net/sched/sch_tbf.c

   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/sch_generic.h>
  23 #include <net/pkt_sched.h>
  24
  25
  26 /*      Simple Token Bucket Filter.
  27         =======================================
  28
  29         SOURCE.
  30         -------
  31
  32         None.
  33
  34         Description.
  35         ------------
  36
  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).
  40
  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:
  44
  45         s_i+....+s_k <= B + R*(t_k - t_i)
  46
  47         Algorithm.
  48         ----------
  49
  50         Let N(t_i) be B/R initially and N(t) grow continuously with time as:
  51
  52         N(t+delta) = min{B/R, N(t) + delta}
  53
  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:
  57
  58         N(t_* + 0) = N(t_* - 0) - S/R.
  59
  60
  61
  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.
  66
  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.
  69
  70         When TBF works in reshaping mode, latency is estimated as:
  71
  72         lat = max ((L-B)/R, (L-M)/P)
  73
  74
  75         NOTES.
  76         ------
  77
  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.
  84
  85
  86         This means, that with depth B, the maximal rate is
  87
  88         R_crit = B*HZ
  89
  90         F.e. for 10Mbit ethernet and HZ=100 the minimal allowed B is ~10Kbytes.
  91
  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 :-)
  95
  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.
  99 */
 100
 101 struct tbf_sched_data {
 102 /* Parameters */
 103         u32             limit;          /* Maximal length of backlog: bytes */
 104         u32             max_size;
 105         s64             buffer;         /* Token bucket depth/rate: MUST BE >= MTU/B */
 106         s64             mtu;
 107         struct psched_ratecfg rate;
 108         struct psched_ratecfg peak;
 109
 110 /* Variables */
 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 */
 116 };
 117
 118
 119 /* Time to Length, convert time in ns to length in bytes
 120  * to determinate how many bytes can be sent in given time.
 121  */
 122 static u64 psched_ns_t2l(const struct psched_ratecfg *r,
 123                          u64 time_in_ns)
 124 {
 125         /* The formula is :
 126          * len = (time_in_ns * r->rate_bytes_ps) / NSEC_PER_SEC
 127          */
 128         u64 len = time_in_ns * r->rate_bytes_ps;
 129
 130         do_div(len, NSEC_PER_SEC);
 131
 132         if (unlikely(r->linklayer == TC_LINKLAYER_ATM)) {
 133                 do_div(len, 53);
 134                 len = len * 48;
 135         }
 136
 137         if (len > r->overhead)
 138                 len -= r->overhead;
 139         else
 140                 len = 0;
 141
 142         return len;
 143 }
 144
 145 /*
 146  * Return length of individual segments of a gso packet,
 147  * including all headers (MAC, IP, TCP/UDP)
 148  */
 149 static unsigned int skb_gso_mac_seglen(const struct sk_buff *skb)
 150 {
 151         unsigned int hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
 152         return hdr_len + skb_gso_transport_seglen(skb);
 153 }
 154
 155 /* GSO packet is too big, segment it so that tbf can transmit
 156  * each segment in time
 157  */
 158 static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch)
 159 {
 160         struct tbf_sched_data *q = qdisc_priv(sch);
 161         struct sk_buff *segs, *nskb;
 162         netdev_features_t features = netif_skb_features(skb);
 163         unsigned int len = 0, prev_len = qdisc_pkt_len(skb);
 164         int ret, nb;
 165
 166         segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 167
 168         if (IS_ERR_OR_NULL(segs))
 169                 return qdisc_reshape_fail(skb, sch);
 170
 171         nb = 0;
 172         while (segs) {
 173                 nskb = segs->next;
 174                 segs->next = NULL;
 175                 qdisc_skb_cb(segs)->pkt_len = segs->len;
 176                 len += segs->len;
 177                 ret = qdisc_enqueue(segs, q->qdisc);
 178                 if (ret != NET_XMIT_SUCCESS) {
