net/sched/sch_cbq.c

   1 /*
   2  * net/sched/sch_cbq.c  Class-Based Queueing discipline.
   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  *
  11  */
  12
  13 #include <linux/module.h>
  14 #include <linux/slab.h>
  15 #include <linux/types.h>
  16 #include <linux/kernel.h>
  17 #include <linux/string.h>
  18 #include <linux/errno.h>
  19 #include <linux/skbuff.h>
  20 #include <net/netlink.h>
  21 #include <net/pkt_sched.h>
  22 #include <net/pkt_cls.h>
  23
  24
  25 /*      Class-Based Queueing (CBQ) algorithm.
  26         =======================================
  27
  28         Sources: [1] Sally Floyd and Van Jacobson, "Link-sharing and Resource
  29                  Management Models for Packet Networks",
  30                  IEEE/ACM Transactions on Networking, Vol.3, No.4, 1995
  31
  32                  [2] Sally Floyd, "Notes on CBQ and Guaranteed Service", 1995
  33
  34                  [3] Sally Floyd, "Notes on Class-Based Queueing: Setting
  35                  Parameters", 1996
  36
  37                  [4] Sally Floyd and Michael Speer, "Experimental Results
  38                  for Class-Based Queueing", 1998, not published.
  39
  40         -----------------------------------------------------------------------
  41
  42         Algorithm skeleton was taken from NS simulator cbq.cc.
  43         If someone wants to check this code against the LBL version,
  44         he should take into account that ONLY the skeleton was borrowed,
  45         the implementation is different. Particularly:
  46
  47         --- The WRR algorithm is different. Our version looks more
  48         reasonable (I hope) and works when quanta are allowed to be
  49         less than MTU, which is always the case when real time classes
  50         have small rates. Note, that the statement of [3] is
  51         incomplete, delay may actually be estimated even if class
  52         per-round allotment is less than MTU. Namely, if per-round
  53         allotment is W*r_i, and r_1+...+r_k = r < 1
  54
  55         delay_i <= ([MTU/(W*r_i)]*W*r + W*r + k*MTU)/B
  56
  57         In the worst case we have IntServ estimate with D = W*r+k*MTU
  58         and C = MTU*r. The proof (if correct at all) is trivial.
  59
  60
  61         --- It seems that cbq-2.0 is not very accurate. At least, I cannot
  62         interpret some places, which look like wrong translations
  63         from NS. Anyone is advised to find these differences
  64         and explain to me, why I am wrong 8).
  65
  66         --- Linux has no EOI event, so that we cannot estimate true class
  67         idle time. Workaround is to consider the next dequeue event
  68         as sign that previous packet is finished. This is wrong because of
  69         internal device queueing, but on a permanently loaded link it is true.
  70         Moreover, combined with clock integrator, this scheme looks
  71         very close to an ideal solution.  */
  72
  73 struct cbq_sched_data;
  74
  75
  76 struct cbq_class {
  77         struct Qdisc_class_common common;
  78         struct cbq_class        *next_alive;    /* next class with backlog in this priority band */
  79
  80 /* Parameters */
  81         unsigned char           priority;       /* class priority */
  82         unsigned char           priority2;      /* priority to be used after overlimit */
  83         unsigned char           ewma_log;       /* time constant for idle time calculation */
  84
  85         u32                     defmap;
  86
  87         /* Link-sharing scheduler parameters */
  88         long                    maxidle;        /* Class parameters: see below. */
  89         long                    offtime;
  90         long                    minidle;
  91         u32                     avpkt;
  92         struct qdisc_rate_table *R_tab;
  93
  94         /* General scheduler (WRR) parameters */
  95         long                    allot;
  96         long                    quantum;        /* Allotment per WRR round */
  97         long                    weight;         /* Relative allotment: see below */
  98
  99         struct Qdisc            *qdisc;         /* Ptr to CBQ discipline */
 100         struct cbq_class        *split;         /* Ptr to split node */
 101         struct cbq_class        *share;         /* Ptr to LS parent in the class tree */
 102         struct cbq_class        *tparent;       /* Ptr to tree parent in the class tree */
 103         struct cbq_class        *borrow;        /* NULL if class is bandwidth limited;
 104                                                    parent otherwise */
 105         struct cbq_class        *sibling;       /* Sibling chain */
 106         struct cbq_class        *children;      /* Pointer to children chain */
 107
 108         struct Qdisc            *q;             /* Elementary queueing discipline */
 109
 110
 111 /* Variables */
 112         unsigned char           cpriority;      /* Effective priority */
 113         unsigned char           delayed;
 114         unsigned char           level;          /* level of the class in hierarchy:
 115                                                    0 for leaf classes, and maximal
 116                                                    level of children + 1 for nodes.
 117                                                  */
 118
 119         psched_time_t           last;           /* Last end of service */
 120         psched_time_t           undertime;
 121         long                    avgidle;
 122         long                    deficit;        /* Saved deficit for WRR */
 123         psched_time_t           penalized;
 124         struct gnet_stats_basic_packed bstats;
 125         struct gnet_stats_queue qstats;
 126         struct net_rate_estimator __rcu *rate_est;
 127         struct tc_cbq_xstats    xstats;
 128
 129         struct tcf_proto __rcu  *filter_list;
 130         struct tcf_block        *block;
 131
 132         int                     refcnt;
 133         int                     filters;
 134
 135         struct cbq_class        *defaults[TC_PRIO_MAX + 1];
 136 };
 137
 138 struct cbq_sched_data {
 139         struct Qdisc_class_hash clhash;                 /* Hash table of all classes */
 140         int                     nclasses[TC_CBQ_MAXPRIO + 1];
 141         unsigned int            quanta[TC_CBQ_MAXPRIO + 1];
 142
 143         struct cbq_class        link;
 144
 145         unsigned int            activemask;
 146         struct cbq_class        *active[TC_CBQ_MAXPRIO + 1];    /* List of all classes
 147                                                                    with backlog */
 148
 149 #ifdef CONFIG_NET_CLS_ACT
 150         struct cbq_class        *rx_class;
 151 #endif
 152         struct cbq_class        *tx_class;
 153         struct cbq_class        *tx_borrowed;
 154         int                     tx_len;
 155         psched_time_t           now;            /* Cached timestamp */
 156         unsigned int            pmask;
 157
 158         struct hrtimer          delay_timer;
 159         struct qdisc_watchdog   watchdog;       /* Watchdog timer,
 160                                                    started when CBQ has
 161                                                    backlog, but cannot
 162                                                    transmit just now */
 163         psched_tdiff_t          wd_expires;
 164         int                     toplevel;
 165         u32                     hgenerator;
 166 };
 167
 168
 169 #define L2T(cl, len)    qdisc_l2t((cl)->R_tab, len)
 170
 171 static inline struct cbq_class *
 172 cbq_class_lookup(struct cbq_sched_data *q, u32 classid)
 173 {
 174         struct Qdisc_class_common *clc;
 175
 176         clc = qdisc_class_find(&q->clhash, classid);
 177         if (clc == NULL)
 178                 return NULL;
 179         return container_of(clc, struct cbq_class, common);
 180 }
 181
 182 #ifdef CONFIG_NET_CLS_ACT
 183
 184 static struct cbq_class *
 185 cbq_reclassify(struct sk_buff *skb, struct cbq_class *this)
 186 {
 187         struct cbq_class *cl;
 188
 189         for (cl = this->tparent; cl; cl = cl->tparent) {
 190                 struct cbq_class *new = cl->defaults[TC_PRIO_BESTEFFORT];
 191
 192                 if (new != NULL && new != this)
 193                         return new;
 194         }
 195         return NULL;
 196 }
 197
 198 #endif
 199
 200 /* Classify packet. The procedure is pretty complicated, but
 201  * it allows us to combine link sharing and priority scheduling
 202  * transparently.
 203  *
 204  * Namely, you can put link sharing rules (f.e. route based) at root of CBQ,
 205  * so that it resolves to split nodes. Then packets are classified
 206  * by logical priority, or a more specific classifier may be attached
 207  * to the split node.
