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 if (K == 1) { 363 CLEAR_A(); 364 break; 365 } 366 EMIT2(0x31, 0xd2); /* xor %edx,%edx */ 367 EMIT1(0xb9);EMIT(K, 4); /* mov imm32,%ecx */ 368 EMIT2(0xf7, 0xf1); /* div %ecx */ 369 EMIT2(0x89, 0xd0); /* mov %edx,%eax */ 370 break; 371 case BPF_S_ALU_DIV_K: /* A /= K */ 372 if (K == 1) 373 break; 374 EMIT2(0x31, 0xd2); /* xor %edx,%edx */ 375 EMIT1(0xb9);EMIT(K, 4); /* mov imm32,%ecx */ 376 EMIT2(0xf7, 0xf1); /* div %ecx */ 377 break; 378 case BPF_S_ALU_AND_X: 379 seen |= SEEN_XREG; 380 EMIT2(0x21, 0xd8); /* and %ebx,%eax */ 381 break; 382 case BPF_S_ALU_AND_K: 383 if (K >= 0xFFFFFF00) { 384 EMIT2(0x24, K & 0xFF); /* and imm8,%al */ 385 } else if (K >= 0xFFFF0000) { 386 EMIT2(0x66, 0x25); /* and imm16,%ax */ 387 EMIT(K, 2); 388 } else { 389 EMIT1_off32(0x25, K); /* and imm32,%eax */ 390 } 391 break; 392 case BPF_S_ALU_OR_X: 393 seen |= SEEN_XREG; 394 EMIT2(0x09, 0xd8); /* or %ebx,%eax */ 395 break; 396 case BPF_S_ALU_OR_K: 397 if (is_imm8(K)) 398 EMIT3(0x83, 0xc8, K); /* or imm8,%eax */ 399 else 400 EMIT1_off32(0x0d, K); /* or imm32,%eax */ 401 break; 402 case BPF_S_ANC_ALU_XOR_X: /* A ^= X; */ 403 case BPF_S_ALU_XOR_X: 404 seen |= SEEN_XREG; 405 EMIT2(0x31, 0xd8); /* xor %ebx,%eax */ 406 break; 407 case BPF_S_ALU_XOR_K: /* A ^= K; */ 408 if (K == 0) 409 break; 410 if (is_imm8(K)) 411 EMIT3(0x83, 0xf0, K); /* xor imm8,%eax */ 412 else 413 EMIT1_off32(0x35, K); /* xor imm32,%eax */ 414 break; 415 case BPF_S_ALU_LSH_X: /* A <<= X; */ 416 seen |= SEEN_XREG; 417 EMIT4(0x89, 0xd9, 0xd3, 0xe0); /* mov %ebx,%ecx; shl %cl,%eax */ 418 break; 419 case BPF_S_ALU_LSH_K: 420 if (K == 0) 421 break; 422 else if (K == 1) 423 EMIT2(0xd1, 0xe0); /* shl %eax */ 424 else 425 EMIT3(0xc1, 0xe0, K); 426 break; 427 case BPF_S_ALU_RSH_X: /* A >>= X; */ 428 seen |= SEEN_XREG; 429 EMIT4(0x89, 0xd9, 0xd3, 0xe8); /* mov %ebx,%ecx; shr %cl,%eax */ 430 break; 431 case BPF_S_ALU_RSH_K: /* A >>= K; */ 432 if (K == 0) 433 break; 434 else if (K == 1) 435 EMIT2(0xd1, 0xe8); /* shr %eax */ 436 else 437 EMIT3(0xc1, 0xe8, K); 438 break; 439 case BPF_S_ALU_NEG: 440 EMIT2(0xf7, 0xd8); /* neg %eax */ 441 break; 442 case BPF_S_RET_K: 443 if (!K) { 444 if (pc_ret0 == -1) 445 pc_ret0 = i; 446 CLEAR_A(); 447 } else { 448 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */ 449 } 450 /* fallinto */ 451 case BPF_S_RET_A: 452 if (seen_or_pass0) { 453 if (i != flen - 1) { 454 EMIT_JMP(cleanup_addr - addrs[i]); 455 break; 456 } 457 if (seen_or_pass0 & SEEN_XREG) 458 EMIT4(0x48, 0x8b, 0x5d, 0xf8); /* mov -8(%rbp),%rbx */ 459 EMIT1(0xc9); /* leaveq */ 460 } 461 EMIT1(0xc3); /* ret */ 462 break; 463 case BPF_S_MISC_TAX: /* X = A */ 464 seen |= SEEN_XREG; 465 EMIT2(0x89, 0xc3); /* mov %eax,%ebx */ 466 break; 467 case BPF_S_MISC_TXA: /* A = X */ 468 seen |= SEEN_XREG; 469 EMIT2(0x89, 0xd8); /* mov %ebx,%eax */ 470 break; 471 case BPF_S_LD_IMM: /* A = K */ 472 if (!K) 473 CLEAR_A(); 474 else 475 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */ 476 break; 477 case BPF_S_LDX_IMM: /* X = K */ 478 seen |= SEEN_XREG; 479 if (!K) 480 CLEAR_X(); 481 else 482 EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */ 483 break; 484 case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */ 485 seen |= SEEN_MEM; 486 EMIT3(0x8b, 0x45, 0xf0 - K*4); 487 break; 488 case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */ 489 seen |= SEEN_XREG | SEEN_MEM; 490 EMIT3(0x8b, 0x5d, 0xf0 - K*4); 491 break; 492 case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */ 493 seen |= SEEN_MEM; 494 EMIT3(0x89, 0x45, 0xf0 - K*4); 495 break; 496 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */ 497 seen |= SEEN_XREG | SEEN_MEM; 498 EMIT3(0x89, 0x5d, 0xf0 - K*4); 499 break; 500 case BPF_S_LD_W_LEN: /* A = skb->len; */ 501 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); 502 if (is_imm8(offsetof(struct sk_buff, len))) 503 /* mov off8(%rdi),%eax */ 504 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len)); 505 else { 506 EMIT2(0x8b, 0x87); 507 EMIT(offsetof(struct sk_buff, len), 4); 508 } 509 break; 510 case BPF_S_LDX_W_LEN: /* X = skb->len; */ 511 seen |= SEEN_XREG; 512 if (is_imm8(offsetof(struct sk_buff, len))) 513 /* mov off8(%rdi),%ebx */ 514 EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len)); 515 else { 516 EMIT2(0x8b, 0x9f); 517 EMIT(offsetof(struct sk_buff, len), 4); 518 } 519 break; 520 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */ 521 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2); 522 if (is_imm8(offsetof(struct sk_buff, protocol))) { 523 /* movzwl off8(%rdi),%eax */ 524 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol)); 525 } else { 526 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 527 EMIT(offsetof(struct sk_buff, protocol), 4); 528 } 529 EMIT2(0x86, 0xc4); /* ntohs() : xchg %al,%ah */ 530 break; 531 case BPF_S_ANC_IFINDEX: 532 if (is_imm8(offsetof(struct sk_buff, dev))) { 533 /* movq off8(%rdi),%rax */ 534 EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev)); 535 } else { 536 EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */ 537 EMIT(offsetof(struct sk_buff, dev), 4); 538 } 539 EMIT3(0x48, 0x85, 0xc0); /* test %rax,%rax */ 540 EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6)); 541 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4); 542 EMIT2(0x8b, 0x80); /* mov off32(%rax),%eax */ 543 EMIT(offsetof(struct net_device, ifindex), 4); 544 break; 545 case BPF_S_ANC_MARK: 546 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 547 if (is_imm8(offsetof(struct sk_buff, mark))) { 548 /* mov off8(%rdi),%eax */ 549 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark)); 550 } else { 551 EMIT2(0x8b, 0x87); 552 EMIT(offsetof(struct sk_buff, mark), 4); 553 } 554 break; 555 case BPF_S_ANC_RXHASH: 556 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4); 557 if (is_imm8(offsetof(struct sk_buff, hash))) { 558 /* mov off8(%rdi),%eax */ 559 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, hash)); 560 } else { 561 EMIT2(0x8b, 0x87); 562 EMIT(offsetof(struct sk_buff, hash), 4); 563 } 564 break; 565 case BPF_S_ANC_QUEUE: 566 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2); 567 if (is_imm8(offsetof(struct sk_buff, queue_mapping))) { 568 /* movzwl off8(%rdi),%eax */ 569 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping)); 570 } else { 571 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 572 EMIT(offsetof(struct sk_buff, queue_mapping), 4); 573 } 574 break; 575 case BPF_S_ANC_CPU: 576 #ifdef