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         struct tcf_proto *filter_list;
  28         u8  prio2band[TC_PRIO_MAX+1];
  29         struct Qdisc *queues[TCQ_PRIO_BANDS];
  30 };
  31
  32
  33 static struct Qdisc *
  34 prio_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
  35 {
  36         struct prio_sched_data *q = qdisc_priv(sch);
  37         u32 band = skb->priority;
  38         struct tcf_result res;
  39         int err;
  40
  41         *qerr = NET_XMIT_BYPASS;
  42         if (TC_H_MAJ(skb->priority) != sch->handle) {
  43                 err = tc_classify(skb, q->filter_list, &res);
  44 #ifdef CONFIG_NET_CLS_ACT
  45                 switch (err) {
  46                 case TC_ACT_STOLEN:
  47                 case TC_ACT_QUEUED:
  48                         *qerr = NET_XMIT_SUCCESS;
  49                 case TC_ACT_SHOT:
  50                         return NULL;
  51                 }
  52 #endif
  53                 if (!q->filter_list || err < 0) {
  54                         if (TC_H_MAJ(band))
  55                                 band = 0;
  56                         return q->queues[q->prio2band[band&TC_PRIO_MAX]];
  57                 }
  58                 band = res.classid;
  59         }
  60         band = TC_H_MIN(band) - 1;
  61         if (band >= q->bands)
  62                 return q->queues[q->prio2band[0]];
  63
  64         return q->queues[band];
  65 }
  66
  67 static int
  68 prio_enqueue(struct sk_buff *skb, struct Qdisc *sch)
  69 {
  70         struct Qdisc *qdisc;
  71         int ret;
  72
  73         qdisc = prio_classify(skb, sch, &ret);
  74 #ifdef CONFIG_NET_CLS_ACT
  75         if (qdisc == NULL) {
  76
  77                 if (ret == NET_XMIT_BYPASS)
  78                         sch->qstats.drops++;
  79                 kfree_skb(skb);
  80                 return ret;
  81         }
  82 #endif
  83
  84         ret = qdisc_enqueue(skb, qdisc);
  85         if (ret == NET_XMIT_SUCCESS) {
  86                 sch->bstats.bytes += qdisc_pkt_len(skb);
  87                 sch->bstats.packets++;
  88                 sch->q.qlen++;
  89                 return NET_XMIT_SUCCESS;
  90         }
  91         sch->qstats.drops++;
  92         return ret;
  93 }
  94
  95
  96 static int
  97 prio_requeue(struct sk_buff *skb, struct Qdisc* sch)
  98 {
  99         struct Qdisc *qdisc;
 100         int ret;
 101
 102         qdisc = prio_classify(skb, sch, &ret);
 103 #ifdef CONFIG_NET_CLS_ACT
 104         if (qdisc == NULL) {
 105                 if (ret == NET_XMIT_BYPASS)
 106                         sch->qstats.drops++;
 107                 kfree_skb(skb);
 108                 return ret;
 109         }
 110 #endif
 111
 112         if ((ret = qdisc->ops->requeue(skb, qdisc)) == NET_XMIT_SUCCESS) {
 113                 sch->q.qlen++;
 114                 sch->qstats.requeues++;
 115                 return 0;
 116         }
 117         sch->qstats.drops++;
 118         return NET_XMIT_DROP;
 119 }
 120
 121
 122 static struct sk_buff *prio_dequeue(struct Qdisc* sch)
 123 {
 124         struct prio_sched_data *q = qdisc_priv(sch);
 125         int prio;
 126
 127         for (prio = 0; prio < q->bands; prio++) {
 128                 struct Qdisc *qdisc = q->queues[prio];
 129                 struct sk_buff *skb = qdisc->dequeue(qdisc);
 130                 if (skb) {
 131                         sch->q.qlen--;
 132                         return skb;
 133                 }
 134         }
 135         return NULL;
 136
 137 }
 138
 139 static unsigned int prio_drop(struct Qdisc* sch)
 140 {
 141         struct prio_sched_data *q = qdisc_priv(sch);
 142         int prio;
 143         unsigned int len;
 144         struct Qdisc *qdisc;
 145
 146         for (prio = q->bands-1; prio >= 0; prio--) {
 147                 qdisc = q->queues[prio];
 148                 if (qdisc->ops->drop && (len = qdisc->ops->drop(qdisc)) != 0) {
 149                         sch->q.qlen--;
 150                         return len;
 151                 }
 152         }
 153         return 0;
 154 }
 155
 156
 157 static void
 158 prio_reset(struct Qdisc* sch)
 159 {
 160         int prio;
 161         struct prio_sched_data *q = qdisc_priv(sch);
 162
 163         for (prio=0; prio<q->bands; prio++)
 164                 qdisc_reset(q->queues[prio]);
 165         sch->q.qlen = 0;
 166 }
 167
 168 static void
 169 prio_destroy(struct Qdisc* sch)
 170 {
 171         int prio;
 172         struct prio_sched_data *q = qdisc_priv(sch);
 173
 174         tcf_destroy_chain(&q->filter_list);
 175         for (prio=0; prio<q->bands; prio++)
