xref: /openbmc/linux/net/sched/sch_tbf.c (revision b6dcefde)
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/pkt_sched.h>
23 
24 
25 /*	Simple Token Bucket Filter.
26 	=======================================
27 
28 	SOURCE.
29 	-------
30 
31 	None.
32 
33 	Description.
34 	------------
35 
36 	A data flow obeys TBF with rate R and depth B, if for any
37 	time interval t_i...t_f the number of transmitted bits
38 	does not exceed B + R*(t_f-t_i).
39 
40 	Packetized version of this definition:
41 	The sequence of packets of sizes s_i served at moments t_i
42 	obeys TBF, if for any i<=k:
43 
44 	s_i+....+s_k <= B + R*(t_k - t_i)
45 
46 	Algorithm.
47 	----------
48 
49 	Let N(t_i) be B/R initially and N(t) grow continuously with time as:
50 
51 	N(t+delta) = min{B/R, N(t) + delta}
52 
53 	If the first packet in queue has length S, it may be
54 	transmitted only at the time t_* when S/R <= N(t_*),
55 	and in this case N(t) jumps:
56 
57 	N(t_* + 0) = N(t_* - 0) - S/R.
58 
59 
60 
61 	Actually, QoS requires two TBF to be applied to a data stream.
62 	One of them controls steady state burst size, another
63 	one with rate P (peak rate) and depth M (equal to link MTU)
64 	limits bursts at a smaller time scale.
65 
66 	It is easy to see that P>R, and B>M. If P is infinity, this double
67 	TBF is equivalent to a single one.
68 
69 	When TBF works in reshaping mode, latency is estimated as:
70 
71 	lat = max ((L-B)/R, (L-M)/P)
72 
73 
74 	NOTES.
75 	------
76 
77 	If TBF throttles, it starts a watchdog timer, which will wake it up
78 	when it is ready to transmit.
79 	Note that the minimal timer resolution is 1/HZ.
80 	If no new packets arrive during this period,
81 	or if the device is not awaken by EOI for some previous packet,
82 	TBF can stop its activity for 1/HZ.
83 
84 
85 	This means, that with depth B, the maximal rate is
86 
87 	R_crit = B*HZ
88 
89 	F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
90 
91 	Note that the peak rate TBF is much more tough: with MTU 1500
92 	P_crit = 150Kbytes/sec. So, if you need greater peak
93 	rates, use alpha with HZ=1000 :-)
94 
95 	With classful TBF, limit is just kept for backwards compatibility.
96 	It is passed to the default bfifo qdisc - if the inner qdisc is
97 	changed the limit is not effective anymore.
98 */
99 
100 struct tbf_sched_data
101 {
102 /* Parameters */
103 	u32		limit;		/* Maximal length of backlog: bytes */
104 	u32		buffer;		/* Token bucket depth/rate: MUST BE >= MTU/B */
105 	u32		mtu;
106 	u32		max_size;
107 	struct qdisc_rate_table	*R_tab;
108 	struct qdisc_rate_table	*P_tab;
109 
110 /* Variables */
111 	long	tokens;			/* Current number of B tokens */
112 	long	ptokens;		/* Current number of P tokens */
113 	psched_time_t	t_c;		/* Time check-point */
114 	struct Qdisc	*qdisc;		/* Inner qdisc, default - bfifo queue */
115 	struct qdisc_watchdog watchdog;	/* Watchdog timer */
116 };
117 
118 #define L2T(q,L)   qdisc_l2t((q)->R_tab,L)
119 #define L2T_P(q,L) qdisc_l2t((q)->P_tab,L)
120 
121 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc* sch)
122 {
123 	struct tbf_sched_data *q = qdisc_priv(sch);
124 	int ret;
125 
126 	if (qdisc_pkt_len(skb) > q->max_size)
127 		return qdisc_reshape_fail(skb, sch);
128 
129 	ret = qdisc_enqueue(skb, q->qdisc);
130 	if (ret != 0) {
131 		if (net_xmit_drop_count(ret))
132 			sch->qstats.drops++;
133 		return ret;
134 	}
135 
136 	sch->q.qlen++;
137 	sch->bstats.bytes += qdisc_pkt_len(skb);
138 	sch->bstats.packets++;
139 	return 0;
140 }
141 
142 static unsigned int tbf_drop(struct Qdisc* sch)
143 {
144 	struct tbf_sched_data *q = qdisc_priv(sch);
145 	unsigned int len = 0;
146 
147 	if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
148 		sch->q.qlen--;
149 		sch->qstats.drops++;
150 	}
151 	return len;
152 }
153 
154 static struct sk_buff *tbf_dequeue(struct Qdisc* sch)
155 {
156 	struct tbf_sched_data *q = qdisc_priv(sch);
157 	struct sk_buff *skb;
158 
159 	skb = q->qdisc->ops->peek(q->qdisc);
160 
161 	if (skb) {
162 		psched_time_t now;
163 		long toks;
164 		long ptoks = 0;
165 		unsigned int len = qdisc_pkt_len(skb);
166 
167 		now = psched_get_time();
168 		toks = psched_tdiff_bounded(now, q->t_c, q->buffer);
169 
170 		if (q->P_tab) {
171 			ptoks = toks + q->ptokens;
172 			if (ptoks > (long)q->mtu)
173 				ptoks = q->mtu;
174 			ptoks -= L2T_P(q, len);
175 		}
176 		toks += q->tokens;
177 		if (toks > (long)q->buffer)
178 			toks = q->buffer;
179 		toks -= L2T(q, len);
180 
181 		if ((toks|ptoks) >= 0) {
182 			skb = qdisc_dequeue_peeked(q->qdisc);
183 			if (unlikely(!skb))
184 				return NULL;
185 
186 			q->t_c = now;
187 			q->tokens = toks;
188 			q->ptokens = ptoks;
189 			sch->q.qlen--;
190 			sch->flags &= ~TCQ_F_THROTTLED;
191 			return skb;
192 		}
193 
194 		qdisc_watchdog_schedule(&q->watchdog,
195 					now + max_t(long, -toks, -ptoks));
196 
197 		/* Maybe we have a shorter packet in the queue,
198 		   which can be sent now. It sounds cool,
199 		   but, however, this is wrong in principle.
