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 <asm/set_memory.h> 16 #include <linux/bpf.h> 17 18 int bpf_jit_enable __read_mostly; 19 20 /* 21 * assembly code in arch/x86/net/bpf_jit.S 22 */ 23 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[]; 24 extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[]; 25 extern u8 sk_load_byte_positive_offset[]; 26 extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[]; 27 extern u8 sk_load_byte_negative_offset[]; 28 29 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len) 30 { 31 if (len == 1) 32 *ptr = bytes; 33 else if (len == 2) 34 *(u16 *)ptr = bytes; 35 else { 36 *(u32 *)ptr = bytes; 37 barrier(); 38 } 39 return ptr + len; 40 } 41 42 #define EMIT(bytes, len) \ 43 do { prog = emit_code(prog, bytes, len); cnt += len; } while (0) 44 45 #define EMIT1(b1) EMIT(b1, 1) 46 #define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2) 47 #define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3) 48 #define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4) 49 #define EMIT1_off32(b1, off) \ 50 do {EMIT1(b1); EMIT(off, 4); } while (0) 51 #define EMIT2_off32(b1, b2, off) \ 52 do {EMIT2(b1, b2); EMIT(off, 4); } while (0) 53 #define EMIT3_off32(b1, b2, b3, off) \ 54 do {EMIT3(b1, b2, b3); EMIT(off, 4); } while (0) 55 #define EMIT4_off32(b1, b2, b3, b4, off) \ 56 do {EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0) 57 58 static bool is_imm8(int value) 59 { 60 return value <= 127 && value >= -128; 61 } 62 63 static bool is_simm32(s64 value) 64 { 65 return value == (s64) (s32) value; 66 } 67 68 /* mov dst, src */ 69 #define EMIT_mov(DST, SRC) \ 70 do {if (DST != SRC) \ 71 EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \ 72 } while (0) 73 74 static int bpf_size_to_x86_bytes(int bpf_size) 75 { 76 if (bpf_size == BPF_W) 77 return 4; 78 else if (bpf_size == BPF_H) 79 return 2; 80 else if (bpf_size == BPF_B) 81 return 1; 82 else if (bpf_size == BPF_DW) 83 return 4; /* imm32 */ 84 else 85 return 0; 86 } 87 88 /* list of x86 cond jumps opcodes (. + s8) 89 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32) 90 */ 91 #define X86_JB 0x72 92 #define X86_JAE 0x73 93 #define X86_JE 0x74 94 #define X86_JNE 0x75 95 #define X86_JBE 0x76 96 #define X86_JA 0x77 97 #define X86_JL 0x7C 98 #define X86_JGE 0x7D 99 #define X86_JLE 0x7E 100 #define X86_JG 0x7F 101 102 static void bpf_flush_icache(void *start, void *end) 103 { 104 mm_segment_t old_fs = get_fs(); 105 106 set_fs(KERNEL_DS); 107 smp_wmb(); 108 flush_icache_range((unsigned long)start, (unsigned long)end); 109 set_fs(old_fs); 110 } 111 112 #define CHOOSE_LOAD_FUNC(K, func) \ 113 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset) 114 115 /* pick a register outside of BPF range for JIT internal work */ 116 #define AUX_REG (MAX_BPF_JIT_REG + 1) 117 118 /* The following table maps BPF registers to x64 registers. 119 * 120 * x64 register r12 is unused, since if used as base address 121 * register in load/store instructions, it always needs an 122 * extra byte of encoding and is callee saved. 123 * 124 * r9 caches skb->len - skb->data_len 125 * r10 caches skb->data, and used for blinding (if enabled) 126 */ 127 static const int reg2hex[] = { 128 [BPF_REG_0] = 0, /* rax */ 129 [BPF_REG_1] = 7, /* rdi */ 130 [BPF_REG_2] = 6, /* rsi */ 131 [BPF_REG_3] = 2, /* rdx */ 132 [BPF_REG_4] = 1, /* rcx */ 133 [BPF_REG_5] = 0, /* r8 */ 134 [BPF_REG_6] = 3, /* rbx callee saved */ 135 [BPF_REG_7] = 5, /* r13 callee saved */ 136 [BPF_REG_8] = 6, /* r14 callee saved */ 137 [BPF_REG_9] = 7, /* r15 callee saved */ 138 [BPF_REG_FP] = 5, /* rbp readonly */ 139 [BPF_REG_AX] = 2, /* r10 temp register */ 140 [AUX_REG] = 3, /* r11 temp register */ 141 }; 142 143 /* is_ereg() == true if BPF register 'reg' maps to x64 r8..r15 144 * which need extra byte of encoding. 