On 11/13/26 5:34 PM, Numan Siddique via dev wrote:
> On Tue, Nov 10, 2026 at 6:38 AM Takao Yatagai via dev
> <[email protected]> wrote:
>>
>> From: Takao Yatagai <[email protected]>
>>
>> A load balancer on a distributed router without a distributed gateway
>> port gets its lr_in_dnat flows on every chassis, but no un-DNAT: only
>> gateway routers and routers with a DGP get one.  Replies leave the
>> router with the backend as source.
>>
>> With options:symmetric_return=true, the CMS guarantees that replies
>> re-enter the router on the chassis that load balanced the request, for
>> example because a downstream router translates the source of the
>> requests with a distributed NAT rule (logical_port + external_mac).
>> northd then adds, on such a router, the same priority 120 lr_out_undnat
>> flow that routers with a DGP get, without the DGP and chassis terms.
>> OVN does not verify the guarantee: replies that reach another chassis
>> are not un-DNATed.
>>
>> Keeping the load balancer and the SNAT on different routers also keeps
>> the configuration compatible with NB_Global use_common_zone.
>>
>> The option has no effect on gateway routers and routers with a DGP.
>>
>> Assisted-by: Claude Opus 5.5
>> Signed-off-by: Takao Yatagai <[email protected]
> 
> 
> Hi Takao,
> 
> Thanks for the patch.  I've a few comments.
> 

Hi Takao, Numan,

> So presently if a distributed logical router with no gateway ports has
> a load balancer,
> as you pointed in the commit message,  it adds the logical flow in lr_in_dnat 
> to
> load balance the traffic destined to the VIP, but it adds no flow in
> lr_out_unndat.
> 
> So it breaks the load balancer feature anyway.  So I think we can just
> drop the option "symmetric_return"
> and add the lr_out_undnat flow regardless.
> 

I agree, the fewer config knobs we have, the better.

> If there is a distributed gateway port, the present code any way
> matches "inport" and "is_chassis_resident".
> 
> So I'd suggest not to match on "inport" and "is_chassis_resident" if a
> router has no gateway ports.
> 
> It would still be the responsibility of the CMS to ensure through the
> proper logical topology
> that the reply packet enters the same chassis.
> 
> @Dumitru Ceara @Ales Musil @Mark Michelson   Please let me know if
> there are any concerns here.
> 
> @Takao Yatagai  Also you  have to update the logical flow
> documentation here -
> https://github.com/ovn-org/ovn/blob/main/Documentation/ref/ovn-logical-flows.7.rst
> 

I think we need to document this behavior properly at a higher level too
though.  In theory ovn-architecture might be the right place but we
don't document load balancers much there.  Maybe a section in the ovn-nb
man page at least?

