net/core/filter.c

   1 /*
   2  * Linux Socket Filter - Kernel level socket filtering
   3  *
   4  * Author:
   5  *     Jay Schulist <jschlst@samba.org>
   6  *
   7  * Based on the design of:
   8  *     - The Berkeley Packet Filter
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License
  12  * as published by the Free Software Foundation; either version
  13  * 2 of the License, or (at your option) any later version.
  14  *
  15  * Andi Kleen - Fix a few bad bugs and races.
  16  * Kris Katterjohn - Added many additional checks in sk_chk_filter()
  17  */
  18
  19 #include <linux/module.h>
  20 #include <linux/types.h>
  21 #include <linux/mm.h>
  22 #include <linux/fcntl.h>
  23 #include <linux/socket.h>
  24 #include <linux/in.h>
  25 #include <linux/inet.h>
  26 #include <linux/netdevice.h>
  27 #include <linux/if_packet.h>
  28 #include <linux/gfp.h>
  29 #include <net/ip.h>
  30 #include <net/protocol.h>
  31 #include <net/netlink.h>
  32 #include <linux/skbuff.h>
  33 #include <net/sock.h>
  34 #include <linux/errno.h>
  35 #include <linux/timer.h>
  36 #include <asm/system.h>
  37 #include <asm/uaccess.h>
  38 #include <asm/unaligned.h>
  39 #include <linux/filter.h>
  40
  41 /* No hurry in this branch */
  42 static void *__load_pointer(struct sk_buff *skb, int k)
  43 {
  44         u8 *ptr = NULL;
  45
  46         if (k >= SKF_NET_OFF)
  47                 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
  48         else if (k >= SKF_LL_OFF)
  49                 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
  50
  51         if (ptr >= skb->head && ptr < skb_tail_pointer(skb))
  52                 return ptr;
  53         return NULL;
  54 }
  55
  56 static inline void *load_pointer(struct sk_buff *skb, int k,
  57                                  unsigned int size, void *buffer)
  58 {
  59         if (k >= 0)
  60                 return skb_header_pointer(skb, k, size, buffer);
  61         else {
  62                 if (k >= SKF_AD_OFF)
  63                         return NULL;
  64                 return __load_pointer(skb, k);
  65         }
  66 }
  67
  68 /**
  69  *      sk_filter - run a packet through a socket filter
  70  *      @sk: sock associated with &sk_buff
  71  *      @skb: buffer to filter
  72  *
  73  * Run the filter code and then cut skb->data to correct size returned by
  74  * sk_run_filter. If pkt_len is 0 we toss packet. If skb->len is smaller
  75  * than pkt_len we keep whole skb->data. This is the socket level
  76  * wrapper to sk_run_filter. It returns 0 if the packet should
  77  * be accepted or -EPERM if the packet should be tossed.
  78  *
  79  */
  80 int sk_filter(struct sock *sk, struct sk_buff *skb)
  81 {
  82         int err;
  83         struct sk_filter *filter;
  84
  85         err = security_sock_rcv_skb(sk, skb);
  86         if (err)
  87                 return err;
  88
  89         rcu_read_lock_bh();
  90         filter = rcu_dereference_bh(sk->sk_filter);
  91         if (filter) {
  92                 unsigned int pkt_len = sk_run_filter(skb, filter->insns, filter->len);
  93
  94                 err = pkt_len ? pskb_trim(skb, pkt_len) : -EPERM;
  95         }
  96         rcu_read_unlock_bh();
  97
  98         return err;
  99 }
 100 EXPORT_SYMBOL(sk_filter);
 101
 102 /**
 103  *      sk_run_filter - run a filter on a socket
 104  *      @skb: buffer to run the filter on
 105  *      @filter: filter to apply
 106  *      @flen: length of filter
 107  *
 108  * Decode and apply filter instructions to the skb->data.
 109  * Return length to keep, 0 for none. skb is the data we are
 110  * filtering, filter is the array of filter instructions, and
 111  * len is the number of filter blocks in the array.
