Re: [PATCH net] dev_forward_skb: do not scrub skb mark within the same name space

[Nicolas Dichtel <nicolas.dichtel@6wind.com>]

Le 25/06/2021 à 10:45, Eyal Birger a écrit :
> 
> 
> On 24/06/2021 18:26, Nicolas Dichtel wrote:
>> Hi Eyal,
>>
>> Le 24/06/2021 à 14:16, Eyal Birger a écrit :
>>> Hi Nicholas,
>>>
>>> On 24/06/2021 11:05, Nicolas Dichtel wrote:
>>>> The goal is to keep the mark during a bpf_redirect(), like it is done for
>>>> legacy encapsulation / decapsulation, when there is no x-netns.
>>>> This was initially done in commit 213dd74aee76 ("skbuff: Do not scrub skb
>>>> mark within the same name space").
>>>>
>>>> When the call to skb_scrub_packet() was added in dev_forward_skb() (commit
>>>> 8b27f27797ca ("skb: allow skb_scrub_packet() to be used by tunnels")), the
>>>> second argument (xnet) was set to true to force a call to skb_orphan(). At
>>>> this time, the mark was always cleanned up by skb_scrub_packet(), whatever
>>>> xnet value was.
>>>> This call to skb_orphan() was removed later in commit
>>>> 9c4c325252c5 ("skbuff: preserve sock reference when scrubbing the skb.").
>>>> But this 'true' stayed here without any real reason.
>>>>
>>>> Let's correctly set xnet in ____dev_forward_skb(), this function has access
>>>> to the previous interface and to the new interface.
>>>
>>> This change was suggested in the past [1] and I think one of the main concerns
>>> was breaking existing callers which assume the mark would be cleared [2].
>> Thank you for the pointers!
>>
>>>
>>> Personally, I think the suggestion made in [3] adding a flag to bpf_redirect()
>>> makes a lot of sense for this use case.
>> I began with this approach, but actually, as I tried to explain in the commit
>> log, this looks more like a bug. This function is called almost everywhere in
>> the kernel (except for openvswitch and wireguard) with the xnet argument
>> reflecting a netns change. In other words, the behavior is different between
>> legacy encapsulation and the one made with ebpf :/
>>
> 
> I agree, and was also surprised that ebpf redirection scrubs the mark in the
> same ns - and only on ingress! - so I think keeping the mark should have been
Yes, there is also this asymmetry :/

> the default behavior. As noted in the thread though it is not clear whether this
> would break existing deployments both for ebpf and veth pairs on the same ns.
To summarize:
 - it's not consistent between legacy tunnels and ebpf;
 - it's not consistent between ingress and egress;
 - it was already fixed in the past for legacy tunnel (commit 213dd74aee76);
 - this kind of patch has already been done in the past, see commit
   + 963a88b31ddb ("tunnels: harmonize cleanup done on skb on xmit path");
   + ea23192e8e57 ("tunnels: harmonize cleanup done on skb on rx path");
 - nobody stated (on netdev at least) that he relied on this behavior.

For all those reasons, I'm inclined to think that this patch is acceptable.
Obviously, it's not my decision ;-)

Note also that clearing the mark is possible directly in the ebpf prog,
preserving it not. It means that, in the worst case of someone relying on this,
the bug can be fixed without patching the kernel. A trade-off could be to reset
the mark in the veth case only.

Side note: the mark was not cleared (for veth) between 2.6.34 and 3.4, see
commit 59b9997baba5 ("Revert "net: maintain namespace isolation between vlan and
real device"") and the revert was not done due to this ;-)

Comments are welcomed.

Regards,
Nicolas