net/sched/em_meta.c

   1 /*
   2  * net/sched/em_meta.c  Metadata ematch
   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:     Thomas Graf <tgraf@suug.ch>
  10  *
  11  * ==========================================================================
  12  *
  13  *      The metadata ematch compares two meta objects where each object
  14  *      represents either a meta value stored in the kernel or a static
  15  *      value provided by userspace. The objects are not provided by
  16  *      userspace itself but rather a definition providing the information
  17  *      to build them. Every object is of a certain type which must be
  18  *      equal to the object it is being compared to.
  19  *
  20  *      The definition of a objects conists of the type (meta type), a
  21  *      identifier (meta id) and additional type specific information.
  22  *      The meta id is either TCF_META_TYPE_VALUE for values provided by
  23  *      userspace or a index to the meta operations table consisting of
  24  *      function pointers to type specific meta data collectors returning
  25  *      the value of the requested meta value.
  26  *
  27  *               lvalue                                   rvalue
  28  *            +-----------+                           +-----------+
  29  *            | type: INT |                           | type: INT |
  30  *       def  | id: DEV   |                           | id: VALUE |
  31  *            | data:     |                           | data: 3   |
  32  *            +-----------+                           +-----------+
  33  *                  |                                       |
  34  *                  ---> meta_ops[INT][DEV](...)            |
  35  *                            |                             |
  36  *                  -----------                             |
  37  *                  V                                       V
  38  *            +-----------+                           +-----------+
  39  *            | type: INT |                           | type: INT |
  40  *       obj  | id: DEV |                             | id: VALUE |
  41  *            | data: 2   |<--data got filled out     | data: 3   |
  42  *            +-----------+                           +-----------+
  43  *                  |                                         |
  44  *                  --------------> 2  equals 3 <--------------
  45  *
  46  *      This is a simplified schema, the complexity varies depending
  47  *      on the meta type. Obviously, the length of the data must also
  48  *      be provided for non-numeric types.
  49  *
  50  *      Additionally, type dependent modifiers such as shift operators
  51  *      or mask may be applied to extend the functionaliy. As of now,
  52  *      the variable length type supports shifting the byte string to
  53  *      the right, eating up any number of octets and thus supporting
  54  *      wildcard interface name comparisons such as "ppp%" matching
  55  *      ppp0..9.
  56  *
  57  *      NOTE: Certain meta values depend on other subsystems and are
  58  *            only available if that subsystem is enabled in the kernel.
  59  */
  60
  61 #include <linux/slab.h>
  62 #include <linux/module.h>
  63 #include <linux/types.h>
  64 #include <linux/kernel.h>
  65 #include <linux/sched.h>
  66 #include <linux/string.h>
  67 #include <linux/skbuff.h>
  68 #include <linux/random.h>
  69 #include <linux/if_vlan.h>
  70 #include <linux/tc_ematch/tc_em_meta.h>
  71 #include <net/dst.h>
  72 #include <net/route.h>
  73 #include <net/pkt_cls.h>
  74 #include <net/sock.h>
  75
  76 struct meta_obj {
  77         unsigned long           value;
  78         unsigned int            len;
  79 };
