Re: [PATCH] Traffic control cgroups subsystem

[Patrick McHardy <kaber@trash.net>]

Paul Menage wrote:
> On Tue, Jul 22, 2008 at 6:35 AM, Patrick McHardy <kaber@trash.net> wrote:
>> Does this really have to be a new skb member? You could
>> simply use skb->sk->sk_cgroup_classid directly, or if
>> that doesn't work, maybe skb->priority.
>>
> 
> We were actually using skb->priority in our internal version of this
> patch. I suggested that the separate cgroup_classid field be added
> since it might be considered an abuse of skb->priority and would
> interfere with existing users of that. If that's not an issue then
> reusing skb->priority is certainly possible.

Using skb->priority for classification would be fine, but it would
probably interfere with the default initialization to sk->sk_priority.

> Regarding skb->sk->sk_cgroup_classid, is it always the case that the
> original sk is still available when we're making traffic control
> decisions? I'd thought that there were cases (e.g. cloning skbs in the
> TCP retransmit path) where the pointer to the original sk is lost.

After cloning, TCP sets the owner of the skb to the socket, so
that should work fine.