kernel/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         u32             limit;          /* HARD maximal queue length    */
  37         u32             DP;             /* the drop parameters */
  38         u32             bytesin;        /* bytes seen on virtualQ so far*/
  39         u32             packetsin;      /* packets seen on virtualQ so far*/
  40         u32             backlog;        /* bytes on the virtualQ */
  41         u8              prio;           /* the prio of this vq */
  42
  43         struct red_parms parms;
  44         struct red_vars  vars;
  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         struct gred_sched_data *tab[MAX_DPs];
  55         unsigned long   flags;
  56         u32             red_flags;
  57         u32             DPs;
  58         u32             def;
  59         struct red_vars wred_set;
  60 };
  61
  62 static inline int gred_wred_mode(struct gred_sched *table)
  63 {
  64         return test_bit(GRED_WRED_MODE, &table->flags);
  65 }
  66
  67 static inline void gred_enable_wred_mode(struct gred_sched *table)
  68 {
  69         __set_bit(GRED_WRED_MODE, &table->flags);
  70 }
  71
  72 static inline void gred_disable_wred_mode(struct gred_sched *table)
  73 {
  74         __clear_bit(GRED_WRED_MODE, &table->flags);
  75 }
  76
  77 static inline int gred_rio_mode(struct gred_sched *table)
  78 {
  79         return test_bit(GRED_RIO_MODE, &table->flags);
  80 }
  81
  82 static inline void gred_enable_rio_mode(struct gred_sched *table)
  83 {
  84         __set_bit(GRED_RIO_MODE, &table->flags);
  85 }
  86
  87 static inline void gred_disable_rio_mode(struct gred_sched *table)
  88 {
  89         __clear_bit(GRED_RIO_MODE, &table->flags);
  90 }
  91
  92 static inline int gred_wred_mode_check(struct Qdisc *sch)
  93 {
  94         struct gred_sched *table = qdisc_priv(sch);
  95         int i;
  96
  97         /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
  98         for (i = 0; i < table->DPs; i++) {
  99                 struct gred_sched_data *q = table->tab[i];
 100                 int n;
 101
 102                 if (q == NULL)
 103                         continue;
 104
 105                 for (n = i + 1; n < table->DPs; n++)
 106                         if (table->tab[n] && table->tab[n]->prio == q->prio)
 107                                 return 1;
 108         }
 109
 110         return 0;
 111 }
 112
 113 static inline unsigned int gred_backlog(struct gred_sched *table,
 114                                         struct gred_sched_data *q,
 115                                         struct Qdisc *sch)
 116 {
 117         if (gred_wred_mode(table))
 118                 return sch->qstats.backlog;
 119         else
 120                 return q->backlog;
 121 }
 122
 123 static inline u16 tc_index_to_dp(struct sk_buff *skb)
 124 {
 125         return skb->tc_index & GRED_VQ_MASK;
 126 }
 127
 128 static inline void gred_load_wred_set(const struct gred_sched *table,
 129                                       struct gred_sched_data *q)
 130 {
 131         q->vars.qavg = table->wred_set.qavg;
 132         q->vars.qidlestart = table->wred_set.qidlestart;
 133 }
 134
 135 static inline void gred_store_wred_set(struct gred_sched *table,
 136                                        struct gred_sched_data *q)
 137 {
 138         table->wred_set.qavg = q->vars.qavg;
 139         table->wred_set.qidlestart = q->vars.qidlestart;
 140 }
 141
 142 static inline int gred_use_ecn(struct gred_sched *t)
 143 {
 144         return t->red_flags & TC_RED_ECN;
 145 }
 146
 147 static inline int gred_use_harddrop(struct gred_sched *t)
 148 {
 149         return t->red_flags & TC_RED_HARDDROP;
 150 }
 151
 152 static int gred_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 153 {
 154         struct gred_sched_data *q = NULL;
 155         struct gred_sched *t = qdisc_priv(sch);
 156         unsigned long qavg = 0;
 157         u16 dp = tc_index_to_dp(skb);
 158
 159         if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 160                 dp = t->def;
 161
 162                 q = t->tab[dp];
 163                 if (!q) {
 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 (likely(sch->qstats.backlog + qdisc_pkt_len(skb) <=
