Loading...
1/*
2 * net/sched/sch_sfb.c Stochastic Fair Blue
3 *
4 * Copyright (c) 2008-2011 Juliusz Chroboczek <jch@pps.jussieu.fr>
5 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
6 *
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.
10 *
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.
14 *
15 * http://www.thefengs.com/wuchang/blue/CSE-TR-387-99.pdf
16 *
17 */
18
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>
30
31/*
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.
36 */
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 */
41
42/* SFB algo uses a virtual queue, named "bin" */
43struct sfb_bucket {
44 u16 qlen; /* length of virtual queue */
45 u16 p_mark; /* marking probability */
46};
47
48/* We use a double buffering right before hash change
49 * (Section 4.4 of SFB reference : moving hash functions)
50 */
51struct sfb_bins {
52 u32 perturbation; /* jhash perturbation */
53 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
54};
55
56struct 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;
71
72 u8 slot; /* current active bins (0 or 1) */
73 bool double_buffering;
74 struct sfb_bins bins[2];
75
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;
84};
85
86/*
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)
90 */
91struct sfb_skb_cb {
92 u32 hashes[2];
93};
94
95static inline struct sfb_skb_cb *sfb_skb_cb(const struct sk_buff *skb)
96{
97 qdisc_cb_private_validate(skb, sizeof(struct sfb_skb_cb));
98 return (struct sfb_skb_cb *)qdisc_skb_cb(skb)->data;
99}
100
101/*
102 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
103 * If using external classifier, hash comes from the classid.
104 */
105static u32 sfb_hash(const struct sk_buff *skb, u32 slot)
106{
107 return sfb_skb_cb(skb)->hashes[slot];
108}
109
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]
113 */
114static u32 prob_plus(u32 p1, u32 p2)
115{
116 u32 res = p1 + p2;
117
118 return min_t(u32, res, SFB_MAX_PROB);
119}
120
121static u32 prob_minus(u32 p1, u32 p2)
122{
123 return p1 > p2 ? p1 - p2 : 0;
124}
125
126static void increment_one_qlen(u32 sfbhash, u32 slot, struct sfb_sched_data *q)
127{
128 int i;
129 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
130
131 for (i = 0; i < SFB_LEVELS; i++) {
132 u32 hash = sfbhash & SFB_BUCKET_MASK;
133
134 sfbhash >>= SFB_BUCKET_SHIFT;
135 if (b[hash].qlen < 0xFFFF)
136 b[hash].qlen++;
137 b += SFB_NUMBUCKETS; /* next level */
138 }
139}
140
141static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
142{
143 u32 sfbhash;
144
145 sfbhash = sfb_hash(skb, 0);
146 if (sfbhash)
147 increment_one_qlen(sfbhash, 0, q);
148
149 sfbhash = sfb_hash(skb, 1);
150 if (sfbhash)
151 increment_one_qlen(sfbhash, 1, q);
152}
153
154static void decrement_one_qlen(u32 sfbhash, u32 slot,
155 struct sfb_sched_data *q)
156{
157 int i;
158 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
159
160 for (i = 0; i < SFB_LEVELS; i++) {
161 u32 hash = sfbhash & SFB_BUCKET_MASK;
162
163 sfbhash >>= SFB_BUCKET_SHIFT;
164 if (b[hash].qlen > 0)
165 b[hash].qlen--;
166 b += SFB_NUMBUCKETS; /* next level */
167 }
168}
169
170static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
171{
172 u32 sfbhash;
173
174 sfbhash = sfb_hash(skb, 0);
175 if (sfbhash)
176 decrement_one_qlen(sfbhash, 0, q);
177
178 sfbhash = sfb_hash(skb, 1);
179 if (sfbhash)
180 decrement_one_qlen(sfbhash, 1, q);
181}
182
183static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
184{
185 b->p_mark = prob_minus(b->p_mark, q->decrement);
186}
187
188static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
189{
190 b->p_mark = prob_plus(b->p_mark, q->increment);
191}
192
193static void sfb_zero_all_buckets(struct sfb_sched_data *q)
194{
195 memset(&q->bins, 0, sizeof(q->bins));
196}
197
198/*
199 * compute max qlen, max p_mark, and avg p_mark
200 */
201static u32 sfb_compute_qlen(u32 *prob_r, u32 *avgpm_r, const struct sfb_sched_data *q)
202{
203 int i;
204 u32 qlen = 0, prob = 0, totalpm = 0;
205 const struct sfb_bucket *b = &q->bins[q->slot].bins[0][0];