145 * rax,rcx,...,rbp have simpler encoding 146 */ 147 static bool is_ereg(u32 reg) 148 { 149 return (1 << reg) & (BIT(BPF_REG_5) | 150 BIT(AUX_REG) | 151 BIT(BPF_REG_7) | 152 BIT(BPF_REG_8) | 153 BIT(BPF_REG_9) | 154 BIT(BPF_REG_AX)); 155 } 156 157 /* add modifiers if 'reg' maps to x64 registers r8..r15 */ 158 static u8 add_1mod(u8 byte, u32 reg) 159 { 160 if (is_ereg(reg)) 161 byte |= 1; 162 return byte; 163 } 164 165 static u8 add_2mod(u8 byte, u32 r1, u32 r2) 166 { 167 if (is_ereg(r1)) 168 byte |= 1; 169 if (is_ereg(r2)) 170 byte |= 4; 171 return byte; 172 } 173 174 /* encode 'dst_reg' register into x64 opcode 'byte' */ 175 static u8 add_1reg(u8 byte, u32 dst_reg) 176 { 177 return byte + reg2hex[dst_reg]; 178 } 179 180 /* encode 'dst_reg' and 'src_reg' registers into x64 opcode 'byte' */ 181 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg) 182 { 183 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3); 184 } 185 186 static void jit_fill_hole(void *area, unsigned int size) 187 { 188 /* fill whole space with int3 instructions */ 189 memset(area, 0xcc, size); 190 } 191 192 struct jit_context { 193 int cleanup_addr; /* epilogue code offset */ 194 bool seen_ld_abs; 195 bool seen_ax_reg; 196 }; 197 198 /* maximum number of bytes emitted while JITing one eBPF insn */ 199 #define BPF_MAX_INSN_SIZE 128 200 #define BPF_INSN_SAFETY 64 201 202 #define AUX_STACK_SPACE \ 203 (32 /* space for rbx, r13, r14, r15 */ + \ 204 8 /* space for skb_copy_bits() buffer */) 205 206 #define PROLOGUE_SIZE 37 207 208 /* emit x64 prologue code for BPF program and check it's size. 209 * bpf_tail_call helper will skip it while jumping into another program 210 */ 211 static void emit_prologue(u8 **pprog, u32 stack_depth) 212 { 213 u8 *prog = *pprog; 214 int cnt = 0; 215 216 EMIT1(0x55); /* push rbp */ 217 EMIT3(0x48, 0x89, 0xE5); /* mov rbp,rsp */ 218 219 /* sub rsp, rounded_stack_depth + AUX_STACK_SPACE */ 220 EMIT3_off32(0x48, 0x81, 0xEC, 221 round_up(stack_depth, 8) + AUX_STACK_SPACE); 222 223 /* sub rbp, AUX_STACK_SPACE */ 224 EMIT4(0x48, 0x83, 0xED, AUX_STACK_SPACE); 225 226 /* all classic BPF filters use R6(rbx) save it */ 227 228 /* mov qword ptr [rbp+0],rbx */ 229 EMIT4(0x48, 0x89, 0x5D, 0); 230 231 /* bpf_convert_filter() maps classic BPF register X to R7 and uses R8 232 * as temporary, so all tcpdump filters need to spill/fill R7(r13) and 233 * R8(r14). R9(r15) spill could be made conditional, but there is only 234 * one 'bpf_error' return path out of helper functions inside bpf_jit.S 235 * The overhead of extra spill is negligible for any filter other 236 * than synthetic ones. Therefore not worth adding complexity. 237 */ 238 239 /* mov qword ptr [rbp+8],r13 */ 240 EMIT4(0x4C, 0x89, 0x6D, 8); 241 /* mov qword ptr [rbp+16],r14 */ 242 EMIT4(0x4C, 0x89, 0x75, 16); 243 /* mov qword ptr [rbp+24],r15 */ 244 EMIT4(0x4C, 0x89, 0x7D, 24); 245 246 /* Clear the tail call counter (tail_call_cnt): for eBPF tail calls 247 * we need to reset the counter to 0. It's done in two instructions, 248 * resetting rax register to 0 (xor on eax gets 0 extended), and 249 * moving it to the counter location. 250 */ 251 252 /* xor eax, eax */ 253 EMIT2(0x31, 0xc0); 254 /* mov qword ptr [rbp+32], rax */ 255 EMIT4(0x48, 0x89, 0x45, 32); 256 257 BUILD_BUG_ON(cnt != PROLOGUE_SIZE); 258 *pprog = prog; 259 } 260 261 /* generate the following code: 262 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ... 263 * if (index >= array->map.max_entries) 264 * goto out; 265 * if (++tail_call_cnt > MAX_TAIL_CALL_CNT) 266 * goto out; 267 * prog = array->ptrs[index]; 268 * if (prog == NULL) 269 * goto out; 270 * goto *(prog->bpf_func + prologue_size); 271 * out: 272 */ 273 static void emit_bpf_tail_call(u8 **pprog) 274 { 275 u8 *prog = *pprog; 276 int label1, label2, label3; 277 int cnt = 0; 278 279 /* rdi - pointer to ctx 280 * rsi - pointer to bpf_array 281 * rdx - index in bpf_array 282 */ 283 284 /* if (index >= array->map.max_entries) 285 * goto out; 286 */ 287 EMIT4(0x48, 0x8B, 0x46, /* mov rax, qword ptr [rsi + 16] */ 288 offsetof(struct bpf_array, map.max_entries)); 289 EMIT3(0x48, 0x39, 0xD0); /* cmp rax, rdx */ 290 #define OFFSET1 47 /* number of bytes to jump */ 291 EMIT2(X86_JBE, OFFSET1); /* jbe out */ 292 label1 = cnt; 293 294 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT) 295 * goto out; 296 */ 297 EMIT2_off32(0x8B, 0x85, 36); /* mov eax, dword ptr [rbp + 36] */ 298 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */ 299 #define OFFSET2 36 300 EMIT2(X86_JA, OFFSET2); /* ja out */ 301 label2 = cnt; 302 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */ 303 EMIT2_off32(0x89, 