net/sched/sch_drr.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * net/sched/sch_drr.c         Deficit Round Robin scheduler
   4  *
   5  * Copyright (c) 2008 Patrick McHardy <kaber@trash.net>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/slab.h>
  10 #include <linux/init.h>
  11 #include <linux/errno.h>
  12 #include <linux/netdevice.h>
  13 #include <linux/pkt_sched.h>
  14 #include <net/sch_generic.h>
  15 #include <net/pkt_sched.h>
  16 #include <net/pkt_cls.h>
  17
  18 struct drr_class {
  19         struct Qdisc_class_common       common;
  20         unsigned int                    filter_cnt;
  21
  22         struct gnet_stats_basic_sync            bstats;
  23         struct gnet_stats_queue         qstats;
  24         struct net_rate_estimator __rcu *rate_est;
  25         struct list_head                alist;
  26         struct Qdisc                    *qdisc;
  27
  28         u32                             quantum;
  29         u32                             deficit;
  30 };
  31
  32 struct drr_sched {
  33         struct list_head                active;
  34         struct tcf_proto __rcu          *filter_list;
  35         struct tcf_block                *block;
  36         struct Qdisc_class_hash         clhash;
  37 };
  38
  39 static struct drr_class *drr_find_class(struct Qdisc *sch, u32 classid)
  40 {
  41         struct drr_sched *q = qdisc_priv(sch);
  42         struct Qdisc_class_common *clc;
  43
  44         clc = qdisc_class_find(&q->clhash, classid);
  45         if (clc == NULL)
  46                 return NULL;
  47         return container_of(clc, struct drr_class, common);
  48 }
  49
  50 static const struct nla_policy drr_policy[TCA_DRR_MAX + 1] = {
  51         [TCA_DRR_QUANTUM]       = { .type = NLA_U32 },
  52 };
  53
  54 static int drr_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
  55                             struct nlattr **tca, unsigned long *arg,
  56                             struct netlink_ext_ack *extack)
  57 {
  58         struct drr_sched *q = qdisc_priv(sch);
  59         struct drr_class *cl = (struct drr_class *)*arg;
  60         struct nlattr *opt = tca[TCA_OPTIONS];
  61         struct nlattr *tb[TCA_DRR_MAX + 1];
  62         u32 quantum;
  63         int err;
  64
  65         if (!opt) {
  66                 NL_SET_ERR_MSG(extack, "DRR options are required for this operation");
  67                 return -EINVAL;
  68         }
  69
  70         err = nla_parse_nested_deprecated(tb, TCA_DRR_MAX, opt, drr_policy,
  71                                           extack);
  72         if (err < 0)
  73                 return err;
  74
  75         if (tb[TCA_DRR_QUANTUM]) {
  76                 quantum = nla_get_u32(tb[TCA_DRR_QUANTUM]);
  77                 if (quantum == 0) {
  78                         NL_SET_ERR_MSG(extack, "Specified DRR quantum cannot be zero");
  79                         return -EINVAL;
  80                 }
  81         } else
  82                 quantum = psched_mtu(qdisc_dev(sch));
  83
  84         if (cl != NULL) {
  85                 if (tca[TCA_RATE]) {
  86                         err = gen_replace_estimator(&cl->bstats, NULL,
  87                                                     &cl->rate_est,
  88                                                     NULL, true,
  89                                                     tca[TCA_RATE]);
  90                         if (err) {
  91                                 NL_SET_ERR_MSG(extack, "Failed to replace estimator");
  92                                 return err;
  93                         }
  94                 }
  95
  96                 sch_tree_lock(sch);
  97                 if (tb[TCA_DRR_QUANTUM])
  98                         cl->quantum = quantum;
  99                 sch_tree_unlock(sch);
 100
 101                 return 0;
 102         }
 103
 104         cl = kzalloc(sizeof(struct drr_class), GFP_KERNEL);
 105         if (cl == NULL)
 106                 return -ENOBUFS;
 107
 108         gnet_stats_basic_sync_init(&cl->bstats);
 109         cl->common.classid = classid;
 110         cl->quantum        = quantum;
 111         cl->qdisc          = qdisc_create_dflt(sch->dev_queue,
 112                                                &pfifo_qdisc_ops, classid,
 113                                                NULL);
 114         if (cl->qdisc == NULL)
 115                 cl->qdisc = &noop_qdisc;
 116         else
 117                 qdisc_hash_add(cl->qdisc, true);
 118
