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: INDEV |                           | id: VALUE |
  31  *            | data:     |                           | data: 3   |
  32  *            +-----------+                           +-----------+
  33  *                  |                                       |
  34  *                  ---> meta_ops[INT][INDEV](...)          |
  35  *                            |                             |
  36  *                  -----------                             |
  37  *                  V                                       V
  38  *            +-----------+                           +-----------+
  39  *            | type: INT |                           | type: INT |
  40  *       obj  | id: INDEV |                           | 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  *      Additionaly, type dependant 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 subsytem is enabled in the kernel.
  59  */
  60
  61 #include <linux/config.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/tc_ematch/tc_em_meta.h>
  70 #include <net/dst.h>
  71 #include <net/route.h>
  72 #include <net/pkt_cls.h>
  73 #include <net/sock.h>
  74
  75 struct meta_obj
  76 {
  77         unsigned long           value;
  78         unsigned int            len;
  79 };
  80
  81 struct meta_value
  82 {
  83         struct tcf_meta_val     hdr;
  84         unsigned long           val;
  85         unsigned int            len;
  86 };
  87
  88 struct meta_match
  89 {
  90         struct meta_value       lvalue;
  91         struct meta_value       rvalue;
  92 };
  93
  94 static inline int meta_id(struct meta_value *v)
  95 {
  96         return TCF_META_ID(v->hdr.kind);
  97 }
  98
  99 static inline int meta_type(struct meta_value *v)
 100 {
 101         return TCF_META_TYPE(v->hdr.kind);
 102 }
 103
 104 #define META_COLLECTOR(FUNC) static void meta_##FUNC(struct sk_buff *skb, \
 105         struct tcf_pkt_info *info, struct meta_value *v, \
 106         struct meta_obj *dst, int *err)
 107
 108 /**************************************************************************
 109  * System status & misc
 110  **************************************************************************/
 111
 112 META_COLLECTOR(int_random)
 113 {
 114         get_random_bytes(&dst->value, sizeof(dst->value));
 115 }
 116
 117 static inline unsigned long fixed_loadavg(int load)
 118 {
 119         int rnd_load = load + (FIXED_1/200);
 120         int rnd_frac = ((rnd_load & (FIXED_1-1)) * 100) >> FSHIFT;
 121
 122         return ((rnd_load >> FSHIFT) * 100) + rnd_frac;
 123 }
 124
 125 META_COLLECTOR(int_loadavg_0)
 126 {
 127         dst->value = fixed_loadavg(avenrun[0]);
 128 }
 129
 130 META_COLLECTOR(int_loadavg_1)
 131 {
 132         dst->value = fixed_loadavg(avenrun[1]);
 133 }
 134
 135 META_COLLECTOR(int_loadavg_2)
 136 {
 137         dst->value = fixed_loadavg(avenrun[2]);
 138 }
 139
 140 /**************************************************************************
 141  * Device names & indices
 142  **************************************************************************/
 143
 144 static inline int int_dev(struct net_device *dev, struct meta_obj *dst)
 145 {
 146         if (unlikely(dev == NULL))
 147                 return -1;
 148
 149         dst->value = dev->ifindex;
 150         return 0;
 151 }
 152
 153 static inline int var_dev(struct net_device *dev, struct meta_obj *dst)
 154 {
 155         if (unlikely(dev == NULL))
 156                 return -1;
 157
 158         dst->value = (unsigned long) dev->name;
 159         dst->len = strlen(dev->name);
 160         return 0;
 161 }
 162
 163 META_COLLECTOR(int_dev)
 164 {
 165         *err = int_dev(skb->dev, dst);
 166 }
 167
 168 META_COLLECTOR(var_dev)
 169 {
 170         *err = var_dev(skb->dev, dst);
 171 }
 172
 173 META_COLLECTOR(int_indev)
 174 {
 175         *err = int_dev(skb->input_dev, dst);
 176 }
 177
 178 META_COLLECTOR(var_indev)
 179 {
 180         *err = var_dev(skb->input_dev, dst);
