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/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 u32 perturbation; /* jhash perturbation */
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 sk_buff *skb, struct sfb_sched_data *q)
146 sfbhash = sfb_hash(skb, 0);
148 increment_one_qlen(sfbhash, 0, q);
150 sfbhash = sfb_hash(skb, 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 q->bins[slot].perturbation = prandom_u32();
227 static void sfb_swap_slot(struct sfb_sched_data *q)
229 sfb_init_perturbation(q->slot, q);
231 q->double_buffering = false;
234 /* Non elastic flows are allowed to use part of the bandwidth, expressed
235 * in "penalty_rate" packets per second, with "penalty_burst" burst
237 static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
239 if (q->penalty_rate == 0 || q->penalty_burst == 0)
242 if (q->tokens_avail < 1) {
243 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
245 q->tokens_avail = (age * q->penalty_rate) / HZ;
246 if (q->tokens_avail > q->penalty_burst)
247 q->tokens_avail = q->penalty_burst;
248 q->token_time = jiffies;
249 if (q->tokens_avail < 1)
257 static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
258 int *qerr, u32 *salt)
260 struct tcf_result res;
263 result = tcf_classify(skb, fl, &res, false);
265 #ifdef CONFIG_NET_CLS_ACT
270 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
275 *salt = TC_H_MIN(res.classid);
281 static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
282 struct sk_buff **to_free)
285 struct sfb_sched_data *q = qdisc_priv(sch);
286 struct Qdisc *child = q->qdisc;
287 struct tcf_proto *fl;
293 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
295 if (unlikely(sch->q.qlen >= q->limit)) {
296 qdisc_qstats_overlimit(sch);
297 q->stats.queuedrop++;
301 if (q->rehash_interval > 0) {
302 unsigned long limit = q->rehash_time + q->rehash_interval;
304 if (unlikely(time_after(jiffies, limit))) {
306 q->rehash_time = jiffies;
307 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
308 time_after(jiffies, limit - q->warmup_time))) {
309 q->double_buffering = true;
313 fl = rcu_dereference_bh(q->filter_list);
317 /* If using external classifiers, get result and record it. */
318 if (!sfb_classify(skb, fl, &ret, &salt))
320 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
322 sfbhash = skb_get_hash_perturb(skb, q->bins[slot].perturbation);
328 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
330 for (i = 0; i < SFB_LEVELS; i++) {
331 u32 hash = sfbhash & SFB_BUCKET_MASK;
332 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
334 sfbhash >>= SFB_BUCKET_SHIFT;
336 decrement_prob(b, q);
337 else if (b->qlen >= q->bin_size)
338 increment_prob(b, q);
339 if (minqlen > b->qlen)
341 if (p_min > b->p_mark)
346 sfb_skb_cb(skb)->hashes[slot] = 0;
348 if (unlikely(minqlen >= q->max)) {
349 qdisc_qstats_overlimit(sch);
350 q->stats.bucketdrop++;
354 if (unlikely(p_min >= SFB_MAX_PROB)) {
356 if (q->double_buffering) {
357 sfbhash = skb_get_hash_perturb(skb,
358 q->bins[slot].perturbation);
361 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
363 for (i = 0; i < SFB_LEVELS; i++) {
364 u32 hash = sfbhash & SFB_BUCKET_MASK;
365 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
367 sfbhash >>= SFB_BUCKET_SHIFT;
369 decrement_prob(b, q);
370 else if (b->qlen >= q->bin_size)
371 increment_prob(b, q);
374 if (sfb_rate_limit(skb, q)) {
375 qdisc_qstats_overlimit(sch);
376 q->stats.penaltydrop++;
382 r = prandom_u32() & SFB_MAX_PROB;
384 if (unlikely(r < p_min)) {
385 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
386 /* If we're marking that many packets, then either
387 * this flow is unresponsive, or we're badly congested.
388 * In either case, we want to start dropping packets.