  80
  81 struct meta_value {
  82         struct tcf_meta_val     hdr;
  83         unsigned long           val;
  84         unsigned int            len;
  85 };
  86
  87 struct meta_match {
  88         struct meta_value       lvalue;
  89         struct meta_value       rvalue;
  90 };
  91
  92 static inline int meta_id(struct meta_value *v)
  93 {
  94         return TCF_META_ID(v->hdr.kind);
  95 }
  96
  97 static inline int meta_type(struct meta_value *v)
  98 {
  99         return TCF_META_TYPE(v->hdr.kind);
 100 }
 101
 102 #define META_COLLECTOR(FUNC) static void meta_##FUNC(struct sk_buff *skb, \
 103         struct tcf_pkt_info *info, struct meta_value *v, \
 104         struct meta_obj *dst, int *err)
 105
 106 /**************************************************************************
 107  * System status & misc
 108  **************************************************************************/
 109
 110 META_COLLECTOR(int_random)
 111 {
 112         get_random_bytes(&dst->value, sizeof(dst->value));
 113 }
 114
 115 static inline unsigned long fixed_loadavg(int load)
 116 {
 117         int rnd_load = load + (FIXED_1/200);
 118         int rnd_frac = ((rnd_load & (FIXED_1-1)) * 100) >> FSHIFT;
 119
 120         return ((rnd_load >> FSHIFT) * 100) + rnd_frac;
 121 }
 122
 123 META_COLLECTOR(int_loadavg_0)
 124 {
 125         dst->value = fixed_loadavg(avenrun[0]);
 126 }
 127
 128 META_COLLECTOR(int_loadavg_1)
 129 {
 130         dst->value = fixed_loadavg(avenrun[1]);
 131 }
 132
 133 META_COLLECTOR(int_loadavg_2)
 134 {
 135         dst->value = fixed_loadavg(avenrun[2]);
 136 }
 137
 138 /**************************************************************************
 139  * Device names & indices
 140  **************************************************************************/
 141
 142 static inline int int_dev(struct net_device *dev, struct meta_obj *dst)
 143 {
 144         if (unlikely(dev == NULL))
 145                 return -1;
 146
 147         dst->value = dev->ifindex;
 148         return 0;
 149 }
 150
 151 static inline int var_dev(struct net_device *dev, struct meta_obj *dst)
 152 {
 153         if (unlikely(dev == NULL))
 154                 return -1;
 155
 156         dst->value = (unsigned long) dev->name;
 157         dst->len = strlen(dev->name);
 158         return 0;
 159 }
 160
 161 META_COLLECTOR(int_dev)
 162 {
 163         *err = int_dev(skb->dev, dst);
 164 }
 165
 166 META_COLLECTOR(var_dev)
 167 {
 168         *err = var_dev(skb->dev, dst);
 169 }
 170
 171 /**************************************************************************
 172  * vlan tag
 173  **************************************************************************/
 174
 175 META_COLLECTOR(int_vlan_tag)
 176 {
 177         unsigned short tag;
 178
 179         tag = vlan_tx_tag_get(skb);
 180         if (!tag && __vlan_get_tag(skb, &tag))
 181                 *err = -1;
 182         else
 183                 dst->value = tag;
 184 }
 185
 186
 187
 188 /**************************************************************************
 189  * skb attributes
 190  **************************************************************************/
 191
 192 META_COLLECTOR(int_priority)
 193 {
 194         dst->value = skb->priority;
 195 }
 196
 197 META_COLLECTOR(int_protocol)
 198 {
 199         /* Let userspace take care of the byte ordering */
 200         dst->value = skb->protocol;
 201 }
 202
 203 META_COLLECTOR(int_pkttype)
 204 {
 205         dst->value = skb->pkt_type;
 206 }
 207
 208 META_COLLECTOR(int_pktlen)
 209 {
 210         dst->value = skb->len;
 211 }
 212
 213 META_COLLECTOR(int_datalen)
 214 {
 215         dst->value = skb->data_len;
 216 }
 217
 218 META_COLLECTOR(int_maclen)
 219 {
