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 *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 = net_random();
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 sfb_sched_data *q,
256 int *qerr, u32 *salt)
258 struct tcf_result res;
261 result = tc_classify(skb, q->filter_list, &res);
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)
281 struct sfb_sched_data *q = qdisc_priv(sch);
282 struct Qdisc *child = q->qdisc;
286 u32 r, slot, salt, sfbhash;
287 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
289 if (unlikely(sch->q.qlen >= q->limit)) {
290 sch->qstats.overlimits++;
291 q->stats.queuedrop++;
295 if (q->rehash_interval > 0) {
296 unsigned long limit = q->rehash_time + q->rehash_interval;
298 if (unlikely(time_after(jiffies, limit))) {
300 q->rehash_time = jiffies;
301 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
302 time_after(jiffies, limit - q->warmup_time))) {
303 q->double_buffering = true;
307 if (q->filter_list) {
308 /* If using external classifiers, get result and record it. */
309 if (!sfb_classify(skb, q, &ret, &salt))
312 salt = skb_get_rxhash(skb);
317 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
320 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
322 for (i = 0; i < SFB_LEVELS; i++) {
323 u32 hash = sfbhash & SFB_BUCKET_MASK;
324 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
326 sfbhash >>= SFB_BUCKET_SHIFT;
328 decrement_prob(b, q);
329 else if (b->qlen >= q->bin_size)
330 increment_prob(b, q);
331 if (minqlen > b->qlen)
333 if (p_min > b->p_mark)
338 sfb_skb_cb(skb)->hashes[slot] = 0;
340 if (unlikely(minqlen >= q->max)) {
341 sch->qstats.overlimits++;
342 q->stats.bucketdrop++;
346 if (unlikely(p_min >= SFB_MAX_PROB)) {
348 if (q->double_buffering) {
349 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
352 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
354 for (i = 0; i < SFB_LEVELS; i++) {
355 u32 hash = sfbhash & SFB_BUCKET_MASK;
356 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
358 sfbhash >>= SFB_BUCKET_SHIFT;
360 decrement_prob(b, q);
361 else if (b->qlen >= q->bin_size)
362 increment_prob(b, q);
365 if (sfb_rate_limit(skb, q)) {
366 sch->qstats.overlimits++;
367 q->stats.penaltydrop++;
373 r = net_random() & SFB_MAX_PROB;
375 if (unlikely(r < p_min)) {
376 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
377 /* If we're marking that many packets, then either
378 * this flow is unresponsive, or we're badly congested.
379 * In either case, we want to start dropping packets.
381 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
382 q->stats.earlydrop++;
386 if (INET_ECN_set_ce(skb)) {
389 q->stats.earlydrop++;
395 ret = qdisc_enqueue(skb, child);
396 if (likely(ret == NET_XMIT_SUCCESS)) {
398 increment_qlen(skb, q);
399 } else if (net_xmit_drop_count(ret)) {
400 q->stats.childdrop++;
406 qdisc_drop(skb, sch);
409 if (ret & __NET_XMIT_BYPASS)
415 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
417 struct sfb_sched_data *q = qdisc_priv(sch);
418 struct Qdisc *child = q->qdisc;
421 skb = child->dequeue(q->qdisc);
424 qdisc_bstats_update(sch, skb);
426 decrement_qlen(skb, q);
432 static struct sk_buff *sfb_peek(struct Qdisc *sch)
434 struct sfb_sched_data *q = qdisc_priv(sch);
435 struct Qdisc *child = q->qdisc;
437 return child->ops->peek(child);
440 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
442 static void sfb_reset(struct Qdisc *sch)
444 struct sfb_sched_data *q = qdisc_priv(sch);
446 qdisc_reset(q->qdisc);
449 q->double_buffering = false;
450 sfb_zero_all_buckets(q);
451 sfb_init_perturbation(0, q);
454 static void sfb_destroy(struct Qdisc *sch)
456 struct sfb_sched_data *q = qdisc_priv(sch);
458 tcf_destroy_chain(&q->filter_list);
459 qdisc_destroy(q->qdisc);
462 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
463 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
466 static const struct tc_sfb_qopt sfb_default_ops = {
467 .rehash_interval = 600 * MSEC_PER_SEC,
468 .warmup_time = 60 * MSEC_PER_SEC,
472 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
473 .decrement = (SFB_MAX_PROB + 3000) / 6000,
478 static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
480 struct sfb_sched_data *q = qdisc_priv(sch);
482 struct nlattr *tb[TCA_SFB_MAX + 1];
