xref: /openbmc/linux/net/sched/sch_tbf.c (revision 0e96cf7f)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * net/sched/sch_tbf.c	Token Bucket Filter queue.
4  *
5  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6  *		Dmitry Torokhov <dtor@mail.ru> - allow attaching inner qdiscs -
7  *						 original idea by Martin Devera
8  */
9 
10 #include <linux/module.h>
11 #include <linux/types.h>
12 #include <linux/kernel.h>
13 #include <linux/string.h>
14 #include <linux/errno.h>
15 #include <linux/skbuff.h>
16 #include <net/netlink.h>
17 #include <net/sch_generic.h>
18 #include <net/pkt_sched.h>
19 
20 
21 /*	Simple Token Bucket Filter.
22 	=======================================
23 
24 	SOURCE.
25 	-------
26 
27 	None.
28 
29 	Description.
30 	------------
31 
32 	A data flow obeys TBF with rate R and depth B, if for any
33 	time interval t_i...t_f the number of transmitted bits
34 	does not exceed B + R*(t_f-t_i).
35 
36 	Packetized version of this definition:
37 	The sequence of packets of sizes s_i served at moments t_i
38 	obeys TBF, if for any i<=k:
39 
40 	s_i+....+s_k <= B + R*(t_k - t_i)
41 
42 	Algorithm.
43 	----------
44 
45 	Let N(t_i) be B/R initially and N(t) grow continuously with time as:
46 
47 	N(t+delta) = min{B/R, N(t) + delta}
48 
49 	If the first packet in queue has length S, it may be
50 	transmitted only at the time t_* when S/R <= N(t_*),
51 	and in this case N(t) jumps:
52 
53 	N(t_* + 0) = N(t_* - 0) - S/R.
54 
55 
56 
57 	Actually, QoS requires two TBF to be applied to a data stream.
58 	One of them controls steady state burst size, another
59 	one with rate P (peak rate) and depth M (equal to link MTU)
60 	limits bursts at a smaller time scale.
61 
62 	It is easy to see that P>R, and B>M. If P is infinity, this double
63 	TBF is equivalent to a single one.
64 
65 	When TBF works in reshaping mode, latency is estimated as:
66 
67 	lat = max ((L-B)/R, (L-M)/P)
68 
69 
70 	NOTES.
71 	------
72 
73 	If TBF throttles, it starts a watchdog timer, which will wake it up
74 	when it is ready to transmit.
75 	Note that the minimal timer resolution is 1/HZ.
76 	If no new packets arrive during this period,
77 	or if the device is not awaken by EOI for some previous packet,
78 	TBF can stop its activity for 1/HZ.
79 
80 
81 	This means, that with depth B, the maximal rate is
82 
83 	R_crit = B*HZ
84 
85 	F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
86 
87 	Note that the peak rate TBF is much more tough: with MTU 1500
88 	P_crit = 150Kbytes/sec. So, if you need greater peak
89 	rates, use alpha with HZ=1000 :-)
90 
91 	With classful TBF, limit is just kept for backwards compatibility.
92 	It is passed to the default bfifo qdisc - if the inner qdisc is
93 	changed the limit is not effective anymore.
94 */
95 
96 struct tbf_sched_data {
97 /* Parameters */
98 	u32		limit;		/* Maximal length of backlog: bytes */
99 	u32		max_size;
100 	s64		buffer;		/* Token bucket depth/rate: MUST BE >= MTU/B */
101 	s64		mtu;
102 	struct psched_ratecfg rate;
103 	struct psched_ratecfg peak;
104 
105 /* Variables */
106 	s64	tokens;			/* Current number of B tokens */
107 	s64	ptokens;		/* Current number of P tokens */
108 	s64	t_c;			/* Time check-point */
109 	struct Qdisc	*qdisc;		/* Inner qdisc, default - bfifo queue */
110 	struct qdisc_watchdog watchdog;	/* Watchdog timer */
111 };
112 
113 
114 /* Time to Length, convert time in ns to length in bytes
115  * to determinate how many bytes can be sent in given time.
