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>
27 #include <net/codel_impl.h>
28 #include <net/codel_qdisc.h>
33 * Packets are classified (internal classifier or external) on flows.
34 * This is a Stochastic model (as we use a hash, several flows
35 * might be hashed on same slot)
36 * Each flow has a CoDel managed queue.
37 * Flows are linked onto two (Round Robin) lists,
38 * so that new flows have priority on old ones.
40 * For a given flow, packets are not reordered (CoDel uses a FIFO)
42 * ECN capability is on by default.
43 * Low memory footprint (64 bytes per flow)
46 struct fq_codel_flow {
49 struct list_head flowchain;
51 u32 dropped; /* number of drops (or ECN marks) on this flow */
52 struct codel_vars cvars;
53 }; /* please try to keep this structure <= 64 bytes */
55 struct fq_codel_sched_data {
56 struct tcf_proto __rcu *filter_list; /* optional external classifier */
57 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
58 u32 *backlogs; /* backlog table [flows_cnt] */
59 u32 flows_cnt; /* number of flows */
60 u32 perturbation; /* hash perturbation */
61 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
64 struct codel_params cparams;
65 struct codel_stats cstats;
71 struct list_head new_flows; /* list of new flows */
72 struct list_head old_flows; /* list of old flows */
75 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
78 u32 hash = skb_get_hash_perturb(skb, q->perturbation);
80 return reciprocal_scale(hash, q->flows_cnt);
83 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
86 struct fq_codel_sched_data *q = qdisc_priv(sch);
87 struct tcf_proto *filter;
88 struct tcf_result res;
91 if (TC_H_MAJ(skb->priority) == sch->handle &&
92 TC_H_MIN(skb->priority) > 0 &&
93 TC_H_MIN(skb->priority) <= q->flows_cnt)
94 return TC_H_MIN(skb->priority);
96 filter = rcu_dereference_bh(q->filter_list);
98 return fq_codel_hash(q, skb) + 1;
100 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
101 result = tc_classify(skb, filter, &res, false);
103 #ifdef CONFIG_NET_CLS_ACT
107 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
112 if (TC_H_MIN(res.classid) <= q->flows_cnt)
113 return TC_H_MIN(res.classid);
118 /* helper functions : might be changed when/if skb use a standard list_head */
120 /* remove one skb from head of slot queue */
121 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
123 struct sk_buff *skb = flow->head;
125 flow->head = skb->next;
130 /* add skb to flow queue (tail add) */
131 static inline void flow_queue_add(struct fq_codel_flow *flow,
134 if (flow->head == NULL)
137 flow->tail->next = skb;
142 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets)
144 struct fq_codel_sched_data *q = qdisc_priv(sch);
146 unsigned int maxbacklog = 0, idx = 0, i, len;
147 struct fq_codel_flow *flow;
148 unsigned int threshold;
149 unsigned int mem = 0;
151 /* Queue is full! Find the fat flow and drop packet(s) from it.
152 * This might sound expensive, but with 1024 flows, we scan
153 * 4KB of memory, and we dont need to handle a complex tree
154 * in fast path (packet queue/enqueue) with many cache misses.
155 * In stress mode, we'll try to drop 64 packets from the flow,
156 * amortizing this linear lookup to one cache line per drop.
