On 11/6/26 5:59 PM, Rukomoinikova Aleksandra wrote:
>> If two AZs connect ports with the same name to a transit switch,
>> ovn-ic retries a failing transaction forever. Port names are unique
>> in both the NB (Logical_Switch_Port) and the IC-SB (Port_Binding):
>> one AZ tries to create a remote LSP that clashes with its local one,
>> and the other AZ tries to bind a port name that is already bound.
>> The failed commit also blocks every other change in the same
>> transaction.
>>
>> Check the name before inserting. On a conflict, skip the port and
>> log a rate-limited warning that names the port and AZs, so the
>> operator knows which port to rename.
>>
>> Fixes: aa3e4436c232 ("ovn-ic: Interconnection port controller.")
>> Assisted-by: Claude Opus 5.5, Claude Code
>> Signed-off-by: Lucas Vargas Dias <lucas.vdias at magalu.cloud
>> <https://mail.openvswitch.org/mailman/listinfo/ovs-dev>>
> Hi Lucas and Rosemarie,
>
> Thanks for the patch. I have a question below.
>
>> ---
>> ic/en-port-binding.c | 75 +++++++++++++++++++++++++++++++++++++++-----
>> ic/ovn-ic.c | 5 +++
>> ic/ovn-ic.h | 1 +
>> tests/ovn-ic.at | 54 +++++++++++++++++++++++++++++++
>> 4 files changed, 127 insertions(+), 8 deletions(-)
>>
>> diff --git a/ic/en-port-binding.c b/ic/en-port-binding.c
>> index 5a2490b6e..638f819e0 100644
>> --- a/ic/en-port-binding.c
>> +++ b/ic/en-port-binding.c
>> @@ -465,6 +465,28 @@ create_nb_lsp(struct ic_context *ctx,
>> const struct icsbrec_port_binding *isb_pb,
>> const struct nbrec_logical_switch *ls)
>> {
>> + /* The NB Logical_Switch_Port name is unique across the whole NB, but a
>> + * remote AZ may bind a port whose name is already used here, e.g. by a
>> + * local port with the same name on this transit switch or by a port of
>> + * an unrelated logical switch. Inserting it would violate the name
>> + * index, abort the whole NB transaction and retry it forever, so skip
>> the
>> + * remote port instead. A row inserted earlier in this same transaction
>> + * (see lsp_create()) is also found here and is silently reused. */
>> + const struct nbrec_logical_switch_port *existing =
>> + get_lsp_by_ts_port_name(ctx, isb_pb->logical_port);
>> + if (existing) {
>> + if (!nbrec_logical_switch_port_is_new(existing)) {
> I'm not sure this check is needed here. I took a look at ovsdb-idl
> and the comment on _is_new() says it can only be used for rows from
> tracked changes, not for uncommitted local rows. Since existing comes
> from an index lookup, it could be exactly that kind of row.
>
> Maybe we could drop the check and always log the warning? The port is
> skipped either way. A row inserted earlier in the same transaction
> would cause an occasional false warning, but I think that's fine. In
> that case the part of the comment about "silently reused" could go too.
>
> That said, this isn't a big deal. It only affects logging, so I don't
> see it as blocking the patch, and I'm fine with keeping it if you prefer.
>
> Otherwise, the patch looks correct to me.
>
> What do you think?
>
> Thanks,
>
> Sasha.
Hi Sasha, thanks for catching that! I definitely agree with your idea.
