From 77879147a3481babffd7e368d977ab682545a6bd Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Mon, 19 Sep 2016 23:39:11 -0400 Subject: net_sched: sch_fq: add low_rate_threshold parameter This commit adds to the fq module a low_rate_threshold parameter to insert a delay after all packets if the socket requests a pacing rate below the threshold. This helps achieve more precise control of the sending rate with low-rate paths, especially policers. The basic issue is that if a congestion control module detects a policer at a certain rate, it may want fq to be able to shape to that policed rate. That way the sender can avoid policer drops by having the packets arrive at the policer at or just under the policed rate. The default threshold of 550Kbps was chosen analytically so that for policers or links at 500Kbps or 512Kbps fq would very likely invoke this mechanism, even if the pacing rate was briefly slightly above the available bandwidth. This value was then empirically validated with two years of production testing on YouTube video servers. Signed-off-by: Van Jacobson Signed-off-by: Neal Cardwell Signed-off-by: Yuchung Cheng Signed-off-by: Nandita Dukkipati Signed-off-by: Eric Dumazet Signed-off-by: Soheil Hassas Yeganeh Signed-off-by: David S. Miller --- net/sched/sch_fq.c | 22 +++++++++++++++++++--- 1 file changed, 19 insertions(+), 3 deletions(-) (limited to 'net/sched/sch_fq.c') diff --git a/net/sched/sch_fq.c b/net/sched/sch_fq.c index dc52cc1..5dd929c 100644 --- a/net/sched/sch_fq.c +++ b/net/sched/sch_fq.c @@ -94,6 +94,7 @@ struct fq_sched_data { u32 flow_max_rate; /* optional max rate per flow */ u32 flow_plimit; /* max packets per flow */ u32 orphan_mask; /* mask for orphaned skb */ + u32 low_rate_threshold; struct rb_root *fq_root; u8 rate_enable; u8 fq_trees_log; @@ -433,7 +434,7 @@ static struct sk_buff *fq_dequeue(struct Qdisc *sch) struct fq_flow_head *head; struct sk_buff *skb; struct fq_flow *f; - u32 rate; + u32 rate, plen; skb = fq_dequeue_head(sch, &q->internal); if (skb) @@ -482,7 +483,7 @@ begin: prefetch(&skb->end); f->credit -= qdisc_pkt_len(skb); - if (f->credit > 0 || !q->rate_enable) + if (!q->rate_enable) goto out; /* Do not pace locally generated ack packets */ @@ -493,8 +494,15 @@ begin: if (skb->sk) rate = min(skb->sk->sk_pacing_rate, rate); + if (rate <= q->low_rate_threshold) { + f->credit = 0; + plen = qdisc_pkt_len(skb); + } else { + plen = max(qdisc_pkt_len(skb), q->quantum); + if (f->credit > 0) + goto out; + } if (rate != ~0U) { - u32 plen = max(qdisc_pkt_len(skb), q->quantum); u64 len = (u64)plen * NSEC_PER_SEC; if (likely(rate)) @@ -662,6 +670,7 @@ static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 }, + [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 }, }; static int fq_change(struct Qdisc *sch, struct nlattr *opt) @@ -716,6 +725,10 @@ static int fq_change(struct Qdisc *sch, struct nlattr *opt) if (tb[TCA_FQ_FLOW_MAX_RATE]) q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); + if (tb[TCA_FQ_LOW_RATE_THRESHOLD]) + q->low_rate_threshold = + nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]); + if (tb[TCA_FQ_RATE_ENABLE]) { u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); @@ -781,6 +794,7 @@ static int fq_init(struct Qdisc *sch, struct nlattr *opt) q->fq_root = NULL; q->fq_trees_log = ilog2(1024); q->orphan_mask = 1024 - 1; + q->low_rate_threshold = 550000 / 8; qdisc_watchdog_init(&q->watchdog, sch); if (opt) @@ -811,6 +825,8 @@ static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY, jiffies_to_usecs(q->flow_refill_delay)) || nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) || + nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD, + q->low_rate_threshold) || nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log)) goto