err.no Git - linux-2.6/blob - net/sched/sch_dsmark.c

   1 /* net/sched/sch_dsmark.c - Differentiated Services field marker */
   2
   3 /* Written 1998-2000 by Werner Almesberger, EPFL ICA */
   4
   5
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/types.h>
   9 #include <linux/string.h>
  10 #include <linux/errno.h>
  11 #include <linux/skbuff.h>
  12 #include <linux/rtnetlink.h>
  13 #include <net/pkt_sched.h>
  14 #include <net/dsfield.h>
  15 #include <net/inet_ecn.h>
  16 #include <asm/byteorder.h>
  17
  18 /*
  19  * classid      class           marking
  20  * -------      -----           -------
  21  *   n/a          0             n/a
  22  *   x:0          1             use entry [0]
  23  *   ...         ...            ...
  24  *   x:y y>0     y+1            use entry [y]
  25  *   ...         ...            ...
  26  * x:indices-1  indices         use entry [indices-1]
  27  *   ...         ...            ...
  28  *   x:y         y+1            use entry [y & (indices-1)]
  29  *   ...         ...            ...
  30  * 0xffff       0x10000         use entry [indices-1]
  31  */
  32
  33
  34 #define NO_DEFAULT_INDEX        (1 << 16)
  35
  36 struct dsmark_qdisc_data {
  37         struct Qdisc            *q;
  38         struct tcf_proto        *filter_list;
  39         u8                      *mask;  /* "owns" the array */
  40         u8                      *value;
  41         u16                     indices;
  42         u32                     default_index;  /* index range is 0...0xffff */
  43         int                     set_tc_index;
  44 };
  45
  46 static inline int dsmark_valid_indices(u16 indices)
  47 {
  48         while (indices != 1) {
  49                 if (indices & 1)
  50                         return 0;
  51                 indices >>= 1;
  52         }
  53
  54         return 1;
  55 }
  56
  57 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
  58 {
  59         return (index <= p->indices && index > 0);
  60 }
  61
  62 /* ------------------------- Class/flow operations ------------------------- */
  63
  64 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
  65                         struct Qdisc *new, struct Qdisc **old)
  66 {
  67         struct dsmark_qdisc_data *p = qdisc_priv(sch);
  68
  69         pr_debug("dsmark_graft(sch %p,[qdisc %p],new %p,old %p)\n",
  70                 sch, p, new, old);
  71
  72         if (new == NULL) {
  73                 new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
  74                                         sch->handle);
  75                 if (new == NULL)
  76                         new = &noop_qdisc;
  77         }
  78
  79         sch_tree_lock(sch);
  80         *old = xchg(&p->q, new);
  81         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
  82         qdisc_reset(*old);
  83         sch_tree_unlock(sch);
  84
  85         return 0;
  86 }
  87
  88 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
  89 {
  90         struct dsmark_qdisc_data *p = qdisc_priv(sch);
  91         return p->q;
  92 }
  93
  94 static unsigned long dsmark_get(struct Qdisc *sch, u32 classid)
  95 {
  96         pr_debug("dsmark_get(sch %p,[qdisc %p],classid %x)\n",
  97                 sch, qdisc_priv(sch), classid);
  98
  99         return TC_H_MIN(classid) + 1;
 100 }
 101
 102 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
 103                                         unsigned long parent, u32 classid)
 104 {
 105         return dsmark_get(sch, classid);
 106 }
 107
 108 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
 109 {
 110 }
 111
 112 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
 113                          struct rtattr **tca, unsigned long *arg)
 114 {
 115         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 116         struct rtattr *opt = tca[TCA_OPTIONS-1];
 117         struct rtattr *tb[TCA_DSMARK_MAX];
 118         int err = -EINVAL;
 119         u8 mask = 0;
 120
 121         pr_debug("dsmark_change(sch %p,[qdisc %p],classid %x,parent %x),"
 122                 "arg 0x%lx\n", sch, p, classid, parent, *arg);
 123
 124         if (!dsmark_valid_index(p, *arg)) {
 125                 err = -ENOENT;
 126                 goto rtattr_failure;
 127         }
 128
 129         if (!opt || rtattr_parse_nested(tb, TCA_DSMARK_MAX, opt))
 130                 goto rtattr_failure;
 131
 132         if (tb[TCA_DSMARK_MASK-1])
 133                 mask = RTA_GET_U8(tb[TCA_DSMARK_MASK-1]);
 134
 135         if (tb[TCA_DSMARK_VALUE-1])
 136                 p->value[*arg-1] = RTA_GET_U8(tb[TCA_DSMARK_VALUE-1]);
 137
 138         if (tb[TCA_DSMARK_MASK-1])
 139                 p->mask[*arg-1] = mask;
 140
 141         err = 0;
 142
 143 rtattr_failure:
 144         return err;
 145 }
 146
 147 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
 148 {
 149         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 150
 151         if (!dsmark_valid_index(p, arg))
 152                 return -EINVAL;
 153