0x85, 36); /* mov dword ptr [rbp + 36], eax */ 304 305 /* prog = array->ptrs[index]; */ 306 EMIT4_off32(0x48, 0x8D, 0x84, 0xD6, /* lea rax, [rsi + rdx * 8 + offsetof(...)] */ 307 offsetof(struct bpf_array, ptrs)); 308 EMIT3(0x48, 0x8B, 0x00); /* mov rax, qword ptr [rax] */ 309 310 /* if (prog == NULL) 311 * goto out; 312 */ 313 EMIT4(0x48, 0x83, 0xF8, 0x00); /* cmp rax, 0 */ 314 #define OFFSET3 10 315 EMIT2(X86_JE, OFFSET3); /* je out */ 316 label3 = cnt; 317 318 /* goto *(prog->bpf_func + prologue_size); */ 319 EMIT4(0x48, 0x8B, 0x40, /* mov rax, qword ptr [rax + 32] */ 320 offsetof(struct bpf_prog, bpf_func)); 321 EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE); /* add rax, prologue_size */ 322 323 /* now we're ready to jump into next BPF program 324 * rdi == ctx (1st arg) 325 * rax == prog->bpf_func + prologue_size 326 */ 327 EMIT2(0xFF, 0xE0); /* jmp rax */ 328 329 /* out: */ 330 BUILD_BUG_ON(cnt - label1 != OFFSET1); 331 BUILD_BUG_ON(cnt - label2 != OFFSET2); 332 BUILD_BUG_ON(cnt - label3 != OFFSET3); 333 *pprog = prog; 334 } 335 336 337 static void emit_load_skb_data_hlen(u8 **pprog) 338 { 339 u8 *prog = *pprog; 340 int cnt = 0; 341 342 /* r9d = skb->len - skb->data_len (headlen) 343 * r10 = skb->data 344 */ 345 /* mov %r9d, off32(%rdi) */ 346 EMIT3_off32(0x44, 0x8b, 0x8f, offsetof(struct sk_buff, len)); 347 348 /* sub %r9d, off32(%rdi) */ 349 EMIT3_off32(0x44, 0x2b, 0x8f, offsetof(struct sk_buff, data_len)); 350 351 /* mov %r10, off32(%rdi) */ 352 EMIT3_off32(0x4c, 0x8b, 0x97, offsetof(struct sk_buff, data)); 353 *pprog = prog; 354 } 355 356 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, 357 int oldproglen, struct jit_context *ctx) 358 { 359 struct bpf_insn *insn = bpf_prog->insnsi; 360 int insn_cnt = bpf_prog->len; 361 bool seen_ld_abs = ctx->seen_ld_abs | (oldproglen == 0); 362 bool seen_ax_reg = ctx->seen_ax_reg | (oldproglen == 0); 363 bool seen_exit = false; 364 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY]; 365 int i, cnt = 0; 366 int proglen = 0; 367 u8 *prog = temp; 368 369 emit_prologue(&prog, bpf_prog->aux->stack_depth); 370 371 if (seen_ld_abs) 372 emit_load_skb_data_hlen(&prog); 373 374 for (i = 0; i < insn_cnt; i++, insn++) { 375 const s32 imm32 = insn->imm; 376 u32 dst_reg = insn->dst_reg; 377 u32 src_reg = insn->src_reg; 378 u8 b1 = 0, b2 = 0, b3 = 0; 379 s64 jmp_offset; 380 u8 jmp_cond; 381 bool reload_skb_data; 382 int ilen; 383 u8 *func; 384 385 if (dst_reg == BPF_REG_AX || src_reg == BPF_REG_AX) 386 ctx->seen_ax_reg = seen_ax_reg = true; 387 388 switch (insn->code) { 389 /* ALU */ 390 case BPF_ALU | BPF_ADD | BPF_X: 391 case BPF_ALU | BPF_SUB | BPF_X: 392 case BPF_ALU | BPF_AND | BPF_X: 393 case BPF_ALU | BPF_OR | BPF_X: 394 case BPF_ALU | BPF_XOR | BPF_X: 395 case BPF_ALU64 | BPF_ADD | BPF_X: 396 case BPF_ALU64 | BPF_SUB | BPF_X: 397 case BPF_ALU64 | BPF_AND | BPF_X: 398 case BPF_ALU64 | BPF_OR | BPF_X: 399 case BPF_ALU64 | BPF_XOR | BPF_X: 400 switch (BPF_OP(insn->code)) { 401 case BPF_ADD: b2 = 0x01; break; 402 case BPF_SUB: b2 = 0x29; break; 403 case BPF_AND: b2 = 0x21; break; 404 case BPF_OR: b2 = 0x09; break; 405 case BPF_XOR: b2 = 0x31; break; 406 } 407 if (BPF_CLASS(insn->code) == BPF_ALU64) 408 EMIT1(add_2mod(0x48, dst_reg, src_reg)); 409 else if (is_ereg(dst_reg) || is_ereg(src_reg)) 410 EMIT1(add_2mod(0x40, dst_reg, src_reg)); 411 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg)); 412 break; 413 414 /* mov dst, src */ 415 case BPF_ALU64 | BPF_MOV | BPF_X: 416 EMIT_mov(dst_reg, src_reg); 417 break; 418 419 /* mov32 dst, src */ 420 case BPF_ALU | BPF_MOV | BPF_X: 421 if (is_ereg(dst_reg) || is_ereg(src_reg)) 422 EMIT1(add_2mod(0x40, dst_reg, src_reg)); 423 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg)); 424 break; 425 426 /* neg dst */ 427 case BPF_ALU | BPF_NEG: 428 case BPF_ALU64 | BPF_NEG: 429 if (BPF_CLASS(insn->code) == BPF_ALU64) 430 EMIT1(add_1mod(0x48, dst_reg)); 431 else if (is_ereg(dst_reg)) 432 EMIT1(add_1mod(0x40, dst_reg)); 433 EMIT2(0xF7, add_1reg(0xD8, dst_reg)); 434 break; 435 436 case BPF_ALU | BPF_ADD | BPF_K: 437 case BPF_ALU | BPF_SUB | BPF_K: 438 case BPF_ALU | BPF_AND | BPF_K: 439 case BPF_ALU | BPF_OR | BPF_K: 440 case BPF_ALU | BPF_XOR | BPF_K: 