 181 }
 182
 183 META_COLLECTOR(int_realdev)
 184 {
 185         *err = int_dev(skb->real_dev, dst);
 186 }
 187
 188 META_COLLECTOR(var_realdev)
 189 {
 190         *err = var_dev(skb->real_dev, dst);
 191 }
 192
 193 /**************************************************************************
 194  * skb attributes
 195  **************************************************************************/
 196
 197 META_COLLECTOR(int_priority)
 198 {
 199         dst->value = skb->priority;
 200 }
 201
 202 META_COLLECTOR(int_protocol)
 203 {
 204         /* Let userspace take care of the byte ordering */
 205         dst->value = skb->protocol;
 206 }
 207
 208 META_COLLECTOR(int_security)
 209 {
 210         dst->value = skb->security;
 211 }
 212
 213 META_COLLECTOR(int_pkttype)
 214 {
 215         dst->value = skb->pkt_type;
 216 }
 217
 218 META_COLLECTOR(int_pktlen)
 219 {
 220         dst->value = skb->len;
 221 }
 222
 223 META_COLLECTOR(int_datalen)
 224 {
 225         dst->value = skb->data_len;
 226 }
 227
 228 META_COLLECTOR(int_maclen)
 229 {
 230         dst->value = skb->mac_len;
 231 }
 232
 233 /**************************************************************************
 234  * Netfilter
 235  **************************************************************************/
 236
 237 #ifdef CONFIG_NETFILTER
 238 META_COLLECTOR(int_nfmark)
 239 {
 240         dst->value = skb->nfmark;
 241 }
 242 #endif
 243
 244 /**************************************************************************
 245  * Traffic Control
 246  **************************************************************************/
 247
 248 META_COLLECTOR(int_tcindex)
 249 {
 250         dst->value = skb->tc_index;
 251 }
 252
 253 #ifdef CONFIG_NET_CLS_ACT
 254 META_COLLECTOR(int_tcverd)
 255 {
 256         dst->value = skb->tc_verd;
 257 }
 258
 259 META_COLLECTOR(int_tcclassid)
 260 {
 261         dst->value = skb->tc_classid;
 262 }
 263 #endif
 264
 265 /**************************************************************************
 266  * Routing
 267  **************************************************************************/
 268
 269 #ifdef CONFIG_NET_CLS_ROUTE
 270 META_COLLECTOR(int_rtclassid)
 271 {
 272         if (unlikely(skb->dst == NULL))
 273                 *err = -1;
 274         else
 275                 dst->value = skb->dst->tclassid;
 276 }
 277 #endif
 278
 279 META_COLLECTOR(int_rtiif)
 280 {
 281         if (unlikely(skb->dst == NULL))
 282                 *err = -1;
 283         else
 284                 dst->value = ((struct rtable*) skb->dst)->fl.iif;
 285 }
 286
 287 /**************************************************************************
 288  * Socket Attributes
 289  **************************************************************************/
 290
 291 #define SKIP_NONLOCAL(skb)                      \
 292         if (unlikely(skb->sk == NULL)) {        \
 293                 *err = -1;                      \
 294                 return;                         \
 295         }
 296
 297 META_COLLECTOR(int_sk_family)
 298 {
 299         SKIP_NONLOCAL(skb);
 300         dst->value = skb->sk->sk_family;
 301 }
 302
 303 META_COLLECTOR(int_sk_state)
 304 {
 305         SKIP_NONLOCAL(skb);
 306         dst->value = skb->sk->sk_state;
 307 }
 308
 309 META_COLLECTOR(int_sk_reuse)
 310 {
 311         SKIP_NONLOCAL(skb);
 312         dst->value = skb->sk->sk_reuse;
 313 }
 314
 315 META_COLLECTOR(int_sk_bound_if)
 316 {
 317         SKIP_NONLOCAL(skb);
 318         /* No error if bound_dev_if is 0, legal userspace check */
 319         dst->value = skb->sk->sk_bound_dev_if;
 320 }
 321
 322 META_COLLECTOR(var_sk_bound_if)
 323 {
 324         SKIP_NONLOCAL(skb);
 325
 326          if (skb->sk->sk_bound_dev_if == 0) {
 327                 dst->value = (unsigned long) "any";
 328                 dst->len = 3;
 329          } else  {
 330                 struct net_device *dev;
 331
 332                 dev = dev_get_by_index(skb->sk->sk_bound_dev_if);