 154         p->mask[arg-1] = 0xff;
 155         p->value[arg-1] = 0;
 156
 157         return 0;
 158 }
 159
 160 static void dsmark_walk(struct Qdisc *sch,struct qdisc_walker *walker)
 161 {
 162         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 163         int i;
 164
 165         pr_debug("dsmark_walk(sch %p,[qdisc %p],walker %p)\n", sch, p, walker);
 166
 167         if (walker->stop)
 168                 return;
 169
 170         for (i = 0; i < p->indices; i++) {
 171                 if (p->mask[i] == 0xff && !p->value[i])
 172                         goto ignore;
 173                 if (walker->count >= walker->skip) {
 174                         if (walker->fn(sch, i+1, walker) < 0) {
 175                                 walker->stop = 1;
 176                                 break;
 177                         }
 178                 }
 179 ignore:
 180                 walker->count++;
 181         }
 182 }
 183
 184 static inline struct tcf_proto **dsmark_find_tcf(struct Qdisc *sch,unsigned long cl)
 185 {
 186         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 187         return &p->filter_list;
 188 }
 189
 190 /* --------------------------- Qdisc operations ---------------------------- */
 191
 192 static int dsmark_enqueue(struct sk_buff *skb,struct Qdisc *sch)
 193 {
 194         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 195         int err;
 196
 197         pr_debug("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
 198
 199         if (p->set_tc_index) {
 200                 /* FIXME: Safe with non-linear skbs? --RR */
 201                 switch (skb->protocol) {
 202                         case __constant_htons(ETH_P_IP):
 203                                 skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
 204                                         & ~INET_ECN_MASK;
 205                                 break;
 206                         case __constant_htons(ETH_P_IPV6):
 207                                 skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
 208                                         & ~INET_ECN_MASK;
 209                                 break;
 210                         default:
 211                                 skb->tc_index = 0;
 212                                 break;
 213                 }
 214         }
 215
 216         if (TC_H_MAJ(skb->priority) == sch->handle)
 217                 skb->tc_index = TC_H_MIN(skb->priority);
 218         else {
 219                 struct tcf_result res;
 220                 int result = tc_classify(skb, p->filter_list, &res);
 221
 222                 pr_debug("result %d class 0x%04x\n", result, res.classid);
 223
 224                 switch (result) {
 225 #ifdef CONFIG_NET_CLS_ACT
 226                 case TC_ACT_QUEUED:
 227                 case TC_ACT_STOLEN:
 228                         kfree_skb(skb);
 229                         return NET_XMIT_SUCCESS;
 230                 case TC_ACT_SHOT:
 231                         kfree_skb(skb);
 232                         sch->qstats.drops++;
 233                         return NET_XMIT_BYPASS;
 234 #endif
 235                 case TC_ACT_OK:
 236                         skb->tc_index = TC_H_MIN(res.classid);
 237                         break;
 238                 default:
 239                         if (p->default_index != NO_DEFAULT_INDEX)
 240                                 skb->tc_index = p->default_index;
 241                         break;
 242                 }
 243         }
 244
 245         err = p->q->enqueue(skb,p->q);
 246         if (err != NET_XMIT_SUCCESS) {
 247                 sch->qstats.drops++;
 248                 return err;
 249         }
 250
 251         sch->bstats.bytes += skb->len;
 252         sch->bstats.packets++;
 253         sch->q.qlen++;
 254
 255         return NET_XMIT_SUCCESS;
 256 }
 257
 258 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
 259 {
 260         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 261         struct sk_buff *skb;
 262         u32 index;
 263
 264         pr_debug("dsmark_dequeue(sch %p,[qdisc %p])\n", sch, p);
 265
 266         skb = p->q->ops->dequeue(p->q);
 267         if (skb == NULL)
 268                 return NULL;
 269
 270         sch->q.qlen--;
 271
 272         index = skb->tc_index & (p->indices - 1);
 273         pr_debug("index %d->%d\n", skb->tc_index, index);
 274
 275         switch (skb->protocol) {
 276                 case __constant_htons(ETH_P_IP):
 277                         ipv4_change_dsfield(ip_hdr(skb), p->mask[index],
 278                                             p->value[index]);
 279                         break;
 280                 case __constant_htons(ETH_P_IPV6):
 281                         ipv6_change_dsfield(ipv6_hdr(skb), p->mask[index],
 282                                             p->value[index]);
 283                         break;
 284                 default:
 285                         /*
 286                          * Only complain if a change was actually attempted.
