s390: add 'install' target to 'make help'
[linux/fpc-iii.git] / net / sched / sch_sfb.c
blob5bbb6332ec5746dd1f33a04f4e22bcde5eb787c0
1 /*
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>
26 #include <net/ip.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" */
42 struct sfb_bucket {
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)
50 struct sfb_bins {
51 u32 perturbation; /* jhash perturbation */
52 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
55 struct sfb_sched_data {
56 struct Qdisc *qdisc;
57 struct tcf_proto __rcu *filter_list;
58 unsigned long rehash_interval;
59 unsigned long warmup_time; /* double buffering warmup time in jiffies */
60 u32 max;
61 u32 bin_size; /* maximum queue length per bin */
62 u32 increment; /* d1 */
63 u32 decrement; /* d2 */
64 u32 limit; /* HARD maximal queue length */
65 u32 penalty_rate;
66 u32 penalty_burst;
67 u32 tokens_avail;
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];
75 struct {
76 u32 earlydrop;
77 u32 penaltydrop;
78 u32 bucketdrop;
79 u32 queuedrop;
80 u32 childdrop; /* drops in child qdisc */
81 u32 marked; /* ECN mark */
82 } stats;
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)
90 struct sfb_skb_cb {
91 u32 hashes[2];
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)
115 u32 res = p1 + 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)
127 int i;
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)
135 b[hash].qlen++;
136 b += SFB_NUMBUCKETS; /* next level */
140 static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
142 u32 sfbhash;
144 sfbhash = sfb_hash(skb, 0);
145 if (sfbhash)
146 increment_one_qlen(sfbhash, 0, q);
148 sfbhash = sfb_hash(skb, 1);
149 if (sfbhash)
150 increment_one_qlen(sfbhash, 1, q);
153 static void decrement_one_qlen(u32 sfbhash, u32 slot,
154 struct sfb_sched_data *q)
156 int i;
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)
164 b[hash].qlen--;
165 b += SFB_NUMBUCKETS; /* next level */
169 static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
171 u32 sfbhash;
173 sfbhash = sfb_hash(skb, 0);
174 if (sfbhash)
175 decrement_one_qlen(sfbhash, 0, q);
177 sfbhash = sfb_hash(skb, 1);
178 if (sfbhash)
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)
202 int i;
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++) {
207 if (qlen < b->qlen)
208 qlen = b->qlen;
209 totalpm += b->p_mark;
210 if (prob < b->p_mark)
211 prob = b->p_mark;
212 b++;
214 *prob_r = prob;
215 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
216 return qlen;
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);
228 q->slot ^= 1;
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)
238 return true;
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)
248 return true;
251 q->tokens_avail--;
252 return false;
255 static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
256 int *qerr, u32 *salt)
258 struct tcf_result res;
259 int result;
261 result = tc_classify(skb, fl, &res, false);
262 if (result >= 0) {
263 #ifdef CONFIG_NET_CLS_ACT
264 switch (result) {
265 case TC_ACT_STOLEN:
266 case TC_ACT_QUEUED:
267 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
268 case TC_ACT_SHOT:
269 return false;
271 #endif
272 *salt = TC_H_MIN(res.classid);
273 return true;
275 return false;
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;
283 struct tcf_proto *fl;
284 int i;
285 u32 p_min = ~0;
286 u32 minqlen = ~0;
287 u32 r, sfbhash;
288 u32 slot = q->slot;
289 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
291 if (unlikely(sch->q.qlen >= q->limit)) {
292 qdisc_qstats_overlimit(sch);
293 q->stats.queuedrop++;
294 goto drop;
297 if (q->rehash_interval > 0) {
298 unsigned long limit = q->rehash_time + q->rehash_interval;
300 if (unlikely(time_after(jiffies, limit))) {
301 sfb_swap_slot(q);
302 q->rehash_time = jiffies;
303 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
304 time_after(jiffies, limit - q->warmup_time))) {
305 q->double_buffering = true;
309 fl = rcu_dereference_bh(q->filter_list);
310 if (fl) {
311 u32 salt;
313 /* If using external classifiers, get result and record it. */
314 if (!sfb_classify(skb, fl, &ret, &salt))
315 goto other_drop;
316 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
317 } else {
318 sfbhash = skb_get_hash_perturb(skb, q->bins[slot].perturbation);
322 if (!sfbhash)
323 sfbhash = 1;
324 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
326 for (i = 0; i < SFB_LEVELS; i++) {
327 u32 hash = sfbhash & SFB_BUCKET_MASK;
328 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
330 sfbhash >>= SFB_BUCKET_SHIFT;
331 if (b->qlen == 0)
332 decrement_prob(b, q);
333 else if (b->qlen >= q->bin_size)
334 increment_prob(b, q);
335 if (minqlen > b->qlen)
336 minqlen = b->qlen;
337 if (p_min > b->p_mark)
338 p_min = b->p_mark;
341 slot ^= 1;
342 sfb_skb_cb(skb)->hashes[slot] = 0;
344 if (unlikely(minqlen >= q->max)) {
345 qdisc_qstats_overlimit(sch);
346 q->stats.bucketdrop++;
347 goto drop;
350 if (unlikely(p_min >= SFB_MAX_PROB)) {
351 /* Inelastic flow */
352 if (q->double_buffering) {
353 sfbhash = skb_get_hash_perturb(skb,
354 q->bins[slot].perturbation);
355 if (!sfbhash)
356 sfbhash = 1;
357 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
359 for (i = 0; i < SFB_LEVELS; i++) {
360 u32 hash = sfbhash & SFB_BUCKET_MASK;
361 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
363 sfbhash >>= SFB_BUCKET_SHIFT;
364 if (b->qlen == 0)
365 decrement_prob(b, q);
366 else if (b->qlen >= q->bin_size)
367 increment_prob(b, q);
370 if (sfb_rate_limit(skb, q)) {
371 qdisc_qstats_overlimit(sch);
372 q->stats.penaltydrop++;
373 goto drop;
375 goto enqueue;
378 r = prandom_u32() & SFB_MAX_PROB;
380 if (unlikely(r < p_min)) {
381 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
382 /* If we're marking that many packets, then either
383 * this flow is unresponsive, or we're badly congested.
