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         err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1143         if (err < 0)
1144                 return err;
1145
1146         if (tb[TCA_CBQ_RTAB] == NULL || tb[TCA_CBQ_RATE] == NULL)
1147                 return -EINVAL;
1148
1149         r = nla_data(tb[TCA_CBQ_RATE]);
1150
1151         if ((q->link.R_tab = qdisc_get_rtab(r, tb[TCA_CBQ_RTAB])) == NULL)
1152                 return -EINVAL;
1153
1154         err = qdisc_class_hash_init(&q->clhash);
1155         if (err < 0)
1156                 goto put_rtab;
1157
1158         q->link.refcnt = 1;
1159         q->link.sibling = &q->link;
1160         q->link.common.classid = sch->handle;
1161         q->link.qdisc = sch;
1162         q->link.q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1163                                       sch->handle);
1164         if (!q->link.q)
1165                 q->link.q = &noop_qdisc;
1166         else
1167                 qdisc_hash_add(q->link.q, true);
1168
1169         q->link.priority = TC_CBQ_MAXPRIO - 1;
1170         q->link.priority2 = TC_CBQ_MAXPRIO - 1;
1171         q->link.cpriority = TC_CBQ_MAXPRIO - 1;
1172         q->link.allot = psched_mtu(qdisc_dev(sch));
1173         q->link.quantum = q->link.allot;
1174         q->link.weight = q->link.R_tab->rate.rate;
1175
1176         q->link.ewma_log = TC_CBQ_DEF_EWMA;
1177         q->link.avpkt = q->link.allot/2;
1178         q->link.minidle = -0x7FFFFFFF;
1179
1180         qdisc_watchdog_init(&q->watchdog, sch);
1181         hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1182         q->delay_timer.function = cbq_undelay;
1183         q->toplevel = TC_CBQ_MAXLEVEL;
1184         q->now = psched_get_time();
1185
1186         cbq_link_class(&q->link);
1187
1188         if (tb[TCA_CBQ_LSSOPT])
1189                 cbq_set_lss(&q->link, nla_data(tb[TCA_CBQ_LSSOPT]));
1190
1191         cbq_addprio(q, &q->link);
1192         return 0;
1193
1194 put_rtab:
1195         qdisc_put_rtab(q->link.R_tab);
1196         return err;
1197 }
1198
1199 static int cbq_dump_rate(struct sk_buff *skb, struct cbq_class *cl)
1200 {
1201         unsigned char *b = skb_tail_pointer(skb);
1202
1203         if (nla_put(skb, TCA_CBQ_RATE, sizeof(cl->R_tab->rate), &cl->R_tab->rate))
1204                 goto nla_put_failure;
1205         return skb->len;
1206
1207 nla_put_failure:
1208         nlmsg_trim(skb, b);
1209         return -1;
1210 }
1211
1212 static int cbq_dump_lss(struct sk_buff *skb, struct cbq_class *cl)
1213 {
1214         unsigned char *b = skb_tail_pointer(skb);
1215         struct tc_cbq_lssopt opt;
1216
1217         opt.flags = 0;
1218         if (cl->borrow == NULL)
1219                 opt.flags |= TCF_CBQ_LSS_BOUNDED;
1220         if (cl->share == NULL)
1221                 opt.flags |= TCF_CBQ_LSS_ISOLATED;
1222         opt.ewma_log = cl->ewma_log;
1223         opt.level = cl->level;
1224         opt.avpkt = cl->avpkt;
1225         opt.maxidle = cl->maxidle;
1226         opt.minidle = (u32)(-cl->minidle);
1227         opt.offtime = cl->offtime;
1228         opt.change = ~0;
1229         if (nla_put(skb, TCA_CBQ_LSSOPT, sizeof(opt), &opt))
1230                 goto nla_put_failure;
1231         return skb->len;
1232
1233 nla_put_failure:
1234         nlmsg_trim(skb, b);
1235         return -1;
1236 }
1237
1238 static int cbq_dump_wrr(struct sk_buff *skb, struct cbq_class *cl)
1239 {
1240         unsigned char *b = skb_tail_pointer(skb);
1241         struct tc_cbq_wrropt opt;
1242
1243         memset(&opt, 0, sizeof(opt));
1244         opt.flags = 0;
1245         opt.allot = cl->allot;
1246         opt.priority = cl->priority + 1;
1247         opt.cpriority = cl->cpriority + 1;
1248         opt.weight = cl->weight;
1249         if (nla_put(skb, TCA_CBQ_WRROPT, sizeof(opt), &opt))
1250                 goto nla_put_failure;
1251         return skb->len;
1252
1253 nla_put_failure:
