xref: /linux/net/sched/sch_sfb.c (revision 985d4a55e64e43bd86eeb896b81ceba453301989) 
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" */
39 struct 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  */
47 struct sfb_bins {
48 	siphash_key_t	  perturbation; /* siphash key */
49 	struct sfb_bucket bins[SFB_LEVELS][SFB_NUMBUCKETS];
50 };
51 
52 struct 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  */
88 struct sfb_skb_cb {
89 	u32 hashes[2];
90 };
91 
92 static 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  */
102 static 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  */
111 static 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 
118 static u32 prob_minus(u32 p1, u32 p2)
119 {
120 	return p1 > p2 ? p1 - p2 : 0;
121 }
122 
123 static 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 			WRITE_ONCE(b[hash].qlen, b[hash].qlen + 1);
134 		b += SFB_NUMBUCKETS; /* next level */
135 	}
136 }
137 
138 static void increment_qlen(const struct sfb_skb_cb *cb, struct sfb_sched_data *q)
139 {
140 	u32 sfbhash;
141 
142 	sfbhash = cb->hashes[0];
143 	if (sfbhash)
144 		increment_one_qlen(sfbhash, 0, q);
145 
146 	sfbhash = cb->hashes[1];
147 	if (sfbhash)
148 		increment_one_qlen(sfbhash, 1, q);
149 }
150 
151 static 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 			WRITE_ONCE(b[hash].qlen, b[hash].qlen - 1);
163 		b += SFB_NUMBUCKETS; /* next level */
164 	}
165 }
166 
167 static 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 
180 static void decrement_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
181 {
182 	WRITE_ONCE(b->p_mark, prob_minus(b->p_mark, q->decrement));
183 }
184 
185 static void increment_prob(struct sfb_bucket *b, struct sfb_sched_data *q)
186 {
187 	WRITE_ONCE(b->p_mark, prob_plus(b->p_mark, q->increment));
188 }
189 
190 static 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  */
198 static 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 		u32 b_qlen = READ_ONCE(b->qlen);
206 		u32 b_mark = READ_ONCE(b->p_mark);
207 
208 		if (qlen < b_qlen)
209 			qlen = b_qlen;
210 		totalpm += b_mark;
211 		if (prob < b_mark)
212 			prob = b_mark;
213 		b++;
214 	}
215 	*prob_r = prob;
216 	*avgpm_r = totalpm / (SFB_LEVELS * SFB_NUMBUCKETS);
217 	return qlen;
218 }
219 
220 
221 static void sfb_init_perturbation(u32 slot, struct sfb_sched_data *q)
222 {
223 	get_random_bytes(&q->bins[slot].perturbation,
224 			 sizeof(q->bins[slot].perturbation));
225 }
226 
227 static void sfb_swap_slot(struct sfb_sched_data *q)
228 {
229 	sfb_init_perturbation(q->slot, q);
230 	q->slot ^= 1;
231 	q->double_buffering = false;
232 }
233 
234 /* Non elastic flows are allowed to use part of the bandwidth, expressed
235  * in "penalty_rate" packets per second, with "penalty_burst" burst
236  */
237 static bool sfb_rate_limit(struct sk_buff *skb, struct sfb_sched_data *q)
238 {
239 	if (q->penalty_rate == 0 || q->penalty_burst == 0)
240 		return true;
241 
242 	if (q->tokens_avail < 1) {
243 		unsigned long age = min(10UL * HZ, jiffies - q->token_time);
244 
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;
251 	}
252 
253 	q->tokens_avail--;
254 	return false;
255 }
256 
257 static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,
258 			 int *qerr, u32 *salt)
259 {
260 	struct tcf_result res;
261 	int result;
262 
263 	result = tcf_classify(skb, NULL, fl, &res, false);
264 	if (result >= 0) {
265 #ifdef CONFIG_NET_CLS_ACT
266 		switch (result) {
267 		case TC_ACT_STOLEN:
268 		case TC_ACT_QUEUED:
269 		case TC_ACT_TRAP:
270 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