 333                 *err = var_dev(dev, dst);
 334                 if (dev)
 335                         dev_put(dev);
 336          }
 337 }
 338
 339 META_COLLECTOR(int_sk_refcnt)
 340 {
 341         SKIP_NONLOCAL(skb);
 342         dst->value = atomic_read(&skb->sk->sk_refcnt);
 343 }
 344
 345 META_COLLECTOR(int_sk_rcvbuf)
 346 {
 347         SKIP_NONLOCAL(skb);
 348         dst->value = skb->sk->sk_rcvbuf;
 349 }
 350
 351 META_COLLECTOR(int_sk_shutdown)
 352 {
 353         SKIP_NONLOCAL(skb);
 354         dst->value = skb->sk->sk_shutdown;
 355 }
 356
 357 META_COLLECTOR(int_sk_proto)
 358 {
 359         SKIP_NONLOCAL(skb);
 360         dst->value = skb->sk->sk_protocol;
 361 }
 362
 363 META_COLLECTOR(int_sk_type)
 364 {
 365         SKIP_NONLOCAL(skb);
 366         dst->value = skb->sk->sk_type;
 367 }
 368
 369 META_COLLECTOR(int_sk_rmem_alloc)
 370 {
 371         SKIP_NONLOCAL(skb);
 372         dst->value = atomic_read(&skb->sk->sk_rmem_alloc);
 373 }
 374
 375 META_COLLECTOR(int_sk_wmem_alloc)
 376 {
 377         SKIP_NONLOCAL(skb);
 378         dst->value = atomic_read(&skb->sk->sk_wmem_alloc);
 379 }
 380
 381 META_COLLECTOR(int_sk_omem_alloc)
 382 {
 383         SKIP_NONLOCAL(skb);
 384         dst->value = atomic_read(&skb->sk->sk_omem_alloc);
 385 }
 386
 387 META_COLLECTOR(int_sk_rcv_qlen)
 388 {
 389         SKIP_NONLOCAL(skb);
 390         dst->value = skb->sk->sk_receive_queue.qlen;
 391 }
 392
 393 META_COLLECTOR(int_sk_snd_qlen)
 394 {
 395         SKIP_NONLOCAL(skb);
 396         dst->value = skb->sk->sk_write_queue.qlen;
 397 }
 398
 399 META_COLLECTOR(int_sk_wmem_queued)
 400 {
 401         SKIP_NONLOCAL(skb);
 402         dst->value = skb->sk->sk_wmem_queued;
 403 }
 404
 405 META_COLLECTOR(int_sk_fwd_alloc)
 406 {
 407         SKIP_NONLOCAL(skb);
 408         dst->value = skb->sk->sk_forward_alloc;
 409 }
 410
 411 META_COLLECTOR(int_sk_sndbuf)
 412 {
 413         SKIP_NONLOCAL(skb);
 414         dst->value = skb->sk->sk_sndbuf;
 415 }
 416
 417 META_COLLECTOR(int_sk_alloc)
 418 {
 419         SKIP_NONLOCAL(skb);
 420         dst->value = skb->sk->sk_allocation;
 421 }
 422
 423 META_COLLECTOR(int_sk_route_caps)
 424 {
 425         SKIP_NONLOCAL(skb);
 426         dst->value = skb->sk->sk_route_caps;
 427 }
 428
 429 META_COLLECTOR(int_sk_hashent)
 430 {
 431         SKIP_NONLOCAL(skb);
 432         dst->value = skb->sk->sk_hashent;
 433 }
 434
 435 META_COLLECTOR(int_sk_lingertime)
 436 {
 437         SKIP_NONLOCAL(skb);
 438         dst->value = skb->sk->sk_lingertime / HZ;
 439 }
 440
 441 META_COLLECTOR(int_sk_err_qlen)
 442 {
 443         SKIP_NONLOCAL(skb);
 444         dst->value = skb->sk->sk_error_queue.qlen;
 445 }
 446
 447 META_COLLECTOR(int_sk_ack_bl)
 448 {
 449         SKIP_NONLOCAL(skb);
 450         dst->value = skb->sk->sk_ack_backlog;
 451 }
 452
 453 META_COLLECTOR(int_sk_max_ack_bl)
 454 {
 455         SKIP_NONLOCAL(skb);
 456         dst->value = skb->sk->sk_max_ack_backlog;
 457 }
 458
 459 META_COLLECTOR(int_sk_prio)
 460 {
 461         SKIP_NONLOCAL(skb);
 462         dst->value = skb->sk->sk_priority;
 463 }
 464
 465 META_COLLECTOR(int_sk_rcvlowat)
 466 {
 467         SKIP_NONLOCAL(skb);
 468         dst->value = skb->sk->sk_rcvlowat;
 469 }
 470
 471 META_COLLECTOR(int_sk_rcvtimeo)
 472 {
 473         SKIP_NONLOCAL(skb);
 474         dst->value = skb->sk->sk_rcvtimeo / HZ;
 475 }
 476
 477 META_COLLECTOR(int_sk_sndtimeo)
 478 {
 479         SKIP_NONLOCAL(skb);
 480         dst->value = skb->sk->sk_sndtimeo / HZ;
 481 }
 482
 483 META_COLLECTOR(int_sk_sendmsg_off)
 484 {
 485         SKIP_NONLOCAL(skb);
 486         dst->value = skb->sk->sk_sndmsg_off;
 487 }
 488
 489 META_COLLECTOR(int_sk_write_pend)
 490 {
 491         SKIP_NONLOCAL(skb);
 492         dst->value = skb->sk->sk_write_pending;
 493 }
 494
 495 /**************************************************************************
 496  * Meta value collectors assignment table
 497  **************************************************************************/
 498
 499 struct meta_ops
