On Tue, Nov 17, 2026 at 1:31 PM Dumitru Ceara <[email protected]> wrote:

> On 11/2/26 3:08 PM, Ales Musil via dev wrote:
> > We could end up with a stale neighbor entry when advertised DGP
> > migrates to a different chassis, that would be the case until the
> > next recompute. Make sure we check the chassis and properly reflect
> > the migration.
> >
> > Fixes: 32d5252ceb65 ("controller, northd: Add an option to advertise FDB
> over EVPN.")
> > Assisted-by: GPT-6.1 Sol, OpenCode
> > Signed-off-by: Ales Musil <[email protected]>
> > ---
>
> Hi Ales, Phillip,
>
> Thanks for the fix and for testing it!
>

Hi Phillip and Dumitru,

thank you for the testing and review.


>
> >  controller/ovn-controller.c | 27 ++++++++++++-
> >  tests/multinode.at          | 77 +++++++++++++++++++++++++++++++++++++
> >  2 files changed, 102 insertions(+), 2 deletions(-)
> >
> > diff --git a/controller/ovn-controller.c b/controller/ovn-controller.c
> > index 7bc3139ab..f4ec67e89 100644
> > --- a/controller/ovn-controller.c
> > +++ b/controller/ovn-controller.c
> > @@ -6742,6 +6742,18 @@ neighbor_sb_port_binding_handler(struct
> engine_node *node, void *data)
> >      struct ed_type_neighbor *ne_data = data;
> >      const struct sbrec_port_binding_table *pb_table =
> >          EN_OVSDB_GET(engine_get_input("SB_port_binding", node));
> > +    const struct ovsrec_open_vswitch_table *ovs_table =
> > +    EN_OVSDB_GET(engine_get_input("OVS_open_vswitch", node));
>
> Nit: indentation.
>
> > +    struct ovsdb_idl_index *sbrec_chassis_by_name =
> > +        engine_ovsdb_node_get_index(
> > +            engine_get_input("SB_chassis", node),
> > +            "name");
> > +
> > +    const char *chassis_id = get_ovs_chassis_id(ovs_table);
> > +    ovs_assert(chassis_id);
> > +    const struct sbrec_chassis *chassis =
> > +        chassis_lookup_by_name(sbrec_chassis_by_name, chassis_id);
> > +    ovs_assert(chassis);
> >
> >      const struct sbrec_port_binding *pb;
> >      SBREC_PORT_BINDING_TABLE_FOR_EACH_TRACKED (pb, pb_table) {
> > @@ -6751,8 +6763,19 @@ neighbor_sb_port_binding_handler(struct
> engine_node *node, void *data)
> >              return EN_HANDLED_UNCHANGED;
> >          }
> >
> > -        if (sbrec_port_binding_is_updated(pb,
> SBREC_PORT_BINDING_COL_MAC) &&
> > -            sset_contains(&ne_data->advertised_pbs, pb->logical_port)) {
> > +        bool local = lport_pb_is_chassis_resident(chassis, pb);
> > +        enum sbrec_port_binding_column_id mac_col =
> SBREC_PORT_BINDING_COL_MAC;
> > +        if (sset_contains(&ne_data->advertised_pbs, pb->logical_port) &&
> > +            (sbrec_port_binding_is_updated(pb, mac_col) || !local)) {
> > +            return EN_UNHANDLED;
> > +        }
> > +
> > +        enum en_lport_type type = get_lport_type(pb);
> > +        enum sbrec_port_binding_column_id chassis_col =
> > +            SBREC_PORT_BINDING_COL_CHASSIS;
> > +        if (type == LP_CHASSISREDIRECT &&
> > +            sbrec_port_binding_is_updated(pb, chassis_col) &&
> > +            local) {
> >              return EN_UNHANDLED;
>
> So for every LP_CHASSISREDIRECT port binding chassis update (even if
> it's completely irrelevant to the local chassis, e.g.,
> ovn-monitor-all=true) we would fall back to recompute.
>
> Can we do better than that?  Would the pb->logical_port be in the
> advertised_pbs sset?  If so we could check for that before failing I-P.
>
> What do you think?
>

We have the && local with local = lport_pb_is_chassis_resident(chassis,
pb);
so it is only when it moves to our chassis from elsewhere.
Any other movement should be ignored.


