BUG/MEDIUM: mux-h1: always apply the timeout on half-closed connections
authorWilly Tarreau <w@1wt.eu>
Tue, 8 Sep 2020 13:40:57 +0000 (15:40 +0200)
committerChristopher Faulet <cfaulet@haproxy.com>
Fri, 11 Sep 2020 08:06:10 +0000 (10:06 +0200)
The condition in h1_refresh_timeout() seems insufficient to properly
take care of the half-closed timeout, because depending on the ordering
of operations when performing the last send() to a client, the stream
may or may not still be there and we may fail to shrink the client
timeout on our last opportunity to do so.

Here we want to make sure that the timeout is always reduced when the
last chunk was sent and the shutdown completed, regardless of the
presence of a stream or not. This is what this patch does.

This should be backported as far as 2.0, and should fix the issue
reported in #541.

(cherry picked from commit 4313d5ae98b90613318da7e1181a6c4a1db29799)
Signed-off-by: Willy Tarreau <w@1wt.eu>
(cherry picked from commit ee6ed47bba75896b6bbdd15f6a622e88a8168390)
Signed-off-by: Christopher Faulet <cfaulet@haproxy.com>

src/mux_h1.c

index e430260..9145808 100644 (file)
@@ -461,11 +461,19 @@ static void h1_refresh_timeout(struct h1c *h1c)
 {
        if (h1c->task) {
                h1c->task->expire = TICK_ETERNITY;
-               if ((!h1c->h1s && !conn_is_back(h1c->conn)) || b_data(&h1c->obuf)) {
+               if (h1c->flags & H1C_F_CS_SHUTDOWN) {
+                       /* half-closed connections switch to clientfin/serverfin
+                        * timeouts so that we don't hang too long on clients
+                        * that have gone away (especially in tunnel mode).
+                        */
+                       h1c->task->expire = tick_add(now_ms, h1c->shut_timeout);
+                       task_queue(h1c->task);
+                       TRACE_DEVEL("refreshing connection's timeout (half-closed)", H1_EV_H1C_SEND, h1c->conn);
+               } else if ((!h1c->h1s && !conn_is_back(h1c->conn)) || b_data(&h1c->obuf)) {
                        /* front connections waiting for a stream, as well as any connection with
                         * pending data, need a timeout.
                         */
-                       h1c->task->expire = tick_add(now_ms, ((h1c->flags & (H1C_F_CS_SHUTW_NOW|H1C_F_CS_SHUTDOWN))
+                       h1c->task->expire = tick_add(now_ms, ((h1c->flags & H1C_F_CS_SHUTW_NOW)
                                                              ? h1c->shut_timeout
                                                              : h1c->timeout));
                        task_queue(h1c->task);