1 /* bpf_jit_comp.c : BPF JIT compiler 2 * 3 * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com) 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public License 7 * as published by the Free Software Foundation; version 2 8 * of the License. 9 */ 10 #include <linux/moduleloader.h> 11 #include <asm/cacheflush.h> 12 #include <linux/netdevice.h> 13 #include <linux/filter.h> 14 #include <linux/if_vlan.h> 15 #include <linux/random.h> 16 17 /* 18 * Conventions : 19 * EAX : BPF A accumulator 20 * EBX : BPF X accumulator 21 * RDI : pointer to skb (first argument given to JIT function) 22 * RBP : frame pointer (even if CONFIG_FRAME_POINTER=n) 23 * ECX,EDX,ESI : scratch registers 24 * r9d : skb->len - skb->data_len (headlen) 25 * r8 : skb->data 26 * -8(RBP) : saved RBX value 27 * -16(RBP)..-80(RBP) : BPF_MEMWORDS values 28 */ 29 int bpf_jit_enable __read_mostly; 30 31 /* 32 * assembly code in arch/x86/net/bpf_jit.S 33 */ 34 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[]; 35 extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[]; 36 extern u8 sk_load_byte_positive_offset[], sk_load_byte_msh_positive_offset[]; 37 extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[]; 38 extern u8 sk_load_byte_negative_offset[], sk_load_byte_msh_negative_offset[]; 39 40 static inline u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len) 41 { 42 if (len == 1) 43 *ptr = bytes; 44 else if (len == 2) 45 *(u16 *)ptr = bytes; 46 else { 47 *(u32 *)ptr = bytes; 48 barrier(); 49 } 50 return ptr + len; 51 } 52 53 #define EMIT(bytes, len) do { prog = emit_code(prog, bytes, len); } while (0) 54 55 #define EMIT1(b1) EMIT(b1, 1) 56 #define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2) 57 #define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3) 58 #define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4) 59 #define EMIT1_off32(b1, off) do { EMIT1(b1); EMIT(off, 4);} while (0) 60 61 #define CLEAR_A() EMIT2(0x31, 0xc0) /* xor %eax,%eax */ 62 #define CLEAR_X() EMIT2(0x31, 0xdb) /* xor %ebx,%ebx */ 63 64 static inline bool is_imm8(int value) 65 { 66 return value <= 127 && value >= -128; 67 } 68 69 static inline bool is_near(int offset) 70 { 71 return offset <= 127 && offset >= -128; 72 } 73 74 #define EMIT_JMP(offset) \ 75 do { \ 76 if (offset) { \ 77 if (is_near(offset)) \ 78 EMIT2(0xeb, offset); /* jmp .+off8 */ \ 79 else \ 80 EMIT1_off32(0xe9, offset); /* jmp .+off32 */ \ 81 } \ 82 } while (0) 83 84 /* list of x86 cond jumps opcodes (. + s8) 85 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32) 86 */ 87 #define X86_JB 0x72 88 #define X86_JAE 0x73 89 #define X86_JE 0x74 90 #define X86_JNE 0x75 91 #define X86_JBE 0x76 92 #define X86_JA 0x77 93 94 #define EMIT_COND_JMP(op, offset) \ 95 do { \ 96 if (is_near(offset)) \ 97 EMIT2(op, offset); /* jxx .+off8 */ \ 98 else { \ 99 EMIT2(0x0f, op + 0x10); \ 100 EMIT(offset, 4); /* jxx .+off32 */ \ 101 } \ 102 } while (0) 103 104 #define COND_SEL(CODE, TOP, FOP) \ 105 case CODE: \ 106 t_op = TOP; \ 107 f_op = FOP; \ 108 goto cond_branch 109 110 111 #define SEEN_DATAREF 1 /* might call external helpers */ 112 #define SEEN_XREG 2 /* ebx is used */ 113 #define SEEN_MEM 4 /* use mem[] for temporary storage */ 114 115 static inline void bpf_flush_icache(void *start, void *end) 116 { 117 mm_segment_t old_fs = get_fs(); 118 119 set_fs(KERNEL_DS); 120 smp_wmb(); 121 flush_icache_range((unsigned long)start, (unsigned long)end); 122 set_fs(old_fs); 123 } 124 125 #define CHOOSE_LOAD_FUNC(K, func) \ 126 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset) 127 128 /* Helper to find the offset of pkt_type in sk_buff 129 * We want to make sure its still a 3bit field starting at a byte boundary. 