 220         dst->value = skb->mac_len;
 221 }
 222
 223 META_COLLECTOR(int_rxhash)
 224 {
 225         dst->value = skb_get_rxhash(skb);
 226 }
 227
 228 /**************************************************************************
 229  * Netfilter
 230  **************************************************************************/
 231
 232 META_COLLECTOR(int_mark)
 233 {
 234         dst->value = skb->mark;
 235 }
 236
 237 /**************************************************************************
 238  * Traffic Control
 239  **************************************************************************/
 240
 241 META_COLLECTOR(int_tcindex)
 242 {
 243         dst->value = skb->tc_index;
 244 }
 245
 246 /**************************************************************************
 247  * Routing
 248  **************************************************************************/
 249
 250 META_COLLECTOR(int_rtclassid)
 251 {
 252         if (unlikely(skb_dst(skb) == NULL))
 253                 *err = -1;
 254         else
 255 #ifdef CONFIG_IP_ROUTE_CLASSID
 256                 dst->value = skb_dst(skb)->tclassid;
 257 #else
 258                 dst->value = 0;
 259 #endif
 260 }
 261
 262 META_COLLECTOR(int_rtiif)
 263 {
 264         if (unlikely(skb_rtable(skb) == NULL))
 265                 *err = -1;
 266         else
 267                 dst->value = skb_rtable(skb)->rt_iif;
 268 }
 269
 270 /**************************************************************************
 271  * Socket Attributes
 272  **************************************************************************/
 273
 274 #define SKIP_NONLOCAL(skb)                      \
 275         if (unlikely(skb->sk == NULL)) {        \
 276                 *err = -1;                      \
 277                 return;                         \
 278         }
 279
 280 META_COLLECTOR(int_sk_family)
 281 {
 282         SKIP_NONLOCAL(skb);
 283         dst->value = skb->sk->sk_family;
 284 }
 285
 286 META_COLLECTOR(int_sk_state)
 287 {
 288         SKIP_NONLOCAL(skb);
 289         dst->value = skb->sk->sk_state;
 290 }
 291
 292 META_COLLECTOR(int_sk_reuse)
 293 {
 294         SKIP_NONLOCAL(skb);
 295         dst->value = skb->sk->sk_reuse;
 296 }
 297
 298 META_COLLECTOR(int_sk_bound_if)
 299 {
 300         SKIP_NONLOCAL(skb);
 301         /* No error if bound_dev_if is 0, legal userspace check */
 302         dst->value = skb->sk->sk_bound_dev_if;
 303 }
 304
 305 META_COLLECTOR(var_sk_bound_if)
 306 {
 307         SKIP_NONLOCAL(skb);
 308
 309         if (skb->sk->sk_bound_dev_if == 0) {
 310                 dst->value = (unsigned long) "any";
 311                 dst->len = 3;
 312         } else {
 313                 struct net_device *dev;
 314
 315                 rcu_read_lock();
 316                 dev = dev_get_by_index_rcu(sock_net(skb->sk),
 317                                            skb->sk->sk_bound_dev_if);
 318                 *err = var_dev(dev, dst);
 319                 rcu_read_unlock();
 320         }
 321 }
 322
 323 META_COLLECTOR(int_sk_refcnt)
 324 {
 325         SKIP_NONLOCAL(skb);
 326         dst->value = atomic_read(&skb->sk->sk_refcnt);
 327 }
 328
 329 META_COLLECTOR(int_sk_rcvbuf)
 330 {
 331         SKIP_NONLOCAL(skb);
 332         dst->value = skb->sk->sk_rcvbuf;
 333 }
 334
 335 META_COLLECTOR(int_sk_shutdown)
 336 {
 337         SKIP_NONLOCAL(skb);
 338         dst->value = skb->sk->sk_shutdown;
 339 }
 340
 341 META_COLLECTOR(int_sk_proto)
 342 {
 343         SKIP_NONLOCAL(skb);
 344         dst->value = skb->sk->sk_protocol;
 345 }
 346
 347 META_COLLECTOR(int_sk_type)
 348 {
 349         SKIP_NONLOCAL(skb);
 350         dst->value = skb->sk->sk_type;
 351 }
 352
 353 META_COLLECTOR(int_sk_rmem_alloc)
 354 {