441 case BPF_ALU64 | BPF_ADD | BPF_K: 442 case BPF_ALU64 | BPF_SUB | BPF_K: 443 case BPF_ALU64 | BPF_AND | BPF_K: 444 case BPF_ALU64 | BPF_OR | BPF_K: 445 case BPF_ALU64 | BPF_XOR | BPF_K: 446 if (BPF_CLASS(insn->code) == BPF_ALU64) 447 EMIT1(add_1mod(0x48, dst_reg)); 448 else if (is_ereg(dst_reg)) 449 EMIT1(add_1mod(0x40, dst_reg)); 450 451 switch (BPF_OP(insn->code)) { 452 case BPF_ADD: b3 = 0xC0; break; 453 case BPF_SUB: b3 = 0xE8; break; 454 case BPF_AND: b3 = 0xE0; break; 455 case BPF_OR: b3 = 0xC8; break; 456 case BPF_XOR: b3 = 0xF0; break; 457 } 458 459 if (is_imm8(imm32)) 460 EMIT3(0x83, add_1reg(b3, dst_reg), imm32); 461 else 462 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32); 463 break; 464 465 case BPF_ALU64 | BPF_MOV | BPF_K: 466 /* optimization: if imm32 is positive, 467 * use 'mov eax, imm32' (which zero-extends imm32) 468 * to save 2 bytes 469 */ 470 if (imm32 < 0) { 471 /* 'mov rax, imm32' sign extends imm32 */ 472 b1 = add_1mod(0x48, dst_reg); 473 b2 = 0xC7; 474 b3 = 0xC0; 475 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32); 476 break; 477 } 478 479 case BPF_ALU | BPF_MOV | BPF_K: 480 /* optimization: if imm32 is zero, use 'xor <dst>,<dst>' 481 * to save 3 bytes. 482 */ 483 if (imm32 == 0) { 484 if (is_ereg(dst_reg)) 485 EMIT1(add_2mod(0x40, dst_reg, dst_reg)); 486 b2 = 0x31; /* xor */ 487 b3 = 0xC0; 488 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg)); 489 break; 490 } 491 492 /* mov %eax, imm32 */ 493 if (is_ereg(dst_reg)) 494 EMIT1(add_1mod(0x40, dst_reg)); 495 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32); 496 break; 497 498 case BPF_LD | BPF_IMM | BPF_DW: 499 /* optimization: if imm64 is zero, use 'xor <dst>,<dst>' 500 * to save 7 bytes. 501 */ 502 if (insn[0].imm == 0 && insn[1].imm == 0) { 503 b1 = add_2mod(0x48, dst_reg, dst_reg); 504 b2 = 0x31; /* xor */ 505 b3 = 0xC0; 506 EMIT3(b1, b2, add_2reg(b3, dst_reg, dst_reg)); 507 508 insn++; 509 i++; 510 break; 511 } 512 513 /* movabsq %rax, imm64 */ 514 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg)); 515 EMIT(insn[0].imm, 4); 516 EMIT(insn[1].imm, 4); 517 518 insn++; 519 i++; 520 break; 521 522 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */ 523 case BPF_ALU | BPF_MOD | BPF_X: 524 case BPF_ALU | BPF_DIV | BPF_X: 525 case BPF_ALU | BPF_MOD | BPF_K: 526 case BPF_ALU | BPF_DIV | BPF_K: 527 case BPF_ALU64 | BPF_MOD | BPF_X: 528 case BPF_ALU64 | BPF_DIV | BPF_X: 529 case BPF_ALU64 | BPF_MOD | BPF_K: 530 case BPF_ALU64 | BPF_DIV | BPF_K: 531 EMIT1(0x50); /* push rax */ 532 EMIT1(0x52); /* push rdx */ 533 534 if (BPF_SRC(insn->code) == BPF_X) 535 /* mov r11, src_reg */ 536 EMIT_mov(AUX_REG, src_reg); 537 else 538 /* mov r11, imm32 */ 539 EMIT3_off32(0x49, 0xC7, 0xC3, imm32); 540 541 /* mov rax, dst_reg */ 542 EMIT_mov(BPF_REG_0, dst_reg); 543 544 /* xor edx, edx 545 * equivalent to 'xor rdx, rdx', but one byte less 546 */ 547 EMIT2(0x31, 0xd2); 548 549 if (BPF_SRC(insn->code) == BPF_X) { 550 /* if (src_reg == 0) return 0 */ 551 552 /* cmp r11, 0 */ 553 EMIT4(0x49, 0x83, 0xFB, 0x00); 554 555 /* jne .+9 (skip over pop, pop, xor and jmp) */ 556 EMIT2(X86_JNE, 1 + 1 + 2 + 5); 557 EMIT1(0x5A); /* pop rdx */ 558 EMIT1(0x58); /* pop rax */ 559 EMIT2(0x31, 0xc0); /* xor eax, eax */ 560 561 /* jmp cleanup_addr 562 * addrs[i] - 11, because there are 11 bytes 563 * after this insn: div, mov, pop, pop, mov 564 */ 565 jmp_offset = ctx->cleanup_addr - (addrs[i] - 11); 566 EMIT1_off32(0xE9, jmp_offset); 567 } 568 569 if (BPF_CLASS(insn->code) == BPF_ALU64) 570 /* div r11 */ 571 EMIT3(0x49, 0xF7, 0xF3); 572 else 573 /* div r11d */ 574 EMIT3(0x41, 0xF7, 0xF3); 575 576 if (BPF_OP(insn->code) == BPF_MOD) 577 /* mov r11, rdx */ 578 EMIT3(0x49, 0x89, 0xD3); 579 else 580 /* mov r11, rax */ 581 EMIT3(0x49, 0x89, 0xC3); 582 583 EMIT1(0x5A); /* pop rdx */ 584 EMIT1(0x58); /* pop rax */ 585 586 /* mov dst_reg, r11 */ 587 EMIT_mov(dst_reg, AUX_REG); 588 break; 589 590 case BPF_ALU | BPF_MUL | BPF_K: 591 case BPF_ALU | BPF_MUL | BPF_X: 592 case BPF_ALU64 | BPF_MUL | BPF_K: 593 case BPF_ALU64 | BPF_MUL | BPF_X: 594 EMIT1(0x50); /* push rax */ 595 EMIT1(0x52); /* push rdx */ 596 597 /* mov r11, dst_reg */ 598 EMIT_mov(AUX_REG, dst_reg); 599 600 if (BPF_SRC(insn->code) == BPF_X) 