> 
> Thanks
> Numan
> 

Thanks,
Dumitru

>> ---
>>  NEWS                |  3 ++
>>  northd/lb.c         |  3 ++
>>  northd/lb.h         |  6 ++++
>>  northd/northd.c     |  9 ++++++
>>  ovn-nb.xml          | 30 ++++++++++++++++++++
>>  tests/ovn-northd.at | 67 +++++++++++++++++++++++++++++++++++++++++++++
>>  6 files changed, 118 insertions(+)
>>
>> diff --git a/NEWS b/NEWS
>> index 4176f8d7f..4e3f932d3 100644
>> --- a/NEWS
>> +++ b/NEWS
>> @@ -44,6 +44,9 @@ Post v26.09.0
>>       table at startup; it defers its kernel route updates until the initial
>>       dump of the advertised routes of the local route-exchange datapaths is
>>       complete.
>> +   - Added the "symmetric_return" option to the Load_Balancer table.  With
>> +     it, a distributed router without a distributed gateway port un-DNATs
>> +     the replies from the load balancer's backends.  See ovn-nb(5).
>>
>>  OVN v26.09.0 - xxx xx xxxx
>>  --------------------------
>> diff --git a/northd/lb.c b/northd/lb.c
>> index 5ff9d1fad..081aa6ded 100644
>> --- a/northd/lb.c
>> +++ b/northd/lb.c
>> @@ -392,6 +392,9 @@ ovn_northd_lb_init(struct ovn_northd_lb *lb,
>>      lb->is_distributed = smap_get_bool(&nbrec_lb->options, "distributed",
>>                                         false);
>>
>> +    lb->symmetric_return = smap_get_bool(&nbrec_lb->options,
>> +                                         "symmetric_return", false);
>> +
>>      sset_init(&lb->ips_v4);
>>      sset_init(&lb->ips_v6);
>>      struct smap_node *node;
>> diff --git a/northd/lb.h b/northd/lb.h
>> index 7a98c2f55..145d13dfa 100644
>> --- a/northd/lb.h
>> +++ b/northd/lb.h
>> @@ -81,6 +81,12 @@ struct ovn_northd_lb {
>>
>>      bool use_stateless_nat;
>>
>> +    /* Set by options:symmetric_return.  The CMS guarantees that replies to
>> +     * this load balancer re-enter the router on the chassis that load
>> +     * balanced the request, so a distributed router without a distributed
>> +     * gateway port un-DNATs them there. */
>> +    bool symmetric_return;
>> +
>>      char *hairpin_snat_ip;
>>  };
>>
>> diff --git a/northd/northd.c b/northd/northd.c
>> index 77c77f273..f88178301 100644
>> --- a/northd/northd.c
>> +++ b/northd/northd.c
>> @@ -14170,6 +14170,15 @@ build_lrouter_nat_flows_for_lb(
>>
>>          if (vector_is_empty(&od->l3dgw_ports)) {
>>              bitmap_set1(gw_dp_bitmap[type], index);
>> +            /* A distributed router without a DGP load balances on every
>> +             * chassis.  With symmetric_return, the CMS guarantees that the
>> +             * replies re-enter the router on the chassis that load balanced
>> +             * the request, so un-DNAT them there. */
>> +            if (!od->is_gw_router && lb->symmetric_return
>> +                && !vector_is_empty(&lb_vip->backends)) {
>> +                build_lrouter_lb_undnat_flow(&ctx, type, od, NULL, NULL,
>> +                                             lb_dps->lflow_ref);
>> +            }
>>          } else {
>>              /* Create stateless LB NAT rules when using multiple DGPs and
>>               * use_stateless_nat is true.
>> diff --git a/ovn-nb.xml b/ovn-nb.xml
>> index d8c63567a..2ee7f1dec 100644
>> --- a/ovn-nb.xml
>> +++ b/ovn-nb.xml
>> @@ -2839,6 +2839,36 @@ or
>>          router.
>>        </column>
>>
>> +      <column name="options" key="symmetric_return"
>> +              type='{"type": "boolean"}'>
>> +        <p>
>> +          On a distributed logical router with no distributed gateway port,
>> +          a load balancer is applied in <code>lr_in_dnat</code> on every
>> +          chassis, but the replies from its backends are not un-DNATed.
>> +          If set to <code>true</code>, such a router un-DNATs those replies
>> +          in its egress pipeline on the chassis where they pass it.  The
>> +          default is <code>false</code>.
>> +        </p>
>> +
>> +        <p>
>> +          The CMS must guarantee that replies to the load balanced traffic
>> +          re-enter the router on the chassis that load balanced the request,
>> +          for example because the source of the requests is translated by
>> +          a distributed NAT rule (see the <ref column="logical_port"
>> +          table="NAT"/> and <ref column="external_mac" table="NAT"/> columns
