2 * net/sched/sch_sfb.c Stochastic Fair Blue
4 * Copyright (c) 2008-2011 Juliusz Chroboczek <jch@pps.jussieu.fr>
5 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * version 2 as published by the Free Software Foundation.
11 * W. Feng, D. Kandlur, D. Saha, K. Shin. Blue:
12 * A New Class of Active Queue Management Algorithms.
13 * U. Michigan CSE-TR-387-99, April 1999.
15 * http://www.thefengs.com/wuchang/blue/CSE-TR-387-99.pdf
19 #include <linux/module.h>
20 #include <linux/types.h>
21 #include <linux/kernel.h>
22 #include <linux/errno.h>
23 #include <linux/skbuff.h>
24 #include <linux/random.h>
25 #include <linux/siphash.h>
27 #include <net/pkt_sched.h>
28 #include <net/pkt_cls.h>
29 #include <net/inet_ecn.h>
32 * SFB uses two B[l][n] : L x N arrays of bins (L levels, N bins per level)
33 * This implementation uses L = 8 and N = 16
34 * This permits us to split one 32bit hash (provided per packet by rxhash or
35 * external classifier) into 8 subhashes of 4 bits.
37 #define SFB_BUCKET_SHIFT 4
38 #define SFB_NUMBUCKETS (1 << SFB_BUCKET_SHIFT) /* N bins per Level */
39 #define SFB_BUCKET_MASK (SFB_NUMBUCKETS - 1)
40 #define SFB_LEVELS (32 / SFB_BUCKET_SHIFT) /* L */
42 /* SFB algo uses a virtual queue, named "bin" */
44 u16 qlen; /* length of virtual queue */
45 u16 p_mark; /* marking probability */
48 /* We use a double buffering right before hash change
49 * (Section 4.4 of SFB reference : moving hash functions)
52 siphash_key_t perturbation; /* siphash key */
53 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
56 struct sfb_sched_data {
58 struct tcf_proto __rcu *filter_list;
59 struct tcf_block *block;
60 unsigned long rehash_interval;
61 unsigned long warmup_time; /* double buffering warmup time in jiffies */
63 u32 bin_size; /* maximum queue length per bin */
64 u32 increment; /* d1 */
65 u32 decrement; /* d2 */
66 u32 limit; /* HARD maximal queue length */
70 unsigned long rehash_time;
71 unsigned long token_time;
73 u8 slot; /* current active bins (0 or 1) */
74 bool double_buffering;
75 struct sfb_bins bins[2];
82 u32 childdrop; /* drops in child qdisc */
83 u32 marked; /* ECN mark */
88 * Each queued skb might be hashed on one or two bins
89 * We store in skb_cb the two hash values.
90 * (A zero value means double buffering was not used)
96 static inline struct sfb_skb_cb *sfb_skb_cb(const struct sk_buff *skb)
98 qdisc_cb_private_validate(skb, sizeof(struct sfb_skb_cb));
99 return (struct sfb_skb_cb *)qdisc_skb_cb(skb)->data;
103 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
104 * If using external classifier, hash comes from the classid.