601 /* mov rax, src_reg */ 602 EMIT_mov(BPF_REG_0, src_reg); 603 else 604 /* mov rax, imm32 */ 605 EMIT3_off32(0x48, 0xC7, 0xC0, imm32); 606 607 if (BPF_CLASS(insn->code) == BPF_ALU64) 608 EMIT1(add_1mod(0x48, AUX_REG)); 609 else if (is_ereg(AUX_REG)) 610 EMIT1(add_1mod(0x40, AUX_REG)); 611 /* mul(q) r11 */ 612 EMIT2(0xF7, add_1reg(0xE0, AUX_REG)); 613 614 /* mov r11, rax */ 615 EMIT_mov(AUX_REG, BPF_REG_0); 616 617 EMIT1(0x5A); /* pop rdx */ 618 EMIT1(0x58); /* pop rax */ 619 620 /* mov dst_reg, r11 */ 621 EMIT_mov(dst_reg, AUX_REG); 622 break; 623 624 /* shifts */ 625 case BPF_ALU | BPF_LSH | BPF_K: 626 case BPF_ALU | BPF_RSH | BPF_K: 627 case BPF_ALU | BPF_ARSH | BPF_K: 628 case BPF_ALU64 | BPF_LSH | BPF_K: 629 case BPF_ALU64 | BPF_RSH | BPF_K: 630 case BPF_ALU64 | BPF_ARSH | BPF_K: 631 if (BPF_CLASS(insn->code) == BPF_ALU64) 632 EMIT1(add_1mod(0x48, dst_reg)); 633 else if (is_ereg(dst_reg)) 634 EMIT1(add_1mod(0x40, dst_reg)); 635 636 switch (BPF_OP(insn->code)) { 637 case BPF_LSH: b3 = 0xE0; break; 638 case BPF_RSH: b3 = 0xE8; break; 639 case BPF_ARSH: b3 = 0xF8; break; 640 } 641 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32); 642 break; 643 644 case BPF_ALU | BPF_LSH | BPF_X: 645 case BPF_ALU | BPF_RSH | BPF_X: 646 case BPF_ALU | BPF_ARSH | BPF_X: 647 case BPF_ALU64 | BPF_LSH | BPF_X: 648 case BPF_ALU64 | BPF_RSH | BPF_X: 649 case BPF_ALU64 | BPF_ARSH | BPF_X: 650 651 /* check for bad case when dst_reg == rcx */ 652 if (dst_reg == BPF_REG_4) { 653 /* mov r11, dst_reg */ 654 EMIT_mov(AUX_REG, dst_reg); 655 dst_reg = AUX_REG; 656 } 657 658 if (src_reg != BPF_REG_4) { /* common case */ 659 EMIT1(0x51); /* push rcx */ 660 661 /* mov rcx, src_reg */ 662 EMIT_mov(BPF_REG_4, src_reg); 663 } 664 665 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */ 666 if (BPF_CLASS(insn->code) == BPF_ALU64) 667 EMIT1(add_1mod(0x48, dst_reg)); 668 else if (is_ereg(dst_reg)) 669 EMIT1(add_1mod(0x40, dst_reg)); 670 671 switch (BPF_OP(insn->code)) { 672 case BPF_LSH: b3 = 0xE0; break; 673 case BPF_RSH: b3 = 0xE8; break; 674 case BPF_ARSH: b3 = 0xF8; break; 675 } 676 EMIT2(0xD3, add_1reg(b3, dst_reg)); 677 678 if (src_reg != BPF_REG_4) 679 EMIT1(0x59); /* pop rcx */ 680 681 if (insn->dst_reg == BPF_REG_4) 682 /* mov dst_reg, r11 */ 683 EMIT_mov(insn->dst_reg, AUX_REG); 684 break; 685 686 case BPF_ALU | BPF_END | BPF_FROM_BE: 687 switch (imm32) { 688 case 16: 689 /* emit 'ror %ax, 8' to swap lower 2 bytes */ 690 EMIT1(0x66); 691 if (is_ereg(dst_reg)) 692 EMIT1(0x41); 693 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8); 694 695 /* emit 'movzwl eax, ax' */ 696 if (is_ereg(dst_reg)) 697 EMIT3(0x45, 0x0F, 0xB7); 698 else 699 EMIT2(0x0F, 0xB7); 700 EMIT1(add_2reg(0xC0, dst_reg, dst_reg)); 701 break; 702 case 32: 703 /* emit 'bswap eax' to swap lower 4 bytes */ 704 if (is_ereg(dst_reg)) 705 EMIT2(0x41, 0x0F); 706 else 707 EMIT1(0x0F); 708 EMIT1(add_1reg(0xC8, dst_reg)); 709 break; 710 case 64: 711 /* emit 'bswap rax' to swap 8 bytes */ 712 EMIT3(add_1mod(0x48, dst_reg), 0x0F, 713 add_1reg(0xC8, dst_reg)); 714 break; 715 } 716 break; 717 718 case BPF_ALU | BPF_END | BPF_FROM_LE: 719 switch (imm32) { 720 case 16: 721 /* emit 'movzwl eax, ax' to zero extend 16-bit 722 * into 64 bit 723 */ 724 if (is_ereg(dst_reg)) 725 EMIT3(0x45, 0x0F, 0xB7); 726 else 727 EMIT2(0x0F, 0xB7); 728 EMIT1(add_2reg(0xC0, dst_reg, dst_reg)); 729 break; 730 case 32: 731 /* emit 'mov eax, eax' to clear upper 32-bits */ 732 if (is_ereg(dst_reg)) 733 EMIT1(0x45); 734 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg)); 735 break; 736 case 64: 737 /* nop */ 738 break; 739 } 740 break; 741 742 /* ST: *(u8*)(dst_reg + off) = imm */ 743 case BPF_ST | BPF_MEM | BPF_B: 744 if (is_ereg(dst_reg)) 745 EMIT2(0x41, 0xC6); 746 else 747 EMIT1(0xC6); 748 goto st; 749 case BPF_ST | BPF_MEM | BPF_H: 750 if (is_ereg(dst_reg)) 751 EMIT3(0x66, 0x41, 0xC7); 752 else 753 EMIT2(0x66, 0xC7); 754 goto st; 755 case BPF_ST | BPF_MEM | BPF_W: 756 if (is_ereg(dst_reg)) 757 EMIT2(0x41, 0xC7); 758 else 759 EMIT1(0xC7); 760 goto st; 761 case BPF_ST | BPF_MEM | BPF_DW: 762 EMIT2(add_1mod(0x48, dst_reg), 0xC7); 763 764 st: if (is_imm8(insn->off)) 765 EMIT2(add_1reg(0x40, dst_reg), insn->off); 766 else 767 