1254         nlmsg_trim(skb, b);
1255         return -1;
1256 }
1257
1258 static int cbq_dump_fopt(struct sk_buff *skb, struct cbq_class *cl)
1259 {
1260         unsigned char *b = skb_tail_pointer(skb);
1261         struct tc_cbq_fopt opt;
1262
1263         if (cl->split || cl->defmap) {
1264                 opt.split = cl->split ? cl->split->common.classid : 0;
1265                 opt.defmap = cl->defmap;
1266                 opt.defchange = ~0;
1267                 if (nla_put(skb, TCA_CBQ_FOPT, sizeof(opt), &opt))
1268                         goto nla_put_failure;
1269         }
1270         return skb->len;
1271
1272 nla_put_failure:
1273         nlmsg_trim(skb, b);
1274         return -1;
1275 }
1276
1277 static int cbq_dump_attr(struct sk_buff *skb, struct cbq_class *cl)
1278 {
1279         if (cbq_dump_lss(skb, cl) < 0 ||
1280             cbq_dump_rate(skb, cl) < 0 ||
1281             cbq_dump_wrr(skb, cl) < 0 ||
1282             cbq_dump_fopt(skb, cl) < 0)
1283                 return -1;
1284         return 0;
1285 }
1286
1287 static int cbq_dump(struct Qdisc *sch, struct sk_buff *skb)
1288 {
1289         struct cbq_sched_data *q = qdisc_priv(sch);
1290         struct nlattr *nest;
1291
1292         nest = nla_nest_start(skb, TCA_OPTIONS);
1293         if (nest == NULL)
1294                 goto nla_put_failure;
1295         if (cbq_dump_attr(skb, &q->link) < 0)
1296                 goto nla_put_failure;
1297         return nla_nest_end(skb, nest);
1298
1299 nla_put_failure:
1300         nla_nest_cancel(skb, nest);
1301         return -1;
1302 }
1303
1304 static int
1305 cbq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1306 {
1307         struct cbq_sched_data *q = qdisc_priv(sch);
1308
1309         q->link.xstats.avgidle = q->link.avgidle;
1310         return gnet_stats_copy_app(d, &q->link.xstats, sizeof(q->link.xstats));
1311 }
1312
1313 static int
1314 cbq_dump_class(struct Qdisc *sch, unsigned long arg,
1315                struct sk_buff *skb, struct tcmsg *tcm)
1316 {
1317         struct cbq_class *cl = (struct cbq_class *)arg;
1318         struct nlattr *nest;
1319
1320         if (cl->tparent)
1321                 tcm->tcm_parent = cl->tparent->common.classid;
1322         else
1323                 tcm->tcm_parent = TC_H_ROOT;
1324         tcm->tcm_handle = cl->common.classid;
1325         tcm->tcm_info = cl->q->handle;
1326
1327         nest = nla_nest_start(skb, TCA_OPTIONS);
1328         if (nest == NULL)
1329                 goto nla_put_failure;
1330         if (cbq_dump_attr(skb, cl) < 0)
1331                 goto nla_put_failure;
1332         return nla_nest_end(skb, nest);
1333
1334 nla_put_failure:
1335         nla_nest_cancel(skb, nest);
1336         return -1;
1337 }
1338
1339 static int
1340 cbq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1341         struct gnet_dump *d)
1342 {
1343         struct cbq_sched_data *q = qdisc_priv(sch);
1344         struct cbq_class *cl = (struct cbq_class *)arg;
1345
1346         cl->xstats.avgidle = cl->avgidle;
1347         cl->xstats.undertime = 0;
1348
1349         if (cl->undertime != PSCHED_PASTPERFECT)
1350                 cl->xstats.undertime = cl->undertime - q->now;
1351
1352         if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1353                                   d, NULL, &cl->bstats) < 0 ||
1354             gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1355             gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->q->q.qlen) < 0)
1356                 return -1;
1357
1358         return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1359 }
1360
1361 static int cbq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1362                      struct Qdisc **old)
1363 {
1364         struct cbq_class *cl = (struct cbq_class *)arg;
1365
1366         if (new == NULL) {
1367                 new = qdisc_create_dflt(sch->dev_queue,