nla_put_failure; -- cgit v1.1 From fefa569a9d4bc4b7758c0fddd75bb0382c95da77 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Thu, 22 Sep 2016 08:58:55 -0700 Subject: net_sched: sch_fq: account for schedule/timers drifts It looks like the following patch can make FQ very precise, even in VM or stressed hosts. It matters at high pacing rates. We take into account the difference between the time that was programmed when last packet was sent, and current time (a drift of tens of usecs is often observed) Add an EWMA of the unthrottle latency to help diagnostics. This latency is the difference between current time and oldest packet in delayed RB-tree. This accounts for the high resolution timer latency, but can be different under stress, as fq_check_throttled() can be opportunistically be called from a dequeue() called after an enqueue() for a different flow. Tested: // Start a 10Gbit flow $ netperf --google-pacing-rate 1250000000 -H lpaa24 -l 10000 -- -K bbr & Before patch : $ sar -n DEV 10 5 | grep eth0 | grep Average Average: eth0 17106.04 756876.84 1102.75 1119049.02 0.00 0.00 0.52 After patch : $ sar -n DEV 10 5 | grep eth0 | grep Average Average: eth0 17867.00 800245.90 1151.77 1183172.12 0.00 0.00 0.52 A new iproute2 tc can output the 'unthrottle latency' : $ tc -s qd sh dev eth0 | grep latency 0 gc, 0 highprio, 32490767 throttled, 2382 ns latency Signed-off-by: Eric Dumazet Signed-off-by: David S. Miller --- net/sched/sch_fq.c | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-) (limited to 'net/sched/sch_fq.c') diff --git a/net/sched/sch_fq.c b/net/sched/sch_fq.c index 5dd929c..18e7524 100644 --- a/net/sched/sch_fq.c +++ b/net/sched/sch_fq.c @@ -86,6 +86,7 @@ struct fq_sched_data { struct rb_root delayed; /* for rate limited flows */ u64 time_next_delayed_flow; + unsigned long unthrottle_latency_ns; struct fq_flow internal; /* for non classified or high prio packets */ u32 quantum; @@ -408,11 +409,19 @@ static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch, static void fq_check_throttled(struct fq_sched_data *q, u64 now) { + unsigned long sample; struct rb_node *p; if (q->time_next_delayed_flow > now) return; + /* Update unthrottle latency EWMA. + * This is cheap and can help diagnosing timer/latency problems. + */ + sample = (unsigned long)(now - q->time_next_delayed_flow); + q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3; + q->unthrottle_latency_ns += sample >> 3; + q->time_next_delayed_flow = ~0ULL; while ((p = rb_first(&q->delayed)) != NULL) { struct fq_flow *f = container_of(p, struct fq_flow, rate_node); @@ -515,7 +524,12 @@ begin: len = NSEC_PER_SEC; q->stat_pkts_too_long++; } - + /* Account for schedule/timers drifts. + * f->time_next_packet was set when prior packet was sent, + * and current time (@now) can be too late by tens of us. + */ + if (f->time_next_packet) + len -= min(len/2, now - f->time_next_packet); f->time_next_packet = now + len; } out: @@ -787,6 +801,7 @@ static int fq_init(struct Qdisc *sch, struct nlattr *opt) q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); q->flow_refill_delay = msecs_to_jiffies(40); q->flow_max_rate = ~0U; + q->time_next_delayed_flow = ~0ULL; q->rate_enable = 1; q->new_flows.first = NULL; q->old_flows.first = NULL; @@ -854,8 +869,8 @@ static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) st.flows = q->flows; st.inactive_flows = q->inactive_flows; st.throttled_flows = q->throttled_flows; - st.pad = 0; - + st.unthrottle_latency_ns = min_t(unsigned long, + q->unthrottle_latency_ns, ~0U); sch_tree_unlock(sch); return gnet_stats_copy_app(d, &st, sizeof(st)); -- cgit v1.1