384 * In either case, we want to start dropping packets.
386 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
387 q->stats.earlydrop++;
388 goto drop;
391 if (INET_ECN_set_ce(skb)) {
392 q->stats.marked++;
393 } else {
394 q->stats.earlydrop++;
395 goto drop;
399 enqueue:
400 ret = qdisc_enqueue(skb, child);
401 if (likely(ret == NET_XMIT_SUCCESS)) {
402 sch->q.qlen++;
403 increment_qlen(skb, q);
404 } else if (net_xmit_drop_count(ret)) {
405 q->stats.childdrop++;
406 qdisc_qstats_drop(sch);
408 return ret;
410 drop:
411 qdisc_drop(skb, sch);
412 return NET_XMIT_CN;
413 other_drop:
414 if (ret & __NET_XMIT_BYPASS)
415 qdisc_qstats_drop(sch);
416 kfree_skb(skb);
417 return ret;
420 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
422 struct sfb_sched_data *q = qdisc_priv(sch);
423 struct Qdisc *child = q->qdisc;
424 struct sk_buff *skb;
426 skb = child->dequeue(q->qdisc);
428 if (skb) {
429 qdisc_bstats_update(sch, skb);
430 sch->q.qlen--;
431 decrement_qlen(skb, q);
434 return skb;
437 static struct sk_buff *sfb_peek(struct Qdisc *sch)
439 struct sfb_sched_data *q = qdisc_priv(sch);
440 struct Qdisc *child = q->qdisc;
442 return child->ops->peek(child);
445 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
447 static void sfb_reset(struct Qdisc *sch)
449 struct sfb_sched_data *q = qdisc_priv(sch);
451 qdisc_reset(q->qdisc);
452 sch->q.qlen = 0;
453 q->slot = 0;
454 q->double_buffering = false;
455 sfb_zero_all_buckets(q);
456 sfb_init_perturbation(0, q);
459 static void sfb_destroy(struct Qdisc *sch)
461 struct sfb_sched_data *q = qdisc_priv(sch);
463 tcf_destroy_chain(&q->filter_list);
464 qdisc_destroy(q->qdisc);
467 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
468 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
471 static const struct tc_sfb_qopt sfb_default_ops = {
472 .rehash_interval = 600 * MSEC_PER_SEC,
473 .warmup_time = 60 * MSEC_PER_SEC,
474 .limit = 0,
475 .max = 25,
476 .bin_size = 20,
477 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
478 .decrement = (SFB_MAX_PROB + 3000) / 6000,
479 .penalty_rate = 10,
480 .penalty_burst = 20,
483 static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
485 struct sfb_sched_data *q = qdisc_priv(sch);
486 struct Qdisc *child;
487 struct nlattr *tb[TCA_SFB_MAX + 1];
488 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
489 u32 limit;
490 int err;
492 if (opt) {
493 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
494 if (err < 0)
495 return -EINVAL;
497 if (tb[TCA_SFB_PARMS] == NULL)
498 return -EINVAL;
500 ctl = nla_data(tb[TCA_SFB_PARMS]);
503 limit = ctl->limit;
504 if (limit == 0)
505 limit = qdisc_dev(sch)->tx_queue_len;
507 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
508 if (IS_ERR(child))
509 return PTR_ERR(child);
511 sch_tree_lock(sch);
513 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
514 qdisc_destroy(q->qdisc);
515 q->qdisc = child;
517 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
518 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
519 q->rehash_time = jiffies;
520 q->limit = limit;
521 q->increment = ctl->increment;
522 q->decrement = ctl->decrement;
523 q->max = ctl->max;
524 q->bin_size = ctl->bin_size;
525 q->penalty_rate = ctl->penalty_rate;
526 q->penalty_burst = ctl->penalty_burst;
527 q->tokens_avail = ctl->penalty_burst;
528 q->token_time = jiffies;
530 q->slot = 0;
531 q->double_buffering = false;
532 sfb_zero_all_buckets(q);
533 sfb_init_perturbation(0, q);
534 sfb_init_perturbation(1, q);
536 sch_tree_unlock(sch);
538 return 0;
541 static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
543 struct sfb_sched_data *q = qdisc_priv(sch);
545 q->qdisc = &noop_qdisc;
546 return sfb_change(sch, opt);
549 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
551 struct sfb_sched_data *q = qdisc_priv(sch);
