Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_sfb.c Stochastic Fair Blue
4 *
5 * Copyright (c) 2008-2011 Juliusz Chroboczek <jch@pps.jussieu.fr>
6 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
7 *
8 * W. Feng, D. Kandlur, D. Saha, K. Shin. Blue:
9 * A New Class of Active Queue Management Algorithms.
10 * U. Michigan CSE-TR-387-99, April 1999.
11 *
12 * http://www.thefengs.com/wuchang/blue/CSE-TR-387-99.pdf
13 */
14
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/errno.h>
19#include <linux/skbuff.h>
20#include <linux/random.h>
21#include <linux/siphash.h>
22#include <net/ip.h>
23#include <net/pkt_sched.h>
24#include <net/pkt_cls.h>
25#include <net/inet_ecn.h>
26
27/*
28 * SFB uses two B[l][n] : L x N arrays of bins (L levels, N bins per level)
29 * This implementation uses L = 8 and N = 16
30 * This permits us to split one 32bit hash (provided per packet by rxhash or
31 * external classifier) into 8 subhashes of 4 bits.
32 */
33#define SFB_BUCKET_SHIFT 4
34#define SFB_NUMBUCKETS (1 << SFB_BUCKET_SHIFT) /* N bins per Level */
35#define SFB_BUCKET_MASK (SFB_NUMBUCKETS - 1)
36#define SFB_LEVELS (32 / SFB_BUCKET_SHIFT) /* L */
37
38/* SFB algo uses a virtual queue, named "bin" */
39struct sfb_bucket {
40 u16 qlen; /* length of virtual queue */
41 u16 p_mark; /* marking probability */
42};
43
44/* We use a double buffering right before hash change
45 * (Section 4.4 of SFB reference : moving hash functions)
46 */
47struct sfb_bins {
48 siphash_key_t perturbation; /* siphash key */
49 struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
50};
51
52struct sfb_sched_data {
53 struct Qdisc *qdisc;
54 struct tcf_proto __rcu *filter_list;
55 struct tcf_block *block;
56 unsigned long rehash_interval;
57 unsigned long warmup_time; /* double buffering warmup time in jiffies */
58 u32 max;
59 u32 bin_size; /* maximum queue length per bin */
60 u32 increment; /* d1 */
61 u32 decrement; /* d2 */
62 u32 limit; /* HARD maximal queue length */
63 u32 penalty_rate;
64 u32 penalty_burst;
65 u32 tokens_avail;
66 unsigned long rehash_time;
67 unsigned long token_time;
68
69 u8 slot; /* current active bins (0 or 1) */
70 bool double_buffering;
71 struct sfb_bins bins[2];
72
73 struct {
74 u32 earlydrop;
75 u32 penaltydrop;
76 u32 bucketdrop;
77 u32 queuedrop;
78 u32 childdrop; /* drops in child qdisc */
79 u32 marked; /* ECN mark */
80 } stats;
81};
82
83/*
84 * Each queued skb might be hashed on one or two bins
85 * We store in skb_cb the two hash values.
86 * (A zero value means double buffering was not used)
87 */
88struct sfb_skb_cb {
89 u32 hashes[2];
90};
91
92static inline struct sfb_skb_cb *sfb_skb_cb(const struct sk_buff *skb)
93{
94 qdisc_cb_private_validate(skb, sizeof(struct sfb_skb_cb));
95 return (struct sfb_skb_cb *)qdisc_skb_cb(skb)->data;
96}
97
98/*
99 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
100 * If using external classifier, hash comes from the classid.
