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 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->bstats.bytes += skb->len;
 111                 sch->bstats.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->qstats.drops++;
 121 #ifdef CONFIG_NET_CLS_ACT
 122         } else {
 123                 sch->qstats.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                 sch->qstats.requeues++;
 143                 return 0;
 144         }
 145 dropped:
 146         sch->qstats.drops++;
 147         return NET_XMIT_DROP;
 148 }
 149
 150
 151 static struct sk_buff *
 152 prio_dequeue(struct Qdisc* sch)
 153 {
 154         struct sk_buff *skb;
 155         struct prio_sched_data *q = qdisc_priv(sch);
 156         int prio;
 157         struct Qdisc *qdisc;
 158
 159         for (prio = 0; prio < q->bands; prio++) {
 160                 qdisc = q->queues[prio];
 161                 skb = qdisc->dequeue(qdisc);
 162                 if (skb) {
 163                         sch->q.qlen--;
 164                         return skb;
 165                 }
 166         }
 167         return NULL;
 168
 169 }
 170
 171 static unsigned int prio_drop(struct Qdisc* sch)
 172 {
 173         struct prio_sched_data *q = qdisc_priv(sch);
 174         int prio;
 175         unsigned int len;
 176         struct Qdisc *qdisc;
 177
 178         for (prio = q->bands-1; prio >= 0; prio--) {
 179                 qdisc = q->queues[prio];
 180                 if ((len = qdisc->ops->drop(qdisc)) != 0) {
 181                         sch->q.qlen--;
 182                         return len;
 183                 }
 184         }
 185         return 0;
 186 }
 187
 188
 189 static void
 190 prio_reset(struct Qdisc* sch)
 191 {
 192         int prio;
 193         struct prio_sched_data *q = qdisc_priv(sch);
 194
 195         for (prio=0; prio<q->bands; prio++)
 196                 qdisc_reset(q->queues[prio]);
 197         sch->q.qlen = 0;
 198 }
 199
 200 static void
 201 prio_destroy(struct Qdisc* sch)
 202 {
 203         int prio;
 204         struct prio_sched_data *q = qdisc_priv(sch);
 205         struct tcf_proto *tp;
 206
 207         while ((tp = q->filter_list) != NULL) {
 208                 q->filter_list = tp->next;
 209                 tcf_destroy(tp);
 210         }
 211
 212         for (prio=0; prio<q->bands; prio++)
 213                 qdisc_destroy(q->queues[prio]);
 214 }
 215
 216 static int prio_tune(struct Qdisc *sch, struct rtattr *opt)
 217 {
 218         struct prio_sched_data *q = qdisc_priv(sch);
 219         struct tc_prio_qopt *qopt = RTA_DATA(opt);
 220         int i;
 221
 222         if (opt->rta_len < RTA_LENGTH(sizeof(*qopt)))
 223                 return -EINVAL;
 224         if (qopt->bands > TCQ_PRIO_BANDS || qopt->bands < 2)
 225                 return -EINVAL;
 226
 227         for (i=0; i<=TC_PRIO_MAX; i++) {
 228                 if (qopt->priomap[i] >= qopt->bands)
 229                         return -EINVAL;
 230         }
 231
 232         sch_tree_lock(sch);
 233         q->bands = qopt->bands;
 234         memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
 235
 236         for (i=q->bands; i<TCQ_PRIO_BANDS; i++) {
 237                 struct Qdisc *child = xchg(&q->queues[i], &noop_qdisc);
 238                 if (child != &noop_qdisc)
 239                         qdisc_destroy(child);
 240         }
 241         sch_tree_unlock(sch);
 242
 243         for (i=0; i<=TC_PRIO_MAX; i++) {
 244                 int band = q->prio2band[i];
 245                 if (q->queues[band] == &noop_qdisc) {
 246                         struct Qdisc *child;
 247                         child = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
 248                         if (child) {
 249                                 sch_tree_lock(sch);
 250                                 child = xchg(&q->queues[band], child);
 251
 252                                 if (child != &noop_qdisc)
 253                                         qdisc_destroy(child);
 254                                 sch_tree_unlock(sch);
 255                         }
 256                 }
 257         }
 258         return 0;
 259 }
 260
 261 static int prio_init(struct Qdisc *sch, struct rtattr *opt)
 262 {
 263         struct prio_sched_data *q = qdisc_priv(sch);
 264         int i;
 265
 266         for (i=0; i<TCQ_PRIO_BANDS; i++)
 267                 q->queues[i] = &noop_qdisc;
 268
 269         if (opt == NULL) {
 270                 return -EINVAL;
 271         } else {
 272                 int err;
 273
 274                 if ((err= prio_tune(sch, opt)) != 0)
 275                         return err;
 276         }
 277         return 0;
 278 }
 279
 280 static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)
 281 {
 282         struct prio_sched_data *q = qdisc_priv(sch);
 283         unsigned char    *b = skb->tail;
 284         struct tc_prio_qopt opt;
 285
 286         opt.bands = q->bands;
 287         memcpy(&opt.priomap, q->prio2band, TC_PRIO_MAX+1);
 288         RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 289         return skb->len;
 290
 291 rtattr_failure:
 292         skb_trim(skb, b - skb->data);
