1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * net/sched/sch_tbf.c Token Bucket Filter queue.
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Dmitry Torokhov <dtor@mail.ru> - allow attaching inner qdiscs -
7 * original idea by Martin Devera
10 #include <linux/module.h>
11 #include <linux/types.h>
12 #include <linux/kernel.h>
13 #include <linux/string.h>
14 #include <linux/errno.h>
15 #include <linux/skbuff.h>
16 #include <net/netlink.h>
17 #include <net/sch_generic.h>
18 #include <net/pkt_cls.h>
19 #include <net/pkt_sched.h>
22 /* Simple Token Bucket Filter.
23 =======================================
33 A data flow obeys TBF with rate R and depth B, if for any
34 time interval t_i...t_f the number of transmitted bits
35 does not exceed B + R*(t_f-t_i).
37 Packetized version of this definition:
38 The sequence of packets of sizes s_i served at moments t_i
39 obeys TBF, if for any i<=k:
41 s_i+....+s_k <= B + R*(t_k - t_i)
46 Let N(t_i) be B/R initially and N(t) grow continuously with time as:
48 N(t+delta) = min{B/R, N(t) + delta}
50 If the first packet in queue has length S, it may be
51 transmitted only at the time t_* when S/R <= N(t_*),
52 and in this case N(t) jumps:
54 N(t_* + 0) = N(t_* - 0) - S/R.
58 Actually, QoS requires two TBF to be applied to a data stream.
59 One of them controls steady state burst size, another
60 one with rate P (peak rate) and depth M (equal to link MTU)
61 limits bursts at a smaller time scale.
63 It is easy to see that P>R, and B>M. If P is infinity, this double
64 TBF is equivalent to a single one.
66 When TBF works in reshaping mode, latency is estimated as:
68 lat = max ((L-B)/R, (L-M)/P)
74 If TBF throttles, it starts a watchdog timer, which will wake it up
75 when it is ready to transmit.
76 Note that the minimal timer resolution is 1/HZ.
77 If no new packets arrive during this period,
78 or if the device is not awaken by EOI for some previous packet,
79 TBF can stop its activity for 1/HZ.
82 This means, that with depth B, the maximal rate is
86 F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
88 Note that the peak rate TBF is much more tough: with MTU 1500
89 P_crit = 150Kbytes/sec. So, if you need greater peak
90 rates, use alpha with HZ=1000 :-)
92 With classful TBF, limit is just kept for backwards compatibility.
93 It is passed to the default bfifo qdisc - if the inner qdisc is
94 changed the limit is not effective anymore.
97 struct tbf_sched_data {
99 u32 limit; /* Maximal length of backlog: bytes */
101 s64 buffer; /* Token bucket depth/rate: MUST BE >= MTU/B */
103 struct psched_ratecfg rate;
104 struct psched_ratecfg peak;
107 s64 tokens; /* Current number of B tokens */
108 s64 ptokens; /* Current number of P tokens */
109 s64 t_c; /* Time check-point */
110 struct Qdisc *qdisc; /* Inner qdisc, default - bfifo queue */
111 struct qdisc_watchdog watchdog; /* Watchdog timer */
115 /* Time to Length, convert time in ns to length in bytes
116 * to determinate how many bytes can be sent in given time.
118 static u64 psched_ns_t2l(const struct psched_ratecfg *r,
122 * len = (time_in_ns * r->rate_bytes_ps) / NSEC_PER_SEC
124 u64 len = time_in_ns * r->rate_bytes_ps;
126 do_div(len, NSEC_PER_SEC);
128 if (unlikely(r->linklayer == TC_LINKLAYER_ATM)) {
133 if (len > r->overhead)
141 static void tbf_offload_change(struct Qdisc *sch)
143 struct tbf_sched_data *q = qdisc_priv(sch);
144 struct net_device *dev = qdisc_dev(sch);
145 struct tc_tbf_qopt_offload qopt;
147 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
150 qopt.command = TC_TBF_REPLACE;
151 qopt.handle = sch->handle;
152 qopt.parent = sch->parent;
153 qopt.replace_params.rate = q->rate;
154 qopt.replace_params.max_size = q->max_size;
155 qopt.replace_params.qstats = &sch->qstats;
157 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TBF, &qopt);
160 static void tbf_offload_destroy(struct Qdisc *sch)
162 struct net_device *dev = qdisc_dev(sch);
163 struct tc_tbf_qopt_offload qopt;
165 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
168 qopt.command = TC_TBF_DESTROY;
169 qopt.handle = sch->handle;
170 qopt.parent = sch->parent;
171 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_TBF, &qopt);
174 static int tbf_offload_dump(struct Qdisc *sch)
176 struct tc_tbf_qopt_offload qopt;
178 qopt.command = TC_TBF_STATS;
179 qopt.handle = sch->handle;
180 qopt.parent = sch->parent;
181 qopt.stats.bstats = &sch->bstats;
182 qopt.stats.qstats = &sch->qstats;
