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  *
  19  *
  20  *  For all the glorious comments look at Alexey's sch_red.c
  21  */
  22
  23 #include <linux/config.h>
  24 #include <linux/module.h>
  25 #include <asm/uaccess.h>
  26 #include <asm/system.h>
  27 #include <linux/bitops.h>
  28 #include <linux/types.h>
  29 #include <linux/kernel.h>
  30 #include <linux/sched.h>
  31 #include <linux/string.h>
  32 #include <linux/mm.h>
  33 #include <linux/socket.h>
  34 #include <linux/sockios.h>
  35 #include <linux/in.h>
  36 #include <linux/errno.h>
  37 #include <linux/interrupt.h>
  38 #include <linux/if_ether.h>
  39 #include <linux/inet.h>
  40 #include <linux/netdevice.h>
  41 #include <linux/etherdevice.h>
  42 #include <linux/notifier.h>
  43 #include <net/ip.h>
  44 #include <net/route.h>
  45 #include <linux/skbuff.h>
  46 #include <net/sock.h>
  47 #include <net/pkt_sched.h>
  48 #include <net/red.h>
  49
  50 #if 1 /* control */
  51 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
  52 #else
  53 #define DPRINTK(format,args...)
  54 #endif
  55
  56 #if 0 /* data */
  57 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
  58 #else
  59 #define D2PRINTK(format,args...)
  60 #endif
  61
  62 #define GRED_DEF_PRIO (MAX_DPs / 2)
  63 #define GRED_VQ_MASK (MAX_DPs - 1)
  64
  65 struct gred_sched_data;
  66 struct gred_sched;
  67
  68 struct gred_sched_data
  69 {
  70         u32             limit;          /* HARD maximal queue length    */
  71         u32             DP;             /* the drop pramaters */
  72         u32             bytesin;        /* bytes seen on virtualQ so far*/
  73         u32             packetsin;      /* packets seen on virtualQ so far*/
  74         u32             backlog;        /* bytes on the virtualQ */
  75         u8              prio;        /* the prio of this vq */
  76
  77         struct red_parms parms;
  78         struct red_stats stats;
  79 };
  80
  81 enum {
  82         GRED_WRED_MODE = 1,
  83         GRED_RIO_MODE,
  84 };
  85
  86 struct gred_sched
  87 {
  88         struct gred_sched_data *tab[MAX_DPs];
  89         unsigned long   flags;
  90         u32             DPs;
  91         u32             def;
  92 };
  93
  94 static inline int gred_wred_mode(struct gred_sched *table)
  95 {
  96         return test_bit(GRED_WRED_MODE, &table->flags);
  97 }
  98
  99 static inline void gred_enable_wred_mode(struct gred_sched *table)
 100 {
 101         __set_bit(GRED_WRED_MODE, &table->flags);
 102 }
 103
 104 static inline void gred_disable_wred_mode(struct gred_sched *table)
 105 {
 106         __clear_bit(GRED_WRED_MODE, &table->flags);
 107 }
 108
 109 static inline int gred_rio_mode(struct gred_sched *table)
 110 {
 111         return test_bit(GRED_RIO_MODE, &table->flags);
 112 }
 113
 114 static inline void gred_enable_rio_mode(struct gred_sched *table)
 115 {
 116         __set_bit(GRED_RIO_MODE, &table->flags);
 117 }
 118
 119 static inline void gred_disable_rio_mode(struct gred_sched *table)
 120 {
 121         __clear_bit(GRED_RIO_MODE, &table->flags);
 122 }
 123
 124 static inline int gred_wred_mode_check(struct Qdisc *sch)
 125 {
 126         struct gred_sched *table = qdisc_priv(sch);
 127         int i;
 128
 129         /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
 130         for (i = 0; i < table->DPs; i++) {
 131                 struct gred_sched_data *q = table->tab[i];
 132                 int n;
 133
 134                 if (q == NULL)
 135                         continue;
 136
 137                 for (n = 0; n < table->DPs; n++)
 138                         if (table->tab[n] && table->tab[n] != q &&
 139                             table->tab[n]->prio == q->prio)
 140                                 return 1;
 141         }
 142
 143         return 0;
 144 }
 145
 146 static inline unsigned int gred_backlog(struct gred_sched *table,
 147                                         struct gred_sched_data *q,
 148                                         struct Qdisc *sch)
 149 {
 150         if (gred_wred_mode(table))
 151                 return sch->qstats.backlog;
 152         else
 153                 return q->backlog;
 154 }
 155
 156 static inline u16 tc_index_to_dp(struct sk_buff *skb)
 157 {
 158         return skb->tc_index & GRED_VQ_MASK;
 159 }
 160
 161 static int
 162 gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 163 {
 164         struct gred_sched_data *q=NULL;
 165         struct gred_sched *t= qdisc_priv(sch);
 166         unsigned long qavg = 0;
 167         int i=0;
 168         u16 dp = tc_index_to_dp(skb);
 169
 170         if (dp >= t->DPs  || (q = t->tab[dp]) == NULL) {
 171                 dp = t->def;
 172
 173                 if ((q = t->tab[dp]) == NULL) {
 174                         /* Pass through packets not assigned to a DP
 175                          * if no default DP has been configured. This
 176                          * allows for DP flows to be left untouched.