101 */
102static u32 sfb_hash(const struct sk_buff *skb, u32 slot)
103{
104 return sfb_skb_cb(skb)->hashes[slot];
105}
106
107/* Probabilities are coded as Q0.16 fixed-point values,
108 * with 0xFFFF representing 65535/65536 (almost 1.0)
109 * Addition and subtraction are saturating in [0, 65535]
110 */
111static u32 prob_plus(u32 p1, u32 p2)
112{
113 u32 res = p1 + p2;
114
115 return min_t(u32, res, SFB_MAX_PROB);
116}
117
118static u32 prob_minus(u32 p1, u32 p2)
119{
120 return p1 > p2 ? p1 - p2 : 0;
121}
122
123static void increment_one_qlen(u32 sfbhash, u32 slot, struct sfb_sched_data *q)
124{
125 int i;
126 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
127
128 for (i = 0; i < SFB_LEVELS; i++) {
129 u32 hash = sfbhash & SFB_BUCKET_MASK;
130
131 sfbhash >>= SFB_BUCKET_SHIFT;
132 if (b[hash].qlen < 0xFFFF)
133 b[hash].qlen++;
134 b += SFB_NUMBUCKETS; /* next level */
135 }
136}
137
138static void increment_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
139{
140 u32 sfbhash;
141
142 sfbhash = sfb_hash(skb, 0);
143 if (sfbhash)
144 increment_one_qlen(sfbhash, 0, q);
145
146 sfbhash = sfb_hash(skb, 1);
147 if (sfbhash)
148 increment_one_qlen(sfbhash, 1, q);
149}
150
151static void decrement_one_qlen(u32 sfbhash, u32 slot,
152 struct sfb_sched_data *q)
153{
154 int i;
155 struct sfb_bucket *b = &q->bins[slot].bins[0][0];
156
157 for (i = 0; i < SFB_LEVELS; i++) {
158 u32 hash = sfbhash & SFB_BUCKET_MASK;
159
160 sfbhash >>= SFB_BUCKET_SHIFT;
161 if (b[hash].qlen > 0)
162 b[hash].qlen--;
163 b += SFB_NUMBUCKETS; /* next level */
164 }
165}
166
167static void decrement_qlen(const struct sk_buff *skb, struct sfb_sched_data *q)
168{
169 u32 sfbhash;
170
171 sfbhash = sfb_hash(skb, 0);
172 if (sfbhash)
173 decrement_one_qlen(sfbhash, 0, q);
174
175 sfbhash = sfb_hash(skb, 1);
176 if (sfbhash)
177 decrement_one_qlen(sfbhash, 1, q);
178}
179
180static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
181{
182 b->p_mark = prob_minus(b->p_mark, q->decrement);
183}
184
185static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
186{
187 b->p_mark = prob_plus(b->p_mark, q->increment);
188}
189
190static void sfb_zero_all_buckets(struct sfb_sched_data *q)
191{
192 memset(&q->bins, 0, sizeof(q->bins));
193}
194
195/*
196 * compute max qlen, max p_mark, and avg p_mark
197 */
198static u32 sfb_compute_qlen(u32 *prob_r, u32 *avgpm_r, const struct sfb_sched_data *q)
199{
200 int i;
201 u32 qlen = 0, prob = 0, totalpm = 0;
202 const struct sfb_bucket *b = &q->bins[q->slot].bins[0][0];
203
204 for (i = 0; i < SFB_LEVELS * SFB_NUMBUCKETS; i++) {
205 if (qlen < b->qlen)
206 qlen = b->qlen;
207 totalpm += b->p_mark;
208 if (prob < b->p_mark)
209 prob = b->p_mark;
210 b++;
211 }
212 *prob_r = prob;
213 *avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
214 return qlen;
215}
216
217
218static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
219{
220 get_random_bytes(&q->bins[slot].perturbation,
221 sizeof(q->bins[slot].perturbation));
222}
223
224static void sfb_swap_slot(struct sfb_sched_data *q)
225{
226 sfb_init_perturbation(q->slot, q);
227 q->slot ^= 1;
228 q->double_buffering = false;
229}
230
231/* Non elastic flows are allowed to use part of the bandwidth, expressed
232 * in "penalty_rate" packets per second, with "penalty_burst" burst
233 */
234static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
235{
236 if (q->penalty_rate == 0 || q->penalty_burst == 0)
237 return true;
238