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off); 768 769 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code))); 770 break; 771 772 /* STX: *(u8*)(dst_reg + off) = src_reg */ 773 case BPF_STX | BPF_MEM | BPF_B: 774 /* emit 'mov byte ptr [rax + off], al' */ 775 if (is_ereg(dst_reg) || is_ereg(src_reg) || 776 /* have to add extra byte for x86 SIL, DIL regs */ 777 src_reg == BPF_REG_1 || src_reg == BPF_REG_2) 778 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88); 779 else 780 EMIT1(0x88); 781 goto stx; 782 case BPF_STX | BPF_MEM | BPF_H: 783 if (is_ereg(dst_reg) || is_ereg(src_reg)) 784 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89); 785 else 786 EMIT2(0x66, 0x89); 787 goto stx; 788 case BPF_STX | BPF_MEM | BPF_W: 789 if (is_ereg(dst_reg) || is_ereg(src_reg)) 790 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89); 791 else 792 EMIT1(0x89); 793 goto stx; 794 case BPF_STX | BPF_MEM | BPF_DW: 795 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89); 796 stx: if (is_imm8(insn->off)) 797 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off); 798 else 799 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg), 800 insn->off); 801 break; 802 803 /* LDX: dst_reg = *(u8*)(src_reg + off) */ 804 case BPF_LDX | BPF_MEM | BPF_B: 805 /* emit 'movzx rax, byte ptr [rax + off]' */ 806 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6); 807 goto ldx; 808 case BPF_LDX | BPF_MEM | BPF_H: 809 /* emit 'movzx rax, word ptr [rax + off]' */ 810 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7); 811 goto ldx; 812 case BPF_LDX | BPF_MEM | BPF_W: 813 /* emit 'mov eax, dword ptr [rax+0x14]' */ 814 if (is_ereg(dst_reg) || is_ereg(src_reg)) 815 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B); 816 else 817 EMIT1(0x8B); 818 goto ldx; 819 case BPF_LDX | BPF_MEM | BPF_DW: 820 /* emit 'mov rax, qword ptr [rax+0x14]' */ 821 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B); 822 ldx: /* if insn->off == 0 we can save one extra byte, but 823 * special case of x86 r13 which always needs an offset 824 * is not worth the hassle 825 */ 826 if (is_imm8(insn->off)) 827 EMIT2(add_2reg(0x40, src_reg, dst_reg), insn->off); 828 else 829 EMIT1_off32(add_2reg(0x80, src_reg, dst_reg), 830 insn->off); 831 break; 832 833 /* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */ 834 case BPF_STX | BPF_XADD | BPF_W: 835 /* emit 'lock add dword ptr [rax + off], eax' */ 836 if (is_ereg(dst_reg) || is_ereg(src_reg)) 837 EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01); 838 else 839 EMIT2(0xF0, 0x01); 840 goto xadd; 841 case BPF_STX | BPF_XADD | BPF_DW: 842 EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01); 843 xadd: if (is_imm8(insn->off)) 844 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off); 845 else 846 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg), 847 insn->off); 848 break; 849 850 /* call */ 851 case BPF_JMP | BPF_CALL: 852 func = (u8 *) __bpf_call_base + imm32; 853 jmp_offset = func - (image + addrs[i]); 854 if (seen_ld_abs) { 855 reload_skb_data = bpf_helper_changes_pkt_data(func); 856 if (reload_skb_data) { 857 EMIT1(0x57); /* push %rdi */ 858 jmp_offset += 22; /* pop, mov, sub, mov */ 859 } else { 860 EMIT2(0x41, 0x52); /* push %r10 */ 861 EMIT2(0x41, 0x51); /* push %r9 */ 862 /* need to adjust jmp offset, since 863 * pop %r9, pop %r10 take 4 bytes after call insn 864 */ 865 jmp_offset += 4; 866 } 867 } 868 if (!imm32 || !is_simm32(jmp_offset)) { 869 pr_err("unsupported bpf func %d addr %p image %p\n", 870 imm32, func, image); 871 return -EINVAL; 872 } 873 EMIT1_off32(0xE8, jmp_offset); 874 if (seen_ld_abs) { 875 if (reload_skb_data) { 876 EMIT1(0x5F); /* pop %rdi */ 877 emit_load_skb_data_hlen(&prog); 878 } else { 879 EMIT2(0x41, 0x59); /* pop %r9 */ 880 EMIT2(0x41, 0x5A); /* pop %r10 */ 881 } 882 } 883 break; 884 885 case BPF_JMP | BPF_TAIL_CALL: 886 emit_bpf_tail_call(&prog); 887 break; 888 889 /* cond jump */ 890 case BPF_JMP | BPF_JEQ | BPF_X: 891 case BPF_JMP | BPF_JNE | BPF_X: 892 case BPF_JMP | BPF_JGT | BPF_X: 893 case BPF_JMP | BPF_JLT | BPF_X: 894 case BPF_JMP | BPF_JGE | BPF_X: 895 case BPF_JMP | BPF_JLE | BPF_X: 896 case BPF_JMP | BPF_JSGT | BPF_X: 897 case BPF_JMP | BPF_JSLT | BPF_X: 