Pileus Git - ~andy/linux/blob - net/sched/sch_prio.c

   1 /*
   2  * net/sched/sch_prio.c Simple 3-band priority "scheduler".
   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  * Fixes:       19990609: J Hadi Salim <hadi@nortelnetworks.com>:
  11  *              Init --  EINVAL when opt undefined
  12  */
  13
  14 #include <linux/module.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
  23
  24 struct prio_sched_data
  25 {
  26         int bands;
  27         int curband; /* for round-robin */
  28         struct tcf_proto *filter_list;
  29         u8  prio2band[TC_PRIO_MAX+1];
  30         struct Qdisc *queues[TCQ_PRIO_BANDS];
  31         int mq;
  32 };
  33
  34
  35 static struct Qdisc *
  36 prio_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
  37 {
  38         struct prio_sched_data *q = qdisc_priv(sch);
  39         u32 band = skb->priority;
  40         struct tcf_result res;
  41         int err;
  42
  43         *qerr = NET_XMIT_BYPASS;
  44         if (TC_H_MAJ(skb->priority) != sch->handle) {
  45                 err = tc_classify(skb, q->filter_list, &res);
  46 #ifdef CONFIG_NET_CLS_ACT
  47                 switch (tc_classify(skb, q->filter_list, &res)) {
  48                 case TC_ACT_STOLEN:
  49                 case TC_ACT_QUEUED:
  50                         *qerr = NET_XMIT_SUCCESS;
  51                 case TC_ACT_SHOT:
  52                         return NULL;
  53                 }
  54 #endif
  55                 if (!q->filter_list || err < 0) {
  56                         if (TC_H_MAJ(band))
  57                                 band = 0;
  58                         band = q->prio2band[band&TC_PRIO_MAX];
  59                         goto out;
  60                 }
  61                 band = res.classid;
  62         }
  63         band = TC_H_MIN(band) - 1;
  64         if (band >= q->bands)
  65                 band = q->prio2band[0];
  66 out:
  67         if (q->mq)
  68                 skb_set_queue_mapping(skb, band);
  69         return q->queues[band];
  70 }
  71
  72 static int
  73 prio_enqueue(struct sk_buff *skb, struct Qdisc *sch)
  74 {
  75         struct Qdisc *qdisc;
  76         int ret;
  77
  78         qdisc = prio_classify(skb, sch, &ret);
  79 #ifdef CONFIG_NET_CLS_ACT
  80         if (qdisc == NULL) {
  81
  82                 if (ret == NET_XMIT_BYPASS)
  83                         sch->qstats.drops++;
  84                 kfree_skb(skb);
  85                 return ret;
  86         }
  87 #endif
  88
  89         if ((ret = qdisc->enqueue(skb, qdisc)) == NET_XMIT_SUCCESS) {
  90                 sch->bstats.bytes += skb->len;
  91                 sch->bstats.packets++;
  92                 sch->q.qlen++;
  93                 return NET_XMIT_SUCCESS;
  94         }
  95         sch->qstats.drops++;
  96         return ret;
  97 }
  98
  99
 100 static int
 101 prio_requeue(struct sk_buff *skb, struct Qdisc* sch)
 102 {
 103         struct Qdisc *qdisc;
 104         int ret;
 105
 106         qdisc = prio_classify(skb, sch, &ret);
 107 #ifdef CONFIG_NET_CLS_ACT
 108         if (qdisc == NULL) {
 109                 if (ret == NET_XMIT_BYPASS)
 110                         sch->qstats.drops++;
 111                 kfree_skb(skb);
 112                 return ret;
 113         }
 114 #endif
 115
 116         if ((ret = qdisc->ops->requeue(skb, qdisc)) == NET_XMIT_SUCCESS) {
 117                 sch->q.qlen++;
 118                 sch->qstats.requeues++;
 119                 return 0;
 120         }
 121         sch->qstats.drops++;
 122         return NET_XMIT_DROP;
 123 }
 124
 125
 126 static struct sk_buff *
 127 prio_dequeue(struct Qdisc* sch)
 128 {
 129         struct sk_buff *skb;
 130         struct prio_sched_data *q = qdisc_priv(sch);
 131         int prio;
 132         struct Qdisc *qdisc;
 133
 134         for (prio = 0; prio < q->bands; prio++) {
 135                 /* Check if the target subqueue is available before
 136                  * pulling an skb.  This way we avoid excessive requeues
 137                  * for slower queues.