 176                 qdisc_destroy(q->queues[prio]);
 177 }
 178
 179 static int prio_tune(struct Qdisc *sch, struct nlattr *opt)
 180 {
 181         struct prio_sched_data *q = qdisc_priv(sch);
 182         struct tc_prio_qopt *qopt;
 183         int i;
 184
 185         if (nla_len(opt) < sizeof(*qopt))
 186                 return -EINVAL;
 187         qopt = nla_data(opt);
 188
 189         if (qopt->bands > TCQ_PRIO_BANDS || qopt->bands < 2)
 190                 return -EINVAL;
 191
 192         for (i=0; i<=TC_PRIO_MAX; i++) {
 193                 if (qopt->priomap[i] >= qopt->bands)
 194                         return -EINVAL;
 195         }
 196
 197         sch_tree_lock(sch);
 198         q->bands = qopt->bands;
 199         memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
 200
 201         for (i=q->bands; i<TCQ_PRIO_BANDS; i++) {
 202                 struct Qdisc *child = xchg(&q->queues[i], &noop_qdisc);
 203                 if (child != &noop_qdisc) {
 204                         qdisc_tree_decrease_qlen(child, child->q.qlen);
 205                         qdisc_destroy(child);
 206                 }
 207         }
 208         sch_tree_unlock(sch);
 209
 210         for (i=0; i<q->bands; i++) {
 211                 if (q->queues[i] == &noop_qdisc) {
 212                         struct Qdisc *child;
 213                         child = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
 214                                                   &pfifo_qdisc_ops,
 215                                                   TC_H_MAKE(sch->handle, i + 1));
 216                         if (child) {
 217                                 sch_tree_lock(sch);
 218                                 child = xchg(&q->queues[i], child);
 219
 220                                 if (child != &noop_qdisc) {
 221                                         qdisc_tree_decrease_qlen(child,
 222                                                                  child->q.qlen);
 223                                         qdisc_destroy(child);
 224                                 }
 225                                 sch_tree_unlock(sch);
 226                         }
 227                 }
 228         }
 229         return 0;
 230 }
 231
 232 static int prio_init(struct Qdisc *sch, struct nlattr *opt)
 233 {
 234         struct prio_sched_data *q = qdisc_priv(sch);
 235         int i;
 236
 237         for (i=0; i<TCQ_PRIO_BANDS; i++)
 238                 q->queues[i] = &noop_qdisc;
 239
 240         if (opt == NULL) {
 241                 return -EINVAL;
 242         } else {
 243                 int err;
 244
 245                 if ((err= prio_tune(sch, opt)) != 0)
 246                         return err;
 247         }
 248         return 0;
 249 }
 250
 251 static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)
 252 {
 253         struct prio_sched_data *q = qdisc_priv(sch);
 254         unsigned char *b = skb_tail_pointer(skb);
 255         struct nlattr *nest;
 256         struct tc_prio_qopt opt;
 257
 258         opt.bands = q->bands;
 259         memcpy(&opt.priomap, q->prio2band, TC_PRIO_MAX+1);
 260
 261         nest = nla_nest_compat_start(skb, TCA_OPTIONS, sizeof(opt), &opt);
 262         if (nest == NULL)
 263                 goto nla_put_failure;
 264         nla_nest_compat_end(skb, nest);
 265
 266         return skb->len;
 267
 268 nla_put_failure:
 269         nlmsg_trim(skb, b);
 270         return -1;
 271 }
 272
 273 static int prio_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 274                       struct Qdisc **old)
 275 {
 276         struct prio_sched_data *q = qdisc_priv(sch);
 277         unsigned long band = arg - 1;
 278
 279         if (band >= q->bands)
 280                 return -EINVAL;
 281
 282         if (new == NULL)
 283                 new = &noop_qdisc;
 284
 285         sch_tree_lock(sch);
 286         *old = q->queues[band];
 287         q->queues[band] = new;
 288         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
 289         qdisc_reset(*old);
 290         sch_tree_unlock(sch);
 291
 292         return 0;
 293 }
 294
 295 static struct Qdisc *
 296 prio_leaf(struct Qdisc *sch, unsigned long arg)
 297 {
 298         struct prio_sched_data *q = qdisc_priv(sch);
 299         unsigned long band = arg - 1;
 300
 301         if (band >= q->bands)
 302                 return NULL;
 303
 304         return q->queues[band];
 305 }
 306
 307 static unsigned long prio_get(struct Qdisc *sch, u32 classid)
 308 {
 309         struct prio_sched_data *q = qdisc_priv(sch);
