1 /* bpf_jit_comp.c : BPF JIT compiler 2 * 3 * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com) 4 * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; version 2 9 * of the License. 10 */ 11 #include <linux/netdevice.h> 12 #include <linux/filter.h> 13 #include <linux/if_vlan.h> 14 #include <asm/cacheflush.h> 15 #include <linux/bpf.h> 16 17 int bpf_jit_enable __read_mostly; 18 19 /* 20 * assembly code in arch/x86/net/bpf_jit.S 21 */ 22 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[]; 23 extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[]; 24 extern u8 sk_load_byte_positive_offset[]; 25 extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[]; 26 extern u8 sk_load_byte_negative_offset[]; 27 28 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len) 29 { 30 if (len == 1) 31 *ptr = bytes; 32 else if (len == 2) 33 *(u16 *)ptr = bytes; 34 else { 35 *(u32 *)ptr = bytes; 36 barrier(); 37 } 38 return ptr + len; 39 } 40 41 #define EMIT(bytes, len) \ 42 do { prog = emit_code(prog, bytes, len); cnt += len; } while (0) 43 44 #define EMIT1(b1) EMIT(b1, 1) 45 #define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2) 46 #define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3) 47 #define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4) 48 #define EMIT1_off32(b1, off) \ 49 do {EMIT1(b1); EMIT(off, 4); } while (0) 50 #define EMIT2_off32(b1, b2, off) \ 51 do {EMIT2(b1, b2); EMIT(off, 4); } while (0) 52 #define EMIT3_off32(b1, b2, b3, off) \ 53 do {EMIT3(b1, b2, b3); EMIT(off, 4); } while (0) 54 #define EMIT4_off32(b1, b2, b3, b4, off) \ 55 do {EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0) 56 57 static bool is_imm8(int value) 58 { 59 return value <= 127 && value >= -128; 60 } 61 62 static bool is_simm32(s64 value) 63 { 64 return value == (s64) (s32) value; 65 } 66 67 /* mov dst, src */ 68 #define EMIT_mov(DST, SRC) \ 69 do {if (DST != SRC) \ 70 EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \ 71 } while (0) 72 73 static int bpf_size_to_x86_bytes(int bpf_size) 74 { 75 if (bpf_size == BPF_W) 76 return 4; 77 else if (bpf_size == BPF_H) 78 return 2; 79 else if (bpf_size == BPF_B) 80 return 1; 81 else if (bpf_size == BPF_DW) 82 return 4; /* imm32 */ 83 else 84 return 0; 85 } 86 87 /* list of x86 cond jumps opcodes (. + s8) 88 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32) 89 */ 90 #define X86_JB 0x72 91 #define X86_JAE 0x73 92 #define X86_JE 0x74 93 #define X86_JNE 0x75 94 #define X86_JBE 0x76 95 #define X86_JA 0x77 96 #define X86_JGE 0x7D 97 #define X86_JG 0x7F 98 99 static void bpf_flush_icache(void *start, void *end) 100 { 101 mm_segment_t old_fs = get_fs(); 102 103 set_fs(KERNEL_DS); 104 smp_wmb(); 105 flush_icache_range((unsigned long)start, (unsigned long)end); 106 set_fs(old_fs); 107 } 108 109 #define CHOOSE_LOAD_FUNC(K, func) \ 110 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset) 111 112 /* pick a register outside of BPF range for JIT internal work */ 113 #define AUX_REG (MAX_BPF_JIT_REG + 1) 114 115 /* The following table maps BPF registers to x64 registers. 116 * 117 * x64 register r12 is unused, since if used as base address 118 * register in load/store instructions, it always needs an 119 * extra byte of encoding and is callee saved. 120 * 121 * r9 caches skb->len - skb->data_len 122 * r10 caches skb->data, and used for blinding (if enabled) 123 */ 124 static const int reg2hex[] = { 125 [BPF_REG_0] = 0, /* rax */ 126 [BPF_REG_1] = 7, /* rdi */ 127 [BPF_REG_2] = 6, /* rsi */ 128 [BPF_REG_3] = 2, /* rdx */ 129 [BPF_REG_4] = 1, /* rcx */ 130 [BPF_REG_5] = 0, /* r8 */ 131 [BPF_REG_6] = 3, /* rbx callee saved */ 132 [BPF_REG_7] = 5, /* r13 callee saved */ 133 [BPF_REG_8] = 6, /* r14 callee saved */ 134 [BPF_REG_9] = 7, /* r15 callee saved */ 135 [BPF_REG_FP] = 5, /* rbp readonly */ 136 [BPF_REG_AX] = 2, /* r10 temp register */ 137 [AUX_REG] = 3, /* r11 temp register */ 138 }; 139 140 /* is_ereg() == true if BPF register 'reg' maps to x64 r8..r15 141 * which need extra byte of encoding. 