Input: uinput - avoid FF flush when destroying device
[linux-2.6/btrfs-unstable.git] / net / sched / sch_sfb.c
blob0f777273ba293dcf2e62ae484dd0033bce942a2b
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/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" */
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 __rcu *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 qdisc_cb_private_validate(skb, sizeof(struct sfb_skb_cb));
98 return (struct sfb_skb_cb *)qdisc_skb_cb(skb)->data;
102 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
103 * If using external classifier, hash comes from the classid.
105 static u32 sfb_hash(const struct sk_buff *skb, u32 slot)
107 return sfb_skb_cb(skb)->hashes[slot];
110 /* Probabilities are coded as Q0.16 fixed-point values,
111 * with 0xFFFF representing 65535/65536 (almost 1.0)
112 * Addition and subtraction are saturating in [0, 65535]
114 static u32 prob_plus(u32 p1, u32 p2)
116 u32 res = p1 + p2;
118 return min_t(u32, res, SFB_MAX_PROB);
121 static u32 prob_minus(u32 p1, u32 p2)
123 return p1 > p2 ? p1 - p2 : 0;
126 static void increment_one_qlen(u32 sfbhash, u32 slot, struct sfb_sched_data *q)
128 int i;
129 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
131 for (i = 0; i < SFB_LEVELS; i++) {
132 u32 hash = sfbhash & SFB_BUCKET_MASK;
134 sfbhash >>= SFB_BUCKET_SHIFT;
135 if (b[hash].qlen < 0xFFFF)
136 b[hash].qlen++;
137 b += SFB_NUMBUCKETS; /* next level */
141 static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
143 u32 sfbhash;
145 sfbhash = sfb_hash(skb, 0);
146 if (sfbhash)
147 increment_one_qlen(sfbhash, 0, q);
149 sfbhash = sfb_hash(skb, 1);
150 if (sfbhash)
151 increment_one_qlen(sfbhash, 1, q);
154 static void decrement_one_qlen(u32 sfbhash, u32 slot,
155 struct sfb_sched_data *q)
157 int i;
158 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
160 for (i = 0; i < SFB_LEVELS; i++) {
161 u32 hash = sfbhash & SFB_BUCKET_MASK;
163 sfbhash >>= SFB_BUCKET_SHIFT;
164 if (b[hash].qlen > 0)
165 b[hash].qlen--;
166 b += SFB_NUMBUCKETS; /* next level */
170 static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
172 u32 sfbhash;
174 sfbhash = sfb_hash(skb, 0);
175 if (sfbhash)
176 decrement_one_qlen(sfbhash, 0, q);
178 sfbhash = sfb_hash(skb, 1);
179 if (sfbhash)
180 decrement_one_qlen(sfbhash, 1, q);
183 static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
185 b->p_mark = prob_minus(b->p_mark, q->decrement);
188 static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
190 b->p_mark = prob_plus(b->p_mark, q->increment);
193 static void sfb_zero_all_buckets(struct sfb_sched_data *q)
195 memset(&q->bins, 0, sizeof(q->bins));
199 * compute max qlen, max p_mark, and avg p_mark
201 static u32 sfb_compute_qlen(u32 *prob_r, u32 *avgpm_r, const struct sfb_sched_data *q)
203 int i;
204 u32 qlen = 0, prob = 0, totalpm = 0;
205 const struct sfb_bucket *b = &q->bins[q->slot].bins[0][0];
207 for (i = 0; i < SFB_LEVELS * SFB_NUMBUCKETS; i++) {
208 if (qlen < b->qlen)
209 qlen = b->qlen;
210 totalpm += b->p_mark;
211 if (prob < b->p_mark)
212 prob = b->p_mark;
213 b++;
215 *prob_r = prob;
216 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
217 return qlen;
221 static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
223 q->bins[slot].perturbation = prandom_u32();
226 static void sfb_swap_slot(struct sfb_sched_data *q)
228 sfb_init_perturbation(q->slot, q);
229 q->slot ^= 1;
230 q->double_buffering = false;
233 /* Non elastic flows are allowed to use part of the bandwidth, expressed
234 * in "penalty_rate" packets per second, with "penalty_burst" burst
236 static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
