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 <asm/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 struct Qdisc *prio_classify(struct sk_buff *skb, struct Qdisc *sch,int *r)
  51 {
  52         struct prio_sched_data *q = qdisc_priv(sch);
  53         u32 band = skb->priority;
  54         struct tcf_result res;
  55
  56         if (TC_H_MAJ(skb->priority) != sch->handle) {
  57 #ifdef CONFIG_NET_CLS_ACT
  58                 int result = 0, terminal = 0;
  59                 result = tc_classify(skb, q->filter_list, &res);
  60
  61                 switch (result) {
  62                         case TC_ACT_SHOT:
  63                                 *r = NET_XMIT_DROP;
  64                                 terminal = 1;
  65                                 break;
  66                         case TC_ACT_STOLEN:
  67                         case TC_ACT_QUEUED:
  68                                 terminal = 1;
  69                                 break;
  70                         case TC_ACT_RECLASSIFY:
  71                         case TC_ACT_OK:
  72                         case TC_ACT_UNSPEC:
  73                         default:
  74                         break;
  75                 };
  76                 if (terminal) {
  77                         kfree_skb(skb);
  78                         return NULL;
  79                 }
  80
  81                 if (!q->filter_list ) {
  82 #else
  83                 if (!q->filter_list || tc_classify(skb, q->filter_list, &res)) {
  84 #endif
  85                         if (TC_H_MAJ(band))
  86                                 band = 0;
  87                         return q->queues[q->prio2band[band&TC_PRIO_MAX]];
  88                 }
  89                 band = res.classid;
  90         }
  91         band = TC_H_MIN(band) - 1;
  92         if (band > q->bands)
  93                 return q->queues[q->prio2band[0]];
  94
  95         return q->queues[band];
  96 }
  97
  98 static int
  99 prio_enqueue(struct sk_buff *skb, struct Qdisc* sch)
 100 {
 101         struct Qdisc *qdisc;
 102         int ret = NET_XMIT_SUCCESS;
 103
 104         qdisc = prio_classify(skb, sch, &ret);
 105
 106         if (NULL == qdisc)
 107                 goto dropped;
 108
 109         if ((ret = qdisc->enqueue(skb, qdisc)) == NET_XMIT_SUCCESS) {
 110                 sch->stats.bytes += skb->len;
 111                 sch->stats.packets++;
 112                 sch->q.qlen++;
 113                 return NET_XMIT_SUCCESS;
 114         }
 115
 116 dropped:
 117 #ifdef CONFIG_NET_CLS_ACT
 118         if (NET_XMIT_DROP == ret) {
 119 #endif
 120                 sch->stats.drops++;
 121 #ifdef CONFIG_NET_CLS_ACT
 122         } else {
 123                 sch->stats.overlimits++; /* abuse, but noone uses it */
 124         }
 125 #endif
 126         return ret;
 127 }
 128
 129
 130 static int
 131 prio_requeue(struct sk_buff *skb, struct Qdisc* sch)
 132 {
 133         struct Qdisc *qdisc;
 134         int ret = NET_XMIT_DROP;
 135
 136         qdisc = prio_classify(skb, sch, &ret);
 137         if (qdisc == NULL)
 138                 goto dropped;
 139
 140         if ((ret = qdisc->ops->requeue(skb, qdisc)) == 0) {
 141                 sch->q.qlen++;
 142                 return 0;
 143         }
 144 dropped:
 145         sch->stats.drops++;
 146         return NET_XMIT_DROP;
 147 }
 148
 149
 150 static struct sk_buff *
 151 prio_dequeue(struct Qdisc* sch)
 152 {
 153         struct sk_buff *skb;
 154         struct prio_sched_data *q = qdisc_priv(sch);
 155         int prio;
 156         struct Qdisc *qdisc;
 157
 158         for (prio = 0; prio < q->bands; prio++) {
 159                 qdisc = q->queues[prio];
 160                 skb = qdisc->dequeue(qdisc);
 161                 if (skb) {
 162                         sch->q.qlen--;
 163                         return skb;
 164                 }
 165         }
 166         return NULL;
 167
 168 }
 169
 170 static unsigned int prio_drop(struct Qdisc* sch)
 171 {
 172         struct prio_sched_data *q = qdisc_priv(sch);
 173         int prio;
 174         unsigned int len;
 175         struct Qdisc *qdisc;
 176
 177         for (prio = q->bands-1; prio >= 0; prio--) {
 178                 qdisc = q->queues[prio];
 179                 if ((len = qdisc->ops->drop(qdisc)) != 0) {
 180                         sch->q.qlen--;
 181                         return len;