 138                  */
 139                 if (!netif_subqueue_stopped(sch->dev, (q->mq ? prio : 0))) {
 140                         qdisc = q->queues[prio];
 141                         skb = qdisc->dequeue(qdisc);
 142                         if (skb) {
 143                                 sch->q.qlen--;
 144                                 return skb;
 145                         }
 146                 }
 147         }
 148         return NULL;
 149
 150 }
 151
 152 static struct sk_buff *rr_dequeue(struct Qdisc* sch)
 153 {
 154         struct sk_buff *skb;
 155         struct prio_sched_data *q = qdisc_priv(sch);
 156         struct Qdisc *qdisc;
 157         int bandcount;
 158
 159         /* Only take one pass through the queues.  If nothing is available,
 160          * return nothing.
 161          */
 162         for (bandcount = 0; bandcount < q->bands; bandcount++) {
 163                 /* Check if the target subqueue is available before
 164                  * pulling an skb.  This way we avoid excessive requeues
 165                  * for slower queues.  If the queue is stopped, try the
 166                  * next queue.
 167                  */
 168                 if (!netif_subqueue_stopped(sch->dev,
 169                                             (q->mq ? q->curband : 0))) {
 170                         qdisc = q->queues[q->curband];
 171                         skb = qdisc->dequeue(qdisc);
 172                         if (skb) {
 173                                 sch->q.qlen--;
 174                                 q->curband++;
 175                                 if (q->curband >= q->bands)
 176                                         q->curband = 0;
 177                                 return skb;
 178                         }
 179                 }
 180                 q->curband++;
 181                 if (q->curband >= q->bands)
 182                         q->curband = 0;
 183         }
 184         return NULL;
 185 }
 186
 187 static unsigned int prio_drop(struct Qdisc* sch)
 188 {
 189         struct prio_sched_data *q = qdisc_priv(sch);
 190         int prio;
 191         unsigned int len;
 192         struct Qdisc *qdisc;
 193
 194         for (prio = q->bands-1; prio >= 0; prio--) {
 195                 qdisc = q->queues[prio];
 196                 if (qdisc->ops->drop && (len = qdisc->ops->drop(qdisc)) != 0) {
 197                         sch->q.qlen--;
 198                         return len;
 199                 }
 200         }
 201         return 0;
 202 }
 203
 204
 205 static void
 206 prio_reset(struct Qdisc* sch)
 207 {
 208         int prio;
 209         struct prio_sched_data *q = qdisc_priv(sch);
 210
 211         for (prio=0; prio<q->bands; prio++)
 212                 qdisc_reset(q->queues[prio]);
 213         sch->q.qlen = 0;
 214 }
 215
 216 static void
 217 prio_destroy(struct Qdisc* sch)
 218 {
 219         int prio;
 220         struct prio_sched_data *q = qdisc_priv(sch);
 221
 222         tcf_destroy_chain(q->filter_list);
 223         for (prio=0; prio<q->bands; prio++)
 224                 qdisc_destroy(q->queues[prio]);
 225 }
 226
 227 static int prio_tune(struct Qdisc *sch, struct rtattr *opt)
 228 {
 229         struct prio_sched_data *q = qdisc_priv(sch);
 230         struct tc_prio_qopt *qopt;
 231         struct rtattr *tb[TCA_PRIO_MAX];
 232         int i;
 233
 234         if (rtattr_parse_nested_compat(tb, TCA_PRIO_MAX, opt, qopt,
 235                                        sizeof(*qopt)))
 236                 return -EINVAL;
 237         q->bands = qopt->bands;
 238         /* If we're multiqueue, make sure the number of incoming bands
 239          * matches the number of queues on the device we're associating with.
 240          * If the number of bands requested is zero, then set q->bands to
 241          * dev->egress_subqueue_count.
