fq_codel: add batch ability to fq_codel_drop()
authorEric Dumazet <edumazet@google.com>
Sun, 1 May 2016 23:47:26 +0000 (16:47 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 3 May 2016 16:47:09 +0000 (12:47 -0400)
In presence of inelastic flows and stress, we can call
fq_codel_drop() for every packet entering fq_codel qdisc.

fq_codel_drop() is quite expensive, as it does a linear scan
of 4 KB of memory to find a fat flow.
Once found, it drops the oldest packet of this flow.

Instead of dropping a single packet, try to drop 50% of the backlog
of this fat flow, with a configurable limit of 64 packets per round.

TCA_FQ_CODEL_DROP_BATCH_SIZE is the new attribute to make this
limit configurable.

With this strategy the 4 KB search is amortized to a single cache line
per drop [1], so fq_codel_drop() no longer appears at the top of kernel
profile in presence of few inelastic flows.

[1] Assuming a 64byte cache line, and 1024 buckets

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dave Taht <dave.taht@gmail.com>
Cc: Jonathan Morton <chromatix99@gmail.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Dave Taht
Signed-off-by: David S. Miller <davem@davemloft.net>
include/uapi/linux/pkt_sched.h
net/sched/sch_fq_codel.c

index 1c78c74..a11afec 100644 (file)
@@ -718,6 +718,7 @@ enum {
        TCA_FQ_CODEL_FLOWS,
        TCA_FQ_CODEL_QUANTUM,
        TCA_FQ_CODEL_CE_THRESHOLD,
+       TCA_FQ_CODEL_DROP_BATCH_SIZE,
        __TCA_FQ_CODEL_MAX
 };
 
index a5e420b..e7b42b0 100644 (file)
@@ -59,6 +59,7 @@ struct fq_codel_sched_data {
        u32             flows_cnt;      /* number of flows */
        u32             perturbation;   /* hash perturbation */
        u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
+       u32             drop_batch_size;
        struct codel_params cparams;
        struct codel_stats cstats;
        u32             drop_overlimit;
@@ -135,17 +136,20 @@ static inline void flow_queue_add(struct fq_codel_flow *flow,
        skb->next = NULL;
 }
 
-static unsigned int fq_codel_drop(struct Qdisc *sch)
+static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets)
 {
        struct fq_codel_sched_data *q = qdisc_priv(sch);
        struct sk_buff *skb;
        unsigned int maxbacklog = 0, idx = 0, i, len;
        struct fq_codel_flow *flow;
+       unsigned int threshold;
 
-       /* Queue is full! Find the fat flow and drop packet from it.
+       /* Queue is full! Find the fat flow and drop packet(s) from it.
         * This might sound expensive, but with 1024 flows, we scan
         * 4KB of memory, and we dont need to handle a complex tree
         * in fast path (packet queue/enqueue) with many cache misses.
+        * In stress mode, we'll try to drop 64 packets from the flow,
+        * amortizing this linear lookup to one cache line per drop.
         */
        for (i = 0; i < q->flows_cnt; i++) {
                if (q->backlogs[i] > maxbacklog) {
@@ -153,15 +157,24 @@ static unsigned int fq_codel_drop(struct Qdisc *sch)
                        idx = i;
                }
        }
+
+       /* Our goal is to drop half of this fat flow backlog */
+       threshold = maxbacklog >> 1;
+
        flow = &q->flows[idx];
-       skb = dequeue_head(flow);
-       len = qdisc_pkt_len(skb);
+       len = 0;
+       i = 0;
+       do {
+               skb = dequeue_head(flow);
+               len += qdisc_pkt_len(skb);
+               kfree_skb(skb);
+       } while (++i < max_packets && len < threshold);
+
+       flow->dropped += i;
        q->backlogs[idx] -= len;
-       sch->q.qlen--;
-       qdisc_qstats_drop(sch);
-       qdisc_qstats_backlog_dec(sch, skb);
-       kfree_skb(skb);
-       flow->dropped++;
+       sch->qstats.drops += i;
+       sch->qstats.backlog -= len;
+       sch->q.qlen -= i;
        return idx;
 }
 
@@ -170,14 +183,14 @@ static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch)
        unsigned int prev_backlog;
 
        prev_backlog = sch->qstats.backlog;
-       fq_codel_drop(sch);
+       fq_codel_drop(sch, 1U);
        return prev_backlog - sch->qstats.backlog;
 }
 
 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 {
        struct fq_codel_sched_data *q = qdisc_priv(sch);
-       unsigned int idx, prev_backlog;
+       unsigned int idx, prev_backlog, prev_qlen;
        struct fq_codel_flow *flow;
        int uninitialized_var(ret);
 
@@ -206,16 +219,22 @@ static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
                return NET_XMIT_SUCCESS;
 
        prev_backlog = sch->qstats.backlog;
-       q->drop_overlimit++;
-       /* Return Congestion Notification only if we dropped a packet
-        * from this flow.
+       prev_qlen = sch->q.qlen;
+
+       /* fq_codel_drop() is quite expensive, as it performs a linear search
+        * in q->backlogs[] to find a fat flow.
+        * So instead of dropping a single packet, drop half of its backlog
+        * with a 64 packets limit to not add a too big cpu spike here.
         */
-       if (fq_codel_drop(sch) == idx)
-               return NET_XMIT_CN;
+       ret = fq_codel_drop(sch, q->drop_batch_size);
+
+       q->drop_overlimit += prev_qlen - sch->q.qlen;
 
-       /* As we dropped a packet, better let upper stack know this */
-       qdisc_tree_reduce_backlog(sch, 1, prev_backlog - sch->qstats.backlog);
-       return NET_XMIT_SUCCESS;
+       /* As we dropped packet(s), better let upper stack know this */
+       qdisc_tree_reduce_backlog(sch, prev_qlen - sch->q.qlen,
+                                 prev_backlog - sch->qstats.backlog);
+
+       return ret == idx ? NET_XMIT_CN : NET_XMIT_SUCCESS;
 }
 
 /* This is the specific function called from codel_dequeue()
@@ -335,6 +354,7 @@ static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
        [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
        [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
        [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
+       [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
 };
 
 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
@@ -386,6 +406,9 @@ static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
        if (tb[TCA_FQ_CODEL_QUANTUM])
                q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
 
+       if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
+               q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
+
        while (sch->q.qlen > sch->limit) {
                struct sk_buff *skb = fq_codel_dequeue(sch);
 
@@ -431,6 +454,7 @@ static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
 
        sch->limit = 10*1024;
        q->flows_cnt = 1024;
+       q->drop_batch_size = 64;
        q->quantum = psched_mtu(qdisc_dev(sch));
        q->perturbation = prandom_u32();
        INIT_LIST_HEAD(&q->new_flows);
@@ -489,6 +513,8 @@ static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
                        q->cparams.ecn) ||
            nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
                        q->quantum) ||
+           nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
+                       q->drop_batch_size) ||
            nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
                        q->flows_cnt))
                goto nla_put_failure;