xref: /openbmc/linux/net/sched/sch_tbf.c (revision cd2a9e62c8a3c5cae7691982667d79a0edc65283) 
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 /*
146  * Return length of individual segments of a gso packet,
147  * including all headers (MAC, IP, TCP/UDP)
148  */
149 static unsigned int skb_gso_mac_seglen(const struct sk_buff *skb)
150 {
151 	unsigned int hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
152 	return hdr_len + skb_gso_transport_seglen(skb);
153 }
154 
155 /* GSO packet is too big, segment it so that tbf can transmit
156  * each segment in time
157  */
158 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch)
159 {
160 	struct tbf_sched_data *q = qdisc_priv(sch);
161 	struct sk_buff *segs, *nskb;
162 	netdev_features_t features = netif_skb_features(skb);
163 	unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
164 	int ret, nb;
165 
166 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
167 
168 	if (IS_ERR_OR_NULL(segs))
169 		return qdisc_drop(skb, sch);
170 
171 	nb = 0;
172 	while (segs) {
173 		nskb = segs->next;
174 		segs->next = NULL;
175 		qdisc_skb_cb(segs)->pkt_len = segs->len;
176 		len += segs->len;
177 		ret = qdisc_enqueue(segs, q->qdisc);
178 		if (ret != NET_XMIT_SUCCESS) {
179 			if (net_xmit_drop_count(ret))
180 				qdisc_qstats_drop(sch);
181 		} else {
182 			nb++;
183 		}
184 		segs = nskb;
185 	}
186 	sch->q.qlen += nb;
187 	if (nb > 1)
188 		qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
189 	consume_skb(skb);
190 	return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
191 }
192 
193 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch)
194 {
195 	struct tbf_sched_data *q = qdisc_priv(sch);
196 	int ret;
197 
198 	if (qdisc_pkt_len(skb) > q->max_size) {
199 		if (skb_is_gso(skb) && skb_gso_mac_seglen(skb) <= q->max_size)
200 			return tbf_segment(skb, sch);
201 		return qdisc_drop(skb, sch);
202 	}
203 	ret = qdisc_enqueue(skb, q->qdisc);
204 	if (ret != NET_XMIT_SUCCESS) {
205 		if (net_xmit_drop_count(ret))
206 			qdisc_qstats_drop(sch);
207 		return ret;
208 	}
209 
210 	qdisc_qstats_backlog_inc(sch, skb);
211 	sch->q.qlen++;
212 	return NET_XMIT_SUCCESS;
213 }
214 
215 static bool tbf_peak_present(const struct tbf_sched_data *q)
216 {
217 	return q->peak.rate_bytes_ps;
218 }
219 
220 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
221 {
222 	struct tbf_sched_data *q = qdisc_priv(sch);
223 	struct sk_buff *skb;
224 
225 	skb = q->qdisc->ops->peek(q->qdisc);
226 
227 	if (skb) {
228 		s64 now;
229 		s64 toks;
230 		s64 ptoks = 0;
231 		unsigned int len = qdisc_pkt_len(skb);
232 
233 		now = ktime_get_ns();
234 		toks = min_t(s64, now - q->t_c, q->buffer);
235 
236 		if (tbf_peak_present(q)) {
237 			ptoks = toks + q->ptokens;
238 			if (ptoks > q->mtu)
239 				ptoks = q->mtu;
240 			ptoks -= (s64) psched_l2t_ns(&q->peak, len);
241 		}
242 		toks += q->tokens;
243 		if (toks > q->buffer)
244 			toks = q->buffer;
245 		toks -= (s64) psched_l2t_ns(&q->rate, len);
246 
247 		if ((toks|ptoks) >= 0) {
248 			skb = qdisc_dequeue_peeked(q->qdisc);
249 			if (unlikely(!skb))
250 				return NULL;
251 
252 			q->t_c = now;
253 			q->tokens = toks;
254 			q->ptokens = ptoks;
255 			qdisc_qstats_backlog_dec(sch, skb);
256 			sch->q.qlen--;
257 			qdisc_bstats_update(sch, skb);
258 			return skb;
259 		}
260 
261 		qdisc_watchdog_schedule_ns(&q->watchdog,
262 					   now + max_t(long, -toks, -ptoks));
263 
264 		/* Maybe we have a shorter packet in the queue,
265 		   which can be sent now. It sounds cool,
266 		   but, however, this is wrong in principle.
267 		   We MUST NOT reorder packets under these circumstances.
268 
269 		   Really, if we split the flow into independent
270 		   subflows, it would be a very good solution.