142 * rax,rcx,...,rbp have simpler encoding 143 */ 144 static bool is_ereg(u32 reg) 145 { 146 return (1 << reg) & (BIT(BPF_REG_5) | 147 BIT(AUX_REG) | 148 BIT(BPF_REG_7) | 149 BIT(BPF_REG_8) | 150 BIT(BPF_REG_9) | 151 BIT(BPF_REG_AX)); 152 } 153 154 /* add modifiers if 'reg' maps to x64 registers r8..r15 */ 155 static u8 add_1mod(u8 byte, u32 reg) 156 { 157 if (is_ereg(reg)) 158 byte |= 1; 159 return byte; 160 } 161 162 static u8 add_2mod(u8 byte, u32 r1, u32 r2) 163 { 164 if (is_ereg(r1)) 165 byte |= 1; 166 if (is_ereg(r2)) 167 byte |= 4; 168 return byte; 169 } 170 171 /* encode 'dst_reg' register into x64 opcode 'byte' */ 172 static u8 add_1reg(u8 byte, u32 dst_reg) 173 { 174 return byte + reg2hex[dst_reg]; 175 } 176 177 /* encode 'dst_reg' and 'src_reg' registers into x64 opcode 'byte' */ 178 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg) 179 { 180 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3); 181 } 182 183 static void jit_fill_hole(void *area, unsigned int size) 184 { 185 /* fill whole space with int3 instructions */ 186 memset(area, 0xcc, size); 187 } 188 189 struct jit_context { 190 int cleanup_addr; /* epilogue code offset */ 191 bool seen_ld_abs; 192 bool seen_ax_reg; 193 }; 194 195 /* maximum number of bytes emitted while JITing one eBPF insn */ 196 #define BPF_MAX_INSN_SIZE 128 197 #define BPF_INSN_SAFETY 64 198 199 #define STACKSIZE \ 200 (MAX_BPF_STACK + \ 201 32 /* space for rbx, r13, r14, r15 */ + \ 202 8 /* space for skb_copy_bits() buffer */) 203 204 #define PROLOGUE_SIZE 48 205 206 /* emit x64 prologue code for BPF program and check it's size. 207 * bpf_tail_call helper will skip it while jumping into another program 208 */ 209 static void emit_prologue(u8 **pprog) 210 { 211 u8 *prog = *pprog; 212 int cnt = 0; 213 214 EMIT1(0x55); /* push rbp */ 215 EMIT3(0x48, 0x89, 0xE5); /* mov rbp,rsp */ 216 217 /* sub rsp, STACKSIZE */ 218 EMIT3_off32(0x48, 0x81, 0xEC, STACKSIZE); 219 220 /* all classic BPF filters use R6(rbx) save it */ 221 222 /* mov qword ptr [rbp-X],rbx */ 223 EMIT3_off32(0x48, 0x89, 0x9D, -STACKSIZE); 224 225 /* bpf_convert_filter() maps classic BPF register X to R7 and uses R8 226 * as temporary, so all tcpdump filters need to spill/fill R7(r13) and 227 * R8(r14). R9(r15) spill could be made conditional, but there is only 228 * one 'bpf_error' return path out of helper functions inside bpf_jit.S 229 * The overhead of extra spill is negligible for any filter other 230 * than synthetic ones. Therefore not worth adding complexity. 231 */ 232 233 /* mov qword ptr [rbp-X],r13 */ 234 EMIT3_off32(0x4C, 0x89, 0xAD, -STACKSIZE + 8); 235 /* mov qword ptr [rbp-X],r14 */ 236 EMIT3_off32(0x4C, 0x89, 0xB5, -STACKSIZE + 16); 237 /* mov qword ptr [rbp-X],r15 */ 238 EMIT3_off32(0x4C, 0x89, 0xBD, -STACKSIZE + 24); 239 240 /* Clear the tail call counter (tail_call_cnt): for eBPF tail calls 241 * we need to reset the counter to 0. It's done in two instructions, 242 * resetting rax register to 0 (xor on eax gets 0 extended), and 243 * moving it to the counter location. 244 */ 245 246 /* xor eax, eax */ 247 EMIT2(0x31, 0xc0); 248 /* mov qword ptr [rbp-X], rax */ 249 EMIT3_off32(0x48, 0x89, 0x85, -STACKSIZE + 32); 250 251 BUILD_BUG_ON(cnt != PROLOGUE_SIZE); 252 *pprog = prog; 253 } 254 255 /* generate the following code: 256 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ... 257 * if (index >= array->map.max_entries) 258 * goto out; 259 * if (++tail_call_cnt > MAX_TAIL_CALL_CNT) 260 * goto out; 261 * prog = array->ptrs[index]; 262 * if (prog == NULL) 263 * goto out; 264 * goto *(prog->bpf_func + prologue_size); 265 * out: 266 */ 267 static void emit_bpf_tail_call(u8 **pprog) 268 { 269 u8 *prog = *pprog; 270 int label1, label2, label3; 271 int cnt = 0; 272 273 /* rdi - pointer to ctx 274 * rsi - pointer to bpf_array 275 * rdx - index in bpf_array 276 */ 277 278 /* if (index >= array->map.max_entries) 279 * goto out; 280 */ 281 EMIT4(0x48, 0x8B, 0x46, /* mov rax, qword ptr [rsi + 16] */ 282 offsetof(struct bpf_array, map.max_entries)); 283 EMIT3(0x48, 0x39, 0xD0); /* cmp rax, rdx */ 284 #define OFFSET1 47 /* number of bytes to jump */ 285 EMIT2(X86_JBE, OFFSET1); /* jbe out */ 286 label1 = cnt; 287 288 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT) 289 * goto out; 290 */ 291 