 182                 }
 183         }
 184         return 0;
 185 }
 186
 187
 188 static void
 189 prio_reset(struct Qdisc* sch)
 190 {
 191         int prio;
 192         struct prio_sched_data *q = qdisc_priv(sch);
 193
 194         for (prio=0; prio<q->bands; prio++)
 195                 qdisc_reset(q->queues[prio]);
 196         sch->q.qlen = 0;
 197 }
 198
 199 static void
 200 prio_destroy(struct Qdisc* sch)
 201 {
 202         int prio;
 203         struct prio_sched_data *q = qdisc_priv(sch);
 204         struct tcf_proto *tp;
 205
 206         while ((tp = q->filter_list) != NULL) {
 207                 q->filter_list = tp->next;
 208                 tcf_destroy(tp);
 209         }
 210
 211         for (prio=0; prio<q->bands; prio++)
 212                 qdisc_destroy(q->queues[prio]);
 213 }
 214
 215 static int prio_tune(struct Qdisc *sch, struct rtattr *opt)
 216 {
 217         struct prio_sched_data *q = qdisc_priv(sch);
 218         struct tc_prio_qopt *qopt = RTA_DATA(opt);
 219         int i;
 220
 221         if (opt->rta_len < RTA_LENGTH(sizeof(*qopt)))
 222                 return -EINVAL;
 223         if (qopt->bands > TCQ_PRIO_BANDS || qopt->bands < 2)
 224                 return -EINVAL;
 225
 226         for (i=0; i<=TC_PRIO_MAX; i++) {
 227                 if (qopt->priomap[i] >= qopt->bands)
 228                         return -EINVAL;
 229         }
 230
 231         sch_tree_lock(sch);
 232         q->bands = qopt->bands;
 233         memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
 234
 235         for (i=q->bands; i<TCQ_PRIO_BANDS; i++) {
 236                 struct Qdisc *child = xchg(&q->queues[i], &noop_qdisc);
 237                 if (child != &noop_qdisc)
 238                         qdisc_destroy(child);
 239         }
 240         sch_tree_unlock(sch);
 241
 242         for (i=0; i<=TC_PRIO_MAX; i++) {
 243                 int band = q->prio2band[i];
 244                 if (q->queues[band] == &noop_qdisc) {
 245                         struct Qdisc *child;
 246                         child = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
 247                         if (child) {
 248                                 sch_tree_lock(sch);
 249                                 child = xchg(&q->queues[band], child);
 250
 251                                 if (child != &noop_qdisc)
 252                                         qdisc_destroy(child);
 253                                 sch_tree_unlock(sch);
 254                         }
 255                 }
 256         }
 257         return 0;
 258 }
 259
 260 static int prio_init(struct Qdisc *sch, struct rtattr *opt)
 261 {
 262         struct prio_sched_data *q = qdisc_priv(sch);
 263         int i;
 264
 265         for (i=0; i<TCQ_PRIO_BANDS; i++)
 266                 q->queues[i] = &noop_qdisc;
 267
 268         if (opt == NULL) {
 269                 return -EINVAL;
 270         } else {
 271                 int err;
 272
 273                 if ((err= prio_tune(sch, opt)) != 0)
 274                         return err;
 275         }
 276         return 0;
 277 }
 278
 279 static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)
 280 {
 281         struct prio_sched_data *q = qdisc_priv(sch);
 282         unsigned char    *b = skb->tail;
 283         struct tc_prio_qopt opt;
 284
 285         opt.bands = q->bands;
 286         memcpy(&opt.priomap, q->prio2band, TC_PRIO_MAX+1);
 287         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 288         return skb->len;
 289
 290 rtattr_failure:
 291         skb_trim(skb, b - skb->data);
 292         return -1;
 293 }
 294
 295 static int prio_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 296                       struct Qdisc **old)
 297 {
 298         struct prio_sched_data *q = qdisc_priv(sch);
 299         unsigned long band = arg - 1;
 300
 301         if (band >= q->bands)
 302                 return -EINVAL;
 303
 304         if (new == NULL)
 305                 new = &noop_qdisc;
 306
 307         sch_tree_lock(sch);
 308         *old = q->queues[band];
 309         q->queues[band] = new;
 310         sch->q.qlen -= (*old)->q.qlen;
 311         qdisc_reset(*old);
 312         sch_tree_unlock(sch);
 313
 314         return 0;
 315 }
 316
 317 static struct Qdisc *
 318 prio_leaf(struct Qdisc *sch, unsigned long arg)
 319 {