200 		   We MUST NOT reorder packets under these circumstances.
201 
202 		   Really, if we split the flow into independent
203 		   subflows, it would be a very good solution.
204 		   This is the main idea of all FQ algorithms
205 		   (cf. CSZ, HPFQ, HFSC)
206 		 */
207 
208 		sch->qstats.overlimits++;
209 	}
210 	return NULL;
211 }
212 
213 static void tbf_reset(struct Qdisc* sch)
214 {
215 	struct tbf_sched_data *q = qdisc_priv(sch);
216 
217 	qdisc_reset(q->qdisc);
218 	sch->q.qlen = 0;
219 	q->t_c = psched_get_time();
220 	q->tokens = q->buffer;
221 	q->ptokens = q->mtu;
222 	qdisc_watchdog_cancel(&q->watchdog);
223 }
224 
225 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
226 	[TCA_TBF_PARMS]	= { .len = sizeof(struct tc_tbf_qopt) },
227 	[TCA_TBF_RTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
228 	[TCA_TBF_PTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
229 };
230 
231 static int tbf_change(struct Qdisc* sch, struct nlattr *opt)
232 {
233 	int err;
234 	struct tbf_sched_data *q = qdisc_priv(sch);
235 	struct nlattr *tb[TCA_TBF_PTAB + 1];
236 	struct tc_tbf_qopt *qopt;
237 	struct qdisc_rate_table *rtab = NULL;
238 	struct qdisc_rate_table *ptab = NULL;
239 	struct Qdisc *child = NULL;
240 	int max_size,n;
241 
242 	err = nla_parse_nested(tb, TCA_TBF_PTAB, opt, tbf_policy);
243 	if (err < 0)
244 		return err;
245 
246 	err = -EINVAL;
247 	if (tb[TCA_TBF_PARMS] == NULL)
248 		goto done;
249 
250 	qopt = nla_data(tb[TCA_TBF_PARMS]);
251 	rtab = qdisc_get_rtab(&qopt->rate, tb[TCA_TBF_RTAB]);
252 	if (rtab == NULL)
253 		goto done;
254 
255 	if (qopt->peakrate.rate) {
256 		if (qopt->peakrate.rate > qopt->rate.rate)
257 			ptab = qdisc_get_rtab(&qopt->peakrate, tb[TCA_TBF_PTAB]);
258 		if (ptab == NULL)
259 			goto done;
260 	}
261 
262 	for (n = 0; n < 256; n++)
263 		if (rtab->data[n] > qopt->buffer) break;
264 	max_size = (n << qopt->rate.cell_log)-1;
265 	if (ptab) {
266 		int size;
267 
268 		for (n = 0; n < 256; n++)
269 			if (ptab->data[n] > qopt->mtu) break;
270 		size = (n << qopt->peakrate.cell_log)-1;
271 		if (size < max_size) max_size = size;
272 	}
273 	if (max_size < 0)
274 		goto done;
275 
276 	if (qopt->limit > 0) {
277 		child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
278 		if (IS_ERR(child)) {
279 			err = PTR_ERR(child);
280 			goto done;
281 		}
282 	}
283 
284 	sch_tree_lock(sch);
285 	if (child) {
286 		qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
287 		qdisc_destroy(q->qdisc);
288 		q->qdisc = child;
289 	}
290 	q->limit = qopt->limit;
291 	q->mtu = qopt->mtu;
292 	q->max_size = max_size;
293 	q->buffer = qopt->buffer;
294 	q->tokens = q->buffer;
295 	q->ptokens = q->mtu;
296 
297 	swap(q->R_tab, rtab);
298 	swap(q->P_tab, ptab);
299 
300 	sch_tree_unlock(sch);
301 	err = 0;
302 done:
303 	if (rtab)
304 		qdisc_put_rtab(rtab);
305 	if (ptab)
306 		qdisc_put_rtab(ptab);
307 	return err;
308 }
309 
310 static int tbf_init(struct Qdisc* sch, struct nlattr *opt)
311 {
312 	struct tbf_sched_data *q = qdisc_priv(sch);
313 
314 	if (opt == NULL)
315 		return -EINVAL;
316 
317 	q->t_c = psched_get_time();
318 	qdisc_watchdog_init(&q->watchdog, sch);
319 	q->qdisc = &noop_qdisc;
320 
321 	return tbf_change(sch, opt);
322 }
323 