 355         SKIP_NONLOCAL(skb);
 356         dst->value = sk_rmem_alloc_get(skb->sk);
 357 }
 358
 359 META_COLLECTOR(int_sk_wmem_alloc)
 360 {
 361         SKIP_NONLOCAL(skb);
 362         dst->value = sk_wmem_alloc_get(skb->sk);
 363 }
 364
 365 META_COLLECTOR(int_sk_omem_alloc)
 366 {
 367         SKIP_NONLOCAL(skb);
 368         dst->value = atomic_read(&skb->sk->sk_omem_alloc);
 369 }
 370
 371 META_COLLECTOR(int_sk_rcv_qlen)
 372 {
 373         SKIP_NONLOCAL(skb);
 374         dst->value = skb->sk->sk_receive_queue.qlen;
 375 }
 376
 377 META_COLLECTOR(int_sk_snd_qlen)
 378 {
 379         SKIP_NONLOCAL(skb);
 380         dst->value = skb->sk->sk_write_queue.qlen;
 381 }
 382
 383 META_COLLECTOR(int_sk_wmem_queued)
 384 {
 385         SKIP_NONLOCAL(skb);
 386         dst->value = skb->sk->sk_wmem_queued;
 387 }
 388
 389 META_COLLECTOR(int_sk_fwd_alloc)
 390 {
 391         SKIP_NONLOCAL(skb);
 392         dst->value = skb->sk->sk_forward_alloc;
 393 }
 394
 395 META_COLLECTOR(int_sk_sndbuf)
 396 {
 397         SKIP_NONLOCAL(skb);
 398         dst->value = skb->sk->sk_sndbuf;
 399 }
 400
 401 META_COLLECTOR(int_sk_alloc)
 402 {
 403         SKIP_NONLOCAL(skb);
 404         dst->value = (__force int) skb->sk->sk_allocation;
 405 }
 406
 407 META_COLLECTOR(int_sk_route_caps)
 408 {
 409         SKIP_NONLOCAL(skb);
 410         dst->value = skb->sk->sk_route_caps;
 411 }
 412
 413 META_COLLECTOR(int_sk_hash)
 414 {
 415         SKIP_NONLOCAL(skb);
 416         dst->value = skb->sk->sk_hash;
 417 }
 418
 419 META_COLLECTOR(int_sk_lingertime)
 420 {
 421         SKIP_NONLOCAL(skb);
 422         dst->value = skb->sk->sk_lingertime / HZ;
 423 }
 424
 425 META_COLLECTOR(int_sk_err_qlen)
 426 {
 427         SKIP_NONLOCAL(skb);
 428         dst->value = skb->sk->sk_error_queue.qlen;
 429 }
 430
 431 META_COLLECTOR(int_sk_ack_bl)
 432 {
 433         SKIP_NONLOCAL(skb);
 434         dst->value = skb->sk->sk_ack_backlog;
 435 }
 436
 437 META_COLLECTOR(int_sk_max_ack_bl)
 438 {
 439         SKIP_NONLOCAL(skb);
 440         dst->value = skb->sk->sk_max_ack_backlog;
 441 }
 442
 443 META_COLLECTOR(int_sk_prio)
 444 {
 445         SKIP_NONLOCAL(skb);
 446         dst->value = skb->sk->sk_priority;
 447 }
 448
 449 META_COLLECTOR(int_sk_rcvlowat)
 450 {
 451         SKIP_NONLOCAL(skb);
 452         dst->value = skb->sk->sk_rcvlowat;
 453 }
 454
 455 META_COLLECTOR(int_sk_rcvtimeo)
 456 {
 457         SKIP_NONLOCAL(skb);
 458         dst->value = skb->sk->sk_rcvtimeo / HZ;
 459 }
 460
 461 META_COLLECTOR(int_sk_sndtimeo)
 462 {
 463         SKIP_NONLOCAL(skb);
 464         dst->value = skb->sk->sk_sndtimeo / HZ;
 465 }
 466
 467 META_COLLECTOR(int_sk_sendmsg_off)
 468 {
 469         SKIP_NONLOCAL(skb);
 470         dst->value = skb->sk->sk_sndmsg_off;
 471 }
 472
 473 META_COLLECTOR(int_sk_write_pend)
 474 {
 475         SKIP_NONLOCAL(skb);
 476         dst->value = skb->sk->sk_write_pending;
 477 }
 478
 479 /**************************************************************************
 480  * Meta value collectors assignment table
 481  **************************************************************************/
 482
 483 struct meta_ops {
 484         void            (*get)(struct sk_buff *, struct tcf_pkt_info *,
 485                                struct meta_value *, struct meta_obj *, int *);
 486 };
 487
 488 #define META_ID(name) TCF_META_ID_##name
 489 #define META_FUNC(name) { .get = meta_##name }
 490
 491 /* Meta value operations table listing all meta value collectors and
 492  * assigns them to a type and meta id. */
 493 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX + 1][TCF_META_ID_MAX + 1] = {
 494         [TCF_META_TYPE_VAR] = {
 495                 [META_ID(DEV)]                  = META_FUNC(var_dev),