238 if (q->penalty_rate == 0 || q->penalty_burst == 0)
239 return true;
241 if (q->tokens_avail < 1) {
242 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
244 q->tokens_avail = (age * q->penalty_rate) / HZ;
245 if (q->tokens_avail > q->penalty_burst)
246 q->tokens_avail = q->penalty_burst;
247 q->token_time = jiffies;
248 if (q->tokens_avail < 1)
249 return true;
252 q->tokens_avail--;
253 return false;
256 static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
257 int *qerr, u32 *salt)
259 struct tcf_result res;
260 int result;
262 result = tc_classify(skb, fl, &res, false);
263 if (result >= 0) {
264 #ifdef CONFIG_NET_CLS_ACT
265 switch (result) {
266 case TC_ACT_STOLEN:
267 case TC_ACT_QUEUED:
268 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
269 case TC_ACT_SHOT:
270 return false;
272 #endif
273 *salt = TC_H_MIN(res.classid);
274 return true;
276 return false;
279 static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
280 struct sk_buff **to_free)
283 struct sfb_sched_data *q = qdisc_priv(sch);
284 struct Qdisc *child = q->qdisc;
285 struct tcf_proto *fl;
286 int i;
287 u32 p_min = ~0;
288 u32 minqlen = ~0;
289 u32 r, sfbhash;
290 u32 slot = q->slot;
291 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
293 if (unlikely(sch->q.qlen >= q->limit)) {
294 qdisc_qstats_overlimit(sch);
295 q->stats.queuedrop++;
296 goto drop;
299 if (q->rehash_interval > 0) {
300 unsigned long limit = q->rehash_time + q->rehash_interval;
302 if (unlikely(time_after(jiffies, limit))) {
303 sfb_swap_slot(q);
304 q->rehash_time = jiffies;
305 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
306 time_after(jiffies, limit - q->warmup_time))) {
307 q->double_buffering = true;
311 fl = rcu_dereference_bh(q->filter_list);
312 if (fl) {
313 u32 salt;
315 /* If using external classifiers, get result and record it. */
316 if (!sfb_classify(skb, fl, &ret, &salt))
317 goto other_drop;
318 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
319 } else {
320 sfbhash = skb_get_hash_perturb(skb, q->bins[slot].perturbation);
324 if (!sfbhash)
325 sfbhash = 1;
326 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
328 for (i = 0; i < SFB_LEVELS; i++) {
329 u32 hash = sfbhash & SFB_BUCKET_MASK;
330 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
332 sfbhash >>= SFB_BUCKET_SHIFT;
333 if (b->qlen == 0)
334 decrement_prob(b, q);
335 else if (b->qlen >= q->bin_size)
336 increment_prob(b, q);
337 if (minqlen > b->qlen)
338 minqlen = b->qlen;
339 if (p_min > b->p_mark)
340 p_min = b->p_mark;
343 slot ^= 1;
344 sfb_skb_cb(skb)->hashes[slot] = 0;
346 if (unlikely(minqlen >= q->max)) {
347 qdisc_qstats_overlimit(sch);
348 q->stats.bucketdrop++;
349 goto drop;
352 if (unlikely(p_min >= SFB_MAX_PROB)) {
353 /* Inelastic flow */
354 if (q->double_buffering) {
355 sfbhash = skb_get_hash_perturb(skb,
356 q->bins[slot].perturbation);
357 if (!sfbhash)
358 sfbhash = 1;
359 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
361 for (i = 0; i < SFB_LEVELS; i++) {
362 u32 hash = sfbhash & SFB_BUCKET_MASK;
363 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
365 sfbhash >>= SFB_BUCKET_SHIFT;
366 if (b->qlen == 0)
367 decrement_prob(b, q);
368 else if (b->qlen >= q->bin_size)
369 increment_prob(b, q);
372 if (sfb_rate_limit(skb, q)) {
373 qdisc_qstats_overlimit(sch);
374 q->stats.penaltydrop++;
375 goto drop;
377 goto enqueue;
380 r = prandom_u32() & SFB_MAX_PROB;
382 if (unlikely(r < p_min)) {
383 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
384 /* If we're marking that many packets, then either
385 * this flow is unresponsive, or we're badly congested.