 287                          * This way, we can send non-IP traffic through dsmark
 288                          * and don't need yet another qdisc as a bypass.
 289                          */
 290                         if (p->mask[index] != 0xff || p->value[index])
 291                                 printk(KERN_WARNING "dsmark_dequeue: "
 292                                        "unsupported protocol %d\n",
 293                                        ntohs(skb->protocol));
 294                         break;
 295         }
 296
 297         return skb;
 298 }
 299
 300 static int dsmark_requeue(struct sk_buff *skb,struct Qdisc *sch)
 301 {
 302         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 303         int err;
 304
 305         pr_debug("dsmark_requeue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
 306
 307         err = p->q->ops->requeue(skb, p->q);
 308         if (err != NET_XMIT_SUCCESS) {
 309                 sch->qstats.drops++;
 310                 return err;
 311         }
 312
 313         sch->q.qlen++;
 314         sch->qstats.requeues++;
 315
 316         return NET_XMIT_SUCCESS;
 317 }
 318
 319 static unsigned int dsmark_drop(struct Qdisc *sch)
 320 {
 321         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 322         unsigned int len;
 323
 324         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
 325
 326         if (p->q->ops->drop == NULL)
 327                 return 0;
 328
 329         len = p->q->ops->drop(p->q);
 330         if (len)
 331                 sch->q.qlen--;
 332
 333         return len;
 334 }
 335
 336 static int dsmark_init(struct Qdisc *sch, struct rtattr *opt)
 337 {
 338         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 339         struct rtattr *tb[TCA_DSMARK_MAX];
 340         int err = -EINVAL;
 341         u32 default_index = NO_DEFAULT_INDEX;
 342         u16 indices;
 343         u8 *mask;
 344
 345         pr_debug("dsmark_init(sch %p,[qdisc %p],opt %p)\n", sch, p, opt);
 346
 347         if (!opt || rtattr_parse_nested(tb, TCA_DSMARK_MAX, opt) < 0)
 348                 goto errout;
 349
 350         indices = RTA_GET_U16(tb[TCA_DSMARK_INDICES-1]);
 351         if (!indices || !dsmark_valid_indices(indices))
 352                 goto errout;
 353
 354         if (tb[TCA_DSMARK_DEFAULT_INDEX-1])
 355                 default_index = RTA_GET_U16(tb[TCA_DSMARK_DEFAULT_INDEX-1]);
 356
 357         mask = kmalloc(indices * 2, GFP_KERNEL);
 358         if (mask == NULL) {
 359                 err = -ENOMEM;
 360                 goto errout;
 361         }
 362
 363         p->mask = mask;
 364         memset(p->mask, 0xff, indices);
 365
 366         p->value = p->mask + indices;
 367         memset(p->value, 0, indices);
 368
 369         p->indices = indices;
 370         p->default_index = default_index;
 371         p->set_tc_index = RTA_GET_FLAG(tb[TCA_DSMARK_SET_TC_INDEX-1]);
 372
 373         p->q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops, sch->handle);
 374         if (p->q == NULL)
 375                 p->q = &noop_qdisc;
 376
 377         pr_debug("dsmark_init: qdisc %p\n", p->q);
 378
 379         err = 0;
 380 errout:
 381 rtattr_failure:
 382         return err;
 383 }
 384
 385 static void dsmark_reset(struct Qdisc *sch)
 386 {
 387         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 388
 389         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
 390         qdisc_reset(p->q);
 391         sch->q.qlen = 0;
 392 }
 393
 394 static void dsmark_destroy(struct Qdisc *sch)
 395 {
