On Wed, Oct 22, 2008 at 10:06:58AM +0400, Badalian Vyacheslav wrote:
Hi!
So it looks like htb_hysteresis change could be not enough, and it's
a pity because this could be easiest to "reverse" in the kernel.
So I guess this patch seems to make some difference, but it also could
be hard to convince people to fix it this way, because it makes
scheduling less exact.
Alas I have still no idea of the real reason. IMHO this should be
rather debugged by hrtimers/NMI people, but there were no respose last
time I Cc-ed them - anyway I added linux-kernel to Cc again here.
I attach below a slightly modified version of the previous patch,
which lets for more exact scheduling, but could be more vulnerable for
this bug. Alas, even if it works, it still could be not the final
solution of this problem.
Thanks,
Jarek P.
--- (testing patch #2)
net/sched/sch_htb.c | 8 +++++++-
1 files changed, 7 insertions(+), 1 deletions(-)
diff --git a/net/sched/sch_htb.c b/net/sched/sch_htb.c
index 30c999c..ff9e965 100644
--- a/net/sched/sch_htb.c
+++ b/net/sched/sch_htb.c
@@ -162,6 +162,7 @@ struct htb_sched {
int rate2quantum; /* quant = rate / rate2quantum */
psched_time_t now; /* cached dequeue time */
+ psched_time_t next_watchdog;
struct qdisc_watchdog watchdog;
/* non shaped skbs; let them go directly thru */
@@ -920,7 +921,11 @@ static struct sk_buff *htb_dequeue(struct Qdisc *sch)
}
}
sch->qstats.overlimits++;
- qdisc_watchdog_schedule(&q->watchdog, next_event);
+ if (q->next_watchdog < q->now || next_event <=
+ q->next_watchdog - PSCHED_TICKS_PER_SEC / (10 * HZ)) {
+ qdisc_watchdog_schedule(&q->watchdog, next_event);
+ q->next_watchdog = next_event;
+ }
fin:
return skb;
}
@@ -973,6 +978,7 @@ static void htb_reset(struct Qdisc *sch)
}
}
qdisc_watchdog_cancel(&q->watchdog);
+ q->next_watchdog = 0;
__skb_queue_purge(&q->direct_queue);
sch->q.qlen = 0;
memset(q->row, 0, sizeof(q->row));
--
