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 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 /* Helper to find the offset of pkt_type in sk_buff 128 * We want to make sure its still a 3bit field starting at a byte boundary. 129 */ 130 #define PKT_TYPE_MAX 7 131 static int pkt_type_offset(void) 132 { 133 struct sk_buff skb_probe = { 134 .pkt_type = ~0, 135 }; 136 char *ct = (char *)&skb_probe; 137 unsigned int off; 138 139 for (off = 0; off < sizeof(struct sk_buff); off++) { 140 if (ct[off] == PKT_TYPE_MAX) 141 return off; 142 } 143 pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be found\n"); 144 return -1; 145 } 146 147 void bpf_jit_compile(struct sk_filter *fp) 148 { 149 u8 temp[64]; 150 u8 *prog; 151 unsigned int proglen, oldproglen = 0; 152 int ilen, i; 153 int t_offset, f_offset; 154 u8 t_op, f_op, seen = 0, pass; 155 u8 *image = NULL; 156 u8 *func; 157 int pc_ret0 = -1; /* bpf index of first RET #0 instruction (if any) */ 158 unsigned int cleanup_addr; /* epilogue code offset */ 159 unsigned int *addrs; 160 const struct sock_filter *filter = fp->insns; 161 int flen = fp->len; 162 163 if (!bpf_jit_enable) 164 return; 165 166 addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL); 167 if (addrs == NULL) 168 return; 169 170 /* Before first pass, make a rough estimation of addrs[] 171 * each bpf instruction is translated to less than 64 bytes 172 */ 173 for (proglen = 0, i = 0; i < flen; i++) { 174 proglen += 64; 175 addrs[i] = proglen; 176 } 177 cleanup_addr = proglen; /* epilogue address */ 178 179 for (pass = 0; pass < 10; pass++) { 180 u8 seen_or_pass0 = (pass == 0) ? (SEEN_XREG | SEEN_DATAREF | SEEN_MEM) : seen; 181 /* no prologue/epilogue for trivial filters (RET something) */ 182 proglen = 0; 183 prog = temp; 184 185 if (seen_or_pass0) { 186 EMIT4(0x55, 0x48, 0x89, 0xe5); /* push %rbp; mov %rsp,%rbp */ 187 EMIT4(0x48, 0x83, 0xec, 96); /* subq $96,%rsp */ 188 /* note : must save %rbx in case bpf_error is hit */ 189 if (seen_or_pass0 & (SEEN_XREG | SEEN_DATAREF)) 190 EMIT4(0x48, 0x89, 0x5d, 0xf8); /* mov %rbx, -8(%rbp) */ 191 if (seen_or_pass0 & SEEN_XREG) 192 CLEAR_X(); /* make sure we dont leek kernel memory */ 193 194 /* 195 * If this filter needs to access skb data, 196 * loads r9 and r8 with : 197 * r9 = skb->len - skb->data_len 198 * r8 = skb->data 199 */ 200 if (seen_or_pass0 & SEEN_DATAREF) { 201 if (offsetof(struct sk_buff, len) <= 127) 202 /* mov off8(%rdi),%r9d */ 203 EMIT4(0x44, 0x8b, 0x4f, offsetof(struct sk_buff, len)); 204 else { 205 /* mov off32(%rdi),%r9d */ 206 EMIT3(0x44, 0x8b, 0x8f); 207 EMIT(offsetof(struct sk_buff, len), 4); 208 } 209 if (is_imm8(offsetof(struct sk_buff, data_len))) 210 /* sub off8(%rdi),%r9d */ 211 EMIT4(0x44, 0x2b, 0x4f, offsetof(struct sk_buff, data_len)); 212 else { 213 EMIT3(0x44, 0x2b, 0x8f); 214 EMIT(offsetof(struct sk_buff, data_len), 4); 215 } 216 217 if (is_imm8(offsetof(struct sk_buff, data))) 218 /* mov off8(%rdi),%r8 */ 219 EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data)); 220 else { 221 /* mov off32(%rdi),%r8 */ 222 EMIT3(0x4c, 0x8b, 0x87); 223 EMIT(offsetof(struct sk_buff, data), 4); 224 } 225 } 226 } 