 293         return -1;
 294 }
 295
 296 static int prio_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 297                       struct Qdisc **old)
 298 {
 299         struct prio_sched_data *q = qdisc_priv(sch);
 300         unsigned long band = arg - 1;
 301
 302         if (band >= q->bands)
 303                 return -EINVAL;
 304
 305         if (new == NULL)
 306                 new = &noop_qdisc;
 307
 308         sch_tree_lock(sch);
 309         *old = q->queues[band];
 310         q->queues[band] = new;
 311         sch->q.qlen -= (*old)->q.qlen;
 312         qdisc_reset(*old);
 313         sch_tree_unlock(sch);
 314
 315         return 0;
 316 }
 317
 318 static struct Qdisc *
 319 prio_leaf(struct Qdisc *sch, unsigned long arg)
 320 {
 321         struct prio_sched_data *q = qdisc_priv(sch);
 322         unsigned long band = arg - 1;
 323
 324         if (band >= q->bands)
 325                 return NULL;
 326
 327         return q->queues[band];
 328 }
 329
 330 static unsigned long prio_get(struct Qdisc *sch, u32 classid)
 331 {
 332         struct prio_sched_data *q = qdisc_priv(sch);
 333         unsigned long band = TC_H_MIN(classid);
 334
 335         if (band - 1 >= q->bands)
 336                 return 0;
 337         return band;
 338 }
 339
 340 static unsigned long prio_bind(struct Qdisc *sch, unsigned long parent, u32 classid)
 341 {
 342         return prio_get(sch, classid);
 343 }
 344
 345
 346 static void prio_put(struct Qdisc *q, unsigned long cl)
 347 {
 348         return;
 349 }
 350
 351 static int prio_change(struct Qdisc *sch, u32 handle, u32 parent, struct rtattr **tca, unsigned long *arg)
 352 {
 353         unsigned long cl = *arg;
 354         struct prio_sched_data *q = qdisc_priv(sch);
 355
 356         if (cl - 1 > q->bands)
 357                 return -ENOENT;
 358         return 0;
 359 }
 360
 361 static int prio_delete(struct Qdisc *sch, unsigned long cl)
 362 {
 363         struct prio_sched_data *q = qdisc_priv(sch);
 364         if (cl - 1 > q->bands)
 365                 return -ENOENT;
 366         return 0;
 367 }
 368
 369
 370 static int prio_dump_class(struct Qdisc *sch, unsigned long cl, struct sk_buff *skb,
 371                            struct tcmsg *tcm)
 372 {
 373         struct prio_sched_data *q = qdisc_priv(sch);
 374
 375         if (cl - 1 > q->bands)
 376                 return -ENOENT;
 377         tcm->tcm_handle |= TC_H_MIN(cl);
 378         if (q->queues[cl-1])
 379                 tcm->tcm_info = q->queues[cl-1]->handle;
 380         return 0;
 381 }
 382
 383 static void prio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 384 {
 385         struct prio_sched_data *q = qdisc_priv(sch);
 386         int prio;
 387
 388         if (arg->stop)
 389                 return;
 390
 391         for (prio = 0; prio < q->bands; prio++) {
 392                 if (arg->count < arg->skip) {
 393                         arg->count++;
 394                         continue;
 395                 }
 396                 if (arg->fn(sch, prio+1, arg) < 0) {
 397                         arg->stop = 1;
 398                         break;
 399                 }
 400                 arg->count++;
 401         }
 402 }
 403
 404 static struct tcf_proto ** prio_find_tcf(struct Qdisc *sch, unsigned long cl)
 405 {
 406         struct prio_sched_data *q = qdisc_priv(sch);
 407
 408         if (cl)
 409                 return NULL;
 410         return &q->filter_list;
 411 }
 412
 413 static struct Qdisc_class_ops prio_class_ops = {
 414         .graft          =       prio_graft,
 415         .leaf           =       prio_leaf,
 416         .get            =       prio_get,
 417         .put            =       prio_put,
 418         .change         =       prio_change,
 419         .delete         =       prio_delete,
 420         .walk           =       prio_walk,
 421         .tcf_chain      =       prio_find_tcf,
 422         .bind_tcf       =       prio_bind,
 423         .unbind_tcf     =       prio_put,
 424         .dump           =       prio_dump_class,
 425 };
 426
 427 static struct Qdisc_ops prio_qdisc_ops = {
 428         .next           =       NULL,
 429         .cl_ops         =       &prio_class_ops,
 430         .id             =       "prio",
 431         .priv_size      =       sizeof(struct prio_sched_data),
 432         .enqueue        =       prio_enqueue,
 433         .dequeue        =       prio_dequeue,
 434         .requeue        =       prio_requeue,
 435         .drop           =       prio_drop,
 436         .init           =       prio_init,
 437         .reset          =       prio_reset,
 438         .destroy        =       prio_destroy,
 439         .change         =       prio_tune,
 440         .dump           =       prio_dump,
 441         .owner          =       THIS_MODULE,
 442 };
 443
 444 static int __init prio_module_init(void)
 445 {
 446         return register_qdisc(&prio_qdisc_ops);
 447 }
 448
 449 static void __exit prio_module_exit(void)
 450 {
 451         unregister_qdisc(&prio_qdisc_ops);
 452 }
 453
 454 module_init(prio_module_init)
 455 module_exit(prio_module_exit)
 456
 457 MODULE_LICENSE("GPL");