When performing feature detection with OVS, we use an rconn connected to
br-int.mgmt to query for features. If the rconn is not connected, then
we would exit early. However, this early exit did not call
rconn_run_wait() or rconn_recv_wait(). Therefore, if the rconn were in
a state such as BACKOFF or CONNECTING, we would not wait on the rconn's
fd for activity.
In most cases, nobody would notice this because ovn-controller would
likely be waking up constantly due to southbound database changes,
handling CLI commands, or other activity. However, in one of Red Hat
QE's tests, they were seeing ovn-controller take 10 seconds between
when it would start up and when OF flows were installed. This was
because the only thing waking ovn-controller up was regularly-scheduled
probe intervals with the southbound database.
To fix this, we call rconn_run_wait() and rconn_recv_wait() even in
the case when the rconn is not connected. This way, if we end up in a
state other than ACTIVE or IDLE, we can handle traffic from the rconn
when it arrives.
Fixes: e9e716ad531e ("controller: Don't artificially limit group and
meter IDs to 16bit.")
Reported-at: https://issues.redhat.com/browse/FDP-357
Signed-off-by: Mark Michelson <[email protected]>
---
lib/features.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/lib/features.c b/lib/features.c
index 82bce10e9..b04437235 100644
--- a/lib/features.c
+++ b/lib/features.c
@@ -150,6 +150,8 @@ ovs_feature_get_openflow_cap(const char *br_name)
rconn_run(swconn);
if (!rconn_is_connected(swconn)) {
+ rconn_run_wait(swconn);
+ rconn_recv_wait(swconn);
return false;
}
--
2.43.0
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
