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)
146 * Return length of individual segments of a gso packet,
147 * including all headers (MAC, IP, TCP/UDP)
149 static unsigned int skb_gso_mac_seglen(const struct sk_buff *skb)
151 unsigned int hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
152 return hdr_len + skb_gso_transport_seglen(skb);
155 /* GSO packet is too big, segment it so that tbf can transmit
156 * each segment in time
158 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,
159 struct sk_buff **to_free)
161 struct tbf_sched_data *q = qdisc_priv(sch);
162 struct sk_buff *segs, *nskb;
163 netdev_features_t features = netif_skb_features(skb);
164 unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
167 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
169 if (IS_ERR_OR_NULL(segs))
170 return qdisc_drop(skb, sch, to_free);
176 qdisc_skb_cb(segs)->pkt_len = segs->len;
178 ret = qdisc_enqueue(segs, q->qdisc, to_free);
179 if (ret != NET_XMIT_SUCCESS) {
180 if (net_xmit_drop_count(ret))
181 qdisc_qstats_drop(sch);
189 qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
191 return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
194 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
195 struct sk_buff **to_free)
197 struct tbf_sched_data *q = qdisc_priv(sch);
200 if (qdisc_pkt_len(skb) > q->max_size) {
201 if (skb_is_gso(skb) && skb_gso_mac_seglen(skb) <= q->max_size)
202 return tbf_segment(skb, sch, to_free);
203 return qdisc_drop(skb, sch, to_free);
205 ret = qdisc_enqueue(skb, q->qdisc, to_free);
206 if (ret != NET_XMIT_SUCCESS) {
207 if (net_xmit_drop_count(ret))
208 qdisc_qstats_drop(sch);
212 qdisc_qstats_backlog_inc(sch, skb);
214 return NET_XMIT_SUCCESS;
217 static bool tbf_peak_present(const struct tbf_sched_data *q)
219 return q->peak.rate_bytes_ps;
222 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
224 struct tbf_sched_data *q = qdisc_priv(sch);
227 skb = q->qdisc->ops->peek(q->qdisc);
233 unsigned int len = qdisc_pkt_len(skb);
235 now = ktime_get_ns();
236 toks = min_t(s64, now - q->t_c, q->buffer);
238 if (tbf_peak_present(q)) {
239 ptoks = toks + q->ptokens;
242 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
245 if (toks > q->buffer)
247 toks -= (s64) psched_l2t_ns(&q->rate, len);
249 if ((toks|ptoks) >= 0) {
250 skb = qdisc_dequeue_peeked(q->qdisc);
257 qdisc_qstats_backlog_dec(sch, skb);
259 qdisc_bstats_update(sch, skb);
263 qdisc_watchdog_schedule_ns(&q->watchdog,
264 now + max_t(long, -toks, -ptoks));
266 /* Maybe we have a shorter packet in the queue,
267 which can be sent now. It sounds cool,
268 but, however, this is wrong in principle.
269 We MUST NOT reorder packets under these circumstances.
271 Really, if we split the flow into independent
272 subflows, it would be a very good solution.
273 This is the main idea of all FQ algorithms
274 (cf. CSZ, HPFQ, HFSC)
277 qdisc_qstats_overlimit(sch);
282 static void tbf_reset(struct Qdisc *sch)
284 struct tbf_sched_data *q = qdisc_priv(sch);
286 qdisc_reset(q->qdisc);
287 sch->qstats.backlog = 0;
289 q->t_c = ktime_get_ns();
290 q->tokens = q->buffer;
292 qdisc_watchdog_cancel(&q->watchdog);
295 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
296 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
297 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
298 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
299 [TCA_TBF_RATE64] = { .type = NLA_U64 },
300 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
301 [TCA_TBF_BURST] = { .type = NLA_U32 },
302 [TCA_TBF_PBURST] = { .type = NLA_U32 },
305 static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
308 struct tbf_sched_data *q = qdisc_priv(sch);
309 struct nlattr *tb[TCA_TBF_MAX + 1];
310 struct tc_tbf_qopt *qopt;
311 struct Qdisc *child = NULL;
312 struct psched_ratecfg rate;
313 struct psched_ratecfg peak;
316 u64 rate64 = 0, prate64 = 0;
318 err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy, NULL);
323 if (tb[TCA_TBF_PARMS] == NULL)
326 qopt = nla_data(tb[TCA_TBF_PARMS]);
327 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
328 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
331 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
332 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
335 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
336 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
338 if (tb[TCA_TBF_RATE64])
339 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
340 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
342 if (tb[TCA_TBF_BURST]) {