 177                          */
 178                         if (skb_queue_len(&sch->q) < sch->dev->tx_queue_len)
 179                                 return qdisc_enqueue_tail(skb, sch);
 180                         else
 181                                 goto drop;
 182                 }
 183
 184                 /* fix tc_index? --could be controvesial but needed for
 185                    requeueing */
 186                 skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
 187         }
 188
 189         /* sum up all the qaves of prios <= to ours to get the new qave*/
 190         if (!gred_wred_mode(t) && gred_rio_mode(t)) {
 191                 for (i=0;i<t->DPs;i++) {
 192                         if ((!t->tab[i]) || (i==q->DP))
 193                                 continue;
 194
 195                         if (t->tab[i]->prio < q->prio &&
 196                             !red_is_idling(&t->tab[i]->parms))
 197                                 qavg +=t->tab[i]->parms.qavg;
 198                 }
 199
 200         }
 201
 202         q->packetsin++;
 203         q->bytesin+=skb->len;
 204
 205         if (gred_wred_mode(t)) {
 206                 qavg = 0;
 207                 q->parms.qavg = t->tab[t->def]->parms.qavg;
 208                 q->parms.qidlestart = t->tab[t->def]->parms.qidlestart;
 209         }
 210
 211         q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
 212
 213         if (red_is_idling(&q->parms))
 214                 red_end_of_idle_period(&q->parms);
 215
 216         if (gred_wred_mode(t))
 217                 t->tab[t->def]->parms.qavg = q->parms.qavg;
 218
 219         switch (red_action(&q->parms, q->parms.qavg + qavg)) {
 220                 case RED_DONT_MARK:
 221                         break;
 222
 223                 case RED_PROB_MARK:
 224                         sch->qstats.overlimits++;
 225                         q->stats.prob_drop++;
 226                         goto congestion_drop;
 227
 228                 case RED_HARD_MARK:
 229                         sch->qstats.overlimits++;
 230                         q->stats.forced_drop++;
 231                         goto congestion_drop;
 232         }
 233
 234         if (q->backlog + skb->len <= q->limit) {
 235                 q->backlog += skb->len;
 236                 return qdisc_enqueue_tail(skb, sch);
 237         }
 238
 239         q->stats.pdrop++;
 240 drop:
 241         return qdisc_drop(skb, sch);
 242
 243 congestion_drop:
 244         qdisc_drop(skb, sch);
 245         return NET_XMIT_CN;
 246 }
 247
 248 static int
 249 gred_requeue(struct sk_buff *skb, struct Qdisc* sch)
 250 {
 251         struct gred_sched *t = qdisc_priv(sch);
 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 VQ 0x%x "
 258                                "for requeue, screwing up backlog.\n",
 259                                tc_index_to_dp(skb));
 260         } else {
 261                 if (red_is_idling(&q->parms))
 262                         red_end_of_idle_period(&q->parms);
 263                 q->backlog += skb->len;
 264         }
 265
 266         return qdisc_requeue(skb, sch);
 267 }
 268
 269 static struct sk_buff *
 270 gred_dequeue(struct Qdisc* sch)
 271 {
 272         struct sk_buff *skb;
 273         struct gred_sched_data *q;
 274         struct gred_sched *t= qdisc_priv(sch);
 275
 276         skb = qdisc_dequeue_head(sch);
 277
 278         if (skb) {
 279                 u16 dp = tc_index_to_dp(skb);
 280
 281                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 282                         if (net_ratelimit())
 283                                 printk(KERN_WARNING "GRED: Unable to relocate "