239 if (q->tokens_avail < 1) {
240 unsigned long age = min(10UL * HZ, jiffies - q->token_time);
241
242 q->tokens_avail = (age * q->penalty_rate) / HZ;
243 if (q->tokens_avail > q->penalty_burst)
244 q->tokens_avail = q->penalty_burst;
245 q->token_time = jiffies;
246 if (q->tokens_avail < 1)
247 return true;
248 }
249
250 q->tokens_avail--;
251 return false;
252}
253
254static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
255 int *qerr, u32 *salt)
256{
257 struct tcf_result res;
258 int result;
259
260 result = tcf_classify(skb, fl, &res, false);
261 if (result >= 0) {
262#ifdef CONFIG_NET_CLS_ACT
263 switch (result) {
264 case TC_ACT_STOLEN:
265 case TC_ACT_QUEUED:
266 case TC_ACT_TRAP:
267 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
268 /* fall through */
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 struct sk_buff **to_free)
281{
282
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;
292
293 if (unlikely(sch->q.qlen >= q->limit)) {
294 qdisc_qstats_overlimit(sch);
295 q->stats.queuedrop++;
296 goto drop;
297 }
298
299 if (q->rehash_interval > 0) {
300 unsigned long limit = q->rehash_time + q->rehash_interval;
301
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;
308 }
309 }
310
311 fl = rcu_dereference_bh(q->filter_list);
312 if (fl) {
313 u32 salt;
314
315 /* If using external classifiers, get result and record it. */
316 if (!sfb_classify(skb, fl, &ret, &salt))
317 goto other_drop;
318 sfbhash = siphash_1u32(salt, &q->bins[slot].perturbation);
319 } else {
320 sfbhash = skb_get_hash_perturb(skb, &q->bins[slot].perturbation);
321 }
322
323
324 if (!sfbhash)
325 sfbhash = 1;
326 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
327
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];
331
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;
341 }
342
343 slot ^= 1;
344 sfb_skb_cb(skb)->hashes[slot] = 0;
345
346 if (unlikely(minqlen >= q->max)) {
347 qdisc_qstats_overlimit(sch);
348 q->stats.bucketdrop++;
349 goto drop;
350 }
351
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;
360
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];
364
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);
370 }
371 }
372 if (sfb_rate_limit(skb, q)) {
373 qdisc_qstats_overlimit(sch);
374 q->stats.penaltydrop++;
375 goto drop;
376 }
377 goto enqueue;
378 }
379
380 r = prandom_u32() & SFB_MAX_PROB;
381
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.
387 */
388 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
389 q->stats.earlydrop++;
390 goto drop;
391 }
392 }
393 if (INET_ECN_set_ce(skb)) {
394 q->stats.marked++;
395 } else {
396 q->stats.earlydrop++;
397 goto drop;
398 }
399 }
400
401enqueue:
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);
410 }
411 return ret;
412
413drop:
414 qdisc_drop(skb, sch, to_free);
415 return NET_XMIT_CN;
416other_drop:
417 if (ret & __NET_XMIT_BYPASS)
418 qdisc_qstats_drop(sch);
419 kfree_skb(skb);
420 return ret;
421}
422
423static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
424{
425 struct sfb_sched_data *q = qdisc_priv(sch);
426 struct Qdisc *child = q->qdisc;
427 struct sk_buff *skb;
428
429 skb = child->dequeue(q->qdisc);
430
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);
436 }
437
438 return skb;
439}
440
441static struct sk_buff *sfb_peek(struct Qdisc *sch)
442{
443 struct sfb_sched_data *q = qdisc_priv(sch);
444 struct Qdisc *child = q->qdisc;
445
446 return child->ops->peek(child);
447}