 112  */
 113 unsigned int sk_run_filter(struct sk_buff *skb, struct sock_filter *filter, int flen)
 114 {
 115         void *ptr;
 116         u32 A = 0;                      /* Accumulator */
 117         u32 X = 0;                      /* Index Register */
 118         u32 mem[BPF_MEMWORDS];          /* Scratch Memory Store */
 119         unsigned long memvalid = 0;
 120         u32 tmp;
 121         int k;
 122         int pc;
 123
 124         BUILD_BUG_ON(BPF_MEMWORDS > BITS_PER_LONG);
 125         /*
 126          * Process array of filter instructions.
 127          */
 128         for (pc = 0; pc < flen; pc++) {
 129                 const struct sock_filter *fentry = &filter[pc];
 130                 u32 f_k = fentry->k;
 131
 132                 switch (fentry->code) {
 133                 case BPF_S_ALU_ADD_X:
 134                         A += X;
 135                         continue;
 136                 case BPF_S_ALU_ADD_K:
 137                         A += f_k;
 138                         continue;
 139                 case BPF_S_ALU_SUB_X:
 140                         A -= X;
 141                         continue;
 142                 case BPF_S_ALU_SUB_K:
 143                         A -= f_k;
 144                         continue;
 145                 case BPF_S_ALU_MUL_X:
 146                         A *= X;
 147                         continue;
 148                 case BPF_S_ALU_MUL_K:
 149                         A *= f_k;
 150                         continue;
 151                 case BPF_S_ALU_DIV_X:
 152                         if (X == 0)
 153                                 return 0;
 154                         A /= X;
 155                         continue;
 156                 case BPF_S_ALU_DIV_K:
 157                         A /= f_k;
 158                         continue;
 159                 case BPF_S_ALU_AND_X:
 160                         A &= X;
 161                         continue;
 162                 case BPF_S_ALU_AND_K:
 163                         A &= f_k;
 164                         continue;
 165                 case BPF_S_ALU_OR_X:
 166                         A |= X;
 167                         continue;
 168                 case BPF_S_ALU_OR_K:
 169                         A |= f_k;
 170                         continue;
 171                 case BPF_S_ALU_LSH_X:
 172                         A <<= X;
 173                         continue;
 174                 case BPF_S_ALU_LSH_K:
 175                         A <<= f_k;
 176                         continue;
 177                 case BPF_S_ALU_RSH_X:
 178                         A >>= X;
 179                         continue;
 180                 case BPF_S_ALU_RSH_K:
 181                         A >>= f_k;
 182                         continue;
 183                 case BPF_S_ALU_NEG:
 184                         A = -A;
 185                         continue;
 186                 case BPF_S_JMP_JA:
 187                         pc += f_k;
 188                         continue;
 189                 case BPF_S_JMP_JGT_K:
 190                         pc += (A > f_k) ? fentry->jt : fentry->jf;
 191                         continue;
 192                 case BPF_S_JMP_JGE_K:
 193                         pc += (A >= f_k) ? fentry->jt : fentry->jf;
 194                         continue;
 195                 case BPF_S_JMP_JEQ_K:
 196                         pc += (A == f_k) ? fentry->jt : fentry->jf;
 197                         continue;
 198                 case BPF_S_JMP_JSET_K:
 199                         pc += (A & f_k) ? fentry->jt : fentry->jf;
 200                         continue;
 201                 case BPF_S_JMP_JGT_X:
 202                         pc += (A > X) ? fentry->jt : fentry->jf;
 203                         continue;
 204                 case BPF_S_JMP_JGE_X:
 205                         pc += (A >= X) ? fentry->jt : fentry->jf;
 206                         continue;
 207                 case BPF_S_JMP_JEQ_X:
 208                         pc += (A == X) ? fentry->jt : fentry->jf;
 209                         continue;
 210                 case BPF_S_JMP_JSET_X:
 211                         pc += (A & X) ? fentry->jt : fentry->jf;
 212                         continue;
 213                 case BPF_S_LD_W_ABS:
 214                         k = f_k;
 215 load_w:
 216                         ptr = load_pointer(skb, k, 4, &tmp);
 217                         if (ptr != NULL) {
 218                                 A = get_unaligned_be32(ptr);
 219                                 continue;
 220                         }
 221                         break;
 222                 case BPF_S_LD_H_ABS:
 223                         k = f_k;
 224 load_h:
 225                         ptr = load_pointer(skb, k, 2, &tmp);
 226                         if (ptr != NULL) {
 227                                 A = get_unaligned_be16(ptr);
 228                                 continue;
 229                         }
 230                         break;
 231                 case BPF_S_LD_B_ABS:
 232                         k = f_k;
 233 load_b:
 234                         ptr = load_pointer(skb, k, 1, &tmp);
 235                         if (ptr != NULL) {
 236                                 A = *(u8 *)ptr;
 237                                 continue;
 238                         }
 239                         break;
 240                 case BPF_S_LD_W_LEN:
 241                         A = skb->len;
 242                         continue;
 243                 case BPF_S_LDX_W_LEN:
 244                         X = skb->len;
 245                         continue;
 246                 case BPF_S_LD_W_IND:
 247                         k = X + f_k;
 248                         goto load_w;
 249                 case BPF_S_LD_H_IND:
 250                         k = X + f_k;
 251                         goto load_h;
 252                 case BPF_S_LD_B_IND:
 253                         k = X + f_k;
 254                         goto load_b;
 255                 case BPF_S_LDX_B_MSH:
 256                         ptr = load_pointer(skb, f_k, 1, &tmp);
 257                         if (ptr != NULL) {
 258                                 X = (*(u8 *)ptr & 0xf) << 2;
 259                                 continue;
 260                         }
 261                         return 0;
 262                 case BPF_S_LD_IMM:
 263                         A = f_k;
 264                         continue;
 265                 case BPF_S_LDX_IMM:
 266                         X = f_k;
 267                         continue;
 268                 case BPF_S_LD_MEM:
 269                         A = (memvalid & (1UL << f_k)) ?
 270                                 mem[f_k] : 0;
 271                         continue;
 272                 case BPF_S_LDX_MEM:
 273                         X = (memvalid & (1UL << f_k)) ?
 274                                 mem[f_k] : 0;
 275                         continue;
 276                 case BPF_S_MISC_TAX:
 277                         X = A;
 278                         continue;
 279                 case BPF_S_MISC_TXA:
 280                         A = X;
 281                         continue;
 282                 case BPF_S_RET_K:
 283                         return f_k;
 284                 case BPF_S_RET_A:
 285                         return A;
 286                 case BPF_S_ST:
 287                         memvalid |= 1UL << f_k;
 288                         mem[f_k] = A;
 289                         continue;
 290                 case BPF_S_STX:
 291                         memvalid |= 1UL << f_k;
 292                         mem[f_k] = X;
 293                         continue;
 294                 default:
 295                         WARN_ON(1);
 296                         return 0;
 297                 }
 298
 299                 /*
 300                  * Handle ancillary data, which are impossible
 301                  * (or very difficult) to get parsing packet contents.
 302                  */