206
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++;
214 }
215 *prob_r = prob;
216 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
217 return qlen;
218}
219
220
221static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
222{
223 q->bins[slot].perturbation = net_random();
224}
225
226static void sfb_swap_slot(struct sfb_sched_data *q)
227{
228 sfb_init_perturbation(q->slot, q);
229 q->slot ^= 1;
230 q->double_buffering = false;
231}
232
233/* Non elastic flows are allowed to use part of the bandwidth, expressed
234 * in "penalty_rate" packets per second, with "penalty_burst" burst
235 */
236static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
237{
238 if (q->penalty_rate == 0 || q->penalty_burst == 0)
239 return true;
240
241 if (q->tokens_avail < 1) {
242 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
243
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;
250 }
251
252 q->tokens_avail--;
253 return false;
254}
255
256static bool sfb_classify(struct sk_buff *skb, struct sfb_sched_data *q,
257 int *qerr, u32 *salt)
258{
259 struct tcf_result res;
260 int result;
261
262 result = tc_classify(skb, q->filter_list, &res);
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;
271 }
272#endif
273 *salt = TC_H_MIN(res.classid);
274 return true;
275 }
276 return false;
277}
278
279static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
280{
281
282 struct sfb_sched_data *q = qdisc_priv(sch);
283 struct Qdisc *child = q->qdisc;
284 int i;
285 u32 p_min = ~0;
286 u32 minqlen = ~0;
287 u32 r, slot, salt, sfbhash;
288 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
289 struct flow_keys keys;
290
291 if (unlikely(sch->q.qlen >= q->limit)) {
292 sch->qstats.overlimits++;
293 q->stats.queuedrop++;
294 goto drop;
295 }
296
297 if (q->rehash_interval > 0) {
298 unsigned long limit = q->rehash_time + q->rehash_interval;
299
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;
306 }
307 }
308
309 if (q->filter_list) {
310 /* If using external classifiers, get result and record it. */
311 if (!sfb_classify(skb, q, &ret, &salt))
312 goto other_drop;
313 keys.src = salt;
314 keys.dst = 0;
315 keys.ports = 0;
316 } else {
317 skb_flow_dissect(skb, &keys);
318 }
319
320 slot = q->slot;
321
322 sfbhash = jhash_3words((__force u32)keys.dst,
323 (__force u32)keys.src,
324 (__force u32)keys.ports,
325 q->bins[slot].perturbation);
326 if (!sfbhash)
327 sfbhash = 1;
328 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
329
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];
333
334 sfbhash >>= SFB_BUCKET_SHIFT;
335 if (b->qlen == 0)
336 decrement_prob(b, q);
337 else if (b->qlen >= q->bin_size)
338 increment_prob(b, q);
339 if (minqlen > b->qlen)
340 minqlen = b->qlen;
341 if (p_min > b->p_mark)
342 p_min = b->p_mark;
343 }
344
345 slot ^= 1;
346 sfb_skb_cb(skb)->hashes[slot] = 0;
347
348 if (unlikely(minqlen >= q->max)) {
349 sch->qstats.overlimits++;
350 q->stats.bucketdrop++;
351 goto drop;
352 }
353
354 if (unlikely(p_min >= SFB_MAX_PROB)) {
355 /* Inelastic flow */
356 if (q->double_buffering) {
357 sfbhash = jhash_3words((__force u32)keys.dst,
358 (__force u32)keys.src,
359 (__force u32)keys.ports,
360 q->bins[slot].perturbation);
361 if (!sfbhash)
362 sfbhash = 1;
363 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
364
365 for (i = 0; i < SFB_LEVELS; i++) {
366 u32 hash = sfbhash & SFB_BUCKET_MASK;
367 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
368
369 sfbhash >>= SFB_BUCKET_SHIFT;
370 if (b->qlen == 0)
371 decrement_prob(b, q);
372 else if (b->qlen >= q->bin_size)
373 increment_prob(b, q);
374 }
375 }
376 if (sfb_rate_limit(skb, q)) {
377 sch->qstats.overlimits++;
378 q->stats.penaltydrop++;
379 goto drop;
380 }
381 goto enqueue;
382 }
383
384 r = net_random() & SFB_MAX_PROB;
385
386 if (unlikely(r < p_min)) {
387 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
388 /* If we're marking that many packets, then either
389 * this flow is unresponsive, or we're badly congested.