227 228 switch (filter[0].code) { 229 case BPF_S_RET_K: 230 case BPF_S_LD_W_LEN: 231 case BPF_S_ANC_PROTOCOL: 232 case BPF_S_ANC_IFINDEX: 233 case BPF_S_ANC_MARK: 234 case BPF_S_ANC_RXHASH: 235 case BPF_S_ANC_CPU: 236 case BPF_S_ANC_VLAN_TAG: 237 case BPF_S_ANC_VLAN_TAG_PRESENT: 238 case BPF_S_ANC_QUEUE: 239 case BPF_S_ANC_PKTTYPE: 240 case BPF_S_LD_W_ABS: 241 case BPF_S_LD_H_ABS: 242 case BPF_S_LD_B_ABS: 243 /* first instruction sets A register (or is RET 'constant') */ 244 break; 245 default: 246 /* make sure we dont leak kernel information to user */ 247 CLEAR_A(); /* A = 0 */ 248 } 249 250 for (i = 0; i < flen; i++) { 251 unsigned int K = filter[i].k; 252 253 switch (filter[i].code) { 254 case BPF_S_ALU_ADD_X: /* A += X; */ 255 seen |= SEEN_XREG; 256 EMIT2(0x01, 0xd8); /* add %ebx,%eax */ 257 break; 258 case BPF_S_ALU_ADD_K: /* A += K; */ 259 if (!K) 260 break; 261 if (is_imm8(K)) 262 EMIT3(0x83, 0xc0, K); /* add imm8,%eax */ 263 else 264 EMIT1_off32(0x05, K); /* add imm32,%eax */ 265 break; 266 case BPF_S_ALU_SUB_X: /* A -= X; */ 267 seen |= SEEN_XREG; 268 EMIT2(0x29, 0xd8); /* sub %ebx,%eax */ 269 break; 270 case BPF_S_ALU_SUB_K: /* A -= K */ 271 if (!K) 272 break; 273 if (is_imm8(K)) 274 EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */ 275 else 276 EMIT1_off32(0x2d, K); /* sub imm32,%eax */ 277 break; 278 case BPF_S_ALU_MUL_X: /* A *= X; */ 279 seen |= SEEN_XREG; 280 EMIT3(0x0f, 0xaf, 0xc3); /* imul %ebx,%eax */ 281 break; 282 case BPF_S_ALU_MUL_K: /* A *= K */ 283 if (is_imm8(K)) 284 EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */ 285 else { 286 EMIT2(0x69, 0xc0); /* imul imm32,%eax */ 287 EMIT(K, 4); 288 } 289 break; 290 case BPF_S_ALU_DIV_X: /* A /= X; */ 291 seen |= SEEN_XREG; 292 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */ 293 if (pc_ret0 > 0) { 294 /* addrs[pc_ret0 - 1] is start address of target 295 * (addrs[i] - 4) is the address following this jmp 296 * ("xor %edx,%edx; div %ebx" being 4 bytes long) 297 */ 298 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] - 299 (addrs[i] - 4)); 300 } else { 301 EMIT_COND_JMP(X86_JNE, 2 + 5); 302 CLEAR_A(); 303 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */ 304 } 305 EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */ 306 break; 307 case BPF_S_ALU_MOD_X: /* A %= X; */ 308 seen |= SEEN_XREG; 309 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */ 310 if (pc_ret0 > 0) { 311 /* addrs[pc_ret0 - 1] is start address of target 312 * (addrs[i] - 6) is the address following this jmp 313 * ("xor %edx,%edx; div %ebx;mov %edx,%eax" being 6 bytes long) 314 */ 315 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] - 316 (addrs[i] - 6)); 317 } else { 318 EMIT_COND_JMP(X86_JNE, 2 + 5); 319 CLEAR_A(); 320 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 6)); /* jmp .