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