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