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_packed          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,
  89                                                     qdisc_root_sleeping_running(sch),
  90                                                     tca[TCA_RATE]);
  91                         if (err) {
  92                                 NL_SET_ERR_MSG(extack, "Failed to replace estimator");
  93                                 return err;
  94                         }
  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->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,
 122                                             qdisc_root_sleeping_running(sch),
 123                                             tca[TCA_RATE]);
 124                 if (err) {
 125                         NL_SET_ERR_MSG(extack, "Failed to replace estimator");
 126                         qdisc_put(cl->qdisc);
 127                         kfree(cl);
 128                         return err;
 129                 }
 130         }
 131
 132         sch_tree_lock(sch);
 133         qdisc_class_hash_insert(&q->clhash, &cl->common);
 134         sch_tree_unlock(sch);
 135
 136         qdisc_class_hash_grow(sch, &q->clhash);
 137
 138         *arg = (unsigned long)cl;
 139         return 0;
 140 }
 141
 142 static void drr_destroy_class(struct Qdisc *sch, struct drr_class *cl)
 143 {
 144         gen_kill_estimator(&cl->rate_est);
 145         qdisc_put(cl->qdisc);
 146         kfree(cl);
 147 }
 148
 149 static int drr_delete_class(struct Qdisc *sch, unsigned long arg,
 150                             struct netlink_ext_ack *extack)
 151 {
 152         struct drr_sched *q = qdisc_priv(sch);
 153         struct drr_class *cl = (struct drr_class *)arg;
 154
 155         if (cl->filter_cnt > 0)
 156                 return -EBUSY;
 157
 158         sch_tree_lock(sch);
 159
 160         qdisc_purge_queue(cl->qdisc);
 161         qdisc_class_hash_remove(&q->clhash, &cl->common);
 162
 163         sch_tree_unlock(sch);
 164
 165         drr_destroy_class(sch, cl);
 166         return 0;
 167 }
 168
 169 static unsigned long drr_search_class(struct Qdisc *sch, u32 classid)
 170 {
 171         return (unsigned long)drr_find_class(sch, classid);
 172 }
 173
 174 static struct tcf_block *drr_tcf_block(struct Qdisc *sch, unsigned long cl,
 175                                        struct netlink_ext_ack *extack)
 176 {
 177         struct drr_sched *q = qdisc_priv(sch);
 178
 179         if (cl) {
 180                 NL_SET_ERR_MSG(extack, "DRR classid must be zero");
 181                 return NULL;
 182         }
 183
 184         return q->block;
 185 }
 186
 187 static unsigned long drr_bind_tcf(struct Qdisc *sch, unsigned long parent,
 188                                   u32 classid)
 189 {
 190         struct drr_class *cl = drr_find_class(sch, classid);
 191
 192         if (cl != NULL)
 193                 cl->filter_cnt++;
 194
 195         return (unsigned long)cl;
 196 }
 197
 198 static void drr_unbind_tcf(struct Qdisc *sch, unsigned long arg)
 199 {
 200         struct drr_class *cl = (struct drr_class *)arg;
 201
 202         cl->filter_cnt--;
 203 }
 204
 205 static int drr_graft_class(struct Qdisc *sch, unsigned long arg,
 206                            struct Qdisc *new, struct Qdisc **old,
 207                            struct netlink_ext_ack *extack)
 208 {
 209         struct drr_class *cl = (struct drr_class *)arg;
 210
 211         if (new == NULL) {
 212                 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
 213                                         cl->common.classid, NULL);
 214                 if (new == NULL)
 215                         new = &noop_qdisc;
 216         }
 217
 218         *old = qdisc_replace(sch, new, &cl->qdisc);
 219         return 0;
 220 }
 221
 222 static struct Qdisc *drr_class_leaf(struct Qdisc *sch, unsigned long arg)
 223 {
 224         struct drr_class *cl = (struct drr_class *)arg;
 225
 226         return cl->qdisc;
 227 }
 228
 229 static void drr_qlen_notify(struct Qdisc *csh, unsigned long arg)
 230 {
 231         struct drr_class *cl = (struct drr_class *)arg;
 232
 233         list_del(&cl->alist);
 234 }
 235
 236 static int drr_dump_class(struct Qdisc *sch, unsigned long arg,
 237                           struct sk_buff *skb, struct tcmsg *tcm)
 238 {
 239         struct drr_class *cl = (struct drr_class *)arg;
 240         struct nlattr *nest;
 241
 242         tcm->tcm_parent = TC_H_ROOT;
 243         tcm->tcm_handle = cl->common.classid;
 244         tcm->tcm_info   = cl->qdisc->handle;
 245