 208  */
 209
 210 static struct cbq_class *
 211 cbq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
 212 {
 213         struct cbq_sched_data *q = qdisc_priv(sch);
 214         struct cbq_class *head = &q->link;
 215         struct cbq_class **defmap;
 216         struct cbq_class *cl = NULL;
 217         u32 prio = skb->priority;
 218         struct tcf_proto *fl;
 219         struct tcf_result res;
 220
 221         /*
 222          *  Step 1. If skb->priority points to one of our classes, use it.
 223          */
 224         if (TC_H_MAJ(prio ^ sch->handle) == 0 &&
 225             (cl = cbq_class_lookup(q, prio)) != NULL)
 226                 return cl;
 227
 228         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 229         for (;;) {
 230                 int result = 0;
 231                 defmap = head->defaults;
 232
 233                 fl = rcu_dereference_bh(head->filter_list);
 234                 /*
 235                  * Step 2+n. Apply classifier.
 236                  */
 237                 result = tcf_classify(skb, fl, &res, true);
 238                 if (!fl || result < 0)
 239                         goto fallback;
 240
 241                 cl = (void *)res.class;
 242                 if (!cl) {
 243                         if (TC_H_MAJ(res.classid))
 244                                 cl = cbq_class_lookup(q, res.classid);
 245                         else if ((cl = defmap[res.classid & TC_PRIO_MAX]) == NULL)
 246                                 cl = defmap[TC_PRIO_BESTEFFORT];
 247
 248                         if (cl == NULL)
 249                                 goto fallback;
 250                 }
 251                 if (cl->level >= head->level)
 252                         goto fallback;
 253 #ifdef CONFIG_NET_CLS_ACT
 254                 switch (result) {
 255                 case TC_ACT_QUEUED:
 256                 case TC_ACT_STOLEN:
 257                 case TC_ACT_TRAP:
 258                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 259                 case TC_ACT_SHOT:
 260                         return NULL;
 261                 case TC_ACT_RECLASSIFY:
 262                         return cbq_reclassify(skb, cl);
 263                 }
 264 #endif
 265                 if (cl->level == 0)
 266                         return cl;
 267
 268                 /*
 269                  * Step 3+n. If classifier selected a link sharing class,
 270                  *         apply agency specific classifier.
 271                  *         Repeat this procdure until we hit a leaf node.
 272                  */
 273                 head = cl;
 274         }
 275
 276 fallback:
 277         cl = head;
 278
 279         /*
 280          * Step 4. No success...
 281          */
 282         if (TC_H_MAJ(prio) == 0 &&
 283             !(cl = head->defaults[prio & TC_PRIO_MAX]) &&
 284             !(cl = head->defaults[TC_PRIO_BESTEFFORT]))
 285                 return head;
 286
 287         return cl;
 288 }
 289
 290 /*
 291  * A packet has just been enqueued on the empty class.
 292  * cbq_activate_class adds it to the tail of active class list
 293  * of its priority band.
 294  */
 295
 296 static inline void cbq_activate_class(struct cbq_class *cl)
 297 {
 298         struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 299         int prio = cl->cpriority;
 300         struct cbq_class *cl_tail;
 301
 302         cl_tail = q->active[prio];
 303         q->active[prio] = cl;
 304
 305         if (cl_tail != NULL) {
 306                 cl->next_alive = cl_tail->next_alive;
 307                 cl_tail->next_alive = cl;
 308         } else {
 309                 cl->next_alive = cl;
 310                 q->activemask |= (1<<prio);
 311         }
 312 }
 313
 314 /*
 315  * Unlink class from active chain.
 316  * Note that this same procedure is done directly in cbq_dequeue*
 317  * during round-robin procedure.
 318  */
 319
 320 static void cbq_deactivate_class(struct cbq_class *this)
 321 {
 322         struct cbq_sched_data *q = qdisc_priv(this->qdisc);
 323         int prio = this->cpriority;
 324         struct cbq_class *cl;
 325         struct cbq_class *cl_prev = q->active[prio];
 326
 327         do {
 328                 cl = cl_prev->next_alive;
 329                 if (cl == this) {
 330                         cl_prev->next_alive = cl->next_alive;
 331                         cl->next_alive = NULL;
 332
 333                         if (cl == q->active[prio]) {
 334                                 q->active[prio] = cl_prev;
 335                                 if (cl == q->active[prio]) {
 336                                         q->active[prio] = NULL;
 337                                         q->activemask &= ~(1<<prio);
 338                                         return;
 339                                 }
 340                         }
 341                         return;
 342                 }
 343         } while ((cl_prev = cl) != q->active[prio]);
 344 }
 345
 346 static void
 347 cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl)
 348 {
 349         int toplevel = q->toplevel;
 350
 351         if (toplevel > cl->level) {
 352                 psched_time_t now = psched_get_time();
 353
 354                 do {
 355                         if (cl->undertime < now) {
 356                                 q->toplevel = cl->level;
 357                                 return;
 358                         }
 359                 } while ((cl = cl->borrow) != NULL && toplevel > cl->level);
 360         }
 361 }
 362
 363 static int
 364 cbq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 365             struct sk_buff **to_free)
 366 {
 367         struct cbq_sched_data *q = qdisc_priv(sch);
 368         int uninitialized_var(ret);
 369         struct cbq_class *cl = cbq_classify(skb, sch, &ret);
 370
 371 #ifdef CONFIG_NET_CLS_ACT
 372         q->rx_class = cl;
 373 #endif
 374         if (cl == NULL) {
 375                 if (ret & __NET_XMIT_BYPASS)
 376                         qdisc_qstats_drop(sch);
 377                 __qdisc_drop(skb, to_free);
 378                 return ret;
 379         }
 380
 381         ret = qdisc_enqueue(skb, cl->q, to_free);
 382         if (ret == NET_XMIT_SUCCESS) {
 383                 sch->q.qlen++;
 384                 cbq_mark_toplevel(q, cl);
 385                 if (!cl->next_alive)
 386                         cbq_activate_class(cl);
 387                 return ret;
 388         }
 389
 390         if (net_xmit_drop_count(ret)) {
 391                 qdisc_qstats_drop(sch);
 392                 cbq_mark_toplevel(q, cl);
 393                 cl->qstats.drops++;
 394         }
 395         return ret;
 396 }
 397
 398 /* Overlimit action: penalize leaf class by adding offtime */
 399 static void cbq_overlimit(struct cbq_class *cl)
 400 {
 401         struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 402         psched_tdiff_t delay = cl->undertime - q->now;
 403
 404         if (!cl->delayed) {
 405                 delay += cl->offtime;
 406
 407                 /*
 408                  * Class goes to sleep, so that it will have no
 409                  * chance to work avgidle. Let's forgive it 8)
 410                  *
 411                  * BTW cbq-2.0 has a crap in this
 412                  * place, apparently they forgot to shift it by cl->ewma_log.
 413                  */
 414                 if (cl->avgidle < 0)
 415                         delay -= (-cl->avgidle) - ((-cl->avgidle) >> cl->ewma_log);
 416                 if (cl->avgidle < cl->minidle)
 417                         cl->avgidle = cl->minidle;
 418                 if (delay <= 0)
 419                         delay = 1;
 420                 cl->undertime = q->now + delay;
 421
 422                 cl->xstats.overactions++;
 423                 cl->delayed = 1;
 424         }
 425         if (q->wd_expires == 0 || q->wd_expires > delay)
 426                 q->wd_expires = delay;
 427
 428         /* Dirty work! We must schedule wakeups based on
 429          * real available rate, rather than leaf rate,
 430          * which may be tiny (even zero).