271 			fallthrough;
272 		case TC_ACT_SHOT:
273 			return false;
274 		}
275 #endif
276 		*salt = TC_H_MIN(res.classid);
277 		return true;
278 	}
279 	return false;
280 }
281 
282 static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
283 		       struct sk_buff **to_free)
284 {
285 
286 	enum qdisc_drop_reason reason = QDISC_DROP_OVERLIMIT;
287 	struct sfb_sched_data *q = qdisc_priv(sch);
288 	unsigned int len = qdisc_pkt_len(skb);
289 	struct Qdisc *child = q->qdisc;
290 	struct tcf_proto *fl;
291 	struct sfb_skb_cb cb;
292 	int i;
293 	u32 p_min = ~0;
294 	u32 minqlen = ~0;
295 	u32 r, sfbhash;
296 	u32 slot = q->slot;
297 	int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
298 
299 	if (unlikely(sch->q.qlen >= q->limit)) {
300 		qdisc_qstats_overlimit(sch);
301 		WRITE_ONCE(q->stats.queuedrop,
302 			   q->stats.queuedrop + 1);
303 		goto drop;
304 	}
305 
306 	if (q->rehash_interval > 0) {
307 		unsigned long limit = q->rehash_time + q->rehash_interval;
308 
309 		if (unlikely(time_after(jiffies, limit))) {
310 			sfb_swap_slot(q);
311 			q->rehash_time = jiffies;
312 		} else if (unlikely(!q->double_buffering && q->warmup_time > 0 &&
313 				    time_after(jiffies, limit - q->warmup_time))) {
314 			q->double_buffering = true;
315 		}
316 	}
317 
318 	fl = rcu_dereference_bh(q->filter_list);
319 	if (fl) {
320 		u32 salt;
321 
322 		/* If using external classifiers, get result and record it. */
323 		if (!sfb_classify(skb, fl, &ret, &salt))
324 			goto other_drop;
325 		sfbhash = siphash_1u32(salt, &q->bins[slot].perturbation);
326 	} else {
327 		sfbhash = skb_get_hash_perturb(skb, &q->bins[slot].perturbation);
328 	}
329 
330 
331 	if (!sfbhash)
332 		sfbhash = 1;
333 	sfb_skb_cb(skb)->hashes[slot] = sfbhash;
334 
335 	for (i = 0; i < SFB_LEVELS; i++) {
336 		u32 hash = sfbhash & SFB_BUCKET_MASK;
337 		struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
338 
339 		sfbhash >>= SFB_BUCKET_SHIFT;
340 		if (b->qlen == 0)
341 			decrement_prob(b, q);
342 		else if (b->qlen >= q->bin_size)
343 			increment_prob(b, q);
344 		if (minqlen > b->qlen)
345 			minqlen = b->qlen;
346 		if (p_min > b->p_mark)
347 			p_min = b->p_mark;
348 	}
349 
350 	slot ^= 1;
351 	sfb_skb_cb(skb)->hashes[slot] = 0;
352 
353 	if (unlikely(minqlen >= q->max)) {
354 		qdisc_qstats_overlimit(sch);
355 		WRITE_ONCE(q->stats.bucketdrop,
356 			   q->stats.bucketdrop + 1);
357 		goto drop;
358 	}
359 
360 	if (unlikely(p_min >= SFB_MAX_PROB)) {
361 		/* Inelastic flow */
362 		if (q->double_buffering) {
363 			sfbhash = skb_get_hash_perturb(skb,
364 			    &q->bins[slot].perturbation);
365 			if (!sfbhash)
366 				sfbhash = 1;
367 			sfb_skb_cb(skb)->hashes[slot] = sfbhash;
368 
369 			for (i = 0; i < SFB_LEVELS; i++) {
370 				u32 hash = sfbhash & SFB_BUCKET_MASK;
371 				struct sfb_bucket *b = &q->bins[slot].bins[i][hash];
372 
373 				sfbhash >>= SFB_BUCKET_SHIFT;
374 				if (b->qlen == 0)
375 					decrement_prob(b, q);
376 				else if (b->qlen >= q->bin_size)
377 					increment_prob(b, q);
378 			}
379 		}
380 		if (sfb_rate_limit(skb, q)) {
381 			qdisc_qstats_overlimit(sch);
382 			WRITE_ONCE(q->stats.penaltydrop,
383 				   q->stats.penaltydrop + 1);
384 			goto drop;
385 		}
386 		goto enqueue;
387 	}
388 
389 	r = get_random_u16() & SFB_MAX_PROB;
390 	reason = QDISC_DROP_CONGESTED;
391 
392 	if (unlikely(r < p_min)) {
393 		if (unlikely(p_min > SFB_MAX_PROB / 2)) {
394 			/* If we're marking that many packets, then either
395 			 * this flow is unresponsive, or we're badly congested.
