net/sched/sch_gred.c

   1 /*
   2  * net/sched/sch_gred.c Generic Random Early Detection queue.
   3  *
   4  *
   5  *              This program is free software; you can redistribute it and/or
   6  *              modify it under the terms of the GNU General Public License
   7  *              as published by the Free Software Foundation; either version
   8  *              2 of the License, or (at your option) any later version.
   9  *
  10  * Authors:    J Hadi Salim (hadi@cyberus.ca) 1998-2002
  11  *
  12  *             991129: -  Bug fix with grio mode
  13  *                     - a better sing. AvgQ mode with Grio(WRED)
  14  *                     - A finer grained VQ dequeue based on sugestion
  15  *                       from Ren Liu
  16  *                     - More error checks
  17  *
  18  *  For all the glorious comments look at include/net/red.h
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/module.h>
  23 #include <linux/types.h>
  24 #include <linux/kernel.h>
  25 #include <linux/skbuff.h>
  26 #include <net/pkt_sched.h>
  27 #include <net/red.h>
  28
  29 #define GRED_DEF_PRIO (MAX_DPs / 2)
  30 #define GRED_VQ_MASK (MAX_DPs - 1)
  31
  32 struct gred_sched_data;
  33 struct gred_sched;
  34
  35 struct gred_sched_data
  36 {
  37         u32             limit;          /* HARD maximal queue length    */
  38         u32             DP;             /* the drop pramaters */
  39         u32             bytesin;        /* bytes seen on virtualQ so far*/
  40         u32             packetsin;      /* packets seen on virtualQ so far*/
  41         u32             backlog;        /* bytes on the virtualQ */
  42         u8              prio;           /* the prio of this vq */
  43
  44         struct red_parms parms;
  45         struct red_stats stats;
  46 };
  47
  48 enum {
  49         GRED_WRED_MODE = 1,
  50         GRED_RIO_MODE,
  51 };
  52
  53 struct gred_sched
  54 {
  55         struct gred_sched_data *tab[MAX_DPs];
  56         unsigned long   flags;
  57         u32             red_flags;
  58         u32             DPs;
  59         u32             def;
  60         struct red_parms wred_set;
  61 };
  62
  63 static inline int gred_wred_mode(struct gred_sched *table)
  64 {
  65         return test_bit(GRED_WRED_MODE, &table->flags);
  66 }
  67
  68 static inline void gred_enable_wred_mode(struct gred_sched *table)
  69 {
  70         __set_bit(GRED_WRED_MODE, &table->flags);
  71 }
  72
  73 static inline void gred_disable_wred_mode(struct gred_sched *table)
  74 {
  75         __clear_bit(GRED_WRED_MODE, &table->flags);
  76 }
  77
  78 static inline int gred_rio_mode(struct gred_sched *table)
  79 {
  80         return test_bit(GRED_RIO_MODE, &table->flags);
  81 }
  82
  83 static inline void gred_enable_rio_mode(struct gred_sched *table)
  84 {
  85         __set_bit(GRED_RIO_MODE, &table->flags);
  86 }
  87
  88 static inline void gred_disable_rio_mode(struct gred_sched *table)
  89 {
  90         __clear_bit(GRED_RIO_MODE, &table->flags);
  91 }
  92
  93 static inline int gred_wred_mode_check(struct Qdisc *sch)
  94 {
  95         struct gred_sched *table = qdisc_priv(sch);
  96         int i;
  97
  98         /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
  99         for (i = 0; i < table->DPs; i++) {
 100                 struct gred_sched_data *q = table->tab[i];
 101                 int n;
 102
 103                 if (q == NULL)
 104                         continue;
 105
 106                 for (n = 0; n < table->DPs; n++)
 107                         if (table->tab[n] && table->tab[n] != q &&
 108                             table->tab[n]->prio == q->prio)
 109                                 return 1;
 110         }
 111
 112         return 0;
 113 }
 114
 115 static inline unsigned int gred_backlog(struct gred_sched *table,
 116                                         struct gred_sched_data *q,
 117                                         struct Qdisc *sch)
 118 {
 119         if (gred_wred_mode(table))
 120                 return sch->qstats.backlog;
 121         else
 122                 return q->backlog;
