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)
160 struct tbf_sched_data *q = qdisc_priv(sch);
161 struct sk_buff *segs, *nskb;
162 netdev_features_t features = netif_skb_features(skb);
165 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
167 if (IS_ERR_OR_NULL(segs))
168 return qdisc_reshape_fail(skb, sch);
174 qdisc_skb_cb(segs)->pkt_len = segs->len;
175 ret = qdisc_enqueue(segs, q->qdisc);
176 if (ret != NET_XMIT_SUCCESS) {
177 if (net_xmit_drop_count(ret))
186 qdisc_tree_decrease_qlen(sch, 1 - nb);
188 return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
191 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch)
193 struct tbf_sched_data *q = qdisc_priv(sch);
196 if (qdisc_pkt_len(skb) > q->max_size) {
197 if (skb_is_gso(skb) && skb_gso_mac_seglen(skb) <= q->max_size)
198 return tbf_segment(skb, sch);
199 return qdisc_reshape_fail(skb, sch);
201 ret = qdisc_enqueue(skb, q->qdisc);
202 if (ret != NET_XMIT_SUCCESS) {
203 if (net_xmit_drop_count(ret))
209 return NET_XMIT_SUCCESS;
212 static unsigned int tbf_drop(struct Qdisc *sch)
214 struct tbf_sched_data *q = qdisc_priv(sch);
215 unsigned int len = 0;
217 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
224 static bool tbf_peak_present(const struct tbf_sched_data *q)
226 return q->peak.rate_bytes_ps;
229 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
231 struct tbf_sched_data *q = qdisc_priv(sch);
234 skb = q->qdisc->ops->peek(q->qdisc);
240 unsigned int len = qdisc_pkt_len(skb);
242 now = ktime_to_ns(ktime_get());
243 toks = min_t(s64, now - q->t_c, q->buffer);
245 if (tbf_peak_present(q)) {
246 ptoks = toks + q->ptokens;
249 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
252 if (toks > q->buffer)
254 toks -= (s64) psched_l2t_ns(&q->rate, len);
256 if ((toks|ptoks) >= 0) {
257 skb = qdisc_dequeue_peeked(q->qdisc);
265 qdisc_unthrottled(sch);
266 qdisc_bstats_update(sch, skb);
270 qdisc_watchdog_schedule_ns(&q->watchdog,
271 now + max_t(long, -toks, -ptoks));
273 /* Maybe we have a shorter packet in the queue,
274 which can be sent now. It sounds cool,
275 but, however, this is wrong in principle.
276 We MUST NOT reorder packets under these circumstances.
278 Really, if we split the flow into independent
279 subflows, it would be a very good solution.
280 This is the main idea of all FQ algorithms
281 (cf. CSZ, HPFQ, HFSC)
284 sch->qstats.overlimits++;
289 static void tbf_reset(struct Qdisc *sch)
291 struct tbf_sched_data *q = qdisc_priv(sch);
293 qdisc_reset(q->qdisc);
295 q->t_c = ktime_to_ns(ktime_get());
296 q->tokens = q->buffer;
298 qdisc_watchdog_cancel(&q->watchdog);
301 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
302 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
303 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
304 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
305 [TCA_TBF_RATE64] = { .type = NLA_U64 },
306 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
307 [TCA_TBF_BURST] = { .type = NLA_U32 },
308 [TCA_TBF_PBURST] = { .type = NLA_U32 },
311 static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
314 struct tbf_sched_data *q = qdisc_priv(sch);
315 struct nlattr *tb[TCA_TBF_MAX + 1];
316 struct tc_tbf_qopt *qopt;
317 struct Qdisc *child = NULL;
318 struct psched_ratecfg rate;
319 struct psched_ratecfg peak;
322 u64 rate64 = 0, prate64 = 0;
324 err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy);
329 if (tb[TCA_TBF_PARMS] == NULL)
332 qopt = nla_data(tb[TCA_TBF_PARMS]);
333 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
334 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
337 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
338 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
341 if (q->qdisc != &noop_qdisc) {
342 err = fifo_set_limit(q->qdisc, qopt->limit);
345 } else if (qopt->limit > 0) {
346 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
348 err = PTR_ERR(child);
353 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
354 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
356 if (tb[TCA_TBF_RATE64])
357 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
358 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
360 if (tb[TCA_TBF_BURST]) {
361 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
