1/* 2 * Linux Socket Filter - Kernel level socket filtering 3 * 4 * Based on the design of the Berkeley Packet Filter. The new 5 * internal format has been designed by PLUMgrid: 6 * 7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 8 * 9 * Authors: 10 * 11 * Jay Schulist <jschlst@samba.org> 12 * Alexei Starovoitov <ast@plumgrid.com> 13 * Daniel Borkmann <dborkman@redhat.com> 14 * 15 * This program is free software; you can redistribute it and/or 16 * modify it under the terms of the GNU General Public License 17 * as published by the Free Software Foundation; either version 18 * 2 of the License, or (at your option) any later version. 19 * 20 * Andi Kleen - Fix a few bad bugs and races. 21 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 22 */ 23 24#include <linux/module.h> 25#include <linux/types.h> 26#include <linux/mm.h> 27#include <linux/fcntl.h> 28#include <linux/socket.h> 29#include <linux/in.h> 30#include <linux/inet.h> 31#include <linux/netdevice.h> 32#include <linux/if_packet.h> 33#include <linux/gfp.h> 34#include <net/ip.h> 35#include <net/protocol.h> 36#include <net/netlink.h> 37#include <linux/skbuff.h> 38#include <net/sock.h> 39#include <linux/errno.h> 40#include <linux/timer.h> 41#include <asm/uaccess.h> 42#include <asm/unaligned.h> 43#include <linux/filter.h> 44#include <linux/ratelimit.h> 45#include <linux/seccomp.h> 46#include <linux/if_vlan.h> 47 48/** 49 * sk_filter - run a packet through a socket filter 50 * @sk: sock associated with &sk_buff 51 * @skb: buffer to filter 52 * 53 * Run the filter code and then cut skb->data to correct size returned by 54 * SK_RUN_FILTER. If pkt_len is 0 we toss packet. If skb->len is smaller 55 * than pkt_len we keep whole skb->data. This is the socket level 56 * wrapper to SK_RUN_FILTER. It returns 0 if the packet should 57 * be accepted or -EPERM if the packet should be tossed. 58 * 59 */ 60int sk_filter(struct sock *sk, struct sk_buff *skb) 61{ 62 int err; 63 struct sk_filter *filter; 64 65 /* 66 * If the skb was allocated from pfmemalloc reserves, only 67 * allow SOCK_MEMALLOC sockets to use it as this socket is 68 * helping free memory 69 */ 70 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC)) 71 return -ENOMEM; 72 73 err = security_sock_rcv_skb(sk, skb); 74 if (err) 75 return err; 76 77 rcu_read_lock(); 78 filter = rcu_dereference(sk->sk_filter); 79 if (filter) { 80 unsigned int pkt_len = SK_RUN_FILTER(filter, skb); 81 82 err = pkt_len ? pskb_trim(skb, pkt_len) : -EPERM; 83 } 84 rcu_read_unlock(); 85 86 return err; 87} 88EXPORT_SYMBOL(sk_filter); 89 90static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5) 91{ 92 return skb_get_poff((struct sk_buff *)(unsigned long) ctx); 93} 94 95static u64 __skb_get_nlattr(u64 ctx, u64 a, u64 x, u64 r4, u64 r5) 96{ 97 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx; 98 struct nlattr *nla; 99 100 if (skb_is_nonlinear(skb)) 101 return 0; 102 103 if (skb->len < sizeof(struct nlattr)) 104 return 0; 105 106 if (a > skb->len - sizeof(struct nlattr)) 107 return 0; 108 109 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x); 110 if (nla) 111 return (void *) nla - (void *) skb->data; 112 113 return 0; 114} 115 116static u64 __skb_get_nlattr_nest(u64 ctx, u64 a, u64 x, u64 r4, u64 r5) 117{ 118 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx; 119 struct nlattr *nla; 120 121 if (skb_is_nonlinear(skb)) 122 return 0; 123 124 if (skb->len < sizeof(struct nlattr)) 125 return 0; 126 127 if (a > skb->len - sizeof(struct nlattr)) 128 return 0; 129 130 nla = (struct nlattr *) &skb->data[a]; 131 if (nla->nla_len > skb->len - a) 132 return 0; 133 134 nla = nla_find_nested(nla, x); 135 if (nla) 136 return (void *) nla - (void *) skb->data; 137 138 return 0; 139} 140 141static u64 __get_raw_cpu_id(u64 ctx, u64 a, u64 x, u64 r4, u64 r5) 142{ 143 return raw_smp_processor_id(); 144} 145 146/* note that this only generates 32-bit random numbers */ 147static u64 __get_random_u32(u64 ctx, u64 a, u64 x, u64 r4, u64 r5) 148{ 149 return prandom_u32(); 150} 151 152static bool convert_bpf_extensions(struct sock_filter *fp, 153 struct bpf_insn **insnp) 154{ 155 struct bpf_insn *insn = *insnp; 156 157 switch (fp->k) { 158 case SKF_AD_OFF + SKF_AD_PROTOCOL: 159 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2); 160 161 /* A = *(u16 *) (CTX + offsetof(protocol)) */ 162 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX, 163 offsetof(struct sk_buff, protocol)); 164 /* A = ntohs(A) [emitting a nop or swap16] */ 165 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16); 166 break; 167 168 case SKF_AD_OFF + SKF_AD_PKTTYPE: 169 *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX, 170 PKT_TYPE_OFFSET()); 171 *insn = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, PKT_TYPE_MAX); 172#ifdef __BIG_ENDIAN_BITFIELD 173 insn++; 174 *insn = BPF_ALU32_IMM(BPF_RSH, BPF_REG_A, 5); 175#endif 176 break; 177 178 case SKF_AD_OFF + SKF_AD_IFINDEX: 179 case SKF_AD_OFF + SKF_AD_HATYPE: 180 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4); 181 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2); 182 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0); 183 184 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)), 185 BPF_REG_TMP, BPF_REG_CTX, 186 offsetof(struct sk_buff, dev)); 187 /* if (tmp != 0) goto pc + 1 */ 188 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1); 189 *insn++ = BPF_EXIT_INSN(); 190 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX) 191 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP, 192 offsetof(struct net_device, ifindex)); 193 else 194 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP, 195 offsetof(struct net_device, type)); 196 break; 197 198 case SKF_AD_OFF + SKF_AD_MARK: 199 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 200 201 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, 202 offsetof(struct sk_buff, mark)); 203 break; 204 205 case SKF_AD_OFF + SKF_AD_RXHASH: 206 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4); 207 208 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, 209 offsetof(struct sk_buff, hash)); 210 break; 211 212 case SKF_AD_OFF + SKF_AD_QUEUE: 213 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2); 214 215 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX, 216 offsetof(struct sk_buff, queue_mapping)); 217 break; 218 219 case SKF_AD_OFF + SKF_AD_VLAN_TAG: 220 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT: 221 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 222 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000); 223 224 /* A = *(u16 *) (CTX + offsetof(vlan_tci)) */ 225 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX, 226 offsetof(struct sk_buff, vlan_tci)); 227 if (fp->k == SKF_AD_OFF + SKF_AD_VLAN_TAG) { 228 *insn = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 229 ~VLAN_TAG_PRESENT); 230 } else { 231 /* A >>= 12 */ 232 *insn++ = BPF_ALU32_IMM(BPF_RSH, BPF_REG_A, 12); 233 /* A &= 1 */ 234 *insn = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 1); 235 } 236 break; 237 238 case SKF_AD_OFF + SKF_AD_PAY_OFFSET: 239 case SKF_AD_OFF + SKF_AD_NLATTR: 240 case SKF_AD_OFF + SKF_AD_NLATTR_NEST: 241 case SKF_AD_OFF + SKF_AD_CPU: 242 case SKF_AD_OFF + SKF_AD_RANDOM: 243 /* arg1 = CTX */ 244 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX); 245 /* arg2 = A */ 246 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A); 247 /* arg3 = X */ 248 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X); 249 /* Emit call(arg1=CTX, arg2=A, arg3=X) */ 250 switch (fp->k) { 251 case SKF_AD_OFF + SKF_AD_PAY_OFFSET: 252 *insn = BPF_EMIT_CALL(__skb_get_pay_offset); 253 break; 254 case SKF_AD_OFF + SKF_AD_NLATTR: 255 *insn = BPF_EMIT_CALL(__skb_get_nlattr); 256 break; 257 case SKF_AD_OFF + SKF_AD_NLATTR_NEST: 258 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest); 259 break; 260 case SKF_AD_OFF + SKF_AD_CPU: 261 *insn = BPF_EMIT_CALL(__get_raw_cpu_id); 262 break; 263 case SKF_AD_OFF + SKF_AD_RANDOM: 264 *insn = BPF_EMIT_CALL(__get_random_u32); 265 break; 266 } 267 break; 268 269 case SKF_AD_OFF + SKF_AD_ALU_XOR_X: 270 /* A ^= X */ 271 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X); 272 break; 273 274 default: 275 /* This is just a dummy call to avoid letting the compiler 276 * evict __bpf_call_base() as an optimization. Placed here 277 * where no-one bothers. 