 431          */
 432         if (q->toplevel == TC_CBQ_MAXLEVEL) {
 433                 struct cbq_class *b;
 434                 psched_tdiff_t base_delay = q->wd_expires;
 435
 436                 for (b = cl->borrow; b; b = b->borrow) {
 437                         delay = b->undertime - q->now;
 438                         if (delay < base_delay) {
 439                                 if (delay <= 0)
 440                                         delay = 1;
 441                                 base_delay = delay;
 442                         }
 443                 }
 444
 445                 q->wd_expires = base_delay;
 446         }
 447 }
 448
 449 static psched_tdiff_t cbq_undelay_prio(struct cbq_sched_data *q, int prio,
 450                                        psched_time_t now)
 451 {
 452         struct cbq_class *cl;
 453         struct cbq_class *cl_prev = q->active[prio];
 454         psched_time_t sched = now;
 455
 456         if (cl_prev == NULL)
 457                 return 0;
 458
 459         do {
 460                 cl = cl_prev->next_alive;
 461                 if (now - cl->penalized > 0) {
 462                         cl_prev->next_alive = cl->next_alive;
 463                         cl->next_alive = NULL;
 464                         cl->cpriority = cl->priority;
 465                         cl->delayed = 0;
 466                         cbq_activate_class(cl);
 467
 468                         if (cl == q->active[prio]) {
 469                                 q->active[prio] = cl_prev;
 470                                 if (cl == q->active[prio]) {
 471                                         q->active[prio] = NULL;
 472                                         return 0;
 473                                 }
 474                         }
 475
 476                         cl = cl_prev->next_alive;
 477                 } else if (sched - cl->penalized > 0)
 478                         sched = cl->penalized;
 479         } while ((cl_prev = cl) != q->active[prio]);
 480
 481         return sched - now;
 482 }
 483
 484 static enum hrtimer_restart cbq_undelay(struct hrtimer *timer)
 485 {
 486         struct cbq_sched_data *q = container_of(timer, struct cbq_sched_data,
 487                                                 delay_timer);
 488         struct Qdisc *sch = q->watchdog.qdisc;
 489         psched_time_t now;
 490         psched_tdiff_t delay = 0;
 491         unsigned int pmask;
 492
 493         now = psched_get_time();
 494
 495         pmask = q->pmask;
 496         q->pmask = 0;
 497
 498         while (pmask) {
 499                 int prio = ffz(~pmask);
 500                 psched_tdiff_t tmp;
 501
 502                 pmask &= ~(1<<prio);
 503
 504                 tmp = cbq_undelay_prio(q, prio, now);
 505                 if (tmp > 0) {
 506                         q->pmask |= 1<<prio;
 507                         if (tmp < delay || delay == 0)
 508                                 delay = tmp;
 509                 }
 510         }
 511
 512         if (delay) {
 513                 ktime_t time;
 514
 515                 time = 0;
 516                 time = ktime_add_ns(time, PSCHED_TICKS2NS(now + delay));
 517                 hrtimer_start(&q->delay_timer, time, HRTIMER_MODE_ABS_PINNED);
 518         }
 519
 520         __netif_schedule(qdisc_root(sch));
 521         return HRTIMER_NORESTART;
 522 }
 523
 524 /*
 525  * It is mission critical procedure.
 526  *
 527  * We "regenerate" toplevel cutoff, if transmitting class
 528  * has backlog and it is not regulated. It is not part of
 529  * original CBQ description, but looks more reasonable.
 530  * Probably, it is wrong. This question needs further investigation.
 531  */
 532
 533 static inline void
 534 cbq_update_toplevel(struct cbq_sched_data *q, struct cbq_class *cl,
 535                     struct cbq_class *borrowed)
 536 {
 537         if (cl && q->toplevel >= borrowed->level) {
 538                 if (cl->q->q.qlen > 1) {
 539                         do {
 540                                 if (borrowed->undertime == PSCHED_PASTPERFECT) {
 541                                         q->toplevel = borrowed->level;
 542                                         return;
 543                                 }
 544                         } while ((borrowed = borrowed->borrow) != NULL);
 545                 }
 546 #if 0
 547         /* It is not necessary now. Uncommenting it
 548            will save CPU cycles, but decrease fairness.
 549          */
 550                 q->toplevel = TC_CBQ_MAXLEVEL;
 551 #endif
 552         }
 553 }
 554
 555 static void
 556 cbq_update(struct cbq_sched_data *q)
 557 {
 558         struct cbq_class *this = q->tx_class;
 559         struct cbq_class *cl = this;
 560         int len = q->tx_len;
 561         psched_time_t now;
 562
 563         q->tx_class = NULL;
 564         /* Time integrator. We calculate EOS time
 565          * by adding expected packet transmission time.
 566          */
 567         now = q->now + L2T(&q->link, len);
 568
 569         for ( ; cl; cl = cl->share) {
 570                 long avgidle = cl->avgidle;
 571                 long idle;
 572
 573                 cl->bstats.packets++;
 574                 cl->bstats.bytes += len;
 575
 576                 /*
 577                  * (now - last) is total time between packet right edges.
 578                  * (last_pktlen/rate) is "virtual" busy time, so that
 579                  *
 580                  *      idle = (now - last) - last_pktlen/rate
 581                  */
 582
 583                 idle = now - cl->last;
 584                 if ((unsigned long)idle > 128*1024*1024) {
 585                         avgidle = cl->maxidle;
 586                 } else {
 587                         idle -= L2T(cl, len);
 588
 589                 /* true_avgidle := (1-W)*true_avgidle + W*idle,
 590                  * where W=2^{-ewma_log}. But cl->avgidle is scaled:
 591                  * cl->avgidle == true_avgidle/W,
 592                  * hence:
 593                  */
 594                         avgidle += idle - (avgidle>>cl->ewma_log);
 595                 }
 596
 597                 if (avgidle <= 0) {
 598                         /* Overlimit or at-limit */
 599
 600                         if (avgidle < cl->minidle)
 601                                 avgidle = cl->minidle;
 602
 603                         cl->avgidle = avgidle;
 604
 605                         /* Calculate expected time, when this class
 606                          * will be allowed to send.
 607                          * It will occur, when:
 608                          * (1-W)*true_avgidle + W*delay = 0, i.e.
 609                          * idle = (1/W - 1)*(-true_avgidle)
 610                          * or
 611                          * idle = (1 - W)*(-cl->avgidle);
 612                          */
 613                         idle = (-avgidle) - ((-avgidle) >> cl->ewma_log);
 614
 615                         /*
 616                          * That is not all.
 617                          * To maintain the rate allocated to the class,
 618                          * we add to undertime virtual clock,
 619                          * necessary to complete transmitted packet.
 620                          * (len/phys_bandwidth has been already passed
 621                          * to the moment of cbq_update)
 622                          */
 623
 624                         idle -= L2T(&q->link, len);
 625                         idle += L2T(cl, len);
 626
 627                         cl->undertime = now + idle;
 628                 } else {
 629                         /* Underlimit */
 630
 631                         cl->undertime = PSCHED_PASTPERFECT;
 632                         if (avgidle > cl->maxidle)
 633                                 cl->avgidle = cl->maxidle;
 634                         else
 635                                 cl->avgidle = avgidle;
 636                 }
 637                 if ((s64)(now - cl->last) > 0)
 638                         cl->last = now;
 639         }
 640
 641         cbq_update_toplevel(q, this, q->tx_borrowed);
 642 }
 643
 644 static inline struct cbq_class *
 645 cbq_under_limit(struct cbq_class *cl)
 646 {
 647         struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 648         struct cbq_class *this_cl = cl;
 649
 650         if (cl->tparent == NULL)
 651                 return cl;
 652
 653         if (cl->undertime == PSCHED_PASTPERFECT || q->now >= cl->undertime) {
 654                 cl->delayed = 0;
 655                 return cl;
 656         }
 657
 658         do {
 659                 /* It is very suspicious place. Now overlimit
 660                  * action is generated for not bounded classes
 661                  * only if link is completely congested.
 662                  * Though it is in agree with ancestor-only paradigm,
 663                  * it looks very stupid. Particularly,
 664                  * it means that this chunk of code will either
 665                  * never be called or result in strong amplification
 666                  * of burstiness. Dangerous, silly, and, however,
 667                  * no another solution exists.
