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/* GSO packet is too big, segment it so that tbf can transmit
146 * each segment in time
147 */
148static 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 segs->next = NULL;
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
184static 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 int ret;
189
190 if (qdisc_pkt_len(skb) > q->max_size) {
191 if (skb_is_gso(skb) &&
192 skb_gso_validate_mac_len(skb, q->max_size))
193 return tbf_segment(skb, sch, to_free);
194 return qdisc_drop(skb, sch, to_free);
195 }
196 ret = qdisc_enqueue(skb, q->qdisc, to_free);
197 if (ret != NET_XMIT_SUCCESS) {
198 if (net_xmit_drop_count(ret))
199 qdisc_qstats_drop(sch);
200 return ret;
201 }
202
203 qdisc_qstats_backlog_inc(sch, skb);
204 sch->q.qlen++;
205 return NET_XMIT_SUCCESS;
206}
207
208static bool tbf_peak_present(const struct tbf_sched_data *q)
209{
210 return q->peak.rate_bytes_ps;
211}
212
213static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
214{
215 struct tbf_sched_data *q = qdisc_priv(sch);
216 struct sk_buff *skb;
217
218 skb = q->qdisc->ops->peek(q->qdisc);
219
220 if (skb) {
221 s64 now;
222 s64 toks;
223 s64 ptoks = 0;
224 unsigned int len = qdisc_pkt_len(skb);
225
226 now = ktime_get_ns();
227 toks = min_t(s64, now - q->t_c, q->buffer);
228
229 if (tbf_peak_present(q)) {
230 ptoks = toks + q->ptokens;
231 if (ptoks > q->mtu)
232 ptoks = q->mtu;
233 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
234 }
235 toks += q->tokens;
236 if (toks > q->buffer)
237 toks = q->buffer;
238 toks -= (s64) psched_l2t_ns(&q->rate, len);
239
240 if ((toks|ptoks) >= 0) {
241 skb = qdisc_dequeue_peeked(q->qdisc);
242 if (unlikely(!skb))
243 return NULL;
244
245 q->t_c = now;
246 q->tokens = toks;
247 q->ptokens = ptoks;
248 qdisc_qstats_backlog_dec(sch, skb);
249 sch->q.qlen--;
250 qdisc_bstats_update(sch, skb);
251 return skb;
252 }
253
254 qdisc_watchdog_schedule_ns(&q->watchdog,
255 now + max_t(long, -toks, -ptoks));
256
257 /* Maybe we have a shorter packet in the queue,
258 which can be sent now. It sounds cool,
259 but, however, this is wrong in principle.
260 We MUST NOT reorder packets under these circumstances.
261
262 Really, if we split the flow into independent
263 subflows, it would be a very good solution.
264 This is the main idea of all FQ algorithms
265 (cf. CSZ, HPFQ, HFSC)
266 */
267
268 qdisc_qstats_overlimit(sch);
269 }
270 return NULL;
271}
272
273static void tbf_reset(struct Qdisc *sch)
274{
275 struct tbf_sched_data *q = qdisc_priv(sch);
276
277 qdisc_reset(q->qdisc);
278 sch->qstats.backlog = 0;
279 sch->q.qlen = 0;
280 q->t_c = ktime_get_ns();
281 q->tokens = q->buffer;
282 q->ptokens = q->mtu;
283 qdisc_watchdog_cancel(&q->watchdog);
284}
285
286static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
287 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
288 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
289 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
290 [TCA_TBF_RATE64] = { .type = NLA_U64 },
291 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
292 [TCA_TBF_BURST] = { .type = NLA_U32 },
293 [TCA_TBF_PBURST] = { .type = NLA_U32 },
294};
295
296static int tbf_change(struct Qdisc *sch, struct nlattr *opt,
297 struct netlink_ext_ack *extack)
298{
299 int err;
300 struct tbf_sched_data *q = qdisc_priv(sch);
301 struct nlattr *tb[TCA_TBF_MAX + 1];
302 struct tc_tbf_qopt *qopt;
303 struct Qdisc *child = NULL;
304 struct psched_ratecfg rate;
305 struct psched_ratecfg peak;
306 u64 max_size;
307 s64 buffer, mtu;
308 u64 rate64 = 0, prate64 = 0;
309
310 err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy, NULL);
311 if (err < 0)
312 return err;
313
314 err = -EINVAL;
315 if (tb[TCA_TBF_PARMS] == NULL)
316 goto done;
317
318 qopt = nla_data(tb[TCA_TBF_PARMS]);
319 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
320 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
321 tb[TCA_TBF_RTAB],
322 NULL));
323
324 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
325 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