 496                 [META_ID(SK_BOUND_IF)]          = META_FUNC(var_sk_bound_if),
 497         },
 498         [TCF_META_TYPE_INT] = {
 499                 [META_ID(RANDOM)]               = META_FUNC(int_random),
 500                 [META_ID(LOADAVG_0)]            = META_FUNC(int_loadavg_0),
 501                 [META_ID(LOADAVG_1)]            = META_FUNC(int_loadavg_1),
 502                 [META_ID(LOADAVG_2)]            = META_FUNC(int_loadavg_2),
 503                 [META_ID(DEV)]                  = META_FUNC(int_dev),
 504                 [META_ID(PRIORITY)]             = META_FUNC(int_priority),
 505                 [META_ID(PROTOCOL)]             = META_FUNC(int_protocol),
 506                 [META_ID(PKTTYPE)]              = META_FUNC(int_pkttype),
 507                 [META_ID(PKTLEN)]               = META_FUNC(int_pktlen),
 508                 [META_ID(DATALEN)]              = META_FUNC(int_datalen),
 509                 [META_ID(MACLEN)]               = META_FUNC(int_maclen),
 510                 [META_ID(NFMARK)]               = META_FUNC(int_mark),
 511                 [META_ID(TCINDEX)]              = META_FUNC(int_tcindex),
 512                 [META_ID(RTCLASSID)]            = META_FUNC(int_rtclassid),
 513                 [META_ID(RTIIF)]                = META_FUNC(int_rtiif),
 514                 [META_ID(SK_FAMILY)]            = META_FUNC(int_sk_family),
 515                 [META_ID(SK_STATE)]             = META_FUNC(int_sk_state),
 516                 [META_ID(SK_REUSE)]             = META_FUNC(int_sk_reuse),
 517                 [META_ID(SK_BOUND_IF)]          = META_FUNC(int_sk_bound_if),
 518                 [META_ID(SK_REFCNT)]            = META_FUNC(int_sk_refcnt),
 519                 [META_ID(SK_RCVBUF)]            = META_FUNC(int_sk_rcvbuf),
 520                 [META_ID(SK_SNDBUF)]            = META_FUNC(int_sk_sndbuf),
 521                 [META_ID(SK_SHUTDOWN)]          = META_FUNC(int_sk_shutdown),
 522                 [META_ID(SK_PROTO)]             = META_FUNC(int_sk_proto),
 523                 [META_ID(SK_TYPE)]              = META_FUNC(int_sk_type),
 524                 [META_ID(SK_RMEM_ALLOC)]        = META_FUNC(int_sk_rmem_alloc),
 525                 [META_ID(SK_WMEM_ALLOC)]        = META_FUNC(int_sk_wmem_alloc),
 526                 [META_ID(SK_OMEM_ALLOC)]        = META_FUNC(int_sk_omem_alloc),
 527                 [META_ID(SK_WMEM_QUEUED)]       = META_FUNC(int_sk_wmem_queued),
 528                 [META_ID(SK_RCV_QLEN)]          = META_FUNC(int_sk_rcv_qlen),
 529                 [META_ID(SK_SND_QLEN)]          = META_FUNC(int_sk_snd_qlen),
 530                 [META_ID(SK_ERR_QLEN)]          = META_FUNC(int_sk_err_qlen),
 531                 [META_ID(SK_FORWARD_ALLOCS)]    = META_FUNC(int_sk_fwd_alloc),
 532                 [META_ID(SK_ALLOCS)]            = META_FUNC(int_sk_alloc),
 533                 [META_ID(SK_ROUTE_CAPS)]        = META_FUNC(int_sk_route_caps),
 534                 [META_ID(SK_HASH)]              = META_FUNC(int_sk_hash),
 535                 [META_ID(SK_LINGERTIME)]        = META_FUNC(int_sk_lingertime),
 536                 [META_ID(SK_ACK_BACKLOG)]       = META_FUNC(int_sk_ack_bl),
 537                 [META_ID(SK_MAX_ACK_BACKLOG)]   = META_FUNC(int_sk_max_ack_bl),
 538                 [META_ID(SK_PRIO)]              = META_FUNC(int_sk_prio),
 539                 [META_ID(SK_RCVLOWAT)]          = META_FUNC(int_sk_rcvlowat),
 540                 [META_ID(SK_RCVTIMEO)]          = META_FUNC(int_sk_rcvtimeo),
 541                 [META_ID(SK_SNDTIMEO)]          = META_FUNC(int_sk_sndtimeo),
 542                 [META_ID(SK_SENDMSG_OFF)]       = META_FUNC(int_sk_sendmsg_off),
 543                 [META_ID(SK_WRITE_PENDING)]     = META_FUNC(int_sk_write_pend),