+off32 */ 321 } 322 EMIT2(0x31, 0xd2); /* xor %edx,%edx */ 323 EMIT2(0xf7, 0xf3); /* div %ebx */ 324 EMIT2(0x89, 0xd0); /* mov %edx,%eax */ 325 break; 326 case BPF_S_ALU_MOD_K: /* A %= K; */ 327 EMIT2(0x31, 0xd2); /* xor %edx,%edx */ 328 EMIT1(0xb9);EMIT(K, 4); /* mov imm32,%ecx */ 329 EMIT2(0xf7, 0xf1); /* div %ecx */ 330 EMIT2(0x89, 0xd0); /* mov %edx,%eax */ 331 break; 332 case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */ 333 EMIT3(0x48, 0x69, 0xc0); /* imul imm32,%rax,%rax */ 334 EMIT(K, 4); 335 EMIT4(0x48, 0xc1, 0xe8, 0x20); /* shr $0x20,%rax */ 336 break; 337 case BPF_S_ALU_AND_X: 338 seen |= SEEN_XREG; 339 EMIT2(0x21, 0xd8); /* and %ebx,%eax */ 340 break; 341 case BPF_S_ALU_AND_K: 342 if (K >= 0xFFFFFF00) { 343 EMIT2(0x24, K & 0xFF); /* and imm8,%al */ 344 } else if (K >= 0xFFFF0000) { 345 EMIT2(0x66, 0x25); /* and imm16,%ax */ 346 EMIT(K, 2); 347 } else { 348 EMIT1_off32(0x25, K); /* and imm32,%eax */ 349 } 350 break; 351 case BPF_S_ALU_OR_X: 352 seen |= SEEN_XREG; 353 EMIT2(0x09, 0xd8); /* or %ebx,%eax */ 354 break; 355 case BPF_S_ALU_OR_K: 356 if (is_imm8(K)) 357 EMIT3(0x83, 0xc8, K); /* or imm8,%eax */ 358 else 359 EMIT1_off32(0x0d, K); /* or imm32,%eax */ 360 break; 361 case BPF_S_ANC_ALU_XOR_X: /* A ^= X; */ 362 case BPF_S_ALU_XOR_X: 363 seen |= SEEN_XREG; 364 EMIT2(0x31, 0xd8); /* xor %ebx,%eax */ 365 break; 366 case BPF_S_ALU_XOR_K: /* A ^= K; */ 367 if (K == 0) 368 break; 369 if (is_imm8(K)) 370 EMIT3(0x83, 0xf0, K); /* xor imm8,%eax */ 371 else 372 EMIT1_off32(0x35, K); /* xor imm32,%eax */ 373 break; 374 case BPF_S_ALU_LSH_X: /* A <<= X; */ 375 seen |= SEEN_XREG; 376 EMIT4(0x89, 0xd9, 0xd3, 0xe0); /* mov %ebx,%ecx; shl %cl,%eax */ 377 break; 378 case BPF_S_ALU_LSH_K: 379 if (K == 0) 380 break; 381 else if (K == 1) 382 EMIT2(0xd1, 0xe0); /* shl %eax */ 383 else 384 EMIT3(0xc1, 0xe0, K); 385 break; 386 case BPF_S_ALU_RSH_X: /* A >>= X; */ 387 seen |= SEEN_XREG; 388 EMIT4(0x89, 0xd9, 0xd3, 0xe8); /* mov %ebx,%ecx; shr %cl,%eax */ 389 break; 390 case BPF_S_ALU_RSH_K: /* A >>= K; */ 391 if (K == 0) 392 break; 393 else if (K == 1) 394 EMIT2(0xd1, 0xe8); /* shr %eax */ 395 else 396 EMIT3(0xc1, 0xe8, K); 397 break; 398 case BPF_S_ALU_NEG: 399 EMIT2(0xf7, 0xd8); /* neg %eax */ 400 break; 401 case BPF_S_RET_K: 402 if (!K) { 403 if (pc_ret0 == -1) 404 pc_ret0 = i; 405 CLEAR_A(); 406 } else { 407 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */ 408 } 409 /* fallinto */ 410 case BPF_S_RET_A: 411 if (seen_or_pass0) { 412 if (i != flen - 1) { 413 EMIT_JMP(cleanup_addr - addrs[i]); 414 break; 415 } 416 if (seen_or_pass0 & SEEN_XREG) 417 EMIT4(0x48, 0x8b, 0x5d, 0xf8); /* mov -8(%rbp),%rbx */ 418 EMIT1(0xc9); /* leaveq */ 419 } 420 EMIT1(0xc3); /* ret */ 421 break; 422 case BPF_S_MISC_TAX: /* X = A */ 423 seen |= SEEN_XREG; 424 EMIT2(0x89, 0xc3); /* mov %eax,%ebx */ 425 break; 426 case BPF_S_MISC_TXA: /* A = X */ 427 seen |= SEEN_XREG; 428 EMIT2(0x89, 0xd8); /* mov %ebx,%eax */ 429 break; 430 case BPF_S_LD_IMM: /* A = K */ 431 if (!K) 432 CLEAR_A(); 433 else 434 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */ 435 break; 436 case BPF_S_LDX_IMM: /* X = K */ 437 seen |= SEEN_XREG; 438 if (!K) 439 CLEAR_X(); 440 else 441 EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */ 442 break; 443 case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */ 444 seen |= SEEN_MEM; 445 EMIT3(0x8b, 0x45, 0xf0 - K*4); 446 break; 447 case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */ 448 seen |= SEEN_XREG | SEEN_MEM; 449 