 310         unsigned long band = TC_H_MIN(classid);
 311
 312         if (band - 1 >= q->bands)
 313                 return 0;
 314         return band;
 315 }
 316
 317 static unsigned long prio_bind(struct Qdisc *sch, unsigned long parent, u32 classid)
 318 {
 319         return prio_get(sch, classid);
 320 }
 321
 322
 323 static void prio_put(struct Qdisc *q, unsigned long cl)
 324 {
 325         return;
 326 }
 327
 328 static int prio_change(struct Qdisc *sch, u32 handle, u32 parent, struct nlattr **tca, unsigned long *arg)
 329 {
 330         unsigned long cl = *arg;
 331         struct prio_sched_data *q = qdisc_priv(sch);
 332
 333         if (cl - 1 > q->bands)
 334                 return -ENOENT;
 335         return 0;
 336 }
 337
 338 static int prio_delete(struct Qdisc *sch, unsigned long cl)
 339 {
 340         struct prio_sched_data *q = qdisc_priv(sch);
 341         if (cl - 1 > q->bands)
 342                 return -ENOENT;
 343         return 0;
 344 }
 345
 346
 347 static int prio_dump_class(struct Qdisc *sch, unsigned long cl, struct sk_buff *skb,
 348                            struct tcmsg *tcm)
 349 {
 350         struct prio_sched_data *q = qdisc_priv(sch);
 351
 352         if (cl - 1 > q->bands)
 353                 return -ENOENT;
 354         tcm->tcm_handle |= TC_H_MIN(cl);
 355         if (q->queues[cl-1])
 356                 tcm->tcm_info = q->queues[cl-1]->handle;
 357         return 0;
 358 }
 359
 360 static int prio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 361                                  struct gnet_dump *d)
 362 {
 363         struct prio_sched_data *q = qdisc_priv(sch);
 364         struct Qdisc *cl_q;
 365
 366         cl_q = q->queues[cl - 1];
 367         if (gnet_stats_copy_basic(d, &cl_q->bstats) < 0 ||
 368             gnet_stats_copy_queue(d, &cl_q->qstats) < 0)
 369                 return -1;
 370
 371         return 0;
 372 }
 373
 374 static void prio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 375 {
 376         struct prio_sched_data *q = qdisc_priv(sch);
 377         int prio;
 378
 379         if (arg->stop)
 380                 return;
 381
 382         for (prio = 0; prio < q->bands; prio++) {
 383                 if (arg->count < arg->skip) {
 384                         arg->count++;
 385                         continue;
 386                 }
 387                 if (arg->fn(sch, prio+1, arg) < 0) {
 388                         arg->stop = 1;
 389                         break;
 390                 }
 391                 arg->count++;
 392         }
 393 }
 394
 395 static struct tcf_proto ** prio_find_tcf(struct Qdisc *sch, unsigned long cl)
 396 {
 397         struct prio_sched_data *q = qdisc_priv(sch);
 398
 399         if (cl)
 400                 return NULL;
 401         return &q->filter_list;
 402 }
 403
 404 static const struct Qdisc_class_ops prio_class_ops = {
 405         .graft          =       prio_graft,
 406         .leaf           =       prio_leaf,
 407         .get            =       prio_get,
 408         .put            =       prio_put,
 409         .change         =       prio_change,
 410         .delete         =       prio_delete,
 411         .walk           =       prio_walk,
 412         .tcf_chain      =       prio_find_tcf,
 413         .bind_tcf       =       prio_bind,
 414         .unbind_tcf     =       prio_put,
 415         .dump           =       prio_dump_class,
 416         .dump_stats     =       prio_dump_class_stats,
 417 };
 418
 419 static struct Qdisc_ops prio_qdisc_ops __read_mostly = {
 420         .next           =       NULL,
 421         .cl_ops         =       &prio_class_ops,
 422         .id             =       "prio",
 423         .priv_size      =       sizeof(struct prio_sched_data),
 424         .enqueue        =       prio_enqueue,
 425         .dequeue        =       prio_dequeue,
 426         .requeue        =       prio_requeue,
 427         .drop           =       prio_drop,
 428         .init           =       prio_init,
 429         .reset          =       prio_reset,
 430         .destroy        =       prio_destroy,
 431         .change         =       prio_tune,
 432         .dump           =       prio_dump,
 433         .owner          =       THIS_MODULE,
 434 };
 435
 436 static int __init prio_module_init(void)
 437 {
 438         return register_qdisc(&prio_qdisc_ops);
 439 }
 440
 441 static void __exit prio_module_exit(void)
 442 {
 443         unregister_qdisc(&prio_qdisc_ops);
 444 }
 445
 446 module_init(prio_module_init)
 447 module_exit(prio_module_exit)
 448
 449 MODULE_LICENSE("GPL");