278 */ 279 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0); 280 return false; 281 } 282 283 *insnp = insn; 284 return true; 285} 286 287/** 288 * bpf_convert_filter - convert filter program 289 * @prog: the user passed filter program 290 * @len: the length of the user passed filter program 291 * @new_prog: buffer where converted program will be stored 292 * @new_len: pointer to store length of converted program 293 * 294 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style. 295 * Conversion workflow: 296 * 297 * 1) First pass for calculating the new program length: 298 * bpf_convert_filter(old_prog, old_len, NULL, &new_len) 299 * 300 * 2) 2nd pass to remap in two passes: 1st pass finds new 301 * jump offsets, 2nd pass remapping: 302 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len); 303 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len); 304 * 305 * User BPF's register A is mapped to our BPF register 6, user BPF 306 * register X is mapped to BPF register 7; frame pointer is always 307 * register 10; Context 'void *ctx' is stored in register 1, that is, 308 * for socket filters: ctx == 'struct sk_buff *', for seccomp: 309 * ctx == 'struct seccomp_data *'. 310 */ 311int bpf_convert_filter(struct sock_filter *prog, int len, 312 struct bpf_insn *new_prog, int *new_len) 313{ 314 int new_flen = 0, pass = 0, target, i; 315 struct bpf_insn *new_insn; 316 struct sock_filter *fp; 317 int *addrs = NULL; 318 u8 bpf_src; 319 320 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK); 321 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG); 322 323 if (len <= 0 || len > BPF_MAXINSNS) 324 return -EINVAL; 325 326 if (new_prog) { 327 addrs = kcalloc(len, sizeof(*addrs), GFP_KERNEL); 328 if (!addrs) 329 return -ENOMEM; 330 } 331 332do_pass: 333 new_insn = new_prog; 334 fp = prog; 335 336 if (new_insn) 337 *new_insn = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1); 338 new_insn++; 339 340 for (i = 0; i < len; fp++, i++) { 341 struct bpf_insn tmp_insns[6] = { }; 342 struct bpf_insn *insn = tmp_insns; 343 344 if (addrs) 345 addrs[i] = new_insn - new_prog; 346 347 switch (fp->code) { 348 /* All arithmetic insns and skb loads map as-is. */ 349 case BPF_ALU | BPF_ADD | BPF_X: 350 case BPF_ALU | BPF_ADD | BPF_K: 351 case BPF_ALU | BPF_SUB | BPF_X: 352 case BPF_ALU | BPF_SUB | BPF_K: 353 case BPF_ALU | BPF_AND | BPF_X: 354 case BPF_ALU | BPF_AND | BPF_K: 355 case BPF_ALU | BPF_OR | BPF_X: 356 case BPF_ALU | BPF_OR | BPF_K: 357 case BPF_ALU | BPF_LSH | BPF_X: 358 case BPF_ALU | BPF_LSH | BPF_K: 359 case BPF_ALU | BPF_RSH | BPF_X: 360 case BPF_ALU | BPF_RSH | BPF_K: 361 case BPF_ALU | BPF_XOR | BPF_X: 362 case BPF_ALU | BPF_XOR | BPF_K: 363 case BPF_ALU | BPF_MUL | BPF_X: 364 case BPF_ALU | BPF_MUL | BPF_K: 365 case BPF_ALU | BPF_DIV | BPF_X: 366 case BPF_ALU | BPF_DIV | BPF_K: 367 case BPF_ALU | BPF_MOD | BPF_X: 368 case BPF_ALU | BPF_MOD | BPF_K: 369 case BPF_ALU | BPF_NEG: 370 case BPF_LD | BPF_ABS | BPF_W: 371 case BPF_LD | BPF_ABS | BPF_H: 372 case BPF_LD | BPF_ABS | BPF_B: 373 case BPF_LD | BPF_IND | BPF_W: 374 case BPF_LD | BPF_IND | BPF_H: 375 case BPF_LD | BPF_IND | BPF_B: 376 /* Check for overloaded BPF extension and 377 * directly convert it if found, otherwise 378 * just move on with mapping. 379 */ 380 if (BPF_CLASS(fp->code) == BPF_LD && 381 BPF_MODE(fp->code) == BPF_ABS && 382 convert_bpf_extensions(fp, &insn)) 383 break; 384 385 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k); 386 break; 387 388 /* Jump transformation cannot use BPF block macros 389 * everywhere as offset calculation and target updates 390 * require a bit more work than the rest, i.e. jump 391 * opcodes map as-is, but offsets need adjustment. 