On Tue, Apr 14, 2026 at 3:41 PM Dumitru Ceara <[email protected]> wrote:

> Every first-time addition of an existing ls_stateful_record to the
> crupdated tracking set must call ls_stateful_record_set_acls() to
> refresh has_stateful_acl and related ACL state.  Two places were
> missing this call:
>
> 1. The LB loop in ls_stateful_northd_handler() added to crupdated
>    without calling set_acls().  When a switch had both changed LBs
>    and changed ACLs, the LB loop added the record first (without
>    refreshing ACL state), then the ACL loop's hmapx_add() returned
>    false and skipped the refresh.
>
> 2. ls_stateful_acl_handler() added to crupdated without calling
>    set_acls().  When an existing ACL's action changed (e.g., allow
>    to allow-related), has_stateful_acl was never updated.
>
> Fixes: fb477aff9286 ("northd: Process ACL changes incrementally.")
> Assisted-by: Claude, with model: claude-opus-4-6
> Signed-off-by: Dumitru Ceara <[email protected]>
> ---
>

Hi Dumitru,

thank you for the patch, I have two comments below.


>  northd/en-ls-stateful.c | 21 +++++++++++--
>  tests/ovn-northd.at     | 69 +++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 87 insertions(+), 3 deletions(-)
>
> diff --git a/northd/en-ls-stateful.c b/northd/en-ls-stateful.c
> index 4daeab20d7..84e58b5deb 100644
> --- a/northd/en-ls-stateful.c
> +++ b/northd/en-ls-stateful.c
> @@ -171,8 +171,14 @@ ls_stateful_northd_handler(struct engine_node *node,
> void *data_)
>          ovs_assert(ls_stateful_rec);
>          ls_stateful_rec->has_lb_vip = ls_has_lb_vip(od);
>
> -        /* Add the ls_stateful_rec to the tracking data. */
> -        hmapx_add(&data->trk_data.crupdated, ls_stateful_rec);
> +        /* Add the ls_stateful_rec to the tracking data.  Refresh ACL
> +         * state when first added so that a switch with both changed LBs
> +         * and changed ACLs gets its ACL state updated regardless of
> +         * which loop runs first. */
> +        if (hmapx_add(&data->trk_data.crupdated, ls_stateful_rec)) {
> +            ls_stateful_record_set_acls(ls_stateful_rec, od->nbs,
> +                                        input_data.ls_port_groups);
> +        }
>      }
>
>      HMAPX_FOR_EACH (hmapx_node, &nd_changes->ls_with_changed_acls) {
> @@ -243,6 +249,7 @@ ls_stateful_port_group_handler(struct engine_node
> *node, void *data_)
>  enum engine_input_handler_result
>  ls_stateful_acl_handler(struct engine_node *node, void *data_)
>  {
> +    struct ls_stateful_input input_data =
> ls_stateful_get_input_data(node);
>      struct ed_type_ls_stateful *data = data_;
>      const struct nbrec_acl_table *nbrec_acl_table =
>          EN_OVSDB_GET(engine_get_input("NB_acl", node));
> @@ -259,7 +266,15 @@ ls_stateful_acl_handler(struct engine_node *node,
> void *data_)
>          LS_STATEFUL_TABLE_FOR_EACH (ls_stateful_rec, &data->table) {
>              if (uuidset_contains(&ls_stateful_rec->related_acls,
>                                   &acl->header_.uuid)) {
> -                hmapx_add(&data->trk_data.crupdated, ls_stateful_rec);
> +                if (hmapx_add(&data->trk_data.crupdated,
> ls_stateful_rec)) {
> +                    const struct ovn_datapath *od = ovn_datapath_find(
> +                        &input_data.ls_datapaths->datapaths,
> +                        &ls_stateful_rec->nbs_uuid);
>

Wouldn't it be better to do "ovn_datapaths_find_by_index()"?
Also I think that at this point the od cannot be NULL,
so let's make it an assert WDYT?

