On 5/15/25 6:24 PM, Han Zhou wrote:
> On Fri, Aug 30, 2024 at 1:04 PM Tim Rozet <tro...@redhat.com> wrote:
>>
>> Hi Han,
>> My understanding is that the unSNAT stage was originally being skipped
> for LB ports, then Dumitru changed it to attempt unSNAT, and this breaks
> HWOL for ingress connections that never need to get unSNAT'ed and therefore
> not committed. In the OVNK use case for pods simply connecting egress,
> almost all packets will be SNAT'ed to the node IP. Therefore on egress
> reply, we always enter the unSNAT stage for these packets. So this skipping
> unSNAT stage seems to be specific to LB related traffic. While I agree
> splitting the port range would work, I think it makes more sense to just
> always commit in the unSNAT stage for all traffic. I get there is a
> performance hit there, but the datapath pipeline seems more consistent and
> I think outweighs the cost of committing LB traffic.
>
> Hi Tim and folks,
>
Hi Han,
> + @Ales Musil <amu...@redhat.com>
>
> We've spent a lot of effort on the alternative solution: Always commit to
> both SNAT and DNAT zones. Unfortunately, the HW offload issue discussed in
> this thread is still unresolved. I'd like to discuss the current situation
> and ask for opinions.
>
Thanks again for working on this!
> A while ago, Ales tried to fix it with this commit: 800fd0681579 ("northd:
> Add LR option to commit all traffic."). When the ct-commit-all option is
> enabled, it commits the connection to both SNAT and DNAT zones. However, in
> the ovn-k8s nodeport scenario (a very typical use case) the HW offload is
> still broken. In that scenario, a packet may come from external network
> with dest IP being a node's IP, which is also used as an OVN gateway router
> IP. The dest IP is firstly DNATed to a LB backend IP, and then the src IP
> is SNATed to another LRP's IP on this router. The src/dst IPs in different
> stages are:
>
> Original: EXT IP <-> GR IP1
> After DNAT: EXT IP <-> BE IP
> After SNAT: GR IP2 <-> BE IP
>
> With the ct-commit-all option, it commits after DNAT and after SNAT:
> DNAT zone: EXT IP <-> GR IP1/BE IP
> SNAT zone: EXT IP/GR IP2 <-> BE IP
>
> So, at the UNSNAT stage, the packet traverses the SNAT zone with the
> original header, which will still always be "new" state, which breaks HW
> offload.
>
Ah, now I understand the problem you were facing.
> I tried to fix it by also committing to the SNAT zone before DNAT (at the
> POST_SNAT stage). Now there are 3 entries committed:
> SNAT zone: EXT IP <-> GR IP1
> DNAT zone: EXT IP <-> GR IP1/BE IP
> SNAT zone: EXT IP/GR IP2 <-> BE IP
>
> HW offload worked, but it breaks at least the scenario of "DNAT and SNAT on
> distributed router - E/W", which tests communication between two workloads
> using "floating IPs". In this scenario, both workloads are under different
> LSes behind the same router, and both have their own floating IPs
> (dnat_and_snat) on the public side of the LR. When one of them sends
> packets to the other using the floating IP, what's expected is:
>
> Original: overlay IP1 <-> floating IP2
> After SNAT: floating IP1 <-> floating IP2
> After DNAT: floating IP1 <-> overlay IP2
>
> With my fix, the packet is firstly SNATed with CT entry in SNAT zone:
> SNAT: overlay IP1/floating IP1 <-> floating IP2
>
> This happens in the egress pipeline of the LR, and then it enters into the
> ingress pipeline again because this is like a hairpin (the dest IP is a
> dnat_and_snat IP on the same router). It will hit the UNSNAT stage first,
> traverse the CT zone, and for the first packet the reply is not seen yet,
> so the CT state is still new, so in the POST_UNSNAT we will try to commit
> it (because of the issue I am trying to fix above). Because the connection
> floating IP1 <-> floating IP2 is conflicting with the above committed entry
> "SNAT: overlay IP1/floating IP1 <-> floating IP2", the CT action fails and
> the packet is dropped.
>
> There is still a way to solve this, probably by skipping the unnecessary
> UNSNAT traversal and the conflict commit if the source IP is one of the
> snat/dnat_and_snat IPs of this same router.
>
> However, as we see, the NAT related pipelines have been very complex today,
> and with the above changes it would become even much more complex. I wonder
> if this is maintainable in the long run.
