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 <linux/bitops.h>
  14 #include <net/pkt_sched.h>
  15 #include <net/dsfield.h>
  16 #include <net/inet_ecn.h>
  17 #include <asm/byteorder.h>
  18
  19 /*
  20  * classid      class           marking
  21  * -------      -----           -------
  22  *   n/a          0             n/a
  23  *   x:0          1             use entry [0]
  24  *   ...         ...            ...
  25  *   x:y y>0     y+1            use entry [y]
  26  *   ...         ...            ...
  27  * x:indices-1  indices         use entry [indices-1]
  28  *   ...         ...            ...
  29  *   x:y         y+1            use entry [y & (indices-1)]
  30  *   ...         ...            ...
  31  * 0xffff       0x10000         use entry [indices-1]
  32  */
  33
  34
  35 #define NO_DEFAULT_INDEX        (1 << 16)
  36
  37 struct dsmark_qdisc_data {
  38         struct Qdisc            *q;
  39         struct tcf_proto        *filter_list;
  40         u8                      *mask;  /* "owns" the array */
  41         u8                      *value;
  42         u16                     indices;
  43         u32                     default_index;  /* index range is 0...0xffff */
  44         int                     set_tc_index;
  45 };
  46
  47 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
  48 {
  49         return (index <= p->indices && index > 0);
  50 }
  51
  52 /* ------------------------- Class/flow operations ------------------------- */
  53
  54 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
  55                         struct Qdisc *new, struct Qdisc **old)
  56 {
  57         struct dsmark_qdisc_data *p = qdisc_priv(sch);
  58
  59         pr_debug("dsmark_graft(sch %p,[qdisc %p],new %p,old %p)\n",
  60                 sch, p, new, old);
  61
  62         if (new == NULL) {
  63                 new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
  64                                         &pfifo_qdisc_ops,
  65                                         sch->handle);
  66                 if (new == NULL)
  67                         new = &noop_qdisc;
  68         }
  69
  70         sch_tree_lock(sch);
  71         *old = xchg(&p->q, new);
  72         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
  73         qdisc_reset(*old);
  74         sch_tree_unlock(sch);
  75
  76         return 0;
  77 }
  78
  79 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
  80 {
  81         struct dsmark_qdisc_data *p = qdisc_priv(sch);
  82         return p->q;
  83 }
  84
  85 static unsigned long dsmark_get(struct Qdisc *sch, u32 classid)
  86 {
  87         pr_debug("dsmark_get(sch %p,[qdisc %p],classid %x)\n",
  88                 sch, qdisc_priv(sch), classid);
  89
  90         return TC_H_MIN(classid) + 1;
  91 }
  92
  93 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
  94                                         unsigned long parent, u32 classid)
  95 {
  96         return dsmark_get(sch, classid);
  97 }
  98
  99 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
 100 {
 101 }
 102
 103 static const struct nla_policy dsmark_policy[TCA_DSMARK_MAX + 1] = {
 104         [TCA_DSMARK_INDICES]            = { .type = NLA_U16 },
 105         [TCA_DSMARK_DEFAULT_INDEX]      = { .type = NLA_U16 },
 106         [TCA_DSMARK_SET_TC_INDEX]       = { .type = NLA_FLAG },
 107         [TCA_DSMARK_MASK]               = { .type = NLA_U8 },
 108         [TCA_DSMARK_VALUE]              = { .type = NLA_U8 },
 109 };
 110
 111 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
 112                          struct nlattr **tca, unsigned long *arg)
 113 {
 114         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 115         struct nlattr *opt = tca[TCA_OPTIONS];
 116         struct nlattr *tb[TCA_DSMARK_MAX + 1];
 117         int err = -EINVAL;
 118         u8 mask = 0;
 119
 120         pr_debug("dsmark_change(sch %p,[qdisc %p],classid %x,parent %x),"
 121                 "arg 0x%lx\n", sch, p, classid, parent, *arg);
 122
 123         if (!dsmark_valid_index(p, *arg)) {
 124                 err = -ENOENT;
 125                 goto errout;
 126         }
 127
 128         if (!opt)
 129                 goto errout;
 130
 131         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
 132         if (err < 0)
 133                 goto errout;
 134
 135         if (tb[TCA_DSMARK_MASK])
 136                 mask = nla_get_u8(tb[TCA_DSMARK_MASK]);
 137
 138         if (tb[TCA_DSMARK_VALUE])
 139                 p->value[*arg-1] = nla_get_u8(tb[TCA_DSMARK_VALUE]);
 140
 141         if (tb[TCA_DSMARK_MASK])
 142                 p->mask[*arg-1] = mask;
 143
 144         err = 0;
 145
 146 errout:
 147         return err;
 148 }
 149
 150 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
 151 {
 152         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 153
 154         if (!dsmark_valid_index(p, arg))
 155                 return -EINVAL;
 156
 157         p->mask[arg-1] = 0xff;
 158         p->value[arg-1] = 0;
 159
 160         return 0;
 161 }
 162