>> +          of the <ref table="NAT"/> table) on a downstream router whose
>> +          external IP is reachable only on that chassis.  OVN does not 
>> verify
>> +          this.  If replies reach another chassis, they are not un-DNATed.
>> +          Unlike <ref column="options" key="distributed"/>, this option does
>> +          not restrict the backends to logical ports on the local chassis.
>> +        </p>
>> +
>> +        <p>
>> +          This option has no effect on gateway routers and on routers with
>> +          a distributed gateway port.
>> +        </p>
>> +      </column>
>> +
>>        <column name="options" key="event" type='{"type": "boolean"}'>
>>          If set to <code>true</code> and the load balancer has no
>>          active backends, whenever a packet matches a VIP on this
>> diff --git a/tests/ovn-northd.at b/tests/ovn-northd.at
>> index 6c37337ad..a81a12fd3 100644
>> --- a/tests/ovn-northd.at
>> +++ b/tests/ovn-northd.at
>> @@ -25340,3 +25340,70 @@ OVN_CLEANUP_NORTHD
>>  AT_CLEANUP
>>  ])
>>
>> +OVN_FOR_EACH_NORTHD_NO_HV([
>> +AT_SETUP([Load balancer with symmetric_return on a router without DGP])
>> +ovn_start
>> +
>> +check ovn-nbctl ls-add sw0
>> +check ovn-nbctl lsp-add sw0 sw0-port1 \
>> +    -- lsp-set-addresses sw0-port1 "50:54:00:00:00:03 10.0.0.3"
>> +
>> +check ovn-nbctl lr-add lr0
>> +check ovn-nbctl lrp-add lr0 lr0-sw0 00:00:00:00:ff:01 10.0.0.1/24
>> +check ovn-nbctl lsp-add sw0 sw0-lr0 \
>> +    -- lsp-set-type sw0-lr0 router \
>> +    -- lsp-set-addresses sw0-lr0 router \
>> +    -- lsp-set-options sw0-lr0 router-port=lr0-sw0
>> +
>> +check ovn-nbctl lb-add lb0 172.168.0.200:80 10.0.0.4:8080
>> +check ovn-nbctl lr-lb-add lr0 lb0
>> +check ovn-nbctl --wait=sb sync
>> +
>> +# Default: the router load balances in lr_in_dnat with no chassis binding,
>> +# but has no un-DNAT for the backend.
>> +ovn-sbctl dump-flows lr0 > lr0flows
>> +AT_CAPTURE_FILE([lr0flows])
>> +AT_CHECK([grep "lr_in_dnat" lr0flows | grep "ip4.dst == 172.168.0.200" | \
>> +          grep -c "ct_lb_mark(backends=10.0.0.4:8080)"], [0], [1
>> +])
>> +AT_CHECK([grep "lr_in_dnat" lr0flows | grep "ip4.dst == 172.168.0.200" | \
>> +          grep -c "is_chassis_resident"], [1], [0
>> +])
>> +AT_CHECK([grep "lr_out_undnat" lr0flows | grep -c "10.0.0.4"], [1], [0
>> +])
>> +
>> +# symmetric_return: the backend's replies are un-DNATed; lr_in_dnat is
>> +# unchanged.
>> +check ovn-nbctl --wait=sb set load_balancer lb0 
>> options:symmetric_return=true
>> +ovn-sbctl dump-flows lr0 > lr0flows
>> +AT_CHECK([grep "lr_in_dnat" lr0flows | grep "ip4.dst == 172.168.0.200" | \
>> +          grep -c "ct_lb_mark(backends=10.0.0.4:8080)"], [0], [1
>> +])
>> +AT_CHECK([grep "lr_out_undnat" lr0flows | grep "priority=120" | \
>> +          grep "ip4.src == 10.0.0.4 && tcp.src == 8080" | \
>> +          grep -c "action=(ct_dnat;)"], [0], [1
>> +])
>> +AT_CHECK([grep "lr_out_undnat" lr0flows | grep "10.0.0.4" | \
>> +          grep -c "is_chassis_resident"], [1], [0
>> +])
>> +
>> +# With use_common_zone the un-DNAT uses the common zone, like the LB's
>> +# own flows on routers with a distributed gateway port.
>> +check ovn-nbctl --wait=sb set NB_Global . options:use_common_zone=true
>> +ovn-sbctl dump-flows lr0 > lr0flows
>> +AT_CHECK([grep "lr_out_undnat" lr0flows | grep "priority=120" | \
>> +          grep "ip4.src == 10.0.0.4 && tcp.src == 8080" | \
>> +          grep -c "action=(ct_dnat_in_czone;)"], [0], [1
>> +])
>> +check ovn-nbctl --wait=sb remove NB_Global . options use_common_zone
>> +
>> +# Removing the option removes the un-DNAT flow.
>> +check ovn-nbctl --wait=sb remove load_balancer lb0 options symmetric_return
>> +ovn-sbctl dump-flows lr0 > lr0flows
>> +AT_CHECK([grep "lr_out_undnat" lr0flows | grep -c "10.0.0.4"], [1], [0
>> +])
>> +
>> +OVN_CLEANUP_NORTHD
>> +AT_CLEANUP
>> +])
>> +
>> --
>> 2.55.0
>>
>> _______________________________________________
>> dev mailing list
>> [email protected]
>> https://mail.openvswitch.org/mailman/listinfo/ovs-dev
> _______________________________________________
> dev mailing list
> [email protected]
> https://mail.openvswitch.org/mailman/listinfo/ovs-dev

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to