 123 }
 124
 125 static inline u16 tc_index_to_dp(struct sk_buff *skb)
 126 {
 127         return skb->tc_index & GRED_VQ_MASK;
 128 }
 129
 130 static inline void gred_load_wred_set(struct gred_sched *table,
 131                                       struct gred_sched_data *q)
 132 {
 133         q->parms.qavg = table->wred_set.qavg;
 134         q->parms.qidlestart = table->wred_set.qidlestart;
 135 }
 136
 137 static inline void gred_store_wred_set(struct gred_sched *table,
 138                                        struct gred_sched_data *q)
 139 {
 140         table->wred_set.qavg = q->parms.qavg;
 141 }
 142
 143 static inline int gred_use_ecn(struct gred_sched *t)
 144 {
 145         return t->red_flags & TC_RED_ECN;
 146 }
 147
 148 static inline int gred_use_harddrop(struct gred_sched *t)
 149 {
 150         return t->red_flags & TC_RED_HARDDROP;
 151 }
 152
 153 static int gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 154 {
 155         struct gred_sched_data *q=NULL;
 156         struct gred_sched *t= qdisc_priv(sch);
 157         unsigned long qavg = 0;
 158         u16 dp = tc_index_to_dp(skb);
 159
 160         if (dp >= t->DPs  || (q = t->tab[dp]) == NULL) {
 161                 dp = t->def;
 162
 163                 if ((q = t->tab[dp]) == NULL) {
 164                         /* Pass through packets not assigned to a DP
 165                          * if no default DP has been configured. This
 166                          * allows for DP flows to be left untouched.
 167                          */
 168                         if (skb_queue_len(&sch->q) < qdisc_dev(sch)->tx_queue_len)
 169                                 return qdisc_enqueue_tail(skb, sch);
 170                         else
 171                                 goto drop;
 172                 }
 173
 174                 /* fix tc_index? --could be controvesial but needed for
 175                    requeueing */
 176                 skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
 177         }
 178
 179         /* sum up all the qaves of prios <= to ours to get the new qave */
 180         if (!gred_wred_mode(t) && gred_rio_mode(t)) {
 181                 int i;
 182
 183                 for (i = 0; i < t->DPs; i++) {
 184                         if (t->tab[i] && t->tab[i]->prio < q->prio &&
 185                             !red_is_idling(&t->tab[i]->parms))
 186                                 qavg +=t->tab[i]->parms.qavg;
 187                 }
 188
 189         }
 190
 191         q->packetsin++;
 192         q->bytesin += qdisc_pkt_len(skb);
 193
 194         if (gred_wred_mode(t))
 195                 gred_load_wred_set(t, q);
 196
 197         q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
 198
 199         if (red_is_idling(&q->parms))
 200                 red_end_of_idle_period(&q->parms);
 201
 202         if (gred_wred_mode(t))
 203                 gred_store_wred_set(t, q);
 204
 205         switch (red_action(&q->parms, q->parms.qavg + qavg)) {
 206                 case RED_DONT_MARK:
 207                         break;
 208
 209                 case RED_PROB_MARK:
 210                         sch->qstats.overlimits++;
 211                         if (!gred_use_ecn(t) || !INET_ECN_set_ce(skb)) {
 212                                 q->stats.prob_drop++;
 213                                 goto congestion_drop;
 214                         }
 215
 216                         q->stats.prob_mark++;
 217                         break;
 218
 219                 case RED_HARD_MARK:
 220                         sch->qstats.overlimits++;
 221                         if (gred_use_harddrop(t) || !gred_use_ecn(t) ||
 222                             !INET_ECN_set_ce(skb)) {
 223                                 q->stats.forced_drop++;
 224                                 goto congestion_drop;
 225                         }
 226                         q->stats.forced_mark++;
 227                         break;
 228         }
 229
 230         if (q->backlog + qdisc_pkt_len(skb) <= q->limit) {
 231                 q->backlog += qdisc_pkt_len(skb);
 232                 return qdisc_enqueue_tail(skb, sch);