 668                  */
 669                 cl = cl->borrow;
 670                 if (!cl) {
 671                         this_cl->qstats.overlimits++;
 672                         cbq_overlimit(this_cl);
 673                         return NULL;
 674                 }
 675                 if (cl->level > q->toplevel)
 676                         return NULL;
 677         } while (cl->undertime != PSCHED_PASTPERFECT && q->now < cl->undertime);
 678
 679         cl->delayed = 0;
 680         return cl;
 681 }
 682
 683 static inline struct sk_buff *
 684 cbq_dequeue_prio(struct Qdisc *sch, int prio)
 685 {
 686         struct cbq_sched_data *q = qdisc_priv(sch);
 687         struct cbq_class *cl_tail, *cl_prev, *cl;
 688         struct sk_buff *skb;
 689         int deficit;
 690
 691         cl_tail = cl_prev = q->active[prio];
 692         cl = cl_prev->next_alive;
 693
 694         do {
 695                 deficit = 0;
 696
 697                 /* Start round */
 698                 do {
 699                         struct cbq_class *borrow = cl;
 700
 701                         if (cl->q->q.qlen &&
 702                             (borrow = cbq_under_limit(cl)) == NULL)
 703                                 goto skip_class;
 704
 705                         if (cl->deficit <= 0) {
 706                                 /* Class exhausted its allotment per
 707                                  * this round. Switch to the next one.
 708                                  */
 709                                 deficit = 1;
 710                                 cl->deficit += cl->quantum;
 711                                 goto next_class;
 712                         }
 713
 714                         skb = cl->q->dequeue(cl->q);
 715
 716                         /* Class did not give us any skb :-(
 717                          * It could occur even if cl->q->q.qlen != 0
 718                          * f.e. if cl->q == "tbf"
 719                          */
 720                         if (skb == NULL)
 721                                 goto skip_class;
 722
 723                         cl->deficit -= qdisc_pkt_len(skb);
 724                         q->tx_class = cl;
 725                         q->tx_borrowed = borrow;
 726                         if (borrow != cl) {
 727 #ifndef CBQ_XSTATS_BORROWS_BYTES
 728                                 borrow->xstats.borrows++;
 729                                 cl->xstats.borrows++;
 730 #else
 731                                 borrow->xstats.borrows += qdisc_pkt_len(skb);
 732                                 cl->xstats.borrows += qdisc_pkt_len(skb);
 733 #endif
 734                         }
 735                         q->tx_len = qdisc_pkt_len(skb);
 736
 737                         if (cl->deficit <= 0) {
 738                                 q->active[prio] = cl;
 739                                 cl = cl->next_alive;
 740                                 cl->deficit += cl->quantum;
 741                         }
 742                         return skb;
 743
 744 skip_class:
 745                         if (cl->q->q.qlen == 0 || prio != cl->cpriority) {
 746                                 /* Class is empty or penalized.
 747                                  * Unlink it from active chain.
 748                                  */
 749                                 cl_prev->next_alive = cl->next_alive;
 750                                 cl->next_alive = NULL;
 751
 752                                 /* Did cl_tail point to it? */
 753                                 if (cl == cl_tail) {
 754                                         /* Repair it! */
 755                                         cl_tail = cl_prev;
 756
 757                                         /* Was it the last class in this band? */
 758                                         if (cl == cl_tail) {
 759                                                 /* Kill the band! */
 760                                                 q->active[prio] = NULL;
 761                                                 q->activemask &= ~(1<<prio);
 762                                                 if (cl->q->q.qlen)
 763                                                         cbq_activate_class(cl);
 764                                                 return NULL;
 765                                         }
 766
 767                                         q->active[prio] = cl_tail;
 768                                 }
 769                                 if (cl->q->q.qlen)
 770                                         cbq_activate_class(cl);
 771
 772                                 cl = cl_prev;
 773                         }
 774
 775 next_class:
 776                         cl_prev = cl;
 777                         cl = cl->next_alive;
 778                 } while (cl_prev != cl_tail);
 779         } while (deficit);
 780
 781         q->active[prio] = cl_prev;
 782
 783         return NULL;
 784 }
 785
 786 static inline struct sk_buff *
 787 cbq_dequeue_1(struct Qdisc *sch)
 788 {
 789         struct cbq_sched_data *q = qdisc_priv(sch);
 790         struct sk_buff *skb;
 791         unsigned int activemask;
 792
 793         activemask = q->activemask & 0xFF;
 794         while (activemask) {
 795                 int prio = ffz(~activemask);
 796                 activemask &= ~(1<<prio);
 797                 skb = cbq_dequeue_prio(sch, prio);
 798                 if (skb)
 799                         return skb;
 800         }
 801         return NULL;
 802 }
 803
 804 static struct sk_buff *
 805 cbq_dequeue(struct Qdisc *sch)
 806 {
 807         struct sk_buff *skb;
 808         struct cbq_sched_data *q = qdisc_priv(sch);
 809         psched_time_t now;
 810
 811         now = psched_get_time();
 812
 813         if (q->tx_class)
 814                 cbq_update(q);
 815
 816         q->now = now;
 817
 818         for (;;) {
 819                 q->wd_expires = 0;
 820
 821                 skb = cbq_dequeue_1(sch);
 822                 if (skb) {
 823                         qdisc_bstats_update(sch, skb);
 824                         sch->q.qlen--;
 825                         return skb;
 826                 }
 827
 828                 /* All the classes are overlimit.
 829                  *
 830                  * It is possible, if:
 831                  *
 832                  * 1. Scheduler is empty.
 833                  * 2. Toplevel cutoff inhibited borrowing.
 834                  * 3. Root class is overlimit.
 835                  *
 836                  * Reset 2d and 3d conditions and retry.
 837                  *
 838                  * Note, that NS and cbq-2.0 are buggy, peeking
 839                  * an arbitrary class is appropriate for ancestor-only
 840                  * sharing, but not for toplevel algorithm.
 841                  *
 842                  * Our version is better, but slower, because it requires
 843                  * two passes, but it is unavoidable with top-level sharing.
 844                  */
 845
 846                 if (q->toplevel == TC_CBQ_MAXLEVEL &&
 847                     q->link.undertime == PSCHED_PASTPERFECT)
 848                         break;
 849
 850                 q->toplevel = TC_CBQ_MAXLEVEL;
 851                 q->link.undertime = PSCHED_PASTPERFECT;
 852         }
 853
 854         /* No packets in scheduler or nobody wants to give them to us :-(
 855          * Sigh... start watchdog timer in the last case.
 856          */
 857
 858         if (sch->q.qlen) {
 859                 qdisc_qstats_overlimit(sch);
 860                 if (q->wd_expires)
 861                         qdisc_watchdog_schedule(&q->watchdog,
 862                                                 now + q->wd_expires);
 863         }
 864         return NULL;
 865 }
 866
 867 /* CBQ class maintanance routines */
 868
 869 static void cbq_adjust_levels(struct cbq_class *this)
 870 {
 871         if (this == NULL)
 872                 return;
 873
 874         do {
 875                 int level = 0;
 876                 struct cbq_class *cl;
 877
 878                 cl = this->children;
 879                 if (cl) {
 880                         do {
 881                                 if (cl->level > level)
 882                                         level = cl->level;
 883                         } while ((cl = cl->sibling) != this->children);
 884                 }
 885                 this->level = level + 1;
 886         } while ((this = this->tparent) != NULL);
 887 }
 888
 889 static void cbq_normalize_quanta(struct cbq_sched_data *q, int prio)
 890 {
 891         struct cbq_class *cl;
 892         unsigned int h;
 893
 894         if (q->quanta[prio] == 0)
 895                 return;
 896
 897         for (h = 0; h < q->clhash.hashsize; h++) {
 898                 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
 899                         /* BUGGGG... Beware! This expression suffer of
 900                          * arithmetic overflows!