 303                 switch (k-SKF_AD_OFF) {
 304                 case SKF_AD_PROTOCOL:
 305                         A = ntohs(skb->protocol);
 306                         continue;
 307                 case SKF_AD_PKTTYPE:
 308                         A = skb->pkt_type;
 309                         continue;
 310                 case SKF_AD_IFINDEX:
 311                         if (!skb->dev)
 312                                 return 0;
 313                         A = skb->dev->ifindex;
 314                         continue;
 315                 case SKF_AD_MARK:
 316                         A = skb->mark;
 317                         continue;
 318                 case SKF_AD_QUEUE:
 319                         A = skb->queue_mapping;
 320                         continue;
 321                 case SKF_AD_HATYPE:
 322                         if (!skb->dev)
 323                                 return 0;
 324                         A = skb->dev->type;
 325                         continue;
 326                 case SKF_AD_NLATTR: {
 327                         struct nlattr *nla;
 328
 329                         if (skb_is_nonlinear(skb))
 330                                 return 0;
 331                         if (A > skb->len - sizeof(struct nlattr))
 332                                 return 0;
 333
 334                         nla = nla_find((struct nlattr *)&skb->data[A],
 335                                        skb->len - A, X);
 336                         if (nla)
 337                                 A = (void *)nla - (void *)skb->data;
 338                         else
 339                                 A = 0;
 340                         continue;
 341                 }
 342                 case SKF_AD_NLATTR_NEST: {
 343                         struct nlattr *nla;
 344
 345                         if (skb_is_nonlinear(skb))
 346                                 return 0;
 347                         if (A > skb->len - sizeof(struct nlattr))
 348                                 return 0;
 349
 350                         nla = (struct nlattr *)&skb->data[A];
 351                         if (nla->nla_len > A - skb->len)
 352                                 return 0;
 353
 354                         nla = nla_find_nested(nla, X);
 355                         if (nla)
 356                                 A = (void *)nla - (void *)skb->data;
 357                         else
 358                                 A = 0;
 359                         continue;
 360                 }
 361                 default:
 362                         return 0;
 363                 }
 364         }
 365
 366         return 0;
 367 }
 368 EXPORT_SYMBOL(sk_run_filter);
 369
 370 /**
 371  *      sk_chk_filter - verify socket filter code
 372  *      @filter: filter to verify
 373  *      @flen: length of filter
 374  *
 375  * Check the user's filter code. If we let some ugly
 376  * filter code slip through kaboom! The filter must contain
 377  * no references or jumps that are out of range, no illegal
 378  * instructions, and must end with a RET instruction.
 379  *
 380  * All jumps are forward as they are not signed.
 381  *
 382  * Returns 0 if the rule set is legal or -EINVAL if not.
 383  */
 384 int sk_chk_filter(struct sock_filter *filter, int flen)
 385 {
 386         /*
 387          * Valid instructions are initialized to non-0.
 388          * Invalid instructions are initialized to 0.
 389          */
 390         static const u8 codes[] = {
 391                 [BPF_ALU|BPF_ADD|BPF_K]  = BPF_S_ALU_ADD_K + 1,
 392                 [BPF_ALU|BPF_ADD|BPF_X]  = BPF_S_ALU_ADD_X + 1,
 393                 [BPF_ALU|BPF_SUB|BPF_K]  = BPF_S_ALU_SUB_K + 1,
 394                 [BPF_ALU|BPF_SUB|BPF_X]  = BPF_S_ALU_SUB_X + 1,
 395                 [BPF_ALU|BPF_MUL|BPF_K]  = BPF_S_ALU_MUL_K + 1,
 396                 [BPF_ALU|BPF_MUL|BPF_X]  = BPF_S_ALU_MUL_X + 1,
 397                 [BPF_ALU|BPF_DIV|BPF_X]  = BPF_S_ALU_DIV_X + 1,
