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/pkt_sched.h>
25 /* Simple Token Bucket Filter.
26 =======================================
36 A data flow obeys TBF with rate R and depth B, if for any
37 time interval t_i...t_f the number of transmitted bits
38 does not exceed B + R*(t_f-t_i).
40 Packetized version of this definition:
41 The sequence of packets of sizes s_i served at moments t_i
42 obeys TBF, if for any i<=k:
44 s_i+....+s_k <= B + R*(t_k - t_i)
49 Let N(t_i) be B/R initially and N(t) grow continuously with time as:
51 N(t+delta) = min{B/R, N(t) + delta}
53 If the first packet in queue has length S, it may be
54 transmitted only at the time t_* when S/R <= N(t_*),
55 and in this case N(t) jumps:
57 N(t_* + 0) = N(t_* - 0) - S/R.
61 Actually, QoS requires two TBF to be applied to a data stream.
62 One of them controls steady state burst size, another
63 one with rate P (peak rate) and depth M (equal to link MTU)
64 limits bursts at a smaller time scale.
66 It is easy to see that P>R, and B>M. If P is infinity, this double
67 TBF is equivalent to a single one.
69 When TBF works in reshaping mode, latency is estimated as:
71 lat = max ((L-B)/R, (L-M)/P)
77 If TBF throttles, it starts a watchdog timer, which will wake it up
78 when it is ready to transmit.
79 Note that the minimal timer resolution is 1/HZ.
80 If no new packets arrive during this period,
81 or if the device is not awaken by EOI for some previous packet,
82 TBF can stop its activity for 1/HZ.
85 This means, that with depth B, the maximal rate is
89 F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
91 Note that the peak rate TBF is much more tough: with MTU 1500
92 P_crit = 150Kbytes/sec. So, if you need greater peak
93 rates, use alpha with HZ=1000 :-)
95 With classful TBF, limit is just kept for backwards compatibility.
96 It is passed to the default bfifo qdisc - if the inner qdisc is
97 changed the limit is not effective anymore.
100 struct tbf_sched_data
103 u32 limit; /* Maximal length of backlog: bytes */
104 u32 buffer; /* Token bucket depth/rate: MUST BE >= MTU/B */
107 struct qdisc_rate_table *R_tab;
108 struct qdisc_rate_table *P_tab;
111 long tokens; /* Current number of B tokens */
112 long ptokens; /* Current number of P tokens */
113 psched_time_t t_c; /* Time check-point */
114 struct Qdisc *qdisc; /* Inner qdisc, default - bfifo queue */
115 struct qdisc_watchdog watchdog; /* Watchdog timer */
118 #define L2T(q,L) ((q)->R_tab->data[(L)>>(q)->R_tab->rate.cell_log])
119 #define L2T_P(q,L) ((q)->P_tab->data[(L)>>(q)->P_tab->rate.cell_log])
121 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc* sch)
123 struct tbf_sched_data *q = qdisc_priv(sch);
126 if (skb->len > q->max_size) {
128 #ifdef CONFIG_NET_CLS_ACT
129 if (sch->reshape_fail == NULL || sch->reshape_fail(skb, sch))
133 return NET_XMIT_DROP;
136 if ((ret = q->qdisc->enqueue(skb, q->qdisc)) != 0) {
142 sch->bstats.bytes += skb->len;
143 sch->bstats.packets++;
147 static int tbf_requeue(struct sk_buff *skb, struct Qdisc* sch)
149 struct tbf_sched_data *q = qdisc_priv(sch);
152 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
154 sch->qstats.requeues++;
160 static unsigned int tbf_drop(struct Qdisc* sch)
162 struct tbf_sched_data *q = qdisc_priv(sch);
163 unsigned int len = 0;
165 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
172 static struct sk_buff *tbf_dequeue(struct Qdisc* sch)
174 struct tbf_sched_data *q = qdisc_priv(sch);
177 skb = q->qdisc->dequeue(q->qdisc);
183 unsigned int len = skb->len;
185 now = psched_get_time();
186 toks = psched_tdiff_bounded(now, q->t_c, q->buffer);
189 ptoks = toks + q->ptokens;
190 if (ptoks > (long)q->mtu)
192 ptoks -= L2T_P(q, len);
195 if (toks > (long)q->buffer)
199 if ((toks|ptoks) >= 0) {
204 sch->flags &= ~TCQ_F_THROTTLED;
208 qdisc_watchdog_schedule(&q->watchdog,
209 now + max_t(long, -toks, -ptoks));
211 /* Maybe we have a shorter packet in the queue,
212 which can be sent now. It sounds cool,
213 but, however, this is wrong in principle.
214 We MUST NOT reorder packets under these circumstances.
216 Really, if we split the flow into independent
217 subflows, it would be a very good solution.