 179                         if (net_xmit_drop_count(ret))
 180                                 qdisc_qstats_drop(sch);
 181                 } else {
 182                         nb++;
 183                 }
 184                 segs = nskb;
 185         }
 186         sch->q.qlen += nb;
 187         if (nb > 1)
 188                 qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
 189         consume_skb(skb);
 190         return nb > 0 ? NET_XMIT_SUCCESS : NET_XMIT_DROP;
 191 }
 192
 193 static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 194 {
 195         struct tbf_sched_data *q = qdisc_priv(sch);
 196         int ret;
 197
 198         if (qdisc_pkt_len(skb) > q->max_size) {
 199                 if (skb_is_gso(skb) && skb_gso_mac_seglen(skb) <= q->max_size)
 200                         return tbf_segment(skb, sch);
 201                 return qdisc_reshape_fail(skb, sch);
 202         }
 203         ret = qdisc_enqueue(skb, q->qdisc);
 204         if (ret != NET_XMIT_SUCCESS) {
 205                 if (net_xmit_drop_count(ret))
 206                         qdisc_qstats_drop(sch);
 207                 return ret;
 208         }
 209
 210         qdisc_qstats_backlog_inc(sch, skb);
 211         sch->q.qlen++;
 212         return NET_XMIT_SUCCESS;
 213 }
 214
 215 static unsigned int tbf_drop(struct Qdisc *sch)
 216 {
 217         struct tbf_sched_data *q = qdisc_priv(sch);
 218         unsigned int len = 0;
 219
 220         if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
 221                 sch->qstats.backlog -= len;
 222                 sch->q.qlen--;
 223                 qdisc_qstats_drop(sch);
 224         }
 225         return len;
 226 }
 227
 228 static bool tbf_peak_present(const struct tbf_sched_data *q)
 229 {
 230         return q->peak.rate_bytes_ps;
 231 }
 232
 233 static struct sk_buff *tbf_dequeue(struct Qdisc *sch)
 234 {
 235         struct tbf_sched_data *q = qdisc_priv(sch);
 236         struct sk_buff *skb;
 237
 238         skb = q->qdisc->ops->peek(q->qdisc);
 239
 240         if (skb) {
 241                 s64 now;
 242                 s64 toks;
 243                 s64 ptoks = 0;
 244                 unsigned int len = qdisc_pkt_len(skb);
 245
 246                 now = ktime_get_ns();
 247                 toks = min_t(s64, now - q->t_c, q->buffer);
 248
 249                 if (tbf_peak_present(q)) {
 250                         ptoks = toks + q->ptokens;
 251                         if (ptoks > q->mtu)
 252                                 ptoks = q->mtu;
 253                         ptoks -= (s64) psched_l2t_ns(&q->peak, len);
 254                 }
 255                 toks += q->tokens;
 256                 if (toks > q->buffer)
 257                         toks = q->buffer;
 258                 toks -= (s64) psched_l2t_ns(&q->rate, len);
 259
 260                 if ((toks|ptoks) >= 0) {
 261                         skb = qdisc_dequeue_peeked(q->qdisc);
 262                         if (unlikely(!skb))
 263                                 return NULL;
 264
 265                         q->t_c = now;
 266                         q->tokens = toks;
 267                         q->ptokens = ptoks;
 268                         qdisc_qstats_backlog_dec(sch, skb);
 269                         sch->q.qlen--;
 270                         qdisc_unthrottled(sch);
 271                         qdisc_bstats_update(sch, skb);
 272                         return skb;
 273                 }
 274
 275                 qdisc_watchdog_schedule_ns(&q->watchdog,
 276                                            now + max_t(long, -toks, -ptoks),
 277                                            true);
 278
 279                 /* Maybe we have a shorter packet in the queue,
 280                    which can be sent now. It sounds cool,
 281                    but, however, this is wrong in principle.
 282                    We MUST NOT reorder packets under these circumstances.
 283
 284                    Really, if we split the flow into independent
 285                    subflows, it would be a very good solution.