 163 static void dsmark_walk(struct Qdisc *sch, struct qdisc_walker *walker)
 164 {
 165         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 166         int i;
 167
 168         pr_debug("dsmark_walk(sch %p,[qdisc %p],walker %p)\n", sch, p, walker);
 169
 170         if (walker->stop)
 171                 return;
 172
 173         for (i = 0; i < p->indices; i++) {
 174                 if (p->mask[i] == 0xff && !p->value[i])
 175                         goto ignore;
 176                 if (walker->count >= walker->skip) {
 177                         if (walker->fn(sch, i+1, walker) < 0) {
 178                                 walker->stop = 1;
 179                                 break;
 180                         }
 181                 }
 182 ignore:
 183                 walker->count++;
 184         }
 185 }
 186
 187 static inline struct tcf_proto **dsmark_find_tcf(struct Qdisc *sch,
 188                                                  unsigned long cl)
 189 {
 190         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 191         return &p->filter_list;
 192 }
 193
 194 /* --------------------------- Qdisc operations ---------------------------- */
 195
 196 static int dsmark_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 197 {
 198         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 199         int err;
 200
 201         pr_debug("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
 202
 203         if (p->set_tc_index) {
 204                 switch (skb->protocol) {
 205                 case htons(ETH_P_IP):
 206                         if (skb_cow_head(skb, sizeof(struct iphdr)))
 207                                 goto drop;
 208
 209                         skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
 210                                 & ~INET_ECN_MASK;
 211                         break;
 212
 213                 case htons(ETH_P_IPV6):
 214                         if (skb_cow_head(skb, sizeof(struct ipv6hdr)))
 215                                 goto drop;
 216
 217                         skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
 218                                 & ~INET_ECN_MASK;
 219                         break;
 220                 default:
 221                         skb->tc_index = 0;
 222                         break;
 223                 }
 224         }
 225
 226         if (TC_H_MAJ(skb->priority) == sch->handle)
 227                 skb->tc_index = TC_H_MIN(skb->priority);
 228         else {
 229                 struct tcf_result res;
 230                 int result = tc_classify(skb, p->filter_list, &res);
 231
 232                 pr_debug("result %d class 0x%04x\n", result, res.classid);
 233
 234                 switch (result) {
 235 #ifdef CONFIG_NET_CLS_ACT
 236                 case TC_ACT_QUEUED:
 237                 case TC_ACT_STOLEN:
 238                         kfree_skb(skb);
 239                         return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 240
 241                 case TC_ACT_SHOT:
 242                         goto drop;
 243 #endif
 244                 case TC_ACT_OK:
 245                         skb->tc_index = TC_H_MIN(res.classid);
 246                         break;
 247
 248                 default:
 249                         if (p->default_index != NO_DEFAULT_INDEX)
 250                                 skb->tc_index = p->default_index;
 251                         break;
 252                 }
 253         }
 254
 255         err = qdisc_enqueue(skb, p->q);
 256         if (err != NET_XMIT_SUCCESS) {
 257                 if (net_xmit_drop_count(err))
 258                         sch->qstats.drops++;
 259                 return err;
 260         }
 261
 262         sch->bstats.bytes += qdisc_pkt_len(skb);
 263         sch->bstats.packets++;
 264         sch->q.qlen++;
 265
 266         return NET_XMIT_SUCCESS;
 267
 268 drop:
 269         kfree_skb(skb);
 270         sch->qstats.drops++;
 271         return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 272 }
 273
 274 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
 275 {
 276         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 277         struct sk_buff *skb;
 278         u32 index;
 279
 280         pr_debug("dsmark_dequeue(sch %p,[qdisc %p])\n", sch, p);
 281
 282         skb = p->q->ops->dequeue(p->q);
 283         if (skb == NULL)
 284                 return NULL;
 285
 286         sch->q.qlen--;
 287
 288         index = skb->tc_index & (p->indices - 1);
 289         pr_debug("index %d->%d\n", skb->tc_index, index);
 290
 291         switch (skb->protocol) {
 292         case htons(ETH_P_IP):
 293                 ipv4_change_dsfield(ip_hdr(skb), p->mask[index],
 294                                     p->value[index]);
 295                         break;
 296         case htons(ETH_P_IPV6):
 297                 ipv6_change_dsfield(ipv6_hdr(skb), p->mask[index],
 298                                     p->value[index]);
 299                         break;
 300         default:
 301                 /*
 302                  * Only complain if a change was actually attempted.