 500 {
 501         void            (*get)(struct sk_buff *, struct tcf_pkt_info *,
 502                                struct meta_value *, struct meta_obj *, int *);
 503 };
 504
 505 #define META_ID(name) TCF_META_ID_##name
 506 #define META_FUNC(name) { .get = meta_##name }
 507
 508 /* Meta value operations table listing all meta value collectors and
 509  * assigns them to a type and meta id. */
 510 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX+1][TCF_META_ID_MAX+1] = {
 511         [TCF_META_TYPE_VAR] = {
 512                 [META_ID(DEV)]                  = META_FUNC(var_dev),
 513                 [META_ID(INDEV)]                = META_FUNC(var_indev),
 514                 [META_ID(REALDEV)]              = META_FUNC(var_realdev),
 515                 [META_ID(SK_BOUND_IF)]          = META_FUNC(var_sk_bound_if),
 516         },
 517         [TCF_META_TYPE_INT] = {
 518                 [META_ID(RANDOM)]               = META_FUNC(int_random),
 519                 [META_ID(LOADAVG_0)]            = META_FUNC(int_loadavg_0),
 520                 [META_ID(LOADAVG_1)]            = META_FUNC(int_loadavg_1),
 521                 [META_ID(LOADAVG_2)]            = META_FUNC(int_loadavg_2),
 522                 [META_ID(DEV)]                  = META_FUNC(int_dev),
 523                 [META_ID(INDEV)]                = META_FUNC(int_indev),
 524                 [META_ID(REALDEV)]              = META_FUNC(int_realdev),
 525                 [META_ID(PRIORITY)]             = META_FUNC(int_priority),
 526                 [META_ID(PROTOCOL)]             = META_FUNC(int_protocol),
 527                 [META_ID(SECURITY)]             = META_FUNC(int_security),
 528                 [META_ID(PKTTYPE)]              = META_FUNC(int_pkttype),
 529                 [META_ID(PKTLEN)]               = META_FUNC(int_pktlen),
 530                 [META_ID(DATALEN)]              = META_FUNC(int_datalen),
 531                 [META_ID(MACLEN)]               = META_FUNC(int_maclen),
 532 #ifdef CONFIG_NETFILTER
 533                 [META_ID(NFMARK)]               = META_FUNC(int_nfmark),
 534 #endif
 535                 [META_ID(TCINDEX)]              = META_FUNC(int_tcindex),
 536 #ifdef CONFIG_NET_CLS_ACT
 537                 [META_ID(TCVERDICT)]            = META_FUNC(int_tcverd),
 538                 [META_ID(TCCLASSID)]            = META_FUNC(int_tcclassid),
 539 #endif
 540 #ifdef CONFIG_NET_CLS_ROUTE
 541                 [META_ID(RTCLASSID)]            = META_FUNC(int_rtclassid),
 542 #endif
 543                 [META_ID(RTIIF)]                = META_FUNC(int_rtiif),
 544                 [META_ID(SK_FAMILY)]            = META_FUNC(int_sk_family),
 545                 [META_ID(SK_STATE)]             = META_FUNC(int_sk_state),
 546                 [META_ID(SK_REUSE)]             = META_FUNC(int_sk_reuse),
 547                 [META_ID(SK_BOUND_IF)]          = META_FUNC(int_sk_bound_if),
 548                 [META_ID(SK_REFCNT)]            = META_FUNC(int_sk_refcnt),
 549                 [META_ID(SK_RCVBUF)]            = META_FUNC(int_sk_rcvbuf),
 550                 [META_ID(SK_SNDBUF)]            = META_FUNC(int_sk_sndbuf),
 551                 [META_ID(SK_SHUTDOWN)]          = META_FUNC(int_sk_shutdown),
 552                 [META_ID(SK_PROTO)]             = META_FUNC(int_sk_proto),
 553                 [META_ID(SK_TYPE)]              = META_FUNC(int_sk_type),
 554                 [META_ID(SK_RMEM_ALLOC)]        = META_FUNC(int_sk_rmem_alloc),
 555                 [META_ID(SK_WMEM_ALLOC)]        = META_FUNC(int_sk_wmem_alloc),
 556                 [META_ID(SK_OMEM_ALLOC)]        = META_FUNC(int_sk_omem_alloc),
 557                 [META_ID(SK_WMEM_QUEUED)]       = META_FUNC(int_sk_wmem_queued),
 558                 [META_ID(SK_RCV_QLEN)]          = META_FUNC(int_sk_rcv_qlen),
 559                 [META_ID(SK_SND_QLEN)]          = META_FUNC(int_sk_snd_qlen),
 560                 [META_ID(SK_ERR_QLEN)]          = META_FUNC(int_sk_err_qlen),
 561                 [META_ID(SK_FORWARD_ALLOCS)]    = META_FUNC(int_sk_fwd_alloc),