 233         }
 234
 235         q->stats.pdrop++;
 236 drop:
 237         return qdisc_drop(skb, sch);
 238
 239 congestion_drop:
 240         qdisc_drop(skb, sch);
 241         return NET_XMIT_CN;
 242 }
 243
 244 static struct sk_buff *gred_dequeue(struct Qdisc* sch)
 245 {
 246         struct sk_buff *skb;
 247         struct gred_sched *t = qdisc_priv(sch);
 248
 249         skb = qdisc_dequeue_head(sch);
 250
 251         if (skb) {
 252                 struct gred_sched_data *q;
 253                 u16 dp = tc_index_to_dp(skb);
 254
 255                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 256                         if (net_ratelimit())
 257                                 printk(KERN_WARNING "GRED: Unable to relocate "
 258                                        "VQ 0x%x after dequeue, screwing up "
 259                                        "backlog.\n", tc_index_to_dp(skb));
 260                 } else {
 261                         q->backlog -= qdisc_pkt_len(skb);
 262
 263                         if (!q->backlog && !gred_wred_mode(t))
 264                                 red_start_of_idle_period(&q->parms);
 265                 }
 266
 267                 return skb;
 268         }
 269
 270         if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
 271                 red_start_of_idle_period(&t->wred_set);
 272
 273         return NULL;
 274 }
 275
 276 static unsigned int gred_drop(struct Qdisc* sch)
 277 {
 278         struct sk_buff *skb;
 279         struct gred_sched *t = qdisc_priv(sch);
 280
 281         skb = qdisc_dequeue_tail(sch);
 282         if (skb) {
 283                 unsigned int len = qdisc_pkt_len(skb);
 284                 struct gred_sched_data *q;
 285                 u16 dp = tc_index_to_dp(skb);
 286
 287                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 288                         if (net_ratelimit())
 289                                 printk(KERN_WARNING "GRED: Unable to relocate "
 290                                        "VQ 0x%x while dropping, screwing up "
 291                                        "backlog.\n", tc_index_to_dp(skb));
 292                 } else {
 293                         q->backlog -= len;
 294                         q->stats.other++;
 295
 296                         if (!q->backlog && !gred_wred_mode(t))
 297                                 red_start_of_idle_period(&q->parms);
 298                 }
 299
 300                 qdisc_drop(skb, sch);
 301                 return len;
 302         }
 303
 304         if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
 305                 red_start_of_idle_period(&t->wred_set);
 306
 307         return 0;
 308
 309 }
 310
 311 static void gred_reset(struct Qdisc* sch)
 312 {
 313         int i;
 314         struct gred_sched *t = qdisc_priv(sch);
 315
 316         qdisc_reset_queue(sch);
 317
 318         for (i = 0; i < t->DPs; i++) {
 319                 struct gred_sched_data *q = t->tab[i];
 320
 321                 if (!q)
 322                         continue;
 323
 324                 red_restart(&q->parms);
 325                 q->backlog = 0;
 326         }
 327 }
 328
 329 static inline void gred_destroy_vq(struct gred_sched_data *q)
 330 {
 331         kfree(q);
 332 }
 333
 334 static inline int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps)
 335 {
 336         struct gred_sched *table = qdisc_priv(sch);
 337         struct tc_gred_sopt *sopt;
 338         int i;
 339
 340         if (dps == NULL)
 341                 return -EINVAL;
 342
 343         sopt = nla_data(dps);
 344
 345         if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
 346                 return -EINVAL;
 347
 348         sch_tree_lock(sch);
 349         table->DPs = sopt->DPs;
 350         table->def = sopt->def_DP;
 351         table->red_flags = sopt->flags;
 352
 353         /*
 354          * Every entry point to GRED is synchronized with the above code
 355          * and the DP is checked against DPs, i.e. shadowed VQs can no
 356          * longer be found so we can unlock right here.