 284                                        "VQ 0x%x after dequeue, screwing up "
 285                                        "backlog.\n", tc_index_to_dp(skb));
 286                 } else {
 287                         q->backlog -= skb->len;
 288
 289                         if (!q->backlog && !gred_wred_mode(t))
 290                                 red_start_of_idle_period(&q->parms);
 291                 }
 292
 293                 return skb;
 294         }
 295
 296         if (gred_wred_mode(t)) {
 297                         q= t->tab[t->def];
 298                         if (!q)
 299                                 D2PRINTK("no default VQ set: Results will be "
 300                                        "screwed up\n");
 301                         else
 302                                 red_start_of_idle_period(&q->parms);
 303         }
 304
 305         return NULL;
 306 }
 307
 308 static unsigned int gred_drop(struct Qdisc* sch)
 309 {
 310         struct sk_buff *skb;
 311
 312         struct gred_sched_data *q;
 313         struct gred_sched *t= qdisc_priv(sch);
 314
 315         skb = qdisc_dequeue_tail(sch);
 316         if (skb) {
 317                 unsigned int len = skb->len;
 318                 u16 dp = tc_index_to_dp(skb);
 319
 320                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
 321                         if (net_ratelimit())
 322                                 printk(KERN_WARNING "GRED: Unable to relocate "
 323                                        "VQ 0x%x while dropping, screwing up "
 324                                        "backlog.\n", tc_index_to_dp(skb));
 325                 } else {
 326                         q->backlog -= len;
 327                         q->stats.other++;
 328
 329                         if (!q->backlog && !gred_wred_mode(t))
 330                                 red_start_of_idle_period(&q->parms);
 331                 }
 332
 333                 qdisc_drop(skb, sch);
 334                 return len;
 335         }
 336
 337         q=t->tab[t->def];
 338         if (!q) {
 339                 D2PRINTK("no default VQ set: Results might be screwed up\n");
 340                 return 0;
 341         }
 342
 343         red_start_of_idle_period(&q->parms);
 344         return 0;
 345
 346 }
 347
 348 static void gred_reset(struct Qdisc* sch)
 349 {
 350         int i;
 351         struct gred_sched_data *q;
 352         struct gred_sched *t= qdisc_priv(sch);
 353
 354         qdisc_reset_queue(sch);
 355
 356         for (i=0;i<t->DPs;i++) {
 357                 q= t->tab[i];
 358                 if (!q)
 359                         continue;
 360                 red_restart(&q->parms);
 361                 q->backlog = 0;
 362         }
 363 }
 364
 365 static inline void gred_destroy_vq(struct gred_sched_data *q)
 366 {
 367         kfree(q);
 368 }
 369
 370 static inline int gred_change_table_def(struct Qdisc *sch, struct rtattr *dps)
 371 {
 372         struct gred_sched *table = qdisc_priv(sch);
 373         struct tc_gred_sopt *sopt;
 374         int i;
 375
 376         if (dps == NULL || RTA_PAYLOAD(dps) < sizeof(*sopt))
 377                 return -EINVAL;
 378
 379         sopt = RTA_DATA(dps);
 380
 381         if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
 382                 return -EINVAL;
 383
 384         sch_tree_lock(sch);
 385         table->DPs = sopt->DPs;
 386         table->def = sopt->def_DP;
 387
 388         /*
 389          * Every entry point to GRED is synchronized with the above code
 390          * and the DP is checked against DPs, i.e. shadowed VQs can no
 391          * longer be found so we can unlock right here.