> Regards,
> Dumitru
>
> >          }
> >      }
> > diff --git a/tests/multinode.at b/tests/multinode.at
> > index bd556b8f5..6292cc40b 100644
> > --- a/tests/multinode.at
> > +++ b/tests/multinode.at
> > @@ -4466,6 +4466,83 @@ OVS_WAIT_FOR_OUTPUT([m_as ovn-gw-2 ovs-ofctl
> dump-flows br-int table=OFTABLE_GET
> >  1
> >  ])
> >
> > +AS_BOX([Check gateway EVPN advertisements follow HA ownership])
> > +# Both gateways already have local workloads on ls.  Migrating only the
> > +# chassisredirect binding must refresh advertisements even when neither
> > +# the local datapaths nor the Advertised_MAC_Binding rows change.
> > +gw1_chassis=$(m_fetch_column Chassis _uuid name=ovn-gw-1)
> > +gw2_chassis=$(m_fetch_column Chassis _uuid name=ovn-gw-2)
> > +check multinode_nbctl \
> > +    -- ha-chassis-group-add evpn-ha \
> > +    -- ha-chassis-group-add-chassis evpn-ha ovn-gw-1 20 \
> > +    -- ha-chassis-group-add-chassis evpn-ha ovn-gw-2 10
> > +evpn_ha=$(m_fetch_column nb:HA_Chassis_Group _uuid name=evpn-ha)
> > +check multinode_nbctl --wait=hv set logical_router_port lr-ls
> ha_chassis_group=$evpn_ha
> > +
> > +priority=20
> > +for gw in 1 2 1; do
> > +    active=ovn-gw-$gw
> > +    standby=ovn-gw-$((3 - gw))
> > +    if test "$gw" = 1; then
> > +        chassis=$gw1_chassis
> > +        host_ip=$host_bgp_ip_gw1
> > +    else
> > +        chassis=$gw2_chassis
> > +        host_ip=$host_bgp_ip_gw2
> > +    fi
> > +    check multinode_nbctl ha-chassis-group-add-chassis evpn-ha $active
> $priority
> > +    m_wait_row_count Port_Binding 1 logical_port=cr-lr-ls
> chassis=$chassis
> > +
> > +    # Ignore VLAN 1 copies, which are not managed by neighbor exchange.
> > +    # A desired static entry may retain extern_learn; do not require it
> > +    # to be cleared in order to recognize the local advertisement.
> > +    OVS_WAIT_UNTIL([m_as $active bridge fdb show dev lo-10 | \
> > +        grep -v ' vlan ' | grep '^00:00:00:01:00:00 ' | grep -q '
> static'])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $standby bridge fdb show dev lo-10 | \
> > +        grep -v ' vlan ' | grep -q '^00:00:00:01:00:00 '], [1])
> > +    OVS_WAIT_UNTIL([m_as $active ip -4 neigh show to 10.0.0.1 dev br-10
> nud all | \
> > +        grep 'lladdr 00:00:00:01:00:00 ' | grep -q NOARP])
> > +    OVS_WAIT_UNTIL([m_as $active ip -6 neigh show to 10::1 dev br-10
> nud all | \
> > +        grep 'lladdr 00:00:00:01:00:00 ' | grep -q NOARP])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $standby ip -4 neigh show to 10.0.0.1 dev
> br-10 nud all | \
> > +        grep -q 'lladdr 00:00:00:01:00:00 '], [1])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $standby ip -6 neigh show to 10::1 dev
> br-10 nud all | \
> > +        grep -q 'lladdr 00:00:00:01:00:00 '], [1])
> > +
> > +    OVS_WAIT_FOR_OUTPUT_UNQUOTED([m_as $active ip netns exec frr-ns \
> > +        vtysh --vty_socket /run/frr/frr-ns -c 'show bgp l2vpn evpn
> route' | \
> > +        grep --no-group-separator -A1 00:00:00:01:00:00 | \
> > +        sed 's/ RD [[0-9.:]]*//g'], [0], [dnl
> > + *>  [[2]]:[[0]]:[[48]]:[[00:00:00:01:00:00]]
> > +                    $host_ip                            0 4210000000 i
> > + *>  [[2]]:[[0]]:[[48]]:[[00:00:00:01:00:00]]:[[32]]:[[10.0.0.1]]
> > +                    $host_ip                            0 4210000000 i
> > + *>  [[2]]:[[0]]:[[48]]:[[00:00:00:01:00:00]]:[[128]]:[[10::1]]
> > +                    $host_ip                            0 4210000000 i
> > +])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $standby ip netns exec frr-ns \
> > +        vtysh --vty_socket /run/frr/frr-ns -c 'show bgp l2vpn evpn
> route' | \
> > +        grep -q 00:00:00:01:00:00], [1])
> > +    priority=$((priority + 10))
> > +done
> > +
> > +# Restore the distributed router before the exact workload cleanup
> checks.
> > +check multinode_nbctl --wait=hv \
> > +    -- clear logical_router_port lr-ls ha_chassis_group \
> > +    -- ha-chassis-group-del evpn-ha
> > +m_wait_row_count Port_Binding 0 logical_port=cr-lr-ls
> > +for gw in ovn-gw-1 ovn-gw-2; do
> > +    OVS_WAIT_FOR_OUTPUT([m_as $gw bridge fdb show dev lo-10 | \
> > +        grep -v ' vlan ' | grep -q '^00:00:00:01:00:00 '], [1])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $gw ip -4 neigh show to 10.0.0.1 dev
> br-10 nud all | \
> > +        grep -q 'lladdr 00:00:00:01:00:00 '], [1])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $gw ip -6 neigh show to 10::1 dev br-10
> nud all | \
> > +        grep -q 'lladdr 00:00:00:01:00:00 '], [1])
> > +    OVS_WAIT_FOR_OUTPUT([m_as $gw ip netns exec frr-ns \
> > +        vtysh --vty_socket /run/frr/frr-ns -c 'show bgp l2vpn evpn
> route' | \
> > +        grep -q 00:00:00:01:00:00], [1])
> > +done
> > +
> >  # Remove "workloads" (VIF LSPs) on both chassis.
> >  check multinode_nbctl --wait=hv lsp-del w1 -- lsp-del w2
> >
>
>
Regards,
Ales
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to