2 * net/sched/sch_tbf.c Token Bucket Filter queue.
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.
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
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>
26 /* Simple Token Bucket Filter.
27 =======================================
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).
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:
45 s_i+....+s_k <= B + R*(t_k - t_i)
50 Let N(t_i) be B/R initially and N(t) grow continuously with time as:
52 N(t+delta) = min{B/R, N(t) + delta}
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:
58 N(t_* + 0) = N(t_* - 0) - S/R.
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.
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.
70 When TBF works in reshaping mode, latency is estimated as:
72 lat = max ((L-B)/R, (L-M)/P)
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.
86 This means, that with depth B, the maximal rate is
90 F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
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 :-)
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.
101 struct tbf_sched_data {
103 u32 limit; /* Maximal length of backlog: bytes */
105 s64 buffer; /* Token bucket depth/rate: MUST BE >= MTU/B */
107 struct psched_ratecfg rate;
108 struct psched_ratecfg peak;
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 */
119 /* Time to Length, convert time in ns to length in bytes
120 * to determinate how many bytes can be sent in given time.
122 static u64 psched_ns_t2l(const struct psched_ratecfg *r,
126 * len = (time_in_ns * r->rate_bytes_ps) / NSEC_PER_SEC
128 u64 len = time_in_ns * r->rate_bytes_ps;
130 do_div(len, NSEC_PER_SEC);
132 if (unlikely(r->linklayer == TC_LINKLAYER_ATM)) {
137 if (len > r->overhead)
145 /* GSO packet is too big, segment it so that tbf can transmit
146 * each segment in time
148 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,
149 struct sk_buff **to_free)
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);
157 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
159 if (IS_ERR_OR_NULL(segs))
160 return qdisc_drop(skb, sch, to_free);
165 skb_mark_not_on_list(segs);
166 qdisc_skb_cb(segs)->pkt_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);
179 qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
181 return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
184 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
185 struct sk_buff **to_free)
187 struct tbf_sched_data *q = qdisc_priv(sch);
188 unsigned int len = qdisc_pkt_len(skb);
191 if (qdisc_pkt_len(skb) > q->max_size) {
192 if (skb_is_gso(skb) &&
193 skb_gso_validate_mac_len(skb, q->max_size))
194 return tbf_segment(skb, sch, to_free);
195 return qdisc_drop(skb, sch, to_free);
197 ret = qdisc_enqueue(skb, q->qdisc, to_free);
198 if (ret != NET_XMIT_SUCCESS) {
199 if (net_xmit_drop_count(ret))
200 qdisc_qstats_drop(sch);
204 sch->qstats.backlog += len;
206 return NET_XMIT_SUCCESS;
209 static bool tbf_peak_present(const struct tbf_sched_data *q)
211 return q->peak.rate_bytes_ps;
214 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
216 struct tbf_sched_data *q = qdisc_priv(sch);
219 skb = q->qdisc->ops->peek(q->qdisc);
225 unsigned int len = qdisc_pkt_len(skb);
227 now = ktime_get_ns();
228 toks = min_t(s64, now - q->t_c, q->buffer);
230 if (tbf_peak_present(q)) {
231 ptoks = toks + q->ptokens;
234 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
237 if (toks > q->buffer)
239 toks -= (s64) psched_l2t_ns(&q->rate, len);
241 if ((toks|ptoks) >= 0) {
242 skb = qdisc_dequeue_peeked(q->qdisc);
249 qdisc_qstats_backlog_dec(sch, skb);
251 qdisc_bstats_update(sch, skb);
255 qdisc_watchdog_schedule_ns(&q->watchdog,
256 now + max_t(long, -toks, -ptoks));
258 /* Maybe we have a shorter packet in the queue,
259 which can be sent now. It sounds cool,
260 but, however, this is wrong in principle.
261 We MUST NOT reorder packets under these circumstances.
263 Really, if we split the flow into independent
264 subflows, it would be a very good solution.