 246         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
 247         if (nest == NULL)
 248                 goto nla_put_failure;
 249         if (nla_put_u32(skb, TCA_DRR_QUANTUM, cl->quantum))
 250                 goto nla_put_failure;
 251         return nla_nest_end(skb, nest);
 252
 253 nla_put_failure:
 254         nla_nest_cancel(skb, nest);
 255         return -EMSGSIZE;
 256 }
 257
 258 static int drr_dump_class_stats(struct Qdisc *sch, unsigned long arg,
 259                                 struct gnet_dump *d)
 260 {
 261         struct drr_class *cl = (struct drr_class *)arg;
 262         __u32 qlen = qdisc_qlen_sum(cl->qdisc);
 263         struct Qdisc *cl_q = cl->qdisc;
 264         struct tc_drr_stats xstats;
 265
 266         memset(&xstats, 0, sizeof(xstats));
 267         if (qlen)
 268                 xstats.deficit = cl->deficit;
 269
 270         if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
 271                                   d, NULL, &cl->bstats) < 0 ||
 272             gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
 273             gnet_stats_copy_queue(d, cl_q->cpu_qstats, &cl_q->qstats, qlen) < 0)
 274                 return -1;
 275
 276         return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
 277 }
 278
 279 static void drr_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 280 {
 281         struct drr_sched *q = qdisc_priv(sch);
 282         struct drr_class *cl;
 283         unsigned int i;
 284
 285         if (arg->stop)
 286                 return;
 287
 288         for (i = 0; i < q->clhash.hashsize; i++) {
 289                 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
 290                         if (arg->count < arg->skip) {
 291                                 arg->count++;
 292                                 continue;
 293                         }
 294                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
 295                                 arg->stop = 1;
 296                                 return;
 297                         }
 298                         arg->count++;
 299                 }
 300         }
 301 }
 302
 303 static struct drr_class *drr_classify(struct sk_buff *skb, struct Qdisc *sch,
 304                                       int *qerr)
 305 {
 306         struct drr_sched *q = qdisc_priv(sch);
 307         struct drr_class *cl;
 308         struct tcf_result res;
 309         struct tcf_proto *fl;
 310         int result;
 311
 312         if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
 313                 cl = drr_find_class(sch, skb->priority);
 314                 if (cl != NULL)
 315                         return cl;
 316         }
 317
 318         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 319         fl = rcu_dereference_bh(q->filter_list);
 320         result = tcf_classify(skb, NULL, fl, &res, false);
 321         if (result >= 0) {
 322 #ifdef CONFIG_NET_CLS_ACT
 323                 switch (result) {
 324                 case TC_ACT_QUEUED:
 325                 case TC_ACT_STOLEN:
 326                 case TC_ACT_TRAP:
 327                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 328                         fallthrough;
 329                 case TC_ACT_SHOT:
 330                         return NULL;
 331                 }
 332 #endif
 333                 cl = (struct drr_class *)res.class;
 334                 if (cl == NULL)
 335                         cl = drr_find_class(sch, res.classid);
 336                 return cl;
 337         }
 338         return NULL;
 339 }
 340
 341 static int drr_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 342                        struct sk_buff **to_free)
 343 {
 344         unsigned int len = qdisc_pkt_len(skb);
 345         struct drr_sched *q = qdisc_priv(sch);
 346         struct drr_class *cl;
 347         int err = 0;
 348         bool first;
 349
 350         cl = drr_classify(skb, sch, &err);
 351         if (cl == NULL) {
 352                 if (err & __NET_XMIT_BYPASS)
 353                         qdisc_qstats_drop(sch);
 354                 __qdisc_drop(skb, to_free);
 355                 return err;
 356         }
 357
 358         first = !cl->qdisc->q.qlen;
 359         err = qdisc_enqueue(skb, cl->qdisc, to_free);
 360         if (unlikely(err != NET_XMIT_SUCCESS)) {
 361                 if (net_xmit_drop_count(err)) {
 362                         cl->qstats.drops++;
 363                         qdisc_qstats_drop(sch);
 364                 }
 365                 return err;
 366         }
 367
 368         if (first) {
 369                 list_add_tail(&cl->alist, &q->active);
