2 * Fair Queue CoDel discipline
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
12 #include <linux/module.h>
13 #include <linux/types.h>
14 #include <linux/kernel.h>
15 #include <linux/jiffies.h>
16 #include <linux/string.h>
18 #include <linux/errno.h>
19 #include <linux/init.h>
20 #include <linux/skbuff.h>
21 #include <linux/jhash.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
24 #include <net/netlink.h>
25 #include <net/pkt_sched.h>
26 #include <net/codel.h>
31 * Packets are classified (internal classifier or external) on flows.
32 * This is a Stochastic model (as we use a hash, several flows
33 * might be hashed on same slot)
34 * Each flow has a CoDel managed queue.
35 * Flows are linked onto two (Round Robin) lists,
36 * so that new flows have priority on old ones.
38 * For a given flow, packets are not reordered (CoDel uses a FIFO)
40 * ECN capability is on by default.
41 * Low memory footprint (64 bytes per flow)
44 struct fq_codel_flow {
47 struct list_head flowchain;
49 u32 dropped; /* number of drops (or ECN marks) on this flow */
50 struct codel_vars cvars;
51 }; /* please try to keep this structure <= 64 bytes */
53 struct fq_codel_sched_data {
54 struct tcf_proto __rcu *filter_list; /* optional external classifier */
55 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
56 u32 *backlogs; /* backlog table [flows_cnt] */
57 u32 flows_cnt; /* number of flows */
58 u32 perturbation; /* hash perturbation */
59 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
60 struct codel_params cparams;
61 struct codel_stats cstats;
65 struct list_head new_flows; /* list of new flows */
66 struct list_head old_flows; /* list of old flows */
69 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
72 u32 hash = skb_get_hash_perturb(skb, q->perturbation);
74 return reciprocal_scale(hash, q->flows_cnt);
77 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
80 struct fq_codel_sched_data *q = qdisc_priv(sch);
81 struct tcf_proto *filter;
82 struct tcf_result res;
85 if (TC_H_MAJ(skb->priority) == sch->handle &&
86 TC_H_MIN(skb->priority) > 0 &&
87 TC_H_MIN(skb->priority) <= q->flows_cnt)
88 return TC_H_MIN(skb->priority);
90 filter = rcu_dereference_bh(q->filter_list);
92 return fq_codel_hash(q, skb) + 1;
94 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
95 result = tc_classify(skb, filter, &res);
97 #ifdef CONFIG_NET_CLS_ACT
101 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
106 if (TC_H_MIN(res.classid) <= q->flows_cnt)
107 return TC_H_MIN(res.classid);
112 /* helper functions : might be changed when/if skb use a standard list_head */
114 /* remove one skb from head of slot queue */
115 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
117 struct sk_buff *skb = flow->head;
119 flow->head = skb->next;
124 /* add skb to flow queue (tail add) */
125 static inline void flow_queue_add(struct fq_codel_flow *flow,
128 if (flow->head == NULL)
131 flow->tail->next = skb;
136 static unsigned int fq_codel_drop(struct Qdisc *sch)
138 struct fq_codel_sched_data *q = qdisc_priv(sch);
140 unsigned int maxbacklog = 0, idx = 0, i, len;
141 struct fq_codel_flow *flow;
143 /* Queue is full! Find the fat flow and drop packet from it.
144 * This might sound expensive, but with 1024 flows, we scan
145 * 4KB of memory, and we dont need to handle a complex tree
146 * in fast path (packet queue/enqueue) with many cache misses.
148 for (i = 0; i < q->flows_cnt; i++) {
149 if (q->backlogs[i] > maxbacklog) {
150 maxbacklog = q->backlogs[i];
154 flow = &q->flows[idx];
155 skb = dequeue_head(flow);
156 len = qdisc_pkt_len(skb);
157 q->backlogs[idx] -= len;
159 qdisc_qstats_drop(sch);
160 qdisc_qstats_backlog_dec(sch, skb);
166 static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch)
168 unsigned int prev_backlog;
170 prev_backlog = sch->qstats.backlog;
172 return prev_backlog - sch->qstats.backlog;
175 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
177 struct fq_codel_sched_data *q = qdisc_priv(sch);
179 struct fq_codel_flow *flow;
180 int uninitialized_var(ret);
182 idx = fq_codel_classify(skb, sch, &ret);
184 if (ret & __NET_XMIT_BYPASS)
185 qdisc_qstats_drop(sch);
191 codel_set_enqueue_time(skb);
192 flow = &q->flows[idx];
193 flow_queue_add(flow, skb);
194 q->backlogs[idx] += qdisc_pkt_len(skb);
195 qdisc_qstats_backlog_inc(sch, skb);
197 if (list_empty(&flow->flowchain)) {
198 list_add_tail(&flow->flowchain, &q->new_flows);
200 flow->deficit = q->quantum;
203 if (++sch->q.qlen <= sch->limit)
204 return NET_XMIT_SUCCESS;
207 /* Return Congestion Notification only if we dropped a packet
210 if (fq_codel_drop(sch) == idx)
213 /* As we dropped a packet, better let upper stack know this */
214 qdisc_tree_decrease_qlen(sch, 1);
215 return NET_XMIT_SUCCESS;
218 /* This is the specific function called from codel_dequeue()
219 * to dequeue a packet from queue. Note: backlog is handled in
220 * codel, we dont need to reduce it here.