-Rosemarie
>> + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
>> + VLOG_WARN_RL(&rl, "Port %s of transit switch %s in availability
>> "
>> + "zone %s conflicts with an existing logical switch
>> "
>> + "port with the same name, skipping it.",
>> + isb_pb->logical_port, isb_pb->transit_switch,
>> + isb_pb->availability_zone
>> + ? isb_pb->availability_zone->name : "<unknown>");
>> + }
>> + return;
>> + }
>> +
>> const struct nbrec_logical_switch_port *lsp =
>> nbrec_logical_switch_port_insert(ctx->ovnnb_txn);
>> nbrec_logical_switch_port_set_name(lsp, isb_pb->logical_port);
>> @@ -521,11 +543,45 @@ find_transit_router_by_name(struct ic_context *ctx,
>> const char *name)
>> }
>>
>> static const struct icsbrec_port_binding *
>> -create_isb_pb(struct ovsdb_idl_txn *isb_txn, const char *logical_port,
>> +find_isb_pb_by_name(struct ic_context *ctx, const char *logical_port)
>> +{
>> + const struct icsbrec_port_binding *key =
>> + icsbrec_port_binding_index_init_row(
>> + ctx->icsbrec_port_binding_by_name);
>> + icsbrec_port_binding_index_set_logical_port(key, logical_port);
>> + const struct icsbrec_port_binding *isb_pb =
>> + icsbrec_port_binding_index_find(ctx->icsbrec_port_binding_by_name,
>> + key);
>> + icsbrec_port_binding_index_destroy_row(key);
>> + return isb_pb;
>> +}
>> +
>> +static const struct icsbrec_port_binding *
>> +create_isb_pb(struct ic_context *ctx, struct ovsdb_idl_txn *isb_txn,
>> + const char *logical_port,
>> const struct icsbrec_availability_zone *az, const char
>> *ts_name,
>> const struct uuid *nb_ic_uuid, const char *type,
>> struct hmap *pb_tnlids)
>> {
>> + /* The IC-SB Port_Binding logical_port is unique across all AZs and
>> + * transit switches/routers. If another AZ (or another transit
>> + * switch/router) already binds a port with this name, inserting would
>> + * violate the index, abort the whole IC-SB transaction and retry it
>> + * forever, so leave this port unbound instead. */
>> + const struct icsbrec_port_binding *existing =
>> + find_isb_pb_by_name(ctx, logical_port);
>> + if (existing) {
>> + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
>> + VLOG_WARN_RL(&rl, "Port %s of %s in availability zone %s conflicts "
>> + "with a port with the same name of %s in availability "
>> + "zone %s, not binding it.",
>> + logical_port, ts_name, az->name,
>> + existing->transit_switch,
>> + existing->availability_zone
>> + ? existing->availability_zone->name : "<unknown>");
>> + return NULL;
>> + }
>> +
>> uint32_t pb_tnl_key = allocate_port_key(pb_tnlids);
>> if (!pb_tnl_key) {
>> return NULL;
>> @@ -720,7 +776,7 @@ port_binding_sync_ts(struct ic_context *ctx,
>> if (ctx->ovnisb_txn && is_az_leader(ctx->ovnisb_txn)) {
>> if (!isb_pb) {
>> isb_pb = create_isb_pb(
>> - ctx->ovnisb_txn, tsp->name, runned_az,
>> + ctx, ctx->ovnisb_txn, tsp->name, runned_az,
>> ts->name, &ts->header_.uuid, "transit-switch-port",