326 tb[TCA_TBF_PTAB],
327 NULL));
328
329 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
330 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
331
332 if (tb[TCA_TBF_RATE64])
333 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
334 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
335
336 if (tb[TCA_TBF_BURST]) {
337 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
338 buffer = psched_l2t_ns(&rate, max_size);
339 } else {
340 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
341 }
342
343 if (qopt->peakrate.rate) {
344 if (tb[TCA_TBF_PRATE64])
345 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
346 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
347 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
348 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
349 peak.rate_bytes_ps, rate.rate_bytes_ps);
350 err = -EINVAL;
351 goto done;
352 }
353
354 if (tb[TCA_TBF_PBURST]) {
355 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
356 max_size = min_t(u32, max_size, pburst);
357 mtu = psched_l2t_ns(&peak, pburst);
358 } else {
359 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
360 }
361 } else {
362 memset(&peak, 0, sizeof(peak));
363 }
364
365 if (max_size < psched_mtu(qdisc_dev(sch)))
366 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
367 max_size, qdisc_dev(sch)->name,
368 psched_mtu(qdisc_dev(sch)));
369
370 if (!max_size) {
371 err = -EINVAL;
372 goto done;
373 }
374
375 if (q->qdisc != &noop_qdisc) {
376 err = fifo_set_limit(q->qdisc, qopt->limit);
377 if (err)
378 goto done;
379 } else if (qopt->limit > 0) {
380 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit,
381 extack);
382 if (IS_ERR(child)) {
383 err = PTR_ERR(child);
384 goto done;
385 }
386
387 /* child is fifo, no need to check for noop_qdisc */
388 qdisc_hash_add(child, true);
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;
416done:
417 return err;
418}
419
420static int tbf_init(struct Qdisc *sch, struct nlattr *opt,
421 struct netlink_ext_ack *extack)
422{
423 struct tbf_sched_data *q = qdisc_priv(sch);
424
425 qdisc_watchdog_init(&q->watchdog, sch);
426 q->qdisc = &noop_qdisc;
427
428 if (!opt)
429 return -EINVAL;
430
431 q->t_c = ktime_get_ns();
432
433 return tbf_change(sch, opt, extack);
434}
435
436static void tbf_destroy(struct Qdisc *sch)
437{
438 struct tbf_sched_data *q = qdisc_priv(sch);
439
440 qdisc_watchdog_cancel(&q->watchdog);
441 qdisc_destroy(q->qdisc);
442}
443
444static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
445{
446 struct tbf_sched_data *q = qdisc_priv(sch);
447 struct nlattr *nest;
448 struct tc_tbf_qopt opt;
449
450 sch->qstats.backlog = q->qdisc->qstats.backlog;
451 nest = nla_nest_start(skb, TCA_OPTIONS);
452 if (nest == NULL)
453 goto nla_put_failure;
454
455 opt.limit = q->limit;
456 psched_ratecfg_getrate(&opt.rate, &q->rate);
457 if (tbf_peak_present(q))
458 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
459 else
460 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
461 opt.mtu = PSCHED_NS2TICKS(q->mtu);
462 opt.buffer = PSCHED_NS2TICKS(q->buffer);
463 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
464 goto nla_put_failure;
465 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
466 nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
467 TCA_TBF_PAD))
468 goto nla_put_failure;
469 if (tbf_peak_present(q) &&
470 q->peak.rate_bytes_ps >= (1ULL << 32) &&
471 nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
472 TCA_TBF_PAD))
473 goto nla_put_failure;
474
475 return nla_nest_end(skb, nest);
476
477nla_put_failure:
478 nla_nest_cancel(skb, nest);
479 return -1;
480}
481
482static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
483 struct sk_buff *skb, struct tcmsg *tcm)
484{
485 struct tbf_sched_data *q = qdisc_priv(sch);
486
487 tcm->tcm_handle |= TC_H_MIN(1);
488 tcm->tcm_info = q->qdisc->handle;
489
490 return 0;
491}
492
493static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
494 struct Qdisc **old, struct netlink_ext_ack *extack)
495{
496 struct tbf_sched_data *q = qdisc_priv(sch);