222 static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch)
224 struct fq_codel_sched_data *q = qdisc_priv(sch);
225 struct fq_codel_flow *flow;
226 struct sk_buff *skb = NULL;
228 flow = container_of(vars, struct fq_codel_flow, cvars);
230 skb = dequeue_head(flow);
231 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
237 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
239 struct fq_codel_sched_data *q = qdisc_priv(sch);
241 struct fq_codel_flow *flow;
242 struct list_head *head;
243 u32 prev_drop_count, prev_ecn_mark;
246 head = &q->new_flows;
247 if (list_empty(head)) {
248 head = &q->old_flows;
249 if (list_empty(head))
252 flow = list_first_entry(head, struct fq_codel_flow, flowchain);
254 if (flow->deficit <= 0) {
255 flow->deficit += q->quantum;
256 list_move_tail(&flow->flowchain, &q->old_flows);
260 prev_drop_count = q->cstats.drop_count;
261 prev_ecn_mark = q->cstats.ecn_mark;
263 skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats,
266 flow->dropped += q->cstats.drop_count - prev_drop_count;
267 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
270 /* force a pass through old_flows to prevent starvation */
271 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
272 list_move_tail(&flow->flowchain, &q->old_flows);
274 list_del_init(&flow->flowchain);
277 qdisc_bstats_update(sch, skb);
278 flow->deficit -= qdisc_pkt_len(skb);
279 /* We cant call qdisc_tree_decrease_qlen() if our qlen is 0,
280 * or HTB crashes. Defer it for next round.
282 if (q->cstats.drop_count && sch->q.qlen) {
283 qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
284 q->cstats.drop_count = 0;
289 static void fq_codel_reset(struct Qdisc *sch)
293 while ((skb = fq_codel_dequeue(sch)) != NULL)
297 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
298 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
299 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
300 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
301 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
302 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
303 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
304 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
307 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
309 struct fq_codel_sched_data *q = qdisc_priv(sch);
310 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
316 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
319 if (tb[TCA_FQ_CODEL_FLOWS]) {
322 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
324 q->flows_cnt > 65536)
329 if (tb[TCA_FQ_CODEL_TARGET]) {
330 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
332 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
335 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
336 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
338 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
341 if (tb[TCA_FQ_CODEL_INTERVAL]) {
342 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
344 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
347 if (tb[TCA_FQ_CODEL_LIMIT])
348 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
350 if (tb[TCA_FQ_CODEL_ECN])
351 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
353 if (tb[TCA_FQ_CODEL_QUANTUM])
354 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
356 while (sch->q.qlen > sch->limit) {
357 struct sk_buff *skb = fq_codel_dequeue(sch);
360 q->cstats.drop_count++;
362 qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
363 q->cstats.drop_count = 0;
365 sch_tree_unlock(sch);
369 static void *fq_codel_zalloc(size_t sz)
371 void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
378 static void fq_codel_free(void *addr)
383 static void fq_codel_destroy(struct Qdisc *sch)
385 struct fq_codel_sched_data *q = qdisc_priv(sch);
387 tcf_destroy_chain(&q->filter_list);
388 fq_codel_free(q->backlogs);
389 fq_codel_free(q->flows);
392 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
394 struct fq_codel_sched_data *q = qdisc_priv(sch);
397 sch->limit = 10*1024;
399 q->quantum = psched_mtu(qdisc_dev(sch));
400 q->perturbation = prandom_u32();
401 INIT_LIST_HEAD(&q->new_flows);
402 INIT_LIST_HEAD(&q->old_flows);
403 codel_params_init(&q->cparams, sch);
404 codel_stats_init(&q->cstats);
405 q->cparams.ecn = true;
408 int err = fq_codel_change(sch, opt);
414 q->flows = fq_codel_zalloc(q->flows_cnt *
415 sizeof(struct fq_codel_flow));
418 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
420 fq_codel_free(q->flows);
423 for (i = 0; i < q->flows_cnt; i++) {
424 struct fq_codel_flow *flow = q->flows + i;
426 INIT_LIST_HEAD(&flow->flowchain);
427 codel_vars_init(&flow->cvars);
431 sch->flags |= TCQ_F_CAN_BYPASS;