392 */ 393 394#define BPF_EMIT_JMP \ 395 do { \ 396 if (target >= len || target < 0) \ 397 goto err; \ 398 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \ 399 /* Adjust pc relative offset for 2nd or 3rd insn. */ \ 400 insn->off -= insn - tmp_insns; \ 401 } while (0) 402 403 case BPF_JMP | BPF_JA: 404 target = i + fp->k + 1; 405 insn->code = fp->code; 406 BPF_EMIT_JMP; 407 break; 408 409 case BPF_JMP | BPF_JEQ | BPF_K: 410 case BPF_JMP | BPF_JEQ | BPF_X: 411 case BPF_JMP | BPF_JSET | BPF_K: 412 case BPF_JMP | BPF_JSET | BPF_X: 413 case BPF_JMP | BPF_JGT | BPF_K: 414 case BPF_JMP | BPF_JGT | BPF_X: 415 case BPF_JMP | BPF_JGE | BPF_K: 416 case BPF_JMP | BPF_JGE | BPF_X: 417 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) { 418 /* BPF immediates are signed, zero extend 419 * immediate into tmp register and use it 420 * in compare insn. 421 */ 422 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k); 423 424 insn->dst_reg = BPF_REG_A; 425 insn->src_reg = BPF_REG_TMP; 426 bpf_src = BPF_X; 427 } else { 428 insn->dst_reg = BPF_REG_A; 429 insn->src_reg = BPF_REG_X; 430 insn->imm = fp->k; 431 bpf_src = BPF_SRC(fp->code); 432 } 433 434 /* Common case where 'jump_false' is next insn. */ 435 if (fp->jf == 0) { 436 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src; 437 target = i + fp->jt + 1; 438 BPF_EMIT_JMP; 439 break; 440 } 441 442 /* Convert JEQ into JNE when 'jump_true' is next insn. */ 443 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) { 444 insn->code = BPF_JMP | BPF_JNE | bpf_src; 445 target = i + fp->jf + 1; 446 BPF_EMIT_JMP; 447 break; 448 } 449 450 /* Other jumps are mapped into two insns: Jxx and JA. */ 451 target = i + fp->jt + 1; 452 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src; 453 BPF_EMIT_JMP; 454 insn++; 455 456 insn->code = BPF_JMP | BPF_JA; 457 target = i + fp->jf + 1; 458 BPF_EMIT_JMP; 459 break; 460 461 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */ 462 case BPF_LDX | BPF_MSH | BPF_B: 463 /* tmp = A */ 464 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A); 465 /* A = BPF_R0 = *(u8 *) (skb->data + K) */ 466 *insn++ = BPF_LD_ABS(BPF_B, fp->k); 467 /* A &= 0xf */ 468 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf); 469 /* A <<= 2 */ 470 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2); 471 /* X = A */ 472 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A); 473 /* A = tmp */ 474 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP); 475 break; 476 477 /* RET_K, RET_A are remaped into 2 insns. */ 478 case BPF_RET | BPF_A: 479 case BPF_RET | BPF_K: 480 *insn++ = BPF_MOV32_RAW(BPF_RVAL(fp->code) == BPF_K ? 481 BPF_K : BPF_X, BPF_REG_0, 482 BPF_REG_A, fp->k); 483 *insn = BPF_EXIT_INSN(); 484 break; 485 486 /* Store to stack. */ 487 case BPF_ST: 488 case BPF_STX: 489 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) == 490 BPF_ST ? BPF_REG_A : BPF_REG_X, 491 -(BPF_MEMWORDS - fp->k) * 4); 492 break; 493 494 /* Load from stack. */ 495 case BPF_LD | BPF_MEM: 496 case BPF_LDX | BPF_MEM: 497 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ? 498 BPF_REG_A : BPF_REG_X, BPF_REG_FP, 499 -(BPF_MEMWORDS - fp->k) * 4); 500 break; 501 502 /* A = K or X = K */ 503 case BPF_LD | BPF_IMM: 504 case BPF_LDX | BPF_IMM: 505 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ? 506 BPF_REG_A : BPF_REG_X, fp->k); 507 break; 508 509 /* X = A */ 510 case BPF_MISC | BPF_TAX: 511 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A); 512 break; 513 514 /* A = X */ 515 case BPF_MISC | BPF_TXA: 516 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X); 517 break; 518 519 /* A = skb->len or X = skb->len */ 520 case BPF_LD | BPF_W | BPF_LEN: 521 case BPF_LDX | BPF_W | BPF_LEN: 522 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ? 523 BPF_REG_A : BPF_REG_X, BPF_REG_CTX, 524 offsetof(struct sk_buff, len)); 525 break; 526 527 /* Access seccomp_data fields. */ 528 case BPF_LDX | BPF_ABS | BPF_W: 529 /* A = *(u32 *) (ctx + K) */ 530 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k); 531 break; 532 533 /* Unkown instruction. */ 534 default: 535 goto err; 536 } 537 538 insn++; 539 if (new_prog) 540 memcpy(new_insn, tmp_insns, 541 sizeof(*insn) * (insn - tmp_insns)); 542 new_insn += insn - tmp_insns; 543 } 544 545 if (!new_prog) { 546 /* Only calculating new length. */ 547 *new_len = new_insn - new_prog; 548 return 0; 549 } 550 551 pass++; 552 if (new_flen != new_insn - new_prog) { 553 new_flen = new_insn - new_prog; 554 if (pass > 2) 555 goto err; 556 goto do_pass; 557 } 558 559 kfree(addrs); 560 BUG_ON(*new_len != new_flen); 561 return 0; 562err: 563 kfree(addrs); 564 return -EINVAL; 565} 566 567/* Security: 568 * 569 * As we dont want to clear mem[] array for each packet going through 570 * __bpf_prog_run(), we check that filter loaded by user never try to read 571 * a cell if not previously written, and we check all branches to be sure 572 * a malicious user doesn't try to abuse us. 573 */ 574static int check_load_and_stores(const struct sock_filter *filter, int flen) 575{ 576 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */ 577 int pc, ret = 0; 578 579 BUILD_BUG_ON(BPF_MEMWORDS > 16); 580 581 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL); 582 if (!masks) 583 return -ENOMEM; 584 585 memset(masks, 0xff, flen * sizeof(*masks)); 586 587 for (pc = 0; pc < flen; pc++) { 588 memvalid &= masks[pc]; 589 590 switch (filter[pc].code) { 591 case BPF_ST: 592 case BPF_STX: 593 memvalid |= (1 << filter[pc].k); 594 break; 595 case BPF_LD | BPF_MEM: 596 case BPF_LDX | BPF_MEM: 597 if (!(memvalid & (1 << filter[pc].k))) { 598 ret = -EINVAL; 599 goto error; 600 } 601 break; 602 case BPF_JMP | BPF_JA: 603 /* A jump must set masks on target */ 604 masks[pc + 1 + filter[pc].k] &= memvalid; 605 memvalid = ~0; 606 break; 607 case BPF_JMP | BPF_JEQ | BPF_K: 608 case BPF_JMP | BPF_JEQ | BPF_X: 609 case BPF_JMP | BPF_JGE | BPF_K: 610 case BPF_JMP | BPF_JGE | BPF_X: 611 case BPF_JMP | BPF_JGT | BPF_K: 612 case BPF_JMP | BPF_JGT | BPF_X: 613 case BPF_JMP | BPF_JSET | BPF_K: 614 case BPF_JMP | BPF_JSET | BPF_X: 615 /* A jump must set masks on targets */ 616 masks[pc + 1 + filter[pc].jt] &= memvalid; 617 masks[pc + 1 + filter[pc].jf] &= memvalid; 618 memvalid = ~0; 619 break; 620 } 621 } 622error: 623 kfree(masks); 624 return ret; 625} 626 627static bool chk_code_allowed(u16 code_to_probe) 628{ 629 static const bool codes[] = { 630 /* 32 bit ALU operations */ 631 [BPF_ALU | BPF_ADD | BPF_K] = true, 632 [BPF_ALU | BPF_ADD | BPF_X] = true, 633 [BPF_ALU | BPF_SUB | BPF_K] = true, 634 [BPF_ALU | BPF_SUB | BPF_X] = true, 635 [BPF_ALU | BPF_MUL | BPF_K] = true, 636 [BPF_ALU | BPF_MUL | BPF_X] = true, 637 [BPF_ALU | BPF_DIV | BPF_K] = true, 638 [BPF_ALU | BPF_DIV | BPF_X] = true, 639 [BPF_ALU | BPF_MOD | BPF_K] = true, 640 [BPF_ALU | BPF_MOD | BPF_X] = true, 641 [BPF_ALU | BPF_AND | BPF_K] = true, 642 [BPF_ALU | BPF_AND | BPF_X] = true, 643 [BPF_ALU | BPF_OR | BPF_K] = true, 644 [BPF_ALU | BPF_OR | BPF_X] = true, 645 [BPF_ALU | BPF_XOR | BPF_K] = true, 646 [BPF_ALU | BPF_XOR | BPF_X] = true, 647 [BPF_ALU | BPF_LSH | BPF_K] = true, 648 [BPF_ALU | BPF_LSH | BPF_X] = true, 649 [BPF_ALU | BPF_RSH | BPF_K] = true, 650 [BPF_ALU | BPF_RSH | BPF_X] = true, 651 [BPF_ALU | BPF_NEG] = true, 652 /* Load instructions */ 653 [BPF_LD | BPF_W | BPF_ABS] = true, 654 [BPF_LD | BPF_H | BPF_ABS] = true, 655 [BPF_LD | BPF_B | BPF_ABS] = true, 656 [BPF_LD | BPF_W | BPF_LEN] = true, 657 [BPF_LD | BPF_W | BPF_IND] = true, 658 [BPF_LD | BPF_H | BPF_IND] = true, 659 [BPF_LD | BPF_B | BPF_IND] = true, 660 [BPF_LD | BPF_IMM] = true, 661 [BPF_LD | BPF_MEM] = true, 662 [BPF_LDX | BPF_W | BPF_LEN] = true, 663 [BPF_LDX | BPF_B | BPF_MSH] = true, 664 [BPF_LDX | BPF_IMM] = true, 665 [BPF_LDX | BPF_MEM] = true, 666 /* Store instructions */ 667 [BPF_ST] = true, 668 [BPF_STX] = true, 669 /* Misc instructions */ 670 [BPF_MISC | BPF_TAX] = true, 671 [BPF_MISC | BPF_TXA] = true, 672 /* Return instructions */ 673 [BPF_RET | BPF_K] = true, 674 [BPF_RET | BPF_A] = true, 675 /* Jump instructions */ 676 [BPF_JMP | BPF_JA] = true, 677 [BPF_JMP | BPF_JEQ | BPF_K] = true, 678 [BPF_JMP | BPF_JEQ | BPF_X] = true, 679 [BPF_JMP | BPF_JGE | BPF_K] = true, 680 [BPF_JMP | BPF_JGE | BPF_X] = true, 681 [BPF_JMP | BPF_JGT | BPF_K] = true, 682 [BPF_JMP | BPF_JGT | BPF_X] = true, 683 [BPF_JMP | BPF_JSET | BPF_K] = true, 684 [BPF_JMP | BPF_JSET | BPF_X] = true, 685 }; 686 687 if (code_to_probe >= ARRAY_SIZE(codes)) 688 return false; 689 690 return codes[code_to_probe]; 691} 692 693/** 694 * bpf_check_classic - verify socket filter code 695 * @filter: filter to verify 696 * @flen: length of filter 697 * 698 * Check the user's filter code. If we let some ugly 699 * filter code slip through kaboom! The filter must contain 700 * no references or jumps that are out of range, no illegal 701 * instructions, and must end with a RET instruction. 