497
498 if (new == NULL)
499 new = &noop_qdisc;
500
501 *old = qdisc_replace(sch, new, &q->qdisc);
502 return 0;
503}
504
505static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
506{
507 struct tbf_sched_data *q = qdisc_priv(sch);
508 return q->qdisc;
509}
510
511static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
512{
513 return 1;
514}
515
516static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
517{
518 if (!walker->stop) {
519 if (walker->count >= walker->skip)
520 if (walker->fn(sch, 1, walker) < 0) {
521 walker->stop = 1;
522 return;
523 }
524 walker->count++;
525 }
526}
527
528static const struct Qdisc_class_ops tbf_class_ops = {
529 .graft = tbf_graft,
530 .leaf = tbf_leaf,
531 .find = tbf_find,
532 .walk = tbf_walk,
533 .dump = tbf_dump_class,
534};
535
536static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
537 .next = NULL,
538 .cl_ops = &tbf_class_ops,
539 .id = "tbf",
540 .priv_size = sizeof(struct tbf_sched_data),
541 .enqueue = tbf_enqueue,
542 .dequeue = tbf_dequeue,
543 .peek = qdisc_peek_dequeued,
544 .init = tbf_init,
545 .reset = tbf_reset,
546 .destroy = tbf_destroy,
547 .change = tbf_change,
548 .dump = tbf_dump,
549 .owner = THIS_MODULE,
550};
551
552static int __init tbf_module_init(void)
553{
554 return register_qdisc(&tbf_qdisc_ops);
555}
556
557static void __exit tbf_module_exit(void)
558{
559 unregister_qdisc(&tbf_qdisc_ops);
560}
561module_init(tbf_module_init)
562module_exit(tbf_module_exit)
563MODULE_LICENSE("GPL");
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
100struct tbf_sched_data {
101/* Parameters */
102 u32 limit; /* Maximal length of backlog: bytes */
103 u32 buffer; /* Token bucket depth/rate: MUST BE >= MTU/B */
104 u32 mtu;
105 u32 max_size;
106 struct qdisc_rate_table *R_tab;
107 struct qdisc_rate_table *P_tab;
108
109/* Variables */
110 long tokens; /* Current number of B tokens */
111 long ptokens; /* Current number of P tokens */
112 psched_time_t t_c; /* Time check-point */
113 struct Qdisc *qdisc; /* Inner qdisc, default - bfifo queue */
114 struct qdisc_watchdog watchdog; /* Watchdog timer */
115};
116
117#define L2T(q, L) qdisc_l2t((q)->R_tab, L)
118#define L2T_P(q, L) qdisc_l2t((q)->P_tab, L)
119
120static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch)
121{
122 struct tbf_sched_data *q = qdisc_priv(sch);
123 int ret;
124
125 if (qdisc_pkt_len(skb) > q->max_size)
126 return qdisc_reshape_fail(skb, sch);
127
128 ret = qdisc_enqueue(skb, q->qdisc);
129 if (ret != NET_XMIT_SUCCESS) {
130 if (net_xmit_drop_count(ret))
131 sch->qstats.drops++;
132 return ret;
133 }
134
135 sch->q.qlen++;
136 return NET_XMIT_SUCCESS;
137}
138
139static unsigned int tbf_drop(struct Qdisc *sch)
140{
141 struct tbf_sched_data *q = qdisc_priv(sch);
142 unsigned int len = 0;
143
144 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
145 sch->q.qlen--;
146 sch->qstats.drops++;
147 }
148 return len;
149}
150
151static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
152{
153 struct tbf_sched_data *q = qdisc_priv(sch);
154 struct sk_buff *skb;
155
156 skb = q->qdisc->ops->peek(q->qdisc);
157
158 if (skb) {
159 psched_time_t now;
160 long toks;
161 long ptoks = 0;
162 unsigned int len = qdisc_pkt_len(skb);
163
164 now = psched_get_time();
165 toks = psched_tdiff_bounded(now, q->t_c, q->buffer);
166
167 if (q->P_tab) {
168 ptoks = toks + q->ptokens;
169 if (ptoks > (long)q->mtu)
170 ptoks = q->mtu;
171 ptoks -= L2T_P(q, len);
172 }
173 toks += q->tokens;
174 if (toks > (long)q->buffer)
175 toks = q->buffer;
176 toks -= L2T(q, len);
177
178 if ((toks|ptoks) >= 0) {
179 skb = qdisc_dequeue_peeked(q->qdisc);
180 if (unlikely(!skb))
181 return NULL;
182
183 q->t_c = now;
184 q->tokens = toks;
185 q->ptokens = ptoks;
186 sch->q.qlen--;
187 qdisc_unthrottled(sch);
188 qdisc_bstats_update(sch, skb);
189 return skb;
190 }
191
192 qdisc_watchdog_schedule(&q->watchdog,
193 now + max_t(long, -toks, -ptoks));
194
195 /* Maybe we have a shorter packet in the queue,
196 which can be sent now. It sounds cool,
197 but, however, this is wrong in principle.