 901                          */
 902                         if (cl->priority == prio) {
 903                                 cl->quantum = (cl->weight*cl->allot*q->nclasses[prio])/
 904                                         q->quanta[prio];
 905                         }
 906                         if (cl->quantum <= 0 ||
 907                             cl->quantum > 32*qdisc_dev(cl->qdisc)->mtu) {
 908                                 pr_warn("CBQ: class %08x has bad quantum==%ld, repaired.\n",
 909                                         cl->common.classid, cl->quantum);
 910                                 cl->quantum = qdisc_dev(cl->qdisc)->mtu/2 + 1;
 911                         }
 912                 }
 913         }
 914 }
 915
 916 static void cbq_sync_defmap(struct cbq_class *cl)
 917 {
 918         struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
 919         struct cbq_class *split = cl->split;
 920         unsigned int h;
 921         int i;
 922
 923         if (split == NULL)
 924                 return;
 925
 926         for (i = 0; i <= TC_PRIO_MAX; i++) {
 927                 if (split->defaults[i] == cl && !(cl->defmap & (1<<i)))
 928                         split->defaults[i] = NULL;
 929         }
 930
 931         for (i = 0; i <= TC_PRIO_MAX; i++) {
 932                 int level = split->level;
 933
 934                 if (split->defaults[i])
 935                         continue;
 936
 937                 for (h = 0; h < q->clhash.hashsize; h++) {
 938                         struct cbq_class *c;
 939
 940                         hlist_for_each_entry(c, &q->clhash.hash[h],
 941                                              common.hnode) {
 942                                 if (c->split == split && c->level < level &&
 943                                     c->defmap & (1<<i)) {
 944                                         split->defaults[i] = c;
 945                                         level = c->level;
 946                                 }
 947                         }
 948                 }
 949         }
 950 }
 951
 952 static void cbq_change_defmap(struct cbq_class *cl, u32 splitid, u32 def, u32 mask)
 953 {
 954         struct cbq_class *split = NULL;
 955
 956         if (splitid == 0) {
 957                 split = cl->split;
 958                 if (!split)
 959                         return;
 960                 splitid = split->common.classid;
 961         }
 962
 963         if (split == NULL || split->common.classid != splitid) {
 964                 for (split = cl->tparent; split; split = split->tparent)
 965                         if (split->common.classid == splitid)
 966                                 break;
 967         }
 968
 969         if (split == NULL)
 970                 return;
 971
 972         if (cl->split != split) {
 973                 cl->defmap = 0;
 974                 cbq_sync_defmap(cl);
 975                 cl->split = split;
 976                 cl->defmap = def & mask;
 977         } else
 978                 cl->defmap = (cl->defmap & ~mask) | (def & mask);
 979
 980         cbq_sync_defmap(cl);
 981 }
 982
 983 static void cbq_unlink_class(struct cbq_class *this)
 984 {
 985         struct cbq_class *cl, **clp;
 986         struct cbq_sched_data *q = qdisc_priv(this->qdisc);
 987
 988         qdisc_class_hash_remove(&q->clhash, &this->common);
 989
 990         if (this->tparent) {
 991                 clp = &this->sibling;
 992                 cl = *clp;
 993                 do {
 994                         if (cl == this) {
 995                                 *clp = cl->sibling;
 996                                 break;
 997                         }
 998                         clp = &cl->sibling;
 999                 } while ((cl = *clp) != this->sibling);
1000
1001                 if (this->tparent->children == this) {
1002                         this->tparent->children = this->sibling;
1003                         if (this->sibling == this)
1004                                 this->tparent->children = NULL;
1005                 }
1006         } else {
1007                 WARN_ON(this->sibling != this);
1008         }
1009 }
1010
1011 static void cbq_link_class(struct cbq_class *this)
1012 {
1013         struct cbq_sched_data *q = qdisc_priv(this->qdisc);
1014         struct cbq_class *parent = this->tparent;
1015
1016         this->sibling = this;
1017         qdisc_class_hash_insert(&q->clhash, &this->common);
1018
1019         if (parent == NULL)
1020                 return;
1021
1022         if (parent->children == NULL) {
1023                 parent->children = this;
1024         } else {
1025                 this->sibling = parent->children->sibling;
1026                 parent->children->sibling = this;
1027         }
1028 }
1029
1030 static void
1031 cbq_reset(struct Qdisc *sch)
1032 {
1033         struct cbq_sched_data *q = qdisc_priv(sch);
1034         struct cbq_class *cl;
1035         int prio;
1036         unsigned int h;
1037
1038         q->activemask = 0;
1039         q->pmask = 0;
1040         q->tx_class = NULL;
1041         q->tx_borrowed = NULL;
1042         qdisc_watchdog_cancel(&q->watchdog);
1043         hrtimer_cancel(&q->delay_timer);
1044         q->toplevel = TC_CBQ_MAXLEVEL;
1045         q->now = psched_get_time();
1046
1047         for (prio = 0; prio <= TC_CBQ_MAXPRIO; prio++)
1048                 q->active[prio] = NULL;
1049
1050         for (h = 0; h < q->clhash.hashsize; h++) {
1051                 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1052                         qdisc_reset(cl->q);
1053
1054                         cl->next_alive = NULL;
1055                         cl->undertime = PSCHED_PASTPERFECT;
1056                         cl->avgidle = cl->maxidle;
1057                         cl->deficit = cl->quantum;
1058                         cl->cpriority = cl->priority;
1059                 }
1060         }
1061         sch->q.qlen = 0;
1062 }
1063
1064
1065 static int cbq_set_lss(struct cbq_class *cl, struct tc_cbq_lssopt *lss)
1066 {
1067         if (lss->change & TCF_CBQ_LSS_FLAGS) {
1068                 cl->share = (lss->flags & TCF_CBQ_LSS_ISOLATED) ? NULL : cl->tparent;
1069                 cl->borrow = (lss->flags & TCF_CBQ_LSS_BOUNDED) ? NULL : cl->tparent;
1070         }
1071         if (lss->change & TCF_CBQ_LSS_EWMA)
1072                 cl->ewma_log = lss->ewma_log;
1073         if (lss->change & TCF_CBQ_LSS_AVPKT)
1074                 cl->avpkt = lss->avpkt;
1075         if (lss->change & TCF_CBQ_LSS_MINIDLE)
1076                 cl->minidle = -(long)lss->minidle;
1077         if (lss->change & TCF_CBQ_LSS_MAXIDLE) {
1078                 cl->maxidle = lss->maxidle;
1079                 cl->avgidle = lss->maxidle;
1080         }
1081         if (lss->change & TCF_CBQ_LSS_OFFTIME)
1082                 cl->offtime = lss->offtime;
1083         return 0;
1084 }
1085
1086 static void cbq_rmprio(struct cbq_sched_data *q, struct cbq_class *cl)
1087 {
1088         q->nclasses[cl->priority]--;
1089         q->quanta[cl->priority] -= cl->weight;
1090         cbq_normalize_quanta(q, cl->priority);
1091 }
1092
1093 static void cbq_addprio(struct cbq_sched_data *q, struct cbq_class *cl)
1094 {
1095         q->nclasses[cl->priority]++;
1096         q->quanta[cl->priority] += cl->weight;
1097         cbq_normalize_quanta(q, cl->priority);
1098 }
1099
1100 static int cbq_set_wrr(struct cbq_class *cl, struct tc_cbq_wrropt *wrr)
1101 {
1102         struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
1103
1104         if (wrr->allot)
1105                 cl->allot = wrr->allot;
1106         if (wrr->weight)
1107                 cl->weight = wrr->weight;
1108         if (wrr->priority) {
1109                 cl->priority = wrr->priority - 1;
1110                 cl->cpriority = cl->priority;
1111                 if (cl->priority >= cl->priority2)
1112                         cl->priority2 = TC_CBQ_MAXPRIO - 1;
1113         }
1114
1115         cbq_addprio(q, cl);
1116         return 0;
1117 }
1118
1119 static int cbq_set_fopt(struct cbq_class *cl, struct tc_cbq_fopt *fopt)
1120 {