 320         struct prio_sched_data *q = qdisc_priv(sch);
 321         unsigned long band = arg - 1;
 322
 323         if (band >= q->bands)
 324                 return NULL;
 325
 326         return q->queues[band];
 327 }
 328
 329 static unsigned long prio_get(struct Qdisc *sch, u32 classid)
 330 {
 331         struct prio_sched_data *q = qdisc_priv(sch);
 332         unsigned long band = TC_H_MIN(classid);
 333
 334         if (band - 1 >= q->bands)
 335                 return 0;
 336         return band;
 337 }
 338
 339 static unsigned long prio_bind(struct Qdisc *sch, unsigned long parent, u32 classid)
 340 {
 341         return prio_get(sch, classid);
 342 }
 343
 344
 345 static void prio_put(struct Qdisc *q, unsigned long cl)
 346 {
 347         return;
 348 }
 349
 350 static int prio_change(struct Qdisc *sch, u32 handle, u32 parent, struct rtattr **tca, unsigned long *arg)
 351 {
 352         unsigned long cl = *arg;
 353         struct prio_sched_data *q = qdisc_priv(sch);
 354
 355         if (cl - 1 > q->bands)
 356                 return -ENOENT;
 357         return 0;
 358 }
 359
 360 static int prio_delete(struct Qdisc *sch, unsigned long cl)
 361 {
 362         struct prio_sched_data *q = qdisc_priv(sch);
 363         if (cl - 1 > q->bands)
 364                 return -ENOENT;
 365         return 0;
 366 }
 367
 368
 369 static int prio_dump_class(struct Qdisc *sch, unsigned long cl, struct sk_buff *skb,
 370                            struct tcmsg *tcm)
 371 {
 372         struct prio_sched_data *q = qdisc_priv(sch);
 373
 374         if (cl - 1 > q->bands)
 375                 return -ENOENT;
 376         tcm->tcm_handle |= TC_H_MIN(cl);
 377         if (q->queues[cl-1])
 378                 tcm->tcm_info = q->queues[cl-1]->handle;
 379         return 0;
 380 }
 381
 382 static void prio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 383 {
 384         struct prio_sched_data *q = qdisc_priv(sch);
 385         int prio;
 386
 387         if (arg->stop)
 388                 return;
 389
 390         for (prio = 0; prio < q->bands; prio++) {
 391                 if (arg->count < arg->skip) {
 392                         arg->count++;
 393                         continue;
 394                 }
 395                 if (arg->fn(sch, prio+1, arg) < 0) {
 396                         arg->stop = 1;
 397                         break;
 398                 }
 399                 arg->count++;
 400         }
 401 }
 402
 403 static struct tcf_proto ** prio_find_tcf(struct Qdisc *sch, unsigned long cl)
 404 {
 405         struct prio_sched_data *q = qdisc_priv(sch);
 406
 407         if (cl)
 408                 return NULL;
 409         return &q->filter_list;
 410 }
 411
 412 static struct Qdisc_class_ops prio_class_ops = {
 413         .graft          =       prio_graft,
 414         .leaf           =       prio_leaf,
 415         .get            =       prio_get,
 416         .put            =       prio_put,
 417         .change         =       prio_change,
 418         .delete         =       prio_delete,
 419         .walk           =       prio_walk,
 420         .tcf_chain      =       prio_find_tcf,
 421         .bind_tcf       =       prio_bind,
 422         .unbind_tcf     =       prio_put,
 423         .dump           =       prio_dump_class,
 424 };
 425
 426 static struct Qdisc_ops prio_qdisc_ops = {
 427         .next           =       NULL,
 428         .cl_ops         =       &prio_class_ops,
 429         .id             =       "prio",
 430         .priv_size      =       sizeof(struct prio_sched_data),
 431         .enqueue        =       prio_enqueue,
 432         .dequeue        =       prio_dequeue,
 433         .requeue        =       prio_requeue,
 434         .drop           =       prio_drop,
 435         .init           =       prio_init,
 436         .reset          =       prio_reset,
 437         .destroy        =       prio_destroy,
 438         .change         =       prio_tune,
 439         .dump           =       prio_dump,
 440         .owner          =       THIS_MODULE,
 441 };
 442
 443 static int __init prio_module_init(void)
 444 {
 445         return register_qdisc(&prio_qdisc_ops);
 446 }
 447
 448 static void __exit prio_module_exit(void)
 449 {
 450         unregister_qdisc(&prio_qdisc_ops);
 451 }
 452
 453 module_init(prio_module_init)
 454 module_exit(prio_module_exit)
 455
 456 MODULE_LICENSE("GPL");