448
449/* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
450
451static void sfb_reset(struct Qdisc *sch)
452{
453 struct sfb_sched_data *q = qdisc_priv(sch);
454
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);
462}
463
464static void sfb_destroy(struct Qdisc *sch)
465{
466 struct sfb_sched_data *q = qdisc_priv(sch);
467
468 tcf_block_put(q->block);
469 qdisc_put(q->qdisc);
470}
471
472static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
473 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
474};
475
476static 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,
486};
487
488static int sfb_change(struct Qdisc *sch, struct nlattr *opt,
489 struct netlink_ext_ack *extack)
490{
491 struct sfb_sched_data *q = qdisc_priv(sch);
492 struct Qdisc *child, *old;
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_deprecated(tb, TCA_SFB_MAX, opt,
500 sfb_policy, NULL);
501 if (err < 0)
502 return -EINVAL;
503
504 if (tb[TCA_SFB_PARMS] == NULL)
505 return -EINVAL;
506
507 ctl = nla_data(tb[TCA_SFB_PARMS]);
508 }
509
510 limit = ctl->limit;
511 if (limit == 0)
512 limit = qdisc_dev(sch)->tx_queue_len;
513
514 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit, extack);
515 if (IS_ERR(child))
516 return PTR_ERR(child);
517
518 if (child != &noop_qdisc)
519 qdisc_hash_add(child, true);
520 sch_tree_lock(sch);
521
522 qdisc_purge_queue(q->qdisc);
523 old = q->qdisc;
524 q->qdisc = child;
525
526 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
527 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
528 q->rehash_time = jiffies;
529 q->limit = limit;
530 q->increment = ctl->increment;
531 q->decrement = ctl->decrement;
532 q->max = ctl->max;
533 q->bin_size = ctl->bin_size;
534 q->penalty_rate = ctl->penalty_rate;
535 q->penalty_burst = ctl->penalty_burst;
536 q->tokens_avail = ctl->penalty_burst;
537 q->token_time = jiffies;
538
539 q->slot = 0;
540 q->double_buffering = false;
541 sfb_zero_all_buckets(q);
542 sfb_init_perturbation(0, q);
543 sfb_init_perturbation(1, q);
544
545 sch_tree_unlock(sch);
546 qdisc_put(old);
547
548 return 0;
549}
550
551static int sfb_init(struct Qdisc *sch, struct nlattr *opt,
552 struct netlink_ext_ack *extack)
553{
554 struct sfb_sched_data *q = qdisc_priv(sch);
555 int err;
556
557 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
558 if (err)
559 return err;
560
561 q->qdisc = &noop_qdisc;
562 return sfb_change(sch, opt, extack);
563}
564
565static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
566{
567 struct sfb_sched_data *q = qdisc_priv(sch);
568 struct nlattr *opts;
569 struct tc_sfb_qopt opt = {
570 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
571 .warmup_time = jiffies_to_msecs(q->warmup_time),
572 .limit = q->limit,
573 .max = q->max,
574 .bin_size = q->bin_size,
575 .increment = q->increment,
576 .decrement = q->decrement,
577 .penalty_rate = q->penalty_rate,
578 .penalty_burst = q->penalty_burst,
579 };
580
581 sch->qstats.backlog = q->qdisc->qstats.backlog;
582 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
583 if (opts == NULL)
584 goto nla_put_failure;
585 if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
586 goto nla_put_failure;
587 return nla_nest_end(skb, opts);
588
589nla_put_failure:
590 nla_nest_cancel(skb, opts);
591 return -EMSGSIZE;
592}
593
594static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
595{
596 struct sfb_sched_data *q = qdisc_priv(sch);
597 struct tc_sfb_xstats st = {