EMIT2_off32(0x8B, 0x85, -STACKSIZE + 36); /* mov eax, dword ptr [rbp - 516] */ 292 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */ 293 #define OFFSET2 36 294 EMIT2(X86_JA, OFFSET2); /* ja out */ 295 label2 = cnt; 296 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */ 297 EMIT2_off32(0x89, 0x85, -STACKSIZE + 36); /* mov dword ptr [rbp - 516], eax */ 298 299 /* prog = array->ptrs[index]; */ 300 EMIT4_off32(0x48, 0x8D, 0x84, 0xD6, /* lea rax, [rsi + rdx * 8 + offsetof(...)] */ 301 offsetof(struct bpf_array, ptrs)); 302 EMIT3(0x48, 0x8B, 0x00); /* mov rax, qword ptr [rax] */ 303 304 /* if (prog == NULL) 305 * goto out; 306 */ 307 EMIT4(0x48, 0x83, 0xF8, 0x00); /* cmp rax, 0 */ 308 #define OFFSET3 10 309 EMIT2(X86_JE, OFFSET3); /* je out */ 310 label3 = cnt; 311 312 /* goto *(prog->bpf_func + prologue_size); */ 313 EMIT4(0x48, 0x8B, 0x40, /* mov rax, qword ptr [rax + 32] */ 314 offsetof(struct bpf_prog, bpf_func)); 315 EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE); /* add rax, prologue_size */ 316 317 /* now we're ready to jump into next BPF program 318 * rdi == ctx (1st arg) 319 * rax == prog->bpf_func + prologue_size 320 */ 321 EMIT2(0xFF, 0xE0); /* jmp rax */ 322 323 /* out: */ 324 BUILD_BUG_ON(cnt - label1 != OFFSET1); 325 BUILD_BUG_ON(cnt - label2 != OFFSET2); 326 BUILD_BUG_ON(cnt - label3 != OFFSET3); 327 *pprog = prog; 328 } 329 330 331 static void emit_load_skb_data_hlen(u8 **pprog) 332 { 333 u8 *prog = *pprog; 334 int cnt = 0; 335 336 /* r9d = skb->len - skb->data_len (headlen) 337 * r10 = skb->data 338 */ 339 /* mov %r9d, off32(%rdi) */ 340 EMIT3_off32(0x44, 0x8b, 0x8f, offsetof(struct sk_buff, len)); 341 342 /* sub %r9d, off32(%rdi) */ 343 EMIT3_off32(0x44, 0x2b, 0x8f, offsetof(struct sk_buff, data_len)); 344 345 /* mov %r10, off32(%rdi) */ 346 EMIT3_off32(0x4c, 0x8b, 0x97, offsetof(struct sk_buff, data)); 347 *pprog = prog; 348 } 349 350 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, 351 int oldproglen, struct jit_context *ctx) 352 { 353 struct bpf_insn *insn = bpf_prog->insnsi; 354 int insn_cnt = bpf_prog->len; 355 bool seen_ld_abs = ctx->seen_ld_abs | (oldproglen == 0); 356 bool seen_ax_reg = ctx->seen_ax_reg | (oldproglen == 0); 357 bool seen_exit = false; 358 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY]; 359 int i, cnt = 0; 360 int proglen = 0; 361 u8 *prog = temp; 362 363 emit_prologue(&prog); 364 365 if (seen_ld_abs) 366 emit_load_skb_data_hlen(&prog); 367 368 for (i = 0; i < insn_cnt; i++, insn++) { 369 const s32 imm32 = insn->imm; 370 u32 dst_reg = insn->dst_reg; 371 u32 src_reg = insn->src_reg; 372 u8 b1 = 0, b2 = 0, b3 = 0; 373 s64 jmp_offset; 374 u8 jmp_cond; 375 bool reload_skb_data; 376 int ilen; 377 u8 *func; 378 379 if (dst_reg == BPF_REG_AX || src_reg == BPF_REG_AX) 380 ctx->seen_ax_reg = seen_ax_reg = true; 381 382 switch (insn->code) { 383 /* ALU */ 384 case BPF_ALU | BPF_ADD | BPF_X: 385 case BPF_ALU | BPF_SUB | BPF_X: 386 case BPF_ALU | BPF_AND | BPF_X: 387 case BPF_ALU | BPF_OR | BPF_X: 388 case BPF_ALU | BPF_XOR | BPF_X: 389 case BPF_ALU64 | BPF_ADD | BPF_X: 390 case BPF_ALU64 | BPF_SUB | BPF_X: 391 case BPF_ALU64 | BPF_AND | BPF_X: 392 case BPF_ALU64 | BPF_OR | BPF_X: 393 case BPF_ALU64 | BPF_XOR | BPF_X: 394 switch (BPF_OP(insn->code)) { 395 case BPF_ADD: b2 = 0x01; break; 396 case BPF_SUB: b2 = 0x29; break; 397 case BPF_AND: b2 = 0x21; break; 398 case BPF_OR: b2 = 0x09; break; 399 case BPF_XOR: b2 = 0x31; break; 400 } 401 if (BPF_CLASS(insn->code) == BPF_ALU64) 402 EMIT1(add_2mod(0x48, dst_reg, src_reg)); 403 else if (is_ereg(dst_reg) || is_ereg(src_reg)) 404 EMIT1(add_2mod(0x40, dst_reg, src_reg)); 405 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg)); 406 break; 407 408 /* mov dst, src */ 409 case BPF_ALU64 | BPF_MOV | BPF_X: 410 EMIT_mov(dst_reg, src_reg); 411 break; 412 413 /* mov32 dst, src */ 414 case BPF_ALU | BPF_MOV | BPF_X: 415 if (is_ereg(dst_reg) || is_ereg(src_reg)) 416 EMIT1(add_2mod(0x40, dst_reg, src_reg)); 417 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg)); 418 break; 419 420 /* neg dst */ 421 case BPF_ALU | BPF_NEG: 422 case BPF_ALU64 | BPF_NEG: 423 if (BPF_CLASS(insn->code) == BPF_ALU64) 