158 for (i = 0; i < q->flows_cnt; i++) {
159 if (q->backlogs[i] > maxbacklog) {
160 maxbacklog = q->backlogs[i];
165 /* Our goal is to drop half of this fat flow backlog */
166 threshold = maxbacklog >> 1;
168 flow = &q->flows[idx];
172 skb = dequeue_head(flow);
173 len += qdisc_pkt_len(skb);
174 mem += skb->truesize;
176 } while (++i < max_packets && len < threshold);
179 q->backlogs[idx] -= len;
180 q->memory_usage -= mem;
181 sch->qstats.drops += i;
182 sch->qstats.backlog -= len;
187 static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch)
189 unsigned int prev_backlog;
191 prev_backlog = sch->qstats.backlog;
192 fq_codel_drop(sch, 1U);
193 return prev_backlog - sch->qstats.backlog;
196 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
198 struct fq_codel_sched_data *q = qdisc_priv(sch);
199 unsigned int idx, prev_backlog, prev_qlen;
200 struct fq_codel_flow *flow;
201 int uninitialized_var(ret);
204 idx = fq_codel_classify(skb, sch, &ret);
206 if (ret & __NET_XMIT_BYPASS)
207 qdisc_qstats_drop(sch);
213 codel_set_enqueue_time(skb);
214 flow = &q->flows[idx];
215 flow_queue_add(flow, skb);
216 q->backlogs[idx] += qdisc_pkt_len(skb);
217 qdisc_qstats_backlog_inc(sch, skb);
219 if (list_empty(&flow->flowchain)) {
220 list_add_tail(&flow->flowchain, &q->new_flows);
222 flow->deficit = q->quantum;
225 q->memory_usage += skb->truesize;
226 memory_limited = q->memory_usage > q->memory_limit;
227 if (++sch->q.qlen <= sch->limit && !memory_limited)
228 return NET_XMIT_SUCCESS;
230 prev_backlog = sch->qstats.backlog;
231 prev_qlen = sch->q.qlen;
233 /* fq_codel_drop() is quite expensive, as it performs a linear search
234 * in q->backlogs[] to find a fat flow.
235 * So instead of dropping a single packet, drop half of its backlog
236 * with a 64 packets limit to not add a too big cpu spike here.
238 ret = fq_codel_drop(sch, q->drop_batch_size);
240 q->drop_overlimit += prev_qlen - sch->q.qlen;
242 q->drop_overmemory += prev_qlen - sch->q.qlen;
243 /* As we dropped packet(s), better let upper stack know this */
244 qdisc_tree_reduce_backlog(sch, prev_qlen - sch->q.qlen,
245 prev_backlog - sch->qstats.backlog);
247 return ret == idx ? NET_XMIT_CN : NET_XMIT_SUCCESS;
250 /* This is the specific function called from codel_dequeue()
251 * to dequeue a packet from queue. Note: backlog is handled in
252 * codel, we dont need to reduce it here.
254 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
256 struct Qdisc *sch = ctx;
257 struct fq_codel_sched_data *q = qdisc_priv(sch);
258 struct fq_codel_flow *flow;
259 struct sk_buff *skb = NULL;
261 flow = container_of(vars, struct fq_codel_flow, cvars);
263 skb = dequeue_head(flow);
264 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
265 q->memory_usage -= skb->truesize;
267 sch->qstats.backlog -= qdisc_pkt_len(skb);
272 static void drop_func(struct sk_buff *skb, void *ctx)
274 struct Qdisc *sch = ctx;
276 qdisc_drop(skb, sch);
279 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
281 struct fq_codel_sched_data *q = qdisc_priv(sch);
283 struct fq_codel_flow *flow;
284 struct list_head *head;
285 u32 prev_drop_count, prev_ecn_mark;
286 unsigned int prev_backlog;
289 head = &q->new_flows;
290 if (list_empty(head)) {
291 head = &q->old_flows;
292 if (list_empty(head))
295 flow = list_first_entry(head, struct fq_codel_flow, flowchain);
297 if (flow->deficit <= 0) {
298 flow->deficit += q->quantum;
299 list_move_tail(&flow->flowchain, &q->old_flows);
303 prev_drop_count = q->cstats.drop_count;
304 prev_ecn_mark = q->cstats.ecn_mark;
305 prev_backlog = sch->qstats.backlog;
307 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
308 &flow->cvars, &q->cstats, qdisc_pkt_len,
309 codel_get_enqueue_time, drop_func, dequeue_func);
311 flow->dropped += q->cstats.drop_count - prev_drop_count;
312 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
315 /* force a pass through old_flows to prevent starvation */
316 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
317 list_move_tail(&flow->flowchain, &q->old_flows);
319 list_del_init(&flow->flowchain);
322 qdisc_bstats_update(sch, skb);
323 flow->deficit -= qdisc_pkt_len(skb);
324 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
325 * or HTB crashes. Defer it for next round.