598 .earlydrop = q->stats.earlydrop,
599 .penaltydrop = q->stats.penaltydrop,
600 .bucketdrop = q->stats.bucketdrop,
601 .queuedrop = q->stats.queuedrop,
602 .childdrop = q->stats.childdrop,
603 .marked = q->stats.marked,
604 };
605
606 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
607
608 return gnet_stats_copy_app(d, &st, sizeof(st));
609}
610
611static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
612 struct sk_buff *skb, struct tcmsg *tcm)
613{
614 return -ENOSYS;
615}
616
617static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
618 struct Qdisc **old, struct netlink_ext_ack *extack)
619{
620 struct sfb_sched_data *q = qdisc_priv(sch);
621
622 if (new == NULL)
623 new = &noop_qdisc;
624
625 *old = qdisc_replace(sch, new, &q->qdisc);
626 return 0;
627}
628
629static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
630{
631 struct sfb_sched_data *q = qdisc_priv(sch);
632
633 return q->qdisc;
634}
635
636static unsigned long sfb_find(struct Qdisc *sch, u32 classid)
637{
638 return 1;
639}
640
641static void sfb_unbind(struct Qdisc *sch, unsigned long arg)
642{
643}
644
645static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
646 struct nlattr **tca, unsigned long *arg,
647 struct netlink_ext_ack *extack)
648{
649 return -ENOSYS;
650}
651
652static int sfb_delete(struct Qdisc *sch, unsigned long cl)
653{
654 return -ENOSYS;
655}
656
657static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
658{
659 if (!walker->stop) {
660 if (walker->count >= walker->skip)
661 if (walker->fn(sch, 1, walker) < 0) {
662 walker->stop = 1;
663 return;
664 }
665 walker->count++;
666 }
667}
668
669static struct tcf_block *sfb_tcf_block(struct Qdisc *sch, unsigned long cl,
670 struct netlink_ext_ack *extack)
671{
672 struct sfb_sched_data *q = qdisc_priv(sch);
673
674 if (cl)
675 return NULL;
676 return q->block;
677}
678
679static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
680 u32 classid)
681{
682 return 0;
683}
684
685
686static const struct Qdisc_class_ops sfb_class_ops = {
687 .graft = sfb_graft,
688 .leaf = sfb_leaf,
689 .find = sfb_find,
690 .change = sfb_change_class,
691 .delete = sfb_delete,
692 .walk = sfb_walk,
693 .tcf_block = sfb_tcf_block,
694 .bind_tcf = sfb_bind,
695 .unbind_tcf = sfb_unbind,
696 .dump = sfb_dump_class,
697};
698
699static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
700 .id = "sfb",
701 .priv_size = sizeof(struct sfb_sched_data),
702 .cl_ops = &sfb_class_ops,
703 .enqueue = sfb_enqueue,
704 .dequeue = sfb_dequeue,
705 .peek = sfb_peek,
706 .init = sfb_init,
707 .reset = sfb_reset,
708 .destroy = sfb_destroy,
709 .change = sfb_change,
710 .dump = sfb_dump,
711 .dump_stats = sfb_dump_stats,
712 .owner = THIS_MODULE,
713};
714
715static int __init sfb_module_init(void)
716{
717 return register_qdisc(&sfb_qdisc_ops);
718}
719
720static void __exit sfb_module_exit(void)
721{
722 unregister_qdisc(&sfb_qdisc_ops);
723}
724
725module_init(sfb_module_init)
726module_exit(sfb_module_exit)
727
728MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
729MODULE_AUTHOR("Juliusz Chroboczek");
730MODULE_AUTHOR("Eric Dumazet");
731MODULE_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 *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 BUILD_BUG_ON(sizeof(skb->cb) <
97 sizeof(struct qdisc_skb_cb) + 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
290 if (q->rehash_interval > 0) {
291 unsigned long limit = q->rehash_time + q->rehash_interval;