271 		   This is the main idea of all FQ algorithms
272 		   (cf. CSZ, HPFQ, HFSC)
273 		 */
274 
275 		qdisc_qstats_overlimit(sch);
276 	}
277 	return NULL;
278 }
279 
280 static void tbf_reset(struct Qdisc *sch)
281 {
282 	struct tbf_sched_data *q = qdisc_priv(sch);
283 
284 	qdisc_reset(q->qdisc);
285 	sch->qstats.backlog = 0;
286 	sch->q.qlen = 0;
287 	q->t_c = ktime_get_ns();
288 	q->tokens = q->buffer;
289 	q->ptokens = q->mtu;
290 	qdisc_watchdog_cancel(&q->watchdog);
291 }
292 
293 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
294 	[TCA_TBF_PARMS]	= { .len = sizeof(struct tc_tbf_qopt) },
295 	[TCA_TBF_RTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
296 	[TCA_TBF_PTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
297 	[TCA_TBF_RATE64]	= { .type = NLA_U64 },
298 	[TCA_TBF_PRATE64]	= { .type = NLA_U64 },
299 	[TCA_TBF_BURST] = { .type = NLA_U32 },
300 	[TCA_TBF_PBURST] = { .type = NLA_U32 },
301 };
302 
303 static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
304 {
305 	int err;
306 	struct tbf_sched_data *q = qdisc_priv(sch);
307 	struct nlattr *tb[TCA_TBF_MAX + 1];
308 	struct tc_tbf_qopt *qopt;
309 	struct Qdisc *child = NULL;
310 	struct psched_ratecfg rate;
311 	struct psched_ratecfg peak;
312 	u64 max_size;
313 	s64 buffer, mtu;
314 	u64 rate64 = 0, prate64 = 0;
315 
316 	err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy);
317 	if (err < 0)
318 		return err;
319 
320 	err = -EINVAL;
321 	if (tb[TCA_TBF_PARMS] == NULL)
322 		goto done;
323 
324 	qopt = nla_data(tb[TCA_TBF_PARMS]);
325 	if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
326 		qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
327 					      tb[TCA_TBF_RTAB]));
328 
329 	if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
330 			qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
331 						      tb[TCA_TBF_PTAB]));
332 
333 	buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
334 	mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
335 
336 	if (tb[TCA_TBF_RATE64])
337 		rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
338 	psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
339 
340 	if (tb[TCA_TBF_BURST]) {
341 		max_size = nla_get_u32(tb[TCA_TBF_BURST]);
342 		buffer = psched_l2t_ns(&rate, max_size);
343 	} else {
344 		max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
345 	}
346 
347 	if (qopt->peakrate.rate) {
348 		if (tb[TCA_TBF_PRATE64])
349 			prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
350 		psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
351 		if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
352 			pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
353 					peak.rate_bytes_ps, rate.rate_bytes_ps);
354 			err = -EINVAL;
355 			goto done;
356 		}
357 
358 		if (tb[TCA_TBF_PBURST]) {
359 			u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
360 			max_size = min_t(u32, max_size, pburst);
361 			mtu = psched_l2t_ns(&peak, pburst);
362 		} else {
363 			max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
364 		}
365 	} else {
366 		memset(&peak, 0, sizeof(peak));
367 	}
368 
369 	if (max_size < psched_mtu(qdisc_dev(sch)))
370 		pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
371 				    max_size, qdisc_dev(sch)->name,
372 				    psched_mtu(qdisc_dev(sch)));
373 
374 	if (!max_size) {
375 		err = -EINVAL;
376 		goto done;
377 	}
378 
379 	if (q->qdisc != &noop_qdisc) {
380 		err = fifo_set_limit(q->qdisc, qopt->limit);
381 		if (err)
382 			goto done;
383 	} else if (qopt->limit > 0) {
384 		child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
385 		if (IS_ERR(child)) {
386 			err = PTR_ERR(child);
387 			goto done;
388 		}
389 	}
390 
391 	sch_tree_lock(sch);
392 	if (child) {
393 		qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
394 					  q->qdisc->qstats.backlog);
395 		qdisc_destroy(q->qdisc);
396 		q->qdisc = child;
397 	}
398 	q->limit = qopt->limit;
399 	if (tb[TCA_TBF_PBURST])
400 		q->mtu = mtu;
401 	else
402 		q->mtu = PSCHED_TICKS2NS(qopt->mtu);
403 	q->max_size = max_size;
404 	if (tb[TCA_TBF_BURST])
405 		q->buffer = buffer;
406 	else
