net_sched: avoid too many hrtimer_start() calls
authorEric Dumazet <edumazet@google.com>
Mon, 23 May 2016 21:24:56 +0000 (14:24 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 24 May 2016 21:49:14 +0000 (14:49 -0700)
I found a serious performance bug in packet schedulers using hrtimers.

sch_htb and sch_fq are definitely impacted by this problem.

We constantly rearm high resolution timers if some packets are throttled
in one (or more) class, and other packets are flying through qdisc on
another (non throttled) class.

hrtimer_start() does not have the mod_timer() trick of doing nothing if
expires value does not change :

if (timer_pending(timer) &&
            timer->expires == expires)
                return 1;

This issue is particularly visible when multiple cpus can queue/dequeue
packets on the same qdisc, as hrtimer code has to lock a remote base.

I used following fix :

1) Change htb to use qdisc_watchdog_schedule_ns() instead of open-coding
it.

2) Cache watchdog prior expiration. hrtimer might provide this, but I
prefer to not rely on some hrtimer internal.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/pkt_sched.h
net/sched/sch_api.c
net/sched/sch_htb.c

index 401038d..fea53f4 100644 (file)
@@ -61,6 +61,7 @@ psched_tdiff_bounded(psched_time_t tv1, psched_time_t tv2, psched_time_t bound)
 }
 
 struct qdisc_watchdog {
+       u64             last_expires;
        struct hrtimer  timer;
        struct Qdisc    *qdisc;
 };
index 64f71a2..ddf047d 100644 (file)
@@ -607,6 +607,10 @@ void qdisc_watchdog_schedule_ns(struct qdisc_watchdog *wd, u64 expires, bool thr
        if (throttle)
                qdisc_throttled(wd->qdisc);
 
+       if (wd->last_expires == expires)
+               return;
+
+       wd->last_expires = expires;
        hrtimer_start(&wd->timer,
                      ns_to_ktime(expires),
                      HRTIMER_MODE_ABS_PINNED);
index f6bf581..d4b4218 100644 (file)
@@ -928,17 +928,10 @@ ok:
                }
        }
        qdisc_qstats_overlimit(sch);
-       if (likely(next_event > q->now)) {
-               if (!test_bit(__QDISC_STATE_DEACTIVATED,
-                             &qdisc_root_sleeping(q->watchdog.qdisc)->state)) {
-                       ktime_t time = ns_to_ktime(next_event);
-                       qdisc_throttled(q->watchdog.qdisc);
-                       hrtimer_start(&q->watchdog.timer, time,
-                                     HRTIMER_MODE_ABS_PINNED);
-               }
-       } else {
+       if (likely(next_event > q->now))
+               qdisc_watchdog_schedule_ns(&q->watchdog, next_event, true);
+       else
                schedule_work(&q->work);
-       }
 fin:
        return skb;
 }