1368                                         &pfifo_qdisc_ops, cl->common.classid);
1369                 if (new == NULL)
1370                         return -ENOBUFS;
1371         }
1372
1373         *old = qdisc_replace(sch, new, &cl->q);
1374         return 0;
1375 }
1376
1377 static struct Qdisc *cbq_leaf(struct Qdisc *sch, unsigned long arg)
1378 {
1379         struct cbq_class *cl = (struct cbq_class *)arg;
1380
1381         return cl->q;
1382 }
1383
1384 static void cbq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1385 {
1386         struct cbq_class *cl = (struct cbq_class *)arg;
1387
1388         if (cl->q->q.qlen == 0)
1389                 cbq_deactivate_class(cl);
1390 }
1391
1392 static unsigned long cbq_get(struct Qdisc *sch, u32 classid)
1393 {
1394         struct cbq_sched_data *q = qdisc_priv(sch);
1395         struct cbq_class *cl = cbq_class_lookup(q, classid);
1396
1397         if (cl) {
1398                 cl->refcnt++;
1399                 return (unsigned long)cl;
1400         }
1401         return 0;
1402 }
1403
1404 static void cbq_destroy_class(struct Qdisc *sch, struct cbq_class *cl)
1405 {
1406         struct cbq_sched_data *q = qdisc_priv(sch);
1407
1408         WARN_ON(cl->filters);
1409
1410         tcf_block_put(cl->block);
1411         qdisc_destroy(cl->q);
1412         qdisc_put_rtab(cl->R_tab);
1413         gen_kill_estimator(&cl->rate_est);
1414         if (cl != &q->link)
1415                 kfree(cl);
1416 }
1417
1418 static void cbq_destroy(struct Qdisc *sch)
1419 {
1420         struct cbq_sched_data *q = qdisc_priv(sch);
1421         struct hlist_node *next;
1422         struct cbq_class *cl;
1423         unsigned int h;
1424
1425 #ifdef CONFIG_NET_CLS_ACT
1426         q->rx_class = NULL;
1427 #endif
1428         /*
1429          * Filters must be destroyed first because we don't destroy the
1430          * classes from root to leafs which means that filters can still
1431          * be bound to classes which have been destroyed already. --TGR '04
1432          */
1433         for (h = 0; h < q->clhash.hashsize; h++) {
1434                 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1435                         tcf_block_put(cl->block);
1436                         cl->block = NULL;
1437                 }
1438         }
1439         for (h = 0; h < q->clhash.hashsize; h++) {
1440                 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[h],
1441                                           common.hnode)
1442                         cbq_destroy_class(sch, cl);
1443         }
1444         qdisc_class_hash_destroy(&q->clhash);
1445 }
1446
1447 static void cbq_put(struct Qdisc *sch, unsigned long arg)
1448 {
1449         struct cbq_class *cl = (struct cbq_class *)arg;
1450
1451         if (--cl->refcnt == 0) {
1452 #ifdef CONFIG_NET_CLS_ACT
1453                 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1454                 struct cbq_sched_data *q = qdisc_priv(sch);
1455
1456                 spin_lock_bh(root_lock);
1457                 if (q->rx_class == cl)
1458                         q->rx_class = NULL;
1459                 spin_unlock_bh(root_lock);
1460 #endif
1461
1462                 cbq_destroy_class(sch, cl);
1463         }
1464 }
1465
1466 static int
1467 cbq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, struct nlattr **tca,
1468                  unsigned long *arg)
1469 {
1470         int err;
1471         struct cbq_sched_data *q = qdisc_priv(sch);
1472         struct cbq_class *cl = (struct cbq_class *)*arg;
1473         struct nlattr *opt = tca[TCA_OPTIONS];
1474         struct nlattr *tb[TCA_CBQ_MAX + 1];
1475         struct cbq_class *parent;
1476         struct qdisc_rate_table *rtab = NULL;
1477
1478         if (opt == NULL)
1479                 return -EINVAL;
1480
1481         err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1482         if (err < 0)
1483                 return err;
1484
