xref: /openbmc/linux/net/sched/sch_tbf.c (revision be709d48)
1 /*
2  * net/sched/sch_tbf.c	Token Bucket Filter queue.
3  *
4  *		This program is free software; you can redistribute it and/or
5  *		modify it under the terms of the GNU General Public License
6  *		as published by the Free Software Foundation; either version
7  *		2 of the License, or (at your option) any later version.
8  *
9  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10  *		Dmitry Torokhov <dtor@mail.ru> - allow attaching inner qdiscs -
11  *						 original idea by Martin Devera
12  *
13  */
14 
15 #include <linux/module.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/string.h>
19 #include <linux/errno.h>
20 #include <linux/skbuff.h>
21 #include <net/netlink.h>
22 #include <net/sch_generic.h>
23 #include <net/pkt_sched.h>
24 
25 
26 /*	Simple Token Bucket Filter.
27 	=======================================
28 
29 	SOURCE.
30 	-------
31 
32 	None.
33 
34 	Description.
35 	------------
36 
37 	A data flow obeys TBF with rate R and depth B, if for any
38 	time interval t_i...t_f the number of transmitted bits
39 	does not exceed B + R*(t_f-t_i).
40 
41 	Packetized version of this definition:
42 	The sequence of packets of sizes s_i served at moments t_i
43 	obeys TBF, if for any i<=k:
44 
45 	s_i+....+s_k <= B + R*(t_k - t_i)
46 
47 	Algorithm.
48 	----------
49 
50 	Let N(t_i) be B/R initially and N(t) grow continuously with time as:
51 
52 	N(t+delta) = min{B/R, N(t) + delta}
53 
54 	If the first packet in queue has length S, it may be
55 	transmitted only at the time t_* when S/R <= N(t_*),
56 	and in this case N(t) jumps:
57 
58 	N(t_* + 0) = N(t_* - 0) - S/R.
59 
60 
61 
62 	Actually, QoS requires two TBF to be applied to a data stream.
63 	One of them controls steady state burst size, another
64 	one with rate P (peak rate) and depth M (equal to link MTU)
65 	limits bursts at a smaller time scale.
66 
67 	It is easy to see that P>R, and B>M. If P is infinity, this double
68 	TBF is equivalent to a single one.
69 
70 	When TBF works in reshaping mode, latency is estimated as:
71 
72 	lat = max ((L-B)/R, (L-M)/P)
73 
74 
75 	NOTES.
76 	------
77 
78 	If TBF throttles, it starts a watchdog timer, which will wake it up
79 	when it is ready to transmit.
80 	Note that the minimal timer resolution is 1/HZ.
81 	If no new packets arrive during this period,
82 	or if the device is not awaken by EOI for some previous packet,
83 	TBF can stop its activity for 1/HZ.
84 
85 
86 	This means, that with depth B, the maximal rate is
87 
88 	R_crit = B*HZ
89 
90 	F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
91 
92 	Note that the peak rate TBF is much more tough: with MTU 1500
93 	P_crit = 150Kbytes/sec. So, if you need greater peak
94 	rates, use alpha with HZ=1000 :-)
95 
96 	With classful TBF, limit is just kept for backwards compatibility.
97 	It is passed to the default bfifo qdisc - if the inner qdisc is
98 	changed the limit is not effective anymore.
99 */
100 
101 struct tbf_sched_data {
102 /* Parameters */
103 	u32		limit;		/* Maximal length of backlog: bytes */
104 	u32		max_size;
105 	s64		buffer;		/* Token bucket depth/rate: MUST BE >= MTU/B */
106 	s64		mtu;
107 	struct psched_ratecfg rate;
108 	struct psched_ratecfg peak;
109 
110 /* Variables */
111 	s64	tokens;			/* Current number of B tokens */
112 	s64	ptokens;		/* Current number of P tokens */
113 	s64	t_c;			/* Time check-point */
114 	struct Qdisc	*qdisc;		/* Inner qdisc, default - bfifo queue */
115 	struct qdisc_watchdog watchdog;	/* Watchdog timer */
116 };
117 
118 
119 /* Time to Length, convert time in ns to length in bytes
120  * to determinate how many bytes can be sent in given time.