390 * In either case, we want to start dropping packets.
391 */
392 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
393 q->stats.earlydrop++;
394 goto drop;
395 }
396 }
397 if (INET_ECN_set_ce(skb)) {
398 q->stats.marked++;
399 } else {
400 q->stats.earlydrop++;
401 goto drop;
402 }
403 }
404
405enqueue:
406 ret = qdisc_enqueue(skb, child);
407 if (likely(ret == NET_XMIT_SUCCESS)) {
408 sch->q.qlen++;
409 increment_qlen(skb, q);
410 } else if (net_xmit_drop_count(ret)) {
411 q->stats.childdrop++;
412 sch->qstats.drops++;
413 }
414 return ret;
415
416drop:
417 qdisc_drop(skb, sch);
418 return NET_XMIT_CN;
419other_drop:
420 if (ret & __NET_XMIT_BYPASS)
421 sch->qstats.drops++;
422 kfree_skb(skb);
423 return ret;
424}
425
426static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
427{
428 struct sfb_sched_data *q = qdisc_priv(sch);
429 struct Qdisc *child = q->qdisc;
430 struct sk_buff *skb;
431
432 skb = child->dequeue(q->qdisc);
433
434 if (skb) {
435 qdisc_bstats_update(sch, skb);
436 sch->q.qlen--;
437 decrement_qlen(skb, q);
438 }
439
440 return skb;
441}
442
443static struct sk_buff *sfb_peek(struct Qdisc *sch)
444{
445 struct sfb_sched_data *q = qdisc_priv(sch);
446 struct Qdisc *child = q->qdisc;
447
448 return child->ops->peek(child);
449}
450
451/* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
452
453static void sfb_reset(struct Qdisc *sch)
454{
455 struct sfb_sched_data *q = qdisc_priv(sch);
456
457 qdisc_reset(q->qdisc);
458 sch->q.qlen = 0;
459 q->slot = 0;
460 q->double_buffering = false;
461 sfb_zero_all_buckets(q);
462 sfb_init_perturbation(0, q);
463}
464
465static void sfb_destroy(struct Qdisc *sch)
466{
467 struct sfb_sched_data *q = qdisc_priv(sch);
468
469 tcf_destroy_chain(&q->filter_list);
470 qdisc_destroy(q->qdisc);
471}
472
473static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
474 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
475};
476
477static const struct tc_sfb_qopt sfb_default_ops = {
478 .rehash_interval = 600 * MSEC_PER_SEC,
479 .warmup_time = 60 * MSEC_PER_SEC,
480 .limit = 0,
481 .max = 25,
482 .bin_size = 20,
483 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
484 .decrement = (SFB_MAX_PROB + 3000) / 6000,
485 .penalty_rate = 10,
486 .penalty_burst = 20,
487};
488
489static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
490{
491 struct sfb_sched_data *q = qdisc_priv(sch);
492 struct Qdisc *child;
493 struct nlattr *tb[TCA_SFB_MAX + 1];
494 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
495 u32 limit;
496 int err;
497
498 if (opt) {
499 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
500 if (err < 0)
501 return -EINVAL;
502
503 if (tb[TCA_SFB_PARMS] == NULL)
504 return -EINVAL;
505
506 ctl = nla_data(tb[TCA_SFB_PARMS]);
507 }
508
509 limit = ctl->limit;
510 if (limit == 0)
511 limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
512
513 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
514 if (IS_ERR(child))
515 return PTR_ERR(child);
516
517 sch_tree_lock(sch);
518
519 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
520 qdisc_destroy(q->qdisc);
521 q->qdisc = child;
522
523 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
524 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
525 q->rehash_time = jiffies;