 562                 [META_ID(SK_ALLOCS)]            = META_FUNC(int_sk_alloc),
 563                 [META_ID(SK_ROUTE_CAPS)]        = META_FUNC(int_sk_route_caps),
 564                 [META_ID(SK_HASHENT)]           = META_FUNC(int_sk_hashent),
 565                 [META_ID(SK_LINGERTIME)]        = META_FUNC(int_sk_lingertime),
 566                 [META_ID(SK_ACK_BACKLOG)]       = META_FUNC(int_sk_ack_bl),
 567                 [META_ID(SK_MAX_ACK_BACKLOG)]   = META_FUNC(int_sk_max_ack_bl),
 568                 [META_ID(SK_PRIO)]              = META_FUNC(int_sk_prio),
 569                 [META_ID(SK_RCVLOWAT)]          = META_FUNC(int_sk_rcvlowat),
 570                 [META_ID(SK_RCVTIMEO)]          = META_FUNC(int_sk_rcvtimeo),
 571                 [META_ID(SK_SNDTIMEO)]          = META_FUNC(int_sk_sndtimeo),
 572                 [META_ID(SK_SENDMSG_OFF)]       = META_FUNC(int_sk_sendmsg_off),
 573                 [META_ID(SK_WRITE_PENDING)]     = META_FUNC(int_sk_write_pend),
 574         }
 575 };
 576
 577 static inline struct meta_ops * meta_ops(struct meta_value *val)
 578 {
 579         return &__meta_ops[meta_type(val)][meta_id(val)];
 580 }
 581
 582 /**************************************************************************
 583  * Type specific operations for TCF_META_TYPE_VAR
 584  **************************************************************************/
 585
 586 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
 587 {
 588         int r = a->len - b->len;
 589
 590         if (r == 0)
 591                 r = memcmp((void *) a->value, (void *) b->value, a->len);
 592
 593         return r;
 594 }
 595
 596 static int meta_var_change(struct meta_value *dst, struct rtattr *rta)
 597 {
 598         int len = RTA_PAYLOAD(rta);
 599
 600         dst->val = (unsigned long) kmalloc(len, GFP_KERNEL);
 601         if (dst->val == 0UL)
 602                 return -ENOMEM;
 603         memcpy((void *) dst->val, RTA_DATA(rta), len);
 604         dst->len = len;
 605         return 0;
 606 }
 607
 608 static void meta_var_destroy(struct meta_value *v)
 609 {
 610         if (v->val)
 611                 kfree((void *) v->val);
 612 }
 613
 614 static void meta_var_apply_extras(struct meta_value *v,
 615                                   struct meta_obj *dst)
 616 {
 617         int shift = v->hdr.shift;
 618
 619         if (shift && shift < dst->len)
 620                 dst->len -= shift;
 621 }
 622
 623 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
 624 {
 625         if (v->val && v->len)
 626                 RTA_PUT(skb, tlv, v->len, (void *) v->val);
 627         return 0;
 628
 629 rtattr_failure:
 630         return -1;
 631 }
 632
 633 /**************************************************************************
 634  * Type specific operations for TCF_META_TYPE_INT
 635  **************************************************************************/
 636
 637 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
 638 {
 639         /* Let gcc optimize it, the unlikely is not really based on
 640          * some numbers but jump free code for mismatches seems
 641          * more logical. */
 642         if (unlikely(a->value == b->value))
 643                 return 0;
 644         else if (a->value < b->value)
 645                 return -1;
 646         else
 647                 return 1;
 648 }
 649
 650 static int meta_int_change(struct meta_value *dst, struct rtattr *rta)
 651 {
 652         if (RTA_PAYLOAD(rta) >= sizeof(unsigned long)) {
 653                 dst->val = *(unsigned long *) RTA_DATA(rta);
 654                 dst->len = sizeof(unsigned long);
 655         } else if (RTA_PAYLOAD(rta) == sizeof(u32)) {
 656                 dst->val = *(u32 *) RTA_DATA(rta);
 657                 dst->len = sizeof(u32);
 658         } else
 659                 return -EINVAL;
 660
 661         return 0;
 662 }
 663
 664 static void meta_int_apply_extras(struct meta_value *v,
 665                                   struct meta_obj *dst)