1121         cbq_change_defmap(cl, fopt->split, fopt->defmap, fopt->defchange);
1122         return 0;
1123 }
1124
1125 static const struct nla_policy cbq_policy[TCA_CBQ_MAX + 1] = {
1126         [TCA_CBQ_LSSOPT]        = { .len = sizeof(struct tc_cbq_lssopt) },
1127         [TCA_CBQ_WRROPT]        = { .len = sizeof(struct tc_cbq_wrropt) },
1128         [TCA_CBQ_FOPT]          = { .len = sizeof(struct tc_cbq_fopt) },
1129         [TCA_CBQ_OVL_STRATEGY]  = { .len = sizeof(struct tc_cbq_ovl) },
1130         [TCA_CBQ_RATE]          = { .len = sizeof(struct tc_ratespec) },
1131         [TCA_CBQ_RTAB]          = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1132         [TCA_CBQ_POLICE]        = { .len = sizeof(struct tc_cbq_police) },
1133 };
1134
1135 static int cbq_init(struct Qdisc *sch, struct nlattr *opt)
1136 {
1137         struct cbq_sched_data *q = qdisc_priv(sch);
1138         struct nlattr *tb[TCA_CBQ_MAX + 1];
1139         struct tc_ratespec *r;
1140         int err;
1141
1142         qdisc_watchdog_init(&q->watchdog, sch);
1143         hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1144         q->delay_timer.function = cbq_undelay;
1145
1146         if (!opt)
1147                 return -EINVAL;
1148
1149         err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1150         if (err < 0)
1151                 return err;
1152
1153         if (tb[TCA_CBQ_RTAB] == NULL || tb[TCA_CBQ_RATE] == NULL)
1154                 return -EINVAL;
1155
1156         r = nla_data(tb[TCA_CBQ_RATE]);
1157
1158         if ((q->link.R_tab = qdisc_get_rtab(r, tb[TCA_CBQ_RTAB])) == NULL)
1159                 return -EINVAL;
1160
1161         err = qdisc_class_hash_init(&q->clhash);
1162         if (err < 0)
1163                 goto put_rtab;
1164
1165         q->link.refcnt = 1;
1166         q->link.sibling = &q->link;
1167         q->link.common.classid = sch->handle;
1168         q->link.qdisc = sch;
1169         q->link.q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1170                                       sch->handle);
1171         if (!q->link.q)
1172                 q->link.q = &noop_qdisc;
1173         else
1174                 qdisc_hash_add(q->link.q, true);
1175
1176         q->link.priority = TC_CBQ_MAXPRIO - 1;
1177         q->link.priority2 = TC_CBQ_MAXPRIO - 1;
1178         q->link.cpriority = TC_CBQ_MAXPRIO - 1;
1179         q->link.allot = psched_mtu(qdisc_dev(sch));
1180         q->link.quantum = q->link.allot;
1181         q->link.weight = q->link.R_tab->rate.rate;
1182
1183         q->link.ewma_log = TC_CBQ_DEF_EWMA;
1184         q->link.avpkt = q->link.allot/2;
1185         q->link.minidle = -0x7FFFFFFF;
1186
1187         q->toplevel = TC_CBQ_MAXLEVEL;
1188         q->now = psched_get_time();
1189
1190         cbq_link_class(&q->link);
1191
1192         if (tb[TCA_CBQ_LSSOPT])
1193                 cbq_set_lss(&q->link, nla_data(tb[TCA_CBQ_LSSOPT]));
1194
1195         cbq_addprio(q, &q->link);
1196         return 0;
1197
1198 put_rtab:
1199         qdisc_put_rtab(q->link.R_tab);
1200         return err;
1201 }
1202
1203 static int cbq_dump_rate(struct sk_buff *skb, struct cbq_class *cl)
1204 {
1205         unsigned char *b = skb_tail_pointer(skb);
1206
1207         if (nla_put(skb, TCA_CBQ_RATE, sizeof(cl->R_tab->rate), &cl->R_tab->rate))
1208                 goto nla_put_failure;
1209         return skb->len;
1210
1211 nla_put_failure:
1212         nlmsg_trim(skb, b);
1213         return -1;
1214 }
1215
1216 static int cbq_dump_lss(struct sk_buff *skb, struct cbq_class *cl)
1217 {
1218         unsigned char *b = skb_tail_pointer(skb);
1219         struct tc_cbq_lssopt opt;
1220
1221         opt.flags = 0;
1222         if (cl->borrow == NULL)
1223                 opt.flags |= TCF_CBQ_LSS_BOUNDED;
1224         if (cl->share == NULL)
1225                 opt.flags |= TCF_CBQ_LSS_ISOLATED;
1226         opt.ewma_log = cl->ewma_log;
1227         opt.level = cl->level;
1228         opt.avpkt = cl->avpkt;
1229         opt.maxidle = cl->maxidle;
1230         opt.minidle = (u32)(-cl->minidle);
1231         opt.offtime = cl->offtime;
1232         opt.change = ~0;
1233         if (nla_put(skb, TCA_CBQ_LSSOPT, sizeof(opt), &opt))
1234                 goto nla_put_failure;
1235         return skb->len;
1236
1237 nla_put_failure:
1238         nlmsg_trim(skb, b);
1239         return -1;
1240 }
1241
1242 static int cbq_dump_wrr(struct sk_buff *skb, struct cbq_class *cl)
1243 {
1244         unsigned char *b = skb_tail_pointer(skb);
1245         struct tc_cbq_wrropt opt;
1246
1247         memset(&opt, 0, sizeof(opt));
1248         opt.flags = 0;
1249         opt.allot = cl->allot;
1250         opt.priority = cl->priority + 1;
1251         opt.cpriority = cl->cpriority + 1;
1252         opt.weight = cl->weight;
1253         if (nla_put(skb, TCA_CBQ_WRROPT, sizeof(opt), &opt))
1254                 goto nla_put_failure;
1255         return skb->len;
1256
1257 nla_put_failure:
1258         nlmsg_trim(skb, b);
1259         return -1;
1260 }
1261
1262 static int cbq_dump_fopt(struct sk_buff *skb, struct cbq_class *cl)
1263 {
1264         unsigned char *b = skb_tail_pointer(skb);
1265         struct tc_cbq_fopt opt;
1266
1267         if (cl->split || cl->defmap) {
1268                 opt.split = cl->split ? cl->split->common.classid : 0;
1269                 opt.defmap = cl->defmap;
1270                 opt.defchange = ~0;
1271                 if (nla_put(skb, TCA_CBQ_FOPT, sizeof(opt), &opt))
1272                         goto nla_put_failure;
1273         }
1274         return skb->len;
1275
1276 nla_put_failure:
1277         nlmsg_trim(skb, b);
1278         return -1;
1279 }
1280
1281 static int cbq_dump_attr(struct sk_buff *skb, struct cbq_class *cl)
1282 {
1283         if (cbq_dump_lss(skb, cl) < 0 ||
1284             cbq_dump_rate(skb, cl) < 0 ||
1285             cbq_dump_wrr(skb, cl) < 0 ||
1286             cbq_dump_fopt(skb, cl) < 0)
1287                 return -1;
1288         return 0;
1289 }
1290
1291 static int cbq_dump(struct Qdisc *sch, struct sk_buff *skb)
1292 {
1293         struct cbq_sched_data *q = qdisc_priv(sch);
1294         struct nlattr *nest;
1295
1296         nest = nla_nest_start(skb, TCA_OPTIONS);
1297         if (nest == NULL)
1298                 goto nla_put_failure;
1299         if (cbq_dump_attr(skb, &q->link) < 0)
1300                 goto nla_put_failure;
1301         return nla_nest_end(skb, nest);
1302
1303 nla_put_failure:
1304         nla_nest_cancel(skb, nest);
1305         return -1;
1306 }
1307
1308 static int
1309 cbq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1310 {
1311         struct cbq_sched_data *q = qdisc_priv(sch);
1312
1313         q->link.xstats.avgidle = q->link.avgidle;
1314         return gnet_stats_copy_app(d, &q->link.xstats, sizeof(q->link.xstats));
1315 }
1316
1317 static int
1318 cbq_dump_class(struct Qdisc *sch, unsigned long arg,
1319                struct sk_buff *skb, struct tcmsg *tcm)
1320 {
1321         struct cbq_class *cl = (struct cbq_class *)arg;
1322         struct nlattr *nest;
1323
1324         if (cl->tparent)
1325                 tcm->tcm_parent = cl->tparent->common.classid;
1326         else
1327                 tcm->tcm_parent = TC_H_ROOT;
1328         tcm->tcm_handle = cl->common.classid;
1329         tcm->tcm_info = cl->q->handle;
1330
1331         nest = nla_nest_start(skb, TCA_OPTIONS);
1332         if (nest == NULL)
1333                 goto nla_put_failure;
1334         if (cbq_dump_attr(skb, cl) < 0)
1335                 goto nla_put_failure;
1336         return nla_nest_end(skb, nest);
1337
1338 nla_put_failure:
1339         nla_nest_cancel(skb, nest);
1340         return -1;
1341 }
1342
1343 static int
1344 cbq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1345         struct gnet_dump *d)
1346 {
1347         struct cbq_sched_data *q = qdisc_priv(sch);
1348         struct cbq_class *cl = (struct cbq_class *)arg;
1349
1350         cl->xstats.avgidle = cl->avgidle;
1351         cl->xstats.undertime = 0;
1352
1353         if (cl->undertime != PSCHED_PASTPERFECT)
1354                 cl->xstats.undertime = cl->undertime - q->now;
1355