526 q->limit = limit;
527 q->increment = ctl->increment;
528 q->decrement = ctl->decrement;
529 q->max = ctl->max;
530 q->bin_size = ctl->bin_size;
531 q->penalty_rate = ctl->penalty_rate;
532 q->penalty_burst = ctl->penalty_burst;
533 q->tokens_avail = ctl->penalty_burst;
534 q->token_time = jiffies;
535
536 q->slot = 0;
537 q->double_buffering = false;
538 sfb_zero_all_buckets(q);
539 sfb_init_perturbation(0, q);
540 sfb_init_perturbation(1, q);
541
542 sch_tree_unlock(sch);
543
544 return 0;
545}
546
547static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
548{
549 struct sfb_sched_data *q = qdisc_priv(sch);
550
551 q->qdisc = &noop_qdisc;
552 return sfb_change(sch, opt);
553}
554
555static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
556{
557 struct sfb_sched_data *q = qdisc_priv(sch);
558 struct nlattr *opts;
559 struct tc_sfb_qopt opt = {
560 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
561 .warmup_time = jiffies_to_msecs(q->warmup_time),
562 .limit = q->limit,
563 .max = q->max,
564 .bin_size = q->bin_size,
565 .increment = q->increment,
566 .decrement = q->decrement,
567 .penalty_rate = q->penalty_rate,
568 .penalty_burst = q->penalty_burst,
569 };
570
571 sch->qstats.backlog = q->qdisc->qstats.backlog;
572 opts = nla_nest_start(skb, TCA_OPTIONS);
573 if (opts == NULL)
574 goto nla_put_failure;
575 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
576 goto nla_put_failure;
577 return nla_nest_end(skb, opts);
578
579nla_put_failure:
580 nla_nest_cancel(skb, opts);
581 return -EMSGSIZE;
582}
583
584static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
585{
586 struct sfb_sched_data *q = qdisc_priv(sch);
587 struct tc_sfb_xstats st = {
588 .earlydrop = q->stats.earlydrop,
589 .penaltydrop = q->stats.penaltydrop,
590 .bucketdrop = q->stats.bucketdrop,
591 .queuedrop = q->stats.queuedrop,
592 .childdrop = q->stats.childdrop,
593 .marked = q->stats.marked,
594 };
595
596 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
597
598 return gnet_stats_copy_app(d, &st, sizeof(st));
599}
600
601static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
602 struct sk_buff *skb, struct tcmsg *tcm)
603{
604 return -ENOSYS;
605}
606
607static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
608 struct Qdisc **old)
609{
610 struct sfb_sched_data *q = qdisc_priv(sch);
611
612 if (new == NULL)
613 new = &noop_qdisc;
614
615 sch_tree_lock(sch);
616 *old = q->qdisc;
617 q->qdisc = new;
618 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
619 qdisc_reset(*old);
620 sch_tree_unlock(sch);
621 return 0;
622}
623
624static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
625{
626 struct sfb_sched_data *q = qdisc_priv(sch);
627
628 return q->qdisc;
629}
630
631static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
632{
633 return 1;
634}
635
636static void sfb_put(struct Qdisc *sch, unsigned long arg)
637{
638}
639
640static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
641 struct nlattr **tca, unsigned long *arg)
642{
643 return -ENOSYS;
644}
645
646static int sfb_delete(struct Qdisc *sch, unsigned long cl)
647{
648 return -ENOSYS;
649}
650
651static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
652{
653 if (!walker->stop) {
654 if (walker->count >= walker->skip)