702 * 703 * All jumps are forward as they are not signed. 704 * 705 * Returns 0 if the rule set is legal or -EINVAL if not. 706 */ 707int bpf_check_classic(const struct sock_filter *filter, unsigned int flen) 708{ 709 bool anc_found; 710 int pc; 711 712 if (flen == 0 || flen > BPF_MAXINSNS) 713 return -EINVAL; 714 715 /* Check the filter code now */ 716 for (pc = 0; pc < flen; pc++) { 717 const struct sock_filter *ftest = &filter[pc]; 718 719 /* May we actually operate on this code? */ 720 if (!chk_code_allowed(ftest->code)) 721 return -EINVAL; 722 723 /* Some instructions need special checks */ 724 switch (ftest->code) { 725 case BPF_ALU | BPF_DIV | BPF_K: 726 case BPF_ALU | BPF_MOD | BPF_K: 727 /* Check for division by zero */ 728 if (ftest->k == 0) 729 return -EINVAL; 730 break; 731 case BPF_LD | BPF_MEM: 732 case BPF_LDX | BPF_MEM: 733 case BPF_ST: 734 case BPF_STX: 735 /* Check for invalid memory addresses */ 736 if (ftest->k >= BPF_MEMWORDS) 737 return -EINVAL; 738 break; 739 case BPF_JMP | BPF_JA: 740 /* Note, the large ftest->k might cause loops. 741 * Compare this with conditional jumps below, 742 * where offsets are limited. --ANK (981016) 743 */ 744 if (ftest->k >= (unsigned int)(flen - pc - 1)) 745 return -EINVAL; 746 break; 747 case BPF_JMP | BPF_JEQ | BPF_K: 748 case BPF_JMP | BPF_JEQ | BPF_X: 749 case BPF_JMP | BPF_JGE | BPF_K: 750 case BPF_JMP | BPF_JGE | BPF_X: 751 case BPF_JMP | BPF_JGT | BPF_K: 752 case BPF_JMP | BPF_JGT | BPF_X: 753 case BPF_JMP | BPF_JSET | BPF_K: 754 case BPF_JMP | BPF_JSET | BPF_X: 755 /* Both conditionals must be safe */ 756 if (pc + ftest->jt + 1 >= flen || 757 pc + ftest->jf + 1 >= flen) 758 return -EINVAL; 759 break; 760 case BPF_LD | BPF_W | BPF_ABS: 761 case BPF_LD | BPF_H | BPF_ABS: 762 case BPF_LD | BPF_B | BPF_ABS: 763 anc_found = false; 764 if (bpf_anc_helper(ftest) & BPF_ANC) 765 anc_found = true; 766 /* Ancillary operation unknown or unsupported */ 767 if (anc_found == false && ftest->k >= SKF_AD_OFF) 768 return -EINVAL; 769 } 770 } 771 772 /* Last instruction must be a RET code */ 773 switch (filter[flen - 1].code) { 774 case BPF_RET | BPF_K: 775 case BPF_RET | BPF_A: 776 return check_load_and_stores(filter, flen); 777 } 778 779 return -EINVAL; 780} 781EXPORT_SYMBOL(bpf_check_classic); 782 783static int bpf_prog_store_orig_filter(struct bpf_prog *fp, 784 const struct sock_fprog *fprog) 785{ 786 unsigned int fsize = bpf_classic_proglen(fprog); 787 struct sock_fprog_kern *fkprog; 788 789 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL); 790 if (!fp->orig_prog) 791 return -ENOMEM; 792 793 fkprog = fp->orig_prog; 794 fkprog->len = fprog->len; 795 fkprog->filter = kmemdup(fp->insns, fsize, GFP_KERNEL); 796 if (!fkprog->filter) { 797 kfree(fp->orig_prog); 798 return -ENOMEM; 799 } 800 801 return 0; 802} 803 804static void bpf_release_orig_filter(struct bpf_prog *fp) 805{ 806 struct sock_fprog_kern *fprog = fp->orig_prog; 807 808 if (fprog) { 809 kfree(fprog->filter); 810 kfree(fprog); 811 } 812} 813 814static void __bpf_prog_release(struct bpf_prog *prog) 815{ 816 bpf_release_orig_filter(prog); 817 bpf_prog_free(prog); 818} 819 820static void __sk_filter_release(struct sk_filter *fp) 821{ 822 __bpf_prog_release(fp->prog); 823 kfree(fp); 824} 825 826/** 827 * sk_filter_release_rcu - Release a socket filter by rcu_head 828 * @rcu: rcu_head that contains the sk_filter to free 829 */ 830static void sk_filter_release_rcu(struct rcu_head *rcu) 831{ 832 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu); 833 834 __sk_filter_release(fp); 835} 836 837/** 838 * sk_filter_release - release a socket filter 839 * @fp: filter to remove 840 * 841 * Remove a filter from a socket and release its resources. 