116  */
117 static u64 psched_ns_t2l(const struct psched_ratecfg *r,
118 			 u64 time_in_ns)
119 {
120 	/* The formula is :
121 	 * len = (time_in_ns * r->rate_bytes_ps) / NSEC_PER_SEC
122 	 */
123 	u64 len = time_in_ns * r->rate_bytes_ps;
124 
125 	do_div(len, NSEC_PER_SEC);
126 
127 	if (unlikely(r->linklayer == TC_LINKLAYER_ATM)) {
128 		do_div(len, 53);
129 		len = len * 48;
130 	}
131 
132 	if (len > r->overhead)
133 		len -= r->overhead;
134 	else
135 		len = 0;
136 
137 	return len;
138 }
139 
140 /* GSO packet is too big, segment it so that tbf can transmit
141  * each segment in time
142  */
143 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,
144 		       struct sk_buff **to_free)
145 {
146 	struct tbf_sched_data *q = qdisc_priv(sch);
147 	struct sk_buff *segs, *nskb;
148 	netdev_features_t features = netif_skb_features(skb);
149 	unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
150 	int ret, nb;
151 
152 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
153 
154 	if (IS_ERR_OR_NULL(segs))
155 		return qdisc_drop(skb, sch, to_free);
156 
157 	nb = 0;
158 	while (segs) {
159 		nskb = segs->next;
160 		skb_mark_not_on_list(segs);
161 		qdisc_skb_cb(segs)->pkt_len = segs->len;
162 		len += segs->len;
163 		ret = qdisc_enqueue(segs, q->qdisc, to_free);
164 		if (ret != NET_XMIT_SUCCESS) {
165 			if (net_xmit_drop_count(ret))
166 				qdisc_qstats_drop(sch);
167 		} else {
168 			nb++;
169 		}
170 		segs = nskb;
171 	}
172 	sch->q.qlen += nb;
173 	if (nb > 1)
174 		qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
175 	consume_skb(skb);
176 	return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
177 }
178 
179 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
180 		       struct sk_buff **to_free)
181 {
182 	struct tbf_sched_data *q = qdisc_priv(sch);
183 	unsigned int len = qdisc_pkt_len(skb);
184 	int ret;
185 
186 	if (qdisc_pkt_len(skb) > q->max_size) {
187 		if (skb_is_gso(skb) &&
188 		    skb_gso_validate_mac_len(skb, q->max_size))
189 			return tbf_segment(skb, sch, to_free);
190 		return qdisc_drop(skb, sch, to_free);
191 	}
192 	ret = qdisc_enqueue(skb, q->qdisc, to_free);
193 	if (ret != NET_XMIT_SUCCESS) {
194 		if (net_xmit_drop_count(ret))
195 			qdisc_qstats_drop(sch);
196 		return ret;
197 	}
198 
199 	sch->qstats.backlog += len;
200 	sch->q.qlen++;
201 	return NET_XMIT_SUCCESS;
202 }
203 
204 static bool tbf_peak_present(const struct tbf_sched_data *q)
205 {
206 	return q->peak.rate_bytes_ps;
207 }
208 
209 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
210 {
211 	struct tbf_sched_data *q = qdisc_priv(sch);
212 	struct sk_buff *skb;
213 
214 	skb = q->qdisc->ops->peek(q->qdisc);
215 
216 	if (skb) {
217 		s64 now;
218 		s64 toks;
219 		s64 ptoks = 0;
220 		unsigned int len = qdisc_pkt_len(skb);
221 
222 		now = ktime_get_ns();
223 		toks = min_t(s64, now - q->t_c, q->buffer);
224 
225 		if (tbf_peak_present(q)) {
226 			ptoks = toks + q->ptokens;
227 			if (ptoks > q->mtu)
228 				ptoks = q->mtu;
229 			ptoks -= (s64) psched_l2t_ns(&q->peak, len);
230 		}
231 		toks += q->tokens;
232 		if (toks > q->buffer)
233 			toks = q->buffer;
234 		toks -= (s64) psched_l2t_ns(&q->rate, len);
235 
236 		if ((toks|ptoks) >= 0) {
237 			skb = qdisc_dequeue_peeked(q->qdisc);
238 			if (unlikely(!skb))
239 				return NULL;
240 
241 			q->t_c = now;
242 			q->tokens = toks;
243 			q->ptokens = ptoks;
244 			qdisc_qstats_backlog_dec(sch, skb);
245 			sch->q.qlen--;
246 			qdisc_bstats_update(sch, skb);
247 			return skb;
248 		}
249 
250 		qdisc_watchdog_schedule_ns(&q->watchdog,
251 					   now + max_t(long, -toks, -ptoks));
252 
253 		/* Maybe we have a shorter packet in the queue,
254 		   which can be sent now. It sounds cool,
255 		   but, however, this is wrong in principle.
256 		   We MUST NOT reorder packets under these circumstances.
257 
258 		   Really, if we split the flow into independent
259 		   subflows, it would be a very good solution.