433 sch->flags &= ~TCQ_F_CAN_BYPASS;
437 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
439 struct fq_codel_sched_data *q = qdisc_priv(sch);
442 opts = nla_nest_start(skb, TCA_OPTIONS);
444 goto nla_put_failure;
446 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
447 codel_time_to_us(q->cparams.target)) ||
448 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
450 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
451 codel_time_to_us(q->cparams.interval)) ||
452 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
454 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
456 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
458 goto nla_put_failure;
460 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
461 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
462 codel_time_to_us(q->cparams.ce_threshold)))
463 goto nla_put_failure;
465 return nla_nest_end(skb, opts);
471 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
473 struct fq_codel_sched_data *q = qdisc_priv(sch);
474 struct tc_fq_codel_xstats st = {
475 .type = TCA_FQ_CODEL_XSTATS_QDISC,
477 struct list_head *pos;
479 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
480 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
481 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
482 st.qdisc_stats.new_flow_count = q->new_flow_count;
483 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
485 list_for_each(pos, &q->new_flows)
486 st.qdisc_stats.new_flows_len++;
488 list_for_each(pos, &q->old_flows)
489 st.qdisc_stats.old_flows_len++;
491 return gnet_stats_copy_app(d, &st, sizeof(st));
494 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
499 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
504 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
507 /* we cannot bypass queue discipline anymore */
508 sch->flags &= ~TCQ_F_CAN_BYPASS;
512 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
516 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
519 struct fq_codel_sched_data *q = qdisc_priv(sch);
523 return &q->filter_list;
526 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
527 struct sk_buff *skb, struct tcmsg *tcm)
529 tcm->tcm_handle |= TC_H_MIN(cl);
533 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
536 struct fq_codel_sched_data *q = qdisc_priv(sch);
538 struct gnet_stats_queue qs = { 0 };
539 struct tc_fq_codel_xstats xstats;
541 if (idx < q->flows_cnt) {
542 const struct fq_codel_flow *flow = &q->flows[idx];
543 const struct sk_buff *skb = flow->head;
545 memset(&xstats, 0, sizeof(xstats));
546 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
547 xstats.class_stats.deficit = flow->deficit;
548 xstats.class_stats.ldelay =
549 codel_time_to_us(flow->cvars.ldelay);
550 xstats.class_stats.count = flow->cvars.count;
551 xstats.class_stats.lastcount = flow->cvars.lastcount;
552 xstats.class_stats.dropping = flow->cvars.dropping;
553 if (flow->cvars.dropping) {
554 codel_tdiff_t delta = flow->cvars.drop_next -
557 xstats.class_stats.drop_next = (delta >= 0) ?
558 codel_time_to_us(delta) :
559 -codel_time_to_us(-delta);
565 qs.backlog = q->backlogs[idx];
566 qs.drops = flow->dropped;
568 if (gnet_stats_copy_queue(d, NULL, &qs, 0) < 0)
570 if (idx < q->flows_cnt)
571 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
575 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
577 struct fq_codel_sched_data *q = qdisc_priv(sch);
583 for (i = 0; i < q->flows_cnt; i++) {
584 if (list_empty(&q->flows[i].flowchain) ||
585 arg->count < arg->skip) {
589 if (arg->fn(sch, i + 1, arg) < 0) {
597 static const struct Qdisc_class_ops fq_codel_class_ops = {
598 .leaf = fq_codel_leaf,
601 .tcf_chain = fq_codel_find_tcf,
602 .bind_tcf = fq_codel_bind,
603 .unbind_tcf = fq_codel_put,
604 .dump = fq_codel_dump_class,
605 .dump_stats = fq_codel_dump_class_stats,
606 .walk = fq_codel_walk,
609 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
610 .cl_ops = &fq_codel_class_ops,
612 .priv_size = sizeof(struct fq_codel_sched_data),
613 .enqueue = fq_codel_enqueue,
614 .dequeue = fq_codel_dequeue,
615 .peek = qdisc_peek_dequeued,
616 .drop = fq_codel_qdisc_drop,
617 .init = fq_codel_init,
618 .reset = fq_codel_reset,
619 .destroy = fq_codel_destroy,
620 .change = fq_codel_change,
621 .dump = fq_codel_dump,
622 .dump_stats = fq_codel_dump_stats,
623 .owner = THIS_MODULE,
626 static int __init fq_codel_module_init(void)
628 return register_qdisc(&fq_codel_qdisc_ops);
631 static void __exit fq_codel_module_exit(void)
633 unregister_qdisc(&fq_codel_qdisc_ops);
636 module_init(fq_codel_module_init)
637 module_exit(fq_codel_module_exit)
638 MODULE_AUTHOR("Eric Dumazet");
639 MODULE_LICENSE("GPL");