121  */
122 static u64 psched_ns_t2l(const struct psched_ratecfg *r,
123 			 u64 time_in_ns)
124 {
125 	/* The formula is :
126 	 * len = (time_in_ns * r->rate_bytes_ps) / NSEC_PER_SEC
127 	 */
128 	u64 len = time_in_ns * r->rate_bytes_ps;
129 
130 	do_div(len, NSEC_PER_SEC);
131 
132 	if (unlikely(r->linklayer == TC_LINKLAYER_ATM)) {
133 		do_div(len, 53);
134 		len = len * 48;
135 	}
136 
137 	if (len > r->overhead)
138 		len -= r->overhead;
139 	else
140 		len = 0;
141 
142 	return len;
143 }
144 
145 /* GSO packet is too big, segment it so that tbf can transmit
146  * each segment in time
147  */
148 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,
149 		       struct sk_buff **to_free)
150 {
151 	struct tbf_sched_data *q = qdisc_priv(sch);
152 	struct sk_buff *segs, *nskb;
153 	netdev_features_t features = netif_skb_features(skb);
154 	unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
155 	int ret, nb;
156 
157 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
158 
159 	if (IS_ERR_OR_NULL(segs))
160 		return qdisc_drop(skb, sch, to_free);
161 
162 	nb = 0;
163 	while (segs) {
164 		nskb = segs->next;
165 		skb_mark_not_on_list(segs);
166 		qdisc_skb_cb(segs)->pkt_len = segs->len;
167 		len += segs->len;
168 		ret = qdisc_enqueue(segs, q->qdisc, to_free);
169 		if (ret != NET_XMIT_SUCCESS) {
170 			if (net_xmit_drop_count(ret))
171 				qdisc_qstats_drop(sch);
172 		} else {
173 			nb++;
174 		}
175 		segs = nskb;
176 	}
177 	sch->q.qlen += nb;
178 	if (nb > 1)
179 		qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
180 	consume_skb(skb);
181 	return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
182 }
183 
184 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
185 		       struct sk_buff **to_free)
186 {
187 	struct tbf_sched_data *q = qdisc_priv(sch);
188 	unsigned int len = qdisc_pkt_len(skb);
189 	int ret;
190 
191 	if (qdisc_pkt_len(skb) > q->max_size) {
192 		if (skb_is_gso(skb) &&
193 		    skb_gso_validate_mac_len(skb, q->max_size))
194 			return tbf_segment(skb, sch, to_free);
195 		return qdisc_drop(skb, sch, to_free);
196 	}
197 	ret = qdisc_enqueue(skb, q->qdisc, to_free);
198 	if (ret != NET_XMIT_SUCCESS) {
199 		if (net_xmit_drop_count(ret))
200 			qdisc_qstats_drop(sch);
201 		return ret;
202 	}
203 
204 	sch->qstats.backlog += len;
205 	sch->q.qlen++;
206 	return NET_XMIT_SUCCESS;
207 }
208 
209 static bool tbf_peak_present(const struct tbf_sched_data *q)
210 {
211 	return q->peak.rate_bytes_ps;
212 }
213 
214 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
215 {
216 	struct tbf_sched_data *q = qdisc_priv(sch);
217 	struct sk_buff *skb;
218 
219 	skb = q->qdisc->ops->peek(q->qdisc);
220 
221 	if (skb) {
222 		s64 now;
223 		s64 toks;
224 		s64 ptoks = 0;
225 		unsigned int len = qdisc_pkt_len(skb);
226 
227 		now = ktime_get_ns();
228 		toks = min_t(s64, now - q->t_c, q->buffer);
229 
230 		if (tbf_peak_present(q)) {
231 			ptoks = toks + q->ptokens;
232 			if (ptoks > q->mtu)
233 				ptoks = q->mtu;
234 			ptoks -= (s64) psched_l2t_ns(&q->peak, len);
235 		}
236 		toks += q->tokens;
237 		if (toks > q->buffer)
238 			toks = q->buffer;
239 		toks -= (s64) psched_l2t_ns(&q->rate, len);
240 
241 		if ((toks|ptoks) >= 0) {
242 			skb = qdisc_dequeue_peeked(q->qdisc);
243 			if (unlikely(!skb))
244 				return NULL;
245 
246 			q->t_c = now;
247 			q->tokens = toks;
248 			q->ptokens = ptoks;
249 			qdisc_qstats_backlog_dec(sch, skb);
250 			sch->q.qlen--;
251 			qdisc_bstats_update(sch, skb);
252 			return skb;
253 		}
254 
255 		qdisc_watchdog_schedule_ns(&q->watchdog,
256 					   now + max_t(long, -toks, -ptoks));
257 
258 		/* Maybe we have a shorter packet in the queue,
259 		   which can be sent now. It sounds cool,
260 		   but, however, this is wrong in principle.
261 		   We MUST NOT reorder packets under these circumstances.
262 
263 		   Really, if we split the flow into independent
264 		   subflows, it would be a very good solution.
