ath6kl: Use vmalloc instead of kmalloc for fw
[linux-2.6.git] / net / sched / sch_sfb.c
blob96e42cae4c7a4580e753c41b9c4f1e5c29cabbcd
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>
29 #include <net/flow_keys.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" */
43 struct sfb_bucket {
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)
51 struct sfb_bins {
52 u32 perturbation; /* jhash perturbation */
53 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
56 struct sfb_sched_data {
57 struct Qdisc *qdisc;
58 struct tcf_proto *filter_list;
59 unsigned long rehash_interval;
60 unsigned long warmup_time; /* double buffering warmup time in jiffies */
61 u32 max;
62 u32 bin_size; /* maximum queue length per bin */
63 u32 increment; /* d1 */
64 u32 decrement; /* d2 */
65 u32 limit; /* HARD maximal queue length */
66 u32 penalty_rate;
67 u32 penalty_burst;
68 u32 tokens_avail;
69 unsigned long rehash_time;
70 unsigned long token_time;
72 u8 slot; /* current active bins (0 or 1) */
73 bool double_buffering;
74 struct sfb_bins bins[2];
76 struct {
77 u32 earlydrop;
78 u32 penaltydrop;
79 u32 bucketdrop;
80 u32 queuedrop;
81 u32 childdrop; /* drops in child qdisc */
82 u32 marked; /* ECN mark */
83 } stats;
87 * Each queued skb might be hashed on one or two bins
88 * We store in skb_cb the two hash values.
89 * (A zero value means double buffering was not used)
91 struct sfb_skb_cb {
92 u32 hashes[2];
95 static inline struct sfb_skb_cb *sfb_skb_cb(const struct sk_buff *skb)
97 BUILD_BUG_ON(sizeof(skb->cb) <
98 sizeof(struct qdisc_skb_cb) + 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)
117 u32 res = p1 + 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)
129 int i;
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)
137 b[hash].qlen++;
138 b += SFB_NUMBUCKETS; /* next level */
142 static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
144 u32 sfbhash;
146 sfbhash = sfb_hash(skb, 0);
147 if (sfbhash)
148 increment_one_qlen(sfbhash, 0, q);
150 sfbhash = sfb_hash(skb, 1);
151 if (sfbhash)
152 increment_one_qlen(sfbhash, 1, q);
155 static void decrement_one_qlen(u32 sfbhash, u32 slot,
156 struct sfb_sched_data *q)
158 int i;
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)
166 b[hash].qlen--;
167 b += SFB_NUMBUCKETS; /* next level */
171 static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
173 u32 sfbhash;
175 sfbhash = sfb_hash(skb, 0);
176 if (sfbhash)
177 decrement_one_qlen(sfbhash, 0, q);
179 sfbhash = sfb_hash(skb, 1);
180 if (sfbhash)
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)
204 int i;
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++) {
209 if (qlen < b->qlen)
210 qlen = b->qlen;
211 totalpm += b->p_mark;
212 if (prob < b->p_mark)
213 prob = b->p_mark;
214 b++;
216 *prob_r = prob;
217 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
218 return qlen;
222 static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
224 q->bins[slot].perturbation = net_random();
227 static void sfb_swap_slot(struct sfb_sched_data *q)
229 sfb_init_perturbation(q->slot, q);
230 q->slot ^= 1;
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)
240 return true;
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)
250 return true;
253 q->tokens_avail--;
254 return false;
257 static bool sfb_classify(struct sk_buff *skb, struct sfb_sched_data *q,
258 int *qerr, u32 *salt)
260 struct tcf_result res;
261 int result;
263 result = tc_classify(skb, q->filter_list, &res);
264 if (result >= 0) {
265 #ifdef CONFIG_NET_CLS_ACT
266 switch (result) {
267 case TC_ACT_STOLEN:
268 case TC_ACT_QUEUED:
269 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
270 case TC_ACT_SHOT:
271 return false;
273 #endif
274 *salt = TC_H_MIN(res.classid);
275 return true;
277 return false;
280 static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
283 struct sfb_sched_data *q = qdisc_priv(sch);
284 struct Qdisc *child = q->qdisc;
285 int i;
286 u32 p_min = ~0;
287 u32 minqlen = ~0;
288 u32 r, slot, salt, sfbhash;
289 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
290 struct flow_keys keys;
292 if (unlikely(sch->q.qlen >= q->limit)) {
293 sch->qstats.overlimits++;
294 q->stats.queuedrop++;
295 goto drop;
298 if (q->rehash_interval > 0) {
299 unsigned long limit = q->rehash_time + q->rehash_interval;
301 if (unlikely(time_after(jiffies, limit))) {
