fq_codel: Fair Queue Codel AQM
authorEric Dumazet <edumazet@google.com>
Fri, 11 May 2012 09:30:50 +0000 (09:30 +0000)
committerDavid S. Miller <davem@davemloft.net>
Sat, 12 May 2012 19:53:42 +0000 (15:53 -0400)
Fair Queue Codel packet scheduler

Principles :

- Packets are classified (internal classifier or external) on flows.
- This is a Stochastic model (as we use a hash, several flows might
                              be hashed on same slot)
- Each flow has a CoDel managed queue.
- Flows are linked onto two (Round Robin) lists,
  so that new flows have priority on old ones.

- For a given flow, packets are not reordered (CoDel uses a FIFO)
- head drops only.
- ECN capability is on by default.
- Very low memory footprint (64 bytes per flow)

tc qdisc ... fq_codel [ limit PACKETS ] [ flows number ]
                      [ target TIME ] [ interval TIME ] [ noecn ]
                      [ quantum BYTES ]

defaults : 1024 flows, 10240 packets limit, quantum : device MTU
           target : 5ms (CoDel default)
           interval : 100ms (CoDel default)

Impressive results on load :

class htb 1:1 root leaf 10: prio 0 quantum 1514 rate 200000Kbit ceil 200000Kbit burst 1475b/8 mpu 0b overhead 0b cburst 1475b/8 mpu 0b overhead 0b level 0
 Sent 43304920109 bytes 33063109 pkt (dropped 0, overlimits 0 requeues 0)
 rate 201691Kbit 28595pps backlog 0b 312p requeues 0
 lended: 33063109 borrowed: 0 giants: 0
 tokens: -912 ctokens: -912

class fq_codel 10:1735 parent 10:
 (dropped 1292, overlimits 0 requeues 0)
 backlog 15140b 10p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 7.1ms
class fq_codel 10:4524 parent 10:
 (dropped 1291, overlimits 0 requeues 0)
 backlog 16654b 11p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 7.1ms
class fq_codel 10:4e74 parent 10:
 (dropped 1290, overlimits 0 requeues 0)
 backlog 6056b 4p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 6.4ms dropping drop_next 92.0ms
class fq_codel 10:628a parent 10:
 (dropped 1289, overlimits 0 requeues 0)
 backlog 7570b 5p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 5.4ms dropping drop_next 90.9ms
class fq_codel 10:a4b3 parent 10:
 (dropped 302, overlimits 0 requeues 0)
 backlog 16654b 11p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 7.1ms
class fq_codel 10:c3c2 parent 10:
 (dropped 1284, overlimits 0 requeues 0)
 backlog 13626b 9p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 5.9ms
class fq_codel 10:d331 parent 10:
 (dropped 299, overlimits 0 requeues 0)
 backlog 15140b 10p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 7.0ms
class fq_codel 10:d526 parent 10:
 (dropped 12160, overlimits 0 requeues 0)
 backlog 35870b 211p requeues 0
  deficit 1508 count 12160 lastcount 1 ldelay 15.3ms dropping drop_next 247us
class fq_codel 10:e2c6 parent 10:
 (dropped 1288, overlimits 0 requeues 0)
 backlog 15140b 10p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 7.1ms
class fq_codel 10:eab5 parent 10:
 (dropped 1285, overlimits 0 requeues 0)
 backlog 16654b 11p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 5.9ms
class fq_codel 10:f220 parent 10:
 (dropped 1289, overlimits 0 requeues 0)
 backlog 15140b 10p requeues 0
  deficit 1514 count 1 lastcount 1 ldelay 7.1ms

qdisc htb 1: root refcnt 6 r2q 10 default 1 direct_packets_stat 0 ver 3.17
 Sent 43331086547 bytes 33092812 pkt (dropped 0, overlimits 66063544 requeues 71)
 rate 201697Kbit 28602pps backlog 0b 260p requeues 71
qdisc fq_codel 10: parent 1:1 limit 10240p flows 65536 target 5.0ms interval 100.0ms ecn
 Sent 43331086547 bytes 33092812 pkt (dropped 949359, overlimits 0 requeues 0)
 rate 201697Kbit 28602pps backlog 189352b 260p requeues 0
  maxpacket 1514 drop_overlimit 0 new_flow_count 5582 ecn_mark 125593
  new_flows_len 0 old_flows_len 11