265 		   This is the main idea of all FQ algorithms
266 		   (cf. CSZ, HPFQ, HFSC)
267 		 */
268 
269 		qdisc_qstats_overlimit(sch);
270 	}
271 	return NULL;
272 }
273 
274 static void tbf_reset(struct Qdisc *sch)
275 {
276 	struct tbf_sched_data *q = qdisc_priv(sch);
277 
278 	qdisc_reset(q->qdisc);
279 	sch->qstats.backlog = 0;
280 	sch->q.qlen = 0;
281 	q->t_c = ktime_get_ns();
282 	q->tokens = q->buffer;
283 	q->ptokens = q->mtu;
284 	qdisc_watchdog_cancel(&q->watchdog);
285 }
286 
287 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
288 	[TCA_TBF_PARMS]	= { .len = sizeof(struct tc_tbf_qopt) },
289 	[TCA_TBF_RTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
290 	[TCA_TBF_PTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
291 	[TCA_TBF_RATE64]	= { .type = NLA_U64 },
292 	[TCA_TBF_PRATE64]	= { .type = NLA_U64 },
293 	[TCA_TBF_BURST] = { .type = NLA_U32 },
294 	[TCA_TBF_PBURST] = { .type = NLA_U32 },
295 };
296 
297 static int tbf_change(struct Qdisc *sch, struct nlattr *opt,
298 		      struct netlink_ext_ack *extack)
299 {
300 	int err;
301 	struct tbf_sched_data *q = qdisc_priv(sch);
302 	struct nlattr *tb[TCA_TBF_MAX + 1];
303 	struct tc_tbf_qopt *qopt;
304 	struct Qdisc *child = NULL;
305 	struct psched_ratecfg rate;
306 	struct psched_ratecfg peak;
307 	u64 max_size;
308 	s64 buffer, mtu;
309 	u64 rate64 = 0, prate64 = 0;
310 
311 	err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy, NULL);
312 	if (err < 0)
313 		return err;
314 
315 	err = -EINVAL;
316 	if (tb[TCA_TBF_PARMS] == NULL)
317 		goto done;
318 
319 	qopt = nla_data(tb[TCA_TBF_PARMS]);
320 	if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
321 		qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
322 					      tb[TCA_TBF_RTAB],
323 					      NULL));
324 
325 	if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
326 			qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
327 						      tb[TCA_TBF_PTAB],
328 						      NULL));
329 
330 	buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
331 	mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
332 
333 	if (tb[TCA_TBF_RATE64])
334 		rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
335 	psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
336 
337 	if (tb[TCA_TBF_BURST]) {
338 		max_size = nla_get_u32(tb[TCA_TBF_BURST]);
339 		buffer = psched_l2t_ns(&rate, max_size);
340 	} else {
341 		max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
342 	}
343 
344 	if (qopt->peakrate.rate) {
345 		if (tb[TCA_TBF_PRATE64])
346 			prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
347 		psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
348 		if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
349 			pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
350 					peak.rate_bytes_ps, rate.rate_bytes_ps);
351 			err = -EINVAL;
352 			goto done;
353 		}
354 
355 		if (tb[TCA_TBF_PBURST]) {
356 			u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
357 			max_size = min_t(u32, max_size, pburst);
358 			mtu = psched_l2t_ns(&peak, pburst);
359 		} else {
360 			max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
361 		}
362 	} else {
363 		memset(&peak, 0, sizeof(peak));
364 	}
365 
366 	if (max_size < psched_mtu(qdisc_dev(sch)))
367 		pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
368 				    max_size, qdisc_dev(sch)->name,
369 				    psched_mtu(qdisc_dev(sch)));
370 
371 	if (!max_size) {
372 		err = -EINVAL;
373 		goto done;
374 	}
375 
376 	if (q->qdisc != &noop_qdisc) {
377 		err = fifo_set_limit(q->qdisc, qopt->limit);
378 		if (err)
379 			goto done;
380 	} else if (qopt->limit > 0) {
381 		child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit,
382 					 extack);
383 		if (IS_ERR(child)) {
384 			err = PTR_ERR(child);
385 			goto done;
386 		}
387 
388 		/* child is fifo, no need to check for noop_qdisc */
389 		qdisc_hash_add(child, true);
390 	}
391 
392 	sch_tree_lock(sch);
393 	if (child) {
394 		qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
395 					  q->qdisc->qstats.backlog);
396 		qdisc_put(q->qdisc);
397 		q->qdisc = child;
398 	}
399 	q->limit = qopt->limit;
400 	if (tb[TCA_TBF_PBURST])
401 		q->mtu = mtu;
402 	else
403 		q->mtu = PSCHED_TICKS2NS(qopt->mtu);
