MEDIUM: servers: Reorganize the way idle connections are cleaned.

Instead of having one task per thread and per server that does clean the
idling connections, have only one global task for every servers.
That tasks parses all the servers that currently have idling connections,
and remove half of them, to put them in a per-thread list of connections
to kill. For each thread that does have connections to kill, wake a task
to do so, so that the cleaning will be done in the context of said thread.

include/proto/server.h

@@ -36,6 +36,13 @@
 #include <proto/log.h>
 #include <proto/freq_ctr.h>
 
+
+__decl_hathreads(extern HA_SPINLOCK_T idle_conn_srv_lock);
+extern struct eb_root idle_conn_srv;
+extern struct task *idle_conn_task;
+extern struct task *idle_conn_cleanup[MAX_THREADS];
+extern struct list toremove_connections[MAX_THREADS];
+
 int srv_downtime(const struct server *s);
 int srv_lastsession(const struct server *s);
 int srv_getinter(const struct check *check);
@@ -59,6 +66,7 @@ int snr_resolution_cb(struct dns_requester *requester, struct dns_nameserver *na
 int snr_resolution_error_cb(struct dns_requester *requester, int error_code);
 struct server *snr_check_ip_callback(struct server *srv, void *ip, unsigned char *ip_family);
 struct task *srv_cleanup_idle_connections(struct task *task, void *ctx, unsigned short state);
+struct task *srv_cleanup_toremove_connections(struct task *task, void *context, unsigned short state);
 
 /* increase the number of cumulated connections on the designated server */
 static void inline srv_inc_sess_ctr(struct server *s)
@@ -256,10 +264,22 @@ static inline int srv_add_to_idle_list(struct server *srv, struct connection *co
 		srv->curr_idle_thr[tid]++;
 
 		conn->idle_time = now_ms;
-		if (!(task_in_wq(srv->idle_task[tid])) &&
-		    !(task_in_rq(srv->idle_task[tid])))
-			task_schedule(srv->idle_task[tid],
-			    tick_add(now_ms, srv->pool_purge_delay));
+		__ha_barrier_full();
+		if ((volatile void *)srv->idle_node.node.leaf_p == NULL) {
+			HA_SPIN_LOCK(OTHER_LOCK, &idle_conn_srv_lock);
+			if ((volatile void *)srv->idle_node.node.leaf_p == NULL) {
+				srv->idle_node.key = tick_add(srv->pool_purge_delay,
+				                              now_ms);
+				eb32_insert(&idle_conn_srv, &srv->idle_node);
+				if (!task_in_wq(idle_conn_task) && !
+				    task_in_rq(idle_conn_task)) {
+					task_schedule(idle_conn_task,
+					              srv->idle_node.key);
+				}
+
+			}
+			HA_SPIN_UNLOCK(OTHER_LOCK, &idle_conn_srv_lock);
+		}
 		return 1;
 	}
 	return 0;

include/types/server.h

@@ -227,7 +227,7 @@ struct server {
 	unsigned int curr_idle_conns;           /* Current number of orphan idling connections */
 	unsigned int *curr_idle_thr;            /* Current number of orphan idling connections per thread */
 	int max_reuse;                          /* Max number of requests on a same connection */
-	struct task **idle_task;                /* task responsible for cleaning idle orphan connections */
+	struct eb32_node idle_node;             /* When to next do cleanup in the idle connections */
 	struct task *warmup;                    /* the task dedicated to the warmup when slowstart is set */
 
 	struct conn_src conn_src;               /* connection source settings */

src/backend.c

@@ -1348,8 +1348,6 @@ int connect_server(struct stream *s)
 		HA_ATOMIC_SUB(&srv->curr_idle_conns, 1);
 		srv->curr_idle_thr[tid]--;
 		LIST_ADDQ(&srv->idle_conns[tid], &srv_conn->list);
-		if (LIST_ISEMPTY(&srv->idle_orphan_conns[tid]))
-			task_unlink_wq(srv->idle_task[tid]);
 	} else if (reuse) {
 		if (srv_conn->flags & CO_FL_SESS_IDLE) {
 			struct session *sess = srv_conn->owner;