424 EMIT1(add_1mod(0x48, dst_reg)); 425 else if (is_ereg(dst_reg)) 426 EMIT1(add_1mod(0x40, dst_reg)); 427 EMIT2(0xF7, add_1reg(0xD8, dst_reg)); 428 break; 429 430 case BPF_ALU | BPF_ADD | BPF_K: 431 case BPF_ALU | BPF_SUB | BPF_K: 432 case BPF_ALU | BPF_AND | BPF_K: 433 case BPF_ALU | BPF_OR | BPF_K: 434 case BPF_ALU | BPF_XOR | BPF_K: 435 case BPF_ALU64 | BPF_ADD | BPF_K: 436 case BPF_ALU64 | BPF_SUB | BPF_K: 437 case BPF_ALU64 | BPF_AND | BPF_K: 438 case BPF_ALU64 | BPF_OR | BPF_K: 439 case BPF_ALU64 | BPF_XOR | BPF_K: 440 if (BPF_CLASS(insn->code) == BPF_ALU64) 441 EMIT1(add_1mod(0x48, dst_reg)); 442 else if (is_ereg(dst_reg)) 443 EMIT1(add_1mod(0x40, dst_reg)); 444 445 switch (BPF_OP(insn->code)) { 446 case BPF_ADD: b3 = 0xC0; break; 447 case BPF_SUB: b3 = 0xE8; break; 448 case BPF_AND: b3 = 0xE0; break; 449 case BPF_OR: b3 = 0xC8; break; 450 case BPF_XOR: b3 = 0xF0; break; 451 } 452 453 if (is_imm8(imm32)) 454 EMIT3(0x83, add_1reg(b3, dst_reg), imm32); 455 else 456 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32); 457 break; 458 459 case BPF_ALU64 | BPF_MOV | BPF_K: 460 /* optimization: if imm32 is positive, 461 * use 'mov eax, imm32' (which zero-extends imm32) 462 * to save 2 bytes 463 */ 464 if (imm32 < 0) { 465 /* 'mov rax, imm32' sign extends imm32 */ 466 b1 = add_1mod(0x48, dst_reg); 467 b2 = 0xC7; 468 b3 = 0xC0; 469 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32); 470 break; 471 } 472 473 case BPF_ALU | BPF_MOV | BPF_K: 474 /* optimization: if imm32 is zero, use 'xor <dst>,<dst>' 475 * to save 3 bytes. 476 */ 477 if (imm32 == 0) { 478 if (is_ereg(dst_reg)) 479 EMIT1(add_2mod(0x40, dst_reg, dst_reg)); 480 b2 = 0x31; /* xor */ 481 b3 = 0xC0; 482 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg)); 483 break; 484 } 485 486 /* mov %eax, imm32 */ 487 if (is_ereg(dst_reg)) 488 EMIT1(add_1mod(0x40, dst_reg)); 489 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32); 490 break; 491 492 case BPF_LD | BPF_IMM | BPF_DW: 493 /* optimization: if imm64 is zero, use 'xor <dst>,<dst>' 494 * to save 7 bytes. 495 */ 496 if (insn[0].imm == 0 && insn[1].imm == 0) { 497 b1 = add_2mod(0x48, dst_reg, dst_reg); 498 b2 = 0x31; /* xor */ 499 b3 = 0xC0; 500 EMIT3(b1, b2, add_2reg(b3, dst_reg, dst_reg)); 501 502 insn++; 503 i++; 504 break; 505 } 506 507 /* movabsq %rax, imm64 */ 508 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg)); 509 EMIT(insn[0].imm, 4); 510 EMIT(insn[1].imm, 4); 511 512 insn++; 513 i++; 514 break; 515 516 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */ 517 case BPF_ALU | BPF_MOD | BPF_X: 518 case BPF_ALU | BPF_DIV | BPF_X: 519 case BPF_ALU | BPF_MOD | BPF_K: 520 case BPF_ALU | BPF_DIV | BPF_K: 521 case BPF_ALU64 | BPF_MOD | BPF_X: 522 case BPF_ALU64 | BPF_DIV | BPF_X: 523 case BPF_ALU64 | BPF_MOD | BPF_K: 524 case BPF_ALU64 | BPF_DIV | BPF_K: 525 EMIT1(0x50); /* push rax */ 526 EMIT1(0x52); /* push rdx */ 527 528 if (BPF_SRC(insn->code) == BPF_X) 529 /* mov r11, src_reg */ 530 EMIT_mov(AUX_REG, src_reg); 531 else 532 /* mov r11, imm32 */ 533 EMIT3_off32(0x49, 0xC7, 0xC3, imm32); 534 535 /* mov rax, dst_reg */ 536 EMIT_mov(BPF_REG_0, dst_reg); 537 538 /* xor edx, edx 539 * equivalent to 'xor rdx, rdx', but one byte less 540 */ 541 EMIT2(0x31, 0xd2); 542 543 if (BPF_SRC(insn->code) == BPF_X) { 544 /* if (src_reg == 0) return 0 */ 545 546 /* cmp r11, 0 */ 547 EMIT4(0x49, 0x83, 0xFB, 0x00); 548 549 /* jne .+9 (skip over pop, pop, xor and jmp) */ 550 EMIT2(X86_JNE, 1 + 1 + 2 + 5); 551 EMIT1(0x5A); /* pop rdx */ 552 EMIT1(0x58); /* pop rax */ 553 EMIT2(0x31, 0xc0); /* xor eax, eax */ 554 555 /* jmp cleanup_addr 556 * addrs[i] - 11, because there are 11 bytes 557 * after this insn: div, mov, pop, pop, mov 558 */ 559 jmp_offset = ctx->cleanup_addr - (addrs[i] - 11); 560 EMIT1_off32(0xE9, jmp_offset); 561 } 562 563 if (BPF_CLASS(insn->code) == BPF_ALU64) 564 /* div r11 */ 565 EMIT3(0x49, 0xF7, 0xF3); 566 else 567 /* div r11d */ 568 EMIT3(0x41, 0xF7, 0xF3); 569 570 if (BPF_OP(insn->code) == BPF_MOD) 571 /* mov r11, rdx */ 572 EMIT3(0x49, 0x89, 0xD3); 573 else 574 /* mov r11, rax */ 575 EMIT3(0x49, 0x89, 0xC3); 576 577 EMIT1(0x5A); /* pop