198 We MUST NOT reorder packets under these circumstances.
199
200 Really, if we split the flow into independent
201 subflows, it would be a very good solution.
202 This is the main idea of all FQ algorithms
203 (cf. CSZ, HPFQ, HFSC)
204 */
205
206 sch->qstats.overlimits++;
207 }
208 return NULL;
209}
210
211static void tbf_reset(struct Qdisc *sch)
212{
213 struct tbf_sched_data *q = qdisc_priv(sch);
214
215 qdisc_reset(q->qdisc);
216 sch->q.qlen = 0;
217 q->t_c = psched_get_time();
218 q->tokens = q->buffer;
219 q->ptokens = q->mtu;
220 qdisc_watchdog_cancel(&q->watchdog);
221}
222
223static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
224 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
225 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
226 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
227};
228
229static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
230{
231 int err;
232 struct tbf_sched_data *q = qdisc_priv(sch);
233 struct nlattr *tb[TCA_TBF_PTAB + 1];
234 struct tc_tbf_qopt *qopt;
235 struct qdisc_rate_table *rtab = NULL;
236 struct qdisc_rate_table *ptab = NULL;
237 struct Qdisc *child = NULL;
238 int max_size, n;
239
240 err = nla_parse_nested(tb, TCA_TBF_PTAB, opt, tbf_policy);
241 if (err < 0)
242 return err;
243
244 err = -EINVAL;
245 if (tb[TCA_TBF_PARMS] == NULL)
246 goto done;
247
248 qopt = nla_data(tb[TCA_TBF_PARMS]);
249 rtab = qdisc_get_rtab(&qopt->rate, tb[TCA_TBF_RTAB]);
250 if (rtab == NULL)
251 goto done;
252
253 if (qopt->peakrate.rate) {
254 if (qopt->peakrate.rate > qopt->rate.rate)
255 ptab = qdisc_get_rtab(&qopt->peakrate, tb[TCA_TBF_PTAB]);
256 if (ptab == NULL)
257 goto done;
258 }
259
260 for (n = 0; n < 256; n++)
261 if (rtab->data[n] > qopt->buffer)
262 break;
263 max_size = (n << qopt->rate.cell_log) - 1;
264 if (ptab) {
265 int size;
266
267 for (n = 0; n < 256; n++)
268 if (ptab->data[n] > qopt->mtu)
269 break;
270 size = (n << qopt->peakrate.cell_log) - 1;
271 if (size < max_size)
272 max_size = size;
273 }
274 if (max_size < 0)
275 goto done;
276
277 if (q->qdisc != &noop_qdisc) {
278 err = fifo_set_limit(q->qdisc, qopt->limit);
279 if (err)
280 goto done;
281 } else if (qopt->limit > 0) {
282 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
283 if (IS_ERR(child)) {
284 err = PTR_ERR(child);
285 goto done;
286 }
287 }
288
289 sch_tree_lock(sch);
290 if (child) {
291 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
292 qdisc_destroy(q->qdisc);
293 q->qdisc = child;
294 }
295 q->limit = qopt->limit;
296 q->mtu = qopt->mtu;
297 q->max_size = max_size;
298 q->buffer = qopt->buffer;
299 q->tokens = q->buffer;
300 q->ptokens = q->mtu;
301
302 swap(q->R_tab, rtab);
303 swap(q->P_tab, ptab);
304
305 sch_tree_unlock(sch);
306 err = 0;
307done:
308 if (rtab)
309 qdisc_put_rtab(rtab);
310 if (ptab)
311 qdisc_put_rtab(ptab);
312 return err;
313}
314
315static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
316{
317 struct tbf_sched_data *q = qdisc_priv(sch);
318
319 if (opt == NULL)
320 return -EINVAL;
321
322 q->t_c = psched_get_time();
323 qdisc_watchdog_init(&q->watchdog, sch);
324 q->qdisc = &noop_qdisc;
325