1485         if (tb[TCA_CBQ_OVL_STRATEGY] || tb[TCA_CBQ_POLICE])
1486                 return -EOPNOTSUPP;
1487
1488         if (cl) {
1489                 /* Check parent */
1490                 if (parentid) {
1491                         if (cl->tparent &&
1492                             cl->tparent->common.classid != parentid)
1493                                 return -EINVAL;
1494                         if (!cl->tparent && parentid != TC_H_ROOT)
1495                                 return -EINVAL;
1496                 }
1497
1498                 if (tb[TCA_CBQ_RATE]) {
1499                         rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]),
1500                                               tb[TCA_CBQ_RTAB]);
1501                         if (rtab == NULL)
1502                                 return -EINVAL;
1503                 }
1504
1505                 if (tca[TCA_RATE]) {
1506                         err = gen_replace_estimator(&cl->bstats, NULL,
1507                                                     &cl->rate_est,
1508                                                     NULL,
1509                                                     qdisc_root_sleeping_running(sch),
1510                                                     tca[TCA_RATE]);
1511                         if (err) {
1512                                 qdisc_put_rtab(rtab);
1513                                 return err;
1514                         }
1515                 }
1516
1517                 /* Change class parameters */
1518                 sch_tree_lock(sch);
1519
1520                 if (cl->next_alive != NULL)
1521                         cbq_deactivate_class(cl);
1522
1523                 if (rtab) {
1524                         qdisc_put_rtab(cl->R_tab);
1525                         cl->R_tab = rtab;
1526                 }
1527
1528                 if (tb[TCA_CBQ_LSSOPT])
1529                         cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1530
1531                 if (tb[TCA_CBQ_WRROPT]) {
1532                         cbq_rmprio(q, cl);
1533                         cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1534                 }
1535
1536                 if (tb[TCA_CBQ_FOPT])
1537                         cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1538
1539                 if (cl->q->q.qlen)
1540                         cbq_activate_class(cl);
1541
1542                 sch_tree_unlock(sch);
1543
1544                 return 0;
1545         }
1546
1547         if (parentid == TC_H_ROOT)
1548                 return -EINVAL;
1549
1550         if (tb[TCA_CBQ_WRROPT] == NULL || tb[TCA_CBQ_RATE] == NULL ||
1551             tb[TCA_CBQ_LSSOPT] == NULL)
1552                 return -EINVAL;
1553
1554         rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]), tb[TCA_CBQ_RTAB]);
1555         if (rtab == NULL)
1556                 return -EINVAL;
1557
1558         if (classid) {
1559                 err = -EINVAL;
1560                 if (TC_H_MAJ(classid ^ sch->handle) ||
1561                     cbq_class_lookup(q, classid))
1562                         goto failure;
1563         } else {
1564                 int i;
1565                 classid = TC_H_MAKE(sch->handle, 0x8000);
1566
1567                 for (i = 0; i < 0x8000; i++) {
1568                         if (++q->hgenerator >= 0x8000)
1569                                 q->hgenerator = 1;
1570                         if (cbq_class_lookup(q, classid|q->hgenerator) == NULL)
1571                                 break;
1572                 }
1573                 err = -ENOSR;
1574                 if (i >= 0x8000)
1575                         goto failure;
1576                 classid = classid|q->hgenerator;
1577         }
1578
1579         parent = &q->link;
1580         if (parentid) {
1581                 parent = cbq_class_lookup(q, parentid);
1582                 err = -EINVAL;