 666 {
 667         if (v->hdr.shift)
 668                 dst->value >>= v->hdr.shift;
 669
 670         if (v->val)
 671                 dst->value &= v->val;
 672 }
 673
 674 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
 675 {
 676         if (v->len == sizeof(unsigned long))
 677                 RTA_PUT(skb, tlv, sizeof(unsigned long), &v->val);
 678         else if (v->len == sizeof(u32)) {
 679                 u32 d = v->val;
 680                 RTA_PUT(skb, tlv, sizeof(d), &d);
 681         }
 682
 683         return 0;
 684
 685 rtattr_failure:
 686         return -1;
 687 }
 688
 689 /**************************************************************************
 690  * Type specific operations table
 691  **************************************************************************/
 692
 693 struct meta_type_ops
 694 {
 695         void    (*destroy)(struct meta_value *);
 696         int     (*compare)(struct meta_obj *, struct meta_obj *);
 697         int     (*change)(struct meta_value *, struct rtattr *);
 698         void    (*apply_extras)(struct meta_value *, struct meta_obj *);
 699         int     (*dump)(struct sk_buff *, struct meta_value *, int);
 700 };
 701
 702 static struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX+1] = {
 703         [TCF_META_TYPE_VAR] = {
 704                 .destroy = meta_var_destroy,
 705                 .compare = meta_var_compare,
 706                 .change = meta_var_change,
 707                 .apply_extras = meta_var_apply_extras,
 708                 .dump = meta_var_dump
 709         },
 710         [TCF_META_TYPE_INT] = {
 711                 .compare = meta_int_compare,
 712                 .change = meta_int_change,
 713                 .apply_extras = meta_int_apply_extras,
 714                 .dump = meta_int_dump
 715         }
 716 };
 717
 718 static inline struct meta_type_ops * meta_type_ops(struct meta_value *v)
 719 {
 720         return &__meta_type_ops[meta_type(v)];
 721 }
 722
 723 /**************************************************************************
 724  * Core
 725  **************************************************************************/
 726
 727 static inline int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
 728                            struct meta_value *v, struct meta_obj *dst)
 729 {
 730         int err = 0;
 731
 732         if (meta_id(v) == TCF_META_ID_VALUE) {
 733                 dst->value = v->val;
 734                 dst->len = v->len;
 735                 return 0;
 736         }
 737
 738         meta_ops(v)->get(skb, info, v, dst, &err);
 739         if (err < 0)
 740                 return err;
 741
 742         if (meta_type_ops(v)->apply_extras)
 743             meta_type_ops(v)->apply_extras(v, dst);
 744
 745         return 0;
 746 }
 747
 748 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
 749                          struct tcf_pkt_info *info)
 750 {
 751         int r;
 752         struct meta_match *meta = (struct meta_match *) m->data;
 753         struct meta_obj l_value, r_value;
 754
 755         if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
 756             meta_get(skb, info, &meta->rvalue, &r_value) < 0)
 757                 return 0;
 758
 759         r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
 760
 761         switch (meta->lvalue.hdr.op) {
 762                 case TCF_EM_OPND_EQ:
 763                         return !r;
 764                 case TCF_EM_OPND_LT:
 765                         return r < 0;
 766                 case TCF_EM_OPND_GT:
 767                         return r > 0;
 768         }
 769
 770         return 0;
 771 }
 772
 773 static inline void meta_delete(struct meta_match *meta)
 774 {
 775         struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
 776
 777         if (ops && ops->destroy) {
 778                 ops->destroy(&meta->lvalue);
 779                 ops->destroy(&meta->rvalue);
 780         }
 781
 782         kfree(meta);
 783 }