 286                    This is the main idea of all FQ algorithms
 287                    (cf. CSZ, HPFQ, HFSC)
 288                  */
 289
 290                 qdisc_qstats_overlimit(sch);
 291         }
 292         return NULL;
 293 }
 294
 295 static void tbf_reset(struct Qdisc *sch)
 296 {
 297         struct tbf_sched_data *q = qdisc_priv(sch);
 298
 299         qdisc_reset(q->qdisc);
 300         sch->qstats.backlog = 0;
 301         sch->q.qlen = 0;
 302         q->t_c = ktime_get_ns();
 303         q->tokens = q->buffer;
 304         q->ptokens = q->mtu;
 305         qdisc_watchdog_cancel(&q->watchdog);
 306 }
 307
 308 static const struct nla_policy tbf_policy[TCA_TBF_MAX + 1] = {
 309         [TCA_TBF_PARMS] = { .len = sizeof(struct tc_tbf_qopt) },
 310         [TCA_TBF_RTAB]  = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
 311         [TCA_TBF_PTAB]  = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
 312         [TCA_TBF_RATE64]        = { .type = NLA_U64 },
 313         [TCA_TBF_PRATE64]       = { .type = NLA_U64 },
 314         [TCA_TBF_BURST] = { .type = NLA_U32 },
 315         [TCA_TBF_PBURST] = { .type = NLA_U32 },
 316 };
 317
 318 static int tbf_change(struct Qdisc *sch, struct nlattr *opt)
 319 {
 320         int err;
 321         struct tbf_sched_data *q = qdisc_priv(sch);
 322         struct nlattr *tb[TCA_TBF_MAX + 1];
 323         struct tc_tbf_qopt *qopt;
 324         struct Qdisc *child = NULL;
 325         struct psched_ratecfg rate;
 326         struct psched_ratecfg peak;
 327         u64 max_size;
 328         s64 buffer, mtu;
 329         u64 rate64 = 0, prate64 = 0;
 330
 331         err = nla_parse_nested(tb, TCA_TBF_MAX, opt, tbf_policy);
 332         if (err < 0)
 333                 return err;
 334
 335         err = -EINVAL;
 336         if (tb[TCA_TBF_PARMS] == NULL)
 337                 goto done;
 338
 339         qopt = nla_data(tb[TCA_TBF_PARMS]);
 340         if (qopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
 341                 qdisc_put_rtab(qdisc_get_rtab(&qopt->rate,
 342                                               tb[TCA_TBF_RTAB]));
 343
 344         if (qopt->peakrate.linklayer == TC_LINKLAYER_UNAWARE)
 345                         qdisc_put_rtab(qdisc_get_rtab(&qopt->peakrate,
 346                                                       tb[TCA_TBF_PTAB]));
 347
 348         buffer = min_t(u64, PSCHED_TICKS2NS(qopt->buffer), ~0U);
 349         mtu = min_t(u64, PSCHED_TICKS2NS(qopt->mtu), ~0U);
 350
 351         if (tb[TCA_TBF_RATE64])
 352                 rate64 = nla_get_u64(tb[TCA_TBF_RATE64]);
 353         psched_ratecfg_precompute(&rate, &qopt->rate, rate64);
 354
 355         if (tb[TCA_TBF_BURST]) {
 356                 max_size = nla_get_u32(tb[TCA_TBF_BURST]);
 357                 buffer = psched_l2t_ns(&rate, max_size);
 358         } else {
 359                 max_size = min_t(u64, psched_ns_t2l(&rate, buffer), ~0U);
 360         }
 361
 362         if (qopt->peakrate.rate) {
 363                 if (tb[TCA_TBF_PRATE64])
 364                         prate64 = nla_get_u64(tb[TCA_TBF_PRATE64]);
 365                 psched_ratecfg_precompute(&peak, &qopt->peakrate, prate64);
 366                 if (peak.rate_bytes_ps <= rate.rate_bytes_ps) {
 367                         pr_warn_ratelimited("sch_tbf: peakrate %llu is lower than or equals to rate %llu !\n",
 368                                         peak.rate_bytes_ps, rate.rate_bytes_ps);
 369                         err = -EINVAL;
 370                         goto done;
 371                 }
 372
 373                 if (tb[TCA_TBF_PBURST]) {
 374                         u32 pburst = nla_get_u32(tb[TCA_TBF_PBURST]);