 169                                         sch->limit))
 170                                 return qdisc_enqueue_tail(skb, sch);
 171                         else
 172                                 goto drop;
 173                 }
 174
 175                 /* fix tc_index? --could be controversial but needed for
 176                    requeueing */
 177                 skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
 178         }
 179
 180         /* sum up all the qaves of prios < ours to get the new qave */
 181         if (!gred_wred_mode(t) && gred_rio_mode(t)) {
 182                 int i;
 183
 184                 for (i = 0; i < t->DPs; i++) {
 185                         if (t->tab[i] && t->tab[i]->prio < q->prio &&
 186                             !red_is_idling(&t->tab[i]->vars))
 187                                 qavg += t->tab[i]->vars.qavg;
 188                 }
 189
 190         }
 191
 192         q->packetsin++;
 193         q->bytesin += qdisc_pkt_len(skb);
 194
 195         if (gred_wred_mode(t))
 196                 gred_load_wred_set(t, q);
 197
 198         q->vars.qavg = red_calc_qavg(&q->parms,
 199                                      &q->vars,
 200                                      gred_backlog(t, q, sch));
 201
 202         if (red_is_idling(&q->vars))
 203                 red_end_of_idle_period(&q->vars);
 204
 205         if (gred_wred_mode(t))
 206                 gred_store_wred_set(t, q);
 207
 208         switch (red_action(&q->parms, &q->vars, q->vars.qavg + qavg)) {
 209         case RED_DONT_MARK:
 210                 break;
 211
 212         case RED_PROB_MARK:
 213                 qdisc_qstats_overlimit(sch);
 214                 if (!gred_use_ecn(t) || !INET_ECN_set_ce(skb)) {
 215                         q->stats.prob_drop++;
 216                         goto congestion_drop;
 217                 }
 218
 219                 q->stats.prob_mark++;
 220                 break;
 221
 222         case RED_HARD_MARK:
 223                 qdisc_qstats_overlimit(sch);
 224                 if (gred_use_harddrop(t) || !gred_use_ecn(t) ||
 225                     !INET_ECN_set_ce(skb)) {
 226                         q->stats.forced_drop++;
 227                         goto congestion_drop;
 228                 }
 229                 q->stats.forced_mark++;
 230                 break;
 231         }
 232
 233         if (gred_backlog(t, q, sch) + qdisc_pkt_len(skb) <= q->limit) {
 234                 q->backlog += qdisc_pkt_len(skb);
 235                 return qdisc_enqueue_tail(skb, sch);
 236         }
 237
 238         q->stats.pdrop++;
 239 drop:
 240         return qdisc_drop(skb, sch);
 241
 242 congestion_drop:
 243         qdisc_drop(skb, sch);
 244         return NET_XMIT_CN;
 245 }
 246
 247 static struct sk_buff *gred_dequeue(struct Qdisc *sch)
 248 {
 249         struct sk_buff *skb;
 250         struct gred_sched *t = qdisc_priv(sch);
 251
 252         skb = qdisc_dequeue_head(sch);
 253
 254         if (skb) {
 255                 struct gred_sched_data *q;
 256                 u16 dp = tc_index_to_dp(skb);
 257
 258                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 259                         net_warn_ratelimited("GRED: Unable to relocate VQ 0x%x after dequeue, screwing up backlog\n",
 260                                              tc_index_to_dp(skb));
 261                 } else {
 262                         q->backlog -= qdisc_pkt_len(skb);
 263
 264                         if (gred_wred_mode(t)) {
 265                                 if (!sch->qstats.backlog)
 266                                         red_start_of_idle_period(&t->wred_set);
 267                         } else {
 268                                 if (!q->backlog)
 269                                         red_start_of_idle_period(&q->vars);
 270                         }
 271                 }
 272
 273                 return skb;
 274         }
 275
 276         return NULL;
 277 }
 278
 279 static unsigned int gred_drop(struct Qdisc *sch)
 280 {
 281         struct sk_buff *skb;
 282         struct gred_sched *t = qdisc_priv(sch);
 283
 284         skb = qdisc_dequeue_tail(sch);
 285         if (skb) {
 286                 unsigned int len = qdisc_pkt_len(skb);
 287                 struct gred_sched_data *q;
