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);
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);
190 if (qdisc_pkt_len(skb) > q->max_size) {
191 if (skb_is_gso(skb) && skb_gso_mac_seglen(skb) <= q->max_size)
192 return tbf_segment(skb, sch, to_free);
193 return qdisc_drop(skb, sch, to_free);
195 ret = qdisc_enqueue(skb, q->qdisc, to_free);
196 if (ret != NET_XMIT_SUCCESS) {
197 if (net_xmit_drop_count(ret))
198 qdisc_qstats_drop(sch);
202 qdisc_qstats_backlog_inc(sch, skb);
204 return NET_XMIT_SUCCESS;
207 static bool tbf_peak_present(const struct tbf_sched_data *q)
209 return q->peak.rate_bytes_ps;
212 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
214 struct tbf_sched_data *q = qdisc_priv(sch);
217 skb = q->qdisc->ops->peek(q->qdisc);
223 unsigned int len = qdisc_pkt_len(skb);
225 now = ktime_get_ns();
226 toks = min_t(s64, now - q->t_c, q->buffer);
228 if (tbf_peak_present(q)) {
229 ptoks = toks + q->ptokens;
232 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
235 if (toks > q->buffer)
237 toks -= (s64) psched_l2t_ns(&q->rate, len);
239 if ((toks|ptoks) >= 0) {
240 skb = qdisc_dequeue_peeked(q->qdisc);
247 qdisc_qstats_backlog_dec(sch, skb);
249 qdisc_bstats_update(sch, skb);
253 qdisc_watchdog_schedule_ns(&q->watchdog,
254 now + max_t(long, -toks, -ptoks));
256 /* Maybe we have a shorter packet in the queue,
257 which can be sent now. It sounds cool,
258 but, however, this is wrong in principle.
259 We MUST NOT reorder packets under these circumstances.
261 Really, if we split the flow into independent
262 subflows, it would be a very good solution.
263 This is the main idea of all FQ algorithms
264 (cf. CSZ, HPFQ, HFSC)
267 qdisc_qstats_overlimit(sch);
272 static void tbf_reset(struct Qdisc *sch)
274 struct tbf_sched_data *q = qdisc_priv(sch);
276 qdisc_reset(q->qdisc);
277 sch->qstats.backlog = 0;
279 q->t_c = ktime_get_ns();
280 q->tokens = q->buffer;
282 qdisc_watchdog_cancel(&q->watchdog);
285 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
286 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
287 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
288 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
289 [TCA_TBF_RATE64] = { .type = NLA_U64 },
290 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
291 [TCA_TBF_BURST] = { .type = NLA_U32 },
292 [TCA_TBF_PBURST] = { .type = NLA_U32 },
295 static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
298 struct tbf_sched_data *q = qdisc_priv(sch);
299 struct nlattr *tb[TCA_TBF_MAX + 1];
300 struct tc_tbf_qopt *qopt;
301 struct Qdisc *child = NULL;
302 struct psched_ratecfg rate;
303 struct psched_ratecfg peak;
306 u64 rate64 = 0, prate64 = 0;
308 err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy);
313 if (tb[TCA_TBF_PARMS] == NULL)
316 qopt = nla_data(tb[TCA_TBF_PARMS]);
317 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
318 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
321 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
322 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
325 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
326 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
328 if (tb[TCA_TBF_RATE64])
329 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
330 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
332 if (tb[TCA_TBF_BURST]) {
333 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
334 buffer = psched_l2t_ns(&rate, max_size);
336 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
339 if (qopt->peakrate.rate) {
340 if (tb[TCA_TBF_PRATE64])
341 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
342 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
343 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
344 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
345 peak.rate_bytes_ps, rate.rate_bytes_ps);
350 if (tb[TCA_TBF_PBURST]) {
351 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