130 */ 131 #define PKT_TYPE_MAX 7 132 static int pkt_type_offset(void) 133 { 134 struct sk_buff skb_probe = { 135 .pkt_type = ~0, 136 }; 137 char *ct = (char *)&skb_probe; 138 unsigned int off; 139 140 for (off = 0; off < sizeof(struct sk_buff); off++) { 141 if (ct[off] == PKT_TYPE_MAX) 142 return off; 143 } 144 pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be found\n"); 145 return -1; 146 } 147 148 struct bpf_binary_header { 149 unsigned int pages; 150 /* Note : for security reasons, bpf code will follow a randomly 151 * sized amount of int3 instructions 152 */ 153 u8 image[]; 154 }; 155 156 static struct bpf_binary_header *bpf_alloc_binary(unsigned int proglen, 157 u8 **image_ptr) 158 { 159 unsigned int sz, hole; 160 struct bpf_binary_header *header; 161 162 /* Most of BPF filters are really small, 163 * but if some of them fill a page, allow at least 164 * 128 extra bytes to insert a random section of int3 165 */ 166 sz = round_up(proglen + sizeof(*header) + 128, PAGE_SIZE); 167 header = module_alloc(sz); 168 if (!header) 169 return NULL; 170 171 memset(header, 0xcc, sz); /* fill whole space with int3 instructions */ 172 173 header->pages = sz / PAGE_SIZE; 174 hole = sz - (proglen + sizeof(*header)); 175 176 /* insert a random number of int3 instructions before BPF code */ 177 *image_ptr = &header->image[prandom_u32() % hole]; 178 return header; 179 } 180 181 void bpf_jit_compile(struct sk_filter *fp) 182 { 183 u8 temp[64]; 184 u8 *prog; 185 unsigned int proglen, oldproglen = 0; 186 int ilen, i; 187 int t_offset, f_offset; 188 u8 t_op, f_op, seen = 0, pass; 189 u8 *image = NULL; 190 struct bpf_binary_header *header = NULL; 191 u8 *func; 192 int pc_ret0 = -1; /* bpf index of first RET #0 instruction (if any) */ 193 unsigned int cleanup_addr; /* epilogue code offset */ 194 unsigned int *addrs; 195 const struct sock_filter *filter = fp->insns; 196 int flen = fp->len; 197 198 if (!bpf_jit_enable) 199 return; 200 201 addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL); 202 if (addrs == NULL) 203 return; 204 205 /* Before first pass, make a rough estimation of addrs[] 206 * each bpf instruction is translated to less than 64 bytes 207 */ 208 for (proglen = 0, i = 0; i < flen; i++) { 209 proglen += 64; 210 addrs[i] = proglen; 211 } 212 cleanup_addr = proglen; /* epilogue address */ 213 214 for (pass = 0; pass < 10; pass++) { 215 u8 seen_or_pass0 = (pass == 0) ? (SEEN_XREG | SEEN_DATAREF | SEEN_MEM) : seen; 216 /* no prologue/epilogue for trivial filters (RET something) */ 217 proglen = 0; 218 prog = temp; 219 220 if (seen_or_pass0) { 221 EMIT4(0x55, 0x48, 0x89, 0xe5); /* push %rbp; mov %rsp,%rbp */ 222 EMIT4(0x48, 0x83, 0xec, 96); /* subq $96,%rsp */ 223 /* note : must save %rbx in case bpf_error is hit */ 224 if (seen_or_pass0 & (SEEN_XREG | SEEN_DATAREF)) 225 EMIT4(0x48, 0x89, 0x5d, 0xf8); /* mov %rbx, -8(%rbp) */ 226 if (seen_or_pass0 & SEEN_XREG) 227 CLEAR_X(); /* make sure we dont leek kernel memory */ 228 229 /* 230 * If this filter needs to access skb data, 