src/cfgparse.c

@@ -3581,18 +3581,26 @@ out_uri_auth_compat:
 			}
 
 			if (newsrv->max_idle_conns != 0) {
-				newsrv->idle_orphan_conns = calloc(global.nbthread, sizeof(*newsrv->idle_orphan_conns));
-				newsrv->idle_task         = calloc(global.nbthread, sizeof(*newsrv->idle_task));
-				if (!newsrv->idle_orphan_conns || !newsrv->idle_task)
-					goto err;
-				for (i = 0; i < global.nbthread; i++) {
-					LIST_INIT(&newsrv->idle_orphan_conns[i]);
-					newsrv->idle_task[i] = task_new(1 << i);
-					if (!newsrv->idle_task[i])
+				if (idle_conn_task == NULL) {
+					idle_conn_task = task_new(MAX_THREADS_MASK);
+					if (!idle_conn_task)
 						goto err;
-					newsrv->idle_task[i]->process = srv_cleanup_idle_connections;
-					newsrv->idle_task[i]->context = newsrv;
+					idle_conn_task->process = srv_cleanup_idle_connections;
+					idle_conn_task->context = NULL;
+					for (i = 0; i < global.nbthread; i++) {
+						idle_conn_cleanup[i] = task_new(1 << i);
+						if (!idle_conn_cleanup[i])
+							goto err;
+						idle_conn_cleanup[i]->process = srv_cleanup_toremove_connections;
+						idle_conn_cleanup[i]->context = NULL;
+						LIST_INIT(&toremove_connections[i]);
+					}
 				}
+				newsrv->idle_orphan_conns = calloc(global.nbthread, sizeof(*newsrv->idle_orphan_conns));
+				if (!newsrv->idle_orphan_conns)
+					goto err;
+				for (i = 0; i < global.nbthread; i++)
+					LIST_INIT(&newsrv->idle_orphan_conns[i]);
 				newsrv->curr_idle_thr = calloc(global.nbthread, sizeof(int));
 				if (!newsrv->curr_idle_thr)
 					goto err;

src/haproxy.c

@@ -2437,13 +2437,6 @@ void deinit(void)
 			free(s->safe_conns);
 			free(s->idle_orphan_conns);
 			free(s->curr_idle_thr);
-			if (s->idle_task) {
-				int i;
-
-				for (i = 0; i < global.nbthread; i++)
-					task_free(s->idle_task[i]);
-				free(s->idle_task);
-			}
 
 			if (s->use_ssl || s->check.use_ssl) {
 				if (xprt_get(XPRT_SSL) && xprt_get(XPRT_SSL)->destroy_srv)
@@ -2537,6 +2530,8 @@ void deinit(void)
 	free(global.desc);    global.desc = NULL;
 	free(oldpids);        oldpids = NULL;
 	task_free(global_listener_queue_task); global_listener_queue_task = NULL;
+	task_free(idle_conn_task);
+	idle_conn_task = NULL;
 
 	list_for_each_entry_safe(log, logb, &global.logsrvs, list) {
 			LIST_DEL(&log->list);

src/server.c

@@ -56,6 +56,12 @@ static struct srv_kw_list srv_keywords = {
 	.list = LIST_HEAD_INIT(srv_keywords.list)
 };
 
+__decl_hathreads(HA_SPINLOCK_T idle_conn_srv_lock);
+struct eb_root idle_conn_srv = EB_ROOT;
+struct task *idle_conn_task = NULL;
+struct task *idle_conn_cleanup[MAX_THREADS] = { NULL };
+struct list toremove_connections[MAX_THREADS];
+
 int srv_downtime(const struct server *s)
 {
 	if ((s->cur_state != SRV_ST_STOPPED) && s->last_change < now.tv_sec)		// ignore negative time
@@ -5305,25 +5311,84 @@ static void srv_update_status(struct server *s)
 	*s->adm_st_chg_cause = 0;
 }
 