265 This is the main idea of all FQ algorithms
266 (cf. CSZ, HPFQ, HFSC)
269 qdisc_qstats_overlimit(sch);
274 static void tbf_reset(struct Qdisc *sch)
276 struct tbf_sched_data *q = qdisc_priv(sch);
278 qdisc_reset(q->qdisc);
279 sch->qstats.backlog = 0;
281 q->t_c = ktime_get_ns();
282 q->tokens = q->buffer;
284 qdisc_watchdog_cancel(&q->watchdog);
287 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
288 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
289 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
290 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
291 [TCA_TBF_RATE64] = { .type = NLA_U64 },
292 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
293 [TCA_TBF_BURST] = { .type = NLA_U32 },
294 [TCA_TBF_PBURST] = { .type = NLA_U32 },
297 static int tbf_change(struct Qdisc *sch, struct nlattr *opt,
298 struct netlink_ext_ack *extack)
301 struct tbf_sched_data *q = qdisc_priv(sch);
302 struct nlattr *tb[TCA_TBF_MAX + 1];
303 struct tc_tbf_qopt *qopt;
304 struct Qdisc *child = NULL;
305 struct psched_ratecfg rate;
306 struct psched_ratecfg peak;
309 u64 rate64 = 0, prate64 = 0;
311 err = nla_parse_nested_deprecated(tb, TCA_TBF_MAX, opt, tbf_policy,
317 if (tb[TCA_TBF_PARMS] == NULL)
320 qopt = nla_data(tb[TCA_TBF_PARMS]);
321 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
322 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
326 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
327 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
331 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
332 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
334 if (tb[TCA_TBF_RATE64])
335 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
336 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
338 if (tb[TCA_TBF_BURST]) {
339 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
340 buffer = psched_l2t_ns(&rate, max_size);
342 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
345 if (qopt->peakrate.rate) {
346 if (tb[TCA_TBF_PRATE64])
347 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
348 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
349 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
350 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
351 peak.rate_bytes_ps, rate.rate_bytes_ps);
356 if (tb[TCA_TBF_PBURST]) {
357 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
358 max_size = min_t(u32, max_size, pburst);
359 mtu = psched_l2t_ns(&peak, pburst);
361 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
364 memset(&peak, 0, sizeof(peak));
367 if (max_size < psched_mtu(qdisc_dev(sch)))
368 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
369 max_size, qdisc_dev(sch)->name,
370 psched_mtu(qdisc_dev(sch)));
377 if (q->qdisc != &noop_qdisc) {
378 err = fifo_set_limit(q->qdisc, qopt->limit);
381 } else if (qopt->limit > 0) {
382 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit,
385 err = PTR_ERR(child);
389 /* child is fifo, no need to check for noop_qdisc */
390 qdisc_hash_add(child, true);
395 qdisc_tree_flush_backlog(q->qdisc);
399 q->limit = qopt->limit;
400 if (tb[TCA_TBF_PBURST])
403 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
404 q->max_size = max_size;
405 if (tb[TCA_TBF_BURST])
408 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
409 q->tokens = q->buffer;
412 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
413 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
415 sch_tree_unlock(sch);
421 static int tbf_init(struct Qdisc *sch, struct nlattr *opt,
422 struct netlink_ext_ack *extack)
424 struct tbf_sched_data *q = qdisc_priv(sch);
426 qdisc_watchdog_init(&q->watchdog, sch);
427 q->qdisc = &noop_qdisc;
432 q->t_c = ktime_get_ns();
434 return tbf_change(sch, opt, extack);
437 static void tbf_destroy(struct Qdisc *sch)
439 struct tbf_sched_data *q = qdisc_priv(sch);
441 qdisc_watchdog_cancel(&q->watchdog);
445 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
447 struct tbf_sched_data *q = qdisc_priv(sch);
449 struct tc_tbf_qopt opt;
451 sch->qstats.backlog = q->qdisc->qstats.backlog;
452 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
454 goto nla_put_failure;
456 opt.limit = q->limit;
457 psched_ratecfg_getrate(&opt.rate, &q->rate);
458 if (tbf_peak_present(q))
459 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
461 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
462 opt.mtu = PSCHED_NS2TICKS(q->mtu);
463 opt.buffer = PSCHED_NS2TICKS(q->buffer);
464 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
465 goto nla_put_failure;
466 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
467 nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
469 goto nla_put_failure;
470 if (tbf_peak_present(q) &&
471 q->peak.rate_bytes_ps >= (1ULL << 32) &&
472 nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
474 goto nla_put_failure;
476 return nla_nest_end(skb, nest);
479 nla_nest_cancel(skb, nest);
483 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
484 struct sk_buff *skb, struct tcmsg *tcm)
486 struct tbf_sched_data *q = qdisc_priv(sch);
488 tcm->tcm_handle |= TC_H_MIN(1);
489 tcm->tcm_info = q->qdisc->handle;
494 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
495 struct Qdisc **old, struct netlink_ext_ack *extack)
497 struct tbf_sched_data *q = qdisc_priv(sch);
502 *old = qdisc_replace(sch, new, &q->qdisc);
506 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
508 struct tbf_sched_data *q = qdisc_priv(sch);
512 static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
517 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
520 if (walker->count >= walker->skip)
521 if (walker->fn(sch, 1, walker) < 0) {
529 static const struct Qdisc_class_ops tbf_class_ops = {
534 .dump = tbf_dump_class,
537 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
539 .cl_ops = &tbf_class_ops,
541 .priv_size = sizeof(struct tbf_sched_data),
542 .enqueue = tbf_enqueue,
543 .dequeue = tbf_dequeue,
544 .peek = qdisc_peek_dequeued,
547 .destroy = tbf_destroy,
548 .change = tbf_change,
550 .owner = THIS_MODULE,
553 static int __init tbf_module_init(void)
555 return register_qdisc(&tbf_qdisc_ops);
558 static void __exit tbf_module_exit(void)
560 unregister_qdisc(&tbf_qdisc_ops);
562 module_init(tbf_module_init)
563 module_exit(tbf_module_exit)
564 MODULE_LICENSE("GPL");