 398                 [BPF_ALU|BPF_AND|BPF_K]  = BPF_S_ALU_AND_K + 1,
 399                 [BPF_ALU|BPF_AND|BPF_X]  = BPF_S_ALU_AND_X + 1,
 400                 [BPF_ALU|BPF_OR|BPF_K]   = BPF_S_ALU_OR_K + 1,
 401                 [BPF_ALU|BPF_OR|BPF_X]   = BPF_S_ALU_OR_X + 1,
 402                 [BPF_ALU|BPF_LSH|BPF_K]  = BPF_S_ALU_LSH_K + 1,
 403                 [BPF_ALU|BPF_LSH|BPF_X]  = BPF_S_ALU_LSH_X + 1,
 404                 [BPF_ALU|BPF_RSH|BPF_K]  = BPF_S_ALU_RSH_K + 1,
 405                 [BPF_ALU|BPF_RSH|BPF_X]  = BPF_S_ALU_RSH_X + 1,
 406                 [BPF_ALU|BPF_NEG]        = BPF_S_ALU_NEG + 1,
 407                 [BPF_LD|BPF_W|BPF_ABS]   = BPF_S_LD_W_ABS + 1,
 408                 [BPF_LD|BPF_H|BPF_ABS]   = BPF_S_LD_H_ABS + 1,
 409                 [BPF_LD|BPF_B|BPF_ABS]   = BPF_S_LD_B_ABS + 1,
 410                 [BPF_LD|BPF_W|BPF_LEN]   = BPF_S_LD_W_LEN + 1,
 411                 [BPF_LD|BPF_W|BPF_IND]   = BPF_S_LD_W_IND + 1,
 412                 [BPF_LD|BPF_H|BPF_IND]   = BPF_S_LD_H_IND + 1,
 413                 [BPF_LD|BPF_B|BPF_IND]   = BPF_S_LD_B_IND + 1,
 414                 [BPF_LD|BPF_IMM]         = BPF_S_LD_IMM + 1,
 415                 [BPF_LDX|BPF_W|BPF_LEN]  = BPF_S_LDX_W_LEN + 1,
 416                 [BPF_LDX|BPF_B|BPF_MSH]  = BPF_S_LDX_B_MSH + 1,
 417                 [BPF_LDX|BPF_IMM]        = BPF_S_LDX_IMM + 1,
 418                 [BPF_MISC|BPF_TAX]       = BPF_S_MISC_TAX + 1,
 419                 [BPF_MISC|BPF_TXA]       = BPF_S_MISC_TXA + 1,
 420                 [BPF_RET|BPF_K]          = BPF_S_RET_K + 1,
 421                 [BPF_RET|BPF_A]          = BPF_S_RET_A + 1,
 422                 [BPF_ALU|BPF_DIV|BPF_K]  = BPF_S_ALU_DIV_K + 1,
 423                 [BPF_LD|BPF_MEM]         = BPF_S_LD_MEM + 1,
 424                 [BPF_LDX|BPF_MEM]        = BPF_S_LDX_MEM + 1,
 425                 [BPF_ST]                 = BPF_S_ST + 1,
 426                 [BPF_STX]                = BPF_S_STX + 1,
 427                 [BPF_JMP|BPF_JA]         = BPF_S_JMP_JA + 1,
 428                 [BPF_JMP|BPF_JEQ|BPF_K]  = BPF_S_JMP_JEQ_K + 1,
 429                 [BPF_JMP|BPF_JEQ|BPF_X]  = BPF_S_JMP_JEQ_X + 1,
 430                 [BPF_JMP|BPF_JGE|BPF_K]  = BPF_S_JMP_JGE_K + 1,
 431                 [BPF_JMP|BPF_JGE|BPF_X]  = BPF_S_JMP_JGE_X + 1,
 432                 [BPF_JMP|BPF_JGT|BPF_K]  = BPF_S_JMP_JGT_K + 1,
 433                 [BPF_JMP|BPF_JGT|BPF_X]  = BPF_S_JMP_JGT_X + 1,
 434                 [BPF_JMP|BPF_JSET|BPF_K] = BPF_S_JMP_JSET_K + 1,
 435                 [BPF_JMP|BPF_JSET|BPF_X] = BPF_S_JMP_JSET_X + 1,
 436         };
 437         int pc;
 438
 439         if (flen == 0 || flen > BPF_MAXINSNS)
 440                 return -EINVAL;
 441
 442         /* check the filter code now */
 443         for (pc = 0; pc < flen; pc++) {
 444                 struct sock_filter *ftest = &filter[pc];
 445                 u16 code = ftest->code;
 446
 447                 if (code >= ARRAY_SIZE(codes))
 448                         return -EINVAL;
 449                 code = codes[code];
 450                 /* Undo the '+ 1' in codes[] after validation. */
 451                 if (!code--)
 452                         return -EINVAL;
 453                 /* Some instructions need special checks */
 454                 switch (code) {
 455                 case BPF_S_ALU_DIV_K:
 456                         /* check for division by zero */
 457                         if (ftest->k == 0)
 458                                 return -EINVAL;
 459                         break;
 460                 case BPF_S_LD_MEM:
 461                 case BPF_S_LDX_MEM:
 462                 case BPF_S_ST:
 463                 case BPF_S_STX:
 464                         /* check for invalid memory addresses */
 465                         if (ftest->k >= BPF_MEMWORDS)
 466                                 return -EINVAL;
 467                         break;
 468                 case BPF_S_JMP_JA:
 469                         /*
 470                          * Note, the large ftest->k might cause loops.