EMIT3(0x8b, 0x5d, 0xf0 - K*4); 450 break; 451 case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */ 452 seen |= SEEN_MEM; 453 EMIT3(0x89, 0x45, 0xf0 - K*4); 454 break; 455 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */ 456 seen |= SEEN_XREG | SEEN_MEM; 457 EMIT3(0x89, 0x5d, 0xf0 - K*4); 458 break; 459 case BPF_S_LD_W_LEN: /* A = skb->len; */ 460 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); 461 if (is_imm8(offsetof(struct sk_buff, len))) 462 /* mov off8(%rdi),%eax */ 463 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len)); 464 else { 465 EMIT2(0x8b, 0x87); 466 EMIT(offsetof(struct sk_buff, len), 4); 467 } 468 break; 469 case BPF_S_LDX_W_LEN: /* X = skb->len; */ 470 seen |= SEEN_XREG; 471 if (is_imm8(offsetof(struct sk_buff, len))) 472 /* mov off8(%rdi),%ebx */ 473 EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len)); 474 else { 475 EMIT2(0x8b, 0x9f); 476 EMIT(offsetof(struct sk_buff, len), 4); 477 } 478 break; 479 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */ 480 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2); 481 if (is_imm8(offsetof(struct sk_buff, protocol))) { 482 /* movzwl off8(%rdi),%eax */ 483 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol)); 484 } else { 485 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 486 EMIT(offsetof(struct sk_buff, protocol), 4); 487 } 488 EMIT2(0x86, 0xc4); /* ntohs() : xchg %al,%ah */ 489 break; 490 case BPF_S_ANC_IFINDEX: 491 if (is_imm8(offsetof(struct sk_buff, dev))) { 492 /* movq off8(%rdi),%rax */ 493 EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev)); 494 } else { 495 EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */ 496 EMIT(offsetof(struct sk_buff, dev), 4); 497 } 498 EMIT3(0x48, 0x85, 0xc0); /* test %rax,%rax */ 499 EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6)); 500 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4); 501 EMIT2(0x8b, 0x80); /* mov off32(%rax),%eax */ 502 EMIT(offsetof(struct net_device, ifindex), 4); 503 break; 504 case BPF_S_ANC_MARK: 505 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); 506 if (is_imm8(offsetof(struct sk_buff, mark))) { 507 /* mov off8(%rdi),%eax */ 508 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark)); 509 } else { 510 EMIT2(0x8b, 0x87); 511 EMIT(offsetof(struct sk_buff, mark), 4); 512 } 513 break; 514 case BPF_S_ANC_RXHASH: 515 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4); 516 if (is_imm8(offsetof(struct sk_buff, rxhash))) { 517 /* mov off8(%rdi),%eax */ 518 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, rxhash)); 519 } else { 520 EMIT2(0x8b, 0x87); 521 EMIT(offsetof(struct sk_buff, rxhash), 4); 522 } 523 break; 524 case BPF_S_ANC_QUEUE: 525 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2); 526 if (is_imm8(offsetof(struct sk_buff, queue_mapping))) { 527 /* movzwl off8(%rdi),%eax */ 528 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping)); 529 } else { 530 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 531 EMIT(offsetof(struct sk_buff, queue_mapping), 4); 532 } 533 break; 534 case BPF_S_ANC_CPU: 535 #ifdef CONFIG_SMP 536 EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */ 