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/jhash.h>
27 #include <net/pkt_sched.h>
28 #include <net/inet_ecn.h>
31 * SFB uses two B[l][n] : L x N arrays of bins (L levels, N bins per level)
32 * This implementation uses L = 8 and N = 16
33 * This permits us to split one 32bit hash (provided per packet by rxhash or
34 * external classifier) into 8 subhashes of 4 bits.
36 #define SFB_BUCKET_SHIFT 4
37 #define SFB_NUMBUCKETS (1 << SFB_BUCKET_SHIFT) /* N bins per Level */
38 #define SFB_BUCKET_MASK (SFB_NUMBUCKETS - 1)
39 #define SFB_LEVELS (32 / SFB_BUCKET_SHIFT) /* L */
41 /* SFB algo uses a virtual queue, named "bin" */
43 u16 qlen; /* length of virtual queue */
44 u16 p_mark; /* marking probability */
47 /* We use a double buffering right before hash change
48 * (Section 4.4 of SFB reference : moving hash functions)
51 u32 perturbation; /* jhash perturbation */
52 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
55 struct sfb_sched_data {
57 struct tcf_proto __rcu *filter_list;
58 unsigned long rehash_interval;
59 unsigned long warmup_time; /* double buffering warmup time in jiffies */
61 u32 bin_size; /* maximum queue length per bin */
62 u32 increment; /* d1 */
63 u32 decrement; /* d2 */
64 u32 limit; /* HARD maximal queue length */
68 unsigned long rehash_time;
69 unsigned long token_time;
71 u8 slot; /* current active bins (0 or 1) */
72 bool double_buffering;
73 struct sfb_bins bins[2];
80 u32 childdrop; /* drops in child qdisc */
81 u32 marked; /* ECN mark */
86 * Each queued skb might be hashed on one or two bins
87 * We store in skb_cb the two hash values.
88 * (A zero value means double buffering was not used)
94 static inline struct sfb_skb_cb *sfb_skb_cb(const struct sk_buff *skb)
96 qdisc_cb_private_validate(skb, sizeof(struct sfb_skb_cb));
97 return (struct sfb_skb_cb *)qdisc_skb_cb(skb)->data;
101 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
102 * If using external classifier, hash comes from the classid.
104 static u32 sfb_hash(const struct sk_buff *skb, u32 slot)
106 return sfb_skb_cb(skb)->hashes[slot];
109 /* Probabilities are coded as Q0.16 fixed-point values,
110 * with 0xFFFF representing 65535/65536 (almost 1.0)
111 * Addition and subtraction are saturating in [0, 65535]
113 static u32 prob_plus(u32 p1, u32 p2)
117 return min_t(u32, res, SFB_MAX_PROB);
120 static u32 prob_minus(u32 p1, u32 p2)
122 return p1 > p2 ? p1 - p2 : 0;
125 static void increment_one_qlen(u32 sfbhash, u32 slot, struct sfb_sched_data *q)
128 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
130 for (i = 0; i < SFB_LEVELS; i++) {
131 u32 hash = sfbhash & SFB_BUCKET_MASK;
133 sfbhash >>= SFB_BUCKET_SHIFT;
134 if (b[hash].qlen < 0xFFFF)
136 b += SFB_NUMBUCKETS; /* next level */
140 static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
144 sfbhash = sfb_hash(skb, 0);
146 increment_one_qlen(sfbhash, 0, q);
148 sfbhash = sfb_hash(skb, 1);
150 increment_one_qlen(sfbhash, 1, q);
153 static void decrement_one_qlen(u32 sfbhash, u32 slot,
154 struct sfb_sched_data *q)
157 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
159 for (i = 0; i < SFB_LEVELS; i++) {
160 u32 hash = sfbhash & SFB_BUCKET_MASK;
162 sfbhash >>= SFB_BUCKET_SHIFT;
163 if (b[hash].qlen > 0)
165 b += SFB_NUMBUCKETS; /* next level */
169 static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
173 sfbhash = sfb_hash(skb, 0);
175 decrement_one_qlen(sfbhash, 0, q);
177 sfbhash = sfb_hash(skb, 1);
179 decrement_one_qlen(sfbhash, 1, q);
182 static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
184 b->p_mark = prob_minus(b->p_mark, q->decrement);
187 static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
189 b->p_mark = prob_plus(b->p_mark, q->increment);
192 static void sfb_zero_all_buckets(struct sfb_sched_data *q)
194 memset(&q->bins, 0, sizeof(q->bins));