>
> In addition, I am not even sure if this is the last problem of my patch. It
> is not easy to tell because today the ct-commit-all feature is an option
> and by default disabled. So all our test cases by default don't test for
> this option except the 4 test cases added by Ales's patch. I had to
> manually set this option to true as default and of course many test cases
> will fail. Most of them fail because of the expected flows won't match, but
> I can't tell for sure unless updating every case with new flows and then
> retest them. Before doing that I just examined ones I think might be
> impacted by the patch and the "DNAT and SNAT on distributed router - E/W"
> is the first broken one I found. So there can be other issues not
> discovered yet.
>
> Moreover, we should also keep in mind the performance penalty of these
> extra CT commit & traversal.
>
But basically this happens because we use the same SNAT zone for all
router ports.
In an ideal world wouldn't using a unique CT SNAT and CT DNAT zone per
router port fix it? E.g.:
host1 ---- LRP1 (SNAT ZONE: 11, DNAT ZONE:12) [LR] LRP2 (SNAT ZONE: 21,
DNAT ZONE: 22) ---- host2
That'd also imply that we commit that session (from host1 to host2 -
potentially via LB and SNAT) to 4 different zones (SNAT and DNAT for
LRP1 and SNAT and DNAT for LRP2). And we'd also have to do 4 different
CT lookups for all packets traversing the router.
I didn't try any of this but I have the impression that this would solve
the HWOL issue: all subsequent packets are part of known and committed
sessions in all 4 conntrack zones.
So what if we change the implementation of the "commit-all" feature to
add the missing ct commits? Wouldn't that work?
It feels like (lookup performance hit aside) this is the correct way to
forward traffic, i.e., never traverse a CT zone without having a
committed entry in that zone.
> I'd like to listen to your opinion before moving forward. Is this worth the
> effort, or shall we just keep the pipeline simple (relatively) and avoid
> the original problem (https://issues.redhat.com/browse/FDP-291) by
> separating the port range of SNAT and DNAT, the first proposal at the
> beginning of this thread?
>
Now that I understand the problem you're describing better it might be
worth reverting the fix for FDP-291 like you said until we find a proper
generic solution.
However, we can't afford to do that without separating the port range of
SNAT and DNAT like you said [0] because we'd be regressing (at least
from OpenShift perspective).
Because selecting the nodeport (DNAT) range is done through a k8s API
server cmdline argument that means that users can in theory choose
whatever range they want. I guess we'd need a command line argument in
ovn-kubernetes to allow users (operators, etc.) to inform ovn-kubernetes
what SNAT port range to use so that it doesn't conflict with the KAPI
nodeport range.
Tim, you're the expert on this side, what are your thoughts on the matter?
Regards,
Dumitru
[0] https://mail.openvswitch.org/pipermail/ovs-dev/2024-August/416974.html
> Thanks,
> Han
>
>>
>> Thanks,
>> Tim Rozet
>> Red Hat OpenShift Networking Team
>>
>>
>> On Fri, Aug 30, 2024 at 5:26 AM Dumitru Ceara <dce...@redhat.com> wrote:
>>>
>>> On 8/29/24 18:14, Han Zhou wrote:
>>>> On Wed, Mar 6, 2024 at 11:13 AM Dumitru Ceara <dce...@redhat.com>
> wrote:
>>>>>
>>>>> On 3/5/24 15:56, Numan Siddique wrote:
>>>>>> On Mon, Feb 26, 2024 at 7:59 AM Dumitru Ceara <dce...@redhat.com>
> wrote:
>>>>>>>>>>>>> Otherwise, in case there's also a SNAT rule that uses the
> VIP as
>>>>>>> external IP, we break sessions initiated from behind the VIP.
>>>>>>>
>>>>>>> This partially reverts 832893bdbb42 ("ovn-northd: Skip unsnat flows
> for
>>>>>>> load balancer vips in router ingress pipeline"). That's OK because
>>>>>>> commit 384a7c6237da ("northd: Refactor Logical Flows for routers
> with
>>>>>>> DNAT/Load Balancers") addressed the original issue in a better way:
>>>>>>>
>>>>>>> In the reply direction, the order of traversal of the tables
>>>>>>> "lr_in_defrag", "lr_in_unsnat" and "lr_in_dnat" adds incorrect
>>>>>>> datapath flows that check ct_state in the wrong conntrack zone.
>>>>>>> This is illustrated below where reply trafic enters the
> physical host
>>>>>>> port (6) and traverses DNAT zone (14), SNAT zone (default),
> back to the
>>>>>>> DNAT zone and then on to Logical Switch Port zone (22). The
> third
>>>>>>> flow is incorrectly checking the state from the SNAT zone
> instead
>>>>>>> of the DNAT zone.
>>>>>>>
>>>>>>> We also add a system test to ensure traffic initiated from behind a
> VIP
>>>>>>> + SNAT is not broken.