898 case BPF_JMP | BPF_JSGE | BPF_X: 899 case BPF_JMP | BPF_JSLE | BPF_X: 900 /* cmp dst_reg, src_reg */ 901 EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x39, 902 add_2reg(0xC0, dst_reg, src_reg)); 903 goto emit_cond_jmp; 904 905 case BPF_JMP | BPF_JSET | BPF_X: 906 /* test dst_reg, src_reg */ 907 EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x85, 908 add_2reg(0xC0, dst_reg, src_reg)); 909 goto emit_cond_jmp; 910 911 case BPF_JMP | BPF_JSET | BPF_K: 912 /* test dst_reg, imm32 */ 913 EMIT1(add_1mod(0x48, dst_reg)); 914 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32); 915 goto emit_cond_jmp; 916 917 case BPF_JMP | BPF_JEQ | BPF_K: 918 case BPF_JMP | BPF_JNE | BPF_K: 919 case BPF_JMP | BPF_JGT | BPF_K: 920 case BPF_JMP | BPF_JLT | BPF_K: 921 case BPF_JMP | BPF_JGE | BPF_K: 922 case BPF_JMP | BPF_JLE | BPF_K: 923 case BPF_JMP | BPF_JSGT | BPF_K: 924 case BPF_JMP | BPF_JSLT | BPF_K: 925 case BPF_JMP | BPF_JSGE | BPF_K: 926 case BPF_JMP | BPF_JSLE | BPF_K: 927 /* cmp dst_reg, imm8/32 */ 928 EMIT1(add_1mod(0x48, dst_reg)); 929 930 if (is_imm8(imm32)) 931 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32); 932 else 933 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32); 934 935 emit_cond_jmp: /* convert BPF opcode to x86 */ 936 switch (BPF_OP(insn->code)) { 937 case BPF_JEQ: 938 jmp_cond = X86_JE; 939 break; 940 case BPF_JSET: 941 case BPF_JNE: 942 jmp_cond = X86_JNE; 943 break; 944 case BPF_JGT: 945 /* GT is unsigned '>', JA in x86 */ 946 jmp_cond = X86_JA; 947 break; 948 case BPF_JLT: 949 /* LT is unsigned '<', JB in x86 */ 950 jmp_cond = X86_JB; 951 break; 952 case BPF_JGE: 953 /* GE is unsigned '>=', JAE in x86 */ 954 jmp_cond = X86_JAE; 955 break; 956 case BPF_JLE: 957 /* LE is unsigned '<=', JBE in x86 */ 958 jmp_cond = X86_JBE; 959 break; 960 case BPF_JSGT: 961 /* signed '>', GT in x86 */ 962 jmp_cond = X86_JG; 963 break; 964 case BPF_JSLT: 965 /* signed '<', LT in x86 */ 966 jmp_cond = X86_JL; 967 break; 968 case BPF_JSGE: 969 /* signed '>=', GE in x86 */ 970 jmp_cond = X86_JGE; 971 break; 972 case BPF_JSLE: 973 /* signed '<=', LE in x86 */ 974 jmp_cond = X86_JLE; 975 break; 976 default: /* to silence gcc warning */ 977 return -EFAULT; 978 } 979 jmp_offset = addrs[i + insn->off] - addrs[i]; 980 if (is_imm8(jmp_offset)) { 981 EMIT2(jmp_cond, jmp_offset); 982 } else if (is_simm32(jmp_offset)) { 983 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset); 984 } else { 985 pr_err("cond_jmp gen bug %llx\n", jmp_offset); 986 return -EFAULT; 987 } 988 989 break; 990 991 case BPF_JMP | BPF_JA: 992 jmp_offset = addrs[i + insn->off] - addrs[i]; 993 if (!jmp_offset) 994 /* optimize out nop jumps */ 995 break; 996 emit_jmp: 997 if (is_imm8(jmp_offset)) { 998 EMIT2(0xEB, jmp_offset); 999 } else if (is_simm32(jmp_offset)) { 1000 EMIT1_off32(0xE9, jmp_offset); 1001 } else { 1002 pr_err("jmp gen bug %llx\n", jmp_offset); 1003 return -EFAULT; 1004 } 1005 break; 1006 1007 case BPF_LD | BPF_IND | BPF_W: 1008 func = sk_load_word; 1009 goto common_load; 1010 case BPF_LD | BPF_ABS | BPF_W: 1011 func = CHOOSE_LOAD_FUNC(imm32, sk_load_word); 1012 common_load: 1013 ctx->seen_ld_abs = seen_ld_abs = true; 1014 jmp_offset = func - (image + addrs[i]); 1015 if (!func || !is_simm32(jmp_offset)) { 1016 pr_err("unsupported bpf func %d addr %p image %p\n", 1017 imm32, func, image); 1018 return -EINVAL; 1019 } 1020 if (BPF_MODE(insn->code) == BPF_ABS) { 1021 /* mov %esi, imm32 */ 1022 EMIT1_off32(0xBE, imm32); 1023 } else { 1024 /* mov %rsi, src_reg */ 1025 EMIT_mov(BPF_REG_2, src_reg); 1026 if (imm32) { 1027 if (is_imm8(imm32)) 1028 /* add %esi, imm8 */ 1029 EMIT3(0x83, 0xC6, imm32); 1030 else 1031 /* add %esi, imm32 */ 1032 EMIT2_off32(0x81, 0xC6, imm32); 1033 } 1034 } 1035 /* skb pointer is in R6 (%rbx), it will be copied into 1036 * %rdi if skb_copy_bits() call is necessary. 