 288                 u16 dp = tc_index_to_dp(skb);
 289
 290                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 291                         net_warn_ratelimited("GRED: Unable to relocate VQ 0x%x while dropping, screwing up backlog\n",
 292                                              tc_index_to_dp(skb));
 293                 } else {
 294                         q->backlog -= len;
 295                         q->stats.other++;
 296
 297                         if (gred_wred_mode(t)) {
 298                                 if (!sch->qstats.backlog)
 299                                         red_start_of_idle_period(&t->wred_set);
 300                         } else {
 301                                 if (!q->backlog)
 302                                         red_start_of_idle_period(&q->vars);
 303                         }
 304                 }
 305
 306                 qdisc_drop(skb, sch);
 307                 return len;
 308         }
 309
 310         return 0;
 311 }
 312
 313 static void gred_reset(struct Qdisc *sch)
 314 {
 315         int i;
 316         struct gred_sched *t = qdisc_priv(sch);
 317
 318         qdisc_reset_queue(sch);
 319
 320         for (i = 0; i < t->DPs; i++) {
 321                 struct gred_sched_data *q = t->tab[i];
 322
 323                 if (!q)
 324                         continue;
 325
 326                 red_restart(&q->vars);
 327                 q->backlog = 0;
 328         }
 329 }
 330
 331 static inline void gred_destroy_vq(struct gred_sched_data *q)
 332 {
 333         kfree(q);
 334 }
 335
 336 static inline int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps)
 337 {
 338         struct gred_sched *table = qdisc_priv(sch);
 339         struct tc_gred_sopt *sopt;
 340         int i;
 341
 342         if (dps == NULL)
 343                 return -EINVAL;
 344
 345         sopt = nla_data(dps);
 346
 347         if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
 348                 return -EINVAL;
 349
 350         sch_tree_lock(sch);
 351         table->DPs = sopt->DPs;
 352         table->def = sopt->def_DP;
 353         table->red_flags = sopt->flags;
 354
 355         /*
 356          * Every entry point to GRED is synchronized with the above code
 357          * and the DP is checked against DPs, i.e. shadowed VQs can no
 358          * longer be found so we can unlock right here.
 359          */
 360         sch_tree_unlock(sch);
 361
 362         if (sopt->grio) {
 363                 gred_enable_rio_mode(table);
 364                 gred_disable_wred_mode(table);
 365                 if (gred_wred_mode_check(sch))
 366                         gred_enable_wred_mode(table);
 367         } else {
 368                 gred_disable_rio_mode(table);
 369                 gred_disable_wred_mode(table);
 370         }
 371
 372         for (i = table->DPs; i < MAX_DPs; i++) {
 373                 if (table->tab[i]) {
 374                         pr_warn("GRED: Warning: Destroying shadowed VQ 0x%x\n",
 375                                 i);
 376                         gred_destroy_vq(table->tab[i]);
 377                         table->tab[i] = NULL;
 378                 }
 379         }
 380
 381         return 0;
 382 }
 383
 384 static inline int gred_change_vq(struct Qdisc *sch, int dp,
 385                                  struct tc_gred_qopt *ctl, int prio,
 386                                  u8 *stab, u32 max_P,
 387                                  struct gred_sched_data **prealloc)
 388 {
 389         struct gred_sched *table = qdisc_priv(sch);
 390         struct gred_sched_data *q = table->tab[dp];
 391
 392         if (!q) {
 393                 table->tab[dp] = q = *prealloc;
 394                 *prealloc = NULL;
 395                 if (!q)
 396                         return -ENOMEM;
 397         }
 398
 399         q->DP = dp;
 400         q->prio = prio;
 401         if (ctl->limit > sch->limit)
 402                 q->limit = sch->limit;
 403         else
 404                 q->limit = ctl->limit;
 405
 406         if (q->backlog == 0)
 407                 red_end_of_idle_period(&q->vars);
 408
 409         red_set_parms(&q->parms,
 410                       ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