343 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
344 buffer = psched_l2t_ns(&rate, max_size);
346 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
349 if (qopt->peakrate.rate) {
350 if (tb[TCA_TBF_PRATE64])
351 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
352 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
353 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
354 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
355 peak.rate_bytes_ps, rate.rate_bytes_ps);
360 if (tb[TCA_TBF_PBURST]) {
361 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
362 max_size = min_t(u32, max_size, pburst);
363 mtu = psched_l2t_ns(&peak, pburst);
365 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
368 memset(&peak, 0, sizeof(peak));
371 if (max_size < psched_mtu(qdisc_dev(sch)))
372 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
373 max_size, qdisc_dev(sch)->name,
374 psched_mtu(qdisc_dev(sch)));
381 if (q->qdisc != &noop_qdisc) {
382 err = fifo_set_limit(q->qdisc, qopt->limit);
385 } else if (qopt->limit > 0) {
386 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
388 err = PTR_ERR(child);
395 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
396 q->qdisc->qstats.backlog);
397 qdisc_destroy(q->qdisc);
399 if (child != &noop_qdisc)
400 qdisc_hash_add(child, true);
402 q->limit = qopt->limit;
403 if (tb[TCA_TBF_PBURST])
406 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
407 q->max_size = max_size;
408 if (tb[TCA_TBF_BURST])
411 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
412 q->tokens = q->buffer;
415 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
416 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
418 sch_tree_unlock(sch);
424 static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
426 struct tbf_sched_data *q = qdisc_priv(sch);
431 q->t_c = ktime_get_ns();
432 qdisc_watchdog_init(&q->watchdog, sch);
433 q->qdisc = &noop_qdisc;
435 return tbf_change(sch, opt);
438 static void tbf_destroy(struct Qdisc *sch)
440 struct tbf_sched_data *q = qdisc_priv(sch);
442 qdisc_watchdog_cancel(&q->watchdog);
443 qdisc_destroy(q->qdisc);
446 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
448 struct tbf_sched_data *q = qdisc_priv(sch);
450 struct tc_tbf_qopt opt;
452 sch->qstats.backlog = q->qdisc->qstats.backlog;
453 nest = nla_nest_start(skb, TCA_OPTIONS);
455 goto nla_put_failure;
457 opt.limit = q->limit;
458 psched_ratecfg_getrate(&opt.rate, &q->rate);
459 if (tbf_peak_present(q))
460 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
462 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
463 opt.mtu = PSCHED_NS2TICKS(q->mtu);
464 opt.buffer = PSCHED_NS2TICKS(q->buffer);
465 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
466 goto nla_put_failure;
467 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
468 nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
470 goto nla_put_failure;
471 if (tbf_peak_present(q) &&
472 q->peak.rate_bytes_ps >= (1ULL << 32) &&
473 nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
475 goto nla_put_failure;
477 return nla_nest_end(skb, nest);
480 nla_nest_cancel(skb, nest);
484 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
485 struct sk_buff *skb, struct tcmsg *tcm)
487 struct tbf_sched_data *q = qdisc_priv(sch);
489 tcm->tcm_handle |= TC_H_MIN(1);
490 tcm->tcm_info = q->qdisc->handle;
495 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
498 struct tbf_sched_data *q = qdisc_priv(sch);
503 *old = qdisc_replace(sch, new, &q->qdisc);
507 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
509 struct tbf_sched_data *q = qdisc_priv(sch);
513 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
518 static void tbf_put(struct Qdisc *sch, unsigned long arg)
522 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
525 if (walker->count >= walker->skip)
526 if (walker->fn(sch, 1, walker) < 0) {
534 static const struct Qdisc_class_ops tbf_class_ops = {
540 .dump = tbf_dump_class,
543 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
545 .cl_ops = &tbf_class_ops,
547 .priv_size = sizeof(struct tbf_sched_data),
548 .enqueue = tbf_enqueue,
549 .dequeue = tbf_dequeue,
550 .peek = qdisc_peek_dequeued,
553 .destroy = tbf_destroy,
554 .change = tbf_change,
556 .owner = THIS_MODULE,
559 static int __init tbf_module_init(void)
561 return register_qdisc(&tbf_qdisc_ops);
564 static void __exit tbf_module_exit(void)
566 unregister_qdisc(&tbf_qdisc_ops);
568 module_init(tbf_module_init)
569 module_exit(tbf_module_exit)
570 MODULE_LICENSE("GPL");