231 * loads r9 and r8 with : 232 * r9 = skb->len - skb->data_len 233 * r8 = skb->data 234 */ 235 if (seen_or_pass0 & SEEN_DATAREF) { 236 if (offsetof(struct sk_buff, len) <= 127) 237 /* mov off8(%rdi),%r9d */ 238 EMIT4(0x44, 0x8b, 0x4f, offsetof(struct sk_buff, len)); 239 else { 240 /* mov off32(%rdi),%r9d */ 241 EMIT3(0x44, 0x8b, 0x8f); 242 EMIT(offsetof(struct sk_buff, len), 4); 243 } 244 if (is_imm8(offsetof(struct sk_buff, data_len))) 245 /* sub off8(%rdi),%r9d */ 246 EMIT4(0x44, 0x2b, 0x4f, offsetof(struct sk_buff, data_len)); 247 else { 248 EMIT3(0x44, 0x2b, 0x8f); 249 EMIT(offsetof(struct sk_buff, data_len), 4); 250 } 251 252 if (is_imm8(offsetof(struct sk_buff, data))) 253 /* mov off8(%rdi),%r8 */ 254 EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data)); 255 else { 256 /* mov off32(%rdi),%r8 */ 257 EMIT3(0x4c, 0x8b, 0x87); 258 EMIT(offsetof(struct sk_buff, data), 4); 259 } 260 } 261 } 262 263 switch (filter[0].code) { 264 case BPF_S_RET_K: 265 case BPF_S_LD_W_LEN: 266 case BPF_S_ANC_PROTOCOL: 267 case BPF_S_ANC_IFINDEX: 268 case BPF_S_ANC_MARK: 269 case BPF_S_ANC_RXHASH: 270 case BPF_S_ANC_CPU: 271 case BPF_S_ANC_VLAN_TAG: 272 case BPF_S_ANC_VLAN_TAG_PRESENT: 273 case BPF_S_ANC_QUEUE: 274 case BPF_S_ANC_PKTTYPE: 275 case BPF_S_LD_W_ABS: 276 case BPF_S_LD_H_ABS: 277 case BPF_S_LD_B_ABS: 278 /* first instruction sets A register (or is RET 'constant') */ 279 break; 280 default: 281 /* make sure we dont leak kernel information to user */ 282 CLEAR_A(); /* A = 0 */ 283 } 284 285 for (i = 0; i < flen; i++) { 286 unsigned int K = filter[i].k; 287 288 switch (filter[i].code) { 289 case BPF_S_ALU_ADD_X: /* A += X; */ 290 seen |= SEEN_XREG; 291 EMIT2(0x01, 0xd8); /* add %ebx,%eax */ 292 break; 293 case BPF_S_ALU_ADD_K: /* A += K; */ 294 if (!K) 295 break; 296 if (is_imm8(K)) 297 EMIT3(0x83, 0xc0, K); /* add imm8,%eax */ 298 else 299 EMIT1_off32(0x05, K); /* add imm32,%eax */ 300 break; 301 case BPF_S_ALU_SUB_X: /* A -= X; */ 302 seen |= SEEN_XREG; 303 EMIT2(0x29, 0xd8); /* sub %ebx,%eax */ 304 break; 305 case BPF_S_ALU_SUB_K: /* A -= K */ 306 if (!K) 307 break; 308 if (is_imm8(K)) 309 EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */ 310 else 311 EMIT1_off32(0x2d, K); /* sub imm32,%eax */ 312 break; 313 case BPF_S_ALU_MUL_X: /* A *= X; */ 314 seen |= SEEN_XREG; 315 EMIT3(0x0f, 0xaf, 0xc3); /* imul %ebx,%eax */ 316 break; 317 case BPF_S_ALU_MUL_K: /* A *= K */ 318 if (is_imm8(K)) 319 EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */ 320 else { 321 EMIT2(0x69, 0xc0); /* imul imm32,%eax */ 322 EMIT(K, 4); 323 } 324 break; 325 case BPF_S_ALU_DIV_X: /* A /= X; */ 326 seen |= SEEN_XREG; 327 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */ 328 if (pc_ret0 > 0) { 329 /* addrs[pc_ret0 - 1] is start address of target 330 * (addrs[i] - 4) is the address following this jmp 331 * ("xor %edx,%edx; div %ebx" being 4 bytes long) 332 */ 333 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] - 334 (addrs[i] - 4)); 335 } else { 336 EMIT_COND_JMP(X86_JNE, 2 + 5); 337 CLEAR_A(); 338 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */ 339 } 340 EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */ 341 break; 342 case BPF_S_ALU_MOD_X: /* A %= X; */ 343 seen |= SEEN_XREG; 344 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */ 345 if (pc_ret0 > 0) { 346 /* addrs[pc_ret0 - 1] is start address of target 347 * (addrs[i] - 6) is the address following this jmp 348 * ("xor %edx,%edx; div %ebx;mov %edx,%eax" being 6 bytes long) 349 */ 350 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] - 351 (addrs[i] - 6)); 352 } else { 353 EMIT_COND_JMP(X86_JNE, 2 + 5); 354 CLEAR_A(); 355 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 6)); /* jmp .