>>>>>>>
>>>>>>> Another nice side effect is that the northd I-P is slightly
> simplified
>>>>>>> because we don't need to track NAT external IPs anymore.
>>>>>>>
>>>>>>> Fixes: 832893bdbb42 ("ovn-northd: Skip unsnat flows for load
> balancer vips in router ingress pipeline")
>>>>>>> Reported-at: https://issues.redhat.com/browse/FDP-291
>>>>>>> Signed-off-by: Dumitru Ceara <dce...@redhat.com>
>>>>>>
>>>>>>
>>>>>> Thanks for the fix. It also simplified the lr-nat-stateful code.
>>>>>>
>>>>>> Acked-by: Numan Siddique <num...@ovn.org>
>>>>>>
>>>>>
>>>>> Thanks, Numan!
>>>>>
>>>>> Applied to main and backported to all branches down to 22.03.
>>>>>
>>>>> Regards,
>>>>> Dumitru
>>>>>
>>>>> _______________________________________________
>>>>> dev mailing list
>>>>> d...@openvswitch.org
>>>>> https://mail.openvswitch.org/mailman/listinfo/ovs-dev
>>>>
>>>> Hi Dumitru, Numan, Tim and folks,
>>>>
>>>
>>> Hi Han,
>>>
>>>> I noticed that the HW offload of k8s nodePort traffic is broken due to
>>>> this change. The reason is that for client to nodePort (LB with VIP
>>>> being the node IP) traffic, when the packet is going through the
>>>> unSNAT stage in the SNAT CT zone, since the entry is never committed
>>>> to the SNAT zone, it will have CT state returned as "new", which
>>>> prevents the HW offload to work for such packets.
>>>>
>>>
>>> Sorry about that, I forgot we don't always commit in the SNAT zone.
>>>
>>>> At the moment I have to revert this change in our downstream. For the
>>>> problem that was fixed by this change [0], I think we can avoid it by
>>>> separating the port range of SNAT and DNAT. For DNAT, the nodePort
>>>> range in k8s is configured by API-server option:
>>>>
>>>> --service-node-port-range <a string in the form 'N1-N2'> Default:
> 30000-32767
>>>>
>>>> For SNAT, it can be configured in the OVN's NAT table's
>>>> external_port_range column, and we can choose something like
>>>> 10000-30000.
>>>>
>>>
>>> Tim, does this look OK to you? If it's acceptable to limit the SNAT
>>> port range this workaround should be fine.
>>>
>>>> An extra benefit of this is that it reduces a CT recirc.
>>>>
>>>> Alternatively solutions are:
>>>> Alternative1: Always commit to both SNAT and DNAT zones. This would
>>>> introduce unnecessary cost of CT entries and extra CT recirc.
>>>
>>> Extra recirculation aside, I would actually love it if we could use this
>>> alternative. I think it's the "most correct" option. I think it would
>>> allow us to avoid other workarounds like:
>>>
>>> https://github.com/ovn-org/ovn/commit/40136a2f2c8
>>>
>>> or
>>>
>>>
> https://patchwork.ozlabs.org/project/ovn/patch/20240827085252.458355-1-amu...@redhat.com/
>>>
>>> I do understand the worry about the extra recirculation though. In the
>>> HWOL context does that cause visible performance impact?
>>>
>>> We'd probably have to do more performance testing without HWOL to figure
>>> out the impact in the software datapath.
>>>
>>>> Alternative2: Use a common zone for SNAT and DNAT. But there are other
>>>> issues reported for using the common zone [1]
>>>>
>>>> Could you let me know if other thoughts on this?
>>>>
>>>
>>> On a related note, I know it has been discussed in different settings
>>> but I don't think this ever moved forward: would it be possible for
>>> NVIDIA to help out with automatically testing HWOL impact for incoming
>>> patches?
>>>
>>> Maybe we could some "simple" system-like tests that ensure that traffic
>>> is correctly offloaded in common scenarios? Alternatively, I guess we
>>> could also tag a subset of the existing system tests and just run those
>>> on actual hardware?
>>>
>>> It's quite simple (AFAIU) for external CIs to report status on each OVN
>>> patch posted on patchwork. That would at least allow us to flag this
>>> kind of breakages early (even before they get merged).
>>>
>>> What do you think?
>>>
>>>> [0] https://issues.redhat.com/browse/FDP-291
>>>> [1] b8c40e7593
> https://github.com/ovn-org/ovn/commit/b8c40e7593a9fa40a057268c507a912d67b99ec4
>>>>
>>>> Thanks,
>>>> Han
>>>>
>>>
>>> Thanks,
>>> Dumitru
> _______________________________________________
> dev mailing list
> d...@openvswitch.org
> https://mail.openvswitch.org/mailman/listinfo/ovs-dev
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
