net/sched/sch_drr.c

   1 /*
   2  * net/sched/sch_drr.c         Deficit Round Robin scheduler
   3  *
   4  * Copyright (c) 2008 Patrick McHardy <kaber@trash.net>
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * version 2 as published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/module.h>
  12 #include <linux/init.h>
  13 #include <linux/errno.h>
  14 #include <linux/netdevice.h>
  15 #include <linux/pkt_sched.h>
  16 #include <net/sch_generic.h>
  17 #include <net/pkt_sched.h>
  18 #include <net/pkt_cls.h>
  19
  20 struct drr_class {
  21         struct Qdisc_class_common       common;
  22         unsigned int                    refcnt;
  23         unsigned int                    filter_cnt;
  24
  25         struct gnet_stats_basic_packed          bstats;
  26         struct gnet_stats_queue         qstats;
  27         struct gnet_stats_rate_est      rate_est;
  28         struct list_head                alist;
  29         struct Qdisc                    *qdisc;
  30
  31         u32                             quantum;
  32         u32                             deficit;
  33 };
  34
  35 struct drr_sched {
  36         struct list_head                active;
  37         struct tcf_proto                *filter_list;
  38         struct Qdisc_class_hash         clhash;
  39 };
  40
  41 static struct drr_class *drr_find_class(struct Qdisc *sch, u32 classid)
  42 {
  43         struct drr_sched *q = qdisc_priv(sch);
  44         struct Qdisc_class_common *clc;
  45
  46         clc = qdisc_class_find(&q->clhash, classid);
  47         if (clc == NULL)
  48                 return NULL;
  49         return container_of(clc, struct drr_class, common);
  50 }
  51
  52 static void drr_purge_queue(struct drr_class *cl)
  53 {
  54         unsigned int len = cl->qdisc->q.qlen;
  55
  56         qdisc_reset(cl->qdisc);
  57         qdisc_tree_decrease_qlen(cl->qdisc, len);
  58 }
  59
  60 static const struct nla_policy drr_policy[TCA_DRR_MAX + 1] = {
  61         [TCA_DRR_QUANTUM]       = { .type = NLA_U32 },
  62 };
  63
  64 static int drr_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
  65                             struct nlattr **tca, unsigned long *arg)
  66 {
  67         struct drr_sched *q = qdisc_priv(sch);
  68         struct drr_class *cl = (struct drr_class *)*arg;
  69         struct nlattr *opt = tca[TCA_OPTIONS];
  70         struct nlattr *tb[TCA_DRR_MAX + 1];
  71         u32 quantum;
  72         int err;
  73
  74         if (!opt)
  75                 return -EINVAL;
  76
  77         err = nla_parse_nested(tb, TCA_DRR_MAX, opt, drr_policy);
  78         if (err < 0)
  79                 return err;
  80
  81         if (tb[TCA_DRR_QUANTUM]) {
  82                 quantum = nla_get_u32(tb[TCA_DRR_QUANTUM]);
  83                 if (quantum == 0)
  84                         return -EINVAL;
  85         } else
  86                 quantum = psched_mtu(qdisc_dev(sch));
  87
  88         if (cl != NULL) {
  89                 if (tca[TCA_RATE]) {
  90                         err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
  91                                                     qdisc_root_sleeping_lock(sch),
  92                                                     tca[TCA_RATE]);
  93                         if (err)
  94                                 return err;
  95                 }
  96
  97                 sch_tree_lock(sch);
  98                 if (tb[TCA_DRR_QUANTUM])
  99                         cl->quantum = quantum;
 100                 sch_tree_unlock(sch);
 101
 102                 return 0;
 103         }
 104
 105         cl = kzalloc(sizeof(struct drr_class), GFP_KERNEL);
 106         if (cl == NULL)
 107                 return -ENOBUFS;
 108
 109         cl->refcnt         = 1;
 110         cl->common.classid = classid;
 111         cl->quantum        = quantum;
 112         cl->qdisc          = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
 113                                                &pfifo_qdisc_ops, classid);
 114         if (cl->qdisc == NULL)
 115                 cl->qdisc = &noop_qdisc;
 116
 117         if (tca[TCA_RATE]) {
 118                 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
 119                                             qdisc_root_sleeping_lock(sch),
 120                                             tca[TCA_RATE]);
 121                 if (err) {
 122                         qdisc_destroy(cl->qdisc);
 123                         kfree(cl);
 124                         return err;
 125                 }
 126         }
 127
 128         sch_tree_lock(sch);
 129         qdisc_class_hash_insert(&q->clhash, &cl->common);
 130         sch_tree_unlock(sch);
 131
 132         qdisc_class_hash_grow(sch, &q->clhash);
 133
 134         *arg = (unsigned long)cl;
 135         return 0;
 136 }
 137
 138 static void drr_destroy_class(struct Qdisc *sch, struct drr_class *cl)
 139 {
 140         gen_kill_estimator(&cl->bstats, &cl->rate_est);
 141         qdisc_destroy(cl->qdisc);
 142         kfree(cl);
 143 }
 144
 145 static int drr_delete_class(struct Qdisc *sch, unsigned long arg)
 146 {
 147         struct drr_sched *q = qdisc_priv(sch);
 148         struct drr_class *cl = (struct drr_class *)arg;
 149
 150         if (cl->filter_cnt > 0)
 151                 return -EBUSY;
 152
 153         sch_tree_lock(sch);
 154
 155         drr_purge_queue(cl);
 156         qdisc_class_hash_remove(&q->clhash, &cl->common);
 157
 158         BUG_ON(--cl->refcnt == 0);
 159         /*
 160          * This shouldn't happen: we "hold" one cops->get() when called
 161          * from tc_ctl_tclass; the destroy method is done from cops->put().
