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