 375                         max_size = min_t(u32, max_size, pburst);
 376                         mtu = psched_l2t_ns(&peak, pburst);
 377                 } else {
 378                         max_size = min_t(u64, max_size, psched_ns_t2l(&peak, mtu));
 379                 }
 380         } else {
 381                 memset(&peak, 0, sizeof(peak));
 382         }
 383
 384         if (max_size < psched_mtu(qdisc_dev(sch)))
 385                 pr_warn_ratelimited("sch_tbf: burst %llu is lower than device %s mtu (%u) !\n",
 386                                     max_size, qdisc_dev(sch)->name,
 387                                     psched_mtu(qdisc_dev(sch)));
 388
 389         if (!max_size) {
 390                 err = -EINVAL;
 391                 goto done;
 392         }
 393
 394         if (q->qdisc != &noop_qdisc) {
 395                 err = fifo_set_limit(q->qdisc, qopt->limit);
 396                 if (err)
 397                         goto done;
 398         } else if (qopt->limit > 0) {
 399                 child = fifo_create_dflt(sch, &bfifo_qdisc_ops, qopt->limit);
 400                 if (IS_ERR(child)) {
 401                         err = PTR_ERR(child);
 402                         goto done;
 403                 }
 404         }
 405
 406         sch_tree_lock(sch);
 407         if (child) {
 408                 qdisc_tree_reduce_backlog(q->qdisc, q->qdisc->q.qlen,
 409                                           q->qdisc->qstats.backlog);
 410                 qdisc_destroy(q->qdisc);
 411                 q->qdisc = child;
 412         }
 413         q->limit = qopt->limit;
 414         if (tb[TCA_TBF_PBURST])
 415                 q->mtu = mtu;
 416         else
 417                 q->mtu = PSCHED_TICKS2NS(qopt->mtu);
 418         q->max_size = max_size;
 419         if (tb[TCA_TBF_BURST])
 420                 q->buffer = buffer;
 421         else
 422                 q->buffer = PSCHED_TICKS2NS(qopt->buffer);
 423         q->tokens = q->buffer;
 424         q->ptokens = q->mtu;
 425
 426         memcpy(&q->rate, &rate, sizeof(struct psched_ratecfg));
 427         memcpy(&q->peak, &peak, sizeof(struct psched_ratecfg));
 428
 429         sch_tree_unlock(sch);
 430         err = 0;
 431 done:
 432         return err;
 433 }
 434
 435 static int tbf_init(struct Qdisc *sch, struct nlattr *opt)
 436 {
 437         struct tbf_sched_data *q = qdisc_priv(sch);
 438
 439         if (opt == NULL)
 440                 return -EINVAL;
 441
 442         q->t_c = ktime_get_ns();
 443         qdisc_watchdog_init(&q->watchdog, sch);
 444         q->qdisc = &noop_qdisc;
 445
 446         return tbf_change(sch, opt);
 447 }
 448
 449 static void tbf_destroy(struct Qdisc *sch)
 450 {
 451         struct tbf_sched_data *q = qdisc_priv(sch);
 452
 453         qdisc_watchdog_cancel(&q->watchdog);
 454         qdisc_destroy(q->qdisc);
 455 }
 456
 457 static int tbf_dump(struct Qdisc *sch, struct sk_buff *skb)
 458 {
 459         struct tbf_sched_data *q = qdisc_priv(sch);
 460         struct nlattr *nest;
 461         struct tc_tbf_qopt opt;
 462
 463         sch->qstats.backlog = q->qdisc->qstats.backlog;
 464         nest = nla_nest_start(skb, TCA_OPTIONS);
 465         if (nest == NULL)
 466                 goto nla_put_failure;
 467
 468         opt.limit = q->limit;
 469         psched_ratecfg_getrate(&opt.rate, &q->rate);
 470         if (tbf_peak_present(q))
 471                 psched_ratecfg_getrate(&opt.peakrate, &q->peak);
 472         else
 473                 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
 474         opt.mtu = PSCHED_NS2TICKS(q->mtu);
 475         opt.buffer = PSCHED_NS2TICKS(q->buffer);