362 buffer = psched_l2t_ns(&rate, max_size);
364 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
367 if (qopt->peakrate.rate) {
368 if (tb[TCA_TBF_PRATE64])
369 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
370 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
371 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
372 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
373 peak.rate_bytes_ps, rate.rate_bytes_ps);
378 if (tb[TCA_TBF_PBURST]) {
379 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
380 max_size = min_t(u32, max_size, pburst);
381 mtu = psched_l2t_ns(&peak, pburst);
383 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
386 memset(&peak, 0, sizeof(peak));
389 if (max_size < psched_mtu(qdisc_dev(sch)))
390 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
391 max_size, qdisc_dev(sch)->name,
392 psched_mtu(qdisc_dev(sch)));
401 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
402 qdisc_destroy(q->qdisc);
405 q->limit = qopt->limit;
406 if (tb[TCA_TBF_PBURST])
409 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
410 q->max_size = max_size;
411 if (tb[TCA_TBF_BURST])
414 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
415 q->tokens = q->buffer;
418 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
419 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
421 sch_tree_unlock(sch);
427 static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
429 struct tbf_sched_data *q = qdisc_priv(sch);
434 q->t_c = ktime_to_ns(ktime_get());
435 qdisc_watchdog_init(&q->watchdog, sch);
436 q->qdisc = &noop_qdisc;
438 return tbf_change(sch, opt);
441 static void tbf_destroy(struct Qdisc *sch)
443 struct tbf_sched_data *q = qdisc_priv(sch);
445 qdisc_watchdog_cancel(&q->watchdog);
446 qdisc_destroy(q->qdisc);
449 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
451 struct tbf_sched_data *q = qdisc_priv(sch);
453 struct tc_tbf_qopt opt;
455 sch->qstats.backlog = q->qdisc->qstats.backlog;
456 nest = nla_nest_start(skb, TCA_OPTIONS);
458 goto nla_put_failure;
460 opt.limit = q->limit;
461 psched_ratecfg_getrate(&opt.rate, &q->rate);
462 if (tbf_peak_present(q))
463 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
465 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
466 opt.mtu = PSCHED_NS2TICKS(q->mtu);
467 opt.buffer = PSCHED_NS2TICKS(q->buffer);
468 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
469 goto nla_put_failure;
470 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
471 nla_put_u64(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps))
472 goto nla_put_failure;
473 if (tbf_peak_present(q) &&
474 q->peak.rate_bytes_ps >= (1ULL << 32) &&
475 nla_put_u64(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps))
476 goto nla_put_failure;
478 nla_nest_end(skb, nest);
482 nla_nest_cancel(skb, nest);
486 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
487 struct sk_buff *skb, struct tcmsg *tcm)
489 struct tbf_sched_data *q = qdisc_priv(sch);
491 tcm->tcm_handle |= TC_H_MIN(1);
492 tcm->tcm_info = q->qdisc->handle;
497 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
500 struct tbf_sched_data *q = qdisc_priv(sch);
508 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
510 sch_tree_unlock(sch);
515 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
517 struct tbf_sched_data *q = qdisc_priv(sch);
521 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
526 static void tbf_put(struct Qdisc *sch, unsigned long arg)
530 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
533 if (walker->count >= walker->skip)
534 if (walker->fn(sch, 1, walker) < 0) {
542 static const struct Qdisc_class_ops tbf_class_ops = {
548 .dump = tbf_dump_class,
551 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
553 .cl_ops = &tbf_class_ops,
555 .priv_size = sizeof(struct tbf_sched_data),
556 .enqueue = tbf_enqueue,
557 .dequeue = tbf_dequeue,
558 .peek = qdisc_peek_dequeued,
562 .destroy = tbf_destroy,
563 .change = tbf_change,
565 .owner = THIS_MODULE,
568 static int __init tbf_module_init(void)
570 return register_qdisc(&tbf_qdisc_ops);
573 static void __exit tbf_module_exit(void)
575 unregister_qdisc(&tbf_qdisc_ops);
577 module_init(tbf_module_init)
578 module_exit(tbf_module_exit)
579 MODULE_LICENSE("GPL");