292
293 if (unlikely(time_after(jiffies, limit))) {
294 sfb_swap_slot(q);
295 q->rehash_time = jiffies;
296 } else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
297 time_after(jiffies, limit - q->warmup_time))) {
298 q->double_buffering = true;
299 }
300 }
301
302 if (q->filter_list) {
303 /* If using external classifiers, get result and record it. */
304 if (!sfb_classify(skb, q, &ret, &salt))
305 goto other_drop;
306 } else {
307 salt = skb_get_rxhash(skb);
308 }
309
310 slot = q->slot;
311
312 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
313 if (!sfbhash)
314 sfbhash = 1;
315 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
316
317 for (i = 0; i < SFB_LEVELS; i++) {
318 u32 hash = sfbhash & SFB_BUCKET_MASK;
319 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
320
321 sfbhash >>= SFB_BUCKET_SHIFT;
322 if (b->qlen == 0)
323 decrement_prob(b, q);
324 else if (b->qlen >= q->bin_size)
325 increment_prob(b, q);
326 if (minqlen > b->qlen)
327 minqlen = b->qlen;
328 if (p_min > b->p_mark)
329 p_min = b->p_mark;
330 }
331
332 slot ^= 1;
333 sfb_skb_cb(skb)->hashes[slot] = 0;
334
335 if (unlikely(minqlen >= q->max || sch->q.qlen >= q->limit)) {
336 sch->qstats.overlimits++;
337 if (minqlen >= q->max)
338 q->stats.bucketdrop++;
339 else
340 q->stats.queuedrop++;
341 goto drop;
342 }
343
344 if (unlikely(p_min >= SFB_MAX_PROB)) {
345 /* Inelastic flow */
346 if (q->double_buffering) {
347 sfbhash = jhash_1word(salt, q->bins[slot].perturbation);
348 if (!sfbhash)
349 sfbhash = 1;
350 sfb_skb_cb(skb)->hashes[slot] = sfbhash;
351
352 for (i = 0; i < SFB_LEVELS; i++) {
353 u32 hash = sfbhash & SFB_BUCKET_MASK;
354 struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
355
356 sfbhash >>= SFB_BUCKET_SHIFT;
357 if (b->qlen == 0)
358 decrement_prob(b, q);
359 else if (b->qlen >= q->bin_size)
360 increment_prob(b, q);
361 }
362 }
363 if (sfb_rate_limit(skb, q)) {
364 sch->qstats.overlimits++;
365 q->stats.penaltydrop++;
366 goto drop;
367 }
368 goto enqueue;
369 }
370
371 r = net_random() & SFB_MAX_PROB;
372
373 if (unlikely(r < p_min)) {
374 if (unlikely(p_min > SFB_MAX_PROB / 2)) {
375 /* If we're marking that many packets, then either
376 * this flow is unresponsive, or we're badly congested.
377 * In either case, we want to start dropping packets.
378 */
379 if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
380 q->stats.earlydrop++;
381 goto drop;
382 }
383 }
384 if (INET_ECN_set_ce(skb)) {
385 q->stats.marked++;
386 } else {
387 q->stats.earlydrop++;
388 goto drop;
389 }
390 }
391
392enqueue:
393 ret = qdisc_enqueue(skb, child);
394 if (likely(ret == NET_XMIT_SUCCESS)) {
395 sch->q.qlen++;
396 increment_qlen(skb, q);
397 } else if (net_xmit_drop_count(ret)) {
398 q->stats.childdrop++;
399 sch->qstats.drops++;
400 }
401 return ret;
402
403drop:
404 qdisc_drop(skb, sch);
405 return NET_XMIT_CN;
406other_drop:
407 if (ret & __NET_XMIT_BYPASS)
408 sch->qstats.drops++;
409 kfree_skb(skb);
410 return ret;
411}
412
413static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
414{
415 struct sfb_sched_data *q = qdisc_priv(sch);
416 struct Qdisc *child = q->qdisc;
417 struct sk_buff *skb;
418
419 skb = child->dequeue(q->qdisc);
420
421 if (skb) {
422 qdisc_bstats_update(sch, skb);
423 sch->q.qlen--;
424 decrement_qlen(skb, q);
425 }
426
427 return skb;
428}
429
430static struct sk_buff *sfb_peek(struct Qdisc *sch)
431{
432 struct sfb_sched_data *q = qdisc_priv(sch);