+off32 */ 356 } 357 EMIT2(0x31, 0xd2); /* xor %edx,%edx */ 358 EMIT2(0xf7, 0xf3); /* div %ebx */ 359 EMIT2(0x89, 0xd0); /* mov %edx,%eax */ 360 break; 361 case BPF_S_ALU_MOD_K: /* A %= K; */ 362 EMIT2(0x31, 0xd2); /* xor %edx,%edx */ 363 EMIT1(0xb9);EMIT(K, 4); /* mov imm32,%ecx */ 364 EMIT2(0xf7, 0xf1); /* div %ecx */ 365 EMIT2(0x89, 0xd0); /* mov %edx,%eax */ 366 break; 367 case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */ 368 EMIT3(0x48, 0x69, 0xc0); /* imul imm32,%rax,%rax */ 369 EMIT(K, 4); 370 EMIT4(0x48, 0xc1, 0xe8, 0x20); /* shr $0x20,%rax */ 371 break; 372 case BPF_S_ALU_AND_X: 373 seen |= SEEN_XREG; 374 EMIT2(0x21, 0xd8); /* and %ebx,%eax */ 375 break; 376 case BPF_S_ALU_AND_K: 377 if (K >= 0xFFFFFF00) { 378 EMIT2(0x24, K & 0xFF); /* and imm8,%al */ 379 } else if (K >= 0xFFFF0000) { 380 EMIT2(0x66, 0x25); /* and imm16,%ax */ 381 EMIT(K, 2); 382 } else { 383 EMIT1_off32(0x25, K); /* and imm32,%eax */ 384 } 385 break; 386 case BPF_S_ALU_OR_X: 387 seen |= SEEN_XREG; 388 EMIT2(0x09, 0xd8); /* or %ebx,%eax */ 389 break; 390 case BPF_S_ALU_OR_K: 391 if (is_imm8(K)) 392 EMIT3(0x83, 0xc8, K); /* or imm8,%eax */ 393 else 394 EMIT1_off32(0x0d, K); /* or imm32,%eax */ 395 break; 396 case BPF_S_ANC_ALU_XOR_X: /* A ^= X; */ 397 case BPF_S_ALU_XOR_X: 398 seen |= SEEN_XREG; 399 EMIT2(0x31, 0xd8); /* xor %ebx,%eax */ 400 break; 401 case BPF_S_ALU_XOR_K: /* A ^= K; */ 402 if (K == 0) 403 break; 404 if (is_imm8(K)) 405 EMIT3(0x83, 0xf0, K); /* xor imm8,%eax */ 406 else 407 EMIT1_off32(0x35, K); /* xor imm32,%eax */ 408 break; 409 case BPF_S_ALU_LSH_X: /* A <<= X; */ 410 seen |= SEEN_XREG; 411 EMIT4(0x89, 0xd9, 0xd3, 0xe0); /* mov %ebx,%ecx; shl %cl,%eax */ 412 break; 413 case BPF_S_ALU_LSH_K: 414 if (K == 0) 415 break; 416 else if (K == 1) 417 EMIT2(0xd1, 0xe0); /* shl %eax */ 418 else 419 EMIT3(0xc1, 0xe0, K); 420 break; 421 case BPF_S_ALU_RSH_X: /* A >>= X; */ 422 seen |= SEEN_XREG; 423 EMIT4(0x89, 0xd9, 0xd3, 0xe8); /* mov %ebx,%ecx; shr %cl,%eax */ 424 break; 425 case BPF_S_ALU_RSH_K: /* A >>= K; */ 426 if (K == 0) 427 break; 428 else if (K == 1) 429 EMIT2(0xd1, 0xe8); /* shr %eax */ 430 else 431 EMIT3(0xc1, 0xe8, K); 432 break; 433 case BPF_S_ALU_NEG: 434 EMIT2(0xf7, 0xd8); /* neg %eax */ 435 break; 436 case BPF_S_RET_K: 437 if (!K) { 438 if (pc_ret0 == -1) 439 pc_ret0 = i; 440 CLEAR_A(); 441 } else { 442 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */ 443 } 444 /* fallinto */ 445 case BPF_S_RET_A: 446 if (seen_or_pass0) { 447 if (i != flen - 1) { 448 EMIT_JMP(cleanup_addr - addrs[i]); 449 break; 450 } 451 if (seen_or_pass0 & SEEN_XREG) 452 EMIT4(0x48, 0x8b, 0x5d, 0xf8); /* mov -8(%rbp),%rbx */ 453 EMIT1(0xc9); /* leaveq */ 454 } 455 EMIT1(0xc3); /* ret */ 456 break; 457 case BPF_S_MISC_TAX: /* X = A */ 458 seen |= SEEN_XREG; 459 EMIT2(0x89, 0xc3); /* mov %eax,%ebx */ 460 break; 461 case BPF_S_MISC_TXA: /* A = X */ 462 seen |= SEEN_XREG; 463 EMIT2(0x89, 0xd8); /* mov %ebx,%eax */ 464 break; 465 case BPF_S_LD_IMM: /* A = K */ 466 if (!K) 