 476         if (nla_put(skb, TCA_TBF_PARMS, sizeof(opt), &opt))
 477                 goto nla_put_failure;
 478         if (q->rate.rate_bytes_ps >= (1ULL << 32) &&
 479             nla_put_u64_64bit(skb, TCA_TBF_RATE64, q->rate.rate_bytes_ps,
 480                               TCA_TBF_PAD))
 481                 goto nla_put_failure;
 482         if (tbf_peak_present(q) &&
 483             q->peak.rate_bytes_ps >= (1ULL << 32) &&
 484             nla_put_u64_64bit(skb, TCA_TBF_PRATE64, q->peak.rate_bytes_ps,
 485                               TCA_TBF_PAD))
 486                 goto nla_put_failure;
 487
 488         return nla_nest_end(skb, nest);
 489
 490 nla_put_failure:
 491         nla_nest_cancel(skb, nest);
 492         return -1;
 493 }
 494
 495 static int tbf_dump_class(struct Qdisc *sch, unsigned long cl,
 496                           struct sk_buff *skb, struct tcmsg *tcm)
 497 {
 498         struct tbf_sched_data *q = qdisc_priv(sch);
 499
 500         tcm->tcm_handle |= TC_H_MIN(1);
 501         tcm->tcm_info = q->qdisc->handle;
 502
 503         return 0;
 504 }
 505
 506 static int tbf_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 507                      struct Qdisc **old)
 508 {
 509         struct tbf_sched_data *q = qdisc_priv(sch);
 510
 511         if (new == NULL)
 512                 new = &noop_qdisc;
 513
 514         *old = qdisc_replace(sch, new, &q->qdisc);
 515         return 0;
 516 }
 517
 518 static struct Qdisc *tbf_leaf(struct Qdisc *sch, unsigned long arg)
 519 {
 520         struct tbf_sched_data *q = qdisc_priv(sch);
 521         return q->qdisc;
 522 }
 523
 524 static unsigned long tbf_get(struct Qdisc *sch, u32 classid)
 525 {
 526         return 1;
 527 }
 528
 529 static void tbf_put(struct Qdisc *sch, unsigned long arg)
 530 {
 531 }
 532
 533 static void tbf_walk(struct Qdisc *sch, struct qdisc_walker *walker)
 534 {
 535         if (!walker->stop) {
 536                 if (walker->count >= walker->skip)
 537                         if (walker->fn(sch, 1, walker) < 0) {
 538                                 walker->stop = 1;
 539                                 return;
 540                         }
 541                 walker->count++;
 542         }
 543 }
 544
 545 static const struct Qdisc_class_ops tbf_class_ops = {
 546         .graft          =       tbf_graft,
 547         .leaf           =       tbf_leaf,
 548         .get            =       tbf_get,
 549         .put            =       tbf_put,
 550         .walk           =       tbf_walk,
 551         .dump           =       tbf_dump_class,
 552 };
 553
 554 static struct Qdisc_ops tbf_qdisc_ops __read_mostly = {
 555         .next           =       NULL,
 556         .cl_ops         =       &tbf_class_ops,
 557         .id             =       "tbf",
 558         .priv_size      =       sizeof(struct tbf_sched_data),
 559         .enqueue        =       tbf_enqueue,
 560         .dequeue        =       tbf_dequeue,
 561         .peek           =       qdisc_peek_dequeued,
 562         .drop           =       tbf_drop,
 563         .init           =       tbf_init,
 564         .reset          =       tbf_reset,
 565         .destroy        =       tbf_destroy,
 566         .change         =       tbf_change,
 567         .dump           =       tbf_dump,
 568         .owner          =       THIS_MODULE,
 569 };
 570
 571 static int __init tbf_module_init(void)
 572 {
 573         return register_qdisc(&tbf_qdisc_ops);
 574 }
 575
 576 static void __exit tbf_module_exit(void)
 577 {
 578         unregister_qdisc(&tbf_qdisc_ops);
 579 }
 580 module_init(tbf_module_init)
 581 module_exit(tbf_module_exit)
 582 MODULE_LICENSE("GPL");