198 * compute max qlen, max p_mark, and avg p_mark
200 static u32 sfb_compute_qlen(u32 *prob_r, u32 *avgpm_r, const struct sfb_sched_data *q)
203 u32 qlen = 0, prob = 0, totalpm = 0;
204 const struct sfb_bucket *b = &q->bins[q->slot].bins[0][0];
206 for (i = 0; i < SFB_LEVELS * SFB_NUMBUCKETS; i++) {
209 totalpm += b->p_mark;
210 if (prob < b->p_mark)
215 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
220 static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
222 q->bins[slot].perturbation = prandom_u32();
225 static void sfb_swap_slot(struct sfb_sched_data *q)
227 sfb_init_perturbation(q->slot, q);
229 q->double_buffering = false;
232 /* Non elastic flows are allowed to use part of the bandwidth, expressed
233 * in "penalty_rate" packets per second, with "penalty_burst" burst
235 static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
237 if (q->penalty_rate == 0 || q->penalty_burst == 0)
240 if (q->tokens_avail < 1) {
241 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
243 q->tokens_avail = (age * q->penalty_rate) / HZ;
244 if (q->tokens_avail > q->penalty_burst)
245 q->tokens_avail = q->penalty_burst;
246 q->token_time = jiffies;
247 if (q->tokens_avail < 1)
255 static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
256 int *qerr, u32 *salt)
258 struct tcf_result res;
261 result = tc_classify(skb, fl, &res, false);
263 #ifdef CONFIG_NET_CLS_ACT
267 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
272 *salt = TC_H_MIN(res.classid);
278 static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
279 struct sk_buff **to_free)
282 struct sfb_sched_data *q = qdisc_priv(sch);
283 struct Qdisc *child = q->qdisc;
284 struct tcf_proto *fl;
290 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
292 if (unlikely(sch->q.qlen >= q->limit)) {
293 qdisc_qstats_overlimit(sch);
294 q->stats.queuedrop++;
298 if (q->rehash_interval > 0) {
299 unsigned long limit = q->rehash_time + q->rehash_interval;
301 if (unlikely(time_after(jiffies, limit))) {
303 q->rehash_time = jiffies;
304 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
305 time_after(jiffies, limit - q->warmup_time))) {
306 q->double_buffering = true;
310 fl = rcu_dereference_bh(q->filter_list);
314 /* If using external classifiers, get result and record it. */
315 if (!sfb_classify(skb, fl, &ret, &salt))
317 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
319 sfbhash = skb_get_hash_perturb(skb, q->bins[slot].perturbation);
325 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
327 for (i = 0; i < SFB_LEVELS; i++) {
328 u32 hash = sfbhash & SFB_BUCKET_MASK;
329 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
331 sfbhash >>= SFB_BUCKET_SHIFT;
333 decrement_prob(b, q);
334 else if (b->qlen >= q->bin_size)
335 increment_prob(b, q);
336 if (minqlen > b->qlen)
338 if (p_min > b->p_mark)
343 sfb_skb_cb(skb)->hashes[slot] = 0;
345 if (unlikely(minqlen >= q->max)) {
346 qdisc_qstats_overlimit(sch);
347 q->stats.bucketdrop++;
351 if (unlikely(p_min >= SFB_MAX_PROB)) {
353 if (q->double_buffering) {
354 sfbhash = skb_get_hash_perturb(skb,
355 q->bins[slot].perturbation);
358 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
360 for (i = 0; i < SFB_LEVELS; i++) {
361 u32 hash = sfbhash & SFB_BUCKET_MASK;
362 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
364 sfbhash >>= SFB_BUCKET_SHIFT;
366 decrement_prob(b, q);
367 else if (b->qlen >= q->bin_size)
368 increment_prob(b, q);
371 if (sfb_rate_limit(skb, q)) {
372 qdisc_qstats_overlimit(sch);
373 q->stats.penaltydrop++;
379 r = prandom_u32() & SFB_MAX_PROB;
381 if (unlikely(r < p_min)) {
382 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
383 /* If we're marking that many packets, then either
384 * this flow is unresponsive, or we're badly congested.
385 * In either case, we want to start dropping packets.