PING 172.30.42.18 (172.30.42.18) 56(84) bytes of data.
64 bytes from 172.30.42.18: icmp_req=1 ttl=64 time=0.227 ms
64 bytes from 172.30.42.18: icmp_req=2 ttl=64 time=0.165 ms
64 bytes from 172.30.42.18: icmp_req=3 ttl=64 time=0.166 ms
64 bytes from 172.30.42.18: icmp_req=4 ttl=64 time=0.151 ms
64 bytes from 172.30.42.18: icmp_req=5 ttl=64 time=0.164 ms
64 bytes from 172.30.42.18: icmp_req=6 ttl=64 time=0.172 ms
64 bytes from 172.30.42.18: icmp_req=7 ttl=64 time=0.175 ms
64 bytes from 172.30.42.18: icmp_req=8 ttl=64 time=0.183 ms
64 bytes from 172.30.42.18: icmp_req=9 ttl=64 time=0.158 ms
64 bytes from 172.30.42.18: icmp_req=10 ttl=64 time=0.200 ms

10 packets transmitted, 10 received, 0% packet loss, time 8999ms
rtt min/avg/max/mdev = 0.151/0.176/0.227/0.022 ms

Much better than SFQ because of priority given to new flows, and fast
path dirtying less cache lines.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/pkt_sched.h
net/sched/Kconfig
net/sched/Makefile
net/sched/sch_fq_codel.c [new file with mode: 0644]

index cde56c22bdabc88da5ba60809ddffd27356f43bc..32aef0a439ef120b14c196d092a6ae67edfb74cb 100644 (file)
@@ -681,4 +681,58 @@ struct tc_codel_xstats {
        __u32   dropping;  /* are we in dropping state ? */
 };
 
+/* FQ_CODEL */
+
+enum {
+       TCA_FQ_CODEL_UNSPEC,
+       TCA_FQ_CODEL_TARGET,
+       TCA_FQ_CODEL_LIMIT,
+       TCA_FQ_CODEL_INTERVAL,
+       TCA_FQ_CODEL_ECN,
+       TCA_FQ_CODEL_FLOWS,
+       TCA_FQ_CODEL_QUANTUM,
+       __TCA_FQ_CODEL_MAX
+};
+
+#define TCA_FQ_CODEL_MAX       (__TCA_FQ_CODEL_MAX - 1)
+
+enum {
+       TCA_FQ_CODEL_XSTATS_QDISC,
+       TCA_FQ_CODEL_XSTATS_CLASS,
+};
+
+struct tc_fq_codel_qd_stats {
+       __u32   maxpacket;      /* largest packet we've seen so far */
+       __u32   drop_overlimit; /* number of time max qdisc
+                                * packet limit was hit
+                                */
+       __u32   ecn_mark;       /* number of packets we ECN marked
+                                * instead of being dropped
+                                */
+       __u32   new_flow_count; /* number of time packets
+                                * created a 'new flow'
+                                */
+       __u32   new_flows_len;  /* count of flows in new list */
+       __u32   old_flows_len;  /* count of flows in old list */
+};
+
+struct tc_fq_codel_cl_stats {
+       __s32   deficit;
+       __u32   ldelay;         /* in-queue delay seen by most recently
+                                * dequeued packet
+                                */
+       __u32   count;
+       __u32   lastcount;
+       __u32   dropping;
+       __s32   drop_next;
+};
+
+struct tc_fq_codel_xstats {
+       __u32   type;
+       union {
+               struct tc_fq_codel_qd_stats qdisc_stats;
+               struct tc_fq_codel_cl_stats class_stats;
+       };
+};
+
 #endif
index fadd2522053d915dc117d75965b11a70924d2227..e7a8976bf25cc0b44da94748061256e4c782efdb 100644 (file)
@@ -261,6 +261,17 @@ config NET_SCH_CODEL
 
          If unsure, say N.
 