106 static u32 sfb_hash(const struct sk_buff *skb, u32 slot)
108 return sfb_skb_cb(skb)->hashes[slot];
111 /* Probabilities are coded as Q0.16 fixed-point values,
112 * with 0xFFFF representing 65535/65536 (almost 1.0)
113 * Addition and subtraction are saturating in [0, 65535]
115 static u32 prob_plus(u32 p1, u32 p2)
119 return min_t(u32, res, SFB_MAX_PROB);
122 static u32 prob_minus(u32 p1, u32 p2)
124 return p1 > p2 ? p1 - p2 : 0;
127 static void increment_one_qlen(u32 sfbhash, u32 slot, struct sfb_sched_data *q)
130 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
132 for (i = 0; i < SFB_LEVELS; i++) {
133 u32 hash = sfbhash & SFB_BUCKET_MASK;
135 sfbhash >>= SFB_BUCKET_SHIFT;
136 if (b[hash].qlen < 0xFFFF)
138 b += SFB_NUMBUCKETS; /* next level */
142 static void increment_qlen(const struct sfb_skb_cb *cb, struct sfb_sched_data *q)
146 sfbhash = cb->hashes[0];
148 increment_one_qlen(sfbhash, 0, q);
150 sfbhash = cb->hashes[1];
152 increment_one_qlen(sfbhash, 1, q);
155 static void decrement_one_qlen(u32 sfbhash, u32 slot,
156 struct sfb_sched_data *q)
159 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
161 for (i = 0; i < SFB_LEVELS; i++) {
162 u32 hash = sfbhash & SFB_BUCKET_MASK;
164 sfbhash >>= SFB_BUCKET_SHIFT;
165 if (b[hash].qlen > 0)
167 b += SFB_NUMBUCKETS; /* next level */
171 static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
175 sfbhash = sfb_hash(skb, 0);
177 decrement_one_qlen(sfbhash, 0, q);
179 sfbhash = sfb_hash(skb, 1);
181 decrement_one_qlen(sfbhash, 1, q);
184 static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
186 b->p_mark = prob_minus(b->p_mark, q->decrement);
189 static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
191 b->p_mark = prob_plus(b->p_mark, q->increment);
194 static void sfb_zero_all_buckets(struct sfb_sched_data *q)
196 memset(&q->bins, 0, sizeof(q->bins));
200 * compute max qlen, max p_mark, and avg p_mark
202 static u32 sfb_compute_qlen(u32 *prob_r, u32 *avgpm_r, const struct sfb_sched_data *q)
205 u32 qlen = 0, prob = 0, totalpm = 0;
206 const struct sfb_bucket *b = &q->bins[q->slot].bins[0][0];
208 for (i = 0; i < SFB_LEVELS * SFB_NUMBUCKETS; i++) {
211 totalpm += b->p_mark;
212 if (prob < b->p_mark)
217 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
222 static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
224 get_random_bytes(&q->bins[slot].perturbation,
225 sizeof(q->bins[slot].perturbation));
228 static void sfb_swap_slot(struct sfb_sched_data *q)
230 sfb_init_perturbation(q->slot, q);
232 q->double_buffering = false;
235 /* Non elastic flows are allowed to use part of the bandwidth, expressed
236 * in "penalty_rate" packets per second, with "penalty_burst" burst
238 static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
240 if (q->penalty_rate == 0 || q->penalty_burst == 0)
243 if (q->tokens_avail < 1) {
244 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
246 q->tokens_avail = (age * q->penalty_rate) / HZ;
247 if (q->tokens_avail > q->penalty_burst)
248 q->tokens_avail = q->penalty_burst;
249 q->token_time = jiffies;
250 if (q->tokens_avail < 1)
258 static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
259 int *qerr, u32 *salt)
261 struct tcf_result res;
264 result = tcf_classify(skb, fl, &res, false);
266 #ifdef CONFIG_NET_CLS_ACT
271 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
277 *salt = TC_H_MIN(res.classid);
283 static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
284 struct sk_buff **to_free)
287 struct sfb_sched_data *q = qdisc_priv(sch);
288 struct Qdisc *child = q->qdisc;
289 struct tcf_proto *fl;
290 struct sfb_skb_cb cb;
296 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
298 if (unlikely(sch->q.qlen >= q->limit)) {
299 qdisc_qstats_overlimit(sch);
300 q->stats.queuedrop++;
304 if (q->rehash_interval > 0) {
305 unsigned long limit = q->rehash_time + q->rehash_interval;
307 if (unlikely(time_after(jiffies, limit))) {
309 q->rehash_time = jiffies;
310 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
311 time_after(jiffies, limit - q->warmup_time))) {
312 q->double_buffering = true;
316 fl = rcu_dereference_bh(q->filter_list);
320 /* If using external classifiers, get result and record it. */
321 if (!sfb_classify(skb, fl, &ret, &salt))
323 sfbhash = siphash_1u32(salt, &q->bins[slot].perturbation);
325 sfbhash = skb_get_hash_perturb(skb, &q->bins[slot].perturbation);
331 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
333 for (i = 0; i < SFB_LEVELS; i++) {
334 u32 hash = sfbhash & SFB_BUCKET_MASK;
335 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
337 sfbhash >>= SFB_BUCKET_SHIFT;
339 decrement_prob(b, q);
340 else if (b->qlen >= q->bin_size)
341 increment_prob(b, q);
342 if (minqlen > b->qlen)
344 if (p_min > b->p_mark)
349 sfb_skb_cb(skb)->hashes[slot] = 0;
351 if (unlikely(minqlen >= q->max)) {
352 qdisc_qstats_overlimit(sch);
353 q->stats.bucketdrop++;
357 if (unlikely(p_min >= SFB_MAX_PROB)) {
359 if (q->double_buffering) {
360 sfbhash = skb_get_hash_perturb(skb,
361 &q->bins[slot].perturbation);
364 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
366 for (i = 0; i < SFB_LEVELS; i++) {
367 u32 hash = sfbhash & SFB_BUCKET_MASK;
368 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
370 sfbhash >>= SFB_BUCKET_SHIFT;
372 decrement_prob(b, q);
373 else if (b->qlen >= q->bin_size)
374 increment_prob(b, q);
377 if (sfb_rate_limit(skb, q)) {
378 qdisc_qstats_overlimit(sch);
379 q->stats.penaltydrop++;
385 r = prandom_u32() & SFB_MAX_PROB;
387 if (unlikely(r < p_min)) {
388 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
389 /* If we're marking that many packets, then either
390 * this flow is unresponsive, or we're badly congested.