 544                 [META_ID(VLAN_TAG)]             = META_FUNC(int_vlan_tag),
 545                 [META_ID(RXHASH)]               = META_FUNC(int_rxhash),
 546         }
 547 };
 548
 549 static inline struct meta_ops *meta_ops(struct meta_value *val)
 550 {
 551         return &__meta_ops[meta_type(val)][meta_id(val)];
 552 }
 553
 554 /**************************************************************************
 555  * Type specific operations for TCF_META_TYPE_VAR
 556  **************************************************************************/
 557
 558 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
 559 {
 560         int r = a->len - b->len;
 561
 562         if (r == 0)
 563                 r = memcmp((void *) a->value, (void *) b->value, a->len);
 564
 565         return r;
 566 }
 567
 568 static int meta_var_change(struct meta_value *dst, struct nlattr *nla)
 569 {
 570         int len = nla_len(nla);
 571
 572         dst->val = (unsigned long)kmemdup(nla_data(nla), len, GFP_KERNEL);
 573         if (dst->val == 0UL)
 574                 return -ENOMEM;
 575         dst->len = len;
 576         return 0;
 577 }
 578
 579 static void meta_var_destroy(struct meta_value *v)
 580 {
 581         kfree((void *) v->val);
 582 }
 583
 584 static void meta_var_apply_extras(struct meta_value *v,
 585                                   struct meta_obj *dst)
 586 {
 587         int shift = v->hdr.shift;
 588
 589         if (shift && shift < dst->len)
 590                 dst->len -= shift;
 591 }
 592
 593 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
 594 {
 595         if (v->val && v->len)
 596                 NLA_PUT(skb, tlv, v->len, (void *) v->val);
 597         return 0;
 598
 599 nla_put_failure:
 600         return -1;
 601 }
 602
 603 /**************************************************************************
 604  * Type specific operations for TCF_META_TYPE_INT
 605  **************************************************************************/
 606
 607 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
 608 {
 609         /* Let gcc optimize it, the unlikely is not really based on
 610          * some numbers but jump free code for mismatches seems
 611          * more logical. */
 612         if (unlikely(a->value == b->value))
 613                 return 0;
 614         else if (a->value < b->value)
 615                 return -1;
 616         else
 617                 return 1;
 618 }
 619
 620 static int meta_int_change(struct meta_value *dst, struct nlattr *nla)
 621 {
 622         if (nla_len(nla) >= sizeof(unsigned long)) {
 623                 dst->val = *(unsigned long *) nla_data(nla);
 624                 dst->len = sizeof(unsigned long);
 625         } else if (nla_len(nla) == sizeof(u32)) {
 626                 dst->val = nla_get_u32(nla);
 627                 dst->len = sizeof(u32);
 628         } else
 629                 return -EINVAL;
 630
 631         return 0;
 632 }
 633
 634 static void meta_int_apply_extras(struct meta_value *v,
 635                                   struct meta_obj *dst)
 636 {
 637         if (v->hdr.shift)
 638                 dst->value >>= v->hdr.shift;
 639
 640         if (v->val)
 641                 dst->value &= v->val;
 642 }
 643
 644 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
 645 {
 646         if (v->len == sizeof(unsigned long))
 647                 NLA_PUT(skb, tlv, sizeof(unsigned long), &v->val);
 648         else if (v->len == sizeof(u32))
 649                 NLA_PUT_U32(skb, tlv, v->val);
 650
 651         return 0;
 652
 653 nla_put_failure:
 654         return -1;
 655 }
 656
 657 /**************************************************************************