+                    if (od) {
> +                        ls_stateful_record_set_acls(ls_stateful_rec,
> od->nbs,
> +
> input_data.ls_port_groups);
> +                    }
> +                }
>              }
>          }
>      }
> diff --git a/tests/ovn-northd.at b/tests/ovn-northd.at
> index 624e08c1df..796c30daf7 100644
> --- a/tests/ovn-northd.at
> +++ b/tests/ovn-northd.at
> @@ -18957,6 +18957,75 @@ OVN_CLEANUP_NORTHD
>  AT_CLEANUP
>  ])
>
> +OVN_FOR_EACH_NORTHD_NO_HV([
> +AT_SETUP([LS ls_stateful incremental processing])
> +AT_KEYWORDS([incremental processing])
> +ovn_start
> +
> +AS_BOX([LB and ACL added in same transaction])
> +
> +dnl Create a switch with a port.
> +check ovn-nbctl --wait=sb \
> +    -- ls-add ls0 \
> +    -- lsp-add ls0 lsp0 \
> +    -- lsp-set-addresses lsp0 "00:00:00:00:00:01 10.0.0.1"
> +
> +dnl Clear engine stats before the combined transaction.
> +check as northd ovn-appctl -t ovn-northd inc-engine/clear-stats
> +
> +dnl In a single transaction, add both a stateful ACL and a load balancer.
> +dnl This exercises the ls_stateful_northd_handler() LB loop path where
> +dnl the switch is added to crupdated tracking and must also refresh ACL
> +dnl state via ls_stateful_record_set_acls().
> +check ovn-nbctl --wait=sb \
> +    -- acl-add ls0 from-lport 100 "ip" allow-related \
> +    -- lb-add lb1 10.0.0.100:80 10.0.0.1:80 \
> +    -- ls-lb-add ls0 lb1
> +
> +dnl Verify incremental processing was used (no recompute).
> +check_engine_stats ls_stateful norecompute compute
> +
> +dnl Verify the conntrack defrag flow exists in pre_acl at priority 100.
> +dnl REGBIT_CONNTRACK_DEFRAG (reg0[0] = 1) is only set when
> +dnl has_stateful_acl is true, proving ls_stateful_record_set_acls()
> +dnl was called during the combined LB+ACL transaction.
> +ovn-sbctl dump-flows ls0 > lflows
> +AT_CHECK([grep 'ls_in_pre_acl' lflows | grep 'priority=100' | grep -q
> 'reg0\[[0\]] = 1'])
> +
> +AS_BOX([ACL action change from allow to allow-related])
> +
> +dnl Create a switch with a non-stateful ACL (allow) and a port.
> +check ovn-nbctl --wait=sb \
> +    -- ls-add ls1 \
> +    -- lsp-add ls1 lsp1 \
> +    -- lsp-set-addresses lsp1 "00:00:00:00:00:02 10.0.0.2" \
> +    -- acl-add ls1 from-lport 100 "ip" allow
> +
> +dnl Verify NO conntrack defrag flow (non-stateful ACL).
> +ovn-sbctl dump-flows ls1 > lflows
> +AT_CHECK([grep 'ls_in_pre_acl' lflows | grep 'priority=100' | grep -q
> 'reg0\[[0\]] = 1'], [1])
> +
> +dnl Clear engine stats before modifying the ACL.
> +check as northd ovn-appctl -t ovn-northd inc-engine/clear-stats
> +
> +dnl Change the ACL action from allow to allow-related.
> +dnl This exercises ls_stateful_acl_handler() where a changed ACL must
> +dnl trigger ls_stateful_record_set_acls() to update has_stateful_acl.
> +acl_uuid=$(fetch_column nb:Acl _uuid action=allow)
> +check ovn-nbctl --wait=sb set acl $acl_uuid action=allow-related
> +
> +dnl Verify incremental processing was used (no recompute).
> +check_engine_stats ls_stateful norecompute compute
> +
> +dnl Verify the conntrack defrag flow now appears, proving
> +dnl has_stateful_acl was correctly updated.
> +ovn-sbctl dump-flows ls1 > lflows
> +AT_CHECK([grep 'ls_in_pre_acl' lflows | grep 'priority=100' | grep -q
> 'reg0\[[0\]] = 1'])
> +
> +OVN_CLEANUP_NORTHD
> +AT_CLEANUP
> +])
> +
>  OVN_FOR_EACH_NORTHD_NO_HV([
>  AT_SETUP([Check network function])
>  ovn_start
> --
> 2.53.0
>
>
Regards,
Ales
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to