467 CLEAR_A(); 468 else 469 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */ 470 break; 471 case BPF_S_LDX_IMM: /* X = K */ 472 seen |= SEEN_XREG; 473 if (!K) 474 CLEAR_X(); 475 else 476 EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */ 477 break; 478 case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */ 479 seen |= SEEN_MEM; 480 EMIT3(0x8b, 0x45, 0xf0 - K*4); 481 break; 482 case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */ 483 seen |= SEEN_XREG | SEEN_MEM; 484 EMIT3(0x8b, 0x5d, 0xf0 - K*4); 485 break; 486 case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */ 487 seen |= SEEN_MEM; 488 EMIT3(0x89, 0x45, 0xf0 - K*4); 489 break; 490 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */ 491 seen |= SEEN_XREG | SEEN_MEM; 492 EMIT3(0x89, 0x5d, 0xf0 - K*4); 493 break; 494 case BPF_S_LD_W_LEN: /* A = skb->len; */ 495 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); 496 if (is_imm8(offsetof(struct sk_buff, len))) 497 /* mov off8(%rdi),%eax */ 498 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len)); 499 else { 500 EMIT2(0x8b, 0x87); 501 EMIT(offsetof(struct sk_buff, len), 4); 502 } 503 break; 504 case BPF_S_LDX_W_LEN: /* X = skb->len; */ 505 seen |= SEEN_XREG; 506 if (is_imm8(offsetof(struct sk_buff, len))) 507 /* mov off8(%rdi),%ebx */ 508 EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len)); 509 else { 510 EMIT2(0x8b, 0x9f); 511 EMIT(offsetof(struct sk_buff, len), 4); 512 } 513 break; 514 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */ 515 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2); 516 if (is_imm8(offsetof(struct sk_buff, protocol))) { 517 /* movzwl off8(%rdi),%eax */ 518 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol)); 519 } else { 520 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 521 EMIT(offsetof(struct sk_buff, protocol), 4); 522 } 523 EMIT2(0x86, 0xc4); /* ntohs() : xchg %al,%ah */ 524 break; 525 case BPF_S_ANC_IFINDEX: 526 if (is_imm8(offsetof(struct sk_buff, dev))) { 527 /* movq off8(%rdi),%rax */ 528 EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev)); 529 } else { 530 EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */ 531 EMIT(offsetof(struct sk_buff, dev), 4); 532 } 533 EMIT3(0x48, 0x85, 0xc0); /* test %rax,%rax */ 534 EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6)); 535 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4); 536 EMIT2(0x8b, 0x80); /* mov off32(%rax),%eax */ 537 EMIT(offsetof(struct net_device, ifindex), 4); 538 break; 539 case BPF_S_ANC_MARK: 540 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 541 if (is_imm8(offsetof(struct sk_buff, mark))) { 542 /* mov off8(%rdi),%eax */ 543 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark)); 544 } else { 545 EMIT2(0x8b, 0x87); 546 EMIT(offsetof(struct sk_buff, mark), 4); 547 } 548 break; 549 case BPF_S_ANC_RXHASH: 550 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4); 551 if (is_imm8(offsetof(struct sk_buff, rxhash))) { 552 /* mov off8(%rdi),%eax */ 553 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, rxhash)); 554 } else { 555 EMIT2(0x8b, 0x87); 556 EMIT(offsetof(struct sk_buff, rxhash), 4); 557 } 558 break; 559 case BPF_S_ANC_QUEUE: 560 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2); 561 if (is_imm8(offsetof(struct sk_buff, queue_mapping))) { 562 /* movzwl off8(%rdi),%eax */ 563 