 658  * Type specific operations table
 659  **************************************************************************/
 660
 661 struct meta_type_ops {
 662         void    (*destroy)(struct meta_value *);
 663         int     (*compare)(struct meta_obj *, struct meta_obj *);
 664         int     (*change)(struct meta_value *, struct nlattr *);
 665         void    (*apply_extras)(struct meta_value *, struct meta_obj *);
 666         int     (*dump)(struct sk_buff *, struct meta_value *, int);
 667 };
 668
 669 static struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX + 1] = {
 670         [TCF_META_TYPE_VAR] = {
 671                 .destroy = meta_var_destroy,
 672                 .compare = meta_var_compare,
 673                 .change = meta_var_change,
 674                 .apply_extras = meta_var_apply_extras,
 675                 .dump = meta_var_dump
 676         },
 677         [TCF_META_TYPE_INT] = {
 678                 .compare = meta_int_compare,
 679                 .change = meta_int_change,
 680                 .apply_extras = meta_int_apply_extras,
 681                 .dump = meta_int_dump
 682         }
 683 };
 684
 685 static inline struct meta_type_ops *meta_type_ops(struct meta_value *v)
 686 {
 687         return &__meta_type_ops[meta_type(v)];
 688 }
 689
 690 /**************************************************************************
 691  * Core
 692  **************************************************************************/
 693
 694 static int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
 695                     struct meta_value *v, struct meta_obj *dst)
 696 {
 697         int err = 0;
 698
 699         if (meta_id(v) == TCF_META_ID_VALUE) {
 700                 dst->value = v->val;
 701                 dst->len = v->len;
 702                 return 0;
 703         }
 704
 705         meta_ops(v)->get(skb, info, v, dst, &err);
 706         if (err < 0)
 707                 return err;
 708
 709         if (meta_type_ops(v)->apply_extras)
 710                 meta_type_ops(v)->apply_extras(v, dst);
 711
 712         return 0;
 713 }
 714
 715 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
 716                          struct tcf_pkt_info *info)
 717 {
 718         int r;
 719         struct meta_match *meta = (struct meta_match *) m->data;
 720         struct meta_obj l_value, r_value;
 721
 722         if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
 723             meta_get(skb, info, &meta->rvalue, &r_value) < 0)
 724                 return 0;
 725
 726         r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
 727
 728         switch (meta->lvalue.hdr.op) {
 729         case TCF_EM_OPND_EQ:
 730                 return !r;
 731         case TCF_EM_OPND_LT:
 732                 return r < 0;
 733         case TCF_EM_OPND_GT:
 734                 return r > 0;
 735         }
 736
 737         return 0;
 738 }
 739
 740 static void meta_delete(struct meta_match *meta)
 741 {
 742         if (meta) {
 743                 struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
 744
 745                 if (ops && ops->destroy) {
 746                         ops->destroy(&meta->lvalue);
 747                         ops->destroy(&meta->rvalue);
 748                 }
 749         }
 750
 751         kfree(meta);
 752 }
 753
 754 static inline int meta_change_data(struct meta_value *dst, struct nlattr *nla)
 755 {
 756         if (nla) {
 757                 if (nla_len(nla) == 0)
 758                         return -EINVAL;
 759
 760                 return meta_type_ops(dst)->change(dst, nla);
 761         }
 762
 763         return 0;
 764 }
 765
 766 static inline int meta_is_supported(struct meta_value *val)
 767 {