352 max_size = min_t(u32, max_size, pburst);
353 mtu = psched_l2t_ns(&peak, pburst);
355 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
358 memset(&peak, 0, sizeof(peak));
361 if (max_size < psched_mtu(qdisc_dev(sch)))
362 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
363 max_size, qdisc_dev(sch)->name,
364 psched_mtu(qdisc_dev(sch)));
371 if (q->qdisc != &noop_qdisc) {
372 err = fifo_set_limit(q->qdisc, qopt->limit);
375 } else if (qopt->limit > 0) {
376 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
378 err = PTR_ERR(child);
385 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
386 q->qdisc->qstats.backlog);
387 qdisc_destroy(q->qdisc);
390 q->limit = qopt->limit;
391 if (tb[TCA_TBF_PBURST])
394 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
395 q->max_size = max_size;
396 if (tb[TCA_TBF_BURST])
399 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
400 q->tokens = q->buffer;
403 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
404 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
406 sch_tree_unlock(sch);
412 static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
414 struct tbf_sched_data *q = qdisc_priv(sch);
416 qdisc_watchdog_init(&q->watchdog, sch);
417 q->qdisc = &noop_qdisc;
422 q->t_c = ktime_get_ns();
424 return tbf_change(sch, opt);
427 static void tbf_destroy(struct Qdisc *sch)
429 struct tbf_sched_data *q = qdisc_priv(sch);
431 qdisc_watchdog_cancel(&q->watchdog);
432 qdisc_destroy(q->qdisc);
435 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
437 struct tbf_sched_data *q = qdisc_priv(sch);
439 struct tc_tbf_qopt opt;
441 sch->qstats.backlog = q->qdisc->qstats.backlog;
442 nest = nla_nest_start(skb, TCA_OPTIONS);
444 goto nla_put_failure;
446 opt.limit = q->limit;
447 psched_ratecfg_getrate(&opt.rate, &q->rate);
448 if (tbf_peak_present(q))
449 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
451 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
452 opt.mtu = PSCHED_NS2TICKS(q->mtu);
453 opt.buffer = PSCHED_NS2TICKS(q->buffer);
454 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
455 goto nla_put_failure;
456 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
457 nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
459 goto nla_put_failure;
460 if (tbf_peak_present(q) &&
461 q->peak.rate_bytes_ps >= (1ULL << 32) &&
462 nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
464 goto nla_put_failure;
466 return nla_nest_end(skb, nest);
469 nla_nest_cancel(skb, nest);
473 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
474 struct sk_buff *skb, struct tcmsg *tcm)
476 struct tbf_sched_data *q = qdisc_priv(sch);
478 tcm->tcm_handle |= TC_H_MIN(1);
479 tcm->tcm_info = q->qdisc->handle;
484 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
487 struct tbf_sched_data *q = qdisc_priv(sch);
492 *old = qdisc_replace(sch, new, &q->qdisc);
496 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
498 struct tbf_sched_data *q = qdisc_priv(sch);
502 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
507 static void tbf_put(struct Qdisc *sch, unsigned long arg)
511 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
514 if (walker->count >= walker->skip)
515 if (walker->fn(sch, 1, walker) < 0) {
523 static const struct Qdisc_class_ops tbf_class_ops = {
529 .dump = tbf_dump_class,
532 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
534 .cl_ops = &tbf_class_ops,
536 .priv_size = sizeof(struct tbf_sched_data),
537 .enqueue = tbf_enqueue,
538 .dequeue = tbf_dequeue,
539 .peek = qdisc_peek_dequeued,
542 .destroy = tbf_destroy,
543 .change = tbf_change,
545 .owner = THIS_MODULE,
548 static int __init tbf_module_init(void)
550 return register_qdisc(&tbf_qdisc_ops);
553 static void __exit tbf_module_exit(void)
555 unregister_qdisc(&tbf_qdisc_ops);
557 module_init(tbf_module_init)
558 module_exit(tbf_module_exit)
559 MODULE_LICENSE("GPL");