 242          */
 243         q->mq = RTA_GET_FLAG(tb[TCA_PRIO_MQ - 1]);
 244         if (q->mq) {
 245                 if (sch->handle != TC_H_ROOT)
 246                         return -EINVAL;
 247                 if (netif_is_multiqueue(sch->dev)) {
 248                         if (q->bands == 0)
 249                                 q->bands = sch->dev->egress_subqueue_count;
 250                         else if (q->bands != sch->dev->egress_subqueue_count)
 251                                 return -EINVAL;
 252                 } else
 253                         return -EOPNOTSUPP;
 254         }
 255
 256         if (q->bands > TCQ_PRIO_BANDS || q->bands < 2)
 257                 return -EINVAL;
 258
 259         for (i=0; i<=TC_PRIO_MAX; i++) {
 260                 if (qopt->priomap[i] >= q->bands)
 261                         return -EINVAL;
 262         }
 263
 264         sch_tree_lock(sch);
 265         memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
 266
 267         for (i=q->bands; i<TCQ_PRIO_BANDS; i++) {
 268                 struct Qdisc *child = xchg(&q->queues[i], &noop_qdisc);
 269                 if (child != &noop_qdisc) {
 270                         qdisc_tree_decrease_qlen(child, child->q.qlen);
 271                         qdisc_destroy(child);
 272                 }
 273         }
 274         sch_tree_unlock(sch);
 275
 276         for (i=0; i<q->bands; i++) {
 277                 if (q->queues[i] == &noop_qdisc) {
 278                         struct Qdisc *child;
 279                         child = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
 280                                                   TC_H_MAKE(sch->handle, i + 1));
 281                         if (child) {
 282                                 sch_tree_lock(sch);
 283                                 child = xchg(&q->queues[i], child);
 284
 285                                 if (child != &noop_qdisc) {
 286                                         qdisc_tree_decrease_qlen(child,
 287                                                                  child->q.qlen);
 288                                         qdisc_destroy(child);
 289                                 }
 290                                 sch_tree_unlock(sch);
 291                         }
 292                 }
 293         }
 294         return 0;
 295 }
 296
 297 static int prio_init(struct Qdisc *sch, struct rtattr *opt)
 298 {
 299         struct prio_sched_data *q = qdisc_priv(sch);
 300         int i;
 301
 302         for (i=0; i<TCQ_PRIO_BANDS; i++)
 303                 q->queues[i] = &noop_qdisc;
 304
 305         if (opt == NULL) {
 306                 return -EINVAL;
 307         } else {
 308                 int err;
 309
 310                 if ((err= prio_tune(sch, opt)) != 0)
 311                         return err;
 312         }
 313         return 0;
 314 }
 315
 316 static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)
 317 {
 318         struct prio_sched_data *q = qdisc_priv(sch);
 319         unsigned char *b = skb_tail_pointer(skb);
 320         struct rtattr *nest;
 321         struct tc_prio_qopt opt;
 322
 323         opt.bands = q->bands;
 324         memcpy(&opt.priomap, q->prio2band, TC_PRIO_MAX+1);
 325
 326         nest = RTA_NEST_COMPAT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 327         if (q->mq)
 328                 RTA_PUT_FLAG(skb, TCA_PRIO_MQ);
 329         RTA_NEST_COMPAT_END(skb, nest);
 330
 331         return skb->len;
 332
 333 rtattr_failure:
 334         nlmsg_trim(skb, b);
 335         return -1;
 336 }
 337
 338 static int prio_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 339                       struct Qdisc **old)
 340 {
 341         struct prio_sched_data *q = qdisc_priv(sch);
 342         unsigned long band = arg - 1;
 343
 344         if (band >= q->bands)
 345                 return -EINVAL;
 346
 347         if (new == NULL)
 348                 new = &noop_qdisc;
 349
 350         sch_tree_lock(sch);
 351         *old = q->queues[band];
 352         q->queues[band] = new;
 353         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
 354         qdisc_reset(*old);
 355         sch_tree_unlock(sch);
 356
 357         return 0;
 358 }
 359
 360 static struct Qdisc *
 361 prio_leaf(struct Qdisc *sch, unsigned long arg)
 362 {
 363         struct prio_sched_data *q = qdisc_priv(sch);
 364         unsigned long band = arg - 1;
 365
 366         if (band >= q->bands)
 367                 return NULL;
 368
 369         return q->queues[band];
 370 }
 371
 372 static unsigned long prio_get(struct Qdisc *sch, u32 classid)
 373 {
 374         struct prio_sched_data *q = qdisc_priv(sch);
 375         unsigned long band = TC_H_MIN(classid);
 376
 377         if (band - 1 >= q->bands)
 378                 return 0;
 379         return band;
 380 }
 381
 382 static unsigned long prio_bind(struct Qdisc *sch, unsigned long parent, u32 classid)
 383 {
 384         return prio_get(sch, classid);
 385 }
 386
 387
 388 static void prio_put(struct Qdisc *q, unsigned long cl)
 389 {
 390         return;
 391 }
 392
 393 static int prio_change(struct Qdisc *sch, u32 handle, u32 parent, struct rtattr **tca, unsigned long *arg)