 162          */
 163
 164         sch_tree_unlock(sch);
 165         return 0;
 166 }
 167
 168 static unsigned long drr_get_class(struct Qdisc *sch, u32 classid)
 169 {
 170         struct drr_class *cl = drr_find_class(sch, classid);
 171
 172         if (cl != NULL)
 173                 cl->refcnt++;
 174
 175         return (unsigned long)cl;
 176 }
 177
 178 static void drr_put_class(struct Qdisc *sch, unsigned long arg)
 179 {
 180         struct drr_class *cl = (struct drr_class *)arg;
 181
 182         if (--cl->refcnt == 0)
 183                 drr_destroy_class(sch, cl);
 184 }
 185
 186 static struct tcf_proto **drr_tcf_chain(struct Qdisc *sch, unsigned long cl)
 187 {
 188         struct drr_sched *q = qdisc_priv(sch);
 189
 190         if (cl)
 191                 return NULL;
 192
 193         return &q->filter_list;
 194 }
 195
 196 static unsigned long drr_bind_tcf(struct Qdisc *sch, unsigned long parent,
 197                                   u32 classid)
 198 {
 199         struct drr_class *cl = drr_find_class(sch, classid);
 200
 201         if (cl != NULL)
 202                 cl->filter_cnt++;
 203
 204         return (unsigned long)cl;
 205 }
 206
 207 static void drr_unbind_tcf(struct Qdisc *sch, unsigned long arg)
 208 {
 209         struct drr_class *cl = (struct drr_class *)arg;
 210
 211         cl->filter_cnt--;
 212 }
 213
 214 static int drr_graft_class(struct Qdisc *sch, unsigned long arg,
 215                            struct Qdisc *new, struct Qdisc **old)
 216 {
 217         struct drr_class *cl = (struct drr_class *)arg;
 218
 219         if (new == NULL) {
 220                 new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
 221                                         &pfifo_qdisc_ops, cl->common.classid);
 222                 if (new == NULL)
 223                         new = &noop_qdisc;
 224         }
 225
 226         sch_tree_lock(sch);
 227         drr_purge_queue(cl);
 228         *old = cl->qdisc;
 229         cl->qdisc = new;
 230         sch_tree_unlock(sch);
 231         return 0;
 232 }
 233
 234 static struct Qdisc *drr_class_leaf(struct Qdisc *sch, unsigned long arg)
 235 {
 236         struct drr_class *cl = (struct drr_class *)arg;
 237
 238         return cl->qdisc;
 239 }
 240
 241 static void drr_qlen_notify(struct Qdisc *csh, unsigned long arg)
 242 {
 243         struct drr_class *cl = (struct drr_class *)arg;
 244
 245         if (cl->qdisc->q.qlen == 0)
 246                 list_del(&cl->alist);
 247 }
 248
 249 static int drr_dump_class(struct Qdisc *sch, unsigned long arg,
 250                           struct sk_buff *skb, struct tcmsg *tcm)
 251 {
 252         struct drr_class *cl = (struct drr_class *)arg;
 253         struct nlattr *nest;
 254
 255         tcm->tcm_parent = TC_H_ROOT;
 256         tcm->tcm_handle = cl->common.classid;
 257         tcm->tcm_info   = cl->qdisc->handle;
 258
 259         nest = nla_nest_start(skb, TCA_OPTIONS);
 260         if (nest == NULL)
 261                 goto nla_put_failure;
 262         NLA_PUT_U32(skb, TCA_DRR_QUANTUM, cl->quantum);
 263         return nla_nest_end(skb, nest);
 264
 265 nla_put_failure:
 266         nla_nest_cancel(skb, nest);
 267         return -EMSGSIZE;
 268 }
 269
 270 static int drr_dump_class_stats(struct Qdisc *sch, unsigned long arg,
 271                                 struct gnet_dump *d)
 272 {
 273         struct drr_class *cl = (struct drr_class *)arg;
 274         struct tc_drr_stats xstats;
 275
 276         memset(&xstats, 0, sizeof(xstats));
 277         if (cl->qdisc->q.qlen)
 278                 xstats.deficit = cl->deficit;
 279
 280         if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
 281             gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
 282             gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
 283                 return -1;
 284
 285         return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
 286 }
 287
 288 static void drr_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 289 {
 290         struct drr_sched *q = qdisc_priv(sch);
 291         struct drr_class *cl;