391 * In either case, we want to start dropping packets.
393 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
394 q->stats.earlydrop++;
398 if (INET_ECN_set_ce(skb)) {
401 q->stats.earlydrop++;
407 memcpy(&cb, sfb_skb_cb(skb), sizeof(cb));
408 ret = qdisc_enqueue(skb, child, to_free);
409 if (likely(ret == NET_XMIT_SUCCESS)) {
410 qdisc_qstats_backlog_inc(sch, skb);
412 increment_qlen(&cb, q);
413 } else if (net_xmit_drop_count(ret)) {
414 q->stats.childdrop++;
415 qdisc_qstats_drop(sch);
420 qdisc_drop(skb, sch, to_free);
423 if (ret & __NET_XMIT_BYPASS)
424 qdisc_qstats_drop(sch);
429 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
431 struct sfb_sched_data *q = qdisc_priv(sch);
432 struct Qdisc *child = q->qdisc;
435 skb = child->dequeue(q->qdisc);
438 qdisc_bstats_update(sch, skb);
439 qdisc_qstats_backlog_dec(sch, skb);
441 decrement_qlen(skb, q);
447 static struct sk_buff *sfb_peek(struct Qdisc *sch)
449 struct sfb_sched_data *q = qdisc_priv(sch);
450 struct Qdisc *child = q->qdisc;
452 return child->ops->peek(child);
455 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
457 static void sfb_reset(struct Qdisc *sch)
459 struct sfb_sched_data *q = qdisc_priv(sch);
461 qdisc_reset(q->qdisc);
462 sch->qstats.backlog = 0;
465 q->double_buffering = false;
466 sfb_zero_all_buckets(q);
467 sfb_init_perturbation(0, q);
470 static void sfb_destroy(struct Qdisc *sch)
472 struct sfb_sched_data *q = qdisc_priv(sch);
474 tcf_block_put(q->block);
478 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
479 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
482 static const struct tc_sfb_qopt sfb_default_ops = {
483 .rehash_interval = 600 * MSEC_PER_SEC,
484 .warmup_time = 60 * MSEC_PER_SEC,
488 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
489 .decrement = (SFB_MAX_PROB + 3000) / 6000,
494 static int sfb_change(struct Qdisc *sch, struct nlattr *opt,
495 struct netlink_ext_ack *extack)
497 struct sfb_sched_data *q = qdisc_priv(sch);
499 struct nlattr *tb[TCA_SFB_MAX + 1];
500 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
505 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy, NULL);
509 if (tb[TCA_SFB_PARMS] == NULL)
512 ctl = nla_data(tb[TCA_SFB_PARMS]);
517 limit = qdisc_dev(sch)->tx_queue_len;
519 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit, extack);
521 return PTR_ERR(child);
523 if (child != &noop_qdisc)
524 qdisc_hash_add(child, true);
527 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
528 q->qdisc->qstats.backlog);
532 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
533 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
534 q->rehash_time = jiffies;
536 q->increment = ctl->increment;
537 q->decrement = ctl->decrement;
539 q->bin_size = ctl->bin_size;
540 q->penalty_rate = ctl->penalty_rate;
541 q->penalty_burst = ctl->penalty_burst;
542 q->tokens_avail = ctl->penalty_burst;
543 q->token_time = jiffies;
546 q->double_buffering = false;
547 sfb_zero_all_buckets(q);
548 sfb_init_perturbation(0, q);
549 sfb_init_perturbation(1, q);
551 sch_tree_unlock(sch);
556 static int sfb_init(struct Qdisc *sch, struct nlattr *opt,
557 struct netlink_ext_ack *extack)
559 struct sfb_sched_data *q = qdisc_priv(sch);
562 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