 392          */
 393         sch_tree_unlock(sch);
 394
 395         if (sopt->grio) {
 396                 gred_enable_rio_mode(table);
 397                 gred_disable_wred_mode(table);
 398                 if (gred_wred_mode_check(sch))
 399                         gred_enable_wred_mode(table);
 400         } else {
 401                 gred_disable_rio_mode(table);
 402                 gred_disable_wred_mode(table);
 403         }
 404
 405         for (i = table->DPs; i < MAX_DPs; i++) {
 406                 if (table->tab[i]) {
 407                         printk(KERN_WARNING "GRED: Warning: Destroying "
 408                                "shadowed VQ 0x%x\n", i);
 409                         gred_destroy_vq(table->tab[i]);
 410                         table->tab[i] = NULL;
 411                 }
 412         }
 413
 414         return 0;
 415 }
 416
 417 static inline int gred_change_vq(struct Qdisc *sch, int dp,
 418                                  struct tc_gred_qopt *ctl, int prio, u8 *stab)
 419 {
 420         struct gred_sched *table = qdisc_priv(sch);
 421         struct gred_sched_data *q;
 422
 423         if (table->tab[dp] == NULL) {
 424                 table->tab[dp] = kmalloc(sizeof(*q), GFP_KERNEL);
 425                 if (table->tab[dp] == NULL)
 426                         return -ENOMEM;
 427                 memset(table->tab[dp], 0, sizeof(*q));
 428         }
 429
 430         q = table->tab[dp];
 431         q->DP = dp;
 432         q->prio = prio;
 433         q->limit = ctl->limit;
 434
 435         if (q->backlog == 0)
 436                 red_end_of_idle_period(&q->parms);
 437
 438         red_set_parms(&q->parms,
 439                       ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
 440                       ctl->Scell_log, stab);
 441
 442         return 0;
 443 }
 444
 445 static int gred_change(struct Qdisc *sch, struct rtattr *opt)
 446 {
 447         struct gred_sched *table = qdisc_priv(sch);
 448         struct tc_gred_qopt *ctl;
 449         struct rtattr *tb[TCA_GRED_MAX];
 450         int err = -EINVAL, prio = GRED_DEF_PRIO;
 451         u8 *stab;
 452
 453         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
 454                 return -EINVAL;
 455
 456         if (tb[TCA_GRED_PARMS-1] == NULL && tb[TCA_GRED_STAB-1] == NULL)
 457                 return gred_change_table_def(sch, opt);
 458
 459         if (tb[TCA_GRED_PARMS-1] == NULL ||
 460             RTA_PAYLOAD(tb[TCA_GRED_PARMS-1]) < sizeof(*ctl) ||
 461             tb[TCA_GRED_STAB-1] == NULL ||
 462             RTA_PAYLOAD(tb[TCA_GRED_STAB-1]) < 256)
 463                 return -EINVAL;
 464
 465         ctl = RTA_DATA(tb[TCA_GRED_PARMS-1]);
 466         stab = RTA_DATA(tb[TCA_GRED_STAB-1]);
 467
 468         if (ctl->DP >= table->DPs)
 469                 goto errout;
 470
 471         if (gred_rio_mode(table)) {
 472                 if (ctl->prio == 0) {
 473                         int def_prio = GRED_DEF_PRIO;
 474
 475                         if (table->tab[table->def])
 476                                 def_prio = table->tab[table->def]->prio;
 477
 478                         printk(KERN_DEBUG "GRED: DP %u does not have a prio "
 479                                "setting default to %d\n", ctl->DP, def_prio);
 480
 481                         prio = def_prio;
 482                 } else
 483                         prio = ctl->prio;
 484         }
 485
 486         sch_tree_lock(sch);
 487
 488         err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
 489         if (err < 0)
 490                 goto errout_locked;
 491
 492         if (table->tab[table->def] == NULL) {
 493                 if (gred_rio_mode(table))
 494                         prio = table->tab[ctl->DP]->prio;
 495
 496                 err = gred_change_vq(sch, table->def, ctl, prio, stab);
 497                 if (err < 0)
 498                         goto errout_locked;
 499         }
 500
 501         if (gred_rio_mode(table)) {
 502                 gred_disable_wred_mode(table);
 503                 if (gred_wred_mode_check(sch))
 504                         gred_enable_wred_mode(table);
 505         }
 506
 507         err = 0;
 508
 509 errout_locked:
 510         sch_tree_unlock(sch);
 511 errout:
 512         return err;
 513 }
 514
 515 static int gred_init(struct Qdisc *sch, struct rtattr *opt)