842 */ 843static void sk_filter_release(struct sk_filter *fp) 844{ 845 if (atomic_dec_and_test(&fp->refcnt)) 846 call_rcu(&fp->rcu, sk_filter_release_rcu); 847} 848 849void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp) 850{ 851 u32 filter_size = bpf_prog_size(fp->prog->len); 852 853 atomic_sub(filter_size, &sk->sk_omem_alloc); 854 sk_filter_release(fp); 855} 856 857/* try to charge the socket memory if there is space available 858 * return true on success 859 */ 860bool sk_filter_charge(struct sock *sk, struct sk_filter *fp) 861{ 862 u32 filter_size = bpf_prog_size(fp->prog->len); 863 864 /* same check as in sock_kmalloc() */ 865 if (filter_size <= sysctl_optmem_max && 866 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) { 867 atomic_inc(&fp->refcnt); 868 atomic_add(filter_size, &sk->sk_omem_alloc); 869 return true; 870 } 871 return false; 872} 873 874static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp) 875{ 876 struct sock_filter *old_prog; 877 struct bpf_prog *old_fp; 878 int err, new_len, old_len = fp->len; 879 880 /* We are free to overwrite insns et al right here as it 881 * won't be used at this point in time anymore internally 882 * after the migration to the internal BPF instruction 883 * representation. 884 */ 885 BUILD_BUG_ON(sizeof(struct sock_filter) != 886 sizeof(struct bpf_insn)); 887 888 /* Conversion cannot happen on overlapping memory areas, 889 * so we need to keep the user BPF around until the 2nd 890 * pass. At this time, the user BPF is stored in fp->insns. 891 */ 892 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter), 893 GFP_KERNEL); 894 if (!old_prog) { 895 err = -ENOMEM; 896 goto out_err; 897 } 898 899 /* 1st pass: calculate the new program length. */ 900 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len); 901 if (err) 902 goto out_err_free; 903 904 /* Expand fp for appending the new filter representation. */ 905 old_fp = fp; 906 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0); 907 if (!fp) { 908 /* The old_fp is still around in case we couldn't 909 * allocate new memory, so uncharge on that one. 910 */ 911 fp = old_fp; 912 err = -ENOMEM; 913 goto out_err_free; 914 } 915 916 fp->len = new_len; 917 918 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */ 919 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len); 920 if (err) 921 /* 2nd bpf_convert_filter() can fail only if it fails 922 * to allocate memory, remapping must succeed. Note, 923 * that at this time old_fp has already been released 924 * by krealloc(). 925 */ 926 goto out_err_free; 927 928 bpf_prog_select_runtime(fp); 929 930 kfree(old_prog); 931 return fp; 932 933out_err_free: 934 kfree(old_prog); 935out_err: 936 __bpf_prog_release(fp); 937 return ERR_PTR(err); 938} 939 940static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp) 941{ 942 int err; 943 944 fp->bpf_func = NULL; 945 fp->jited = false; 946 947 err = bpf_check_classic(fp->insns, fp->len); 948 if (err) { 949 __bpf_prog_release(fp); 950 return ERR_PTR(err); 951 } 952 953 /* Probe if we can JIT compile the filter and if so, do 954 * the compilation of the filter. 955 */ 956 bpf_jit_compile(fp); 957 958 /* JIT compiler couldn't process this filter, so do the 959 * internal BPF translation for the optimized interpreter. 960 */ 961 if (!fp->jited) 962 fp = bpf_migrate_filter(fp); 963 964 return fp; 965} 966 967/** 968 * bpf_prog_create - create an unattached filter 969 * @pfp: the unattached filter that is created 970 * @fprog: the filter program 971 * 972 * Create a filter independent of any socket. We first run some 973 * sanity checks on it to make sure it does not explode on us later. 974 * If an error occurs or there is insufficient memory for the filter 975 * a negative errno code is returned. On success the return is zero. 976 */ 977int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog) 978{ 979 unsigned int fsize = bpf_classic_proglen(fprog); 980 struct bpf_prog *fp; 981 982 /* Make sure new filter is there and in the right amounts. */ 983 if (fprog->filter == NULL) 984 return -EINVAL; 985 986 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0); 987 if (!fp) 988 return -ENOMEM; 989 990 memcpy(fp->insns, fprog->filter, fsize); 991 992 fp->len = fprog->len; 993 /* Since unattached filters are not copied back to user 994 * space through sk_get_filter(), we do not need to hold 995 * a copy here, and can spare us the work. 996 */ 997 fp->orig_prog = NULL; 998 999 /* bpf_prepare_filter() already takes care of freeing 1000 * memory in case something goes wrong. 