655 if (walker->fn(sch, 1, walker) < 0) {
656 walker->stop = 1;
657 return;
658 }
659 walker->count++;
660 }
661}
662
663static struct tcf_proto **sfb_find_tcf(struct Qdisc *sch, unsigned long cl)
664{
665 struct sfb_sched_data *q = qdisc_priv(sch);
666
667 if (cl)
668 return NULL;
669 return &q->filter_list;
670}
671
672static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
673 u32 classid)
674{
675 return 0;
676}
677
678
679static const struct Qdisc_class_ops sfb_class_ops = {
680 .graft = sfb_graft,
681 .leaf = sfb_leaf,
682 .get = sfb_get,
683 .put = sfb_put,
684 .change = sfb_change_class,
685 .delete = sfb_delete,
686 .walk = sfb_walk,
687 .tcf_chain = sfb_find_tcf,
688 .bind_tcf = sfb_bind,
689 .unbind_tcf = sfb_put,
690 .dump = sfb_dump_class,
691};
692
693static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
694 .id = "sfb",
695 .priv_size = sizeof(struct sfb_sched_data),
696 .cl_ops = &sfb_class_ops,
697 .enqueue = sfb_enqueue,
698 .dequeue = sfb_dequeue,
699 .peek = sfb_peek,
700 .init = sfb_init,
701 .reset = sfb_reset,
702 .destroy = sfb_destroy,
703 .change = sfb_change,
704 .dump = sfb_dump,
705 .dump_stats = sfb_dump_stats,
706 .owner = THIS_MODULE,
707};
708
709static int __init sfb_module_init(void)
710{
711 return register_qdisc(&sfb_qdisc_ops);
712}
713
714static void __exit sfb_module_exit(void)
715{
716 unregister_qdisc(&sfb_qdisc_ops);
717}
718
719module_init(sfb_module_init)
720module_exit(sfb_module_exit)
721
722MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
723MODULE_AUTHOR("Juliusz Chroboczek");
724MODULE_AUTHOR("Eric Dumazet");
725MODULE_LICENSE("GPL");
1/*
2 * net/sched/sch_sfb.c Stochastic Fair Blue
3 *
4 * Copyright (c) 2008-2011 Juliusz Chroboczek <jch@pps.jussieu.fr>
5 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
6 *
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.
10 *
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.
14 *
15 * http://www.thefengs.com/wuchang/blue/CSE-TR-387-99.pdf
16 *
17 */
18
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
30/*
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.
35 */
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 */
40
41/* SFB algo uses a virtual queue, named "bin" */
42struct sfb_bucket {
43 u16 qlen; /* length of virtual queue */
44 u16 p_mark; /* marking probability */
45};
46
47/* We use a double buffering right before hash change
48 * (Section 4.4 of SFB reference : moving hash functions)
49 */
50struct sfb_bins {
51 u32 perturbation; /* jhash perturbation */
52 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
53};
54
55struct 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;
70
71 u8 slot; /* current active bins (0 or 1) */
72 bool double_buffering;
73 struct sfb_bins bins[2];
74
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;
83};
84
85/*
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)
89 */
90struct sfb_skb_cb {
91 u32 hashes[2];
92};
93
94static inline struct sfb_skb_cb *sfb_skb_cb(const struct sk_buff *skb)
95{
96 qdisc_cb_private_validate(skb, sizeof(struct sfb_skb_cb));
97 return (struct sfb_skb_cb *)qdisc_skb_cb(skb)->data;
98}
99
100/*
101 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
102 * If using external classifier, hash comes from the classid.