+config NET_SCH_FQ_CODEL
+       tristate "Fair Queue Controlled Delay AQM (FQ_CODEL)"
+       help
+         Say Y here if you want to use the FQ Controlled Delay (FQ_CODEL)
+         packet scheduling algorithm.
+
+         To compile this driver as a module, choose M here: the module
+         will be called sch_fq_codel.
+
+         If unsure, say N.
+
 config NET_SCH_INGRESS
        tristate "Ingress Qdisc"
        depends on NET_CLS_ACT
index 30fab03b8516e087aa27ebbeb9833df36c609317..5940a1992f0dec04f378f2d6af76aac3bd8bf4fc 100644 (file)
@@ -38,6 +38,7 @@ obj-$(CONFIG_NET_SCH_MQPRIO)  += sch_mqprio.o
 obj-$(CONFIG_NET_SCH_CHOKE)    += sch_choke.o
 obj-$(CONFIG_NET_SCH_QFQ)      += sch_qfq.o
 obj-$(CONFIG_NET_SCH_CODEL)    += sch_codel.o
+obj-$(CONFIG_NET_SCH_FQ_CODEL) += sch_fq_codel.o
 
 obj-$(CONFIG_NET_CLS_U32)      += cls_u32.o
 obj-$(CONFIG_NET_CLS_ROUTE4)   += cls_route.o
diff --git a/net/sched/sch_fq_codel.c b/net/sched/sch_fq_codel.c
new file mode 100644 (file)
index 0000000..a7b3754
--- /dev/null
@@ -0,0 +1,624 @@
+/*
+ * Fair Queue CoDel discipline
+ *
+ *     This program is free software; you can redistribute it and/or
+ *     modify it under the terms of the GNU General Public License
+ *     as published by the Free Software Foundation; either version
+ *     2 of the License, or (at your option) any later version.
+ *
+ *  Copyright (C) 2012 Eric Dumazet <edumazet@google.com>
+ */
+
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/jiffies.h>
+#include <linux/string.h>
+#include <linux/in.h>
+#include <linux/errno.h>
+#include <linux/init.h>
+#include <linux/skbuff.h>
+#include <linux/jhash.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+#include <net/netlink.h>
+#include <net/pkt_sched.h>
+#include <net/flow_keys.h>
+#include <net/codel.h>
+
+/*     Fair Queue CoDel.
+ *
+ * Principles :
+ * Packets are classified (internal classifier or external) on flows.
+ * This is a Stochastic model (as we use a hash, several flows
+ *                            might be hashed on same slot)
+ * Each flow has a CoDel managed queue.
+ * Flows are linked onto two (Round Robin) lists,
+ * so that new flows have priority on old ones.
+ *
+ * For a given flow, packets are not reordered (CoDel uses a FIFO)
+ * head drops only.
+ * ECN capability is on by default.
+ * Low memory footprint (64 bytes per flow)
+ */
+
+struct fq_codel_flow {
+       struct sk_buff    *head;
+       struct sk_buff    *tail;
+       struct list_head  flowchain;
+       int               deficit;
+       u32               dropped; /* number of drops (or ECN marks) on this flow */
+       struct codel_vars cvars;
+}; /* please try to keep this structure <= 64 bytes */
+
+struct fq_codel_sched_data {
+       struct tcf_proto *filter_list;  /* optional external classifier */
+       struct fq_codel_flow *flows;    /* Flows table [flows_cnt] */
+       u32             *backlogs;      /* backlog table [flows_cnt] */
+       u32             flows_cnt;      /* number of flows */
+       u32             perturbation;   /* hash perturbation */
+       u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
+       struct codel_params cparams;
+       struct codel_stats cstats;
+       u32             drop_overlimit;
+       u32             new_flow_count;
+
+       struct list_head new_flows;     /* list of new flows */
+       struct list_head old_flows;     /* list of old flows */
+};
+
+static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
+                                 const struct sk_buff *skb)
+{
+       struct flow_keys keys;
+       unsigned int hash;
+
+       skb_flow_dissect(skb, &keys);
+       hash = jhash_3words((__force u32)keys.dst,
+                           (__force u32)keys.src ^ keys.ip_proto,
+                           (__force u32)keys.ports, q->perturbation);
+       return ((u64)hash * q->flows_cnt) >> 32;
+}
+
+static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
+                                     int *qerr)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       struct tcf_result res;
+       int result;
+
+       if (TC_H_MAJ(skb->priority) == sch->handle &&
+           TC_H_MIN(skb->priority) > 0 &&