390 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
391 q->stats.earlydrop++;
395 if (INET_ECN_set_ce(skb)) {
398 q->stats.earlydrop++;
404 ret = qdisc_enqueue(skb, child, to_free);
405 if (likely(ret == NET_XMIT_SUCCESS)) {
406 qdisc_qstats_backlog_inc(sch, skb);
408 increment_qlen(skb, q);
409 } else if (net_xmit_drop_count(ret)) {
410 q->stats.childdrop++;
411 qdisc_qstats_drop(sch);
416 qdisc_drop(skb, sch, to_free);
419 if (ret & __NET_XMIT_BYPASS)
420 qdisc_qstats_drop(sch);
425 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
427 struct sfb_sched_data *q = qdisc_priv(sch);
428 struct Qdisc *child = q->qdisc;
431 skb = child->dequeue(q->qdisc);
434 qdisc_bstats_update(sch, skb);
435 qdisc_qstats_backlog_dec(sch, skb);
437 decrement_qlen(skb, q);
443 static struct sk_buff *sfb_peek(struct Qdisc *sch)
445 struct sfb_sched_data *q = qdisc_priv(sch);
446 struct Qdisc *child = q->qdisc;
448 return child->ops->peek(child);
451 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
453 static void sfb_reset(struct Qdisc *sch)
455 struct sfb_sched_data *q = qdisc_priv(sch);
457 qdisc_reset(q->qdisc);
458 sch->qstats.backlog = 0;
461 q->double_buffering = false;
462 sfb_zero_all_buckets(q);
463 sfb_init_perturbation(0, q);
466 static void sfb_destroy(struct Qdisc *sch)
468 struct sfb_sched_data *q = qdisc_priv(sch);
470 tcf_block_put(q->block);
471 qdisc_destroy(q->qdisc);
474 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
475 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
478 static const struct tc_sfb_qopt sfb_default_ops = {
479 .rehash_interval = 600 * MSEC_PER_SEC,
480 .warmup_time = 60 * MSEC_PER_SEC,
484 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
485 .decrement = (SFB_MAX_PROB + 3000) / 6000,
490 static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
492 struct sfb_sched_data *q = qdisc_priv(sch);
494 struct nlattr *tb[TCA_SFB_MAX + 1];
495 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
500 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy, NULL);
504 if (tb[TCA_SFB_PARMS] == NULL)
507 ctl = nla_data(tb[TCA_SFB_PARMS]);
512 limit = qdisc_dev(sch)->tx_queue_len;
514 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
516 return PTR_ERR(child);
518 if (child != &noop_qdisc)
519 qdisc_hash_add(child, true);
522 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
523 q->qdisc->qstats.backlog);
524 qdisc_destroy(q->qdisc);
527 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
528 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
529 q->rehash_time = jiffies;
531 q->increment = ctl->increment;
532 q->decrement = ctl->decrement;
534 q->bin_size = ctl->bin_size;
535 q->penalty_rate = ctl->penalty_rate;
536 q->penalty_burst = ctl->penalty_burst;
537 q->tokens_avail = ctl->penalty_burst;
538 q->token_time = jiffies;
541 q->double_buffering = false;
542 sfb_zero_all_buckets(q);
543 sfb_init_perturbation(0, q);
544 sfb_init_perturbation(1, q);
546 sch_tree_unlock(sch);
551 static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
553 struct sfb_sched_data *q = qdisc_priv(sch);
556 err = tcf_block_get(&q->block, &q->filter_list);
560 q->qdisc = &noop_qdisc;
561 return sfb_change(sch, opt);
564 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
566 struct sfb_sched_data *q = qdisc_priv(sch);
568 struct tc_sfb_qopt opt = {
569 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
570 .warmup_time = jiffies_to_msecs(q->warmup_time),
573 .bin_size = q->bin_size,
574 .increment = q->increment,
575 .decrement = q->decrement,
576 .penalty_rate = q->penalty_rate,
577 .penalty_burst = q->penalty_burst,
580 sch->qstats.backlog = q->qdisc->qstats.backlog;
581 opts = nla_nest_start(skb, TCA_OPTIONS);
583 goto nla_put_failure;
584 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
585 goto nla_put_failure;
586 return nla_nest_end(skb, opts);
589 nla_nest_cancel(skb, opts);
593 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
595 struct sfb_sched_data *q = qdisc_priv(sch);
596 struct tc_sfb_xstats st = {
597 .earlydrop = q->stats.earlydrop,
598 .penaltydrop = q->stats.penaltydrop,
599 .bucketdrop = q->stats.bucketdrop,
600 .queuedrop = q->stats.queuedrop,
601 .childdrop = q->stats.childdrop,
602 .marked = q->stats.marked,
605 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
607 return gnet_stats_copy_app(d, &st, sizeof(st));
610 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
611 struct sk_buff *skb, struct tcmsg *tcm)
616 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
619 struct sfb_sched_data *q = qdisc_priv(sch);
624 *old = qdisc_replace(sch, new, &q->qdisc);
628 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
630 struct sfb_sched_data *q = qdisc_priv(sch);
635 static unsigned long sfb_find(struct Qdisc *sch, u32 classid)
640 static void sfb_unbind(struct Qdisc *sch, unsigned long arg)
644 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
645 struct nlattr **tca, unsigned long *arg)
650 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
655 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
658 if (walker->count >= walker->skip)
659 if (walker->fn(sch, 1, walker) < 0) {
667 static struct tcf_block *sfb_tcf_block(struct Qdisc *sch, unsigned long cl)
669 struct sfb_sched_data *q = qdisc_priv(sch);
676 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
683 static const struct Qdisc_class_ops sfb_class_ops = {
687 .change = sfb_change_class,
688 .delete = sfb_delete,
690 .tcf_block = sfb_tcf_block,
691 .bind_tcf = sfb_bind,
692 .unbind_tcf = sfb_unbind,
693 .dump = sfb_dump_class,
696 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
698 .priv_size = sizeof(struct sfb_sched_data),
699 .cl_ops = &sfb_class_ops,
700 .enqueue = sfb_enqueue,
701 .dequeue = sfb_dequeue,
705 .destroy = sfb_destroy,
706 .change = sfb_change,
708 .dump_stats = sfb_dump_stats,
709 .owner = THIS_MODULE,
712 static int __init sfb_module_init(void)
714 return register_qdisc(&sfb_qdisc_ops);
717 static void __exit sfb_module_exit(void)
719 unregister_qdisc(&sfb_qdisc_ops);
722 module_init(sfb_module_init)
723 module_exit(sfb_module_exit)
725 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
726 MODULE_AUTHOR("Juliusz Chroboczek");
727 MODULE_AUTHOR("Eric Dumazet");
728 MODULE_LICENSE("GPL");