103 */
104static u32 sfb_hash(const struct sk_buff *skb, u32 slot)
105{
106 return sfb_skb_cb(skb)->hashes[slot];
107}
108
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]
112 */
113static u32 prob_plus(u32 p1, u32 p2)
114{
115 u32 res = p1 + p2;
116
117 return min_t(u32, res, SFB_MAX_PROB);
118}
119
120static u32 prob_minus(u32 p1, u32 p2)
121{
122 return p1 > p2 ? p1 - p2 : 0;
123}
124
125static void increment_one_qlen(u32 sfbhash, u32 slot, struct sfb_sched_data *q)
126{
127 int i;
128 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
129
130 for (i = 0; i < SFB_LEVELS; i++) {
131 u32 hash = sfbhash & SFB_BUCKET_MASK;
132
133 sfbhash >>= SFB_BUCKET_SHIFT;
134 if (b[hash].qlen < 0xFFFF)
135 b[hash].qlen++;
136 b += SFB_NUMBUCKETS; /* next level */
137 }
138}
139
140static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
141{
142 u32 sfbhash;
143
144 sfbhash = sfb_hash(skb, 0);
145 if (sfbhash)
146 increment_one_qlen(sfbhash, 0, q);
147
148 sfbhash = sfb_hash(skb, 1);
149 if (sfbhash)
150 increment_one_qlen(sfbhash, 1, q);
151}
152
153static void decrement_one_qlen(u32 sfbhash, u32 slot,
154 struct sfb_sched_data *q)
155{
156 int i;
157 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
158
159 for (i = 0; i < SFB_LEVELS; i++) {
160 u32 hash = sfbhash & SFB_BUCKET_MASK;
161
162 sfbhash >>= SFB_BUCKET_SHIFT;
163 if (b[hash].qlen > 0)
164 b[hash].qlen--;
165 b += SFB_NUMBUCKETS; /* next level */
166 }
167}
168
169static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
170{
171 u32 sfbhash;
172
173 sfbhash = sfb_hash(skb, 0);
174 if (sfbhash)
175 decrement_one_qlen(sfbhash, 0, q);
176
177 sfbhash = sfb_hash(skb, 1);
178 if (sfbhash)
179 decrement_one_qlen(sfbhash, 1, q);
180}
181
182static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
183{
184 b->p_mark = prob_minus(b->p_mark, q->decrement);
185}
186
187static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
188{
189 b->p_mark = prob_plus(b->p_mark, q->increment);
190}
191
192static void sfb_zero_all_buckets(struct sfb_sched_data *q)
193{
194 memset(&q->bins, 0, sizeof(q->bins));
195}
196
197/*
198 * compute max qlen, max p_mark, and avg p_mark
199 */
200static u32 sfb_compute_qlen(u32 *prob_r, u32 *avgpm_r, const struct sfb_sched_data *q)
201{
202 int i;
203 u32 qlen = 0, prob = 0, totalpm = 0;
204 const struct sfb_bucket *b = &q->bins[q->slot].bins[0][0];
205
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++;
213 }
214 *prob_r = prob;
215 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
216 return qlen;
217}
218
219
220static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
221{
222 q->bins[slot].perturbation = prandom_u32();
223}
224
225static void sfb_swap_slot(struct sfb_sched_data *q)
226{
227 sfb_init_perturbation(q->slot, q);
228 q->slot ^= 1;
229 q->double_buffering = false;
230}
231
232/* Non elastic flows are allowed to use part of the bandwidth, expressed
233 * in "penalty_rate" packets per second, with "penalty_burst" burst
234 */
235static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
236{
237 if (q->penalty_rate == 0 || q->penalty_burst == 0)
238 return true;
239
240 if (q->tokens_avail < 1) {
241 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
242
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;
249 }
250
251 q->tokens_avail--;
252 return false;
253}
254
255static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
256 int *qerr, u32 *salt)
257{
258 struct tcf_result res;
259 int result;
260
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;
270 }
271#endif
272 *salt = TC_H_MIN(res.classid);
273 return true;
274 }
275 return false;
276}
277
278static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
279{
280
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;
290
291 if (unlikely(sch->q.qlen >= q->limit)) {
292 qdisc_qstats_overlimit(sch);
293 q->stats.queuedrop++;
294 goto drop;
295 }
296
297 if (q->rehash_interval > 0) {
298 unsigned long limit = q->rehash_time + q->rehash_interval;
299
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;
306 }
307 }
308
309 fl = rcu_dereference_bh(q->filter_list);
310 if (fl) {
311 u32 salt;
312
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);
319 }
320
321
322 if (!sfbhash)
323 sfbhash = 1;
324 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
325
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];
329
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;
339 }
340
341 slot ^= 1;
342 sfb_skb_cb(skb)->hashes[slot] = 0;
343
344 if (unlikely(minqlen >= q->max)) {
345 qdisc_qstats_overlimit(sch);
346 q->stats.bucketdrop++;
347 goto drop;
348 }
349
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;
358
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];
362
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);
368 }
369 }
370 if (sfb_rate_limit(skb, q)) {
371 qdisc_qstats_overlimit(sch);
372 q->stats.penaltydrop++;
373 goto drop;
374 }
375 goto enqueue;
376 }
377
378 r = prandom_u32() & SFB_MAX_PROB;
379
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.