260 		   This is the main idea of all FQ algorithms
261 		   (cf. CSZ, HPFQ, HFSC)
262 		 */
263 
264 		qdisc_qstats_overlimit(sch);
265 	}
266 	return NULL;
267 }
268 
269 static void tbf_reset(struct Qdisc *sch)
270 {
271 	struct tbf_sched_data *q = qdisc_priv(sch);
272 
273 	qdisc_reset(q->qdisc);
274 	sch->qstats.backlog = 0;
275 	sch->q.qlen = 0;
276 	q->t_c = ktime_get_ns();
277 	q->tokens = q->buffer;
278 	q->ptokens = q->mtu;
279 	qdisc_watchdog_cancel(&q->watchdog);
280 }
281 
282 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
283 	[TCA_TBF_PARMS]	= { .len = sizeof(struct tc_tbf_qopt) },
284 	[TCA_TBF_RTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
285 	[TCA_TBF_PTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
286 	[TCA_TBF_RATE64]	= { .type = NLA_U64 },
287 	[TCA_TBF_PRATE64]	= { .type = NLA_U64 },
288 	[TCA_TBF_BURST] = { .type = NLA_U32 },
289 	[TCA_TBF_PBURST] = { .type = NLA_U32 },
290 };
291 
292 static int tbf_change(struct Qdisc *sch, struct nlattr *opt,
293 		      struct netlink_ext_ack *extack)
294 {
295 	int err;
296 	struct tbf_sched_data *q = qdisc_priv(sch);
297 	struct nlattr *tb[TCA_TBF_MAX + 1];
298 	struct tc_tbf_qopt *qopt;
299 	struct Qdisc *child = NULL;
300 	struct psched_ratecfg rate;
301 	struct psched_ratecfg peak;
302 	u64 max_size;
303 	s64 buffer, mtu;
304 	u64 rate64 = 0, prate64 = 0;
305 
306 	err = nla_parse_nested_deprecated(tb, TCA_TBF_MAX, opt, tbf_policy,
307 					  NULL);
308 	if (err < 0)
309 		return err;
310 
311 	err = -EINVAL;
312 	if (tb[TCA_TBF_PARMS] == NULL)
313 		goto done;
314 
315 	qopt = nla_data(tb[TCA_TBF_PARMS]);
316 	if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
317 		qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
318 					      tb[TCA_TBF_RTAB],
319 					      NULL));
320 
321 	if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
322 			qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
323 						      tb[TCA_TBF_PTAB],
324 						      NULL));
325 
326 	buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
327 	mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
328 
329 	if (tb[TCA_TBF_RATE64])
330 		rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
331 	psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
332 
333 	if (tb[TCA_TBF_BURST]) {
334 		max_size = nla_get_u32(tb[TCA_TBF_BURST]);
335 		buffer = psched_l2t_ns(&rate, max_size);
336 	} else {
337 		max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
338 	}
339 
340 	if (qopt->peakrate.rate) {
341 		if (tb[TCA_TBF_PRATE64])
342 			prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
343 		psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
344 		if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
345 			pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
346 					peak.rate_bytes_ps, rate.rate_bytes_ps);
347 			err = -EINVAL;
348 			goto done;
349 		}
350 
351 		if (tb[TCA_TBF_PBURST]) {
352 			u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
353 			max_size = min_t(u32, max_size, pburst);
354 			mtu = psched_l2t_ns(&peak, pburst);
355 		} else {
356 			max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
357 		}
358 	} else {
359 		memset(&peak, 0, sizeof(peak));
360 	}
361 
362 	if (max_size < psched_mtu(qdisc_dev(sch)))
363 		pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
364 				    max_size, qdisc_dev(sch)->name,
365 				    psched_mtu(qdisc_dev(sch)));
366 
367 	if (!max_size) {
368 		err = -EINVAL;
369 		goto done;
370 	}
371 
372 	if (q->qdisc != &noop_qdisc) {
373 		err = fifo_set_limit(q->qdisc, qopt->limit);
374 		if (err)
375 			goto done;
376 	} else if (qopt->limit > 0) {
377 		child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit,
378 					 extack);
379 		if (IS_ERR(child)) {
380 			err = PTR_ERR(child);
381 			goto done;
382 		}
383 
384 		/* child is fifo, no need to check for noop_qdisc */
385 		qdisc_hash_add(child, true);
386 	}
387 
388 	sch_tree_lock(sch);
389 	if (child) {
390 		qdisc_tree_flush_backlog(q->qdisc);
391 		qdisc_put(q->qdisc);
392 		q->qdisc = child;
393 	}
394 	q->limit = qopt->limit;
395 	if (tb[TCA_TBF_PBURST])
396 		q->mtu = mtu;
397 	else
398 		q->mtu = PSCHED_TICKS2NS(qopt->mtu);