 411                       ctl->Scell_log, stab, max_P);
 412         red_set_vars(&q->vars);
 413         return 0;
 414 }
 415
 416 static const struct nla_policy gred_policy[TCA_GRED_MAX + 1] = {
 417         [TCA_GRED_PARMS]        = { .len = sizeof(struct tc_gred_qopt) },
 418         [TCA_GRED_STAB]         = { .len = 256 },
 419         [TCA_GRED_DPS]          = { .len = sizeof(struct tc_gred_sopt) },
 420         [TCA_GRED_MAX_P]        = { .type = NLA_U32 },
 421         [TCA_GRED_LIMIT]        = { .type = NLA_U32 },
 422 };
 423
 424 static int gred_change(struct Qdisc *sch, struct nlattr *opt)
 425 {
 426         struct gred_sched *table = qdisc_priv(sch);
 427         struct tc_gred_qopt *ctl;
 428         struct nlattr *tb[TCA_GRED_MAX + 1];
 429         int err, prio = GRED_DEF_PRIO;
 430         u8 *stab;
 431         u32 max_P;
 432         struct gred_sched_data *prealloc;
 433
 434         if (opt == NULL)
 435                 return -EINVAL;
 436
 437         err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
 438         if (err < 0)
 439                 return err;
 440
 441         if (tb[TCA_GRED_PARMS] == NULL && tb[TCA_GRED_STAB] == NULL) {
 442                 if (tb[TCA_GRED_LIMIT] != NULL)
 443                         sch->limit = nla_get_u32(tb[TCA_GRED_LIMIT]);
 444                 return gred_change_table_def(sch, opt);
 445         }
 446
 447         if (tb[TCA_GRED_PARMS] == NULL ||
 448             tb[TCA_GRED_STAB] == NULL ||
 449             tb[TCA_GRED_LIMIT] != NULL)
 450                 return -EINVAL;
 451
 452         max_P = tb[TCA_GRED_MAX_P] ? nla_get_u32(tb[TCA_GRED_MAX_P]) : 0;
 453
 454         err = -EINVAL;
 455         ctl = nla_data(tb[TCA_GRED_PARMS]);
 456         stab = nla_data(tb[TCA_GRED_STAB]);
 457
 458         if (ctl->DP >= table->DPs)
 459                 goto errout;
 460
 461         if (gred_rio_mode(table)) {
 462                 if (ctl->prio == 0) {
 463                         int def_prio = GRED_DEF_PRIO;
 464
 465                         if (table->tab[table->def])
 466                                 def_prio = table->tab[table->def]->prio;
 467
 468                         printk(KERN_DEBUG "GRED: DP %u does not have a prio "
 469                                "setting default to %d\n", ctl->DP, def_prio);
 470
 471                         prio = def_prio;
 472                 } else
 473                         prio = ctl->prio;
 474         }
 475
 476         prealloc = kzalloc(sizeof(*prealloc), GFP_KERNEL);
 477         sch_tree_lock(sch);
 478
 479         err = gred_change_vq(sch, ctl->DP, ctl, prio, stab, max_P, &prealloc);
 480         if (err < 0)
 481                 goto errout_locked;
 482
 483         if (gred_rio_mode(table)) {
 484                 gred_disable_wred_mode(table);
 485                 if (gred_wred_mode_check(sch))
 486                         gred_enable_wred_mode(table);
 487         }
 488
 489         err = 0;
 490
 491 errout_locked:
 492         sch_tree_unlock(sch);
 493         kfree(prealloc);
 494 errout:
 495         return err;
 496 }
 497
 498 static int gred_init(struct Qdisc *sch, struct nlattr *opt)
 499 {
 500         struct nlattr *tb[TCA_GRED_MAX + 1];
 501         int err;
 502
 503         if (opt == NULL)
 504                 return -EINVAL;
 505
 506         err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
 507         if (err < 0)
 508                 return err;
 509
 510         if (tb[TCA_GRED_PARMS] || tb[TCA_GRED_STAB])
 511                 return -EINVAL;
 512
 513         if (tb[TCA_GRED_LIMIT])
 514                 sch->limit = nla_get_u32(tb[TCA_GRED_LIMIT]);
 515         else
 516                 sch->limit = qdisc_dev(sch)->tx_queue_len
 517                              * psched_mtu(qdisc_dev(sch));
 518
 519         return gred_change_table_def(sch, tb[TCA_GRED_DPS]);
 520 }
 521
 522 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
 523 {
 524         struct gred_sched *table = qdisc_priv(sch);
 525         struct nlattr *parms, *opts = NULL;
 526         int i;
 527         u32 max_p[MAX_DPs];
 528         struct tc_gred_sopt sopt = {
 529                 .DPs    = table->DPs,
 530                 .def_DP = table->def,
 531                 .grio   = gred_rio_mode(table),
 532                 .flags  = table->red_flags,