 292         struct hlist_node *n;
 293         unsigned int i;
 294
 295         if (arg->stop)
 296                 return;
 297
 298         for (i = 0; i < q->clhash.hashsize; i++) {
 299                 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
 300                         if (arg->count < arg->skip) {
 301                                 arg->count++;
 302                                 continue;
 303                         }
 304                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
 305                                 arg->stop = 1;
 306                                 return;
 307                         }
 308                         arg->count++;
 309                 }
 310         }
 311 }
 312
 313 static struct drr_class *drr_classify(struct sk_buff *skb, struct Qdisc *sch,
 314                                       int *qerr)
 315 {
 316         struct drr_sched *q = qdisc_priv(sch);
 317         struct drr_class *cl;
 318         struct tcf_result res;
 319         int result;
 320
 321         if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
 322                 cl = drr_find_class(sch, skb->priority);
 323                 if (cl != NULL)
 324                         return cl;
 325         }
 326
 327         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 328         result = tc_classify(skb, q->filter_list, &res);
 329         if (result >= 0) {
 330 #ifdef CONFIG_NET_CLS_ACT
 331                 switch (result) {
 332                 case TC_ACT_QUEUED:
 333                 case TC_ACT_STOLEN:
 334                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 335                 case TC_ACT_SHOT:
 336                         return NULL;
 337                 }
 338 #endif
 339                 cl = (struct drr_class *)res.class;
 340                 if (cl == NULL)
 341                         cl = drr_find_class(sch, res.classid);
 342                 return cl;
 343         }
 344         return NULL;
 345 }
 346
 347 static int drr_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 348 {
 349         struct drr_sched *q = qdisc_priv(sch);
 350         struct drr_class *cl;
 351         unsigned int len;
 352         int err;
 353
 354         cl = drr_classify(skb, sch, &err);
 355         if (cl == NULL) {
 356                 if (err & __NET_XMIT_BYPASS)
 357                         sch->qstats.drops++;
 358                 kfree_skb(skb);
 359                 return err;
 360         }
 361
 362         len = qdisc_pkt_len(skb);
 363         err = qdisc_enqueue(skb, cl->qdisc);
 364         if (unlikely(err != NET_XMIT_SUCCESS)) {
 365                 if (net_xmit_drop_count(err)) {
 366                         cl->qstats.drops++;
 367                         sch->qstats.drops++;
 368                 }
 369                 return err;
 370         }
 371
 372         if (cl->qdisc->q.qlen == 1) {
 373                 list_add_tail(&cl->alist, &q->active);
 374                 cl->deficit = cl->quantum;
 375         }
 376
 377         cl->bstats.packets++;
 378         cl->bstats.bytes += len;
 379         sch->bstats.packets++;
 380         sch->bstats.bytes += len;
 381
 382         sch->q.qlen++;
 383         return err;
 384 }
 385
 386 static struct sk_buff *drr_dequeue(struct Qdisc *sch)
 387 {
 388         struct drr_sched *q = qdisc_priv(sch);
 389         struct drr_class *cl;
 390         struct sk_buff *skb;
 391         unsigned int len;
 392
 393         if (list_empty(&q->active))
 394                 goto out;
 395         while (1) {
 396                 cl = list_first_entry(&q->active, struct drr_class, alist);
 397                 skb = cl->qdisc->ops->peek(cl->qdisc);
 398                 if (skb == NULL)
 399                         goto out;
 400
 401                 len = qdisc_pkt_len(skb);
 402                 if (len <= cl->deficit) {
 403                         cl->deficit -= len;
 404                         skb = qdisc_dequeue_peeked(cl->qdisc);
 405                         if (cl->qdisc->q.qlen == 0)
 406                                 list_del(&cl->alist);