184 return qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_TBF, &qopt);
187 /* GSO packet is too big, segment it so that tbf can transmit
188 * each segment in time
190 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,
191 struct sk_buff **to_free)
193 struct tbf_sched_data *q = qdisc_priv(sch);
194 struct sk_buff *segs, *nskb;
195 netdev_features_t features = netif_skb_features(skb);
196 unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
199 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
201 if (IS_ERR_OR_NULL(segs))
202 return qdisc_drop(skb, sch, to_free);
205 skb_list_walk_safe(segs, segs, nskb) {
206 skb_mark_not_on_list(segs);
207 qdisc_skb_cb(segs)->pkt_len = segs->len;
209 ret = qdisc_enqueue(segs, q->qdisc, to_free);
210 if (ret != NET_XMIT_SUCCESS) {
211 if (net_xmit_drop_count(ret))
212 qdisc_qstats_drop(sch);
219 qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
221 return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
224 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
225 struct sk_buff **to_free)
227 struct tbf_sched_data *q = qdisc_priv(sch);
228 unsigned int len = qdisc_pkt_len(skb);
231 if (qdisc_pkt_len(skb) > q->max_size) {
232 if (skb_is_gso(skb) &&
233 skb_gso_validate_mac_len(skb, q->max_size))
234 return tbf_segment(skb, sch, to_free);
235 return qdisc_drop(skb, sch, to_free);
237 ret = qdisc_enqueue(skb, q->qdisc, to_free);
238 if (ret != NET_XMIT_SUCCESS) {
239 if (net_xmit_drop_count(ret))
240 qdisc_qstats_drop(sch);
244 sch->qstats.backlog += len;
246 return NET_XMIT_SUCCESS;
249 static bool tbf_peak_present(const struct tbf_sched_data *q)
251 return q->peak.rate_bytes_ps;
254 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
256 struct tbf_sched_data *q = qdisc_priv(sch);
259 skb = q->qdisc->ops->peek(q->qdisc);
265 unsigned int len = qdisc_pkt_len(skb);
267 now = ktime_get_ns();
268 toks = min_t(s64, now - q->t_c, q->buffer);
270 if (tbf_peak_present(q)) {
271 ptoks = toks + q->ptokens;
274 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
277 if (toks > q->buffer)
279 toks -= (s64) psched_l2t_ns(&q->rate, len);
281 if ((toks|ptoks) >= 0) {
282 skb = qdisc_dequeue_peeked(q->qdisc);
289 qdisc_qstats_backlog_dec(sch, skb);
291 qdisc_bstats_update(sch, skb);
295 qdisc_watchdog_schedule_ns(&q->watchdog,
296 now + max_t(long, -toks, -ptoks));
298 /* Maybe we have a shorter packet in the queue,
299 which can be sent now. It sounds cool,
300 but, however, this is wrong in principle.
301 We MUST NOT reorder packets under these circumstances.
303 Really, if we split the flow into independent
304 subflows, it would be a very good solution.
305 This is the main idea of all FQ algorithms
306 (cf. CSZ, HPFQ, HFSC)
309 qdisc_qstats_overlimit(sch);
314 static void tbf_reset(struct Qdisc *sch)
316 struct tbf_sched_data *q = qdisc_priv(sch);
318 qdisc_reset(q->qdisc);
319 sch->qstats.backlog = 0;
321 q->t_c = ktime_get_ns();
322 q->tokens = q->buffer;
324 qdisc_watchdog_cancel(&q->watchdog);
327 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
328 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
329 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
330 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
331 [TCA_TBF_RATE64] = { .type = NLA_U64 },
332 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
333 [TCA_TBF_BURST] = { .type = NLA_U32 },
334 [TCA_TBF_PBURST] = { .type = NLA_U32 },
337 static int tbf_change(struct Qdisc *sch, struct nlattr *opt,
338 struct netlink_ext_ack *extack)
341 struct tbf_sched_data *q = qdisc_priv(sch);
342 struct nlattr *tb[TCA_TBF_MAX + 1];
343 struct tc_tbf_qopt *qopt;
344 struct Qdisc *child = NULL;
345 struct psched_ratecfg rate;
346 struct psched_ratecfg peak;
349 u64 rate64 = 0, prate64 = 0;
351 err = nla_parse_nested_deprecated(tb, TCA_TBF_MAX, opt, tbf_policy,
357 if (tb[TCA_TBF_PARMS] == NULL)
360 qopt = nla_data(tb[TCA_TBF_PARMS]);
361 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
362 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
366 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
367 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
371 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
372 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
374 if (tb[TCA_TBF_RATE64])
375 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
376 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
378 if (tb[TCA_TBF_BURST]) {
379 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
380 buffer = psched_l2t_ns(&rate, max_size);
382 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
385 if (qopt->peakrate.rate) {
386 if (tb[TCA_TBF_PRATE64])
387 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
388 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
389 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
390 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
391 peak.rate_bytes_ps, rate.rate_bytes_ps);
396 if (tb[TCA_TBF_PBURST]) {
397 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
398 max_size = min_t(u32, max_size, pburst);
399 mtu = psched_l2t_ns(&peak, pburst);
401 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
404 memset(&peak, 0, sizeof(peak));
407 if (max_size < psched_mtu(qdisc_dev(sch)))
408 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
409 max_size, qdisc_dev(sch)->name,
410 psched_mtu(qdisc_dev(sch)));
417 if (q->qdisc != &noop_qdisc) {
418 err = fifo_set_limit(q->qdisc, qopt->limit);
421 } else if (qopt->limit > 0) {
422 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit,
425 err = PTR_ERR(child);
429 /* child is fifo, no need to check for noop_qdisc */
430 qdisc_hash_add(child, true);
435 qdisc_tree_flush_backlog(q->qdisc);
439 q->limit = qopt->limit;
440 if (tb[TCA_TBF_PBURST])
443 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
444 q->max_size = max_size;
445 if (tb[TCA_TBF_BURST])
448 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
449 q->tokens = q->buffer;
452 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
453 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
455 sch_tree_unlock(sch);
458 tbf_offload_change(sch);
463 static int tbf_init(struct Qdisc *sch, struct nlattr *opt,
464 struct netlink_ext_ack *extack)
466 struct tbf_sched_data *q = qdisc_priv(sch);
468 qdisc_watchdog_init(&q->watchdog, sch);
469 q->qdisc = &noop_qdisc;
474 q->t_c = ktime_get_ns();
476 return tbf_change(sch, opt, extack);
479 static void tbf_destroy(struct Qdisc *sch)
481 struct tbf_sched_data *q = qdisc_priv(sch);
483 qdisc_watchdog_cancel(&q->watchdog);
484 tbf_offload_destroy(sch);
488 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
490 struct tbf_sched_data *q = qdisc_priv(sch);
492 struct tc_tbf_qopt opt;
495 err = tbf_offload_dump(sch);
499 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
501 goto nla_put_failure;
503 opt.limit = q->limit;
504 psched_ratecfg_getrate(&opt.rate, &q->rate);
505 if (tbf_peak_present(q))
506 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
508 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
509 opt.mtu = PSCHED_NS2TICKS(q->mtu);
510 opt.buffer = PSCHED_NS2TICKS(q->buffer);
511 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
512 goto nla_put_failure;
513 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
514 nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
516 goto nla_put_failure;
517 if (tbf_peak_present(q) &&
518 q->peak.rate_bytes_ps >= (1ULL << 32) &&
519 nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
521 goto nla_put_failure;
523 return nla_nest_end(skb, nest);
526 nla_nest_cancel(skb, nest);
530 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
531 struct sk_buff *skb, struct tcmsg *tcm)
533 struct tbf_sched_data *q = qdisc_priv(sch);
535 tcm->tcm_handle |= TC_H_MIN(1);
536 tcm->tcm_info = q->qdisc->handle;
541 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
542 struct Qdisc **old, struct netlink_ext_ack *extack)
544 struct tbf_sched_data *q = qdisc_priv(sch);
549 *old = qdisc_replace(sch, new, &q->qdisc);
553 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
555 struct tbf_sched_data *q = qdisc_priv(sch);
559 static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
564 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
567 if (walker->count >= walker->skip)
568 if (walker->fn(sch, 1, walker) < 0) {
576 static const struct Qdisc_class_ops tbf_class_ops = {
581 .dump = tbf_dump_class,
584 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
586 .cl_ops = &tbf_class_ops,
588 .priv_size = sizeof(struct tbf_sched_data),
589 .enqueue = tbf_enqueue,
590 .dequeue = tbf_dequeue,
591 .peek = qdisc_peek_dequeued,
594 .destroy = tbf_destroy,
595 .change = tbf_change,
597 .owner = THIS_MODULE,
600 static int __init tbf_module_init(void)
602 return register_qdisc(&tbf_qdisc_ops);
605 static void __exit tbf_module_exit(void)
607 unregister_qdisc(&tbf_qdisc_ops);
609 module_init(tbf_module_init)
610 module_exit(tbf_module_exit)
611 MODULE_LICENSE("GPL");