cgroup filter on physics interface can't container

cgroup filter on physics interface can't container

[Libo Chen]

hello network hackers,

	A linux container was builded with veth pair(veth0 inside container, veth1 outside container),

the config as below:

  lxc.network.type = veth
  lxc.network.flags = up
  lxc.network.link = br0  // base on eth0
  lxc.network.name = eth0
  lxc.network.ipv4 = 128.5.130.26/24

then I use tc command with cgroup filter on veth0, it works well. But when setting on eth0, it doesn`t work.

The reason is dev_forward_skb() in veth_xmit will call skb_scrub_packet and clean all information including skb->sk
in the skb, so if cls_cgroup_classify is working in serving softirq state, it will return failer, see below:

	if (in_serving_softirq()) {
		/* If there is an sk_classid we'll use that. */
		if (!skb->sk)
			return -1;
		classid = skb->sk->sk_classid;
	}


Qdisc with cgroup filter on physics interface can not control a container network, it is disappointed.

we can save sk_classid before skb_scrub_packet and restore it after that. Is it reasonable?  or any way to achieve this?

thanks,
Libo

Re: cgroup filter on physics interface can't control container

[Libo Chen]

ping...

On 2013/12/9 10:32, Libo Chen wrote:
> hello network hackers,
> 
> 	A linux container was builded with veth pair(veth0 inside container, veth1 outside container),
> 
> the config as below:
> 
>   lxc.network.type = veth
>   lxc.network.flags = up
>   lxc.network.link = br0  // base on eth0
>   lxc.network.name = eth0
>   lxc.network.ipv4 = 128.5.130.26/24
> 
> then I use tc command with cgroup filter on veth0, it works well. But when setting on eth0, it doesn`t work.
> 
> The reason is dev_forward_skb() in veth_xmit will call skb_scrub_packet and clean all information including skb->sk
> in the skb, so if cls_cgroup_classify is working in serving softirq state, it will return failer, see below:
> 
> 	if (in_serving_softirq()) {
> 		/* If there is an sk_classid we'll use that. */
> 		if (!skb->sk)
> 			return -1;
> 		classid = skb->sk->sk_classid;
> 	}
> 
> 
> Qdisc with cgroup filter on physics interface can not control a container network, it is disappointed.
> 
> we can save sk_classid before skb_scrub_packet and restore it after that. Is it reasonable?  or any way to achieve this?
> 
> thanks,
> Libo
> 

Re: cgroup filter on physics interface can't control container

[Libo Chen]

hello guys,

I said before "we can save sk_classid before skb_scrub_packet and
restore it after that"  since skb->sk had been freed in
skb_scrub_packet(), so it is not reasonable. yes?

I have another idea.
commit:f84517253(cls_cgroup: Store classid in struct sock) introduces
sk_classid and put it in skb->sk pointer.  can we put sk_classid form
struct sock to struct sk_buff? then sk_classid will not be affected
by dev_forward_skb()->skb_scrub_packet() ?

any comment are welcome!


thanks,
Libo


On 2013/12/12 20:18, Libo Chen wrote:
> ping...
> 
> On 2013/12/9 10:32, Libo Chen wrote:
>> hello network hackers,
>>
>> 	A linux container was builded with veth pair(veth0 inside container, veth1 outside container),
>>
>> the config as below:
>>
>>   lxc.network.type = veth
>>   lxc.network.flags = up
>>   lxc.network.link = br0  // base on eth0
>>   lxc.network.name = eth0
>>   lxc.network.ipv4 = 128.5.130.26/24
>>
>> then I use tc command with cgroup filter on veth0, it works well. But when setting on eth0, it doesn`t work.
>>
>> The reason is dev_forward_skb() in veth_xmit will call skb_scrub_packet and clean all information including skb->sk
>> in the skb, so if cls_cgroup_classify is working in serving softirq state, it will return failer, see below:
>>
>> 	if (in_serving_softirq()) {
>> 		/* If there is an sk_classid we'll use that. */
>> 		if (!skb->sk)
>> 			return -1;
>> 		classid = skb->sk->sk_classid;
>> 	}
>>
>>
>> Qdisc with cgroup filter on physics interface can not control a container network, it is disappointed.
>>
>> we can save sk_classid before skb_scrub_packet and restore it after that. Is it reasonable?  or any way to achieve this?
>>
>> thanks,
>> Libo
>>
>