385 */
386 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
387 q->stats.earlydrop++;
388 goto drop;
389 }
390 }
391 if (INET_ECN_set_ce(skb)) {
392 q->stats.marked++;
393 } else {
394 q->stats.earlydrop++;
395 goto drop;
396 }
397 }
398
399enqueue:
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);
407 }
408 return ret;
409
410drop:
411 qdisc_drop(skb, sch);
412 return NET_XMIT_CN;
413other_drop:
414 if (ret & __NET_XMIT_BYPASS)
415 qdisc_qstats_drop(sch);
416 kfree_skb(skb);
417 return ret;
418}
419
420static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
421{
422 struct sfb_sched_data *q = qdisc_priv(sch);
423 struct Qdisc *child = q->qdisc;
424 struct sk_buff *skb;
425
426 skb = child->dequeue(q->qdisc);
427
428 if (skb) {
429 qdisc_bstats_update(sch, skb);
430 sch->q.qlen--;
431 decrement_qlen(skb, q);
432 }
433
434 return skb;
435}
436
437static struct sk_buff *sfb_peek(struct Qdisc *sch)
438{
439 struct sfb_sched_data *q = qdisc_priv(sch);
440 struct Qdisc *child = q->qdisc;
441
442 return child->ops->peek(child);
443}
444
445/* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
446
447static void sfb_reset(struct Qdisc *sch)
448{
449 struct sfb_sched_data *q = qdisc_priv(sch);
450
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);
457}
458
459static void sfb_destroy(struct Qdisc *sch)
460{
461 struct sfb_sched_data *q = qdisc_priv(sch);
462
463 tcf_destroy_chain(&q->filter_list);
464 qdisc_destroy(q->qdisc);
465}
466
467static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
468 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
469};
470
471static 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,
481};
482
483static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
484{
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;
491
492 if (opt) {
493 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
494 if (err < 0)
495 return -EINVAL;
496
497 if (tb[TCA_SFB_PARMS] == NULL)
498 return -EINVAL;
499
500 ctl = nla_data(tb[TCA_SFB_PARMS]);
501 }
502
503 limit = ctl->limit;
504 if (limit == 0)
505 limit = qdisc_dev(sch)->tx_queue_len;
506
507 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
508 if (IS_ERR(child))
509 return PTR_ERR(child);
510
511 sch_tree_lock(sch);
512
513 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
514 q->qdisc->qstats.backlog);
515 qdisc_destroy(q->qdisc);
516 q->qdisc = child;
517
518 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
519 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
520 q->rehash_time = jiffies;
521 q->limit = limit;
522 q->increment = ctl->increment;
523 q->decrement = ctl->decrement;
524 q->max = ctl->max;
525 q->bin_size = ctl->bin_size;
526 q->penalty_rate = ctl->penalty_rate;
527 q->penalty_burst = ctl->penalty_burst;
528 q->tokens_avail = ctl->penalty_burst;
529 q->token_time = jiffies;
530
531 q->slot = 0;
532 q->double_buffering = false;
533 sfb_zero_all_buckets(q);
534 sfb_init_perturbation(0, q);
535 sfb_init_perturbation(1, q);
536
537 sch_tree_unlock(sch);
538
539 return 0;
540}
541
542static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
543{
544 struct sfb_sched_data *q = qdisc_priv(sch);
545
546 q->qdisc = &noop_qdisc;
547 return sfb_change(sch, opt);
548}
549
550static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
551{
552 struct sfb_sched_data *q = qdisc_priv(sch);
553 struct nlattr *opts;
554 struct tc_sfb_qopt opt = {
555 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
556 .warmup_time = jiffies_to_msecs(q->warmup_time),
557 .limit = q->limit,
558 .max = q->max,
559 .bin_size = q->bin_size,
560 .increment = q->increment,
561 .decrement = q->decrement,
