The "ovn_ic_sb" lock is requested by every non-paused ovn-ic instance,
regardless of whether it holds the local AZ "ovn_ic" lock. With several
ovn-ic instances per AZ, a standby one can win the race for it.

That deadlocks IC-SB writes: the standby holds the lock but never reaches
inc_proc_ic_run(), which is gated on the local SB lock, while the active
instances run ts_run() with is_az_leader() false and skip datapath
creation and deletion entirely.

Request the lock only while holding the local SB lock, and drop the
request when that lock is lost. Also report ovn-ic-sb lock in
ovn-appctl.

Fixes: 052a298bb90e ("ovn-ic: Use dual IC-SB connections to prevent constraint 
violations.")
Signed-off-by: Alexandra Rukomoinikova <[email protected]>
---
 ic/ovn-ic.c     | 27 ++++++++++++------
 tests/ovn-ic.at | 73 +++++++++++++++++++++++++++++++++++++++++++++----
 2 files changed, 87 insertions(+), 13 deletions(-)

diff --git a/ic/ovn-ic.c b/ic/ovn-ic.c
index be03d1781..16682ccbe 100644
--- a/ic/ovn-ic.c
+++ b/ic/ovn-ic.c
@@ -4375,14 +4375,23 @@ main(int argc, char *argv[])
                 ovsdb_idl_set_lock(ovnsb_idl_loop.idl, "ovn_ic");
             }
 
-            if (!ovsdb_idl_has_lock(ovnisb_idl_loop.idl) &&
-                !ovsdb_idl_is_lock_contended(ovnisb_idl_loop.idl)) {
-                /*
-                 * Ensure that only a single ovn-ic has the permission to
-                 * write to IC-SB.
-                 */
-                VLOG_INFO("Acquiring OVN ISB lock.");
-                ovsdb_idl_set_lock(ovnisb_idl_loop.idl, "ovn_ic_sb");
+            if (ovsdb_idl_has_lock(ovnsb_idl_loop.idl)) {
+                if (!ovsdb_idl_has_lock(ovnisb_idl_loop.idl) &&
+                    !ovsdb_idl_is_lock_contended(ovnisb_idl_loop.idl)) {
+                    /*
+                     * Ensure that only a single ovn-ic has the permission to
+                     * write to IC-SB. Only the instance that is active in
+                     * its own AZ may request it: a standby instance never
+                     * runs the engine, so if it held the IC-SB lock nobody
+                     * would ever write to IC-SB.
+                     */
+                    VLOG_INFO("Acquiring OVN ISB lock.");
+                    ovsdb_idl_set_lock(ovnisb_idl_loop.idl, "ovn_ic_sb");
+                }
+            } else {
+                /* Release the IC-SB lock, or withdraw a pending request for
+                 * it, as soon as the local SB lock is gone. */
+                ovsdb_idl_set_lock(ovnisb_idl_loop.idl, NULL);
             }
 
             struct ovsdb_idl_txn *ovnnb_txn =
@@ -4720,6 +4729,8 @@ ovn_ic_status(struct unixctl_conn *conn, int argc 
OVS_UNUSED,
      */
     struct ds s = DS_EMPTY_INITIALIZER;
     ds_put_format(&s, "Status: %s\n", status);
+    ds_put_format(&s, "IC-SB lock: %s\n",
+                  state->had_isb_lock ? "held" : "not held");
     unixctl_command_reply(conn, ds_cstr(&s));
     ds_destroy(&s);
 }
diff --git a/tests/ovn-ic.at b/tests/ovn-ic.at
index c9e07250b..68b981925 100644
--- a/tests/ovn-ic.at
+++ b/tests/ovn-ic.at
@@ -6152,13 +6152,13 @@ check_column "ts1 ts2 ts3" ic-sb:Datapath_Binding 
transit_switch
 as az2 check ovn-appctl -t ic/ovn-ic resume
 
 # Now do the same on a "slow" ovn-ic-sb
-OVS_WAIT_UNTIL([test "x$(as az2 ovn-appctl -t ic/ovn-ic status)" = "xStatus: 
active"])
+OVS_WAIT_UNTIL([test "x$(as az2 ovn-appctl -t ic/ovn-ic status | grep 
"Status")" = "xStatus: active"])
 
 AS_BOX([$(date +%H:%M:%S.%03N) Pausing ovn-ic-sb])
 AT_CHECK([kill -STOP $(cat ovn-ic-sb/ovsdb-server.pid)])
 
 as az1 check ovn-appctl -t ic/ovn-ic pause
-OVS_WAIT_UNTIL([test "x$(as az1 ovn-appctl -t ic/ovn-ic status)" = "xStatus: 
paused"])
+OVS_WAIT_UNTIL([test "x$(as az1 ovn-appctl -t ic/ovn-ic status | grep 
"Status")" = "xStatus: paused"])
 
 check ovn-ic-nbctl ts-add ts4
 
@@ -6353,21 +6353,21 @@ net_add n1
 
 check ovn-ic-nbctl ts-add ts1
 ovn_start az1
-OVS_WAIT_UNTIL([test "x$(as az1 ovn-appctl -t ic/ovn-ic status)" = "xStatus: 
active"])
+OVS_WAIT_UNTIL([test "x$(as az1 ovn-appctl -t ic/ovn-ic status | grep 
"Status")" = "xStatus: active"])
 OVS_WAIT_UNTIL([grep -q "OVN ISB lock acquired" az1/ic/ovn-ic.log])
 
 ovn_start az2
 
 AS_BOX([az2 paused])
 check as az2 ovn-appctl -t ic/ovn-ic pause
-OVS_WAIT_UNTIL([test "x$(as az2 ovn-appctl -t ic/ovn-ic status)" = "xStatus: 
paused"])
+OVS_WAIT_UNTIL([test "x$(as az2 ovn-appctl -t ic/ovn-ic status | grep 
"Status")" = "xStatus: paused"])
 # Wait for az1 receiving "locked" notification.
 OVS_WAIT_UNTIL([grep -q 'send notification, method="locked"' 
ovn-ic-sb/ovsdb-server.log])
 n1_lock_notif=$(grep -c 'send notification, method="locked"' 
ovn-ic-sb/ovsdb-server.log)
 
