Loading...
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
101struct 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 */
122static 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 */
149static 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 */
158static 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_reshape_fail(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
193static 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_reshape_fail(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 sch->q.qlen++;
211 return NET_XMIT_SUCCESS;
212}
213
214static unsigned int tbf_drop(struct Qdisc *sch)
215{
216 struct tbf_sched_data *q = qdisc_priv(sch);
217 unsigned int len = 0;
218
219 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
220 sch->q.qlen--;
221 qdisc_qstats_drop(sch);
222 }
223 return len;
224}
225
226static bool tbf_peak_present(const struct tbf_sched_data *q)
227{
228 return q->peak.rate_bytes_ps;
229}
230
231static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
232{
233 struct tbf_sched_data *q = qdisc_priv(sch);
234 struct sk_buff *skb;
235
236 skb = q->qdisc->ops->peek(q->qdisc);
237
238 if (skb) {
239 s64 now;
240 s64 toks;
241 s64 ptoks = 0;
242 unsigned int len = qdisc_pkt_len(skb);
243
244 now = ktime_get_ns();
245 toks = min_t(s64, now - q->t_c, q->buffer);
246
247 if (tbf_peak_present(q)) {
248 ptoks = toks + q->ptokens;
249 if (ptoks > q->mtu)
250 ptoks = q->mtu;
251 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
252 }
253 toks += q->tokens;
254 if (toks > q->buffer)
255 toks = q->buffer;
256 toks -= (s64) psched_l2t_ns(&q->rate, len);
257
258 if ((toks|ptoks) >= 0) {
259 skb = qdisc_dequeue_peeked(q->qdisc);
260 if (unlikely(!skb))
261 return NULL;
262
263 q->t_c = now;
264 q->tokens = toks;
265 q->ptokens = ptoks;
266 sch->q.qlen--;
267 qdisc_unthrottled(sch);
268 qdisc_bstats_update(sch, skb);
269 return skb;
270 }
271
272 qdisc_watchdog_schedule_ns(&q->watchdog,
273 now + max_t(long, -toks, -ptoks),
274 true);
275
276 /* Maybe we have a shorter packet in the queue,
277 which can be sent now. It sounds cool,
278 but, however, this is wrong in principle.
279 We MUST NOT reorder packets under these circumstances.
280
281 Really, if we split the flow into independent
282 subflows, it would be a very good solution.
283 This is the main idea of all FQ algorithms
284 (cf. CSZ, HPFQ, HFSC)
285 */
286
287 qdisc_qstats_overlimit(sch);
288 }
289 return NULL;
290}
291
292static void tbf_reset(struct Qdisc *sch)
293{
294 struct tbf_sched_data *q = qdisc_priv(sch);
295
296 qdisc_reset(q->qdisc);
297 sch->q.qlen = 0;
298 q->t_c = ktime_get_ns();
299 q->tokens = q->buffer;
300 q->ptokens = q->mtu;
301 qdisc_watchdog_cancel(&q->watchdog);
302}
303
304static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
305 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
306 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
307 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
308 [TCA_TBF_RATE64] = { .type = NLA_U64 },
309 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
310 [TCA_TBF_BURST] = { .type = NLA_U32 },
311 [TCA_TBF_PBURST] = { .type = NLA_U32 },
312};
313
314static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
315{
316 int err;
317 struct tbf_sched_data *q = qdisc_priv(sch);
318 struct nlattr *tb[TCA_TBF_MAX + 1];
319 struct tc_tbf_qopt *qopt;
320 struct Qdisc *child = NULL;
321 struct psched_ratecfg rate;
322 struct psched_ratecfg peak;
323 u64 max_size;
324 s64 buffer, mtu;
325 u64 rate64 = 0, prate64 = 0;
326
327 err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy);
328 if (err < 0)
329 return err;
330
331 err = -EINVAL;
332 if (tb[TCA_TBF_PARMS] == NULL)
333 goto done;
334
335 qopt = nla_data(tb[TCA_TBF_PARMS]);
336 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
337 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
338 tb[TCA_TBF_RTAB]));
339
340 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
341 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
342 tb[TCA_TBF_PTAB]));
343
344 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
345 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
346
347 if (tb[TCA_TBF_RATE64])
348 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
349 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
350
351 if (tb[TCA_TBF_BURST]) {
352 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
353 buffer = psched_l2t_ns(&rate, max_size);
354 } else {
355 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
356 }
357
358 if (qopt->peakrate.rate) {
359 if (tb[TCA_TBF_PRATE64])
360 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
361 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
362 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
363 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
364 peak.rate_bytes_ps, rate.rate_bytes_ps);
365 err = -EINVAL;
366 goto done;
367 }
368
369 if (tb[TCA_TBF_PBURST]) {
370 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
371 max_size = min_t(u32, max_size, pburst);
372 mtu = psched_l2t_ns(&peak, pburst);
373 } else {
374 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
375 }
376 } else {
377 memset(&peak, 0, sizeof(peak));
378 }
379
380 if (max_size < psched_mtu(qdisc_dev(sch)))
381 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
382 max_size, qdisc_dev(sch)->name,
383 psched_mtu(qdisc_dev(sch)));
384
385 if (!max_size) {
386 err = -EINVAL;
387 goto done;
388 }
389
390 if (q->qdisc != &noop_qdisc) {
391 err = fifo_set_limit(q->qdisc, qopt->limit);
392 if (err)
393 goto done;
394 } else if (qopt->limit > 0) {
395 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
396 if (IS_ERR(child)) {
397 err = PTR_ERR(child);
398 goto done;
399 }
400 }
401
402 sch_tree_lock(sch);
403 if (child) {
404 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
405 q->qdisc->qstats.backlog);
406 qdisc_destroy(q->qdisc);
407 q->qdisc = child;
408 }