396 			 * In either case, we want to start dropping packets.
397 			 */
398 			if (r < (p_min - SFB_MAX_PROB / 2) * 2) {
399 				WRITE_ONCE(q->stats.earlydrop,
400 					   q->stats.earlydrop + 1);
401 				goto drop;
402 			}
403 		}
404 		if (INET_ECN_set_ce(skb)) {
405 			WRITE_ONCE(q->stats.marked,
406 				   q->stats.marked + 1);
407 		} else {
408 			WRITE_ONCE(q->stats.earlydrop,
409 				   q->stats.earlydrop + 1);
410 			goto drop;
411 		}
412 	}
413 
414 enqueue:
415 	memcpy(&cb, sfb_skb_cb(skb), sizeof(cb));
416 	ret = qdisc_enqueue(skb, child, to_free);
417 	if (likely(ret == NET_XMIT_SUCCESS)) {
418 		sch->qstats.backlog += len;
419 		sch->q.qlen++;
420 		increment_qlen(&cb, q);
421 	} else if (net_xmit_drop_count(ret)) {
422 		WRITE_ONCE(q->stats.childdrop,
423 			   q->stats.childdrop + 1);
424 		qdisc_qstats_drop(sch);
425 	}
426 	return ret;
427 
428 drop:
429 	qdisc_drop_reason(skb, sch, to_free, reason);
430 	return NET_XMIT_CN;
431 other_drop:
432 	if (ret & __NET_XMIT_BYPASS)
433 		qdisc_qstats_drop(sch);
434 	kfree_skb(skb);
435 	return ret;
436 }
437 
438 static struct sk_buff *sfb_dequeue(struct Qdisc *sch)
439 {
440 	struct sfb_sched_data *q = qdisc_priv(sch);
441 	struct Qdisc *child = q->qdisc;
442 	struct sk_buff *skb;
443 
444 	skb = qdisc_dequeue_peeked(child);
445 
446 	if (skb) {
447 		qdisc_bstats_update(sch, skb);
448 		qdisc_qstats_backlog_dec(sch, skb);
449 		sch->q.qlen--;
450 		decrement_qlen(skb, q);
451 	}
452 
453 	return skb;
454 }
455 
456 static struct sk_buff *sfb_peek(struct Qdisc *sch)
457 {
458 	struct sfb_sched_data *q = qdisc_priv(sch);
459 	struct Qdisc *child = q->qdisc;
460 
461 	return child->ops->peek(child);
462 }
463 
464 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
465 
466 static void sfb_reset(struct Qdisc *sch)
467 {
468 	struct sfb_sched_data *q = qdisc_priv(sch);
469 
470 	if (likely(q->qdisc))
471 		qdisc_reset(q->qdisc);
472 	q->slot = 0;
473 	q->double_buffering = false;
474 	sfb_zero_all_buckets(q);
475 	sfb_init_perturbation(0, q);
476 }
477 
478 static void sfb_destroy(struct Qdisc *sch)
479 {
480 	struct sfb_sched_data *q = qdisc_priv(sch);
481 
482 	tcf_block_put(q->block);
483 	qdisc_put(q->qdisc);
484 }
485 
486 static const struct nla_policy sfb_policy[TCA_SFB_MAX + 1] = {
487 	[TCA_SFB_PARMS]	= { .len = sizeof(struct tc_sfb_qopt) },
488 };
489 
490 static const struct tc_sfb_qopt sfb_default_ops = {
491 	.rehash_interval = 600 * MSEC_PER_SEC,
492 	.warmup_time = 60 * MSEC_PER_SEC,
493 	.limit = 0,
494 	.max = 25,
495 	.bin_size = 20,
496 	.increment = (SFB_MAX_PROB + 500) / 1000, /* 0.1 % */
497 	.decrement = (SFB_MAX_PROB + 3000) / 6000,
498 	.penalty_rate = 10,
499 	.penalty_burst = 20,
500 };
501 
502 static int sfb_change(struct Qdisc *sch, struct nlattr *opt,
503 		      struct netlink_ext_ack *extack)
504 {
505 	struct sfb_sched_data *q = qdisc_priv(sch);
506 	struct Qdisc *child, *old;
507 	struct nlattr *tb[TCA_SFB_MAX + 1];
508 	const struct tc_sfb_qopt *ctl = &sfb_default_ops;
509 	u32 limit;
510 	int err;
511 
512 	if (opt) {