404 	q->max_size = max_size;
405 	if (tb[TCA_TBF_BURST])
406 		q->buffer = buffer;
407 	else
408 		q->buffer = PSCHED_TICKS2NS(qopt->buffer);
409 	q->tokens = q->buffer;
410 	q->ptokens = q->mtu;
411 
412 	memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
413 	memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
414 
415 	sch_tree_unlock(sch);
416 	err = 0;
417 done:
418 	return err;
419 }
420 
421 static int tbf_init(struct Qdisc *sch, struct nlattr *opt,
422 		    struct netlink_ext_ack *extack)
423 {
424 	struct tbf_sched_data *q = qdisc_priv(sch);
425 
426 	qdisc_watchdog_init(&q->watchdog, sch);
427 	q->qdisc = &noop_qdisc;
428 
429 	if (!opt)
430 		return -EINVAL;
431 
432 	q->t_c = ktime_get_ns();
433 
434 	return tbf_change(sch, opt, extack);
435 }
436 
437 static void tbf_destroy(struct Qdisc *sch)
438 {
439 	struct tbf_sched_data *q = qdisc_priv(sch);
440 
441 	qdisc_watchdog_cancel(&q->watchdog);
442 	qdisc_put(q->qdisc);
443 }
444 
445 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
446 {
447 	struct tbf_sched_data *q = qdisc_priv(sch);
448 	struct nlattr *nest;
449 	struct tc_tbf_qopt opt;
450 
451 	sch->qstats.backlog = q->qdisc->qstats.backlog;
452 	nest = nla_nest_start(skb, TCA_OPTIONS);
453 	if (nest == NULL)
454 		goto nla_put_failure;
455 
456 	opt.limit = q->limit;
457 	psched_ratecfg_getrate(&opt.rate, &q->rate);
458 	if (tbf_peak_present(q))
459 		psched_ratecfg_getrate(&opt.peakrate, &q->peak);
460 	else
461 		memset(&opt.peakrate, 0, sizeof(opt.peakrate));
462 	opt.mtu = PSCHED_NS2TICKS(q->mtu);
463 	opt.buffer = PSCHED_NS2TICKS(q->buffer);
464 	if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
465 		goto nla_put_failure;
466 	if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
467 	    nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
468 			      TCA_TBF_PAD))
469 		goto nla_put_failure;
470 	if (tbf_peak_present(q) &&
471 	    q->peak.rate_bytes_ps >= (1ULL << 32) &&
472 	    nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
473 			      TCA_TBF_PAD))
474 		goto nla_put_failure;
475 
476 	return nla_nest_end(skb, nest);
477 
478 nla_put_failure:
479 	nla_nest_cancel(skb, nest);
480 	return -1;
481 }
482 
483 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
484 			  struct sk_buff *skb, struct tcmsg *tcm)
485 {
486 	struct tbf_sched_data *q = qdisc_priv(sch);
487 
488 	tcm->tcm_handle |= TC_H_MIN(1);
489 	tcm->tcm_info = q->qdisc->handle;
490 
491 	return 0;
492 }
493 
494 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
495 		     struct Qdisc **old, struct netlink_ext_ack *extack)
496 {
497 	struct tbf_sched_data *q = qdisc_priv(sch);
498 
499 	if (new == NULL)
500 		new = &noop_qdisc;
501 
502 	*old = qdisc_replace(sch, new, &q->qdisc);
503 	return 0;
504 }
505 
506 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
507 {
508 	struct tbf_sched_data *q = qdisc_priv(sch);
509 	return q->qdisc;
510 }
511 
512 static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
513 {
514 	return 1;
515 }
516 
517 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
518 {
519 	if (!walker->stop) {
520 		if (walker->count >= walker->skip)
521 			if (walker->fn(sch, 1, walker) < 0) {
522 				walker->stop = 1;
523 				return;
524 			}
525 		walker->count++;
526 	}
527 }
528 
529 static const struct Qdisc_class_ops tbf_class_ops = {
530 	.graft		=	tbf_graft,
531 	.leaf		=	tbf_leaf,
532 	.find		=	tbf_find,
533 	.walk		=	tbf_walk,
534 	.dump		=	tbf_dump_class,
535 };
536 
537 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
538 	.next		=	NULL,
539 	.cl_ops		=	&tbf_class_ops,
540 	.id		=	"tbf",
541 	.priv_size	=	sizeof(struct tbf_sched_data),
542 	.enqueue	=	tbf_enqueue,
543 	.dequeue	=	tbf_dequeue,
544 	.peek		=	qdisc_peek_dequeued,
545 	.init		=	tbf_init,
546 	.reset		=	tbf_reset,
547 	.destroy	=	tbf_destroy,
548 	.change		=	tbf_change,
549 	.dump		=	tbf_dump,
550 	.owner		=	THIS_MODULE,
551 };
552 
553 static int __init tbf_module_init(void)
554 {
555 	return register_qdisc(&tbf_qdisc_ops);
556 }
557 
558 static void __exit tbf_module_exit(void)
559 {
560 	unregister_qdisc(&tbf_qdisc_ops);
561 }
562 module_init(tbf_module_init)
563 module_exit(tbf_module_exit)
564 MODULE_LICENSE("GPL");
565