>> &pb_tnlids);
>> }
>> @@ -730,8 +786,8 @@ port_binding_sync_ts(struct ic_context *ctx,
>> /* Create ISB port_binding as its chassis is local. */
>> isb_pb = shash_find_and_delete(&local_pbs, tsp->name);
>> if (!isb_pb) {
>> - isb_pb = create_isb_pb(ctx->ovnisb_unlocked_txn, tsp->name,
>> - runned_az,
>> + isb_pb = create_isb_pb(ctx, ctx->ovnisb_unlocked_txn,
>> + tsp->name, runned_az,
>> ts->name, &ts->header_.uuid,
>> "transit-switch-port", &pb_tnlids);
>> }
>> @@ -773,9 +829,12 @@ port_binding_sync_ts(struct ic_context *ctx,
>> isb_pb = shash_find_and_delete(&local_pbs, lsp->name);
>> if (!isb_pb) {
>> isb_pb = create_isb_pb(
>> - ctx->ovnisb_unlocked_txn, sb_pb->logical_port,
>> + ctx, ctx->ovnisb_unlocked_txn, sb_pb->logical_port,
>> runned_az, ts->name, &ts->header_.uuid,
>> "transit-switch-port", &pb_tnlids);
>> + if (!isb_pb) {
>> + continue;
>> + }
>> sync_ts_isb_pb(ctx, sb_pb, isb_pb);
>> } else {
>> sync_local_port(ctx, isb_pb, sb_pb, lsp);
>> @@ -899,7 +958,7 @@ port_binding_sync_tr(struct ic_context *ctx,
>>
>> if (ctx->ovnisb_txn && is_az_leader(ctx->ovnisb_txn)) {
>> if (!isb_pb) {
>> - isb_pb = create_isb_pb(ctx->ovnisb_txn, trp->name,
>> + isb_pb = create_isb_pb(ctx, ctx->ovnisb_txn, trp->name,
>> runned_az,
>> tr->name, &tr->header_.uuid,
>> "transit-router-port",
>> @@ -912,8 +971,8 @@ port_binding_sync_tr(struct ic_context *ctx,
>> } else {
>> isb_pb = shash_find_and_delete(&local_pbs, trp->name);
>> if (!isb_pb) {
>> - isb_pb = create_isb_pb(ctx->ovnisb_unlocked_txn, trp->name,
>> - runned_az,
>> + isb_pb = create_isb_pb(ctx, ctx->ovnisb_unlocked_txn,
>> + trp->name, runned_az,
>> tr->name, &tr->header_.uuid,
>> "transit-router-port", &pb_tnlids);
>> }
>> diff --git a/ic/ovn-ic.c b/ic/ovn-ic.c
>> index e65256577..87a266b44 100644
>> --- a/ic/ovn-ic.c
>> +++ b/ic/ovn-ic.c
>> @@ -1191,6 +1191,10 @@ main(int argc, char *argv[])
>> = ovsdb_idl_index_create1(ovnisb_unlocked_idl_loop.idl,
>>
>> &icsbrec_port_binding_col_availability_zone);
>>
>> + struct ovsdb_idl_index *icsbrec_port_binding_by_name
>> + = ovsdb_idl_index_create1(ovnisb_unlocked_idl_loop.idl,
>> + &icsbrec_port_binding_col_logical_port);
>> +
>> struct ovsdb_idl_index *icsbrec_port_binding_by_ts
>> = ovsdb_idl_index_create1(ovnisb_unlocked_idl_loop.idl,
>>
>> &icsbrec_port_binding_col_transit_switch);
>> @@ -1408,6 +1412,7 @@ main(int argc, char *argv[])
>> .icnbrec_transit_router_by_name =
>> icnbrec_transit_router_by_name,
>> .icsbrec_port_binding_by_az = icsbrec_port_binding_by_az,
>> + .icsbrec_port_binding_by_name =
>> icsbrec_port_binding_by_name,
>> .icsbrec_port_binding_by_ts = icsbrec_port_binding_by_ts,
>> .icsbrec_port_binding_by_ts_az =
>> icsbrec_port_binding_by_ts_az,
>> .icsbrec_route_by_az = icsbrec_route_by_az,
>> diff --git a/ic/ovn-ic.h b/ic/ovn-ic.h
>> index 8adb17ad7..b1cfad08f 100644
>> --- a/ic/ovn-ic.h
