1 /* 2 * net/sched/sch_tbf.c Token Bucket Filter queue. 3 * 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. 8 * 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 12 * 13 */ 14 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> 23 24 25 /* Simple Token Bucket Filter. 26 ======================================= 27 28 SOURCE. 29 ------- 30 31 None. 32 33 Description. 34 ------------ 35 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). 39 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: 43 44 s_i+....+s_k <= B + R*(t_k - t_i) 45 46 Algorithm. 47 ---------- 48 49 Let N(t_i) be B/R initially and N(t) grow continuously with time as: 50 51 N(t+delta) = min{B/R, N(t) + delta} 52 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: 56 57 N(t_* + 0) = N(t_* - 0) - S/R. 58 59 60 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. 65 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. 68 69 When TBF works in reshaping mode, latency is estimated as: 70 71 lat = max ((L-B)/R, (L-M)/P) 72 73 74 NOTES. 75 ------ 76 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. 83 84 85 This means, that with depth B, the maximal rate is 86 87 R_crit = B*HZ 88 89 F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes. 90 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 :-) 94 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. 98 */ 99 100 struct tbf_sched_data 101 { 102 /* Parameters */ 103 u32 limit; /* Maximal length of backlog: bytes */ 104 u32 buffer; /* Token bucket depth/rate: MUST BE >= MTU/B */ 105 u32 mtu; 106 u32 max_size; 107 struct qdisc_rate_table *R_tab; 108 struct qdisc_rate_table *P_tab; 109 110 /* Variables */ 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 */ 116 }; 117 118 #define L2T(q,L) qdisc_l2t((q)->R_tab,L) 119 #define L2T_P(q,L) qdisc_l2t((q)->P_tab,L) 120 121 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc* sch) 122 { 123 struct tbf_sched_data *q = qdisc_priv(sch); 124 int ret; 125 126 if (qdisc_pkt_len(skb) > q->max_size) 127 return qdisc_reshape_fail(skb, sch); 128 129 ret = qdisc_enqueue(skb, q->qdisc); 130 if (ret != NET_XMIT_SUCCESS) { 131 if (net_xmit_drop_count(ret)) 132 sch->qstats.drops++; 133 return ret; 134 } 135 136 sch->q.qlen++; 137 return NET_XMIT_SUCCESS; 138 } 139 140 static unsigned int tbf_drop(struct Qdisc* sch) 141 { 142 struct tbf_sched_data *q = qdisc_priv(sch); 143 unsigned int len = 0; 144 145 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) { 146 sch->q.qlen--; 147 sch->qstats.drops++; 148 } 149 return len; 150 } 151 152 static struct sk_buff *tbf_dequeue(struct Qdisc* sch) 153 { 154 struct tbf_sched_data *q = qdisc_priv(sch); 155 struct sk_buff *skb; 156 157 skb = q->qdisc->ops->peek(q->qdisc); 158 159 if (skb) { 160 psched_time_t now; 161 long toks; 162 long ptoks = 0; 163 unsigned int len = qdisc_pkt_len(skb); 164 165 now = psched_get_time(); 166 toks = psched_tdiff_bounded(now, q->t_c, q->buffer); 167 168 if (q->P_tab) { 169 ptoks = toks + q->ptokens; 170 if (ptoks > (long)q->mtu) 171 ptoks = q->mtu; 172 ptoks -= L2T_P(q, len); 173 } 174 toks += q->tokens; 175 if (toks > (long)q->buffer) 176 toks = q->buffer; 177 toks -= L2T(q, len); 178 179 if ((toks|ptoks) >= 0) { 180 skb = qdisc_dequeue_peeked(q->qdisc); 181 if (unlikely(!skb)) 182 return NULL; 183 184 q->t_c = now; 185 q->tokens = toks; 186 q->ptokens = ptoks; 187 sch->q.qlen--; 188 sch->flags &= ~TCQ_F_THROTTLED; 189 qdisc_bstats_update(sch, skb); 190 return skb; 191 } 192 193 qdisc_watchdog_schedule(&q->watchdog, 194 now + max_t(long, -toks, -ptoks)); 195 196 /* Maybe we have a shorter packet in the queue, 197 which can be sent now. It sounds cool, 198 but, however, this is wrong in principle. 