 784
 785 static inline int meta_change_data(struct meta_value *dst, struct rtattr *rta)
 786 {
 787         if (rta) {
 788                 if (RTA_PAYLOAD(rta) == 0)
 789                         return -EINVAL;
 790
 791                 return meta_type_ops(dst)->change(dst, rta);
 792         }
 793
 794         return 0;
 795 }
 796
 797 static inline int meta_is_supported(struct meta_value *val)
 798 {
 799         return (!meta_id(val) || meta_ops(val)->get);
 800 }
 801
 802 static int em_meta_change(struct tcf_proto *tp, void *data, int len,
 803                           struct tcf_ematch *m)
 804 {
 805         int err = -EINVAL;
 806         struct rtattr *tb[TCA_EM_META_MAX];
 807         struct tcf_meta_hdr *hdr;
 808         struct meta_match *meta = NULL;
 809
 810         if (rtattr_parse(tb, TCA_EM_META_MAX, data, len) < 0)
 811                 goto errout;
 812
 813         if (tb[TCA_EM_META_HDR-1] == NULL ||
 814             RTA_PAYLOAD(tb[TCA_EM_META_HDR-1]) < sizeof(*hdr))
 815                 goto errout;
 816         hdr = RTA_DATA(tb[TCA_EM_META_HDR-1]);
 817
 818         if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
 819             TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
 820             TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
 821             TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
 822                 goto errout;
 823
 824         meta = kmalloc(sizeof(*meta), GFP_KERNEL);
 825         if (meta == NULL)
 826                 goto errout;
 827         memset(meta, 0, sizeof(*meta));
 828
 829         memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
 830         memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
 831
 832         if (!meta_is_supported(&meta->lvalue) ||
 833             !meta_is_supported(&meta->rvalue)) {
 834                 err = -EOPNOTSUPP;
 835                 goto errout;
 836         }
 837
 838         if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE-1]) < 0 ||
 839             meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE-1]) < 0)
 840                 goto errout;
 841
 842         m->datalen = sizeof(*meta);
 843         m->data = (unsigned long) meta;
 844
 845         err = 0;
 846 errout:
 847         if (err && meta)
 848                 meta_delete(meta);
 849         return err;
 850 }
 851
 852 static void em_meta_destroy(struct tcf_proto *tp, struct tcf_ematch *m)
 853 {
 854         if (m)
 855                 meta_delete((struct meta_match *) m->data);
 856 }
 857
 858 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
 859 {
 860         struct meta_match *meta = (struct meta_match *) em->data;
 861         struct tcf_meta_hdr hdr;
 862         struct meta_type_ops *ops;
 863
 864         memset(&hdr, 0, sizeof(hdr));
 865         memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
 866         memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
 867
 868         RTA_PUT(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr);
 869
 870         ops = meta_type_ops(&meta->lvalue);
 871         if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
 872             ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
 873                 goto rtattr_failure;
 874
 875         return 0;
 876
 877 rtattr_failure:
 878         return -1;
 879 }
 880
 881 static struct tcf_ematch_ops em_meta_ops = {
 882         .kind     = TCF_EM_META,
 883         .change   = em_meta_change,
 884         .match    = em_meta_match,
 885         .destroy  = em_meta_destroy,
 886         .dump     = em_meta_dump,
 887         .owner    = THIS_MODULE,
 888         .link     = LIST_HEAD_INIT(em_meta_ops.link)
 889 };
 890
 891 static int __init init_em_meta(void)
 892 {
 893         return tcf_em_register(&em_meta_ops);
 894 }
 895
 896 static void __exit exit_em_meta(void)
 897 {
 898         tcf_em_unregister(&em_meta_ops);
 899 }
 900
 901 MODULE_LICENSE("GPL");
 902
 903 module_init(init_em_meta);
 904 module_exit(exit_em_meta);