386 * In either case, we want to start dropping packets.
388 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
389 q->stats.earlydrop++;
390 goto drop;
393 if (INET_ECN_set_ce(skb)) {
394 q->stats.marked++;
395 } else {
396 q->stats.earlydrop++;
397 goto drop;
401 enqueue:
402 ret = qdisc_enqueue(skb, child, to_free);
403 if (likely(ret == NET_XMIT_SUCCESS)) {
404 qdisc_qstats_backlog_inc(sch, skb);
405 sch->q.qlen++;
406 increment_qlen(skb, q);
407 } else if (net_xmit_drop_count(ret)) {
408 q->stats.childdrop++;
409 qdisc_qstats_drop(sch);
411 return ret;
413 drop:
414 qdisc_drop(skb, sch, to_free);
415 return NET_XMIT_CN;
416 other_drop:
417 if (ret & __NET_XMIT_BYPASS)
418 qdisc_qstats_drop(sch);
419 kfree_skb(skb);
420 return ret;
423 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
425 struct sfb_sched_data *q = qdisc_priv(sch);
426 struct Qdisc *child = q->qdisc;
427 struct sk_buff *skb;
429 skb = child->dequeue(q->qdisc);
431 if (skb) {
432 qdisc_bstats_update(sch, skb);
433 qdisc_qstats_backlog_dec(sch, skb);
434 sch->q.qlen--;
435 decrement_qlen(skb, q);
438 return skb;
441 static struct sk_buff *sfb_peek(struct Qdisc *sch)
443 struct sfb_sched_data *q = qdisc_priv(sch);
444 struct Qdisc *child = q->qdisc;
446 return child->ops->peek(child);
449 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
451 static void sfb_reset(struct Qdisc *sch)
453 struct sfb_sched_data *q = qdisc_priv(sch);
455 qdisc_reset(q->qdisc);
456 sch->qstats.backlog = 0;
457 sch->q.qlen = 0;
458 q->slot = 0;
459 q->double_buffering = false;
460 sfb_zero_all_buckets(q);
461 sfb_init_perturbation(0, q);
464 static void sfb_destroy(struct Qdisc *sch)
466 struct sfb_sched_data *q = qdisc_priv(sch);
468 tcf_destroy_chain(&q->filter_list);
469 qdisc_destroy(q->qdisc);
472 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
473 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
476 static const struct tc_sfb_qopt sfb_default_ops = {
477 .rehash_interval = 600 * MSEC_PER_SEC,
478 .warmup_time = 60 * MSEC_PER_SEC,
479 .limit = 0,
480 .max = 25,
481 .bin_size = 20,
482 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
483 .decrement = (SFB_MAX_PROB + 3000) / 6000,
484 .penalty_rate = 10,
485 .penalty_burst = 20,
488 static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
490 struct sfb_sched_data *q = qdisc_priv(sch);
491 struct Qdisc *child;
492 struct nlattr *tb[TCA_SFB_MAX + 1];
493 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
494 u32 limit;
495 int err;
497 if (opt) {
498 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy, NULL);
499 if (err < 0)
500 return -EINVAL;
502 if (tb[TCA_SFB_PARMS] == NULL)
503 return -EINVAL;
505 ctl = nla_data(tb[TCA_SFB_PARMS]);
508 limit = ctl->limit;
509 if (limit == 0)
510 limit = qdisc_dev(sch)->tx_queue_len;
512 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
513 if (IS_ERR(child))
514 return PTR_ERR(child);
516 if (child != &noop_qdisc)
517 qdisc_hash_add(child, true);
518 sch_tree_lock(sch);
520 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
521 q->qdisc->qstats.backlog);
522 qdisc_destroy(q->qdisc);
523 q->qdisc = child;
525 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