537 EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */ 538 #else 539 CLEAR_A(); 540 #endif 541 break; 542 case BPF_S_ANC_VLAN_TAG: 543 case BPF_S_ANC_VLAN_TAG_PRESENT: 544 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); 545 if (is_imm8(offsetof(struct sk_buff, vlan_tci))) { 546 /* movzwl off8(%rdi),%eax */ 547 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, vlan_tci)); 548 } else { 549 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */ 550 EMIT(offsetof(struct sk_buff, vlan_tci), 4); 551 } 552 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000); 553 if (filter[i].code == BPF_S_ANC_VLAN_TAG) { 554 EMIT3(0x80, 0xe4, 0xef); /* and $0xef,%ah */ 555 } else { 556 EMIT3(0xc1, 0xe8, 0x0c); /* shr $0xc,%eax */ 557 EMIT3(0x83, 0xe0, 0x01); /* and $0x1,%eax */ 558 } 559 break; 560 case BPF_S_ANC_PKTTYPE: 561 { 562 int off = pkt_type_offset(); 563 564 if (off < 0) 565 goto out; 566 if (is_imm8(off)) { 567 /* movzbl off8(%rdi),%eax */ 568 EMIT4(0x0f, 0xb6, 0x47, off); 569 } else { 570 /* movbl off32(%rdi),%eax */ 571 EMIT3(0x0f, 0xb6, 0x87); 572 EMIT(off, 4); 573 } 574 EMIT3(0x83, 0xe0, PKT_TYPE_MAX); /* and $0x7,%eax */ 575 break; 576 } 577 case BPF_S_LD_W_ABS: 578 func = CHOOSE_LOAD_FUNC(K, sk_load_word); 579 common_load: seen |= SEEN_DATAREF; 580 t_offset = func - (image + addrs[i]); 581 EMIT1_off32(0xbe, K); /* mov imm32,%esi */ 582 EMIT1_off32(0xe8, t_offset); /* call */ 583 break; 584 case BPF_S_LD_H_ABS: 585 func = CHOOSE_LOAD_FUNC(K, sk_load_half); 586 goto common_load; 587 case BPF_S_LD_B_ABS: 588 func = CHOOSE_LOAD_FUNC(K, sk_load_byte); 589 goto common_load; 590 case BPF_S_LDX_B_MSH: 591 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh); 592 seen |= SEEN_DATAREF | SEEN_XREG; 593 t_offset = func - (image + addrs[i]); 594 EMIT1_off32(0xbe, K); /* mov imm32,%esi */ 595 EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */ 596 break; 597 case BPF_S_LD_W_IND: 598 func = sk_load_word; 599 common_load_ind: seen |= SEEN_DATAREF | SEEN_XREG; 600 t_offset = func - (image + addrs[i]); 601 if (K) { 602 if (is_imm8(K)) { 603 EMIT3(0x8d, 0x73, K); /* lea imm8(%rbx), %esi */ 604 } else { 605 EMIT2(0x8d, 0xb3); /* lea imm32(%rbx),%esi */ 606 EMIT(K, 4); 607 } 608 } else { 609 EMIT2(0x89,0xde); /* mov %ebx,%esi */ 610 } 611 EMIT1_off32(0xe8, t_offset); /* call sk_load_xxx_ind */ 612 break; 613 case BPF_S_LD_H_IND: 614 func = sk_load_half; 615 goto common_load_ind; 616 case BPF_S_LD_B_IND: 617 func = sk_load_byte; 618 goto common_load_ind; 619 case BPF_S_JMP_JA: 620 t_offset = addrs[i + K] - addrs[i]; 621 EMIT_JMP(t_offset); 622 break; 623 COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE); 624 COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB); 625 COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE); 626 COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE); 627 COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE); 628 COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB); 629 COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE); 630 COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE); 631 632 cond_branch: f_offset = addrs[i + filter[i].jf] - addrs[i]; 633 t_offset = addrs[i + filter[i].jt] - addrs[i]; 634 635 /* same targets, can avoid doing the test :) */ 636 if (filter[i].jt == filter[i].jf) { 637 EMIT_JMP(t_offset); 638 break; 639 } 640 641 switch (filter[i].code) { 642 case BPF_S_JMP_JGT_X: 643 case BPF_S_JMP_JGE_X: 644 case BPF_S_JMP_JEQ_X: 645 seen |= SEEN_XREG; 646 EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */ 647 break; 648 case BPF_S_JMP_JSET_X: 649 seen |= SEEN_XREG; 650 EMIT2(0x85, 0xd8); /* test %ebx,%eax */ 651 break; 652 case BPF_S_JMP_JEQ_K: 653 if (K == 0) { 654 EMIT2(0x85, 0xc0); /* test %eax,%eax */ 655 break; 656 } 657 case BPF_S_JMP_JGT_K: 658 case BPF_S_JMP_JGE_K: 659 if (K <= 127) 660 EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */ 661 else 662 EMIT1_off32(0x3d, K); /* cmp imm32,%eax */ 663 break; 664 case BPF_S_JMP_JSET_K: 665 if (K <= 0xFF) 666 EMIT2(0xa8, K); /* test imm8,%al */ 667 else if (!(K & 0xFFFF00FF)) 668 EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */ 669 else if (K <= 0xFFFF) { 670 EMIT2(0x66, 0xa9); /* test imm16,%ax */ 671 EMIT(K, 2); 672 } else { 673 EMIT1_off32(0xa9, K); /* test imm32,%eax */ 674 } 675 break; 676 } 677 if (filter[i].jt != 0) { 678 if (filter[i].jf && f_offset) 679 t_offset += is_near(f_offset) ? 2 : 5; 680 EMIT_COND_JMP(t_op, t_offset); 681 if (filter[i].jf) 682 EMIT_JMP(f_offset); 683 break; 684 } 685 EMIT_COND_JMP(f_op, f_offset); 686 break; 687 default: 688 /* hmm, too complex filter, give up with jit compiler */ 689 goto out; 690 } 691 ilen = prog - temp; 692 if (image) { 693 if (unlikely(proglen + ilen > oldproglen)) { 694 pr_err("bpb_jit_compile fatal error\n"); 695 kfree(addrs); 696 module_free(NULL, image); 697 return; 698 } 699 memcpy(image + proglen, temp, ilen); 700 } 701 proglen += ilen; 702 addrs[i] = proglen; 703 prog = temp; 704 } 705 /* last bpf instruction is always a RET : 706 * use it to give the cleanup instruction(s) addr 707 */ 708 cleanup_addr = proglen - 1; /* ret */ 709 if (seen_or_pass0) 710 cleanup_addr -= 1; /* leaveq */ 711 if (seen_or_pass0 & SEEN_XREG) 712 cleanup_addr -= 4; /* mov -8(%rbp),%rbx */ 713 714 if (image) { 715 if (proglen != oldproglen) 716 pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n", proglen, oldproglen); 717 break; 718 } 719 if (proglen == oldproglen) { 720 image = module_alloc(max_t(unsigned int, 721 proglen, 722 sizeof(struct work_struct))); 723 if (!image) 724 goto out; 725 } 726 oldproglen = proglen; 727 } 728 729 if (bpf_jit_enable > 1) 730 bpf_jit_dump(flen, proglen, pass, image); 731 732 if (image) { 733 bpf_flush_icache(image, image + proglen); 734 fp->bpf_func = (void *)image; 735 } 736 out: 737 kfree(addrs); 738 return; 739 } 740 741 static void jit_free_defer(struct work_struct *arg) 742 { 743 module_free(NULL, arg); 744 } 745 746 /* run from softirq, we must use a work_struct to call 747 * module_free() from process context 748 */ 749 void bpf_jit_free(struct sk_filter *fp) 750 { 751 if (fp->bpf_func != sk_run_filter) { 752 struct work_struct *work = (struct work_struct *)fp->bpf_func; 753 754 INIT_WORK(work, jit_free_defer); 755 schedule_work(work); 756 } 757 } 758