1583                 if (parent == NULL)
1584                         goto failure;
1585         }
1586
1587         err = -ENOBUFS;
1588         cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1589         if (cl == NULL)
1590                 goto failure;
1591
1592         err = tcf_block_get(&cl->block, &cl->filter_list);
1593         if (err) {
1594                 kfree(cl);
1595                 return err;
1596         }
1597
1598         if (tca[TCA_RATE]) {
1599                 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1600                                         NULL,
1601                                         qdisc_root_sleeping_running(sch),
1602                                         tca[TCA_RATE]);
1603                 if (err) {
1604                         tcf_block_put(cl->block);
1605                         kfree(cl);
1606                         goto failure;
1607                 }
1608         }
1609
1610         cl->R_tab = rtab;
1611         rtab = NULL;
1612         cl->refcnt = 1;
1613         cl->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, classid);
1614         if (!cl->q)
1615                 cl->q = &noop_qdisc;
1616         else
1617                 qdisc_hash_add(cl->q, true);
1618
1619         cl->common.classid = classid;
1620         cl->tparent = parent;
1621         cl->qdisc = sch;
1622         cl->allot = parent->allot;
1623         cl->quantum = cl->allot;
1624         cl->weight = cl->R_tab->rate.rate;
1625
1626         sch_tree_lock(sch);
1627         cbq_link_class(cl);
1628         cl->borrow = cl->tparent;
1629         if (cl->tparent != &q->link)
1630                 cl->share = cl->tparent;
1631         cbq_adjust_levels(parent);
1632         cl->minidle = -0x7FFFFFFF;
1633         cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1634         cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1635         if (cl->ewma_log == 0)
1636                 cl->ewma_log = q->link.ewma_log;
1637         if (cl->maxidle == 0)
1638                 cl->maxidle = q->link.maxidle;
1639         if (cl->avpkt == 0)
1640                 cl->avpkt = q->link.avpkt;
1641         if (tb[TCA_CBQ_FOPT])
1642                 cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1643         sch_tree_unlock(sch);
1644
1645         qdisc_class_hash_grow(sch, &q->clhash);
1646
1647         *arg = (unsigned long)cl;
1648         return 0;
1649
1650 failure:
1651         qdisc_put_rtab(rtab);
1652         return err;
1653 }
1654
1655 static int cbq_delete(struct Qdisc *sch, unsigned long arg)
1656 {
1657         struct cbq_sched_data *q = qdisc_priv(sch);
1658         struct cbq_class *cl = (struct cbq_class *)arg;
1659         unsigned int qlen, backlog;
1660
1661         if (cl->filters || cl->children || cl == &q->link)
1662                 return -EBUSY;
1663
1664         sch_tree_lock(sch);
1665
1666         qlen = cl->q->q.qlen;
1667         backlog = cl->q->qstats.backlog;
1668         qdisc_reset(cl->q);
1669         qdisc_tree_reduce_backlog(cl->q, qlen, backlog);
1670
1671         if (cl->next_alive)
1672                 cbq_deactivate_class(cl);
1673
1674         if (q->tx_borrowed == cl)
1675                 q->tx_borrowed = q->tx_class;
1676         if (q->tx_class == cl) {
1677                 q->tx_class = NULL;
1678                 q->tx_borrowed = NULL;
1679         }
1680 #ifdef CONFIG_NET_CLS_ACT
1681         if (q->rx_class == cl)
1682                 q->rx_class = NULL;
1683 #endif
1684
1685         cbq_unlink_class(cl);
1686         cbq_adjust_levels(cl->tparent);
1687         cl->defmap = 0;
1688         cbq_sync_defmap(cl);
1689
1690         cbq_rmprio(q, cl);
1691         sch_tree_unlock(sch);
1692
1693         BUG_ON(--cl->refcnt == 0);
1694         /*
1695          * This shouldn't happen: we "hold" one cops->get() when called
1696          * from tc_ctl_tclass; the destroy method is done from cops->put().