1037 * sk_load_* helpers also use %r10 and %r9d. 1038 * See bpf_jit.S 1039 */ 1040 if (seen_ax_reg) 1041 /* r10 = skb->data, mov %r10, off32(%rbx) */ 1042 EMIT3_off32(0x4c, 0x8b, 0x93, 1043 offsetof(struct sk_buff, data)); 1044 EMIT1_off32(0xE8, jmp_offset); /* call */ 1045 break; 1046 1047 case BPF_LD | BPF_IND | BPF_H: 1048 func = sk_load_half; 1049 goto common_load; 1050 case BPF_LD | BPF_ABS | BPF_H: 1051 func = CHOOSE_LOAD_FUNC(imm32, sk_load_half); 1052 goto common_load; 1053 case BPF_LD | BPF_IND | BPF_B: 1054 func = sk_load_byte; 1055 goto common_load; 1056 case BPF_LD | BPF_ABS | BPF_B: 1057 func = CHOOSE_LOAD_FUNC(imm32, sk_load_byte); 1058 goto common_load; 1059 1060 case BPF_JMP | BPF_EXIT: 1061 if (seen_exit) { 1062 jmp_offset = ctx->cleanup_addr - addrs[i]; 1063 goto emit_jmp; 1064 } 1065 seen_exit = true; 1066 /* update cleanup_addr */ 1067 ctx->cleanup_addr = proglen; 1068 /* mov rbx, qword ptr [rbp+0] */ 1069 EMIT4(0x48, 0x8B, 0x5D, 0); 1070 /* mov r13, qword ptr [rbp+8] */ 1071 EMIT4(0x4C, 0x8B, 0x6D, 8); 1072 /* mov r14, qword ptr [rbp+16] */ 1073 EMIT4(0x4C, 0x8B, 0x75, 16); 1074 /* mov r15, qword ptr [rbp+24] */ 1075 EMIT4(0x4C, 0x8B, 0x7D, 24); 1076 1077 /* add rbp, AUX_STACK_SPACE */ 1078 EMIT4(0x48, 0x83, 0xC5, AUX_STACK_SPACE); 1079 EMIT1(0xC9); /* leave */ 1080 EMIT1(0xC3); /* ret */ 1081 break; 1082 1083 default: 1084 /* By design x64 JIT should support all BPF instructions 1085 * This error will be seen if new instruction was added 1086 * to interpreter, but not to JIT 1087 * or if there is junk in bpf_prog 1088 */ 1089 pr_err("bpf_jit: unknown opcode %02x\n", insn->code); 1090 return -EINVAL; 1091 } 1092 1093 ilen = prog - temp; 1094 if (ilen > BPF_MAX_INSN_SIZE) { 1095 pr_err("bpf_jit: fatal insn size error\n"); 1096 return -EFAULT; 1097 } 1098 1099 if (image) { 1100 if (unlikely(proglen + ilen > oldproglen)) { 1101 pr_err("bpf_jit: fatal error\n"); 1102 return -EFAULT; 1103 } 1104 memcpy(image + proglen, temp, ilen); 1105 } 1106 proglen += ilen; 1107 addrs[i] = proglen; 1108 prog = temp; 1109 } 1110 return proglen; 1111 } 1112 1113 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) 1114 { 1115 struct bpf_binary_header *header = NULL; 1116 struct bpf_prog *tmp, *orig_prog = prog; 1117 int proglen, oldproglen = 0; 1118 struct jit_context ctx = {}; 1119 bool tmp_blinded = false; 1120 u8 *image = NULL; 1121 int *addrs; 1122 int pass; 1123 int i; 1124 1125 if (!bpf_jit_enable) 1126 return orig_prog; 1127 1128 tmp = bpf_jit_blind_constants(prog); 1129 /* If blinding was requested and we failed during blinding, 1130 * we must fall back to the interpreter. 1131 */ 1132 if (IS_ERR(tmp)) 1133 return orig_prog; 1134 if (tmp != prog) { 1135 tmp_blinded = true; 1136 prog = tmp; 1137 } 1138 1139 addrs = kmalloc(prog->len * sizeof(*addrs), GFP_KERNEL); 1140 if (!addrs) { 1141 prog = orig_prog; 1142 goto out; 1143 } 1144 1145 /* Before first pass, make a rough estimation of addrs[] 1146 * each bpf instruction is translated to less than 64 bytes 1147 */ 1148 for (proglen = 0, i = 0; i < prog->len; i++) { 1149 proglen += 64; 1150 addrs[i] = proglen; 1151 } 1152 ctx.cleanup_addr = proglen; 1153 1154 /* JITed image shrinks with every pass and the loop iterates 1155 * until the image stops shrinking. Very large bpf programs 1156 * may converge on the last pass. In such case do one more 1157 * pass to emit the final image 1158 */ 1159 for (pass = 0; pass < 10 || image; pass++) { 1160 proglen = do_jit(prog, addrs, image, oldproglen, &ctx); 1161 if (proglen <= 0) { 1162 image = NULL; 1163 if (header) 1164 bpf_jit_binary_free(header); 1165 prog = orig_prog; 1166 goto out_addrs; 1167 } 1168 if (image) { 1169 if (proglen != oldproglen) { 1170 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n", 1171 proglen, oldproglen); 1172 prog = orig_prog; 1173 goto out_addrs; 1174 } 1175 break; 1176 } 1177 if (proglen == oldproglen) { 1178 header = bpf_jit_binary_alloc(proglen, &image, 1179 1, jit_fill_hole); 1180 if (!header) { 1181 prog = orig_prog; 1182 goto out_addrs; 1183 } 1184 } 1185 oldproglen = proglen; 1186 } 1187 1188 if (bpf_jit_enable > 1) 1189 bpf_jit_dump(prog->len, proglen, pass + 1, image); 1190 1191 if (image) { 1192 bpf_flush_icache(header, image + proglen); 1193 bpf_jit_binary_lock_ro(header); 1194 prog->bpf_func = (void *)image; 1195 prog->jited = 1; 1196 prog->jited_len = proglen; 1197 } else { 1198 prog = orig_prog; 1199 } 1200 1201 out_addrs: 1202 kfree(addrs); 1203 out: 1204 if (tmp_blinded) 1205 bpf_jit_prog_release_other(prog, prog == orig_prog ? 1206 tmp : orig_prog); 1207 return prog; 1208 } 1209