 357          */
 358         sch_tree_unlock(sch);
 359
 360         if (sopt->grio) {
 361                 gred_enable_rio_mode(table);
 362                 gred_disable_wred_mode(table);
 363                 if (gred_wred_mode_check(sch))
 364                         gred_enable_wred_mode(table);
 365         } else {
 366                 gred_disable_rio_mode(table);
 367                 gred_disable_wred_mode(table);
 368         }
 369
 370         for (i = table->DPs; i < MAX_DPs; i++) {
 371                 if (table->tab[i]) {
 372                         printk(KERN_WARNING "GRED: Warning: Destroying "
 373                                "shadowed VQ 0x%x\n", i);
 374                         gred_destroy_vq(table->tab[i]);
 375                         table->tab[i] = NULL;
 376                 }
 377         }
 378
 379         return 0;
 380 }
 381
 382 static inline int gred_change_vq(struct Qdisc *sch, int dp,
 383                                  struct tc_gred_qopt *ctl, int prio, u8 *stab)
 384 {
 385         struct gred_sched *table = qdisc_priv(sch);
 386         struct gred_sched_data *q;
 387
 388         if (table->tab[dp] == NULL) {
 389                 table->tab[dp] = kzalloc(sizeof(*q), GFP_KERNEL);
 390                 if (table->tab[dp] == NULL)
 391                         return -ENOMEM;
 392         }
 393
 394         q = table->tab[dp];
 395         q->DP = dp;
 396         q->prio = prio;
 397         q->limit = ctl->limit;
 398
 399         if (q->backlog == 0)
 400                 red_end_of_idle_period(&q->parms);
 401
 402         red_set_parms(&q->parms,
 403                       ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
 404                       ctl->Scell_log, stab);
 405
 406         return 0;
 407 }
 408
 409 static const struct nla_policy gred_policy[TCA_GRED_MAX + 1] = {
 410         [TCA_GRED_PARMS]        = { .len = sizeof(struct tc_gred_qopt) },
 411         [TCA_GRED_STAB]         = { .len = 256 },
 412         [TCA_GRED_DPS]          = { .len = sizeof(struct tc_gred_sopt) },
 413 };
 414
 415 static int gred_change(struct Qdisc *sch, struct nlattr *opt)
 416 {
 417         struct gred_sched *table = qdisc_priv(sch);
 418         struct tc_gred_qopt *ctl;
 419         struct nlattr *tb[TCA_GRED_MAX + 1];
 420         int err, prio = GRED_DEF_PRIO;
 421         u8 *stab;
 422
 423         if (opt == NULL)
 424                 return -EINVAL;
 425
 426         err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
 427         if (err < 0)
 428                 return err;
 429
 430         if (tb[TCA_GRED_PARMS] == NULL && tb[TCA_GRED_STAB] == NULL)
 431                 return gred_change_table_def(sch, opt);
 432
 433         if (tb[TCA_GRED_PARMS] == NULL ||
 434             tb[TCA_GRED_STAB] == NULL)
 435                 return -EINVAL;
 436
 437         err = -EINVAL;
 438         ctl = nla_data(tb[TCA_GRED_PARMS]);
 439         stab = nla_data(tb[TCA_GRED_STAB]);
 440
 441         if (ctl->DP >= table->DPs)
 442                 goto errout;
 443
 444         if (gred_rio_mode(table)) {
 445                 if (ctl->prio == 0) {
 446                         int def_prio = GRED_DEF_PRIO;
 447
 448                         if (table->tab[table->def])
 449                                 def_prio = table->tab[table->def]->prio;
 450
 451                         printk(KERN_DEBUG "GRED: DP %u does not have a prio "
 452                                "setting default to %d\n", ctl->DP, def_prio);
 453
 454                         prio = def_prio;
 455                 } else
 456                         prio = ctl->prio;
 457         }
 458
 459         sch_tree_lock(sch);
 460
 461         err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
 462         if (err < 0)
 463                 goto errout_locked;
 464
 465         if (gred_rio_mode(table)) {
 466                 gred_disable_wred_mode(table);
 467                 if (gred_wred_mode_check(sch))
 468                         gred_enable_wred_mode(table);
 469         }
 470
 471         err = 0;
 472
 473 errout_locked:
 474         sch_tree_unlock(sch);
 475 errout:
 476         return err;
 477 }
 478
 479 static int gred_init(struct Qdisc *sch, struct nlattr *opt)
 480 {
 481         struct nlattr *tb[TCA_GRED_MAX + 1];
 482         int err;
 483
 484         if (opt == NULL)
 485                 return -EINVAL;
 486
 487         err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