1356         if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1357                                   d, NULL, &cl->bstats) < 0 ||
1358             gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1359             gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->q->q.qlen) < 0)
1360                 return -1;
1361
1362         return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1363 }
1364
1365 static int cbq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1366                      struct Qdisc **old)
1367 {
1368         struct cbq_class *cl = (struct cbq_class *)arg;
1369
1370         if (new == NULL) {
1371                 new = qdisc_create_dflt(sch->dev_queue,
1372                                         &pfifo_qdisc_ops, cl->common.classid);
1373                 if (new == NULL)
1374                         return -ENOBUFS;
1375         }
1376
1377         *old = qdisc_replace(sch, new, &cl->q);
1378         return 0;
1379 }
1380
1381 static struct Qdisc *cbq_leaf(struct Qdisc *sch, unsigned long arg)
1382 {
1383         struct cbq_class *cl = (struct cbq_class *)arg;
1384
1385         return cl->q;
1386 }
1387
1388 static void cbq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1389 {
1390         struct cbq_class *cl = (struct cbq_class *)arg;
1391
1392         if (cl->q->q.qlen == 0)
1393                 cbq_deactivate_class(cl);
1394 }
1395
1396 static unsigned long cbq_get(struct Qdisc *sch, u32 classid)
1397 {
1398         struct cbq_sched_data *q = qdisc_priv(sch);
1399         struct cbq_class *cl = cbq_class_lookup(q, classid);
1400
1401         if (cl) {
1402                 cl->refcnt++;
1403                 return (unsigned long)cl;
1404         }
1405         return 0;
1406 }
1407
1408 static void cbq_destroy_class(struct Qdisc *sch, struct cbq_class *cl)
1409 {
1410         struct cbq_sched_data *q = qdisc_priv(sch);
1411
1412         WARN_ON(cl->filters);
1413
1414         tcf_block_put(cl->block);
1415         qdisc_destroy(cl->q);
1416         qdisc_put_rtab(cl->R_tab);
1417         gen_kill_estimator(&cl->rate_est);
1418         if (cl != &q->link)
1419                 kfree(cl);
1420 }
1421
1422 static void cbq_destroy(struct Qdisc *sch)
1423 {
1424         struct cbq_sched_data *q = qdisc_priv(sch);
1425         struct hlist_node *next;
1426         struct cbq_class *cl;
1427         unsigned int h;
1428
1429 #ifdef CONFIG_NET_CLS_ACT
1430         q->rx_class = NULL;
1431 #endif
1432         /*
1433          * Filters must be destroyed first because we don't destroy the
1434          * classes from root to leafs which means that filters can still
1435          * be bound to classes which have been destroyed already. --TGR '04
1436          */
1437         for (h = 0; h < q->clhash.hashsize; h++) {
1438                 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1439                         tcf_block_put(cl->block);
1440                         cl->block = NULL;
1441                 }
1442         }
1443         for (h = 0; h < q->clhash.hashsize; h++) {
1444                 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[h],
1445                                           common.hnode)
1446                         cbq_destroy_class(sch, cl);
1447         }
1448         qdisc_class_hash_destroy(&q->clhash);
1449 }
1450
1451 static void cbq_put(struct Qdisc *sch, unsigned long arg)
1452 {
1453         struct cbq_class *cl = (struct cbq_class *)arg;
1454
1455         if (--cl->refcnt == 0) {
1456 #ifdef CONFIG_NET_CLS_ACT
1457                 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1458                 struct cbq_sched_data *q = qdisc_priv(sch);
1459
1460                 spin_lock_bh(root_lock);
1461                 if (q->rx_class == cl)
1462                         q->rx_class = NULL;
1463                 spin_unlock_bh(root_lock);
1464 #endif
1465
1466                 cbq_destroy_class(sch, cl);
1467         }
1468 }
1469
1470 static int
1471 cbq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, struct nlattr **tca,
1472                  unsigned long *arg)
1473 {
1474         int err;
1475         struct cbq_sched_data *q = qdisc_priv(sch);
1476         struct cbq_class *cl = (struct cbq_class *)*arg;
1477         struct nlattr *opt = tca[TCA_OPTIONS];
1478         struct nlattr *tb[TCA_CBQ_MAX + 1];
1479         struct cbq_class *parent;
1480         struct qdisc_rate_table *rtab = NULL;
1481
1482         if (opt == NULL)
1483                 return -EINVAL;
1484
1485         err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1486         if (err < 0)
1487                 return err;
1488
1489         if (tb[TCA_CBQ_OVL_STRATEGY] || tb[TCA_CBQ_POLICE])
1490                 return -EOPNOTSUPP;
1491
1492         if (cl) {
1493                 /* Check parent */
1494                 if (parentid) {
1495                         if (cl->tparent &&
1496                             cl->tparent->common.classid != parentid)
1497                                 return -EINVAL;
1498                         if (!cl->tparent && parentid != TC_H_ROOT)
1499                                 return -EINVAL;
1500                 }
1501
1502                 if (tb[TCA_CBQ_RATE]) {
1503                         rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]),
1504                                               tb[TCA_CBQ_RTAB]);
1505                         if (rtab == NULL)
1506                                 return -EINVAL;
1507                 }
1508
1509                 if (tca[TCA_RATE]) {
1510                         err = gen_replace_estimator(&cl->bstats, NULL,
1511                                                     &cl->rate_est,
1512                                                     NULL,
1513                                                     qdisc_root_sleeping_running(sch),
1514                                                     tca[TCA_RATE]);
1515                         if (err) {
1516                                 qdisc_put_rtab(rtab);
1517                                 return err;
1518                         }
1519                 }
1520
1521                 /* Change class parameters */
1522                 sch_tree_lock(sch);
1523
1524                 if (cl->next_alive != NULL)
1525                         cbq_deactivate_class(cl);
1526
1527                 if (rtab) {
1528                         qdisc_put_rtab(cl->R_tab);
1529                         cl->R_tab = rtab;
1530                 }
1531
1532                 if (tb[TCA_CBQ_LSSOPT])
1533                         cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1534
1535                 if (tb[TCA_CBQ_WRROPT]) {
1536                         cbq_rmprio(q, cl);
1537                         cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1538                 }
1539
1540                 if (tb[TCA_CBQ_FOPT])
1541                         cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1542
1543                 if (cl->q->q.qlen)
1544                         cbq_activate_class(cl);
1545
1546                 sch_tree_unlock(sch);
1547
1548                 return 0;
1549         }
1550
1551         if (parentid == TC_H_ROOT)
1552                 return -EINVAL;
1553
1554         if (tb[TCA_CBQ_WRROPT] == NULL || tb[TCA_CBQ_RATE] == NULL ||
1555             tb[TCA_CBQ_LSSOPT] == NULL)
1556                 return -EINVAL;
1557
1558         rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]), tb[TCA_CBQ_RTAB]);
1559         if (rtab == NULL)
1560                 return -EINVAL;
1561
1562         if (classid) {
1563                 err = -EINVAL;
1564                 if (TC_H_MAJ(classid ^ sch->handle) ||
1565                     cbq_class_lookup(q, classid))
1566                         goto failure;
1567         } else {
1568                 int i;
1569                 classid = TC_H_MAKE(sch->handle, 0x8000);
1570
1571                 for (i = 0; i < 0x8000; i++) {
1572                         if (++q->hgenerator >= 0x8000)
1573                                 q->hgenerator = 1;
1574                         if (cbq_class_lookup(q, classid|q->hgenerator) == NULL)
1575                                 break;