 370                 cl->deficit = cl->quantum;
 371         }
 372
 373         sch->qstats.backlog += len;
 374         sch->q.qlen++;
 375         return err;
 376 }
 377
 378 static struct sk_buff *drr_dequeue(struct Qdisc *sch)
 379 {
 380         struct drr_sched *q = qdisc_priv(sch);
 381         struct drr_class *cl;
 382         struct sk_buff *skb;
 383         unsigned int len;
 384
 385         if (list_empty(&q->active))
 386                 goto out;
 387         while (1) {
 388                 cl = list_first_entry(&q->active, struct drr_class, alist);
 389                 skb = cl->qdisc->ops->peek(cl->qdisc);
 390                 if (skb == NULL) {
 391                         qdisc_warn_nonwc(__func__, cl->qdisc);
 392                         goto out;
 393                 }
 394
 395                 len = qdisc_pkt_len(skb);
 396                 if (len <= cl->deficit) {
 397                         cl->deficit -= len;
 398                         skb = qdisc_dequeue_peeked(cl->qdisc);
 399                         if (unlikely(skb == NULL))
 400                                 goto out;
 401                         if (cl->qdisc->q.qlen == 0)
 402                                 list_del(&cl->alist);
 403
 404                         bstats_update(&cl->bstats, skb);
 405                         qdisc_bstats_update(sch, skb);
 406                         qdisc_qstats_backlog_dec(sch, skb);
 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 int drr_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
 419                           struct netlink_ext_ack *extack)
 420 {
 421         struct drr_sched *q = qdisc_priv(sch);
 422         int err;
 423
 424         err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
 425         if (err)
 426                 return err;
 427         err = qdisc_class_hash_init(&q->clhash);
 428         if (err < 0)
 429                 return err;
 430         INIT_LIST_HEAD(&q->active);
 431         return 0;
 432 }
 433
 434 static void drr_reset_qdisc(struct Qdisc *sch)
 435 {
 436         struct drr_sched *q = qdisc_priv(sch);
 437         struct drr_class *cl;
 438         unsigned int i;
 439
 440         for (i = 0; i < q->clhash.hashsize; i++) {
 441                 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
 442                         if (cl->qdisc->q.qlen)
 443                                 list_del(&cl->alist);
 444                         qdisc_reset(cl->qdisc);
 445                 }
 446         }
 447         sch->qstats.backlog = 0;
 448         sch->q.qlen = 0;
 449 }
 450
 451 static void drr_destroy_qdisc(struct Qdisc *sch)
 452 {
 453         struct drr_sched *q = qdisc_priv(sch);
 454         struct drr_class *cl;
 455         struct hlist_node *next;
 456         unsigned int i;
 457
 458         tcf_block_put(q->block);
 459
 460         for (i = 0; i < q->clhash.hashsize; i++) {
 461                 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
 462                                           common.hnode)
 463                         drr_destroy_class(sch, cl);
 464         }
 465         qdisc_class_hash_destroy(&q->clhash);
 466 }
 467
 468 static const struct Qdisc_class_ops drr_class_ops = {
 469         .change         = drr_change_class,
 470         .delete         = drr_delete_class,
 471         .find           = drr_search_class,
 472         .tcf_block      = drr_tcf_block,
 473         .bind_tcf       = drr_bind_tcf,
 474         .unbind_tcf     = drr_unbind_tcf,
 475         .graft          = drr_graft_class,
 476         .leaf           = drr_class_leaf,
 477         .qlen_notify    = drr_qlen_notify,
 478         .dump           = drr_dump_class,
 479         .dump_stats     = drr_dump_class_stats,
 480         .walk           = drr_walk,
 481 };
 482
 483 static struct Qdisc_ops drr_qdisc_ops __read_mostly = {
 484         .cl_ops         = &drr_class_ops,
 485         .id             = "drr",
 486         .priv_size      = sizeof(struct drr_sched),
 487         .enqueue        = drr_enqueue,
 488         .dequeue        = drr_dequeue,
 489         .peek           = qdisc_peek_dequeued,
 490         .init           = drr_init_qdisc,
 491         .reset          = drr_reset_qdisc,
 492         .destroy        = drr_destroy_qdisc,
 493         .owner          = THIS_MODULE,
 494 };
 495
 496 static int __init drr_init(void)
 497 {
 498         return register_qdisc(&drr_qdisc_ops);
 499 }
 500
 501 static void __exit drr_exit(void)
 502 {
 503         unregister_qdisc(&drr_qdisc_ops);
 504 }
 505
 506 module_init(drr_init);
 507 module_exit(drr_exit);
 508 MODULE_LICENSE("GPL");