327 if (q->cstats.drop_count && sch->q.qlen) {
328 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
330 q->cstats.drop_count = 0;
331 q->cstats.drop_len = 0;
336 static void fq_codel_reset(struct Qdisc *sch)
338 struct fq_codel_sched_data *q = qdisc_priv(sch);
341 INIT_LIST_HEAD(&q->new_flows);
342 INIT_LIST_HEAD(&q->old_flows);
343 for (i = 0; i < q->flows_cnt; i++) {
344 struct fq_codel_flow *flow = q->flows + i;
347 struct sk_buff *skb = dequeue_head(flow);
349 qdisc_qstats_backlog_dec(sch, skb);
353 INIT_LIST_HEAD(&flow->flowchain);
354 codel_vars_init(&flow->cvars);
356 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
361 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
362 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
363 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
364 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
365 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
366 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
367 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
368 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
369 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
370 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
373 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
375 struct fq_codel_sched_data *q = qdisc_priv(sch);
376 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
382 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
385 if (tb[TCA_FQ_CODEL_FLOWS]) {
388 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
390 q->flows_cnt > 65536)
395 if (tb[TCA_FQ_CODEL_TARGET]) {
396 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
398 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
401 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
402 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
404 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
407 if (tb[TCA_FQ_CODEL_INTERVAL]) {
408 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
410 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
413 if (tb[TCA_FQ_CODEL_LIMIT])
414 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
416 if (tb[TCA_FQ_CODEL_ECN])
417 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
419 if (tb[TCA_FQ_CODEL_QUANTUM])
420 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
422 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
423 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
425 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
426 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
428 while (sch->q.qlen > sch->limit ||
429 q->memory_usage > q->memory_limit) {
430 struct sk_buff *skb = fq_codel_dequeue(sch);
432 q->cstats.drop_len += qdisc_pkt_len(skb);
434 q->cstats.drop_count++;
436 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
437 q->cstats.drop_count = 0;
438 q->cstats.drop_len = 0;
440 sch_tree_unlock(sch);
444 static void *fq_codel_zalloc(size_t sz)
446 void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
453 static void fq_codel_free(void *addr)
458 static void fq_codel_destroy(struct Qdisc *sch)
460 struct fq_codel_sched_data *q = qdisc_priv(sch);
462 tcf_destroy_chain(&q->filter_list);
463 fq_codel_free(q->backlogs);
464 fq_codel_free(q->flows);
467 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
469 struct fq_codel_sched_data *q = qdisc_priv(sch);
472 sch->limit = 10*1024;
474 q->memory_limit = 32 << 20; /* 32 MBytes */
475 q->drop_batch_size = 64;
476 q->quantum = psched_mtu(qdisc_dev(sch));
477 q->perturbation = prandom_u32();
478 INIT_LIST_HEAD(&q->new_flows);
479 INIT_LIST_HEAD(&q->old_flows);
480 codel_params_init(&q->cparams);
481 codel_stats_init(&q->cstats);
482 q->cparams.ecn = true;
483 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
486 int err = fq_codel_change(sch, opt);
492 q->flows = fq_codel_zalloc(q->flows_cnt *
493 sizeof(struct fq_codel_flow));
496 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
498 fq_codel_free(q->flows);
501 for (i = 0; i < q->flows_cnt; i++) {
502 struct fq_codel_flow *flow = q->flows + i;
504 INIT_LIST_HEAD(&flow->flowchain);
505 codel_vars_init(&flow->cvars);
509 sch->flags |= TCQ_F_CAN_BYPASS;
511 sch->flags &= ~TCQ_F_CAN_BYPASS;
515 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
517 struct fq_codel_sched_data *q = qdisc_priv(sch);
520 opts = nla_nest_start(skb, TCA_OPTIONS);
522 goto nla_put_failure;
524 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
525 codel_time_to_us(q->cparams.target)) ||
526 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