CONFIG_SMP 577 EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */ 578 EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */ 579 #else 580 CLEAR_A(); 581 #endif 582 break; 583 case BPF_S_ANC_VLAN_TAG: 584 case BPF_S_ANC_VLAN_TAG_PRESENT: 585 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 586 if (is_imm8(offsetof(struct sk_buff, vlan_tci))) { 587 /* movzwl off8(%rdi),%eax */ 588 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, vlan_tci)); 589 } else { 590 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 591 EMIT(offsetof(struct sk_buff, vlan_tci), 4); 592 } 593 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000); 594 if (filter[i].code == BPF_S_ANC_VLAN_TAG) { 595 EMIT3(0x80, 0xe4, 0xef); /* and $0xef,%ah */ 596 } else { 597 EMIT3(0xc1, 0xe8, 0x0c); /* shr $0xc,%eax */ 598 EMIT3(0x83, 0xe0, 0x01); /* and $0x1,%eax */ 599 } 600 break; 601 case BPF_S_ANC_PKTTYPE: 602 { 603 int off = pkt_type_offset(); 604 605 if (off < 0) 606 goto out; 607 if (is_imm8(off)) { 608 /* movzbl off8(%rdi),%eax */ 609 EMIT4(0x0f, 0xb6, 0x47, off); 610 } else { 611 /* movbl off32(%rdi),%eax */ 612 EMIT3(0x0f, 0xb6, 0x87); 613 EMIT(off, 4); 614 } 615 EMIT3(0x83, 0xe0, PKT_TYPE_MAX); /* and $0x7,%eax */ 616 break; 617 } 618 case BPF_S_LD_W_ABS: 619 func = CHOOSE_LOAD_FUNC(K, sk_load_word); 620 common_load: seen |= SEEN_DATAREF; 621 t_offset = func - (image + addrs[i]); 622 EMIT1_off32(0xbe, K); /* mov imm32,%esi */ 623 EMIT1_off32(0xe8, t_offset); /* call */ 624 break; 625 case BPF_S_LD_H_ABS: 626 func = CHOOSE_LOAD_FUNC(K, sk_load_half); 627 goto common_load; 628 case BPF_S_LD_B_ABS: 629 func = CHOOSE_LOAD_FUNC(K, sk_load_byte); 630 goto common_load; 631 case BPF_S_LDX_B_MSH: 632 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh); 633 seen |= SEEN_DATAREF | SEEN_XREG; 634 t_offset = func - (image + addrs[i]); 635 EMIT1_off32(0xbe, K); /* mov imm32,%esi */ 636 EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */ 637 break; 638 case BPF_S_LD_W_IND: 639 func = sk_load_word; 640 common_load_ind: seen |= SEEN_DATAREF | SEEN_XREG; 641 t_offset = func - (image + addrs[i]); 642 if (K) { 643 if (is_imm8(K)) { 644 EMIT3(0x8d, 0x73, K); /* lea imm8(%rbx), %esi */ 645 } else { 646 EMIT2(0x8d, 0xb3); /* lea imm32(%rbx),%esi */ 647 EMIT(K, 4); 648 } 649 } else { 650 EMIT2(0x89,0xde); /* mov %ebx,%esi */ 651 } 652 EMIT1_off32(0xe8, t_offset); /* call sk_load_xxx_ind */ 653 break; 654 case BPF_S_LD_H_IND: 655 func = sk_load_half; 656 goto common_load_ind; 657 case BPF_S_LD_B_IND: 658 func = sk_load_byte; 659 goto common_load_ind; 660 case BPF_S_JMP_JA: 661 t_offset = addrs[i + K] - addrs[i]; 662 EMIT_JMP(t_offset); 663 break; 664 COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE); 665 COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB); 666 COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE); 667 COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE); 668 COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE); 669 COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB); 670 COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE); 671 COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE); 672 673 cond_branch: f_offset = addrs[i + filter[i].jf] - addrs[i]; 674 t_offset = addrs[i + filter[i].jt] - addrs[i]; 675 676 /* same targets, can avoid doing the test :) */ 677 if (filter[i].jt == filter[i].jf) { 678 EMIT_JMP(t_offset); 679 break; 680 } 681 682 switch (filter[i].code) { 683 case BPF_S_JMP_JGT_X: 684 case BPF_S_JMP_JGE_X: 685 case BPF_S_JMP_JEQ_X: 686 seen |= SEEN_XREG; 687 EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */ 688 break; 689 case BPF_S_JMP_JSET_X: 690 seen |= SEEN_XREG; 691 EMIT2(0x85, 0xd8); /* test %ebx,%eax */ 692 break; 693 case BPF_S_JMP_JEQ_K: 694 if (K == 0) { 695 EMIT2(0x85, 0xc0); /* test %eax,%eax */ 696 break; 697 } 698 case BPF_S_JMP_JGT_K: 699 case BPF_S_JMP_JGE_K: 700 if (K <= 127) 701 EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */ 702 else 703 EMIT1_off32(0x3d, K); /* cmp imm32,%eax */ 704 break; 705 case BPF_S_JMP_JSET_K: 706 if (K <= 0xFF) 707 EMIT2(0xa8, K); /* test imm8,%al */ 708 else if (!(K & 0xFFFF00FF)) 709 EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */ 710 else if (K <= 0xFFFF) { 711 EMIT2(0x66, 0xa9); /* test imm16,%ax */ 712 EMIT(K, 2); 713 } else { 714 EMIT1_off32(0xa9, K); /* test imm32,%eax */ 715 } 716 break; 717 } 718 if (filter[i].jt != 0) { 719 if (filter[i].jf && f_offset) 720 t_offset += is_near(f_offset) ? 2 : 5; 721 EMIT_COND_JMP(t_op, t_offset); 722 if (filter[i].jf) 723 EMIT_JMP(f_offset); 724 break; 725 } 726 EMIT_COND_JMP(f_op, f_offset); 727 break; 728 default: 729 /* hmm, too complex filter, give up with jit compiler */ 730 goto out; 731 } 732 ilen = prog - temp; 733 if (image) { 734 if (unlikely(proglen + ilen > oldproglen)) { 735 pr_err("bpb_jit_compile fatal error\n"); 736 kfree(addrs); 737 module_free(NULL, header); 738 return; 739 } 740 memcpy(image + proglen, temp, ilen); 741 } 742 proglen += ilen; 743 addrs[i] = proglen; 744 prog = temp; 745 } 746 /* last bpf instruction is always a RET : 747 * use it to give the cleanup instruction(s) addr 748 */ 749 cleanup_addr = proglen - 1; /* ret */ 750 if (seen_or_pass0) 751 cleanup_addr -= 1; /* leaveq */ 752 if (seen_or_pass0 & SEEN_XREG) 753 cleanup_addr -= 4; /* mov -8(%rbp),%rbx */ 754 755 if (image) { 756 if (proglen != oldproglen) 757 pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n", proglen, oldproglen); 758 break; 759 } 760 if (proglen == oldproglen) { 761 header = bpf_alloc_binary(proglen, &image); 762 if (!header) 763 goto out; 764 } 765 oldproglen = proglen; 766 } 767 768 if (bpf_jit_enable > 1) 769 bpf_jit_dump(flen, proglen, pass, image); 770 771 if (image) { 772 bpf_flush_icache(header, image + proglen); 773 set_memory_ro((unsigned long)header, header->pages); 774 fp->bpf_func = (void *)image; 775 fp->jited = 1; 776 } 777 out: 778 kfree(addrs); 779 return; 780 } 781 782 static void bpf_jit_free_deferred(struct work_struct *work) 783 { 784 struct sk_filter *fp = container_of(work, struct sk_filter, work); 785 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK; 786 struct bpf_binary_header *header = (void *)addr; 787 788 set_memory_rw(addr, header->pages); 789 module_free(NULL, header); 790 kfree(fp); 791 } 792 793 void bpf_jit_free(struct sk_filter *fp) 794 { 795 if (fp->jited) { 796 INIT_WORK(&fp->work, bpf_jit_free_deferred); 797 schedule_work(&fp->work); 798 } else { 799 kfree(fp); 800 } 801 } 802