rdx */ 578 EMIT1(0x58); /* pop rax */ 579 580 /* mov dst_reg, r11 */ 581 EMIT_mov(dst_reg, AUX_REG); 582 break; 583 584 case BPF_ALU | BPF_MUL | BPF_K: 585 case BPF_ALU | BPF_MUL | BPF_X: 586 case BPF_ALU64 | BPF_MUL | BPF_K: 587 case BPF_ALU64 | BPF_MUL | BPF_X: 588 EMIT1(0x50); /* push rax */ 589 EMIT1(0x52); /* push rdx */ 590 591 /* mov r11, dst_reg */ 592 EMIT_mov(AUX_REG, dst_reg); 593 594 if (BPF_SRC(insn->code) == BPF_X) 595 /* mov rax, src_reg */ 596 EMIT_mov(BPF_REG_0, src_reg); 597 else 598 /* mov rax, imm32 */ 599 EMIT3_off32(0x48, 0xC7, 0xC0, imm32); 600 601 if (BPF_CLASS(insn->code) == BPF_ALU64) 602 EMIT1(add_1mod(0x48, AUX_REG)); 603 else if (is_ereg(AUX_REG)) 604 EMIT1(add_1mod(0x40, AUX_REG)); 605 /* mul(q) r11 */ 606 EMIT2(0xF7, add_1reg(0xE0, AUX_REG)); 607 608 /* mov r11, rax */ 609 EMIT_mov(AUX_REG, BPF_REG_0); 610 611 EMIT1(0x5A); /* pop rdx */ 612 EMIT1(0x58); /* pop rax */ 613 614 /* mov dst_reg, r11 */ 615 EMIT_mov(dst_reg, AUX_REG); 616 break; 617 618 /* shifts */ 619 case BPF_ALU | BPF_LSH | BPF_K: 620 case BPF_ALU | BPF_RSH | BPF_K: 621 case BPF_ALU | BPF_ARSH | BPF_K: 622 case BPF_ALU64 | BPF_LSH | BPF_K: 623 case BPF_ALU64 | BPF_RSH | BPF_K: 624 case BPF_ALU64 | BPF_ARSH | BPF_K: 625 if (BPF_CLASS(insn->code) == BPF_ALU64) 626 EMIT1(add_1mod(0x48, dst_reg)); 627 else if (is_ereg(dst_reg)) 628 EMIT1(add_1mod(0x40, dst_reg)); 629 630 switch (BPF_OP(insn->code)) { 631 case BPF_LSH: b3 = 0xE0; break; 632 case BPF_RSH: b3 = 0xE8; break; 633 case BPF_ARSH: b3 = 0xF8; break; 634 } 635 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32); 636 break; 637 638 case BPF_ALU | BPF_LSH | BPF_X: 639 case BPF_ALU | BPF_RSH | BPF_X: 640 case BPF_ALU | BPF_ARSH | BPF_X: 641 case BPF_ALU64 | BPF_LSH | BPF_X: 642 case BPF_ALU64 | BPF_RSH | BPF_X: 643 case BPF_ALU64 | BPF_ARSH | BPF_X: 644 645 /* check for bad case when dst_reg == rcx */ 646 if (dst_reg == BPF_REG_4) { 647 /* mov r11, dst_reg */ 648 EMIT_mov(AUX_REG, dst_reg); 649 dst_reg = AUX_REG; 650 } 651 652 if (src_reg != BPF_REG_4) { /* common case */ 653 EMIT1(0x51); /* push rcx */ 654 655 /* mov rcx, src_reg */ 656 EMIT_mov(BPF_REG_4, src_reg); 657 } 658 659 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */ 660 if (BPF_CLASS(insn->code) == BPF_ALU64) 661 EMIT1(add_1mod(0x48, dst_reg)); 662 else if (is_ereg(dst_reg)) 663 EMIT1(add_1mod(0x40, dst_reg)); 664 665 switch (BPF_OP(insn->code)) { 666 case BPF_LSH: b3 = 0xE0; break; 667 case BPF_RSH: b3 = 0xE8; break; 668 case BPF_ARSH: b3 = 0xF8; break; 669 } 670 EMIT2(0xD3, add_1reg(b3, dst_reg)); 671 672 if (src_reg != BPF_REG_4) 673 EMIT1(0x59); /* pop rcx */ 674 675 if (insn->dst_reg == BPF_REG_4) 676 /* mov dst_reg, r11 */ 677 EMIT_mov(insn->dst_reg, AUX_REG); 678 break; 679 680 case BPF_ALU | BPF_END | BPF_FROM_BE: 681 switch (imm32) { 682 case 16: 683 /* emit 'ror %ax, 8' to swap lower 2 bytes */ 684 EMIT1(0x66); 685 if (is_ereg(dst_reg)) 686 EMIT1(0x41); 687 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8); 688 689 /* emit 'movzwl eax, ax' */ 690 if (is_ereg(dst_reg)) 691 EMIT3(0x45, 0x0F, 0xB7); 692 else 693 EMIT2(0x0F, 0xB7); 694 EMIT1(add_2reg(0xC0, dst_reg, dst_reg)); 695 break; 696 case 32: 697 /* emit 'bswap eax' to swap lower 4 bytes */ 698 if (is_ereg(dst_reg)) 699 EMIT2(0x41, 0x0F); 700 else 701 EMIT1(0x0F); 702 EMIT1(add_1reg(0xC8, dst_reg)); 703 break; 704 case 64: 705 /* emit 'bswap rax' to swap 8 bytes */ 706 EMIT3(add_1mod(0x48, dst_reg), 0x0F, 707 add_1reg(0xC8, dst_reg)); 708 break; 709 } 710 break; 711 712 case BPF_ALU | BPF_END | BPF_FROM_LE: 713 switch (imm32) { 714 case 16: 715 /* emit 'movzwl eax, ax' to zero extend 16-bit 716 * into 64 bit 717 */ 718 if (is_ereg(dst_reg)) 719 EMIT3(0x45, 0x0F, 0xB7); 720 else 721 EMIT2(0x0F, 0xB7); 722 EMIT1(add_2reg(0xC0, dst_reg, dst_reg)); 723 break; 724 case 32: 725 /* emit 'mov eax, eax' to clear upper 32-bits */ 726 if (is_ereg(dst_reg)) 727 EMIT1(0x45); 728 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg)); 729 break; 730 case 64: 731 /* nop */ 732 break; 733 } 734 break; 735 736 /* ST: *(u8*)(dst_reg + off) = imm */ 737 case BPF_ST | BPF_MEM | BPF_B: 738 if (is_ereg(dst_reg)) 