 488         if (err < 0)
 489                 return err;
 490
 491         if (tb[TCA_GRED_PARMS] || tb[TCA_GRED_STAB])
 492                 return -EINVAL;
 493
 494         return gred_change_table_def(sch, tb[TCA_GRED_DPS]);
 495 }
 496
 497 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
 498 {
 499         struct gred_sched *table = qdisc_priv(sch);
 500         struct nlattr *parms, *opts = NULL;
 501         int i;
 502         struct tc_gred_sopt sopt = {
 503                 .DPs    = table->DPs,
 504                 .def_DP = table->def,
 505                 .grio   = gred_rio_mode(table),
 506                 .flags  = table->red_flags,
 507         };
 508
 509         opts = nla_nest_start(skb, TCA_OPTIONS);
 510         if (opts == NULL)
 511                 goto nla_put_failure;
 512         NLA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
 513         parms = nla_nest_start(skb, TCA_GRED_PARMS);
 514         if (parms == NULL)
 515                 goto nla_put_failure;
 516
 517         for (i = 0; i < MAX_DPs; i++) {
 518                 struct gred_sched_data *q = table->tab[i];
 519                 struct tc_gred_qopt opt;
 520
 521                 memset(&opt, 0, sizeof(opt));
 522
 523                 if (!q) {
 524                         /* hack -- fix at some point with proper message
 525                            This is how we indicate to tc that there is no VQ
 526                            at this DP */
 527
 528                         opt.DP = MAX_DPs + i;
 529                         goto append_opt;
 530                 }
 531
 532                 opt.limit       = q->limit;
 533                 opt.DP          = q->DP;
 534                 opt.backlog     = q->backlog;
 535                 opt.prio        = q->prio;
 536                 opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
 537                 opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
 538                 opt.Wlog        = q->parms.Wlog;
 539                 opt.Plog        = q->parms.Plog;
 540                 opt.Scell_log   = q->parms.Scell_log;
 541                 opt.other       = q->stats.other;
 542                 opt.early       = q->stats.prob_drop;
 543                 opt.forced      = q->stats.forced_drop;
 544                 opt.pdrop       = q->stats.pdrop;
 545                 opt.packets     = q->packetsin;
 546                 opt.bytesin     = q->bytesin;
 547
 548                 if (gred_wred_mode(table)) {
 549                         q->parms.qidlestart =
 550                                 table->tab[table->def]->parms.qidlestart;
 551                         q->parms.qavg = table->tab[table->def]->parms.qavg;
 552                 }
 553
 554                 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
 555
 556 append_opt:
 557                 if (nla_append(skb, sizeof(opt), &opt) < 0)
 558                         goto nla_put_failure;
 559         }
 560
 561         nla_nest_end(skb, parms);
 562
 563         return nla_nest_end(skb, opts);
 564
 565 nla_put_failure:
 566         nla_nest_cancel(skb, opts);
 567         return -EMSGSIZE;
 568 }
 569
 570 static void gred_destroy(struct Qdisc *sch)
 571 {
 572         struct gred_sched *table = qdisc_priv(sch);
 573         int i;
 574
 575         for (i = 0; i < table->DPs; i++) {
 576                 if (table->tab[i])
 577                         gred_destroy_vq(table->tab[i]);
 578         }
 579 }
 580
 581 static struct Qdisc_ops gred_qdisc_ops __read_mostly = {
 582         .id             =       "gred",
 583         .priv_size      =       sizeof(struct gred_sched),
 584         .enqueue        =       gred_enqueue,
 585         .dequeue        =       gred_dequeue,
 586         .peek           =       qdisc_peek_head,
 587         .drop           =       gred_drop,
 588         .init           =       gred_init,
 589         .reset          =       gred_reset,
 590         .destroy        =       gred_destroy,
 591         .change         =       gred_change,
 592         .dump           =       gred_dump,
 593         .owner          =       THIS_MODULE,
 594 };
 595
 596 static int __init gred_module_init(void)
 597 {
 598         return register_qdisc(&gred_qdisc_ops);
 599 }
 600
 601 static void __exit gred_module_exit(void)
 602 {
 603         unregister_qdisc(&gred_qdisc_ops);
 604 }
 605
 606 module_init(gred_module_init)
 607 module_exit(gred_module_exit)
 608
 609 MODULE_LICENSE("GPL");