDEV_REGISTER and NETDEV_UNREGISTER are paired,
> unfortunately, as reported by jeffy, netdev_wait_allrefs()
> could rebroadcast NETDEV_UNREGISTER event until all refs are
> gone.
> 
> We have to add an additional check to avoid this corner case.
> For netdev_wait_allrefs() dev->reg_state is NETREG_UNREGISTERED,
> for dev_change_net_namespace(), dev->reg_state is
> NETREG_REGISTERED. So check for dev->reg_state != NETREG_UNREGISTERED.
> 
> Fixes: 242d3a49a2a1 ("ipv6: reorder ip6_route_dev_notifier after ipv6_dev_notf")
> Reported-by: jeffy <jeffy.chen@rock-chips.com>
> Cc: David Ahern <dsahern@gmail.com>
> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>

Applied.