 AS_BOX([az1 paused])
 check as az1 ovn-appctl -t ic/ovn-ic pause
-OVS_WAIT_UNTIL([test "x$(as az1 ovn-appctl -t ic/ovn-ic status)" = "xStatus: 
paused"])
+OVS_WAIT_UNTIL([test "x$(as az1 ovn-appctl -t ic/ovn-ic status | | grep 
"Status")" = "xStatus: paused"])
 n2_lock_notif=$(grep -c 'send notification, method="locked"' 
ovn-ic-sb/ovsdb-server.log)
 
 # Pausing az1 should not cause az2 to own the lock: az2 is paused.
@@ -6387,3 +6387,66 @@ OVN_CLEANUP_IC([az1], [az2])
 AT_CLEANUP
 ])
 
+OVN_FOR_EACH_NORTHD([
+AT_SETUP([ovn-ic - IC-SB lock is only requested by the active instance])
+ovn_init_ic_db
+
+ovn_start az1
+ovn_start az2
+wait_row_count ic-sb:Availability_Zone 2
+
+check ovn-ic-nbctl ts-add ts1
+wait_row_count ic-sb:Datapath_Binding 1 transit_switch=ts1
+
+# Start a second ovn-ic in az1. It can't get the SB lock, so it must stay
+# standby and must not compete for the IC-SB lock either.
+ovn_as az1
+mkdir az1/ic2
+as az1/ic2 start_daemon ovn-ic -v \
+    --ovnnb-db=$OVN_NB_DB --ovnsb-db=$OVN_SB_DB \
+    --ic-nb-db=unix:"$ovs_base"/ovn-ic-nb/ovn-ic-nb.sock \
+    --ic-sb-db=unix:"$ovs_base"/ovn-ic-sb/ovn-ic-sb.sock
+OVS_WAIT_UNTIL([as az1 ovn-appctl -t ic2/ovn-ic status | grep -q "^Status: 
standby$"])
+OVS_WAIT_UNTIL([as az1 ovn-appctl -t ic2/ovn-ic ic-sb-connection-status | grep 
-q "^connected$"])
+AT_CHECK([as az1 ovn-appctl -t ic2/ovn-ic status | grep -q "^IC-SB lock: not 
held$"])
+
+# Simulate an IC-SB failover: freeze both active instances, drop every
+# IC-SB connection and let the standby reconnect first. A standby that
+# competes for the IC-SB lock wins it here and, never running the engine,
+# blocks IC-SB datapath creation for everyone.
+AS_BOX([Dropping IC-SB connections with the active instances frozen])
+AT_CHECK([kill -STOP $(cat az1/ic/ovn-ic.pid) $(cat az2/ic/ovn-ic.pid)])
+check as ovn-ic-sb ovs-appctl -t ovsdb-server ovsdb-server/reconnect
+OVS_WAIT_UNTIL([as az1 ovn-appctl -t ic2/ovn-ic ic-sb-connection-status | grep 
-q "^not connected$"])
+OVS_WAIT_UNTIL([as az1 ovn-appctl -t ic2/ovn-ic ic-sb-connection-status | grep 
-q "^connected$"])
+AT_CHECK([kill -CONT $(cat az1/ic/ovn-ic.pid) $(cat az2/ic/ovn-ic.pid)])
+
+check ovn-ic-nbctl ts-add ts2
+wait_row_count ic-sb:Datapath_Binding 1 transit_switch=ts2
+check_column "ts1 ts2" ic-sb:Datapath_Binding transit_switch
+
+AT_CHECK([as az1 ovn-appctl -t ic2/ovn-ic status | grep -q "^Status: 
standby$"])
+AT_CHECK([as az1 ovn-appctl -t ic2/ovn-ic status | grep -q "^IC-SB lock: not 
held$"])
+AT_CHECK([grep -q "OVN ISB lock acquired" az1/ic2/ovn-ic.log], [1])
+
+# Once the standby becomes active in its AZ it should take the IC-SB lock.
+AS_BOX([Pausing both original instances])
+check as az1 ovn-appctl -t ic/ovn-ic pause
+check as az2 ovn-appctl -t ic/ovn-ic pause
+OVS_WAIT_UNTIL([as az1 ovn-appctl -t ic2/ovn-ic status | grep -q "^Status: 
active$"])
+OVS_WAIT_UNTIL([as az1 ovn-appctl -t ic2/ovn-ic status | grep -q "^IC-SB lock: 
held$"])
+
+check ovn-ic-nbctl ts-add ts3
+wait_row_count ic-sb:Datapath_Binding 1 transit_switch=ts3
+check_column "ts1 ts2 ts3" ic-sb:Datapath_Binding transit_switch
+
+check as az1 ovn-appctl -t ic/ovn-ic resume
+check as az2 ovn-appctl -t ic/ovn-ic resume
+
+as az1/ic2
+OVS_APP_EXIT_AND_WAIT([ovn-ic])
+
+OVN_CLEANUP_IC([az1], [az2])
+
+AT_CLEANUP
+])
-- 
2.48.1

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

Reply via email to