399 	q->max_size = max_size;
400 	if (tb[TCA_TBF_BURST])
401 		q->buffer = buffer;
402 	else
403 		q->buffer = PSCHED_TICKS2NS(qopt->buffer);
404 	q->tokens = q->buffer;
405 	q->ptokens = q->mtu;
406 
407 	memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
408 	memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
409 
410 	sch_tree_unlock(sch);
411 	err = 0;
412 done:
413 	return err;
414 }
415 
416 static int tbf_init(struct Qdisc *sch, struct nlattr *opt,
417 		    struct netlink_ext_ack *extack)
418 {
419 	struct tbf_sched_data *q = qdisc_priv(sch);
420 
421 	qdisc_watchdog_init(&q->watchdog, sch);
422 	q->qdisc = &noop_qdisc;
423 
424 	if (!opt)
425 		return -EINVAL;
426 
427 	q->t_c = ktime_get_ns();
428 
429 	return tbf_change(sch, opt, extack);
430 }
431 
432 static void tbf_destroy(struct Qdisc *sch)
433 {
434 	struct tbf_sched_data *q = qdisc_priv(sch);
435 
436 	qdisc_watchdog_cancel(&q->watchdog);
437 	qdisc_put(q->qdisc);
438 }
439 
440 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
441 {
442 	struct tbf_sched_data *q = qdisc_priv(sch);
443 	struct nlattr *nest;
444 	struct tc_tbf_qopt opt;
445 
446 	sch->qstats.backlog = q->qdisc->qstats.backlog;
447 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
448 	if (nest == NULL)
449 		goto nla_put_failure;
450 
451 	opt.limit = q->limit;
452 	psched_ratecfg_getrate(&opt.rate, &q->rate);
453 	if (tbf_peak_present(q))
454 		psched_ratecfg_getrate(&opt.peakrate, &q->peak);
455 	else
456 		memset(&opt.peakrate, 0, sizeof(opt.peakrate));
457 	opt.mtu = PSCHED_NS2TICKS(q->mtu);
458 	opt.buffer = PSCHED_NS2TICKS(q->buffer);
459 	if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
460 		goto nla_put_failure;
461 	if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
462 	    nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
463 			      TCA_TBF_PAD))
464 		goto nla_put_failure;
465 	if (tbf_peak_present(q) &&
466 	    q->peak.rate_bytes_ps >= (1ULL << 32) &&
467 	    nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
468 			      TCA_TBF_PAD))
469 		goto nla_put_failure;
470 
471 	return nla_nest_end(skb, nest);
472 
473 nla_put_failure:
474 	nla_nest_cancel(skb, nest);
475 	return -1;
476 }
477 
478 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
479 			  struct sk_buff *skb, struct tcmsg *tcm)
480 {
481 	struct tbf_sched_data *q = qdisc_priv(sch);
482 
483 	tcm->tcm_handle |= TC_H_MIN(1);
484 	tcm->tcm_info = q->qdisc->handle;
485 
486 	return 0;
487 }
488 
489 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
490 		     struct Qdisc **old, struct netlink_ext_ack *extack)
491 {
492 	struct tbf_sched_data *q = qdisc_priv(sch);
493 
494 	if (new == NULL)
495 		new = &noop_qdisc;
496 
497 	*old = qdisc_replace(sch, new, &q->qdisc);
498 	return 0;
499 }
500 
501 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
502 {
503 	struct tbf_sched_data *q = qdisc_priv(sch);
504 	return q->qdisc;
505 }
506 
507 static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
508 {
509 	return 1;
510 }
511 
512 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
513 {
514 	if (!walker->stop) {
515 		if (walker->count >= walker->skip)
516 			if (walker->fn(sch, 1, walker) < 0) {
517 				walker->stop = 1;
518 				return;
519 			}
520 		walker->count++;
521 	}
522 }
523 
524 static const struct Qdisc_class_ops tbf_class_ops = {
525 	.graft		=	tbf_graft,
526 	.leaf		=	tbf_leaf,
527 	.find		=	tbf_find,
528 	.walk		=	tbf_walk,
529 	.dump		=	tbf_dump_class,
530 };
531 
532 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
533 	.next		=	NULL,
534 	.cl_ops		=	&tbf_class_ops,
535 	.id		=	"tbf",
536 	.priv_size	=	sizeof(struct tbf_sched_data),
537 	.enqueue	=	tbf_enqueue,
538 	.dequeue	=	tbf_dequeue,
539 	.peek		=	qdisc_peek_dequeued,
540 	.init		=	tbf_init,
541 	.reset		=	tbf_reset,
542 	.destroy	=	tbf_destroy,
543 	.change		=	tbf_change,
544 	.dump		=	tbf_dump,
545 	.owner		=	THIS_MODULE,
546 };
547 
548 static int __init tbf_module_init(void)
549 {
550 	return register_qdisc(&tbf_qdisc_ops);
551 }
552 
553 static void __exit tbf_module_exit(void)
554 {
555 	unregister_qdisc(&tbf_qdisc_ops);
556 }
557 module_init(tbf_module_init)
558 module_exit(tbf_module_exit)
559 MODULE_LICENSE("GPL");
560