1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * BPF JIT compiler 4 * 5 * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com) 6 * Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com 7 */ 8 #include <linux/netdevice.h> 9 #include <linux/filter.h> 10 #include <linux/if_vlan.h> 11 #include <linux/bpf.h> 12 #include <linux/memory.h> 13 #include <linux/sort.h> 14 #include <asm/extable.h> 15 #include <asm/ftrace.h> 16 #include <asm/set_memory.h> 17 #include <asm/nospec-branch.h> 18 #include <asm/text-patching.h> 19 #include <asm/unwind.h> 20 #include <asm/cfi.h> 21 22 static bool all_callee_regs_used[4] = {true, true, true, true}; 23 24 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len) 25 { 26 if (len == 1) 27 *ptr = bytes; 28 else if (len == 2) 29 *(u16 *)ptr = bytes; 30 else { 31 *(u32 *)ptr = bytes; 32 barrier(); 33 } 34 return ptr + len; 35 } 36 37 #define EMIT(bytes, len) \ 38 do { prog = emit_code(prog, bytes, len); } while (0) 39 40 #define EMIT1(b1) EMIT(b1, 1) 41 #define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2) 42 #define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3) 43 #define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4) 44 45 #define EMIT1_off32(b1, off) \ 46 do { EMIT1(b1); EMIT(off, 4); } while (0) 47 #define EMIT2_off32(b1, b2, off) \ 48 do { EMIT2(b1, b2); EMIT(off, 4); } while (0) 49 #define EMIT3_off32(b1, b2, b3, off) \ 50 do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0) 51 #define EMIT4_off32(b1, b2, b3, b4, off) \ 52 do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0) 53 54 #ifdef CONFIG_X86_KERNEL_IBT 55 #define EMIT_ENDBR() EMIT(gen_endbr(), 4) 56 #define EMIT_ENDBR_POISON() EMIT(gen_endbr_poison(), 4) 57 #else 58 #define EMIT_ENDBR() 59 #define EMIT_ENDBR_POISON() 60 #endif 61 62 static bool is_imm8(int value) 63 { 64 return value <= 127 && value >= -128; 65 } 66 67 static bool is_simm32(s64 value) 68 { 69 return value == (s64)(s32)value; 70 } 71 72 static bool is_uimm32(u64 value) 73 { 74 return value == (u64)(u32)value; 75 } 76 77 /* mov dst, src */ 78 #define EMIT_mov(DST, SRC) \ 79 do { \ 80 if (DST != SRC) \ 81 EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \ 82 } while (0) 83 84 static int bpf_size_to_x86_bytes(int bpf_size) 85 { 86 if (bpf_size == BPF_W) 87 return 4; 88 else if (bpf_size == BPF_H) 89 return 2; 90 else if (bpf_size == BPF_B) 91 return 1; 92 else if (bpf_size == BPF_DW) 93 return 4; /* imm32 */ 94 else 95 return 0; 96 } 97 98 /* 99 * List of x86 cond jumps opcodes (. + s8) 100 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32) 101 */ 102 #define X86_JB 0x72 103 #define X86_JAE 0x73 104 #define X86_JE 0x74 105 #define X86_JNE 0x75 106 #define X86_JBE 0x76 107 #define X86_JA 0x77 108 #define X86_JL 0x7C 109 #define X86_JGE 0x7D 110 #define X86_JLE 0x7E 111 #define X86_JG 0x7F 112 113 /* Pick a register outside of BPF range for JIT internal work */ 114 #define AUX_REG (MAX_BPF_JIT_REG + 1) 115 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2) 116 #define X86_REG_R12 (MAX_BPF_JIT_REG + 3) 117 118 /* 119 * The following table maps BPF registers to x86-64 registers. 120 * 121 * x86-64 register R12 is unused, since if used as base address 122 * register in load/store instructions, it always needs an 123 * extra byte of encoding and is callee saved. 124 * 125 * x86-64 register R9 is not used by BPF programs, but can be used by BPF 126 * trampoline. x86-64 register R10 is used for blinding (if enabled). 127 */ 128 static const int reg2hex[] = { 129 [BPF_REG_0] = 0, /* RAX */ 130 [BPF_REG_1] = 7, /* RDI */ 131 [BPF_REG_2] = 6, /* RSI */ 132 [BPF_REG_3] = 2, /* RDX */ 133 [BPF_REG_4] = 1, /* RCX */ 134 [BPF_REG_5] = 0, /* R8 */ 135 [BPF_REG_6] = 3, /* RBX callee saved */ 136 [BPF_REG_7] = 5, /* R13 callee saved */ 137 [BPF_REG_8] = 6, /* R14 callee saved */ 138 [BPF_REG_9] = 7, /* R15 callee saved */ 139 [BPF_REG_FP] = 5, /* RBP readonly */ 140 [BPF_REG_AX] = 2, /* R10 temp register */ 141 [AUX_REG] = 3, /* R11 temp register */ 142 [X86_REG_R9] = 1, /* R9 register, 6th function argument */ 143 [X86_REG_R12] = 4, /* R12 callee saved */ 144 }; 145 146 static const int reg2pt_regs[] = { 147 [BPF_REG_0] = offsetof(struct pt_regs, ax), 148 [BPF_REG_1] = offsetof(struct pt_regs, di), 149 [BPF_REG_2] = offsetof(struct pt_regs, si), 150 [BPF_REG_3] = offsetof(struct pt_regs, dx), 151 [BPF_REG_4] = offsetof(struct pt_regs, cx), 152 [BPF_REG_5] = offsetof(struct pt_regs, r8), 153 [BPF_REG_6] = offsetof(struct pt_regs, bx), 154 [BPF_REG_7] = offsetof(struct pt_regs, r13), 155 [BPF_REG_8] = offsetof(struct pt_regs, r14), 156 [BPF_REG_9] = offsetof(struct pt_regs, r15), 157 }; 158 159 /* 160 * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15 161 * which need extra byte of encoding. 