513 		err = nla_parse_nested_deprecated(tb, TCA_SFB_MAX, opt,
514 						  sfb_policy, NULL);
515 		if (err < 0)
516 			return -EINVAL;
517 
518 		if (tb[TCA_SFB_PARMS] == NULL)
519 			return -EINVAL;
520 
521 		ctl = nla_data(tb[TCA_SFB_PARMS]);
522 	}
523 
524 	limit = ctl->limit;
525 	if (limit == 0)
526 		limit = qdisc_dev(sch)->tx_queue_len;
527 
528 	child = fifo_create_dflt(sch, &pfifo_qdisc_ops, limit, extack);
529 	if (IS_ERR(child))
530 		return PTR_ERR(child);
531 
532 	if (child != &noop_qdisc)
533 		qdisc_hash_add(child, true);
534 	sch_tree_lock(sch);
535 
536 	qdisc_purge_queue(q->qdisc);
537 	old = q->qdisc;
538 	q->qdisc = child;
539 
540 	q->rehash_interval = msecs_to_jiffies(ctl->rehash_interval);
541 	q->warmup_time = msecs_to_jiffies(ctl->warmup_time);
542 	q->rehash_time = jiffies;
543 	q->limit = limit;
544 	q->increment = ctl->increment;
545 	q->decrement = ctl->decrement;
546 	q->max = ctl->max;
547 	q->bin_size = ctl->bin_size;
548 	q->penalty_rate = ctl->penalty_rate;
549 	q->penalty_burst = ctl->penalty_burst;
550 	q->tokens_avail = ctl->penalty_burst;
551 	q->token_time = jiffies;
552 
553 	q->slot = 0;
554 	q->double_buffering = false;
555 	sfb_zero_all_buckets(q);
556 	sfb_init_perturbation(0, q);
557 	sfb_init_perturbation(1, q);
558 
559 	sch_tree_unlock(sch);
560 	qdisc_put(old);
561 
562 	return 0;
563 }
564 
565 static int sfb_init(struct Qdisc *sch, struct nlattr *opt,
566 		    struct netlink_ext_ack *extack)
567 {
568 	struct sfb_sched_data *q = qdisc_priv(sch);
569 	int err;
570 
571 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
572 	if (err)
573 		return err;
574 
575 	q->qdisc = &noop_qdisc;
576 	return sfb_change(sch, opt, extack);
577 }
578 
579 static int sfb_dump(struct Qdisc *sch, struct sk_buff *skb)
580 {
581 	struct sfb_sched_data *q = qdisc_priv(sch);
582 	struct nlattr *opts;
583 	struct tc_sfb_qopt opt = {
584 		.rehash_interval = jiffies_to_msecs(q->rehash_interval),
585 		.warmup_time = jiffies_to_msecs(q->warmup_time),
586 		.limit = q->limit,
587 		.max = q->max,
588 		.bin_size = q->bin_size,
589 		.increment = q->increment,
590 		.decrement = q->decrement,
591 		.penalty_rate = q->penalty_rate,
592 		.penalty_burst = q->penalty_burst,
593 	};
594 
595 	sch->qstats.backlog = q->qdisc->qstats.backlog;
596 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
597 	if (opts == NULL)
598 		goto nla_put_failure;
599 	if (nla_put(skb, TCA_SFB_PARMS, sizeof(opt), &opt))
600 		goto nla_put_failure;
601 	return nla_nest_end(skb, opts);
602 
603 nla_put_failure:
604 	nla_nest_cancel(skb, opts);
605 	return -EMSGSIZE;
606 }
607 
608 static int sfb_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
609 {
610 	struct sfb_sched_data *q = qdisc_priv(sch);
611 	struct tc_sfb_xstats st = {
612 		.earlydrop = READ_ONCE(q->stats.earlydrop),
613 		.penaltydrop = READ_ONCE(q->stats.penaltydrop),
614 		.bucketdrop = READ_ONCE(q->stats.bucketdrop),
615 		.queuedrop = READ_ONCE(q->stats.queuedrop),
616 		.childdrop = READ_ONCE(q->stats.childdrop),
617 		.marked = READ_ONCE(q->stats.marked),
618 	};
619 