1576                 }
1577                 err = -ENOSR;
1578                 if (i >= 0x8000)
1579                         goto failure;
1580                 classid = classid|q->hgenerator;
1581         }
1582
1583         parent = &q->link;
1584         if (parentid) {
1585                 parent = cbq_class_lookup(q, parentid);
1586                 err = -EINVAL;
1587                 if (parent == NULL)
1588                         goto failure;
1589         }
1590
1591         err = -ENOBUFS;
1592         cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1593         if (cl == NULL)
1594                 goto failure;
1595
1596         err = tcf_block_get(&cl->block, &cl->filter_list);
1597         if (err) {
1598                 kfree(cl);
1599                 return err;
1600         }
1601
1602         if (tca[TCA_RATE]) {
1603                 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1604                                         NULL,
1605                                         qdisc_root_sleeping_running(sch),
1606                                         tca[TCA_RATE]);
1607                 if (err) {
1608                         tcf_block_put(cl->block);
1609                         kfree(cl);
1610                         goto failure;
1611                 }
1612         }
1613
1614         cl->R_tab = rtab;
1615         rtab = NULL;
1616         cl->refcnt = 1;
1617         cl->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, classid);
1618         if (!cl->q)
1619                 cl->q = &noop_qdisc;
1620         else
1621                 qdisc_hash_add(cl->q, true);
1622
1623         cl->common.classid = classid;
1624         cl->tparent = parent;
1625         cl->qdisc = sch;
1626         cl->allot = parent->allot;
1627         cl->quantum = cl->allot;
1628         cl->weight = cl->R_tab->rate.rate;
1629
1630         sch_tree_lock(sch);
1631         cbq_link_class(cl);
1632         cl->borrow = cl->tparent;
1633         if (cl->tparent != &q->link)
1634                 cl->share = cl->tparent;
1635         cbq_adjust_levels(parent);
1636         cl->minidle = -0x7FFFFFFF;
1637         cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1638         cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1639         if (cl->ewma_log == 0)
1640                 cl->ewma_log = q->link.ewma_log;
1641         if (cl->maxidle == 0)
1642                 cl->maxidle = q->link.maxidle;
1643         if (cl->avpkt == 0)
1644                 cl->avpkt = q->link.avpkt;
1645         if (tb[TCA_CBQ_FOPT])
1646                 cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1647         sch_tree_unlock(sch);
1648
1649         qdisc_class_hash_grow(sch, &q->clhash);
1650
1651         *arg = (unsigned long)cl;
1652         return 0;
1653
1654 failure:
1655         qdisc_put_rtab(rtab);
1656         return err;
1657 }
1658
1659 static int cbq_delete(struct Qdisc *sch, unsigned long arg)
1660 {
1661         struct cbq_sched_data *q = qdisc_priv(sch);
1662         struct cbq_class *cl = (struct cbq_class *)arg;
1663         unsigned int qlen, backlog;
1664
1665         if (cl->filters || cl->children || cl == &q->link)
1666                 return -EBUSY;
1667
1668         sch_tree_lock(sch);
1669
1670         qlen = cl->q->q.qlen;
1671         backlog = cl->q->qstats.backlog;
1672         qdisc_reset(cl->q);
1673         qdisc_tree_reduce_backlog(cl->q, qlen, backlog);
1674
1675         if (cl->next_alive)
1676                 cbq_deactivate_class(cl);
1677
1678         if (q->tx_borrowed == cl)
1679                 q->tx_borrowed = q->tx_class;
1680         if (q->tx_class == cl) {
1681                 q->tx_class = NULL;
1682                 q->tx_borrowed = NULL;
1683         }
1684 #ifdef CONFIG_NET_CLS_ACT
1685         if (q->rx_class == cl)
1686                 q->rx_class = NULL;
1687 #endif
1688
1689         cbq_unlink_class(cl);
1690         cbq_adjust_levels(cl->tparent);
1691         cl->defmap = 0;
1692         cbq_sync_defmap(cl);
1693
1694         cbq_rmprio(q, cl);
1695         sch_tree_unlock(sch);
1696
1697         BUG_ON(--cl->refcnt == 0);
1698         /*
1699          * This shouldn't happen: we "hold" one cops->get() when called
1700          * from tc_ctl_tclass; the destroy method is done from cops->put().
1701          */
1702
1703         return 0;
1704 }
1705
1706 static struct tcf_block *cbq_tcf_block(struct Qdisc *sch, unsigned long arg)
1707 {
1708         struct cbq_sched_data *q = qdisc_priv(sch);
1709         struct cbq_class *cl = (struct cbq_class *)arg;
1710
1711         if (cl == NULL)
1712                 cl = &q->link;
1713
1714         return cl->block;
1715 }
1716
1717 static unsigned long cbq_bind_filter(struct Qdisc *sch, unsigned long parent,
1718                                      u32 classid)
1719 {
1720         struct cbq_sched_data *q = qdisc_priv(sch);
1721         struct cbq_class *p = (struct cbq_class *)parent;
1722         struct cbq_class *cl = cbq_class_lookup(q, classid);
1723
1724         if (cl) {
1725                 if (p && p->level <= cl->level)
1726                         return 0;
1727                 cl->filters++;
1728                 return (unsigned long)cl;
1729         }
1730         return 0;
1731 }
1732
1733 static void cbq_unbind_filter(struct Qdisc *sch, unsigned long arg)
1734 {
1735         struct cbq_class *cl = (struct cbq_class *)arg;
1736
1737         cl->filters--;
1738 }
1739
1740 static void cbq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1741 {
1742         struct cbq_sched_data *q = qdisc_priv(sch);
1743         struct cbq_class *cl;
1744         unsigned int h;
1745
1746         if (arg->stop)
1747                 return;
1748
1749         for (h = 0; h < q->clhash.hashsize; h++) {
1750                 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1751                         if (arg->count < arg->skip) {
1752                                 arg->count++;
1753                                 continue;
1754                         }
1755                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1756                                 arg->stop = 1;
1757                                 return;
1758                         }
1759                         arg->count++;
1760                 }
1761         }
1762 }
1763
1764 static const struct Qdisc_class_ops cbq_class_ops = {
1765         .graft          =       cbq_graft,
1766         .leaf           =       cbq_leaf,
1767         .qlen_notify    =       cbq_qlen_notify,
1768         .get            =       cbq_get,
1769         .put            =       cbq_put,
1770         .change         =       cbq_change_class,
1771         .delete         =       cbq_delete,
1772         .walk           =       cbq_walk,
1773         .tcf_block      =       cbq_tcf_block,
1774         .bind_tcf       =       cbq_bind_filter,
1775         .unbind_tcf     =       cbq_unbind_filter,
1776         .dump           =       cbq_dump_class,
1777         .dump_stats     =       cbq_dump_class_stats,
1778 };
1779
1780 static struct Qdisc_ops cbq_qdisc_ops __read_mostly = {
1781         .next           =       NULL,
1782         .cl_ops         =       &cbq_class_ops,
1783         .id             =       "cbq",
1784         .priv_size      =       sizeof(struct cbq_sched_data),
1785         .enqueue        =       cbq_enqueue,
1786         .dequeue        =       cbq_dequeue,
1787         .peek           =       qdisc_peek_dequeued,
1788         .init           =       cbq_init,
1789         .reset          =       cbq_reset,
1790         .destroy        =       cbq_destroy,
1791         .change         =       NULL,
1792         .dump           =       cbq_dump,
1793         .dump_stats     =       cbq_dump_stats,
1794         .owner          =       THIS_MODULE,
1795 };
1796
1797 static int __init cbq_module_init(void)
1798 {
1799         return register_qdisc(&cbq_qdisc_ops);
1800 }
1801 static void __exit cbq_module_exit(void)
1802 {
1803         unregister_qdisc(&cbq_qdisc_ops);
1804 }
1805 module_init(cbq_module_init)
1806 module_exit(cbq_module_exit)
1807 MODULE_LICENSE("GPL");