1697          */
1698
1699         return 0;
1700 }
1701
1702 static struct tcf_block *cbq_tcf_block(struct Qdisc *sch, unsigned long arg)
1703 {
1704         struct cbq_sched_data *q = qdisc_priv(sch);
1705         struct cbq_class *cl = (struct cbq_class *)arg;
1706
1707         if (cl == NULL)
1708                 cl = &q->link;
1709
1710         return cl->block;
1711 }
1712
1713 static unsigned long cbq_bind_filter(struct Qdisc *sch, unsigned long parent,
1714                                      u32 classid)
1715 {
1716         struct cbq_sched_data *q = qdisc_priv(sch);
1717         struct cbq_class *p = (struct cbq_class *)parent;
1718         struct cbq_class *cl = cbq_class_lookup(q, classid);
1719
1720         if (cl) {
1721                 if (p && p->level <= cl->level)
1722                         return 0;
1723                 cl->filters++;
1724                 return (unsigned long)cl;
1725         }
1726         return 0;
1727 }
1728
1729 static void cbq_unbind_filter(struct Qdisc *sch, unsigned long arg)
1730 {
1731         struct cbq_class *cl = (struct cbq_class *)arg;
1732
1733         cl->filters--;
1734 }
1735
1736 static void cbq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1737 {
1738         struct cbq_sched_data *q = qdisc_priv(sch);
1739         struct cbq_class *cl;
1740         unsigned int h;
1741
1742         if (arg->stop)
1743                 return;
1744
1745         for (h = 0; h < q->clhash.hashsize; h++) {
1746                 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1747                         if (arg->count < arg->skip) {
1748                                 arg->count++;
1749                                 continue;
1750                         }
1751                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1752                                 arg->stop = 1;
1753                                 return;
1754                         }
1755                         arg->count++;
1756                 }
1757         }
1758 }
1759
1760 static const struct Qdisc_class_ops cbq_class_ops = {
1761         .graft          =       cbq_graft,
1762         .leaf           =       cbq_leaf,
1763         .qlen_notify    =       cbq_qlen_notify,
1764         .get            =       cbq_get,
1765         .put            =       cbq_put,
1766         .change         =       cbq_change_class,
1767         .delete         =       cbq_delete,
1768         .walk           =       cbq_walk,
1769         .tcf_block      =       cbq_tcf_block,
1770         .bind_tcf       =       cbq_bind_filter,
1771         .unbind_tcf     =       cbq_unbind_filter,
1772         .dump           =       cbq_dump_class,
1773         .dump_stats     =       cbq_dump_class_stats,
1774 };
1775
1776 static struct Qdisc_ops cbq_qdisc_ops __read_mostly = {
1777         .next           =       NULL,
1778         .cl_ops         =       &cbq_class_ops,
1779         .id             =       "cbq",
1780         .priv_size      =       sizeof(struct cbq_sched_data),
1781         .enqueue        =       cbq_enqueue,
1782         .dequeue        =       cbq_dequeue,
1783         .peek           =       qdisc_peek_dequeued,
1784         .init           =       cbq_init,
1785         .reset          =       cbq_reset,
1786         .destroy        =       cbq_destroy,
1787         .change         =       NULL,
1788         .dump           =       cbq_dump,
1789         .dump_stats     =       cbq_dump_stats,
1790         .owner          =       THIS_MODULE,
1791 };
1792
1793 static int __init cbq_module_init(void)
1794 {
1795         return register_qdisc(&cbq_qdisc_ops);
1796 }
1797 static void __exit cbq_module_exit(void)
1798 {
1799         unregister_qdisc(&cbq_qdisc_ops);
1800 }
1801 module_init(cbq_module_init)
1802 module_exit(cbq_module_exit)
1803 MODULE_LICENSE("GPL");