324 static void tbf_destroy(struct Qdisc *sch)
325 {
326 	struct tbf_sched_data *q = qdisc_priv(sch);
327 
328 	qdisc_watchdog_cancel(&q->watchdog);
329 
330 	if (q->P_tab)
331 		qdisc_put_rtab(q->P_tab);
332 	if (q->R_tab)
333 		qdisc_put_rtab(q->R_tab);
334 
335 	qdisc_destroy(q->qdisc);
336 }
337 
338 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
339 {
340 	struct tbf_sched_data *q = qdisc_priv(sch);
341 	struct nlattr *nest;
342 	struct tc_tbf_qopt opt;
343 
344 	nest = nla_nest_start(skb, TCA_OPTIONS);
345 	if (nest == NULL)
346 		goto nla_put_failure;
347 
348 	opt.limit = q->limit;
349 	opt.rate = q->R_tab->rate;
350 	if (q->P_tab)
351 		opt.peakrate = q->P_tab->rate;
352 	else
353 		memset(&opt.peakrate, 0, sizeof(opt.peakrate));
354 	opt.mtu = q->mtu;
355 	opt.buffer = q->buffer;
356 	NLA_PUT(skb, TCA_TBF_PARMS, sizeof(opt), &opt);
357 
358 	nla_nest_end(skb, nest);
359 	return skb->len;
360 
361 nla_put_failure:
362 	nla_nest_cancel(skb, nest);
363 	return -1;
364 }
365 
366 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
367 			  struct sk_buff *skb, struct tcmsg *tcm)
368 {
369 	struct tbf_sched_data *q = qdisc_priv(sch);
370 
371 	tcm->tcm_handle |= TC_H_MIN(1);
372 	tcm->tcm_info = q->qdisc->handle;
373 
374 	return 0;
375 }
376 
377 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
378 		     struct Qdisc **old)
379 {
380 	struct tbf_sched_data *q = qdisc_priv(sch);
381 
382 	if (new == NULL)
383 		new = &noop_qdisc;
384 
385 	sch_tree_lock(sch);
386 	*old = q->qdisc;
387 	q->qdisc = new;
388 	qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
389 	qdisc_reset(*old);
390 	sch_tree_unlock(sch);
391 
392 	return 0;
393 }
394 
395 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
396 {
397 	struct tbf_sched_data *q = qdisc_priv(sch);
398 	return q->qdisc;
399 }
400 
401 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
402 {
403 	return 1;
404 }
405 
406 static void tbf_put(struct Qdisc *sch, unsigned long arg)
407 {
408 }
409 
410 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
411 {
412 	if (!walker->stop) {
413 		if (walker->count >= walker->skip)
414 			if (walker->fn(sch, 1, walker) < 0) {
415 				walker->stop = 1;
416 				return;
417 			}
418 		walker->count++;
419 	}
420 }
421 
422 static const struct Qdisc_class_ops tbf_class_ops =
423 {
424 	.graft		=	tbf_graft,
425 	.leaf		=	tbf_leaf,
426 	.get		=	tbf_get,
427 	.put		=	tbf_put,
428 	.walk		=	tbf_walk,
429 	.dump		=	tbf_dump_class,
430 };
431 
432 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
433 	.next		=	NULL,
434 	.cl_ops		=	&tbf_class_ops,
435 	.id		=	"tbf",
436 	.priv_size	=	sizeof(struct tbf_sched_data),
437 	.enqueue	=	tbf_enqueue,
438 	.dequeue	=	tbf_dequeue,
439 	.peek		=	qdisc_peek_dequeued,
440 	.drop		=	tbf_drop,
441 	.init		=	tbf_init,
442 	.reset		=	tbf_reset,
443 	.destroy	=	tbf_destroy,
444 	.change		=	tbf_change,
445 	.dump		=	tbf_dump,
446 	.owner		=	THIS_MODULE,
447 };
448 
449 static int __init tbf_module_init(void)
450 {
451 	return register_qdisc(&tbf_qdisc_ops);
452 }
453 
454 static void __exit tbf_module_exit(void)
455 {
456 	unregister_qdisc(&tbf_qdisc_ops);
457 }
458 module_init(tbf_module_init)
459 module_exit(tbf_module_exit)
460 MODULE_LICENSE("GPL");
461