739 EMIT2(0x41, 0xC6); 740 else 741 EMIT1(0xC6); 742 goto st; 743 case BPF_ST | BPF_MEM | BPF_H: 744 if (is_ereg(dst_reg)) 745 EMIT3(0x66, 0x41, 0xC7); 746 else 747 EMIT2(0x66, 0xC7); 748 goto st; 749 case BPF_ST | BPF_MEM | BPF_W: 750 if (is_ereg(dst_reg)) 751 EMIT2(0x41, 0xC7); 752 else 753 EMIT1(0xC7); 754 goto st; 755 case BPF_ST | BPF_MEM | BPF_DW: 756 EMIT2(add_1mod(0x48, dst_reg), 0xC7); 757 758 st: if (is_imm8(insn->off)) 759 EMIT2(add_1reg(0x40, dst_reg), insn->off); 760 else 761 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off); 762 763 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code))); 764 break; 765 766 /* STX: *(u8*)(dst_reg + off) = src_reg */ 767 case BPF_STX | BPF_MEM | BPF_B: 768 /* emit 'mov byte ptr [rax + off], al' */ 769 if (is_ereg(dst_reg) || is_ereg(src_reg) || 770 /* have to add extra byte for x86 SIL, DIL regs */ 771 src_reg == BPF_REG_1 || src_reg == BPF_REG_2) 772 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88); 773 else 774 EMIT1(0x88); 775 goto stx; 776 case BPF_STX | BPF_MEM | BPF_H: 777 if (is_ereg(dst_reg) || is_ereg(src_reg)) 778 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89); 779 else 780 EMIT2(0x66, 0x89); 781 goto stx; 782 case BPF_STX | BPF_MEM | BPF_W: 783 if (is_ereg(dst_reg) || is_ereg(src_reg)) 784 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89); 785 else 786 EMIT1(0x89); 787 goto stx; 788 case BPF_STX | BPF_MEM | BPF_DW: 789 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89); 790 stx: if (is_imm8(insn->off)) 791 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off); 792 else 793 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg), 794 insn->off); 795 break; 796 797 /* LDX: dst_reg = *(u8*)(src_reg + off) */ 798 case BPF_LDX | BPF_MEM | BPF_B: 799 /* emit 'movzx rax, byte ptr [rax + off]' */ 800 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6); 801 goto ldx; 802 case BPF_LDX | BPF_MEM | BPF_H: 803 /* emit 'movzx rax, word ptr [rax + off]' */ 804 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7); 805 goto ldx; 806 case BPF_LDX | BPF_MEM | BPF_W: 807 /* emit 'mov eax, dword ptr [rax+0x14]' */ 808 if (is_ereg(dst_reg) || is_ereg(src_reg)) 809 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B); 810 else 811 EMIT1(0x8B); 812 goto ldx; 813 case BPF_LDX | BPF_MEM | BPF_DW: 814 /* emit 'mov rax, qword ptr [rax+0x14]' */ 815 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B); 816 ldx: /* if insn->off == 0 we can save one extra byte, but 817 * special case of x86 r13 which always needs an offset 818 * is not worth the hassle 819 */ 820 if (is_imm8(insn->off)) 821 EMIT2(add_2reg(0x40, src_reg, dst_reg), insn->off); 822 else 823 EMIT1_off32(add_2reg(0x80, src_reg, dst_reg), 824 insn->off); 825 break; 826 827 /* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */ 828 case BPF_STX | BPF_XADD | BPF_W: 829 /* emit 'lock add dword ptr [rax + off], eax' */ 830 if (is_ereg(dst_reg) || is_ereg(src_reg)) 831 EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01); 832 else 833 EMIT2(0xF0, 0x01); 834 goto xadd; 835 case BPF_STX | BPF_XADD | BPF_DW: 836 EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01); 837 xadd: if (is_imm8(insn->off)) 838 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off); 839 else 840 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg), 841 insn->off); 842 break; 843 844 /* call */ 845 case BPF_JMP | BPF_CALL: 846 func = (u8 *) __bpf_call_base + imm32; 847 jmp_offset = func - (image + addrs[i]); 848 if (seen_ld_abs) { 849 reload_skb_data = bpf_helper_changes_pkt_data(func); 850 if (reload_skb_data) { 851 EMIT1(0x57); /* push %rdi */ 852 jmp_offset += 22; /* pop, mov, sub, mov */ 853 } else { 854 EMIT2(0x41, 0x52); /* push %r10 */ 855 EMIT2(0x41, 0x51); /* push %r9 */ 856 /* need to adjust jmp offset, since 857 * pop %r9, pop %r10 take 4 bytes after call insn 858 */ 859 jmp_offset += 4; 860 } 861 } 862 if (!imm32 || !is_simm32(jmp_offset)) { 863 pr_err("unsupported bpf func %d addr %p image %p\n", 864 imm32, func, image); 865 return -EINVAL; 866 } 867 EMIT1_off32(0xE8, jmp_offset); 868 if (seen_ld_abs) { 869 if (reload_skb_data) { 