+struct task *srv_cleanup_toremove_connections(struct task *task, void *context, unsigned short state)
+{
+	struct connection *conn;
+
+	while ((conn = LIST_POP_LOCKED(&toremove_connections[tid],
+	                               struct connection *, list)) != NULL) {
+		LIST_INIT(&conn->list);
+		conn->mux->destroy(conn);
+	}
+
+	return task;
+}
+
 struct task *srv_cleanup_idle_connections(struct task *task, void *context, unsigned short state)
 {
-	struct server *srv = context;
-	struct connection *conn, *conn_back;
-	unsigned int to_destroy = srv->curr_idle_thr[tid] / 2 + (srv->curr_idle_thr[tid] & 1);
-	unsigned int i = 0;
+	struct server *srv;
+	struct eb32_node *eb;
+	int i;
+	unsigned int next_wakeup;
+	int need_wakeup = 0;
 
+	HA_SPIN_LOCK(OTHER_LOCK, &idle_conn_srv_lock);
+	while (1) {
+		int srv_is_empty = 1;
 
+		eb = eb32_lookup_ge(&idle_conn_srv, now_ms - TIMER_LOOK_BACK);
+		if (!eb) {
+			/* we might have reached the end of the tree, typically because
+			 * <now_ms> is in the first half and we're first scanning the last
+			* half. Let's loop back to the beginning of the tree now.
+			*/
 
-	list_for_each_entry_safe(conn, conn_back, &srv->idle_orphan_conns[tid], list) {
-		if (i == to_destroy)
+			eb = eb32_first(&idle_conn_srv);
+			if (likely(!eb))
+				break;
+		}
+		if (tick_is_lt(now_ms, eb->key)) {
+			/* timer not expired yet, revisit it later */
+			next_wakeup = eb->key;
+			need_wakeup = 1;
 			break;
-		conn->mux->destroy(conn);
-		i++;
+		}
+		srv = eb32_entry(eb, struct server, idle_node);
+		for (i = 0; i < global.nbthread; i++) {
+			int max_conn = (srv->curr_idle_thr[i] / 2) +
+			    (srv->curr_idle_thr[i] & 1);
+			int j;
+			int did_remove = 0;
+
+			for (j = 0; j < max_conn; j++) {
+				struct connection *conn = LIST_POP_LOCKED(&srv->idle_orphan_conns[i], struct connection *, list);
+				if (!conn)
+					break;
+				did_remove = 1;
+				LIST_ADDQ_LOCKED(&toremove_connections[i], &conn->list);
+			}
+			if (did_remove && max_conn < srv->curr_idle_thr[i])
+				srv_is_empty = 0;
+			if (did_remove)
+				task_wakeup(idle_conn_cleanup[i], TASK_WOKEN_OTHER);
+		}
+		eb32_delete(&srv->idle_node);
+		if (!srv_is_empty) {
+			/* There are still more idle connections, add the
+			 * server back in the tree.
+			 */
+			srv->idle_node.key = tick_add(srv->pool_purge_delay,
+			    now_ms);
+			eb32_insert(&idle_conn_srv, &srv->idle_node);
+		}
 	}
-	if (!LIST_ISEMPTY(&srv->idle_orphan_conns[tid]))
-		task_schedule(task, tick_add(now_ms, srv->pool_purge_delay));
+	HA_SPIN_UNLOCK(OTHER_LOCK, &idle_conn_srv_lock);
+
+	if (need_wakeup)
+		task->expire = next_wakeup;
 	else
 		task->expire = TICK_ETERNITY;
+
 	return task;
 }
 /*