409 q->limit = qopt->limit;
410 if (tb[TCA_TBF_PBURST])
411 q->mtu = mtu;
412 else
413 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
414 q->max_size = max_size;
415 if (tb[TCA_TBF_BURST])
416 q->buffer = buffer;
417 else
418 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
419 q->tokens = q->buffer;
420 q->ptokens = q->mtu;
421
422 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
423 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
424
425 sch_tree_unlock(sch);
426 err = 0;
427done:
428 return err;
429}
430
431static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
432{
433 struct tbf_sched_data *q = qdisc_priv(sch);
434
435 if (opt == NULL)
436 return -EINVAL;
437
438 q->t_c = ktime_get_ns();
439 qdisc_watchdog_init(&q->watchdog, sch);
440 q->qdisc = &noop_qdisc;
441
442 return tbf_change(sch, opt);
443}
444
445static void tbf_destroy(struct Qdisc *sch)
446{
447 struct tbf_sched_data *q = qdisc_priv(sch);
448
449 qdisc_watchdog_cancel(&q->watchdog);
450 qdisc_destroy(q->qdisc);
451}
452
453static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
454{
455 struct tbf_sched_data *q = qdisc_priv(sch);
456 struct nlattr *nest;
457 struct tc_tbf_qopt opt;
458
459 sch->qstats.backlog = q->qdisc->qstats.backlog;
460 nest = nla_nest_start(skb, TCA_OPTIONS);
461 if (nest == NULL)
462 goto nla_put_failure;
463
464 opt.limit = q->limit;
465 psched_ratecfg_getrate(&opt.rate, &q->rate);
466 if (tbf_peak_present(q))
467 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
468 else
469 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
470 opt.mtu = PSCHED_NS2TICKS(q->mtu);
471 opt.buffer = PSCHED_NS2TICKS(q->buffer);
472 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
473 goto nla_put_failure;
474 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
475 nla_put_u64(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps))
476 goto nla_put_failure;
477 if (tbf_peak_present(q) &&
478 q->peak.rate_bytes_ps >= (1ULL << 32) &&
479 nla_put_u64(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps))
480 goto nla_put_failure;
481
482 return nla_nest_end(skb, nest);
483
484nla_put_failure:
485 nla_nest_cancel(skb, nest);
486 return -1;
487}
488
489static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
490 struct sk_buff *skb, struct tcmsg *tcm)
491{
492 struct tbf_sched_data *q = qdisc_priv(sch);
493
494 tcm->tcm_handle |= TC_H_MIN(1);
495 tcm->tcm_info = q->qdisc->handle;
496
497 return 0;
498}
499
500static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
501 struct Qdisc **old)
502{
503 struct tbf_sched_data *q = qdisc_priv(sch);
504
505 if (new == NULL)
506 new = &noop_qdisc;
507
508 *old = qdisc_replace(sch, new, &q->qdisc);
509 return 0;
510}
511
512static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
513{
514 struct tbf_sched_data *q = qdisc_priv(sch);
515 return q->qdisc;
516}
517
518static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
519{
520 return 1;
521}
522
523static void tbf_put(struct Qdisc *sch, unsigned long arg)
524{
525}
526
527static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
528{
529 if (!walker->stop) {
530 if (walker->count >= walker->skip)
531 if (walker->fn(sch, 1, walker) < 0) {
532 walker->stop = 1;
533 return;
534 }
535 walker->count++;
536 }
537}
538
539static const struct Qdisc_class_ops tbf_class_ops = {
540 .graft = tbf_graft,
541 .leaf = tbf_leaf,
542 .get = tbf_get,
543 .put = tbf_put,
544 .walk = tbf_walk,
545 .dump = tbf_dump_class,
546};
547
548static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
549 .next = NULL,
550 .cl_ops = &tbf_class_ops,
551 .id = "tbf",
552 .priv_size = sizeof(struct tbf_sched_data),
553 .enqueue = tbf_enqueue,
554 .dequeue = tbf_dequeue,
555 .peek = qdisc_peek_dequeued,
556 .drop = tbf_drop,
557 .init = tbf_init,
558 .reset = tbf_reset,
559 .destroy = tbf_destroy,
560 .change = tbf_change,
561 .dump = tbf_dump,
562 .owner = THIS_MODULE,
563};
564
565static int __init tbf_module_init(void)
566{
567 return register_qdisc(&tbf_qdisc_ops);
568}
569
570static void __exit tbf_module_exit(void)
571{
572 unregister_qdisc(&tbf_qdisc_ops);
573}
574module_init(tbf_module_init)
575module_exit(tbf_module_exit)
576MODULE_LICENSE("GPL");
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
96struct 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 */
117static 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 */
143static 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
179static 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
204static bool tbf_peak_present(const struct tbf_sched_data *q)
205{
206 return q->peak.rate_bytes_ps;
207}
208
209static 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
269static 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
282static 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
292static 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;
412done:
413 return err;
414}
415
416static 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
432static 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
440static 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
473nla_put_failure:
474 nla_nest_cancel(skb, nest);
475 return -1;
476}
477
478static 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
489static 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
501static 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
507static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
508{
509 return 1;
510}
511
512static 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
524static 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
532static 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
548static int __init tbf_module_init(void)
549{
550 return register_qdisc(&tbf_qdisc_ops);
551}
552
553static void __exit tbf_module_exit(void)
554{
555 unregister_qdisc(&tbf_qdisc_ops);
556}
557module_init(tbf_module_init)
558module_exit(tbf_module_exit)
559MODULE_LICENSE("GPL");