 119         if (tca[TCA_RATE]) {
 120                 err = gen_replace_estimator(&cl->bstats, NULL, &cl->rate_est,
 121                                             NULL, true, tca[TCA_RATE]);
 122                 if (err) {
 123                         NL_SET_ERR_MSG(extack, "Failed to replace estimator");
 124                         qdisc_put(cl->qdisc);
 125                         kfree(cl);
 126                         return err;
 127                 }
 128         }
 129
 130         sch_tree_lock(sch);
 131         qdisc_class_hash_insert(&q->clhash, &cl->common);
 132         sch_tree_unlock(sch);
 133
 134         qdisc_class_hash_grow(sch, &q->clhash);
 135
 136         *arg = (unsigned long)cl;
 137         return 0;
 138 }
 139
 140 static void drr_destroy_class(struct Qdisc *sch, struct drr_class *cl)
 141 {
 142         gen_kill_estimator(&cl->rate_est);
 143         qdisc_put(cl->qdisc);
 144         kfree(cl);
 145 }
 146
 147 static int drr_delete_class(struct Qdisc *sch, unsigned long arg,
 148                             struct netlink_ext_ack *extack)
 149 {
 150         struct drr_sched *q = qdisc_priv(sch);
 151         struct drr_class *cl = (struct drr_class *)arg;
 152
 153         if (cl->filter_cnt > 0)
 154                 return -EBUSY;
 155
 156         sch_tree_lock(sch);
 157
 158         qdisc_purge_queue(cl->qdisc);
 159         qdisc_class_hash_remove(&q->clhash, &cl->common);
 160
 161         sch_tree_unlock(sch);
 162
 163         drr_destroy_class(sch, cl);
 164         return 0;
 165 }
 166
 167 static unsigned long drr_search_class(struct Qdisc *sch, u32 classid)
 168 {
 169         return (unsigned long)drr_find_class(sch, classid);
 170 }
 171
 172 static struct tcf_block *drr_tcf_block(struct Qdisc *sch, unsigned long cl,
 173                                        struct netlink_ext_ack *extack)
 174 {
 175         struct drr_sched *q = qdisc_priv(sch);
 176
 177         if (cl) {
 178                 NL_SET_ERR_MSG(extack, "DRR classid must be zero");
 179                 return NULL;
 180         }
 181
 182         return q->block;
 183 }
 184
 185 static unsigned long drr_bind_tcf(struct Qdisc *sch, unsigned long parent,
 186                                   u32 classid)
 187 {
 188         struct drr_class *cl = drr_find_class(sch, classid);
 189
 190         if (cl != NULL)
 191                 cl->filter_cnt++;
 192
 193         return (unsigned long)cl;
 194 }
 195
 196 static void drr_unbind_tcf(struct Qdisc *sch, unsigned long arg)
 197 {
 198         struct drr_class *cl = (struct drr_class *)arg;
 199
 200         cl->filter_cnt--;
 201 }
 202
 203 static int drr_graft_class(struct Qdisc *sch, unsigned long arg,
 204                            struct Qdisc *new, struct Qdisc **old,
 205                            struct netlink_ext_ack *extack)
 206 {
 207         struct drr_class *cl = (struct drr_class *)arg;
 208
 209         if (new == NULL) {
 210                 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
 211                                         cl->common.classid, NULL);
 212                 if (new == NULL)
 213                         new = &noop_qdisc;
 214         }
 215
 216         *old = qdisc_replace(sch, new, &cl->qdisc);
 217         return 0;
 218 }
 219
 220 static struct Qdisc *drr_class_leaf(struct Qdisc *sch, unsigned long arg)
 221 {
 222         struct drr_class *cl = (struct drr_class *)arg;
 223
 224         return cl->qdisc;
 225 }
 226
 227 static void drr_qlen_notify(struct Qdisc *csh, unsigned long arg)
 228 {
 229         struct drr_class *cl = (struct drr_class *)arg;
 230
 231         list_del(&cl->alist);
 232 }
 233
 234 static int drr_dump_class(struct Qdisc *sch, unsigned long arg,
 235                           struct sk_buff *skb, struct tcmsg *tcm)
 236 {
 237         struct drr_class *cl = (struct drr_class *)arg;
 238         struct nlattr *nest;
 239
 240         tcm->tcm_parent = TC_H_ROOT;
 241         tcm->tcm_handle = cl->common.classid;
 242         tcm->tcm_info   = cl->qdisc->handle;
 243
 244         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
 245         if (nest == NULL)
 246                 goto nla_put_failure;
 247         if (nla_put_u32(skb, TCA_DRR_QUANTUM, cl->quantum))
 248                 goto nla_put_failure;
 249         return nla_nest_end(skb, nest);
 250
 251 nla_put_failure:
 252         nla_nest_cancel(skb, nest);
 253         return -EMSGSIZE;
 254 }
 255
 256 static int drr_dump_class_stats(struct Qdisc *sch, unsigned long arg,