526 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
527 q->rehash_time = jiffies;
528 q->limit = limit;
529 q->increment = ctl->increment;
530 q->decrement = ctl->decrement;
531 q->max = ctl->max;
532 q->bin_size = ctl->bin_size;
533 q->penalty_rate = ctl->penalty_rate;
534 q->penalty_burst = ctl->penalty_burst;
535 q->tokens_avail = ctl->penalty_burst;
536 q->token_time = jiffies;
538 q->slot = 0;
539 q->double_buffering = false;
540 sfb_zero_all_buckets(q);
541 sfb_init_perturbation(0, q);
542 sfb_init_perturbation(1, q);
544 sch_tree_unlock(sch);
546 return 0;
549 static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
551 struct sfb_sched_data *q = qdisc_priv(sch);
553 q->qdisc = &noop_qdisc;
554 return sfb_change(sch, opt);
557 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
559 struct sfb_sched_data *q = qdisc_priv(sch);
560 struct nlattr *opts;
561 struct tc_sfb_qopt opt = {
562 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
563 .warmup_time = jiffies_to_msecs(q->warmup_time),
564 .limit = q->limit,
565 .max = q->max,
566 .bin_size = q->bin_size,
567 .increment = q->increment,
568 .decrement = q->decrement,
569 .penalty_rate = q->penalty_rate,
570 .penalty_burst = q->penalty_burst,
573 sch->qstats.backlog = q->qdisc->qstats.backlog;
574 opts = nla_nest_start(skb, TCA_OPTIONS);
575 if (opts == NULL)
576 goto nla_put_failure;
577 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
578 goto nla_put_failure;
579 return nla_nest_end(skb, opts);
581 nla_put_failure:
582 nla_nest_cancel(skb, opts);
583 return -EMSGSIZE;
586 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
588 struct sfb_sched_data *q = qdisc_priv(sch);
589 struct tc_sfb_xstats st = {
590 .earlydrop = q->stats.earlydrop,
591 .penaltydrop = q->stats.penaltydrop,
592 .bucketdrop = q->stats.bucketdrop,
593 .queuedrop = q->stats.queuedrop,
594 .childdrop = q->stats.childdrop,
595 .marked = q->stats.marked,
598 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
600 return gnet_stats_copy_app(d, &st, sizeof(st));
603 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
604 struct sk_buff *skb, struct tcmsg *tcm)
606 return -ENOSYS;
609 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
610 struct Qdisc **old)
612 struct sfb_sched_data *q = qdisc_priv(sch);
614 if (new == NULL)
615 new = &noop_qdisc;
617 *old = qdisc_replace(sch, new, &q->qdisc);
618 return 0;
621 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
623 struct sfb_sched_data *q = qdisc_priv(sch);
625 return q->qdisc;
628 static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
630 return 1;
633 static void sfb_put(struct Qdisc *sch, unsigned long arg)
637 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
638 struct nlattr **tca, unsigned long *arg)
640 return -ENOSYS;
643 static int sfb_delete(struct Qdisc *sch, unsigned long cl)
645 return -ENOSYS;
648 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
650 if (!walker->stop) {
651 if (walker->count >= walker->skip)
652 if (walker->fn(sch, 1, walker) < 0) {
653 walker->stop = 1;
654 return;
656 walker->count++;
660 static struct tcf_proto __rcu **sfb_find_tcf(struct Qdisc *sch,
661 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");