218 This is the main idea of all FQ algorithms
219 (cf. CSZ, HPFQ, HFSC)
222 if (q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS) {
223 /* When requeue fails skb is dropped */
224 qdisc_tree_decrease_qlen(q->qdisc, 1);
228 sch->qstats.overlimits++;
233 static void tbf_reset(struct Qdisc* sch)
235 struct tbf_sched_data *q = qdisc_priv(sch);
237 qdisc_reset(q->qdisc);
239 q->t_c = psched_get_time();
240 q->tokens = q->buffer;
242 qdisc_watchdog_cancel(&q->watchdog);
245 static struct Qdisc *tbf_create_dflt_qdisc(struct Qdisc *sch, u32 limit)
251 q = qdisc_create_dflt(sch->dev, &bfifo_qdisc_ops,
252 TC_H_MAKE(sch->handle, 1));
254 rta = kmalloc(RTA_LENGTH(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
256 rta->rta_type = RTM_NEWQDISC;
257 rta->rta_len = RTA_LENGTH(sizeof(struct tc_fifo_qopt));
258 ((struct tc_fifo_qopt *)RTA_DATA(rta))->limit = limit;
260 ret = q->ops->change(q, rta);
272 static int tbf_change(struct Qdisc* sch, struct rtattr *opt)
275 struct tbf_sched_data *q = qdisc_priv(sch);
276 struct rtattr *tb[TCA_TBF_PTAB];
277 struct tc_tbf_qopt *qopt;
278 struct qdisc_rate_table *rtab = NULL;
279 struct qdisc_rate_table *ptab = NULL;
280 struct Qdisc *child = NULL;
283 if (rtattr_parse_nested(tb, TCA_TBF_PTAB, opt) ||
284 tb[TCA_TBF_PARMS-1] == NULL ||
285 RTA_PAYLOAD(tb[TCA_TBF_PARMS-1]) < sizeof(*qopt))
288 qopt = RTA_DATA(tb[TCA_TBF_PARMS-1]);
289 rtab = qdisc_get_rtab(&qopt->rate, tb[TCA_TBF_RTAB-1]);
293 if (qopt->peakrate.rate) {
294 if (qopt->peakrate.rate > qopt->rate.rate)
295 ptab = qdisc_get_rtab(&qopt->peakrate, tb[TCA_TBF_PTAB-1]);
300 for (n = 0; n < 256; n++)
301 if (rtab->data[n] > qopt->buffer) break;
302 max_size = (n << qopt->rate.cell_log)-1;
306 for (n = 0; n < 256; n++)
307 if (ptab->data[n] > qopt->mtu) break;
308 size = (n << qopt->peakrate.cell_log)-1;
309 if (size < max_size) max_size = size;
314 if (qopt->limit > 0) {
315 if ((child = tbf_create_dflt_qdisc(sch, qopt->limit)) == NULL)
321 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen);
322 qdisc_destroy(xchg(&q->qdisc, child));
324 q->limit = qopt->limit;
326 q->max_size = max_size;
327 q->buffer = qopt->buffer;
328 q->tokens = q->buffer;
330 rtab = xchg(&q->R_tab, rtab);
331 ptab = xchg(&q->P_tab, ptab);
332 sch_tree_unlock(sch);
336 qdisc_put_rtab(rtab);
338 qdisc_put_rtab(ptab);
342 static int tbf_init(struct Qdisc* sch, struct rtattr *opt)
344 struct tbf_sched_data *q = qdisc_priv(sch);
349 q->t_c = psched_get_time();
350 qdisc_watchdog_init(&q->watchdog, sch);
351 q->qdisc = &noop_qdisc;
353 return tbf_change(sch, opt);
356 static void tbf_destroy(struct Qdisc *sch)
358 struct tbf_sched_data *q = qdisc_priv(sch);
360 qdisc_watchdog_cancel(&q->watchdog);
363 qdisc_put_rtab(q->P_tab);
365 qdisc_put_rtab(q->R_tab);
367 qdisc_destroy(q->qdisc);
370 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
372 struct tbf_sched_data *q = qdisc_priv(sch);
373 unsigned char *b = skb_tail_pointer(skb);
375 struct tc_tbf_qopt opt;
377 rta = (struct rtattr*)b;
378 RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
380 opt.limit = q->limit;
381 opt.rate = q->R_tab->rate;
383 opt.peakrate = q->P_tab->rate;
385 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
387 opt.buffer = q->buffer;
388 RTA_PUT(skb, TCA_TBF_PARMS, sizeof(opt), &opt);
389 rta->rta_len = skb_tail_pointer(skb) - b;
398 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
399 struct sk_buff *skb, struct tcmsg *tcm)
401 struct tbf_sched_data *q = qdisc_priv(sch);
403 if (cl != 1) /* only one class */
406 tcm->tcm_handle |= TC_H_MIN(1);
407 tcm->tcm_info = q->qdisc->handle;
412 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
415 struct tbf_sched_data *q = qdisc_priv(sch);
421 *old = xchg(&q->qdisc, new);
422 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
424 sch_tree_unlock(sch);
429 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
431 struct tbf_sched_data *q = qdisc_priv(sch);
435 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
440 static void tbf_put(struct Qdisc *sch, unsigned long arg)
444 static int tbf_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
445 struct rtattr **tca, unsigned long *arg)
450 static int tbf_delete(struct Qdisc *sch, unsigned long arg)
455 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
458 if (walker->count >= walker->skip)
459 if (walker->fn(sch, 1, walker) < 0) {
467 static struct tcf_proto **tbf_find_tcf(struct Qdisc *sch, unsigned long cl)
472 static struct Qdisc_class_ops tbf_class_ops =
478 .change = tbf_change_class,
479 .delete = tbf_delete,
481 .tcf_chain = tbf_find_tcf,
482 .dump = tbf_dump_class,
485 static struct Qdisc_ops tbf_qdisc_ops = {
487 .cl_ops = &tbf_class_ops,
489 .priv_size = sizeof(struct tbf_sched_data),
490 .enqueue = tbf_enqueue,
491 .dequeue = tbf_dequeue,
492 .requeue = tbf_requeue,
496 .destroy = tbf_destroy,
497 .change = tbf_change,
499 .owner = THIS_MODULE,
502 static int __init tbf_module_init(void)
504 return register_qdisc(&tbf_qdisc_ops);
507 static void __exit tbf_module_exit(void)
509 unregister_qdisc(&tbf_qdisc_ops);
511 module_init(tbf_module_init)
512 module_exit(tbf_module_exit)
513 MODULE_LICENSE("GPL");