 303                  * This way, we can send non-IP traffic through dsmark
 304                  * and don't need yet another qdisc as a bypass.
 305                  */
 306                 if (p->mask[index] != 0xff || p->value[index])
 307                         printk(KERN_WARNING
 308                                "dsmark_dequeue: unsupported protocol %d\n",
 309                                ntohs(skb->protocol));
 310                 break;
 311         }
 312
 313         return skb;
 314 }
 315
 316 static int dsmark_requeue(struct sk_buff *skb, struct Qdisc *sch)
 317 {
 318         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 319         int err;
 320
 321         pr_debug("dsmark_requeue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
 322
 323         err = p->q->ops->requeue(skb, p->q);
 324         if (err != NET_XMIT_SUCCESS) {
 325                 if (net_xmit_drop_count(err))
 326                         sch->qstats.drops++;
 327                 return err;
 328         }
 329
 330         sch->q.qlen++;
 331         sch->qstats.requeues++;
 332
 333         return NET_XMIT_SUCCESS;
 334 }
 335
 336 static unsigned int dsmark_drop(struct Qdisc *sch)
 337 {
 338         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 339         unsigned int len;
 340
 341         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
 342
 343         if (p->q->ops->drop == NULL)
 344                 return 0;
 345
 346         len = p->q->ops->drop(p->q);
 347         if (len)
 348                 sch->q.qlen--;
 349
 350         return len;
 351 }
 352
 353 static int dsmark_init(struct Qdisc *sch, struct nlattr *opt)
 354 {
 355         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 356         struct nlattr *tb[TCA_DSMARK_MAX + 1];
 357         int err = -EINVAL;
 358         u32 default_index = NO_DEFAULT_INDEX;
 359         u16 indices;
 360         u8 *mask;
 361
 362         pr_debug("dsmark_init(sch %p,[qdisc %p],opt %p)\n", sch, p, opt);
 363
 364         if (!opt)
 365                 goto errout;
 366
 367         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
 368         if (err < 0)
 369                 goto errout;
 370
 371         err = -EINVAL;
 372         indices = nla_get_u16(tb[TCA_DSMARK_INDICES]);
 373
 374         if (hweight32(indices) != 1)
 375                 goto errout;
 376
 377         if (tb[TCA_DSMARK_DEFAULT_INDEX])
 378                 default_index = nla_get_u16(tb[TCA_DSMARK_DEFAULT_INDEX]);
 379
 380         mask = kmalloc(indices * 2, GFP_KERNEL);
 381         if (mask == NULL) {
 382                 err = -ENOMEM;
 383                 goto errout;
 384         }
 385
 386         p->mask = mask;
 387         memset(p->mask, 0xff, indices);
 388
 389         p->value = p->mask + indices;
 390         memset(p->value, 0, indices);
 391
 392         p->indices = indices;
 393         p->default_index = default_index;
 394         p->set_tc_index = nla_get_flag(tb[TCA_DSMARK_SET_TC_INDEX]);
 395
 396         p->q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
 397                                  &pfifo_qdisc_ops, sch->handle);
 398         if (p->q == NULL)
 399                 p->q = &noop_qdisc;
 400
 401         pr_debug("dsmark_init: qdisc %p\n", p->q);
 402
 403         err = 0;
 404 errout:
 405         return err;
 406 }
 407
 408 static void dsmark_reset(struct Qdisc *sch)
 409 {
 410         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 411
 412         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
 413         qdisc_reset(p->q);
 414         sch->q.qlen = 0;
 415 }
 416
 417 static void dsmark_destroy(struct Qdisc *sch)
 418 {
 419         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 420