 257                                 struct gnet_dump *d)
 258 {
 259         struct drr_class *cl = (struct drr_class *)arg;
 260         __u32 qlen = qdisc_qlen_sum(cl->qdisc);
 261         struct Qdisc *cl_q = cl->qdisc;
 262         struct tc_drr_stats xstats;
 263
 264         memset(&xstats, 0, sizeof(xstats));
 265         if (qlen)
 266                 xstats.deficit = cl->deficit;
 267
 268         if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
 269             gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
 270             gnet_stats_copy_queue(d, cl_q->cpu_qstats, &cl_q->qstats, qlen) < 0)
 271                 return -1;
 272
 273         return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
 274 }
 275
 276 static void drr_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 277 {
 278         struct drr_sched *q = qdisc_priv(sch);
 279         struct drr_class *cl;
 280         unsigned int i;
 281
 282         if (arg->stop)
 283                 return;
 284
 285         for (i = 0; i < q->clhash.hashsize; i++) {
 286                 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
 287                         if (arg->count < arg->skip) {
 288                                 arg->count++;
 289                                 continue;
 290                         }
 291                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
 292                                 arg->stop = 1;
 293                                 return;
 294                         }
 295                         arg->count++;
 296                 }
 297         }
 298 }
 299
 300 static struct drr_class *drr_classify(struct sk_buff *skb, struct Qdisc *sch,
 301                                       int *qerr)
 302 {
 303         struct drr_sched *q = qdisc_priv(sch);
 304         struct drr_class *cl;
 305         struct tcf_result res;
 306         struct tcf_proto *fl;
 307         int result;
 308
 309         if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
 310                 cl = drr_find_class(sch, skb->priority);
 311                 if (cl != NULL)
 312                         return cl;
 313         }
 314
 315         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 316         fl = rcu_dereference_bh(q->filter_list);
 317         result = tcf_classify(skb, NULL, fl, &res, false);
 318         if (result >= 0) {
 319 #ifdef CONFIG_NET_CLS_ACT
 320                 switch (result) {
 321                 case TC_ACT_QUEUED:
 322                 case TC_ACT_STOLEN:
 323                 case TC_ACT_TRAP:
 324                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 325                         fallthrough;
 326                 case TC_ACT_SHOT:
 327                         return NULL;
 328                 }
 329 #endif
 330                 cl = (struct drr_class *)res.class;
 331                 if (cl == NULL)
 332                         cl = drr_find_class(sch, res.classid);
 333                 return cl;
 334         }
 335         return NULL;
 336 }
 337
 338 static int drr_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 339                        struct sk_buff **to_free)
 340 {
 341         unsigned int len = qdisc_pkt_len(skb);
 342         struct drr_sched *q = qdisc_priv(sch);
 343         struct drr_class *cl;
 344         int err = 0;
 345         bool first;
 346
 347         cl = drr_classify(skb, sch, &err);
 348         if (cl == NULL) {
 349                 if (err & __NET_XMIT_BYPASS)
 350                         qdisc_qstats_drop(sch);
 351                 __qdisc_drop(skb, to_free);
 352                 return err;
 353         }
 354
 355         first = !cl->qdisc->q.qlen;
 356         err = qdisc_enqueue(skb, cl->qdisc, to_free);
 357         if (unlikely(err != NET_XMIT_SUCCESS)) {
 358                 if (net_xmit_drop_count(err)) {
 359                         cl->qstats.drops++;
 360                         qdisc_qstats_drop(sch);
 361                 }
 362                 return err;
 363         }
 364
 365         if (first) {
 366                 list_add_tail(&cl->alist, &q->active);
 367                 cl->deficit = cl->quantum;
 368         }
 369
 370         sch->qstats.backlog += len;
 371         sch->q.qlen++;
 372         return err;
 373 }
 374
 375 static struct sk_buff *drr_dequeue(struct Qdisc *sch)
 376 {
 377         struct drr_sched *q = qdisc_priv(sch);
 378         struct drr_class *cl;
 379         struct sk_buff *skb;