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping)); 564 } else { 565 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 566 EMIT(offsetof(struct sk_buff, queue_mapping), 4); 567 } 568 break; 569 case BPF_S_ANC_CPU: 570 #ifdef CONFIG_SMP 571 EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */ 572 EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */ 573 #else 574 CLEAR_A(); 575 #endif 576 break; 577 case BPF_S_ANC_VLAN_TAG: 578 case BPF_S_ANC_VLAN_TAG_PRESENT: 579 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 580 if (is_imm8(offsetof(struct sk_buff, vlan_tci))) { 581 /* movzwl off8(%rdi),%eax */ 582 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, vlan_tci)); 583 } else { 584 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 585 EMIT(offsetof(struct sk_buff, vlan_tci), 4); 586 } 587 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000); 588 if (filter[i].code == BPF_S_ANC_VLAN_TAG) { 589 EMIT3(0x80, 0xe4, 0xef); /* and $0xef,%ah */ 590 } else { 591 EMIT3(0xc1, 0xe8, 0x0c); /* shr $0xc,%eax */ 592 EMIT3(0x83, 0xe0, 0x01); /* and $0x1,%eax */ 593 } 594 break; 595 case BPF_S_ANC_PKTTYPE: 596 { 597 int off = pkt_type_offset(); 598 599 if (off < 0) 600 goto out; 601 if (is_imm8(off)) { 602 /* movzbl off8(%rdi),%eax */ 603 EMIT4(0x0f, 0xb6, 0x47, off); 604 } else { 605 /* movbl off32(%rdi),%eax */ 606 EMIT3(0x0f, 0xb6, 0x87); 607 EMIT(off, 4); 608 } 609 EMIT3(0x83, 0xe0, PKT_TYPE_MAX); /* and $0x7,%eax */ 610 break; 611 } 612 case BPF_S_LD_W_ABS: 613 func = CHOOSE_LOAD_FUNC(K, sk_load_word); 614 common_load: seen |= SEEN_DATAREF; 615 t_offset = func - (image + addrs[i]); 616 EMIT1_off32(0xbe, K); /* mov imm32,%esi */ 617 EMIT1_off32(0xe8, t_offset); /* call */ 618 break; 619 case BPF_S_LD_H_ABS: 620 func = CHOOSE_LOAD_FUNC(K, sk_load_half); 621 goto common_load; 622 case BPF_S_LD_B_ABS: 623 func = CHOOSE_LOAD_FUNC(K, sk_load_byte); 624 goto common_load; 625 case BPF_S_LDX_B_MSH: 626 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh); 627 seen |= SEEN_DATAREF | SEEN_XREG; 628 t_offset = func - (image + addrs[i]); 629 EMIT1_off32(0xbe, K); /* mov imm32,%esi */ 630 EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */ 631 break; 632 case BPF_S_LD_W_IND: 633 func = sk_load_word; 634 common_load_ind: seen |= SEEN_DATAREF | SEEN_XREG; 635 t_offset = func - (image + addrs[i]); 636 if (K) { 637 if (is_imm8(K)) { 638 EMIT3(0x8d, 0x73, K); /* lea imm8(%rbx), %esi */ 639 } else { 640 EMIT2(0x8d, 0xb3); /* lea imm32(%rbx),%esi */ 641 EMIT(K, 4); 642 } 643 } else { 644 EMIT2(0x89,0xde); /* mov %ebx,%esi */ 645 } 646 EMIT1_off32(0xe8, t_offset); /* call sk_load_xxx_ind */ 647 break; 648 case BPF_S_LD_H_IND: 649 func = sk_load_half; 650 goto common_load_ind; 651 case BPF_S_LD_B_IND: 652 func = sk_load_byte; 653 goto common_load_ind; 654 case BPF_S_JMP_JA: 655 t_offset = addrs[i + K] - addrs[i]; 656 EMIT_JMP(t_offset); 657 break; 658 COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE); 659 COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB); 660 COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE); 661 COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE); 662 COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE); 663 COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB); 664 COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE); 665 COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE); 666 667 cond_branch: f_offset = addrs[i + filter[i].jf] - addrs[i]; 668 t_offset = addrs[i + filter[i].jt] - addrs[i]; 669 670 /* same targets, can avoid doing the test :) */ 671 if (filter[i].jt == filter[i].jf) { 672 EMIT_JMP(t_offset); 673 break; 674 } 675 676 switch (filter[i].code) { 677 case BPF_S_JMP_JGT_X: 678 case BPF_S_JMP_JGE_X: 679 case BPF_S_JMP_JEQ_X: 680 seen |= SEEN_XREG; 681 EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */ 682 break; 683 case BPF_S_JMP_JSET_X: 684 seen |= SEEN_XREG; 685 EMIT2(0x85, 0xd8); /* test %ebx,%eax */ 686 break; 687 case BPF_S_JMP_JEQ_K: 688 if (K == 0) { 689 EMIT2(0x85, 0xc0); /* test %eax,%eax */ 690 break; 691 } 692 case BPF_S_JMP_JGT_K: 693 case BPF_S_JMP_JGE_K: 694 if (K <= 127) 695 EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */ 696 else 697 EMIT1_off32(0x3d, K); /* cmp imm32,%eax */ 698 break; 699 case BPF_S_JMP_JSET_K: 700 if (K <= 0xFF) 701 EMIT2(0xa8, K); /* test imm8,%al */ 702 else if (!(K & 0xFFFF00FF)) 703 EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */ 704 else if (K <= 0xFFFF) { 705 EMIT2(0x66, 0xa9); /* test imm16,%ax */ 706 EMIT(K, 2); 707 } else { 708 EMIT1_off32(0xa9, K); /* test imm32,%eax */ 709 } 710 break; 711 } 712 if (filter[i].jt != 0) { 713 if (filter[i].jf && f_offset) 714 t_offset += is_near(f_offset) ? 2 : 5; 715 EMIT_COND_JMP(t_op, t_offset); 716 if (filter[i].jf) 717 EMIT_JMP(f_offset); 718 break; 719 } 720 EMIT_COND_JMP(f_op, f_offset); 721 break; 722 default: 723 /* hmm, too complex filter, give up with jit compiler */ 724 goto out; 725 } 726 ilen = prog - temp; 727 if (image) { 728 if (unlikely(proglen + ilen > oldproglen)) { 729 pr_err("bpb_jit_compile fatal error\n"); 730 kfree(addrs); 731 module_free(NULL, header); 732 return; 733 } 734 memcpy(image + proglen, temp, ilen); 735 } 736 proglen += ilen; 737 addrs[i] = proglen; 738 prog = temp; 739 } 740 /* last bpf instruction is always a RET : 741 * use it to give the cleanup instruction(s) addr 742 */ 743 cleanup_addr = proglen - 1; /* ret */ 744 if (seen_or_pass0) 745 cleanup_addr -= 1; /* leaveq */ 746 if (seen_or_pass0 & SEEN_XREG) 747 cleanup_addr -= 4; /* mov -8(%rbp),%rbx */ 748 749 if (image) { 750 if (proglen != oldproglen) 751 pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n", proglen, oldproglen); 752 break; 753 } 754 if (proglen == oldproglen) { 755 header = bpf_alloc_binary(proglen, &image); 756 if (!header) 757 goto out; 758 } 759 oldproglen = proglen; 760 } 761 762 if (bpf_jit_enable > 1) 763 bpf_jit_dump(flen, proglen, pass, image); 764 765 if (image) { 766 bpf_flush_icache(header, image + proglen); 767 set_memory_ro((unsigned long)header, header->pages); 768 fp->bpf_func = (void *)image; 769 } 770 out: 771 kfree(addrs); 772 return; 773 } 774 775 static void bpf_jit_free_deferred(struct work_struct *work) 776 { 777 struct sk_filter *fp = container_of(work, struct sk_filter, work); 778 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK; 779 struct bpf_binary_header *header = (void *)addr; 780 781 set_memory_rw(addr, header->pages); 782 module_free(NULL, header); 783 kfree(fp); 784 } 785 786 void bpf_jit_free(struct sk_filter *fp) 787 { 788 if (fp->bpf_func != sk_run_filter) { 789 INIT_WORK(&fp->work, bpf_jit_free_deferred); 790 schedule_work(&fp->work); 791 } else { 792 kfree(fp); 793 } 794 } 795