 516 {
 517         struct rtattr *tb[TCA_GRED_MAX];
 518
 519         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
 520                 return -EINVAL;
 521
 522         if (tb[TCA_GRED_PARMS-1] || tb[TCA_GRED_STAB-1])
 523                 return -EINVAL;
 524
 525         return gred_change_table_def(sch, tb[TCA_GRED_DPS-1]);
 526 }
 527
 528 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
 529 {
 530         struct gred_sched *table = qdisc_priv(sch);
 531         struct rtattr *parms, *opts = NULL;
 532         int i;
 533         struct tc_gred_sopt sopt = {
 534                 .DPs    = table->DPs,
 535                 .def_DP = table->def,
 536                 .grio   = gred_rio_mode(table),
 537         };
 538
 539         opts = RTA_NEST(skb, TCA_OPTIONS);
 540         RTA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
 541         parms = RTA_NEST(skb, TCA_GRED_PARMS);
 542
 543         for (i = 0; i < MAX_DPs; i++) {
 544                 struct gred_sched_data *q = table->tab[i];
 545                 struct tc_gred_qopt opt;
 546
 547                 memset(&opt, 0, sizeof(opt));
 548
 549                 if (!q) {
 550                         /* hack -- fix at some point with proper message
 551                            This is how we indicate to tc that there is no VQ
 552                            at this DP */
 553
 554                         opt.DP = MAX_DPs + i;
 555                         goto append_opt;
 556                 }
 557
 558                 opt.limit       = q->limit;
 559                 opt.DP          = q->DP;
 560                 opt.backlog     = q->backlog;
 561                 opt.prio        = q->prio;
 562                 opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
 563                 opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
 564                 opt.Wlog        = q->parms.Wlog;
 565                 opt.Plog        = q->parms.Plog;
 566                 opt.Scell_log   = q->parms.Scell_log;
 567                 opt.other       = q->stats.other;
 568                 opt.early       = q->stats.prob_drop;
 569                 opt.forced      = q->stats.forced_drop;
 570                 opt.pdrop       = q->stats.pdrop;
 571                 opt.packets     = q->packetsin;
 572                 opt.bytesin     = q->bytesin;
 573
 574                 if (gred_wred_mode(table)) {
 575                         q->parms.qidlestart =
 576                                 table->tab[table->def]->parms.qidlestart;
 577                         q->parms.qavg = table->tab[table->def]->parms.qavg;
 578                 }
 579
 580                 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
 581
 582 append_opt:
 583                 RTA_APPEND(skb, sizeof(opt), &opt);
 584         }
 585
 586         RTA_NEST_END(skb, parms);
 587
 588         return RTA_NEST_END(skb, opts);
 589
 590 rtattr_failure:
 591         return RTA_NEST_CANCEL(skb, opts);
 592 }
 593
 594 static void gred_destroy(struct Qdisc *sch)
 595 {
 596         struct gred_sched *table = qdisc_priv(sch);
 597         int i;
 598
 599         for (i = 0;i < table->DPs; i++) {
 600                 if (table->tab[i])
 601                         gred_destroy_vq(table->tab[i]);
 602         }
 603 }
 604
 605 static struct Qdisc_ops gred_qdisc_ops = {
 606         .next           =       NULL,
 607         .cl_ops         =       NULL,
 608         .id             =       "gred",
 609         .priv_size      =       sizeof(struct gred_sched),
 610         .enqueue        =       gred_enqueue,
 611         .dequeue        =       gred_dequeue,
 612         .requeue        =       gred_requeue,
 613         .drop           =       gred_drop,
 614         .init           =       gred_init,
 615         .reset          =       gred_reset,
 616         .destroy        =       gred_destroy,
 617         .change         =       gred_change,
 618         .dump           =       gred_dump,
 619         .owner          =       THIS_MODULE,
 620 };
 621
 622 static int __init gred_module_init(void)
 623 {
 624         return register_qdisc(&gred_qdisc_ops);
 625 }
 626 static void __exit gred_module_exit(void)
 627 {
 628         unregister_qdisc(&gred_qdisc_ops);
 629 }
 630 module_init(gred_module_init)
 631 module_exit(gred_module_exit)
 632 MODULE_LICENSE("GPL");