326 return tbf_change(sch, opt);
327}
328
329static void tbf_destroy(struct Qdisc *sch)
330{
331 struct tbf_sched_data *q = qdisc_priv(sch);
332
333 qdisc_watchdog_cancel(&q->watchdog);
334
335 if (q->P_tab)
336 qdisc_put_rtab(q->P_tab);
337 if (q->R_tab)
338 qdisc_put_rtab(q->R_tab);
339
340 qdisc_destroy(q->qdisc);
341}
342
343static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
344{
345 struct tbf_sched_data *q = qdisc_priv(sch);
346 struct nlattr *nest;
347 struct tc_tbf_qopt opt;
348
349 nest = nla_nest_start(skb, TCA_OPTIONS);
350 if (nest == NULL)
351 goto nla_put_failure;
352
353 opt.limit = q->limit;
354 opt.rate = q->R_tab->rate;
355 if (q->P_tab)
356 opt.peakrate = q->P_tab->rate;
357 else
358 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
359 opt.mtu = q->mtu;
360 opt.buffer = q->buffer;
361 NLA_PUT(skb, TCA_TBF_PARMS, sizeof(opt), &opt);
362
363 nla_nest_end(skb, nest);
364 return skb->len;
365
366nla_put_failure:
367 nla_nest_cancel(skb, nest);
368 return -1;
369}
370
371static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
372 struct sk_buff *skb, struct tcmsg *tcm)
373{
374 struct tbf_sched_data *q = qdisc_priv(sch);
375
376 tcm->tcm_handle |= TC_H_MIN(1);
377 tcm->tcm_info = q->qdisc->handle;
378
379 return 0;
380}
381
382static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
383 struct Qdisc **old)
384{
385 struct tbf_sched_data *q = qdisc_priv(sch);
386
387 if (new == NULL)
388 new = &noop_qdisc;
389
390 sch_tree_lock(sch);
391 *old = q->qdisc;
392 q->qdisc = new;
393 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
394 qdisc_reset(*old);
395 sch_tree_unlock(sch);
396
397 return 0;
398}
399
400static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
401{
402 struct tbf_sched_data *q = qdisc_priv(sch);
403 return q->qdisc;
404}
405
406static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
407{
408 return 1;
409}
410
411static void tbf_put(struct Qdisc *sch, unsigned long arg)
412{
413}
414
415static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
416{
417 if (!walker->stop) {
418 if (walker->count >= walker->skip)
419 if (walker->fn(sch, 1, walker) < 0) {
420 walker->stop = 1;
421 return;
422 }
423 walker->count++;
424 }
425}
426
427static const struct Qdisc_class_ops tbf_class_ops = {
428 .graft = tbf_graft,
429 .leaf = tbf_leaf,
430 .get = tbf_get,
431 .put = tbf_put,
432 .walk = tbf_walk,
433 .dump = tbf_dump_class,
434};
435
436static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
437 .next = NULL,
438 .cl_ops = &tbf_class_ops,
439 .id = "tbf",
440 .priv_size = sizeof(struct tbf_sched_data),
441 .enqueue = tbf_enqueue,
442 .dequeue = tbf_dequeue,
443 .peek = qdisc_peek_dequeued,
444 .drop = tbf_drop,
445 .init = tbf_init,
446 .reset = tbf_reset,
447 .destroy = tbf_destroy,
448 .change = tbf_change,
449 .dump = tbf_dump,
450 .owner = THIS_MODULE,
451};
452
453static int __init tbf_module_init(void)
454{
455 return register_qdisc(&tbf_qdisc_ops);
456}
457
458static void __exit tbf_module_exit(void)
459{
460 unregister_qdisc(&tbf_qdisc_ops);
461}
462module_init(tbf_module_init)
463module_exit(tbf_module_exit)
464MODULE_LICENSE("GPL");