566 q->qdisc = &noop_qdisc;
567 return sfb_change(sch, opt, extack);
570 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
572 struct sfb_sched_data *q = qdisc_priv(sch);
574 struct tc_sfb_qopt opt = {
575 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
576 .warmup_time = jiffies_to_msecs(q->warmup_time),
579 .bin_size = q->bin_size,
580 .increment = q->increment,
581 .decrement = q->decrement,
582 .penalty_rate = q->penalty_rate,
583 .penalty_burst = q->penalty_burst,
586 sch->qstats.backlog = q->qdisc->qstats.backlog;
587 opts = nla_nest_start(skb, TCA_OPTIONS);
589 goto nla_put_failure;
590 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
591 goto nla_put_failure;
592 return nla_nest_end(skb, opts);
595 nla_nest_cancel(skb, opts);
599 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
601 struct sfb_sched_data *q = qdisc_priv(sch);
602 struct tc_sfb_xstats st = {
603 .earlydrop = q->stats.earlydrop,
604 .penaltydrop = q->stats.penaltydrop,
605 .bucketdrop = q->stats.bucketdrop,
606 .queuedrop = q->stats.queuedrop,
607 .childdrop = q->stats.childdrop,
608 .marked = q->stats.marked,
611 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
613 return gnet_stats_copy_app(d, &st, sizeof(st));
616 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
617 struct sk_buff *skb, struct tcmsg *tcm)
622 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
623 struct Qdisc **old, struct netlink_ext_ack *extack)
625 struct sfb_sched_data *q = qdisc_priv(sch);
630 *old = qdisc_replace(sch, new, &q->qdisc);
634 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
636 struct sfb_sched_data *q = qdisc_priv(sch);
641 static unsigned long sfb_find(struct Qdisc *sch, u32 classid)
646 static void sfb_unbind(struct Qdisc *sch, unsigned long arg)
650 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
651 struct nlattr **tca, unsigned long *arg,
652 struct netlink_ext_ack *extack)
657 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
662 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
665 if (walker->count >= walker->skip)
666 if (walker->fn(sch, 1, walker) < 0) {
674 static struct tcf_block *sfb_tcf_block(struct Qdisc *sch, unsigned long cl,
675 struct netlink_ext_ack *extack)
677 struct sfb_sched_data *q = qdisc_priv(sch);
684 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
691 static const struct Qdisc_class_ops sfb_class_ops = {
695 .change = sfb_change_class,
696 .delete = sfb_delete,
698 .tcf_block = sfb_tcf_block,
699 .bind_tcf = sfb_bind,
700 .unbind_tcf = sfb_unbind,
701 .dump = sfb_dump_class,
704 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
706 .priv_size = sizeof(struct sfb_sched_data),
707 .cl_ops = &sfb_class_ops,
708 .enqueue = sfb_enqueue,
709 .dequeue = sfb_dequeue,
713 .destroy = sfb_destroy,
714 .change = sfb_change,
716 .dump_stats = sfb_dump_stats,
717 .owner = THIS_MODULE,
720 static int __init sfb_module_init(void)
722 return register_qdisc(&sfb_qdisc_ops);
725 static void __exit sfb_module_exit(void)
727 unregister_qdisc(&sfb_qdisc_ops);
730 module_init(sfb_module_init)
731 module_exit(sfb_module_exit)
733 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
734 MODULE_AUTHOR("Juliusz Chroboczek");
735 MODULE_AUTHOR("Eric Dumazet");
736 MODULE_LICENSE("GPL");