302 sfb_swap_slot(q);
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 if (q->filter_list) {
311 /* If using external classifiers, get result and record it. */
312 if (!sfb_classify(skb, q, &ret, &salt))
313 goto other_drop;
314 keys.src = salt;
315 keys.dst = 0;
316 keys.ports = 0;
317 } else {
318 skb_flow_dissect(skb, &keys);
321 slot = q->slot;
323 sfbhash = jhash_3words((__force u32)keys.dst,
324 (__force u32)keys.src,
325 (__force u32)keys.ports,
326 q->bins[slot].perturbation);
327 if (!sfbhash)
328 sfbhash = 1;
329 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
331 for (i = 0; i < SFB_LEVELS; i++) {
332 u32 hash = sfbhash & SFB_BUCKET_MASK;
333 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
335 sfbhash >>= SFB_BUCKET_SHIFT;
336 if (b->qlen == 0)
337 decrement_prob(b, q);
338 else if (b->qlen >= q->bin_size)
339 increment_prob(b, q);
340 if (minqlen > b->qlen)
341 minqlen = b->qlen;
342 if (p_min > b->p_mark)
343 p_min = b->p_mark;
346 slot ^= 1;
347 sfb_skb_cb(skb)->hashes[slot] = 0;
349 if (unlikely(minqlen >= q->max)) {
350 sch->qstats.overlimits++;
351 q->stats.bucketdrop++;
352 goto drop;
355 if (unlikely(p_min >= SFB_MAX_PROB)) {
356 /* Inelastic flow */
357 if (q->double_buffering) {
358 sfbhash = jhash_3words((__force u32)keys.dst,
359 (__force u32)keys.src,
360 (__force u32)keys.ports,
361 q->bins[slot].perturbation);
362 if (!sfbhash)
363 sfbhash = 1;
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;
371 if (b->qlen == 0)
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 sch->qstats.overlimits++;
379 q->stats.penaltydrop++;
380 goto drop;
382 goto enqueue;
385 r = net_random() & 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++;
395 goto drop;
398 if (INET_ECN_set_ce(skb)) {
399 q->stats.marked++;
400 } else {
401 q->stats.earlydrop++;
402 goto drop;
406 enqueue:
407 ret = qdisc_enqueue(skb, child);
408 if (likely(ret == NET_XMIT_SUCCESS)) {
409 sch->q.qlen++;
410 increment_qlen(skb, q);
411 } else if (net_xmit_drop_count(ret)) {
412 q->stats.childdrop++;
413 sch->qstats.drops++;
415 return ret;
417 drop:
418 qdisc_drop(skb, sch);
419 return NET_XMIT_CN;
420 other_drop:
421 if (ret & __NET_XMIT_BYPASS)
422 sch->qstats.drops++;
423 kfree_skb(skb);
424 return ret;
427 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
429 struct sfb_sched_data *q = qdisc_priv(sch);
430 struct Qdisc *child = q->qdisc;
431 struct sk_buff *skb;
433 skb = child->dequeue(q->qdisc);
435 if (skb) {
436 qdisc_bstats_update(sch, skb);
437 sch->q.qlen--;
438 decrement_qlen(skb, q);
441 return skb;
444 static struct sk_buff *sfb_peek(struct Qdisc *sch)
446 struct sfb_sched_data *q = qdisc_priv(sch);
447 struct Qdisc *child = q->qdisc;
449 return child->ops->peek(child);
452 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
454 static void sfb_reset(struct Qdisc *sch)
456 struct sfb_sched_data *q = qdisc_priv(sch);
458 qdisc_reset(q->qdisc);
459 sch->q.qlen = 0;
460 q->slot = 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_destroy_chain(&q->filter_list);
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,
481 .limit = 0,
482 .max = 25,
483 .bin_size = 20,
484 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
485 .decrement = (SFB_MAX_PROB + 3000) / 6000,
486 .penalty_rate = 10,
487 .penalty_burst = 20,
490 static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
492 struct sfb_sched_data *q = qdisc_priv(sch);
493 struct Qdisc *child;
494 struct nlattr *tb[TCA_SFB_MAX + 1];
495 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
496 u32 limit;
497 int err;
499 if (opt) {
500 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
501 if (err < 0)
502 return -EINVAL;
504 if (tb[TCA_SFB_PARMS] == NULL)
505 return -EINVAL;
507 ctl = nla_data(tb[TCA_SFB_PARMS]);
510 limit = ctl->limit;
511 if (limit == 0)
512 limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
514 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
515 if (IS_ERR(child))
516 return PTR_ERR(child);
518 sch_tree_lock(sch);
520 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
521 qdisc_destroy(q->qdisc);
522 q->qdisc = child;
524 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
525 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
526 q->rehash_time = jiffies;
527 q->limit = limit;
528 q->increment = ctl->increment;
529 q->decrement = ctl->decrement;