 407                         sch->q.qlen--;
 408                         return skb;
 409                 }
 410
 411                 cl->deficit += cl->quantum;
 412                 list_move_tail(&cl->alist, &q->active);
 413         }
 414 out:
 415         return NULL;
 416 }
 417
 418 static unsigned int drr_drop(struct Qdisc *sch)
 419 {
 420         struct drr_sched *q = qdisc_priv(sch);
 421         struct drr_class *cl;
 422         unsigned int len;
 423
 424         list_for_each_entry(cl, &q->active, alist) {
 425                 if (cl->qdisc->ops->drop) {
 426                         len = cl->qdisc->ops->drop(cl->qdisc);
 427                         if (len > 0) {
 428                                 sch->q.qlen--;
 429                                 if (cl->qdisc->q.qlen == 0)
 430                                         list_del(&cl->alist);
 431                                 return len;
 432                         }
 433                 }
 434         }
 435         return 0;
 436 }
 437
 438 static int drr_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
 439 {
 440         struct drr_sched *q = qdisc_priv(sch);
 441         int err;
 442
 443         err = qdisc_class_hash_init(&q->clhash);
 444         if (err < 0)
 445                 return err;
 446         INIT_LIST_HEAD(&q->active);
 447         return 0;
 448 }
 449
 450 static void drr_reset_qdisc(struct Qdisc *sch)
 451 {
 452         struct drr_sched *q = qdisc_priv(sch);
 453         struct drr_class *cl;
 454         struct hlist_node *n;
 455         unsigned int i;
 456
 457         for (i = 0; i < q->clhash.hashsize; i++) {
 458                 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
 459                         if (cl->qdisc->q.qlen)
 460                                 list_del(&cl->alist);
 461                         qdisc_reset(cl->qdisc);
 462                 }
 463         }
 464         sch->q.qlen = 0;
 465 }
 466
 467 static void drr_destroy_qdisc(struct Qdisc *sch)
 468 {
 469         struct drr_sched *q = qdisc_priv(sch);
 470         struct drr_class *cl;
 471         struct hlist_node *n, *next;
 472         unsigned int i;
 473
 474         tcf_destroy_chain(&q->filter_list);
 475
 476         for (i = 0; i < q->clhash.hashsize; i++) {
 477                 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
 478                                           common.hnode)
 479                         drr_destroy_class(sch, cl);
 480         }
 481         qdisc_class_hash_destroy(&q->clhash);
 482 }
 483
 484 static const struct Qdisc_class_ops drr_class_ops = {
 485         .change         = drr_change_class,
 486         .delete         = drr_delete_class,
 487         .get            = drr_get_class,
 488         .put            = drr_put_class,
 489         .tcf_chain      = drr_tcf_chain,
 490         .bind_tcf       = drr_bind_tcf,
 491         .unbind_tcf     = drr_unbind_tcf,
 492         .graft          = drr_graft_class,
 493         .leaf           = drr_class_leaf,
 494         .qlen_notify    = drr_qlen_notify,
 495         .dump           = drr_dump_class,
 496         .dump_stats     = drr_dump_class_stats,
 497         .walk           = drr_walk,
 498 };
 499
 500 static struct Qdisc_ops drr_qdisc_ops __read_mostly = {
 501         .cl_ops         = &drr_class_ops,
 502         .id             = "drr",
 503         .priv_size      = sizeof(struct drr_sched),
 504         .enqueue        = drr_enqueue,
 505         .dequeue        = drr_dequeue,
 506         .peek           = qdisc_peek_dequeued,
 507         .drop           = drr_drop,
 508         .init           = drr_init_qdisc,
 509         .reset          = drr_reset_qdisc,
 510         .destroy        = drr_destroy_qdisc,
 511         .owner          = THIS_MODULE,
 512 };
 513
 514 static int __init drr_init(void)
 515 {
 516         return register_qdisc(&drr_qdisc_ops);
 517 }
 518
 519 static void __exit drr_exit(void)
 520 {
 521         unregister_qdisc(&drr_qdisc_ops);
 522 }
 523
 524 module_init(drr_init);
 525 module_exit(drr_exit);
 526 MODULE_LICENSE("GPL");