 394 {
 395         unsigned long cl = *arg;
 396         struct prio_sched_data *q = qdisc_priv(sch);
 397
 398         if (cl - 1 > q->bands)
 399                 return -ENOENT;
 400         return 0;
 401 }
 402
 403 static int prio_delete(struct Qdisc *sch, unsigned long cl)
 404 {
 405         struct prio_sched_data *q = qdisc_priv(sch);
 406         if (cl - 1 > q->bands)
 407                 return -ENOENT;
 408         return 0;
 409 }
 410
 411
 412 static int prio_dump_class(struct Qdisc *sch, unsigned long cl, struct sk_buff *skb,
 413                            struct tcmsg *tcm)
 414 {
 415         struct prio_sched_data *q = qdisc_priv(sch);
 416
 417         if (cl - 1 > q->bands)
 418                 return -ENOENT;
 419         tcm->tcm_handle |= TC_H_MIN(cl);
 420         if (q->queues[cl-1])
 421                 tcm->tcm_info = q->queues[cl-1]->handle;
 422         return 0;
 423 }
 424
 425 static int prio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 426                                  struct gnet_dump *d)
 427 {
 428         struct prio_sched_data *q = qdisc_priv(sch);
 429         struct Qdisc *cl_q;
 430
 431         cl_q = q->queues[cl - 1];
 432         if (gnet_stats_copy_basic(d, &cl_q->bstats) < 0 ||
 433             gnet_stats_copy_queue(d, &cl_q->qstats) < 0)
 434                 return -1;
 435
 436         return 0;
 437 }
 438
 439 static void prio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 440 {
 441         struct prio_sched_data *q = qdisc_priv(sch);
 442         int prio;
 443
 444         if (arg->stop)
 445                 return;
 446
 447         for (prio = 0; prio < q->bands; prio++) {
 448                 if (arg->count < arg->skip) {
 449                         arg->count++;
 450                         continue;
 451                 }
 452                 if (arg->fn(sch, prio+1, arg) < 0) {
 453                         arg->stop = 1;
 454                         break;
 455                 }
 456                 arg->count++;
 457         }
 458 }
 459
 460 static struct tcf_proto ** prio_find_tcf(struct Qdisc *sch, unsigned long cl)
 461 {
 462         struct prio_sched_data *q = qdisc_priv(sch);
 463
 464         if (cl)
 465                 return NULL;
 466         return &q->filter_list;
 467 }
 468
 469 static struct Qdisc_class_ops prio_class_ops = {
 470         .graft          =       prio_graft,
 471         .leaf           =       prio_leaf,
 472         .get            =       prio_get,
 473         .put            =       prio_put,
 474         .change         =       prio_change,
 475         .delete         =       prio_delete,
 476         .walk           =       prio_walk,
 477         .tcf_chain      =       prio_find_tcf,
 478         .bind_tcf       =       prio_bind,
 479         .unbind_tcf     =       prio_put,
 480         .dump           =       prio_dump_class,
 481         .dump_stats     =       prio_dump_class_stats,
 482 };
 483
 484 static struct Qdisc_ops prio_qdisc_ops = {
 485         .next           =       NULL,
 486         .cl_ops         =       &prio_class_ops,
 487         .id             =       "prio",
 488         .priv_size      =       sizeof(struct prio_sched_data),
 489         .enqueue        =       prio_enqueue,
 490         .dequeue        =       prio_dequeue,
 491         .requeue        =       prio_requeue,
 492         .drop           =       prio_drop,
 493         .init           =       prio_init,
 494         .reset          =       prio_reset,
 495         .destroy        =       prio_destroy,
 496         .change         =       prio_tune,
 497         .dump           =       prio_dump,
 498         .owner          =       THIS_MODULE,
 499 };
 500
 501 static struct Qdisc_ops rr_qdisc_ops = {
 502         .next           =       NULL,
 503         .cl_ops         =       &prio_class_ops,
 504         .id             =       "rr",
 505         .priv_size      =       sizeof(struct prio_sched_data),
 506         .enqueue        =       prio_enqueue,
 507         .dequeue        =       rr_dequeue,
 508         .requeue        =       prio_requeue,
 509         .drop           =       prio_drop,
 510         .init           =       prio_init,
 511         .reset          =       prio_reset,
 512         .destroy        =       prio_destroy,
 513         .change         =       prio_tune,
 514         .dump           =       prio_dump,
 515         .owner          =       THIS_MODULE,
 516 };
 517
 518 static int __init prio_module_init(void)
 519 {
 520         int err;
 521
 522         err = register_qdisc(&prio_qdisc_ops);
 523         if (err < 0)
 524                 return err;
 525         err = register_qdisc(&rr_qdisc_ops);
 526         if (err < 0)
 527                 unregister_qdisc(&prio_qdisc_ops);
 528         return err;
 529 }
 530
 531 static void __exit prio_module_exit(void)
 532 {
 533         unregister_qdisc(&prio_qdisc_ops);
 534         unregister_qdisc(&rr_qdisc_ops);
 535 }
 536
 537 module_init(prio_module_init)
 538 module_exit(prio_module_exit)
 539
 540 MODULE_LICENSE("GPL");
 541 MODULE_ALIAS("sch_rr");