530 q->max = ctl->max;
531 q->bin_size = ctl->bin_size;
532 q->penalty_rate = ctl->penalty_rate;
533 q->penalty_burst = ctl->penalty_burst;
534 q->tokens_avail = ctl->penalty_burst;
535 q->token_time = jiffies;
537 q->slot = 0;
538 q->double_buffering = false;
539 sfb_zero_all_buckets(q);
540 sfb_init_perturbation(0, q);
541 sfb_init_perturbation(1, q);
543 sch_tree_unlock(sch);
545 return 0;
548 static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
550 struct sfb_sched_data *q = qdisc_priv(sch);
552 q->qdisc = &noop_qdisc;
553 return sfb_change(sch, opt);
556 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
558 struct sfb_sched_data *q = qdisc_priv(sch);
559 struct nlattr *opts;
560 struct tc_sfb_qopt opt = {
561 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
562 .warmup_time = jiffies_to_msecs(q->warmup_time),
563 .limit = q->limit,
564 .max = q->max,
565 .bin_size = q->bin_size,
566 .increment = q->increment,
567 .decrement = q->decrement,
568 .penalty_rate = q->penalty_rate,
569 .penalty_burst = q->penalty_burst,
572 sch->qstats.backlog = q->qdisc->qstats.backlog;
573 opts = nla_nest_start(skb, TCA_OPTIONS);
574 NLA_PUT(skb, TCA_SFB_PARMS, sizeof(opt), &opt);
575 return nla_nest_end(skb, opts);
577 nla_put_failure:
578 nla_nest_cancel(skb, opts);
579 return -EMSGSIZE;
582 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
584 struct sfb_sched_data *q = qdisc_priv(sch);
585 struct tc_sfb_xstats st = {
586 .earlydrop = q->stats.earlydrop,
587 .penaltydrop = q->stats.penaltydrop,
588 .bucketdrop = q->stats.bucketdrop,
589 .queuedrop = q->stats.queuedrop,
590 .childdrop = q->stats.childdrop,
591 .marked = q->stats.marked,
594 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
596 return gnet_stats_copy_app(d, &st, sizeof(st));
599 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
600 struct sk_buff *skb, struct tcmsg *tcm)
602 return -ENOSYS;
605 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
606 struct Qdisc **old)
608 struct sfb_sched_data *q = qdisc_priv(sch);
610 if (new == NULL)
611 new = &noop_qdisc;
613 sch_tree_lock(sch);
614 *old = q->qdisc;
615 q->qdisc = new;
616 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
617 qdisc_reset(*old);
618 sch_tree_unlock(sch);
619 return 0;
622 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
624 struct sfb_sched_data *q = qdisc_priv(sch);
626 return q->qdisc;
629 static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
631 return 1;
634 static void sfb_put(struct Qdisc *sch, unsigned long arg)
638 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
639 struct nlattr **tca, unsigned long *arg)
641 return -ENOSYS;
644 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
646 return -ENOSYS;
649 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
651 if (!walker->stop) {
652 if (walker->count >= walker->skip)
653 if (walker->fn(sch, 1, walker) < 0) {
654 walker->stop = 1;
655 return;
657 walker->count++;
661 static struct tcf_proto **sfb_find_tcf(struct Qdisc *sch, unsigned long cl)
663 struct sfb_sched_data *q = qdisc_priv(sch);
665 if (cl)
666 return NULL;
667 return &q->filter_list;
670 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
671 u32 classid)
673 return 0;
677 static const struct Qdisc_class_ops sfb_class_ops = {
678 .graft = sfb_graft,
679 .leaf = sfb_leaf,
680 .get = sfb_get,
681 .put = sfb_put,
682 .change = sfb_change_class,
683 .delete = sfb_delete,
684 .walk = sfb_walk,
685 .tcf_chain = sfb_find_tcf,
686 .bind_tcf = sfb_bind,
687 .unbind_tcf = sfb_put,
688 .dump = sfb_dump_class,
691 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
692 .id = "sfb",
693 .priv_size = sizeof(struct sfb_sched_data),
694 .cl_ops = &sfb_class_ops,
695 .enqueue = sfb_enqueue,
696 .dequeue = sfb_dequeue,
697 .peek = sfb_peek,
698 .init = sfb_init,
699 .reset = sfb_reset,
700 .destroy = sfb_destroy,
701 .change = sfb_change,
702 .dump = sfb_dump,
703 .dump_stats = sfb_dump_stats,
704 .owner = THIS_MODULE,
707 static int __init sfb_module_init(void)
709 return register_qdisc(&sfb_qdisc_ops);
712 static void __exit sfb_module_exit(void)
714 unregister_qdisc(&sfb_qdisc_ops);
717 module_init(sfb_module_init)
718 module_exit(sfb_module_exit)
720 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
721 MODULE_AUTHOR("Juliusz Chroboczek");
722 MODULE_AUTHOR("Eric Dumazet");
723 MODULE_LICENSE("GPL");