+           TC_H_MIN(skb->priority) <= q->flows_cnt)
+               return TC_H_MIN(skb->priority);
+
+       if (!q->filter_list)
+               return fq_codel_hash(q, skb) + 1;
+
+       *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
+       result = tc_classify(skb, q->filter_list, &res);
+       if (result >= 0) {
+#ifdef CONFIG_NET_CLS_ACT
+               switch (result) {
+               case TC_ACT_STOLEN:
+               case TC_ACT_QUEUED:
+                       *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
+               case TC_ACT_SHOT:
+                       return 0;
+               }
+#endif
+               if (TC_H_MIN(res.classid) <= q->flows_cnt)
+                       return TC_H_MIN(res.classid);
+       }
+       return 0;
+}
+
+/* helper functions : might be changed when/if skb use a standard list_head */
+
+/* remove one skb from head of slot queue */
+static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
+{
+       struct sk_buff *skb = flow->head;
+
+       flow->head = skb->next;
+       skb->next = NULL;
+       return skb;
+}
+
+/* add skb to flow queue (tail add) */
+static inline void flow_queue_add(struct fq_codel_flow *flow,
+                                 struct sk_buff *skb)
+{
+       if (flow->head == NULL)
+               flow->head = skb;
+       else
+               flow->tail->next = skb;
+       flow->tail = skb;
+       skb->next = NULL;
+}
+
+static unsigned int fq_codel_drop(struct Qdisc *sch)
+{
+       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;
+
+       /* Queue is full! Find the fat flow and drop packet 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.
+        */
+       for (i = 0; i < q->flows_cnt; i++) {
+               if (q->backlogs[i] > maxbacklog) {
+                       maxbacklog = q->backlogs[i];
+                       idx = i;
+               }
+       }
+       flow = &q->flows[idx];
+       skb = dequeue_head(flow);
+       len = qdisc_pkt_len(skb);
+       q->backlogs[idx] -= len;
+       kfree_skb(skb);
+       sch->q.qlen--;
+       sch->qstats.drops++;
+       sch->qstats.backlog -= len;
+       flow->dropped++;
+       return idx;
+}
+
+static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       unsigned int idx;
+       struct fq_codel_flow *flow;
+       int uninitialized_var(ret);
+
+       idx = fq_codel_classify(skb, sch, &ret);
+       if (idx == 0) {
+               if (ret & __NET_XMIT_BYPASS)
+                       sch->qstats.drops++;
+               kfree_skb(skb);
+               return ret;
+       }
+       idx--;
+
+       codel_set_enqueue_time(skb);
+       flow = &q->flows[idx];
+       flow_queue_add(flow, skb);
+       q->backlogs[idx] += qdisc_pkt_len(skb);
+       sch->qstats.backlog += qdisc_pkt_len(skb);
+
+       if (list_empty(&flow->flowchain)) {
+               list_add_tail(&flow->flowchain, &q->new_flows);
+               codel_vars_init(&flow->cvars);
+               q->new_flow_count++;
+               flow->deficit = q->quantum;
+               flow->dropped = 0;
+       }
+       if (++sch->q.qlen < sch->limit)
+               return NET_XMIT_SUCCESS;
+
+       q->drop_overlimit++;
+       /* Return Congestion Notification only if we dropped a packet
+        * from this flow.
+        */
+       if (fq_codel_drop(sch) == idx)
+               return NET_XMIT_CN;
+
+       /* As we dropped a packet, better let upper stack know this */
+       qdisc_tree_decrease_qlen(sch, 1);
+       return NET_XMIT_SUCCESS;
+}
+
+/* This is the specific function called from codel_dequeue()
+ * to dequeue a packet from queue. Note: backlog is handled in
+ * codel, we dont need to reduce it here.
+ */
+static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch)
+{
+       struct fq_codel_flow *flow;
+       struct sk_buff *skb = NULL;
+
+       flow = container_of(vars, struct fq_codel_flow, cvars);
+       if (flow->head) {
+               skb = dequeue_head(flow);
+               sch->qstats.backlog -= qdisc_pkt_len(skb);
+               sch->q.qlen--;
+       }
+       return skb;
+}
+
+static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       struct sk_buff *skb;