 533         };
 534
 535         opts = nla_nest_start(skb, TCA_OPTIONS);
 536         if (opts == NULL)
 537                 goto nla_put_failure;
 538         if (nla_put(skb, TCA_GRED_DPS, sizeof(sopt), &sopt))
 539                 goto nla_put_failure;
 540
 541         for (i = 0; i < MAX_DPs; i++) {
 542                 struct gred_sched_data *q = table->tab[i];
 543
 544                 max_p[i] = q ? q->parms.max_P : 0;
 545         }
 546         if (nla_put(skb, TCA_GRED_MAX_P, sizeof(max_p), max_p))
 547                 goto nla_put_failure;
 548
 549         if (nla_put_u32(skb, TCA_GRED_LIMIT, sch->limit))
 550                 goto nla_put_failure;
 551
 552         parms = nla_nest_start(skb, TCA_GRED_PARMS);
 553         if (parms == NULL)
 554                 goto nla_put_failure;
 555
 556         for (i = 0; i < MAX_DPs; i++) {
 557                 struct gred_sched_data *q = table->tab[i];
 558                 struct tc_gred_qopt opt;
 559                 unsigned long qavg;
 560
 561                 memset(&opt, 0, sizeof(opt));
 562
 563                 if (!q) {
 564                         /* hack -- fix at some point with proper message
 565                            This is how we indicate to tc that there is no VQ
 566                            at this DP */
 567
 568                         opt.DP = MAX_DPs + i;
 569                         goto append_opt;
 570                 }
 571
 572                 opt.limit       = q->limit;
 573                 opt.DP          = q->DP;
 574                 opt.backlog     = gred_backlog(table, q, sch);
 575                 opt.prio        = q->prio;
 576                 opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
 577                 opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
 578                 opt.Wlog        = q->parms.Wlog;
 579                 opt.Plog        = q->parms.Plog;
 580                 opt.Scell_log   = q->parms.Scell_log;
 581                 opt.other       = q->stats.other;
 582                 opt.early       = q->stats.prob_drop;
 583                 opt.forced      = q->stats.forced_drop;
 584                 opt.pdrop       = q->stats.pdrop;
 585                 opt.packets     = q->packetsin;
 586                 opt.bytesin     = q->bytesin;
 587
 588                 if (gred_wred_mode(table))
 589                         gred_load_wred_set(table, q);
 590
 591                 qavg = red_calc_qavg(&q->parms, &q->vars,
 592                                      q->vars.qavg >> q->parms.Wlog);
 593                 opt.qave = qavg >> q->parms.Wlog;
 594
 595 append_opt:
 596                 if (nla_append(skb, sizeof(opt), &opt) < 0)
 597                         goto nla_put_failure;
 598         }
 599
 600         nla_nest_end(skb, parms);
 601
 602         return nla_nest_end(skb, opts);
 603
 604 nla_put_failure:
 605         nla_nest_cancel(skb, opts);
 606         return -EMSGSIZE;
 607 }
 608
 609 static void gred_destroy(struct Qdisc *sch)
 610 {
 611         struct gred_sched *table = qdisc_priv(sch);
 612         int i;
 613
 614         for (i = 0; i < table->DPs; i++) {
 615                 if (table->tab[i])
 616                         gred_destroy_vq(table->tab[i]);
 617         }
 618 }
 619
 620 static struct Qdisc_ops gred_qdisc_ops __read_mostly = {
 621         .id             =       "gred",
 622         .priv_size      =       sizeof(struct gred_sched),
 623         .enqueue        =       gred_enqueue,
 624         .dequeue        =       gred_dequeue,
 625         .peek           =       qdisc_peek_head,
 626         .drop           =       gred_drop,
 627         .init           =       gred_init,
 628         .reset          =       gred_reset,
 629         .destroy        =       gred_destroy,
 630         .change         =       gred_change,
 631         .dump           =       gred_dump,
 632         .owner          =       THIS_MODULE,
 633 };
 634
 635 static int __init gred_module_init(void)
 636 {
 637         return register_qdisc(&gred_qdisc_ops);
 638 }
 639
 640 static void __exit gred_module_exit(void)
 641 {
 642         unregister_qdisc(&gred_qdisc_ops);
 643 }
 644
 645 module_init(gred_module_init)
 646 module_exit(gred_module_exit)
 647
 648 MODULE_LICENSE("GPL");