552 struct nlattr *opts;
553 struct tc_sfb_qopt opt = {
554 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
555 .warmup_time = jiffies_to_msecs(q->warmup_time),
556 .limit = q->limit,
557 .max = q->max,
558 .bin_size = q->bin_size,
559 .increment = q->increment,
560 .decrement = q->decrement,
561 .penalty_rate = q->penalty_rate,
562 .penalty_burst = q->penalty_burst,
565 sch->qstats.backlog = q->qdisc->qstats.backlog;
566 opts = nla_nest_start(skb, TCA_OPTIONS);
567 if (opts == NULL)
568 goto nla_put_failure;
569 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
570 goto nla_put_failure;
571 return nla_nest_end(skb, opts);
573 nla_put_failure:
574 nla_nest_cancel(skb, opts);
575 return -EMSGSIZE;
578 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
580 struct sfb_sched_data *q = qdisc_priv(sch);
581 struct tc_sfb_xstats st = {
582 .earlydrop = q->stats.earlydrop,
583 .penaltydrop = q->stats.penaltydrop,
584 .bucketdrop = q->stats.bucketdrop,
585 .queuedrop = q->stats.queuedrop,
586 .childdrop = q->stats.childdrop,
587 .marked = q->stats.marked,
590 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
592 return gnet_stats_copy_app(d, &st, sizeof(st));
595 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
596 struct sk_buff *skb, struct tcmsg *tcm)
598 return -ENOSYS;
601 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
602 struct Qdisc **old)
604 struct sfb_sched_data *q = qdisc_priv(sch);
606 if (new == NULL)
607 new = &noop_qdisc;
609 sch_tree_lock(sch);
610 *old = q->qdisc;
611 q->qdisc = new;
612 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
613 qdisc_reset(*old);
614 sch_tree_unlock(sch);
615 return 0;
618 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
620 struct sfb_sched_data *q = qdisc_priv(sch);
622 return q->qdisc;
625 static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
627 return 1;
630 static void sfb_put(struct Qdisc *sch, unsigned long arg)
634 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
635 struct nlattr **tca, unsigned long *arg)
637 return -ENOSYS;
640 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
642 return -ENOSYS;
645 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
647 if (!walker->stop) {
648 if (walker->count >= walker->skip)
649 if (walker->fn(sch, 1, walker) < 0) {
650 walker->stop = 1;
651 return;
653 walker->count++;
657 static struct tcf_proto __rcu **sfb_find_tcf(struct Qdisc *sch,
658 unsigned long cl)
660 struct sfb_sched_data *q = qdisc_priv(sch);
662 if (cl)
663 return NULL;
664 return &q->filter_list;
667 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
668 u32 classid)
670 return 0;
674 static const struct Qdisc_class_ops sfb_class_ops = {
675 .graft = sfb_graft,
676 .leaf = sfb_leaf,
677 .get = sfb_get,
678 .put = sfb_put,
679 .change = sfb_change_class,
680 .delete = sfb_delete,
681 .walk = sfb_walk,
682 .tcf_chain = sfb_find_tcf,
683 .bind_tcf = sfb_bind,
684 .unbind_tcf = sfb_put,
685 .dump = sfb_dump_class,
688 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
689 .id = "sfb",
690 .priv_size = sizeof(struct sfb_sched_data),
691 .cl_ops = &sfb_class_ops,
692 .enqueue = sfb_enqueue,
693 .dequeue = sfb_dequeue,
694 .peek = sfb_peek,
695 .init = sfb_init,
696 .reset = sfb_reset,
697 .destroy = sfb_destroy,
698 .change = sfb_change,
699 .dump = sfb_dump,
700 .dump_stats = sfb_dump_stats,
701 .owner = THIS_MODULE,
704 static int __init sfb_module_init(void)
706 return register_qdisc(&sfb_qdisc_ops);
709 static void __exit sfb_module_exit(void)
711 unregister_qdisc(&sfb_qdisc_ops);
714 module_init(sfb_module_init)
715 module_exit(sfb_module_exit)
717 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
718 MODULE_AUTHOR("Juliusz Chroboczek");
719 MODULE_AUTHOR("Eric Dumazet");
720 MODULE_LICENSE("GPL");