+       struct fq_codel_flow *flow;
+       struct list_head *head;
+       u32 prev_drop_count, prev_ecn_mark;
+
+begin:
+       head = &q->new_flows;
+       if (list_empty(head)) {
+               head = &q->old_flows;
+               if (list_empty(head))
+                       return NULL;
+       }
+       flow = list_first_entry(head, struct fq_codel_flow, flowchain);
+
+       if (flow->deficit <= 0) {
+               flow->deficit += q->quantum;
+               list_move_tail(&flow->flowchain, &q->old_flows);
+               goto begin;
+       }
+
+       prev_drop_count = q->cstats.drop_count;
+       prev_ecn_mark = q->cstats.ecn_mark;
+
+       skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats,
+                           dequeue, &q->backlogs[flow - q->flows]);
+
+       flow->dropped += q->cstats.drop_count - prev_drop_count;
+       flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
+
+       if (!skb) {
+               /* force a pass through old_flows to prevent starvation */
+               if ((head == &q->new_flows) && !list_empty(&q->old_flows))
+                       list_move_tail(&flow->flowchain, &q->old_flows);
+               else
+                       list_del_init(&flow->flowchain);
+               goto begin;
+       }
+       qdisc_bstats_update(sch, skb);
+       flow->deficit -= qdisc_pkt_len(skb);
+       /* We cant call qdisc_tree_decrease_qlen() if our qlen is 0,
+        * or HTB crashes. Defer it for next round.
+        */
+       if (q->cstats.drop_count && sch->q.qlen) {
+               qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
+               q->cstats.drop_count = 0;
+       }
+       return skb;
+}
+
+static void fq_codel_reset(struct Qdisc *sch)
+{
+       struct sk_buff *skb;
+
+       while ((skb = fq_codel_dequeue(sch)) != NULL)
+               kfree_skb(skb);
+}
+
+static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
+       [TCA_FQ_CODEL_TARGET]   = { .type = NLA_U32 },
+       [TCA_FQ_CODEL_LIMIT]    = { .type = NLA_U32 },
+       [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
+       [TCA_FQ_CODEL_ECN]      = { .type = NLA_U32 },
+       [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
+       [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
+};
+
+static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
+       int err;
+
+       if (!opt)
+               return -EINVAL;
+
+       err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
+       if (err < 0)
+               return err;
+       if (tb[TCA_FQ_CODEL_FLOWS]) {
+               if (q->flows)
+                       return -EINVAL;
+               q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
+               if (!q->flows_cnt ||
+                   q->flows_cnt > 65536)
+                       return -EINVAL;
+       }
+       sch_tree_lock(sch);
+
+       if (tb[TCA_FQ_CODEL_TARGET]) {
+               u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
+
+               q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
+       }
+
+       if (tb[TCA_FQ_CODEL_INTERVAL]) {
+               u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
+
+               q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
+       }
+
+       if (tb[TCA_FQ_CODEL_LIMIT])
+               sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
+
+       if (tb[TCA_FQ_CODEL_ECN])
+               q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
+
+       if (tb[TCA_FQ_CODEL_QUANTUM])
+               q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
+
+       while (sch->q.qlen > sch->limit) {
+               struct sk_buff *skb = fq_codel_dequeue(sch);
+
+               kfree_skb(skb);
+               q->cstats.drop_count++;
+       }
+       qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
+       q->cstats.drop_count = 0;
+
+       sch_tree_unlock(sch);
+       return 0;
+}
+
+static void *fq_codel_zalloc(size_t sz)
+{
+       void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
+
+       if (!ptr)
+               ptr = vzalloc(sz);
+       return ptr;
+}
+
+static void fq_codel_free(void *addr)