 768         return !meta_id(val) || meta_ops(val)->get;
 769 }
 770
 771 static const struct nla_policy meta_policy[TCA_EM_META_MAX + 1] = {
 772         [TCA_EM_META_HDR]       = { .len = sizeof(struct tcf_meta_hdr) },
 773 };
 774
 775 static int em_meta_change(struct tcf_proto *tp, void *data, int len,
 776                           struct tcf_ematch *m)
 777 {
 778         int err;
 779         struct nlattr *tb[TCA_EM_META_MAX + 1];
 780         struct tcf_meta_hdr *hdr;
 781         struct meta_match *meta = NULL;
 782
 783         err = nla_parse(tb, TCA_EM_META_MAX, data, len, meta_policy);
 784         if (err < 0)
 785                 goto errout;
 786
 787         err = -EINVAL;
 788         if (tb[TCA_EM_META_HDR] == NULL)
 789                 goto errout;
 790         hdr = nla_data(tb[TCA_EM_META_HDR]);
 791
 792         if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
 793             TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
 794             TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
 795             TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
 796                 goto errout;
 797
 798         meta = kzalloc(sizeof(*meta), GFP_KERNEL);
 799         if (meta == NULL)
 800                 goto errout;
 801
 802         memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
 803         memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
 804
 805         if (!meta_is_supported(&meta->lvalue) ||
 806             !meta_is_supported(&meta->rvalue)) {
 807                 err = -EOPNOTSUPP;
 808                 goto errout;
 809         }
 810
 811         if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE]) < 0 ||
 812             meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE]) < 0)
 813                 goto errout;
 814
 815         m->datalen = sizeof(*meta);
 816         m->data = (unsigned long) meta;
 817
 818         err = 0;
 819 errout:
 820         if (err && meta)
 821                 meta_delete(meta);
 822         return err;
 823 }
 824
 825 static void em_meta_destroy(struct tcf_proto *tp, struct tcf_ematch *m)
 826 {
 827         if (m)
 828                 meta_delete((struct meta_match *) m->data);
 829 }
 830
 831 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
 832 {
 833         struct meta_match *meta = (struct meta_match *) em->data;
 834         struct tcf_meta_hdr hdr;
 835         struct meta_type_ops *ops;
 836
 837         memset(&hdr, 0, sizeof(hdr));
 838         memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
 839         memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
 840
 841         NLA_PUT(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr);
 842
 843         ops = meta_type_ops(&meta->lvalue);
 844         if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
 845             ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
 846                 goto nla_put_failure;
 847
 848         return 0;
 849
 850 nla_put_failure:
 851         return -1;
 852 }
 853
 854 static struct tcf_ematch_ops em_meta_ops = {
 855         .kind     = TCF_EM_META,
 856         .change   = em_meta_change,
 857         .match    = em_meta_match,
 858         .destroy  = em_meta_destroy,
 859         .dump     = em_meta_dump,
 860         .owner    = THIS_MODULE,
 861         .link     = LIST_HEAD_INIT(em_meta_ops.link)
 862 };
 863
 864 static int __init init_em_meta(void)
 865 {
 866         return tcf_em_register(&em_meta_ops);
 867 }
 868
 869 static void __exit exit_em_meta(void)
 870 {
 871         tcf_em_unregister(&em_meta_ops);
 872 }
 873
 874 MODULE_LICENSE("GPL");
 875
 876 module_init(init_em_meta);
 877 module_exit(exit_em_meta);
 878
 879 MODULE_ALIAS_TCF_EMATCH(TCF_EM_META);