Hi guys,
Now, lxc created with veth can not be under control by
cls_cgroup.
the former discussion:
http://lkml.indiana.edu/hypermail/linux/kernel/1312.1/00214.html
In short, because cls_cgroup relys classid attached to sock
filter skb, but sock will be cleared inside dev_forward_skb()
in veth_xmit().
so I add backup_classid in struct sk_buffer to save classid
before dev_forward_skb(). In cls_cgroup_classify(), if skb->sk
is NULL, we can try to restore classid form skb->bk_classid.
Libo Chen (4):
net: introduce bk_classid to struct sk_buff
cls_cgroup: introduce a helper: bk_cls_classid()
veth: backup classid befor switch net_ns
cls_cgroup: restore classid from skb->sk_classid
drivers/net/veth.c | 5 +++++
include/linux/skbuff.h | 3 +++
include/net/cls_cgroup.h | 11 +++++++++++
net/sched/cls_cgroup.c | 8 ++++----
4 files changed, 23 insertions(+), 4 deletions(-)
On Thu, Jan 2, 2014 at 7:11 PM, Libo Chen <[email protected]> wrote:
> Hi guys,
>
> Now, lxc created with veth can not be under control by
> cls_cgroup.
>
> the former discussion:
> http://lkml.indiana.edu/hypermail/linux/kernel/1312.1/00214.html
>
> In short, because cls_cgroup relys classid attached to sock
> filter skb, but sock will be cleared inside dev_forward_skb()
> in veth_xmit().
So what are you trying to achieve here?
On 2014/1/3 13:20, Cong Wang wrote:
> On Thu, Jan 2, 2014 at 7:11 PM, Libo Chen <[email protected]> wrote:
>> Hi guys,
>>
>> Now, lxc created with veth can not be under control by
>> cls_cgroup.
>>
>> the former discussion:
>> http://lkml.indiana.edu/hypermail/linux/kernel/1312.1/00214.html
>>
>> In short, because cls_cgroup relys classid attached to sock
>> filter skb, but sock will be cleared inside dev_forward_skb()
>> in veth_xmit().
>
>
> So what are you trying to achieve here?
sys container using veth can be controlled by cls_cgroup basing on physic network interface
thanks,
Libo
>
> .
>
On 01/06/2014 03:54 PM, Libo Chen wrote:
> On 2014/1/3 13:20, Cong Wang wrote:
>> On Thu, Jan 2, 2014 at 7:11 PM, Libo Chen <[email protected]> wrote:
>>> Hi guys,
>>>
>>> Now, lxc created with veth can not be under control by
>>> cls_cgroup.
>>>
>>> the former discussion:
>>> http://lkml.indiana.edu/hypermail/linux/kernel/1312.1/00214.html
>>>
>>> In short, because cls_cgroup relys classid attached to sock
>>> filter skb, but sock will be cleared inside dev_forward_skb()
>>> in veth_xmit().
>>
>>
>> So what are you trying to achieve here?
>
> sys container using veth can be controlled by cls_cgroup basing on physic network interface
>
It's a problem about virtual nic, not container/net namespace.
If veth device is running in host. the skb is transmitted firstly by veth device and then delivered
by physical device. if you set both qdisc rule on veth and physical device. which qdisc rule will take
effect?
In your patch, both qdisc rule are effective. it looks strange.
yes
On 2014/1/6 16:42, Gao feng wrote:
> On 01/06/2014 03:54 PM, Libo Chen wrote:
>> On 2014/1/3 13:20, Cong Wang wrote:
>>> On Thu, Jan 2, 2014 at 7:11 PM, Libo Chen <[email protected]> wrote:
>>>> Hi guys,
>>>>
>>>> Now, lxc created with veth can not be under control by
>>>> cls_cgroup.
>>>>
>>>> the former discussion:
>>>> http://lkml.indiana.edu/hypermail/linux/kernel/1312.1/00214.html
>>>>
>>>> In short, because cls_cgroup relys classid attached to sock
>>>> filter skb, but sock will be cleared inside dev_forward_skb()
>>>> in veth_xmit().
>>>
>>>
>>> So what are you trying to achieve here?
>>
>> sys container using veth can be controlled by cls_cgroup basing on physic network interface
>>
>
> It's a problem about virtual nic, not container/net namespace.
yes
>
> If veth device is running in host. the skb is transmitted firstly by veth device and then delivered
> by physical device. if you set both qdisc rule on veth and physical device. which qdisc rule will take
> effect?
both, the end result depends on a smaller.
>
> In your patch, both qdisc rule are effective. it looks strange.
>
qdisc is based nic, does not distinguish virtual or physics. if you are all set,
it means that what you want. so the logic is not the problemI and this appears to be normal.
thanks,
Libo
> .
>
Libo Chen <[email protected]> writes:
> yes
> On 2014/1/6 16:42, Gao feng wrote:
>> On 01/06/2014 03:54 PM, Libo Chen wrote:
>>> On 2014/1/3 13:20, Cong Wang wrote:
>>>> On Thu, Jan 2, 2014 at 7:11 PM, Libo Chen <[email protected]> wrote:
>>>>> Hi guys,
>>>>>
>>>>> Now, lxc created with veth can not be under control by
>>>>> cls_cgroup.
>>>>>
>>>>> the former discussion:
>>>>> http://lkml.indiana.edu/hypermail/linux/kernel/1312.1/00214.html
>>>>>
>>>>> In short, because cls_cgroup relys classid attached to sock
>>>>> filter skb, but sock will be cleared inside dev_forward_skb()
>>>>> in veth_xmit().
>>>>
>>>>
>>>> So what are you trying to achieve here?
>>>
>>> sys container using veth can be controlled by cls_cgroup basing on physic network interface
>>>
>>
>> It's a problem about virtual nic, not container/net namespace.
>
> yes
>
>>
>> If veth device is running in host. the skb is transmitted firstly by veth device and then delivered
>> by physical device. if you set both qdisc rule on veth and physical device. which qdisc rule will take
>> effect?
>
> both, the end result depends on a smaller.
>
>>
>> In your patch, both qdisc rule are effective. it looks strange.
>>
>
> qdisc is based nic, does not distinguish virtual or physics. if you are all set,
> it means that what you want. so the logic is not the problemI and this appears to be normal.
My personal opinion on the matter is that the network class cgroup is
brain dead and should not be used. It is impossible to use for
incomming packets, and it is part of the the problem plagued cgroup
subsystem.
You have real network interfaces to do your classification with you
don't need to enhance the network class cgroup.
The more this is asked about the more I think the network class cgroup
should be be taken out into the woods some dark night and left in a
shallow grave, never to bother us again.
Eric