 396         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 397
 398         pr_debug("dsmark_destroy(sch %p,[qdisc %p])\n", sch, p);
 399
 400         tcf_destroy_chain(p->filter_list);
 401         qdisc_destroy(p->q);
 402         kfree(p->mask);
 403 }
 404
 405 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
 406                              struct sk_buff *skb, struct tcmsg *tcm)
 407 {
 408         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 409         struct rtattr *opts = NULL;
 410
 411         pr_debug("dsmark_dump_class(sch %p,[qdisc %p],class %ld\n", sch, p, cl);
 412
 413         if (!dsmark_valid_index(p, cl))
 414                 return -EINVAL;
 415
 416         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl-1);
 417         tcm->tcm_info = p->q->handle;
 418
 419         opts = RTA_NEST(skb, TCA_OPTIONS);
 420         RTA_PUT_U8(skb,TCA_DSMARK_MASK, p->mask[cl-1]);
 421         RTA_PUT_U8(skb,TCA_DSMARK_VALUE, p->value[cl-1]);
 422
 423         return RTA_NEST_END(skb, opts);
 424
 425 rtattr_failure:
 426         return RTA_NEST_CANCEL(skb, opts);
 427 }
 428
 429 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
 430 {
 431         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 432         struct rtattr *opts = NULL;
 433
 434         opts = RTA_NEST(skb, TCA_OPTIONS);
 435         RTA_PUT_U16(skb, TCA_DSMARK_INDICES, p->indices);
 436
 437         if (p->default_index != NO_DEFAULT_INDEX)
 438                 RTA_PUT_U16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index);
 439
 440         if (p->set_tc_index)
 441                 RTA_PUT_FLAG(skb, TCA_DSMARK_SET_TC_INDEX);
 442
 443         return RTA_NEST_END(skb, opts);
 444
 445 rtattr_failure:
 446         return RTA_NEST_CANCEL(skb, opts);
 447 }
 448
 449 static const struct Qdisc_class_ops dsmark_class_ops = {
 450         .graft          =       dsmark_graft,
 451         .leaf           =       dsmark_leaf,
 452         .get            =       dsmark_get,
 453         .put            =       dsmark_put,
 454         .change         =       dsmark_change,
 455         .delete         =       dsmark_delete,
 456         .walk           =       dsmark_walk,
 457         .tcf_chain      =       dsmark_find_tcf,
 458         .bind_tcf       =       dsmark_bind_filter,
 459         .unbind_tcf     =       dsmark_put,
 460         .dump           =       dsmark_dump_class,
 461 };
 462
 463 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
 464         .next           =       NULL,
 465         .cl_ops         =       &dsmark_class_ops,
 466         .id             =       "dsmark",
 467         .priv_size      =       sizeof(struct dsmark_qdisc_data),
 468         .enqueue        =       dsmark_enqueue,
 469         .dequeue        =       dsmark_dequeue,
 470         .requeue        =       dsmark_requeue,
 471         .drop           =       dsmark_drop,
 472         .init           =       dsmark_init,
 473         .reset          =       dsmark_reset,
 474         .destroy        =       dsmark_destroy,
 475         .change         =       NULL,
 476         .dump           =       dsmark_dump,
 477         .owner          =       THIS_MODULE,
 478 };
 479
 480 static int __init dsmark_module_init(void)
 481 {
 482         return register_qdisc(&dsmark_qdisc_ops);
 483 }
 484
 485 static void __exit dsmark_module_exit(void)
 486 {
 487         unregister_qdisc(&dsmark_qdisc_ops);
 488 }
 489
 490 module_init(dsmark_module_init)
 491 module_exit(dsmark_module_exit)
 492
 493 MODULE_LICENSE("GPL");