870 EMIT1(0x5F); /* pop %rdi */ 871 emit_load_skb_data_hlen(&prog); 872 } else { 873 EMIT2(0x41, 0x59); /* pop %r9 */ 874 EMIT2(0x41, 0x5A); /* pop %r10 */ 875 } 876 } 877 break; 878 879 case BPF_JMP | BPF_CALL | BPF_X: 880 emit_bpf_tail_call(&prog); 881 break; 882 883 /* cond jump */ 884 case BPF_JMP | BPF_JEQ | BPF_X: 885 case BPF_JMP | BPF_JNE | BPF_X: 886 case BPF_JMP | BPF_JGT | BPF_X: 887 case BPF_JMP | BPF_JGE | BPF_X: 888 case BPF_JMP | BPF_JSGT | BPF_X: 889 case BPF_JMP | BPF_JSGE | BPF_X: 890 /* cmp dst_reg, src_reg */ 891 EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x39, 892 add_2reg(0xC0, dst_reg, src_reg)); 893 goto emit_cond_jmp; 894 895 case BPF_JMP | BPF_JSET | BPF_X: 896 /* test dst_reg, src_reg */ 897 EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x85, 898 add_2reg(0xC0, dst_reg, src_reg)); 899 goto emit_cond_jmp; 900 901 case BPF_JMP | BPF_JSET | BPF_K: 902 /* test dst_reg, imm32 */ 903 EMIT1(add_1mod(0x48, dst_reg)); 904 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32); 905 goto emit_cond_jmp; 906 907 case BPF_JMP | BPF_JEQ | BPF_K: 908 case BPF_JMP | BPF_JNE | BPF_K: 909 case BPF_JMP | BPF_JGT | BPF_K: 910 case BPF_JMP | BPF_JGE | BPF_K: 911 case BPF_JMP | BPF_JSGT | BPF_K: 912 case BPF_JMP | BPF_JSGE | BPF_K: 913 /* cmp dst_reg, imm8/32 */ 914 EMIT1(add_1mod(0x48, dst_reg)); 915 916 if (is_imm8(imm32)) 917 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32); 918 else 919 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32); 920 921 emit_cond_jmp: /* convert BPF opcode to x86 */ 922 switch (BPF_OP(insn->code)) { 923 case BPF_JEQ: 924 jmp_cond = X86_JE; 925 break; 926 case BPF_JSET: 927 case BPF_JNE: 928 jmp_cond = X86_JNE; 929 break; 930 case BPF_JGT: 931 /* GT is unsigned '>', JA in x86 */ 932 jmp_cond = X86_JA; 933 break; 934 case BPF_JGE: 935 /* GE is unsigned '>=', JAE in x86 */ 936 jmp_cond = X86_JAE; 937 break; 938 case BPF_JSGT: 939 /* signed '>', GT in x86 */ 940 jmp_cond = X86_JG; 941 break; 942 case BPF_JSGE: 943 /* signed '>=', GE in x86 */ 944 jmp_cond = X86_JGE; 945 break; 946 default: /* to silence gcc warning */ 947 return -EFAULT; 948 } 949 jmp_offset = addrs[i + insn->off] - addrs[i]; 950 if (is_imm8(jmp_offset)) { 951 EMIT2(jmp_cond, jmp_offset); 952 } else if (is_simm32(jmp_offset)) { 953 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset); 954 } else { 955 pr_err("cond_jmp gen bug %llx\n", jmp_offset); 956 return -EFAULT; 957 } 958 959 break; 960 961 case BPF_JMP | BPF_JA: 962 jmp_offset = addrs[i + insn->off] - addrs[i]; 963 if (!jmp_offset) 964 /* optimize out nop jumps */ 965 break; 966 emit_jmp: 967 if (is_imm8(jmp_offset)) { 968 EMIT2(0xEB, jmp_offset); 969 } else if (is_simm32(jmp_offset)) { 970 EMIT1_off32(0xE9, jmp_offset); 971 } else { 972 pr_err("jmp gen bug %llx\n", jmp_offset); 973 return -EFAULT; 974 } 975 break; 976 977 case BPF_LD | BPF_IND | BPF_W: 978 func = sk_load_word; 979 goto common_load; 980 case BPF_LD | BPF_ABS | BPF_W: 981 func = CHOOSE_LOAD_FUNC(imm32, sk_load_word); 982 common_load: 983 ctx->seen_ld_abs = seen_ld_abs = true; 984 jmp_offset = func - (image + addrs[i]); 985 if (!func || !is_simm32(jmp_offset)) { 986 pr_err("unsupported bpf func %d addr %p image %p\n", 987 imm32, func, image); 988 return -EINVAL; 989 } 990 if (BPF_MODE(insn->code) == BPF_ABS) { 991 /* mov %esi, imm32 */ 992 EMIT1_off32(0xBE, imm32); 993 } else { 994 /* mov %rsi, src_reg */ 995 EMIT_mov(BPF_REG_2, src_reg); 996 if (imm32) { 997 if (is_imm8(imm32)) 998 /* add %esi, imm8 */ 999 EMIT3(0x83, 0xC6, imm32); 1000 else 1001 /* add %esi, imm32 */ 1002 EMIT2_off32(0x81, 0xC6, imm32); 1003 } 1004 } 1005 /* skb pointer is in R6 (%rbx), it will be copied into 1006 * %rdi if skb_copy_bits() call is necessary. 