433 struct Qdisc *child = q->qdisc;
434
435 return child->ops->peek(child);
436}
437
438/* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
439
440static void sfb_reset(struct Qdisc *sch)
441{
442 struct sfb_sched_data *q = qdisc_priv(sch);
443
444 qdisc_reset(q->qdisc);
445 sch->q.qlen = 0;
446 q->slot = 0;
447 q->double_buffering = false;
448 sfb_zero_all_buckets(q);
449 sfb_init_perturbation(0, q);
450}
451
452static void sfb_destroy(struct Qdisc *sch)
453{
454 struct sfb_sched_data *q = qdisc_priv(sch);
455
456 tcf_destroy_chain(&q->filter_list);
457 qdisc_destroy(q->qdisc);
458}
459
460static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
461 [TCA_SFB_PARMS] = { .len = sizeof(struct tc_sfb_qopt) },
462};
463
464static const struct tc_sfb_qopt sfb_default_ops = {
465 .rehash_interval = 600 * MSEC_PER_SEC,
466 .warmup_time = 60 * MSEC_PER_SEC,
467 .limit = 0,
468 .max = 25,
469 .bin_size = 20,
470 .increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
471 .decrement = (SFB_MAX_PROB + 3000) / 6000,
472 .penalty_rate = 10,
473 .penalty_burst = 20,
474};
475
476static int sfb_change(struct Qdisc *sch, struct nlattr *opt)
477{
478 struct sfb_sched_data *q = qdisc_priv(sch);
479 struct Qdisc *child;
480 struct nlattr *tb[TCA_SFB_MAX + 1];
481 const struct tc_sfb_qopt *ctl = &sfb_default_ops;
482 u32 limit;
483 int err;
484
485 if (opt) {
486 err = nla_parse_nested(tb, TCA_SFB_MAX, opt, sfb_policy);
487 if (err < 0)
488 return -EINVAL;
489
490 if (tb[TCA_SFB_PARMS] == NULL)
491 return -EINVAL;
492
493 ctl = nla_data(tb[TCA_SFB_PARMS]);
494 }
495
496 limit = ctl->limit;
497 if (limit == 0)
498 limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
499
500 child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit);
501 if (IS_ERR(child))
502 return PTR_ERR(child);
503
504 sch_tree_lock(sch);
505
506 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
507 qdisc_destroy(q->qdisc);
508 q->qdisc = child;
509
510 q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
511 q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
512 q->rehash_time = jiffies;
513 q->limit = limit;
514 q->increment = ctl->increment;
515 q->decrement = ctl->decrement;
516 q->max = ctl->max;
517 q->bin_size = ctl->bin_size;
518 q->penalty_rate = ctl->penalty_rate;
519 q->penalty_burst = ctl->penalty_burst;
520 q->tokens_avail = ctl->penalty_burst;
521 q->token_time = jiffies;
522
523 q->slot = 0;
524 q->double_buffering = false;
525 sfb_zero_all_buckets(q);
526 sfb_init_perturbation(0, q);
527 sfb_init_perturbation(1, q);
528
529 sch_tree_unlock(sch);
530
531 return 0;
532}
533
534static int sfb_init(struct Qdisc *sch, struct nlattr *opt)
535{
536 struct sfb_sched_data *q = qdisc_priv(sch);
537
538 q->qdisc = &noop_qdisc;
539 return sfb_change(sch, opt);
540}
541
542static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
543{
544 struct sfb_sched_data *q = qdisc_priv(sch);
545 struct nlattr *opts;
546 struct tc_sfb_qopt opt = {
547 .rehash_interval = jiffies_to_msecs(q->rehash_interval),
548 .warmup_time = jiffies_to_msecs(q->warmup_time),
549 .limit = q->limit,
550 .max = q->max,
551 .bin_size = q->bin_size,
552 .increment = q->increment,
553 .decrement = q->decrement,
554 .penalty_rate = q->penalty_rate,
555 .penalty_burst = q->penalty_burst,
556 };
557
558 sch->qstats.backlog = q->qdisc->qstats.backlog;
559 opts = nla_nest_start(skb, TCA_OPTIONS);
560 NLA_PUT(skb, TCA_SFB_PARMS, sizeof(opt), &opt);