407 		q->buffer = PSCHED_TICKS2NS(qopt->buffer);
408 	q->tokens = q->buffer;
409 	q->ptokens = q->mtu;
410 
411 	memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
412 	memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
413 
414 	sch_tree_unlock(sch);
415 	err = 0;
416 done:
417 	return err;
418 }
419 
420 static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
421 {
422 	struct tbf_sched_data *q = qdisc_priv(sch);
423 
424 	if (opt == NULL)
425 		return -EINVAL;
426 
427 	q->t_c = ktime_get_ns();
428 	qdisc_watchdog_init(&q->watchdog, sch);
429 	q->qdisc = &noop_qdisc;
430 
431 	return tbf_change(sch, opt);
432 }
433 
434 static void tbf_destroy(struct Qdisc *sch)
435 {
436 	struct tbf_sched_data *q = qdisc_priv(sch);
437 
438 	qdisc_watchdog_cancel(&q->watchdog);
439 	qdisc_destroy(q->qdisc);
440 }
441 
442 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
443 {
444 	struct tbf_sched_data *q = qdisc_priv(sch);
445 	struct nlattr *nest;
446 	struct tc_tbf_qopt opt;
447 
448 	sch->qstats.backlog = q->qdisc->qstats.backlog;
449 	nest = nla_nest_start(skb, TCA_OPTIONS);
450 	if (nest == NULL)
451 		goto nla_put_failure;
452 
453 	opt.limit = q->limit;
454 	psched_ratecfg_getrate(&opt.rate, &q->rate);
455 	if (tbf_peak_present(q))
456 		psched_ratecfg_getrate(&opt.peakrate, &q->peak);
457 	else
458 		memset(&opt.peakrate, 0, sizeof(opt.peakrate));
459 	opt.mtu = PSCHED_NS2TICKS(q->mtu);
460 	opt.buffer = PSCHED_NS2TICKS(q->buffer);
461 	if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
462 		goto nla_put_failure;
463 	if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
464 	    nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
465 			      TCA_TBF_PAD))
466 		goto nla_put_failure;
467 	if (tbf_peak_present(q) &&
468 	    q->peak.rate_bytes_ps >= (1ULL << 32) &&
469 	    nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
470 			      TCA_TBF_PAD))
471 		goto nla_put_failure;
472 
473 	return nla_nest_end(skb, nest);
474 
475 nla_put_failure:
476 	nla_nest_cancel(skb, nest);
477 	return -1;
478 }
479 
480 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
481 			  struct sk_buff *skb, struct tcmsg *tcm)
482 {
483 	struct tbf_sched_data *q = qdisc_priv(sch);
484 
485 	tcm->tcm_handle |= TC_H_MIN(1);
486 	tcm->tcm_info = q->qdisc->handle;
487 
488 	return 0;
489 }
490 
491 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
492 		     struct Qdisc **old)
493 {
494 	struct tbf_sched_data *q = qdisc_priv(sch);
495 
496 	if (new == NULL)
497 		new = &noop_qdisc;
498 
499 	*old = qdisc_replace(sch, new, &q->qdisc);
500 	return 0;
501 }
502 
503 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
504 {
505 	struct tbf_sched_data *q = qdisc_priv(sch);
506 	return q->qdisc;
507 }
508 
509 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
510 {
511 	return 1;
512 }
513 
514 static void tbf_put(struct Qdisc *sch, unsigned long arg)
515 {
516 }
517 
518 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
519 {
520 	if (!walker->stop) {
521 		if (walker->count >= walker->skip)
522 			if (walker->fn(sch, 1, walker) < 0) {
523 				walker->stop = 1;
524 				return;
525 			}
526 		walker->count++;
527 	}
528 }
529 
530 static const struct Qdisc_class_ops tbf_class_ops = {
531 	.graft		=	tbf_graft,
532 	.leaf		=	tbf_leaf,
533 	.get		=	tbf_get,
534 	.put		=	tbf_put,
535 	.walk		=	tbf_walk,
536 	.dump		=	tbf_dump_class,
537 };
538 
539 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
540 	.next		=	NULL,
541 	.cl_ops		=	&tbf_class_ops,
542 	.id		=	"tbf",
543 	.priv_size	=	sizeof(struct tbf_sched_data),
544 	.enqueue	=	tbf_enqueue,
545 	.dequeue	=	tbf_dequeue,
546 	.peek		=	qdisc_peek_dequeued,
547 	.init		=	tbf_init,
548 	.reset		=	tbf_reset,
549 	.destroy	=	tbf_destroy,
550 	.change		=	tbf_change,
551 	.dump		=	tbf_dump,
552 	.owner		=	THIS_MODULE,
553 };
554 
555 static int __init tbf_module_init(void)
556 {
557 	return register_qdisc(&tbf_qdisc_ops);
558 }
559 
560 static void __exit tbf_module_exit(void)
561 {
562 	unregister_qdisc(&tbf_qdisc_ops);
563 }
564 module_init(tbf_module_init)
565 module_exit(tbf_module_exit)
566 MODULE_LICENSE("GPL");
567