1007 * sk_load_* helpers also use %r10 and %r9d. 1008 * See bpf_jit.S 1009 */ 1010 if (seen_ax_reg) 1011 /* r10 = skb->data, mov %r10, off32(%rbx) */ 1012 EMIT3_off32(0x4c, 0x8b, 0x93, 1013 offsetof(struct sk_buff, data)); 1014 EMIT1_off32(0xE8, jmp_offset); /* call */ 1015 break; 1016 1017 case BPF_LD | BPF_IND | BPF_H: 1018 func = sk_load_half; 1019 goto common_load; 1020 case BPF_LD | BPF_ABS | BPF_H: 1021 func = CHOOSE_LOAD_FUNC(imm32, sk_load_half); 1022 goto common_load; 1023 case BPF_LD | BPF_IND | BPF_B: 1024 func = sk_load_byte; 1025 goto common_load; 1026 case BPF_LD | BPF_ABS | BPF_B: 1027 func = CHOOSE_LOAD_FUNC(imm32, sk_load_byte); 1028 goto common_load; 1029 1030 case BPF_JMP | BPF_EXIT: 1031 if (seen_exit) { 1032 jmp_offset = ctx->cleanup_addr - addrs[i]; 1033 goto emit_jmp; 1034 } 1035 seen_exit = true; 1036 /* update cleanup_addr */ 1037 ctx->cleanup_addr = proglen; 1038 /* mov rbx, qword ptr [rbp-X] */ 1039 EMIT3_off32(0x48, 0x8B, 0x9D, -STACKSIZE); 1040 /* mov r13, qword ptr [rbp-X] */ 1041 EMIT3_off32(0x4C, 0x8B, 0xAD, -STACKSIZE + 8); 1042 /* mov r14, qword ptr [rbp-X] */ 1043 EMIT3_off32(0x4C, 0x8B, 0xB5, -STACKSIZE + 16); 1044 /* mov r15, qword ptr [rbp-X] */ 1045 EMIT3_off32(0x4C, 0x8B, 0xBD, -STACKSIZE + 24); 1046 1047 EMIT1(0xC9); /* leave */ 1048 EMIT1(0xC3); /* ret */ 1049 break; 1050 1051 default: 1052 /* By design x64 JIT should support all BPF instructions 1053 * This error will be seen if new instruction was added 1054 * to interpreter, but not to JIT 1055 * or if there is junk in bpf_prog 1056 */ 1057 pr_err("bpf_jit: unknown opcode %02x\n", insn->code); 1058 return -EINVAL; 1059 } 1060 1061 ilen = prog - temp; 1062 if (ilen > BPF_MAX_INSN_SIZE) { 1063 pr_err("bpf_jit: fatal insn size error\n"); 1064 return -EFAULT; 1065 } 1066 1067 if (image) { 1068 if (unlikely(proglen + ilen > oldproglen)) { 1069 pr_err("bpf_jit: fatal error\n"); 1070 return -EFAULT; 1071 } 1072 memcpy(image + proglen, temp, ilen); 1073 } 1074 proglen += ilen; 1075 addrs[i] = proglen; 1076 prog = temp; 1077 } 1078 return proglen; 1079 } 1080 1081 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) 1082 { 1083 struct bpf_binary_header *header = NULL; 1084 struct bpf_prog *tmp, *orig_prog = prog; 1085 int proglen, oldproglen = 0; 1086 struct jit_context ctx = {}; 1087 bool tmp_blinded = false; 1088 u8 *image = NULL; 1089 int *addrs; 1090 int pass; 1091 int i; 1092 1093 if (!bpf_jit_enable) 1094 return orig_prog; 1095 1096 tmp = bpf_jit_blind_constants(prog); 1097 /* If blinding was requested and we failed during blinding, 1098 * we must fall back to the interpreter. 1099 */ 1100 if (IS_ERR(tmp)) 1101 return orig_prog; 1102 if (tmp != prog) { 1103 tmp_blinded = true; 1104 prog = tmp; 1105 } 1106 1107 addrs = kmalloc(prog->len * sizeof(*addrs), GFP_KERNEL); 1108 if (!addrs) { 1109 prog = orig_prog; 1110 goto out; 1111 } 1112 1113 /* Before first pass, make a rough estimation of addrs[] 1114 * each bpf instruction is translated to less than 64 bytes 1115 */ 1116 for (proglen = 0, i = 0; i < prog->len; i++) { 1117 proglen += 64; 1118 addrs[i] = proglen; 1119 } 1120 ctx.cleanup_addr = proglen; 1121 1122 /* JITed image shrinks with every pass and the loop iterates 1123 * until the image stops shrinking. Very large bpf programs 1124 * may converge on the last pass. In such case do one more 1125 * pass to emit the final image 1126 */ 1127 for (pass = 0; pass < 10 || image; pass++) { 1128 proglen = do_jit(prog, addrs, image, oldproglen, &ctx); 1129 if (proglen <= 0) { 1130 image = NULL; 1131 if (header) 1132 bpf_jit_binary_free(header); 1133 prog = orig_prog; 1134 goto out_addrs; 1135 } 1136 if (image) { 1137 if (proglen != oldproglen) { 1138 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n", 1139 proglen, oldproglen); 1140 prog = orig_prog; 1141 goto out_addrs; 1142 } 1143 break; 1144 } 1145 if (proglen == oldproglen) { 1146 header = bpf_jit_binary_alloc(proglen, &image, 1147 1, jit_fill_hole); 1148 if (!header) { 1149 prog = orig_prog; 1150 goto out_addrs; 1151 } 1152 } 1153 oldproglen = proglen; 1154 } 1155 1156 if (bpf_jit_enable > 1) 1157 bpf_jit_dump(prog->len, proglen, pass + 1, image); 1158 1159 if (image) { 1160 bpf_flush_icache(header, image + proglen); 1161 bpf_jit_binary_lock_ro(header); 1162 prog->bpf_func = (void *)image; 1163 prog->jited = 1; 1164 } else { 1165 prog = orig_prog; 1166 } 1167 1168 out_addrs: 1169 kfree(addrs); 1170 out: 1171 if (tmp_blinded) 1172 bpf_jit_prog_release_other(prog, prog == orig_prog ? 1173 tmp : orig_prog); 1174 return prog; 1175 } 1176