>> +++ b/ic/ovn-ic.h
>> @@ -43,6 +43,7 @@ struct ic_context {
>> struct ovsdb_idl_index *icnbrec_transit_switch_by_name;
>> struct ovsdb_idl_index *icnbrec_transit_router_by_name;
>> struct ovsdb_idl_index *icsbrec_port_binding_by_az;
>> + struct ovsdb_idl_index *icsbrec_port_binding_by_name;
>> struct ovsdb_idl_index *icsbrec_port_binding_by_ts;
>> struct ovsdb_idl_index *icsbrec_port_binding_by_ts_az;
>> struct ovsdb_idl_index *icsbrec_route_by_az;
>> diff --git a/tests/ovn-ic.at b/tests/ovn-ic.at
>> index 646026817..4576c725e 100644
>> --- a/tests/ovn-ic.at
>> +++ b/tests/ovn-ic.at
>> @@ -1689,6 +1689,60 @@ OVN_CLEANUP_IC([az1], [az2])
>> AT_CLEANUP
>> ])
>>
>> +OVN_FOR_EACH_NORTHD([
>> +AT_SETUP([ovn-ic -- same port name in two AZs])
>> +
>> +ovn_init_ic_db
>> +ovn-ic-nbctl ts-add ts1
>> +ovn_start az1
>> +ovn_start az2
>> +check ovn-ic-nbctl --wait=sb sync
>> +
>> +dnl Each AZ connects a router to ts1 through a legacy port that has the
>> +dnl same name in both AZs. Pause ovn-ic so that neither AZ learns about the
>> +dnl other's port before creating its own.
>> +for az in az1 az2; do
>> + as $az check ovn-appctl -t ic/ovn-ic pause
>> +done
>> +for az in az1 az2; do
>> + ovn_as $az
>> + check ovn-nbctl lr-add lr1
>> + check ovn-nbctl lrp-add lr1 lrp-lr1-ts1 aa:aa:aa:aa:aa:01
>> 169.254.100.1/24
>> + check ovn-nbctl --wait=sb lsp-add-router-port ts1 lsp-dup lrp-lr1-ts1
>> +done
>> +for az in az1 az2; do
>> + as $az check ovn-appctl -t ic/ovn-ic resume
>> +done
>> +
>> +dnl IC-SB port names are unique, so only one AZ gets to bind the port; the
>> +dnl other one must give up instead of retrying the insert forever.
>> +wait_row_count ic-sb:Port_Binding 1 logical_port=lsp-dup
>> +OVS_WAIT_UNTIL([grep -q "Port lsp-dup of ts1 .* conflicts with a port" \
>> + az1/ic/ovn-ic.log az2/ic/ovn-ic.log])
>> +
>> +dnl And the AZ that got the binding must not try to create a remote NB port
>> +dnl over the other AZ's local port with the same name.
>> +OVS_WAIT_UNTIL([grep -q "Port lsp-dup of transit switch ts1 .* conflicts" \
>> + az1/ic/ovn-ic.log az2/ic/ovn-ic.log])
>> +for az in az1 az2; do
>> + ovn_as $az
>> + check_row_count nb:Logical_Switch_Port 1 name=lsp-dup type=router
>> +done
>> +
>> +dnl ovn-ic must keep committing NB transactions: a port added in az1 shows
>> +dnl up as remote in az2.
>> +ovn_as az1
>> +check ovn-nbctl lrp-add lr1 lrp-lr1-ts1-2 aa:aa:aa:aa:aa:02 169.254.101.1/24
>> +check ovn-nbctl --wait=sb lsp-add-router-port ts1 lsp-az1-only lrp-lr1-ts1-2
>> +ovn_as az2
>> +wait_row_count nb:Logical_Switch_Port 1 name=lsp-az1-only type=remote
>> +
>> +AT_CHECK([grep -q "constraint violation" az1/ic/ovn-ic.log
>> az2/ic/ovn-ic.log], [1])
>> +
>> +OVN_CLEANUP_IC([az1], [az2])
>> +AT_CLEANUP
>> +])
>> +
>> OVN_FOR_EACH_NORTHD([
>> AT_SETUP([ovn-ic -- address-set sync])
>>
>
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev