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_sched.h>
21 /* Simple Token Bucket Filter.
22 =======================================
32 A data flow obeys TBF with rate R and depth B, if for any
33 time interval t_i...t_f the number of transmitted bits
34 does not exceed B + R*(t_f-t_i).
36 Packetized version of this definition:
37 The sequence of packets of sizes s_i served at moments t_i
38 obeys TBF, if for any i<=k:
40 s_i+....+s_k <= B + R*(t_k - t_i)
45 Let N(t_i) be B/R initially and N(t) grow continuously with time as:
47 N(t+delta) = min{B/R, N(t) + delta}
49 If the first packet in queue has length S, it may be
50 transmitted only at the time t_* when S/R <= N(t_*),
51 and in this case N(t) jumps:
53 N(t_* + 0) = N(t_* - 0) - S/R.
57 Actually, QoS requires two TBF to be applied to a data stream.
58 One of them controls steady state burst size, another
59 one with rate P (peak rate) and depth M (equal to link MTU)
60 limits bursts at a smaller time scale.
62 It is easy to see that P>R, and B>M. If P is infinity, this double
63 TBF is equivalent to a single one.
65 When TBF works in reshaping mode, latency is estimated as:
67 lat = max ((L-B)/R, (L-M)/P)
73 If TBF throttles, it starts a watchdog timer, which will wake it up
74 when it is ready to transmit.
75 Note that the minimal timer resolution is 1/HZ.
76 If no new packets arrive during this period,
77 or if the device is not awaken by EOI for some previous packet,
78 TBF can stop its activity for 1/HZ.
81 This means, that with depth B, the maximal rate is
85 F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
87 Note that the peak rate TBF is much more tough: with MTU 1500
88 P_crit = 150Kbytes/sec. So, if you need greater peak
89 rates, use alpha with HZ=1000 :-)
91 With classful TBF, limit is just kept for backwards compatibility.
92 It is passed to the default bfifo qdisc - if the inner qdisc is
93 changed the limit is not effective anymore.
96 struct tbf_sched_data {
98 u32 limit; /* Maximal length of backlog: bytes */
100 s64 buffer; /* Token bucket depth/rate: MUST BE >= MTU/B */
102 struct psched_ratecfg rate;
103 struct psched_ratecfg peak;
106 s64 tokens; /* Current number of B tokens */
107 s64 ptokens; /* Current number of P tokens */
108 s64 t_c; /* Time check-point */
109 struct Qdisc *qdisc; /* Inner qdisc, default - bfifo queue */
110 struct qdisc_watchdog watchdog; /* Watchdog timer */
114 /* Time to Length, convert time in ns to length in bytes
115 * to determinate how many bytes can be sent in given time.
117 static u64 psched_ns_t2l(const struct psched_ratecfg *r,
121 * len = (time_in_ns * r->rate_bytes_ps) / NSEC_PER_SEC
123 u64 len = time_in_ns * r->rate_bytes_ps;
125 do_div(len, NSEC_PER_SEC);
127 if (unlikely(r->linklayer == TC_LINKLAYER_ATM)) {
132 if (len > r->overhead)
140 /* GSO packet is too big, segment it so that tbf can transmit
141 * each segment in time
143 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,
144 struct sk_buff **to_free)
146 struct tbf_sched_data *q = qdisc_priv(sch);
147 struct sk_buff *segs, *nskb;
148 netdev_features_t features = netif_skb_features(skb);
149 unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
152 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
154 if (IS_ERR_OR_NULL(segs))
155 return qdisc_drop(skb, sch, to_free);
158 skb_list_walk_safe(segs, segs, nskb) {
159 skb_mark_not_on_list(segs);
160 qdisc_skb_cb(segs)->pkt_len = segs->len;
162 ret = qdisc_enqueue(segs, q->qdisc, to_free);
163 if (ret != NET_XMIT_SUCCESS) {
164 if (net_xmit_drop_count(ret))
165 qdisc_qstats_drop(sch);
172 qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
174 return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
177 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,
178 struct sk_buff **to_free)
180 struct tbf_sched_data *q = qdisc_priv(sch);
181 unsigned int len = qdisc_pkt_len(skb);
184 if (qdisc_pkt_len(skb) > q->max_size) {
185 if (skb_is_gso(skb) &&
186 skb_gso_validate_mac_len(skb, q->max_size))
187 return tbf_segment(skb, sch, to_free);
188 return qdisc_drop(skb, sch, to_free);
190 ret = qdisc_enqueue(skb, q->qdisc, to_free);
191 if (ret != NET_XMIT_SUCCESS) {
192 if (net_xmit_drop_count(ret))
193 qdisc_qstats_drop(sch);
197 sch->qstats.backlog += len;
199 return NET_XMIT_SUCCESS;
202 static bool tbf_peak_present(const struct tbf_sched_data *q)
204 return q->peak.rate_bytes_ps;
207 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
209 struct tbf_sched_data *q = qdisc_priv(sch);
212 skb = q->qdisc->ops->peek(q->qdisc);
218 unsigned int len = qdisc_pkt_len(skb);
220 now = ktime_get_ns();
221 toks = min_t(s64, now - q->t_c, q->buffer);
223 if (tbf_peak_present(q)) {
224 ptoks = toks + q->ptokens;
227 ptoks -= (s64) psched_l2t_ns(&q->peak, len);
230 if (toks > q->buffer)
232 toks -= (s64) psched_l2t_ns(&q->rate, len);
234 if ((toks|ptoks) >= 0) {
235 skb = qdisc_dequeue_peeked(q->qdisc);
242 qdisc_qstats_backlog_dec(sch, skb);
244 qdisc_bstats_update(sch, skb);
248 qdisc_watchdog_schedule_ns(&q->watchdog,
249 now + max_t(long, -toks, -ptoks));
251 /* Maybe we have a shorter packet in the queue,
252 which can be sent now. It sounds cool,
253 but, however, this is wrong in principle.
254 We MUST NOT reorder packets under these circumstances.
256 Really, if we split the flow into independent
257 subflows, it would be a very good solution.
258 This is the main idea of all FQ algorithms
259 (cf. CSZ, HPFQ, HFSC)
262 qdisc_qstats_overlimit(sch);
267 static void tbf_reset(struct Qdisc *sch)
269 struct tbf_sched_data *q = qdisc_priv(sch);
271 qdisc_reset(q->qdisc);
272 sch->qstats.backlog = 0;
274 q->t_c = ktime_get_ns();
275 q->tokens = q->buffer;
277 qdisc_watchdog_cancel(&q->watchdog);
280 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
281 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
282 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
283 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
284 [TCA_TBF_RATE64] = { .type = NLA_U64 },
285 [TCA_TBF_PRATE64] = { .type = NLA_U64 },
286 [TCA_TBF_BURST] = { .type = NLA_U32 },
287 [TCA_TBF_PBURST] = { .type = NLA_U32 },
290 static int tbf_change(struct Qdisc *sch, struct nlattr *opt,
291 struct netlink_ext_ack *extack)
294 struct tbf_sched_data *q = qdisc_priv(sch);
295 struct nlattr *tb[TCA_TBF_MAX + 1];
296 struct tc_tbf_qopt *qopt;
297 struct Qdisc *child = NULL;
298 struct psched_ratecfg rate;
299 struct psched_ratecfg peak;
302 u64 rate64 = 0, prate64 = 0;
304 err = nla_parse_nested_deprecated(tb, TCA_TBF_MAX, opt, tbf_policy,
310 if (tb[TCA_TBF_PARMS] == NULL)
313 qopt = nla_data(tb[TCA_TBF_PARMS]);
314 if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
315 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
319 if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
320 qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
324 buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
325 mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
327 if (tb[TCA_TBF_RATE64])
328 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
329 psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
331 if (tb[TCA_TBF_BURST]) {
332 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
333 buffer = psched_l2t_ns(&rate, max_size);
335 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
338 if (qopt->peakrate.rate) {
339 if (tb[TCA_TBF_PRATE64])
340 prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
341 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
342 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
343 pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
344 peak.rate_bytes_ps, rate.rate_bytes_ps);
349 if (tb[TCA_TBF_PBURST]) {
350 u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
351 max_size = min_t(u32, max_size, pburst);
352 mtu = psched_l2t_ns(&peak, pburst);
354 max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
357 memset(&peak, 0, sizeof(peak));
360 if (max_size < psched_mtu(qdisc_dev(sch)))
361 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
362 max_size, qdisc_dev(sch)->name,
363 psched_mtu(qdisc_dev(sch)));
370 if (q->qdisc != &noop_qdisc) {
371 err = fifo_set_limit(q->qdisc, qopt->limit);
374 } else if (qopt->limit > 0) {
375 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit,
378 err = PTR_ERR(child);
382 /* child is fifo, no need to check for noop_qdisc */
383 qdisc_hash_add(child, true);
388 qdisc_tree_flush_backlog(q->qdisc);
392 q->limit = qopt->limit;
393 if (tb[TCA_TBF_PBURST])
396 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
397 q->max_size = max_size;
398 if (tb[TCA_TBF_BURST])
401 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
402 q->tokens = q->buffer;
405 memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
406 memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
408 sch_tree_unlock(sch);
414 static int tbf_init(struct Qdisc *sch, struct nlattr *opt,
415 struct netlink_ext_ack *extack)
417 struct tbf_sched_data *q = qdisc_priv(sch);
419 qdisc_watchdog_init(&q->watchdog, sch);
420 q->qdisc = &noop_qdisc;
425 q->t_c = ktime_get_ns();
427 return tbf_change(sch, opt, extack);
430 static void tbf_destroy(struct Qdisc *sch)
432 struct tbf_sched_data *q = qdisc_priv(sch);
434 qdisc_watchdog_cancel(&q->watchdog);
438 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
440 struct tbf_sched_data *q = qdisc_priv(sch);
442 struct tc_tbf_qopt opt;
444 sch->qstats.backlog = q->qdisc->qstats.backlog;
445 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
447 goto nla_put_failure;
449 opt.limit = q->limit;
450 psched_ratecfg_getrate(&opt.rate, &q->rate);
451 if (tbf_peak_present(q))
452 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
454 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
455 opt.mtu = PSCHED_NS2TICKS(q->mtu);
456 opt.buffer = PSCHED_NS2TICKS(q->buffer);
457 if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
458 goto nla_put_failure;
459 if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
460 nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
462 goto nla_put_failure;
463 if (tbf_peak_present(q) &&
464 q->peak.rate_bytes_ps >= (1ULL << 32) &&
465 nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
467 goto nla_put_failure;
469 return nla_nest_end(skb, nest);
472 nla_nest_cancel(skb, nest);
476 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
477 struct sk_buff *skb, struct tcmsg *tcm)
479 struct tbf_sched_data *q = qdisc_priv(sch);
481 tcm->tcm_handle |= TC_H_MIN(1);
482 tcm->tcm_info = q->qdisc->handle;
487 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
488 struct Qdisc **old, struct netlink_ext_ack *extack)
490 struct tbf_sched_data *q = qdisc_priv(sch);
495 *old = qdisc_replace(sch, new, &q->qdisc);
499 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
501 struct tbf_sched_data *q = qdisc_priv(sch);
505 static unsigned long tbf_find(struct Qdisc *sch, u32 classid)
510 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
513 if (walker->count >= walker->skip)
514 if (walker->fn(sch, 1, walker) < 0) {
522 static const struct Qdisc_class_ops tbf_class_ops = {
527 .dump = tbf_dump_class,
530 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
532 .cl_ops = &tbf_class_ops,
534 .priv_size = sizeof(struct tbf_sched_data),
535 .enqueue = tbf_enqueue,
536 .dequeue = tbf_dequeue,
537 .peek = qdisc_peek_dequeued,
540 .destroy = tbf_destroy,
541 .change = tbf_change,
543 .owner = THIS_MODULE,
546 static int __init tbf_module_init(void)
548 return register_qdisc(&tbf_qdisc_ops);
551 static void __exit tbf_module_exit(void)
553 unregister_qdisc(&tbf_qdisc_ops);
555 module_init(tbf_module_init)
556 module_exit(tbf_module_exit)
557 MODULE_LICENSE("GPL");