BUG/MINOR: mux-h2: prevent past scheduling with idle connections

While reviewing HTTP/2 MUX timeout, it seems there is a possibility that
MUX task is requeued via h2c_update_timeout() with an already expired
date. This can happens with idle connections on two cases :
* first with shut timeout, as timer is not refreshed if already set
* second with http-request and keep-alive timers, which are based on
  idle_start

Queuing an already expired task is an undefined behavior. Fix this by
using task_wakeup() instead of task_queue() at the end of
h2c_update_timeout() if such case occurs.

This should be backported up to 2.6.

(cherry picked from commit 3ebdd3ae509236af1f19d80cab98b385a751426b)
Signed-off-by: Willy Tarreau <w@1wt.eu>
(cherry picked from commit 516121499c8edafb67a861bcd1bf41b3cc6f12d2)
Signed-off-by: Willy Tarreau <w@1wt.eu>

diff --git a/src/mux_h2.c b/src/mux_h2.c
index f823b92..e3116fd 100644 (file)
--- a/src/mux_h2.c
+++ b/src/mux_h2.c
@@ -732,6 +732,16 @@ static void h2c_update_timeout(struct h2c *h2c)
        } else {
                h2c->task->expire = TICK_ETERNITY;
        }
+
+       /* Timer may already be expired, in this case it must not be requeued.
+        * Currently it may happen with idle conn calculation based on shut or
+        * http-req/ka timeouts.
+        */
+       if (unlikely(tick_is_expired(h2c->task->expire, now_ms))) {
+               task_wakeup(h2c->task, TASK_WOKEN_TIMER);
+               goto leave;
+       }
+
        task_queue(h2c->task);
  leave:
        TRACE_LEAVE(H2_EV_H2C_WAKE);