561 return nla_nest_end(skb, opts);
562
563nla_put_failure:
564 nla_nest_cancel(skb, opts);
565 return -EMSGSIZE;
566}
567
568static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
569{
570 struct sfb_sched_data *q = qdisc_priv(sch);
571 struct tc_sfb_xstats st = {
572 .earlydrop = q->stats.earlydrop,
573 .penaltydrop = q->stats.penaltydrop,
574 .bucketdrop = q->stats.bucketdrop,
575 .queuedrop = q->stats.queuedrop,
576 .childdrop = q->stats.childdrop,
577 .marked = q->stats.marked,
578 };
579
580 st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
581
582 return gnet_stats_copy_app(d, &st, sizeof(st));
583}
584
585static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
586 struct sk_buff *skb, struct tcmsg *tcm)
587{
588 return -ENOSYS;
589}
590
591static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
592 struct Qdisc **old)
593{
594 struct sfb_sched_data *q = qdisc_priv(sch);
595
596 if (new == NULL)
597 new = &noop_qdisc;
598
599 sch_tree_lock(sch);
600 *old = q->qdisc;
601 q->qdisc = new;
602 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
603 qdisc_reset(*old);
604 sch_tree_unlock(sch);
605 return 0;
606}
607
608static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
609{
610 struct sfb_sched_data *q = qdisc_priv(sch);
611
612 return q->qdisc;
613}
614
615static unsigned long sfb_get(struct Qdisc *sch, u32 classid)
616{
617 return 1;
618}
619
620static void sfb_put(struct Qdisc *sch, unsigned long arg)
621{
622}
623
624static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
625 struct nlattr **tca, unsigned long *arg)
626{
627 return -ENOSYS;
628}
629
630static int sfb_delete(struct Qdisc *sch, unsigned long cl)
631{
632 return -ENOSYS;
633}
634
635static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
636{
637 if (!walker->stop) {
638 if (walker->count >= walker->skip)
639 if (walker->fn(sch, 1, walker) < 0) {
640 walker->stop = 1;
641 return;
642 }
643 walker->count++;
644 }
645}
646
647static struct tcf_proto **sfb_find_tcf(struct Qdisc *sch, unsigned long cl)
648{
649 struct sfb_sched_data *q = qdisc_priv(sch);
650
651 if (cl)
652 return NULL;
653 return &q->filter_list;
654}
655
656static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
657 u32 classid)
658{
659 return 0;
660}
661
662
663static const struct Qdisc_class_ops sfb_class_ops = {
664 .graft = sfb_graft,
665 .leaf = sfb_leaf,
666 .get = sfb_get,
667 .put = sfb_put,
668 .change = sfb_change_class,
669 .delete = sfb_delete,
670 .walk = sfb_walk,
671 .tcf_chain = sfb_find_tcf,
672 .bind_tcf = sfb_bind,
673 .unbind_tcf = sfb_put,
674 .dump = sfb_dump_class,
675};
676
677static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
678 .id = "sfb",
679 .priv_size = sizeof(struct sfb_sched_data),
680 .cl_ops = &sfb_class_ops,
681 .enqueue = sfb_enqueue,
682 .dequeue = sfb_dequeue,
683 .peek = sfb_peek,
684 .init = sfb_init,
685 .reset = sfb_reset,
686 .destroy = sfb_destroy,
687 .change = sfb_change,
688 .dump = sfb_dump,
689 .dump_stats = sfb_dump_stats,
690 .owner = THIS_MODULE,
691};
692
693static int __init sfb_module_init(void)
694{
695 return register_qdisc(&sfb_qdisc_ops);
696}
697
698static void __exit sfb_module_exit(void)
699{
700 unregister_qdisc(&sfb_qdisc_ops);
701}
702
703module_init(sfb_module_init)
704module_exit(sfb_module_exit)
705
706MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
707MODULE_AUTHOR("Juliusz Chroboczek");
708MODULE_AUTHOR("Eric Dumazet");
709MODULE_LICENSE("GPL");