1001 */ 1002 fp = bpf_prepare_filter(fp); 1003 if (IS_ERR(fp)) 1004 return PTR_ERR(fp); 1005 1006 *pfp = fp; 1007 return 0; 1008} 1009EXPORT_SYMBOL_GPL(bpf_prog_create); 1010 1011void bpf_prog_destroy(struct bpf_prog *fp) 1012{ 1013 __bpf_prog_release(fp); 1014} 1015EXPORT_SYMBOL_GPL(bpf_prog_destroy); 1016 1017/** 1018 * sk_attach_filter - attach a socket filter 1019 * @fprog: the filter program 1020 * @sk: the socket to use 1021 * 1022 * Attach the user's filter code. We first run some sanity checks on 1023 * it to make sure it does not explode on us later. If an error 1024 * occurs or there is insufficient memory for the filter a negative 1025 * errno code is returned. On success the return is zero. 1026 */ 1027int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk) 1028{ 1029 struct sk_filter *fp, *old_fp; 1030 unsigned int fsize = bpf_classic_proglen(fprog); 1031 unsigned int bpf_fsize = bpf_prog_size(fprog->len); 1032 struct bpf_prog *prog; 1033 int err; 1034 1035 if (sock_flag(sk, SOCK_FILTER_LOCKED)) 1036 return -EPERM; 1037 1038 /* Make sure new filter is there and in the right amounts. */ 1039 if (fprog->filter == NULL) 1040 return -EINVAL; 1041 1042 prog = bpf_prog_alloc(bpf_fsize, 0); 1043 if (!prog) 1044 return -ENOMEM; 1045 1046 if (copy_from_user(prog->insns, fprog->filter, fsize)) { 1047 __bpf_prog_free(prog); 1048 return -EFAULT; 1049 } 1050 1051 prog->len = fprog->len; 1052 1053 err = bpf_prog_store_orig_filter(prog, fprog); 1054 if (err) { 1055 __bpf_prog_free(prog); 1056 return -ENOMEM; 1057 } 1058 1059 /* bpf_prepare_filter() already takes care of freeing 1060 * memory in case something goes wrong. 1061 */ 1062 prog = bpf_prepare_filter(prog); 1063 if (IS_ERR(prog)) 1064 return PTR_ERR(prog); 1065 1066 fp = kmalloc(sizeof(*fp), GFP_KERNEL); 1067 if (!fp) { 1068 __bpf_prog_release(prog); 1069 return -ENOMEM; 1070 } 1071 fp->prog = prog; 1072 1073 atomic_set(&fp->refcnt, 0); 1074 1075 if (!sk_filter_charge(sk, fp)) { 1076 __sk_filter_release(fp); 1077 return -ENOMEM; 1078 } 1079 1080 old_fp = rcu_dereference_protected(sk->sk_filter, 1081 sock_owned_by_user(sk)); 1082 rcu_assign_pointer(sk->sk_filter, fp); 1083 1084 if (old_fp) 1085 sk_filter_uncharge(sk, old_fp); 1086 1087 return 0; 1088} 1089EXPORT_SYMBOL_GPL(sk_attach_filter); 1090 1091int sk_detach_filter(struct sock *sk) 1092{ 1093 int ret = -ENOENT; 1094 struct sk_filter *filter; 1095 1096 if (sock_flag(sk, SOCK_FILTER_LOCKED)) 1097 return -EPERM; 1098 1099 filter = rcu_dereference_protected(sk->sk_filter, 1100 sock_owned_by_user(sk)); 1101 if (filter) { 1102 RCU_INIT_POINTER(sk->sk_filter, NULL); 1103 sk_filter_uncharge(sk, filter); 1104 ret = 0; 1105 } 1106 1107 return ret; 1108} 1109EXPORT_SYMBOL_GPL(sk_detach_filter); 1110 1111int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf, 1112 unsigned int len) 1113{ 1114 struct sock_fprog_kern *fprog; 1115 struct sk_filter *filter; 1116 int ret = 0; 1117 1118 lock_sock(sk); 1119 filter = rcu_dereference_protected(sk->sk_filter, 1120 sock_owned_by_user(sk)); 1121 if (!filter) 1122 goto out; 1123 1124 /* We're copying the filter that has been originally attached, 1125 * so no conversion/decode needed anymore. 1126 */ 1127 fprog = filter->prog->orig_prog; 1128 1129 ret = fprog->len; 1130 if (!len) 1131 /* User space only enquires number of filter blocks. */ 1132 goto out; 1133 1134 ret = -EINVAL; 1135 if (len < fprog->len) 1136 goto out; 1137 1138 ret = -EFAULT; 1139 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog))) 1140 goto out; 1141 1142 /* Instead of bytes, the API requests to return the number 1143 * of filter blocks. 1144 */ 1145 ret = fprog->len; 1146out: 1147 release_sock(sk); 1148 return ret; 1149} 1150