387 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
388 q->stats.earlydrop++;
392 if (INET_ECN_set_ce(skb)) {
395 q->stats.earlydrop++;
401 ret = qdisc_enqueue(skb, child, to_free);
402 if (likely(ret == NET_XMIT_SUCCESS)) {
404 increment_qlen(skb, q);
405 } else if (net_xmit_drop_count(ret)) {
406 q->stats.childdrop++;
407 qdisc_qstats_drop(sch);
412 qdisc_drop(skb, sch, to_free);
415 if (ret & __NET_XMIT_BYPASS)
416 qdisc_qstats_drop(sch);
421 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
423 struct sfb_sched_data *q = qdisc_priv(sch);
424 struct Qdisc *child = q->qdisc;
427 skb = child->dequeue(q->qdisc);
430 qdisc_bstats_update(sch, skb);
432 decrement_qlen(skb, q);
438 static struct sk_buff *sfb_peek(struct Qdisc *sch)
440 struct sfb_sched_data *q = qdisc_priv(sch);
441 struct Qdisc *child = q->qdisc;
443 return child->ops->peek(child);
446 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
448 static void sfb_reset(struct Qdisc *sch)
450 struct sfb_sched_data *q = qdisc_priv(sch);
452 qdisc_reset(q->qdisc);
455 q->double_buffering = false;
456 sfb_zero_all_buckets(q);
457 sfb_init_perturbation(0, q);
460 static void sfb_destroy(struct Qdisc *sch)
462 struct sfb_sched_data *q = qdisc_priv(sch);
464 tcf_destroy_chain(&q->filter_list);
465 qdisc_destroy(q->qdisc);
468 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
469 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
472 static const struct tc_sfb_qopt sfb_default_ops = {
473 .rehash_interval = 600 * MSEC_PER_SEC,
474 .warmup_time = 60 * MSEC_PER_SEC,
478 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
479 .decrement = (SFB_MAX_PROB + 3000) / 6000,
484 static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
486 struct sfb_sched_data *q = qdisc_priv(sch);
488 struct nlattr *tb[TCA_SFB_MAX + 1];
489 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
494 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
498 if (tb[TCA_SFB_PARMS] == NULL)
501 ctl = nla_data(tb[TCA_SFB_PARMS]);
506 limit = qdisc_dev(sch)->tx_queue_len;
508 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
510 return PTR_ERR(child);
514 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
515 q->qdisc->qstats.backlog);
516 qdisc_destroy(q->qdisc);
519 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
520 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
521 q->rehash_time = jiffies;
523 q->increment = ctl->increment;
524 q->decrement = ctl->decrement;
526 q->bin_size = ctl->bin_size;
527 q->penalty_rate = ctl->penalty_rate;
528 q->penalty_burst = ctl->penalty_burst;
529 q->tokens_avail = ctl->penalty_burst;
530 q->token_time = jiffies;
533 q->double_buffering = false;
534 sfb_zero_all_buckets(q);
535 sfb_init_perturbation(0, q);
536 sfb_init_perturbation(1, q);
538 sch_tree_unlock(sch);
543 static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
545 struct sfb_sched_data *q = qdisc_priv(sch);
547 q->qdisc = &noop_qdisc;
548 return sfb_change(sch, opt);
551 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
553 struct sfb_sched_data *q = qdisc_priv(sch);
555 struct tc_sfb_qopt opt = {
556 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
557 .warmup_time = jiffies_to_msecs(q->warmup_time),
560 .bin_size = q->bin_size,
561 .increment = q->increment,
562 .decrement = q->decrement,
563 .penalty_rate = q->penalty_rate,
564 .penalty_burst = q->penalty_burst,
567 sch->qstats.backlog = q->qdisc->qstats.backlog;
568 opts = nla_nest_start(skb, TCA_OPTIONS);
570 goto nla_put_failure;
571 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
572 goto nla_put_failure;
573 return nla_nest_end(skb, opts);
576 nla_nest_cancel(skb, opts);
580 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
582 struct sfb_sched_data *q = qdisc_priv(sch);
583 struct tc_sfb_xstats st = {
584 .earlydrop = q->stats.earlydrop,
585 .penaltydrop = q->stats.penaltydrop,
586 .bucketdrop = q->stats.bucketdrop,
587 .queuedrop = q->stats.queuedrop,
588 .childdrop = q->stats.childdrop,
589 .marked = q->stats.marked,
592 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
594 return gnet_stats_copy_app(d, &st, sizeof(st));
597 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
598 struct sk_buff *skb, struct tcmsg *tcm)
603 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
606 struct sfb_sched_data *q = qdisc_priv(sch);
611 *old = qdisc_replace(sch, new, &q->qdisc);
615 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
617 struct sfb_sched_data *q = qdisc_priv(sch);
622 static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
627 static void sfb_put(struct Qdisc *sch, unsigned long arg)
631 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
632 struct nlattr **tca, unsigned long *arg)
637 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
642 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
645 if (walker->count >= walker->skip)
646 if (walker->fn(sch, 1, walker) < 0) {
654 static struct tcf_proto __rcu **sfb_find_tcf(struct Qdisc *sch,
657 struct sfb_sched_data *q = qdisc_priv(sch);
661 return &q->filter_list;
664 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
671 static const struct Qdisc_class_ops sfb_class_ops = {
676 .change = sfb_change_class,
677 .delete = sfb_delete,
679 .tcf_chain = sfb_find_tcf,
680 .bind_tcf = sfb_bind,
681 .unbind_tcf = sfb_put,
682 .dump = sfb_dump_class,
685 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
687 .priv_size = sizeof(struct sfb_sched_data),
688 .cl_ops = &sfb_class_ops,
689 .enqueue = sfb_enqueue,
690 .dequeue = sfb_dequeue,
694 .destroy = sfb_destroy,
695 .change = sfb_change,
697 .dump_stats = sfb_dump_stats,
698 .owner = THIS_MODULE,
701 static int __init sfb_module_init(void)
703 return register_qdisc(&sfb_qdisc_ops);
706 static void __exit sfb_module_exit(void)
708 unregister_qdisc(&sfb_qdisc_ops);
711 module_init(sfb_module_init)
712 module_exit(sfb_module_exit)
714 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
715 MODULE_AUTHOR("Juliusz Chroboczek");
716 MODULE_AUTHOR("Eric Dumazet");
717 MODULE_LICENSE("GPL");