528 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
529 codel_time_to_us(q->cparams.interval)) ||
530 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
532 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
534 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
535 q->drop_batch_size) ||
536 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
538 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
540 goto nla_put_failure;
542 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
543 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
544 codel_time_to_us(q->cparams.ce_threshold)))
545 goto nla_put_failure;
547 return nla_nest_end(skb, opts);
553 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
555 struct fq_codel_sched_data *q = qdisc_priv(sch);
556 struct tc_fq_codel_xstats st = {
557 .type = TCA_FQ_CODEL_XSTATS_QDISC,
559 struct list_head *pos;
561 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
562 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
563 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
564 st.qdisc_stats.new_flow_count = q->new_flow_count;
565 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
566 st.qdisc_stats.memory_usage = q->memory_usage;
567 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
569 list_for_each(pos, &q->new_flows)
570 st.qdisc_stats.new_flows_len++;
572 list_for_each(pos, &q->old_flows)
573 st.qdisc_stats.old_flows_len++;
575 return gnet_stats_copy_app(d, &st, sizeof(st));
578 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
583 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
588 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
591 /* we cannot bypass queue discipline anymore */
592 sch->flags &= ~TCQ_F_CAN_BYPASS;
596 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
600 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
603 struct fq_codel_sched_data *q = qdisc_priv(sch);
607 return &q->filter_list;
610 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
611 struct sk_buff *skb, struct tcmsg *tcm)
613 tcm->tcm_handle |= TC_H_MIN(cl);
617 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
620 struct fq_codel_sched_data *q = qdisc_priv(sch);
622 struct gnet_stats_queue qs = { 0 };
623 struct tc_fq_codel_xstats xstats;
625 if (idx < q->flows_cnt) {
626 const struct fq_codel_flow *flow = &q->flows[idx];
627 const struct sk_buff *skb = flow->head;
629 memset(&xstats, 0, sizeof(xstats));
630 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
631 xstats.class_stats.deficit = flow->deficit;
632 xstats.class_stats.ldelay =
633 codel_time_to_us(flow->cvars.ldelay);
634 xstats.class_stats.count = flow->cvars.count;
635 xstats.class_stats.lastcount = flow->cvars.lastcount;
636 xstats.class_stats.dropping = flow->cvars.dropping;
637 if (flow->cvars.dropping) {
638 codel_tdiff_t delta = flow->cvars.drop_next -
641 xstats.class_stats.drop_next = (delta >= 0) ?
642 codel_time_to_us(delta) :
643 -codel_time_to_us(-delta);
649 qs.backlog = q->backlogs[idx];
650 qs.drops = flow->dropped;
652 if (gnet_stats_copy_queue(d, NULL, &qs, 0) < 0)
654 if (idx < q->flows_cnt)
655 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
659 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
661 struct fq_codel_sched_data *q = qdisc_priv(sch);
667 for (i = 0; i < q->flows_cnt; i++) {
668 if (list_empty(&q->flows[i].flowchain) ||
669 arg->count < arg->skip) {
673 if (arg->fn(sch, i + 1, arg) < 0) {
681 static const struct Qdisc_class_ops fq_codel_class_ops = {
682 .leaf = fq_codel_leaf,
685 .tcf_chain = fq_codel_find_tcf,
686 .bind_tcf = fq_codel_bind,
687 .unbind_tcf = fq_codel_put,
688 .dump = fq_codel_dump_class,
689 .dump_stats = fq_codel_dump_class_stats,
690 .walk = fq_codel_walk,
693 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
694 .cl_ops = &fq_codel_class_ops,
696 .priv_size = sizeof(struct fq_codel_sched_data),
697 .enqueue = fq_codel_enqueue,
698 .dequeue = fq_codel_dequeue,
699 .peek = qdisc_peek_dequeued,
700 .drop = fq_codel_qdisc_drop,
701 .init = fq_codel_init,
702 .reset = fq_codel_reset,
703 .destroy = fq_codel_destroy,
704 .change = fq_codel_change,
705 .dump = fq_codel_dump,
706 .dump_stats = fq_codel_dump_stats,
707 .owner = THIS_MODULE,
710 static int __init fq_codel_module_init(void)
712 return register_qdisc(&fq_codel_qdisc_ops);
715 static void __exit fq_codel_module_exit(void)
717 unregister_qdisc(&fq_codel_qdisc_ops);
720 module_init(fq_codel_module_init)
721 module_exit(fq_codel_module_exit)
722 MODULE_AUTHOR("Eric Dumazet");
723 MODULE_LICENSE("GPL");