+{
+       if (addr) {
+               if (is_vmalloc_addr(addr))
+                       vfree(addr);
+               else
+                       kfree(addr);
+       }
+}
+
+static void fq_codel_destroy(struct Qdisc *sch)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+
+       tcf_destroy_chain(&q->filter_list);
+       fq_codel_free(q->backlogs);
+       fq_codel_free(q->flows);
+}
+
+static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       int i;
+
+       sch->limit = 10*1024;
+       q->flows_cnt = 1024;
+       q->quantum = psched_mtu(qdisc_dev(sch));
+       q->perturbation = net_random();
+       INIT_LIST_HEAD(&q->new_flows);
+       INIT_LIST_HEAD(&q->old_flows);
+       codel_params_init(&q->cparams);
+       codel_stats_init(&q->cstats);
+       q->cparams.ecn = true;
+
+       if (opt) {
+               int err = fq_codel_change(sch, opt);
+               if (err)
+                       return err;
+       }
+
+       if (!q->flows) {
+               q->flows = fq_codel_zalloc(q->flows_cnt *
+                                          sizeof(struct fq_codel_flow));
+               if (!q->flows)
+                       return -ENOMEM;
+               q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
+               if (!q->backlogs) {
+                       fq_codel_free(q->flows);
+                       return -ENOMEM;
+               }
+               for (i = 0; i < q->flows_cnt; i++) {
+                       struct fq_codel_flow *flow = q->flows + i;
+
+                       INIT_LIST_HEAD(&flow->flowchain);
+               }
+       }
+       if (sch->limit >= 1)
+               sch->flags |= TCQ_F_CAN_BYPASS;
+       else
+               sch->flags &= ~TCQ_F_CAN_BYPASS;
+       return 0;
+}
+
+static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       struct nlattr *opts;
+
+       opts = nla_nest_start(skb, TCA_OPTIONS);
+       if (opts == NULL)
+               goto nla_put_failure;
+
+       if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
+                       codel_time_to_us(q->cparams.target)) ||
+           nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
+                       sch->limit) ||
+           nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
+                       codel_time_to_us(q->cparams.interval)) ||
+           nla_put_u32(skb, TCA_FQ_CODEL_ECN,
+                       q->cparams.ecn) ||
+           nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
+                       q->quantum) ||
+           nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
+                       q->flows_cnt))
+               goto nla_put_failure;
+
+       nla_nest_end(skb, opts);
+       return skb->len;
+
+nla_put_failure:
+       return -1;
+}
+
+static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       struct tc_fq_codel_xstats st = {
+               .type                           = TCA_FQ_CODEL_XSTATS_QDISC,
+               .qdisc_stats.maxpacket          = q->cstats.maxpacket,
+               .qdisc_stats.drop_overlimit     = q->drop_overlimit,
+               .qdisc_stats.ecn_mark           = q->cstats.ecn_mark,
+               .qdisc_stats.new_flow_count     = q->new_flow_count,
+       };
+       struct list_head *pos;
+
+       list_for_each(pos, &q->new_flows)
+               st.qdisc_stats.new_flows_len++;
+
+       list_for_each(pos, &q->old_flows)
+               st.qdisc_stats.old_flows_len++;
+
+       return gnet_stats_copy_app(d, &st, sizeof(st));
+}
+
+static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
+{
+       return NULL;
+}
+
+static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
+{
+       return 0;
+}
+
+static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
+                             u32 classid)
+{
+       /* we cannot bypass queue discipline anymore */
+       sch->flags &= ~TCQ_F_CAN_BYPASS;
+       return 0;
+}
+
+static void fq_codel_put(struct Qdisc *q, unsigned long cl)
+{
+}
+
+static struct tcf_proto **fq_codel_find_tcf(struct Qdisc *sch, unsigned long cl)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+
+       if (cl)
+               return NULL;