562 .penalty_rate = q->penalty_rate,
563 .penalty_burst = q->penalty_burst,
564 };
565
566 sch->qstats.backlog = q->qdisc->qstats.backlog;
567 opts = nla_nest_start(skb, TCA_OPTIONS);
568 if (opts == NULL)
569 goto nla_put_failure;
570 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
571 goto nla_put_failure;
572 return nla_nest_end(skb, opts);
573
574nla_put_failure:
575 nla_nest_cancel(skb, opts);
576 return -EMSGSIZE;
577}
578
579static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
580{
581 struct sfb_sched_data *q = qdisc_priv(sch);
582 struct tc_sfb_xstats st = {
583 .earlydrop = q->stats.earlydrop,
584 .penaltydrop = q->stats.penaltydrop,
585 .bucketdrop = q->stats.bucketdrop,
586 .queuedrop = q->stats.queuedrop,
587 .childdrop = q->stats.childdrop,
588 .marked = q->stats.marked,
589 };
590
591 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
592
593 return gnet_stats_copy_app(d, &st, sizeof(st));
594}
595
596static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
597 struct sk_buff *skb, struct tcmsg *tcm)
598{
599 return -ENOSYS;
600}
601
602static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
603 struct Qdisc **old)
604{
605 struct sfb_sched_data *q = qdisc_priv(sch);
606
607 if (new == NULL)
608 new = &noop_qdisc;
609
610 *old = qdisc_replace(sch, new, &q->qdisc);
611 return 0;
612}
613
614static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
615{
616 struct sfb_sched_data *q = qdisc_priv(sch);
617
618 return q->qdisc;
619}
620
621static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
622{
623 return 1;
624}
625
626static void sfb_put(struct Qdisc *sch, unsigned long arg)
627{
628}
629
630static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
631 struct nlattr **tca, unsigned long *arg)
632{
633 return -ENOSYS;
634}
635
636static int sfb_delete(struct Qdisc *sch, unsigned long cl)
637{
638 return -ENOSYS;
639}
640
641static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
642{
643 if (!walker->stop) {
644 if (walker->count >= walker->skip)
645 if (walker->fn(sch, 1, walker) < 0) {
646 walker->stop = 1;
647 return;
648 }
649 walker->count++;
650 }
651}
652
653static struct tcf_proto __rcu **sfb_find_tcf(struct Qdisc *sch,
654 unsigned long cl)
655{
656 struct sfb_sched_data *q = qdisc_priv(sch);
657
658 if (cl)
659 return NULL;
660 return &q->filter_list;
661}
662
663static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
664 u32 classid)
665{
666 return 0;
667}
668
669
670static const struct Qdisc_class_ops sfb_class_ops = {
671 .graft = sfb_graft,
672 .leaf = sfb_leaf,
673 .get = sfb_get,
674 .put = sfb_put,
675 .change = sfb_change_class,
676 .delete = sfb_delete,
677 .walk = sfb_walk,
678 .tcf_chain = sfb_find_tcf,
679 .bind_tcf = sfb_bind,
680 .unbind_tcf = sfb_put,
681 .dump = sfb_dump_class,
682};
683
684static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
685 .id = "sfb",
686 .priv_size = sizeof(struct sfb_sched_data),
687 .cl_ops = &sfb_class_ops,
688 .enqueue = sfb_enqueue,
689 .dequeue = sfb_dequeue,
690 .peek = sfb_peek,
691 .init = sfb_init,
692 .reset = sfb_reset,
693 .destroy = sfb_destroy,
694 .change = sfb_change,
695 .dump = sfb_dump,
696 .dump_stats = sfb_dump_stats,
697 .owner = THIS_MODULE,
698};
699
700static int __init sfb_module_init(void)
701{
702 return register_qdisc(&sfb_qdisc_ops);
703}
704
705static void __exit sfb_module_exit(void)
706{
707 unregister_qdisc(&sfb_qdisc_ops);
708}
709
710module_init(sfb_module_init)
711module_exit(sfb_module_exit)
712
713MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
714MODULE_AUTHOR("Juliusz Chroboczek");
715MODULE_AUTHOR("Eric Dumazet");
716MODULE_LICENSE("GPL");