620 	st.maxqlen = sfb_compute_qlen(&st.maxprob, &st.avgprob, q);
621 
622 	return gnet_stats_copy_app(d, &st, sizeof(st));
623 }
624 
625 static int sfb_dump_class(struct Qdisc *sch, unsigned long cl,
626 			  struct sk_buff *skb, struct tcmsg *tcm)
627 {
628 	return -ENOSYS;
629 }
630 
631 static int sfb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
632 		     struct Qdisc **old, struct netlink_ext_ack *extack)
633 {
634 	struct sfb_sched_data *q = qdisc_priv(sch);
635 
636 	if (new == NULL)
637 		new = &noop_qdisc;
638 
639 	*old = qdisc_replace(sch, new, &q->qdisc);
640 	return 0;
641 }
642 
643 static struct Qdisc *sfb_leaf(struct Qdisc *sch, unsigned long arg)
644 {
645 	struct sfb_sched_data *q = qdisc_priv(sch);
646 
647 	return q->qdisc;
648 }
649 
650 static unsigned long sfb_find(struct Qdisc *sch, u32 classid)
651 {
652 	return 1;
653 }
654 
655 static void sfb_unbind(struct Qdisc *sch, unsigned long arg)
656 {
657 }
658 
659 static int sfb_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
660 			    struct nlattr **tca, unsigned long *arg,
661 			    struct netlink_ext_ack *extack)
662 {
663 	return -ENOSYS;
664 }
665 
666 static int sfb_delete(struct Qdisc *sch, unsigned long cl,
667 		      struct netlink_ext_ack *extack)
668 {
669 	return -ENOSYS;
670 }
671 
672 static void sfb_walk(struct Qdisc *sch, struct qdisc_walker *walker)
673 {
674 	if (!walker->stop) {
675 		tc_qdisc_stats_dump(sch, 1, walker);
676 	}
677 }
678 
679 static struct tcf_block *sfb_tcf_block(struct Qdisc *sch, unsigned long cl,
680 				       struct netlink_ext_ack *extack)
681 {
682 	struct sfb_sched_data *q = qdisc_priv(sch);
683 
684 	if (cl)
685 		return NULL;
686 	return q->block;
687 }
688 
689 static unsigned long sfb_bind(struct Qdisc *sch, unsigned long parent,
690 			      u32 classid)
691 {
692 	return 0;
693 }
694 
695 
696 static const struct Qdisc_class_ops sfb_class_ops = {
697 	.graft		=	sfb_graft,
698 	.leaf		=	sfb_leaf,
699 	.find		=	sfb_find,
700 	.change		=	sfb_change_class,
701 	.delete		=	sfb_delete,
702 	.walk		=	sfb_walk,
703 	.tcf_block	=	sfb_tcf_block,
704 	.bind_tcf	=	sfb_bind,
705 	.unbind_tcf	=	sfb_unbind,
706 	.dump		=	sfb_dump_class,
707 };
708 
709 static struct Qdisc_ops sfb_qdisc_ops __read_mostly = {
710 	.id		=	"sfb",
711 	.priv_size	=	sizeof(struct sfb_sched_data),
712 	.cl_ops		=	&sfb_class_ops,
713 	.enqueue	=	sfb_enqueue,
714 	.dequeue	=	sfb_dequeue,
715 	.peek		=	sfb_peek,
716 	.init		=	sfb_init,
717 	.reset		=	sfb_reset,
718 	.destroy	=	sfb_destroy,
719 	.change		=	sfb_change,
720 	.dump		=	sfb_dump,
721 	.dump_stats	=	sfb_dump_stats,
722 	.owner		=	THIS_MODULE,
723 };
724 MODULE_ALIAS_NET_SCH("sfb");
725 
726 static int __init sfb_module_init(void)
727 {
728 	return register_qdisc(&sfb_qdisc_ops);
729 }
730 
731 static void __exit sfb_module_exit(void)
732 {
733 	unregister_qdisc(&sfb_qdisc_ops);
734 }
735 
736 module_init(sfb_module_init)
737 module_exit(sfb_module_exit)
738 
739 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
740 MODULE_AUTHOR("Juliusz Chroboczek");
741 MODULE_AUTHOR("Eric Dumazet");
742 MODULE_LICENSE("GPL");
743