483 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
488 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
492 if (tb[TCA_SFB_PARMS] == NULL)
495 ctl = nla_data(tb[TCA_SFB_PARMS]);
500 limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
502 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
504 return PTR_ERR(child);
508 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
509 qdisc_destroy(q->qdisc);
512 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
513 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
514 q->rehash_time = jiffies;
516 q->increment = ctl->increment;
517 q->decrement = ctl->decrement;
519 q->bin_size = ctl->bin_size;
520 q->penalty_rate = ctl->penalty_rate;
521 q->penalty_burst = ctl->penalty_burst;
522 q->tokens_avail = ctl->penalty_burst;
523 q->token_time = jiffies;
526 q->double_buffering = false;
527 sfb_zero_all_buckets(q);
528 sfb_init_perturbation(0, q);
529 sfb_init_perturbation(1, q);
531 sch_tree_unlock(sch);
536 static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
538 struct sfb_sched_data *q = qdisc_priv(sch);
540 q->qdisc = &noop_qdisc;
541 return sfb_change(sch, opt);
544 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
546 struct sfb_sched_data *q = qdisc_priv(sch);
548 struct tc_sfb_qopt opt = {
549 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
550 .warmup_time = jiffies_to_msecs(q->warmup_time),
553 .bin_size = q->bin_size,
554 .increment = q->increment,
555 .decrement = q->decrement,
556 .penalty_rate = q->penalty_rate,
557 .penalty_burst = q->penalty_burst,
560 sch->qstats.backlog = q->qdisc->qstats.backlog;
561 opts = nla_nest_start(skb, TCA_OPTIONS);
563 goto nla_put_failure;
564 NLA_PUT(skb, TCA_SFB_PARMS, sizeof(opt), &opt);
565 return nla_nest_end(skb, opts);
568 nla_nest_cancel(skb, opts);
572 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
574 struct sfb_sched_data *q = qdisc_priv(sch);
575 struct tc_sfb_xstats st = {
576 .earlydrop = q->stats.earlydrop,
577 .penaltydrop = q->stats.penaltydrop,
578 .bucketdrop = q->stats.bucketdrop,
579 .queuedrop = q->stats.queuedrop,
580 .childdrop = q->stats.childdrop,
581 .marked = q->stats.marked,
584 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
586 return gnet_stats_copy_app(d, &st, sizeof(st));
589 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
590 struct sk_buff *skb, struct tcmsg *tcm)
595 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
598 struct sfb_sched_data *q = qdisc_priv(sch);
606 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
608 sch_tree_unlock(sch);
612 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
614 struct sfb_sched_data *q = qdisc_priv(sch);
619 static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
624 static void sfb_put(struct Qdisc *sch, unsigned long arg)
628 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
629 struct nlattr **tca, unsigned long *arg)
634 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
639 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
642 if (walker->count >= walker->skip)
643 if (walker->fn(sch, 1, walker) < 0) {
651 static struct tcf_proto **sfb_find_tcf(struct Qdisc *sch, unsigned long cl)
653 struct sfb_sched_data *q = qdisc_priv(sch);
657 return &q->filter_list;
660 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
667 static const struct Qdisc_class_ops sfb_class_ops = {
672 .change = sfb_change_class,
673 .delete = sfb_delete,
675 .tcf_chain = sfb_find_tcf,
676 .bind_tcf = sfb_bind,
677 .unbind_tcf = sfb_put,
678 .dump = sfb_dump_class,
681 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
683 .priv_size = sizeof(struct sfb_sched_data),
684 .cl_ops = &sfb_class_ops,
685 .enqueue = sfb_enqueue,
686 .dequeue = sfb_dequeue,
690 .destroy = sfb_destroy,
691 .change = sfb_change,
693 .dump_stats = sfb_dump_stats,
694 .owner = THIS_MODULE,
697 static int __init sfb_module_init(void)
699 return register_qdisc(&sfb_qdisc_ops);
702 static void __exit sfb_module_exit(void)
704 unregister_qdisc(&sfb_qdisc_ops);
707 module_init(sfb_module_init)
708 module_exit(sfb_module_exit)
710 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
711 MODULE_AUTHOR("Juliusz Chroboczek");
712 MODULE_AUTHOR("Eric Dumazet");
713 MODULE_LICENSE("GPL");