 421         pr_debug("dsmark_destroy(sch %p,[qdisc %p])\n", sch, p);
 422
 423         tcf_destroy_chain(&p->filter_list);
 424         qdisc_destroy(p->q);
 425         kfree(p->mask);
 426 }
 427
 428 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
 429                              struct sk_buff *skb, struct tcmsg *tcm)
 430 {
 431         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 432         struct nlattr *opts = NULL;
 433
 434         pr_debug("dsmark_dump_class(sch %p,[qdisc %p],class %ld\n", sch, p, cl);
 435
 436         if (!dsmark_valid_index(p, cl))
 437                 return -EINVAL;
 438
 439         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl-1);
 440         tcm->tcm_info = p->q->handle;
 441
 442         opts = nla_nest_start(skb, TCA_OPTIONS);
 443         if (opts == NULL)
 444                 goto nla_put_failure;
 445         NLA_PUT_U8(skb, TCA_DSMARK_MASK, p->mask[cl-1]);
 446         NLA_PUT_U8(skb, TCA_DSMARK_VALUE, p->value[cl-1]);
 447
 448         return nla_nest_end(skb, opts);
 449
 450 nla_put_failure:
 451         nla_nest_cancel(skb, opts);
 452         return -EMSGSIZE;
 453 }
 454
 455 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
 456 {
 457         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 458         struct nlattr *opts = NULL;
 459
 460         opts = nla_nest_start(skb, TCA_OPTIONS);
 461         if (opts == NULL)
 462                 goto nla_put_failure;
 463         NLA_PUT_U16(skb, TCA_DSMARK_INDICES, p->indices);
 464
 465         if (p->default_index != NO_DEFAULT_INDEX)
 466                 NLA_PUT_U16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index);
 467
 468         if (p->set_tc_index)
 469                 NLA_PUT_FLAG(skb, TCA_DSMARK_SET_TC_INDEX);
 470
 471         return nla_nest_end(skb, opts);
 472
 473 nla_put_failure:
 474         nla_nest_cancel(skb, opts);
 475         return -EMSGSIZE;
 476 }
 477
 478 static const struct Qdisc_class_ops dsmark_class_ops = {
 479         .graft          =       dsmark_graft,
 480         .leaf           =       dsmark_leaf,
 481         .get            =       dsmark_get,
 482         .put            =       dsmark_put,
 483         .change         =       dsmark_change,
 484         .delete         =       dsmark_delete,
 485         .walk           =       dsmark_walk,
 486         .tcf_chain      =       dsmark_find_tcf,
 487         .bind_tcf       =       dsmark_bind_filter,
 488         .unbind_tcf     =       dsmark_put,
 489         .dump           =       dsmark_dump_class,
 490 };
 491
 492 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
 493         .next           =       NULL,
 494         .cl_ops         =       &dsmark_class_ops,
 495         .id             =       "dsmark",
 496         .priv_size      =       sizeof(struct dsmark_qdisc_data),
 497         .enqueue        =       dsmark_enqueue,
 498         .dequeue        =       dsmark_dequeue,
 499         .requeue        =       dsmark_requeue,
 500         .drop           =       dsmark_drop,
 501         .init           =       dsmark_init,
 502         .reset          =       dsmark_reset,
 503         .destroy        =       dsmark_destroy,
 504         .change         =       NULL,
 505         .dump           =       dsmark_dump,
 506         .owner          =       THIS_MODULE,
 507 };
 508
 509 static int __init dsmark_module_init(void)
 510 {
 511         return register_qdisc(&dsmark_qdisc_ops);
 512 }
 513
 514 static void __exit dsmark_module_exit(void)
 515 {
 516         unregister_qdisc(&dsmark_qdisc_ops);
 517 }
 518
 519 module_init(dsmark_module_init)
 520 module_exit(dsmark_module_exit)
 521
 522 MODULE_LICENSE("GPL");