BUG/MINOR: listener: keep accept rate counters accurate under saturation
The test on l->nbconn forces to exit the loop before updating the freq
counters, so the last session which reaches a listener's limit will not
be accounted for in the session rate measurement.
Let's move the test at the beginning of the loop and mark the listener
as saturated on exit.
This may be backported to 1.9 and 1.8.
diff --git a/src/listener.c b/src/listener.c
index 4d08031..f8745e3 100644
--- a/src/listener.c
+++ b/src/listener.c
@@ -464,11 +464,6 @@
if (HA_SPIN_TRYLOCK(LISTENER_LOCK, &l->lock))
return;
- if (unlikely(l->nbconn >= l->maxconn)) {
- listener_full(l);
- goto end;
- }
-
if (!(l->options & LI_O_UNLIMITED) && global.sps_lim) {
int max = freq_ctr_remain(&global.sess_per_sec, global.sps_lim, 0);
@@ -527,7 +522,7 @@
* worst case. If we fail due to system limits or temporary resource
* shortage, we try again 100ms later in the worst case.
*/
- while (max_accept--) {
+ while (l->nbconn < l->maxconn && max_accept--) {
struct sockaddr_storage addr;
socklen_t laddr = sizeof(addr);
unsigned int count;
@@ -647,11 +642,6 @@
goto transient_error;
}
- if (l->nbconn >= l->maxconn) {
- listener_full(l);
- goto end;
- }
-
/* increase the per-process number of cumulated connections */
if (!(l->options & LI_O_UNLIMITED)) {
count = update_freq_ctr(&global.sess_per_sec, 1);
@@ -679,6 +669,9 @@
limit_listener(l, &global_listener_queue);
task_schedule(global_listener_queue_task, tick_first(expire, global_listener_queue_task->expire));
end:
+ if (l->nbconn >= l->maxconn)
+ listener_full(l);
+
HA_SPIN_UNLOCK(LISTENER_LOCK, &l->lock);
}