 380         unsigned int len;
 381
 382         if (list_empty(&q->active))
 383                 goto out;
 384         while (1) {
 385                 cl = list_first_entry(&q->active, struct drr_class, alist);
 386                 skb = cl->qdisc->ops->peek(cl->qdisc);
 387                 if (skb == NULL) {
 388                         qdisc_warn_nonwc(__func__, cl->qdisc);
 389                         goto out;
 390                 }
 391
 392                 len = qdisc_pkt_len(skb);
 393                 if (len <= cl->deficit) {
 394                         cl->deficit -= len;
 395                         skb = qdisc_dequeue_peeked(cl->qdisc);
 396                         if (unlikely(skb == NULL))
 397                                 goto out;
 398                         if (cl->qdisc->q.qlen == 0)
 399                                 list_del(&cl->alist);
 400
 401                         bstats_update(&cl->bstats, skb);
 402                         qdisc_bstats_update(sch, skb);
 403                         qdisc_qstats_backlog_dec(sch, skb);
 404                         sch->q.qlen--;
 405                         return skb;
 406                 }
 407
 408                 cl->deficit += cl->quantum;
 409                 list_move_tail(&cl->alist, &q->active);
 410         }
 411 out:
 412         return NULL;
 413 }
 414
 415 static int drr_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
 416                           struct netlink_ext_ack *extack)
 417 {
 418         struct drr_sched *q = qdisc_priv(sch);
 419         int err;
 420
 421         err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
 422         if (err)
 423                 return err;
 424         err = qdisc_class_hash_init(&q->clhash);
 425         if (err < 0)
 426                 return err;
 427         INIT_LIST_HEAD(&q->active);
 428         return 0;
 429 }
 430
 431 static void drr_reset_qdisc(struct Qdisc *sch)
 432 {
 433         struct drr_sched *q = qdisc_priv(sch);
 434         struct drr_class *cl;
 435         unsigned int i;
 436
 437         for (i = 0; i < q->clhash.hashsize; i++) {
 438                 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
 439                         if (cl->qdisc->q.qlen)
 440                                 list_del(&cl->alist);
 441                         qdisc_reset(cl->qdisc);
 442                 }
 443         }
 444         sch->qstats.backlog = 0;
 445         sch->q.qlen = 0;
 446 }
 447
 448 static void drr_destroy_qdisc(struct Qdisc *sch)
 449 {
 450         struct drr_sched *q = qdisc_priv(sch);
 451         struct drr_class *cl;
 452         struct hlist_node *next;
 453         unsigned int i;
 454
 455         tcf_block_put(q->block);
 456
 457         for (i = 0; i < q->clhash.hashsize; i++) {
 458                 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
 459                                           common.hnode)
 460                         drr_destroy_class(sch, cl);
 461         }
 462         qdisc_class_hash_destroy(&q->clhash);
 463 }
 464
 465 static const struct Qdisc_class_ops drr_class_ops = {
 466         .change         = drr_change_class,
 467         .delete         = drr_delete_class,
 468         .find           = drr_search_class,
 469         .tcf_block      = drr_tcf_block,
 470         .bind_tcf       = drr_bind_tcf,
 471         .unbind_tcf     = drr_unbind_tcf,
 472         .graft          = drr_graft_class,
 473         .leaf           = drr_class_leaf,
 474         .qlen_notify    = drr_qlen_notify,
 475         .dump           = drr_dump_class,
 476         .dump_stats     = drr_dump_class_stats,
 477         .walk           = drr_walk,
 478 };
 479
 480 static struct Qdisc_ops drr_qdisc_ops __read_mostly = {
 481         .cl_ops         = &drr_class_ops,
 482         .id             = "drr",
 483         .priv_size      = sizeof(struct drr_sched),
 484         .enqueue        = drr_enqueue,
 485         .dequeue        = drr_dequeue,
 486         .peek           = qdisc_peek_dequeued,
 487         .init           = drr_init_qdisc,
 488         .reset          = drr_reset_qdisc,
 489         .destroy        = drr_destroy_qdisc,
 490         .owner          = THIS_MODULE,
 491 };
 492
 493 static int __init drr_init(void)
 494 {
 495         return register_qdisc(&drr_qdisc_ops);
 496 }
 497
 498 static void __exit drr_exit(void)
 499 {
 500         unregister_qdisc(&drr_qdisc_ops);
 501 }
 502
 503 module_init(drr_init);
 504 module_exit(drr_exit);
 505 MODULE_LICENSE("GPL");