 471                          * Compare this with conditional jumps below,
 472                          * where offsets are limited. --ANK (981016)
 473                          */
 474                         if (ftest->k >= (unsigned)(flen-pc-1))
 475                                 return -EINVAL;
 476                         break;
 477                 case BPF_S_JMP_JEQ_K:
 478                 case BPF_S_JMP_JEQ_X:
 479                 case BPF_S_JMP_JGE_K:
 480                 case BPF_S_JMP_JGE_X:
 481                 case BPF_S_JMP_JGT_K:
 482                 case BPF_S_JMP_JGT_X:
 483                 case BPF_S_JMP_JSET_X:
 484                 case BPF_S_JMP_JSET_K:
 485                         /* for conditionals both must be safe */
 486                         if (pc + ftest->jt + 1 >= flen ||
 487                             pc + ftest->jf + 1 >= flen)
 488                                 return -EINVAL;
 489                         break;
 490                 }
 491                 ftest->code = code;
 492         }
 493
 494         /* last instruction must be a RET code */
 495         switch (filter[flen - 1].code) {
 496         case BPF_S_RET_K:
 497         case BPF_S_RET_A:
 498                 return 0;
 499         }
 500         return -EINVAL;
 501 }
 502 EXPORT_SYMBOL(sk_chk_filter);
 503
 504 /**
 505  *      sk_filter_rcu_release: Release a socket filter by rcu_head
 506  *      @rcu: rcu_head that contains the sk_filter to free
 507  */
 508 static void sk_filter_rcu_release(struct rcu_head *rcu)
 509 {
 510         struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
 511
 512         sk_filter_release(fp);
 513 }
 514
 515 static void sk_filter_delayed_uncharge(struct sock *sk, struct sk_filter *fp)
 516 {
 517         unsigned int size = sk_filter_len(fp);
 518
 519         atomic_sub(size, &sk->sk_omem_alloc);
 520         call_rcu_bh(&fp->rcu, sk_filter_rcu_release);
 521 }
 522
 523 /**
 524  *      sk_attach_filter - attach a socket filter
 525  *      @fprog: the filter program
 526  *      @sk: the socket to use
 527  *
 528  * Attach the user's filter code. We first run some sanity checks on
 529  * it to make sure it does not explode on us later. If an error
 530  * occurs or there is insufficient memory for the filter a negative
 531  * errno code is returned. On success the return is zero.
 532  */
 533 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
 534 {
 535         struct sk_filter *fp, *old_fp;
 536         unsigned int fsize = sizeof(struct sock_filter) * fprog->len;
 537         int err;
 538
 539         /* Make sure new filter is there and in the right amounts. */
 540         if (fprog->filter == NULL)
 541                 return -EINVAL;
 542
 543         fp = sock_kmalloc(sk, fsize+sizeof(*fp), GFP_KERNEL);
 544         if (!fp)
 545                 return -ENOMEM;
 546         if (copy_from_user(fp->insns, fprog->filter, fsize)) {
 547                 sock_kfree_s(sk, fp, fsize+sizeof(*fp));
 548                 return -EFAULT;
 549         }
 550
 551         atomic_set(&fp->refcnt, 1);
 552         fp->len = fprog->len;
 553
 554         err = sk_chk_filter(fp->insns, fp->len);
 555         if (err) {
 556                 sk_filter_uncharge(sk, fp);
 557                 return err;
 558         }
 559
 560         old_fp = rcu_dereference_protected(sk->sk_filter,
 561                                            sock_owned_by_user(sk));
 562         rcu_assign_pointer(sk->sk_filter, fp);
 563
 564         if (old_fp)
 565                 sk_filter_delayed_uncharge(sk, old_fp);
 566         return 0;
 567 }
 568 EXPORT_SYMBOL_GPL(sk_attach_filter);
 569
 570 int sk_detach_filter(struct sock *sk)
 571 {
 572         int ret = -ENOENT;
 573         struct sk_filter *filter;
 574
 575         filter = rcu_dereference_protected(sk->sk_filter,
 576                                            sock_owned_by_user(sk));
 577         if (filter) {
 578                 rcu_assign_pointer(sk->sk_filter, NULL);
 579                 sk_filter_delayed_uncharge(sk, filter);
 580                 ret = 0;
 581         }
 582         return ret;
 583 }
 584 EXPORT_SYMBOL_GPL(sk_detach_filter);