199 We MUST NOT reorder packets under these circumstances. 200 201 Really, if we split the flow into independent 202 subflows, it would be a very good solution. 203 This is the main idea of all FQ algorithms 204 (cf. CSZ, HPFQ, HFSC) 205 */ 206 207 sch->qstats.overlimits++; 208 } 209 return NULL; 210 } 211 212 static void tbf_reset(struct Qdisc* sch) 213 { 214 struct tbf_sched_data *q = qdisc_priv(sch); 215 216 qdisc_reset(q->qdisc); 217 sch->q.qlen = 0; 218 q->t_c = psched_get_time(); 219 q->tokens = q->buffer; 220 q->ptokens = q->mtu; 221 qdisc_watchdog_cancel(&q->watchdog); 222 } 223 224 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = { 225 [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) }, 226 [TCA_TBF_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE }, 227 [TCA_TBF_PTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE }, 228 }; 229 230 static int tbf_change(struct Qdisc* sch, struct nlattr *opt) 231 { 232 int err; 233 struct tbf_sched_data *q = qdisc_priv(sch); 234 struct nlattr *tb[TCA_TBF_PTAB + 1]; 235 struct tc_tbf_qopt *qopt; 236 struct qdisc_rate_table *rtab = NULL; 237 struct qdisc_rate_table *ptab = NULL; 238 struct Qdisc *child = NULL; 239 int max_size,n; 240 241 err = nla_parse_nested(tb, TCA_TBF_PTAB, opt, tbf_policy); 242 if (err < 0) 243 return err; 244 245 err = -EINVAL; 246 if (tb[TCA_TBF_PARMS] == NULL) 247 goto done; 248 249 qopt = nla_data(tb[TCA_TBF_PARMS]); 250 rtab = qdisc_get_rtab(&qopt->rate, tb[TCA_TBF_RTAB]); 251 if (rtab == NULL) 252 goto done; 253 254 if (qopt->peakrate.rate) { 255 if (qopt->peakrate.rate > qopt->rate.rate) 256 ptab = qdisc_get_rtab(&qopt->peakrate, tb[TCA_TBF_PTAB]); 257 if (ptab == NULL) 258 goto done; 259 } 260 261 for (n = 0; n < 256; n++) 262 if (rtab->data[n] > qopt->buffer) break; 263 max_size = (n << qopt->rate.cell_log)-1; 264 if (ptab) { 265 int size; 266 267 for (n = 0; n < 256; n++) 268 if (ptab->data[n] > qopt->mtu) break; 269 size = (n << qopt->peakrate.cell_log)-1; 270 if (size < max_size) max_size = size; 271 } 272 if (max_size < 0) 273 goto done; 274 275 if (q->qdisc != &noop_qdisc) { 276 err = fifo_set_limit(q->qdisc, qopt->limit); 277 if (err) 278 goto done; 279 } else if (qopt->limit > 0) { 280 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit); 281 if (IS_ERR(child)) { 282 err = PTR_ERR(child); 283 goto done; 284 } 285 } 286 287 sch_tree_lock(sch); 288 if (child) { 289 qdisc_tree_decrease_qlen(q->qdisc, q->qdisc->q.qlen); 290 qdisc_destroy(q->qdisc); 291 q->qdisc = child; 292 } 293 q->limit = qopt->limit; 294 q->mtu = qopt->mtu; 295 q->max_size = max_size; 296 q->buffer = qopt->buffer; 297 q->tokens = q->buffer; 298 q->ptokens = q->mtu; 299 300 swap(q->R_tab, rtab); 301 swap(q->P_tab, ptab); 302 303 sch_tree_unlock(sch); 304 err = 0; 305 done: 306 if (rtab) 307 qdisc_put_rtab(rtab); 308 if (ptab) 309 qdisc_put_rtab(ptab); 310 return err; 311 } 312 313 static int tbf_init(struct Qdisc* sch, struct nlattr *opt) 314 { 315 struct tbf_sched_data *q = qdisc_priv(sch); 316 317 if (opt == NULL) 318 return -EINVAL; 319 320 q->t_c = psched_get_time(); 321 qdisc_watchdog_init(&q->watchdog, sch); 322 q->qdisc = &noop_qdisc; 323 324 return tbf_change(sch, opt); 325 } 326 327 static void tbf_destroy(struct Qdisc *sch) 328 { 329 struct tbf_sched_data *q = qdisc_priv(sch); 330 331 qdisc_watchdog_cancel(&q->watchdog); 332 333 if (q->P_tab) 334 qdisc_put_rtab(q->P_tab); 335 if (q->R_tab) 336 qdisc_put_rtab(q->R_tab); 337 338 qdisc_destroy(q->qdisc); 339 } 340 341 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb) 342 { 343 struct tbf_sched_data *q = qdisc_priv(sch); 344 struct nlattr *nest; 345 struct tc_tbf_qopt opt; 346 347 nest = nla_nest_start(skb, TCA_OPTIONS); 348 if (nest == NULL) 349 goto nla_put_failure; 350 351 opt.limit = q->limit; 352 opt.rate = q->R_tab->rate; 353 if (q->P_tab) 354 opt.peakrate = q->P_tab->rate; 355 else 356 memset(&opt.peakrate, 0, sizeof(opt.peakrate)); 357 opt.mtu = q->mtu; 358 opt.buffer = q->buffer; 359 NLA_PUT(skb, TCA_TBF_PARMS, sizeof(opt), &opt); 360 361 nla_nest_end(skb, nest); 362 return skb->len; 363 364 nla_put_failure: 365 nla_nest_cancel(skb, nest); 366 return -1; 367 } 368 369 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl, 370 struct sk_buff *skb, struct tcmsg *tcm) 371 { 372 struct tbf_sched_data *q = qdisc_priv(sch); 373 374 tcm->tcm_handle |= TC_H_MIN(1); 375 tcm->tcm_info = q->qdisc->handle; 376 377 return 0; 378 } 379 380 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 381 struct Qdisc **old) 382 { 383 struct tbf_sched_data *q = qdisc_priv(sch); 384 385 if (new == NULL) 386 new = &noop_qdisc; 387 388 sch_tree_lock(sch); 389 *old = q->qdisc; 390 q->qdisc = new; 391 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen); 392 qdisc_reset(*old); 393 sch_tree_unlock(sch); 394 395 return 0; 396 } 397 398 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg) 399 { 400 struct tbf_sched_data *q = qdisc_priv(sch); 401 return q->qdisc; 402 } 403 404 static unsigned long tbf_get(struct Qdisc *sch, u32 classid) 405 { 406 return 1; 407 } 408 409 static void tbf_put(struct Qdisc *sch, unsigned long arg) 410 { 411 } 412 413 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker) 414 { 415 if (!walker->stop) { 416 if (walker->count >= walker->skip) 417 if (walker->fn(sch, 1, walker) < 0) { 418 walker->stop = 1; 419 return; 420 } 421 walker->count++; 422 } 423 } 424 425 static const struct Qdisc_class_ops tbf_class_ops = 426 { 427 .graft = tbf_graft, 428 .leaf = tbf_leaf, 429 .get = tbf_get, 430 .put = tbf_put, 431 .walk = tbf_walk, 432 .dump = tbf_dump_class, 433 }; 434 435 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = { 436 .next = NULL, 437 .cl_ops = &tbf_class_ops, 438 .id = "tbf", 439 .priv_size = sizeof(struct tbf_sched_data), 440 .enqueue = tbf_enqueue, 441 .dequeue = tbf_dequeue, 442 .peek = qdisc_peek_dequeued, 443 .drop = tbf_drop, 444 .init = tbf_init, 445 .reset = tbf_reset, 446 .destroy = tbf_destroy, 447 .change = tbf_change, 448 .dump = tbf_dump, 449 .owner = THIS_MODULE, 450 }; 451 452 static int __init tbf_module_init(void) 453 { 454 return register_qdisc(&tbf_qdisc_ops); 455 } 456 457 static void __exit tbf_module_exit(void) 458 { 459 unregister_qdisc(&tbf_qdisc_ops); 460 } 461 module_init(tbf_module_init) 462 module_exit(tbf_module_exit) 463 MODULE_LICENSE("GPL"); 464