+       return &q->filter_list;
+}
+
+static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
+                         struct sk_buff *skb, struct tcmsg *tcm)
+{
+       tcm->tcm_handle |= TC_H_MIN(cl);
+       return 0;
+}
+
+static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
+                                    struct gnet_dump *d)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       u32 idx = cl - 1;
+       struct gnet_stats_queue qs = { 0 };
+       struct tc_fq_codel_xstats xstats;
+
+       if (idx < q->flows_cnt) {
+               const struct fq_codel_flow *flow = &q->flows[idx];
+               const struct sk_buff *skb = flow->head;
+
+               memset(&xstats, 0, sizeof(xstats));
+               xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
+               xstats.class_stats.deficit = flow->deficit;
+               xstats.class_stats.ldelay =
+                       codel_time_to_us(flow->cvars.ldelay);
+               xstats.class_stats.count = flow->cvars.count;
+               xstats.class_stats.lastcount = flow->cvars.lastcount;
+               xstats.class_stats.dropping = flow->cvars.dropping;
+               if (flow->cvars.dropping) {
+                       codel_tdiff_t delta = flow->cvars.drop_next -
+                                             codel_get_time();
+
+                       xstats.class_stats.drop_next = (delta >= 0) ?
+                               codel_time_to_us(delta) :
+                               -codel_time_to_us(-delta);
+               }
+               while (skb) {
+                       qs.qlen++;
+                       skb = skb->next;
+               }
+               qs.backlog = q->backlogs[idx];
+               qs.drops = flow->dropped;
+       }
+       if (gnet_stats_copy_queue(d, &qs) < 0)
+               return -1;
+       if (idx < q->flows_cnt)
+               return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
+       return 0;
+}
+
+static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
+{
+       struct fq_codel_sched_data *q = qdisc_priv(sch);
+       unsigned int i;
+
+       if (arg->stop)
+               return;
+
+       for (i = 0; i < q->flows_cnt; i++) {
+               if (list_empty(&q->flows[i].flowchain) ||
+                   arg->count < arg->skip) {
+                       arg->count++;
+                       continue;
+               }
+               if (arg->fn(sch, i + 1, arg) < 0) {
+                       arg->stop = 1;
+                       break;
+               }
+               arg->count++;
+       }
+}
+
+static const struct Qdisc_class_ops fq_codel_class_ops = {
+       .leaf           =       fq_codel_leaf,
+       .get            =       fq_codel_get,
+       .put            =       fq_codel_put,
+       .tcf_chain      =       fq_codel_find_tcf,
+       .bind_tcf       =       fq_codel_bind,
+       .unbind_tcf     =       fq_codel_put,
+       .dump           =       fq_codel_dump_class,
+       .dump_stats     =       fq_codel_dump_class_stats,
+       .walk           =       fq_codel_walk,
+};
+
+static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
+       .cl_ops         =       &fq_codel_class_ops,
+       .id             =       "fq_codel",
+       .priv_size      =       sizeof(struct fq_codel_sched_data),
+       .enqueue        =       fq_codel_enqueue,
+       .dequeue        =       fq_codel_dequeue,
+       .peek           =       qdisc_peek_dequeued,
+       .drop           =       fq_codel_drop,
+       .init           =       fq_codel_init,
+       .reset          =       fq_codel_reset,
+       .destroy        =       fq_codel_destroy,
+       .change         =       fq_codel_change,
+       .dump           =       fq_codel_dump,
+       .dump_stats =   fq_codel_dump_stats,
+       .owner          =       THIS_MODULE,
+};
+
+static int __init fq_codel_module_init(void)
+{
+       return register_qdisc(&fq_codel_qdisc_ops);
+}
+
+static void __exit fq_codel_module_exit(void)
+{
+       unregister_qdisc(&fq_codel_qdisc_ops);
+}
+
+module_init(fq_codel_module_init)
+module_exit(fq_codel_module_exit)
+MODULE_AUTHOR("Eric Dumazet");
+MODULE_LICENSE("GPL");