162 * rax,rcx,...,rbp have simpler encoding 163 */ 164 static bool is_ereg(u32 reg) 165 { 166 return (1 << reg) & (BIT(BPF_REG_5) | 167 BIT(AUX_REG) | 168 BIT(BPF_REG_7) | 169 BIT(BPF_REG_8) | 170 BIT(BPF_REG_9) | 171 BIT(X86_REG_R9) | 172 BIT(X86_REG_R12) | 173 BIT(BPF_REG_AX)); 174 } 175 176 /* 177 * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64 178 * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte 179 * of encoding. al,cl,dl,bl have simpler encoding. 180 */ 181 static bool is_ereg_8l(u32 reg) 182 { 183 return is_ereg(reg) || 184 (1 << reg) & (BIT(BPF_REG_1) | 185 BIT(BPF_REG_2) | 186 BIT(BPF_REG_FP)); 187 } 188 189 static bool is_axreg(u32 reg) 190 { 191 return reg == BPF_REG_0; 192 } 193 194 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */ 195 static u8 add_1mod(u8 byte, u32 reg) 196 { 197 if (is_ereg(reg)) 198 byte |= 1; 199 return byte; 200 } 201 202 static u8 add_2mod(u8 byte, u32 r1, u32 r2) 203 { 204 if (is_ereg(r1)) 205 byte |= 1; 206 if (is_ereg(r2)) 207 byte |= 4; 208 return byte; 209 } 210 211 static u8 add_3mod(u8 byte, u32 r1, u32 r2, u32 index) 212 { 213 if (is_ereg(r1)) 214 byte |= 1; 215 if (is_ereg(index)) 216 byte |= 2; 217 if (is_ereg(r2)) 218 byte |= 4; 219 return byte; 220 } 221 222 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */ 223 static u8 add_1reg(u8 byte, u32 dst_reg) 224 { 225 return byte + reg2hex[dst_reg]; 226 } 227 228 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */ 229 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg) 230 { 231 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3); 232 } 233 234 /* Some 1-byte opcodes for binary ALU operations */ 235 static u8 simple_alu_opcodes[] = { 236 [BPF_ADD] = 0x01, 237 [BPF_SUB] = 0x29, 238 [BPF_AND] = 0x21, 239 [BPF_OR] = 0x09, 240 [BPF_XOR] = 0x31, 241 [BPF_LSH] = 0xE0, 242 [BPF_RSH] = 0xE8, 243 [BPF_ARSH] = 0xF8, 244 }; 245 246 static void jit_fill_hole(void *area, unsigned int size) 247 { 248 /* Fill whole space with INT3 instructions */ 249 memset(area, 0xcc, size); 250 } 251 252 int bpf_arch_text_invalidate(void *dst, size_t len) 253 { 254 return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len)); 255 } 256 257 struct jit_context { 258 int cleanup_addr; /* Epilogue code offset */ 259 260 /* 261 * Program specific offsets of labels in the code; these rely on the 262 * JIT doing at least 2 passes, recording the position on the first 263 * pass, only to generate the correct offset on the second pass. 264 */ 265 int tail_call_direct_label; 266 int tail_call_indirect_label; 267 }; 268 269 /* Maximum number of bytes emitted while JITing one eBPF insn */ 270 #define BPF_MAX_INSN_SIZE 128 271 #define BPF_INSN_SAFETY 64 272 273 /* Number of bytes emit_patch() needs to generate instructions */ 274 #define X86_PATCH_SIZE 5 275 /* Number of bytes that will be skipped on tailcall */ 276 #define X86_TAIL_CALL_OFFSET (11 + ENDBR_INSN_SIZE) 277 278 static void push_r12(u8 **pprog) 279 { 280 u8 *prog = *pprog; 281 282 EMIT2(0x41, 0x54); /* push r12 */ 283 *pprog = prog; 284 } 285 286 static void push_callee_regs(u8 **pprog, bool *callee_regs_used) 287 { 288 u8 *prog = *pprog; 289 290 if (callee_regs_used[0]) 291 EMIT1(0x53); /* push rbx */ 292 if (callee_regs_used[1]) 293 EMIT2(0x41, 0x55); /* push r13 */ 294 if (callee_regs_used[2]) 295 EMIT2(0x41, 0x56); /* push r14 */ 296 if (callee_regs_used[3]) 297 EMIT2(0x41, 0x57); /* push r15 */ 298 *pprog = prog; 299 } 300 301 static void pop_r12(u8 **pprog) 302 { 303 u8 *prog = *pprog; 304 305 EMIT2(0x41, 0x5C); /* pop r12 */ 306 *pprog = prog; 307 } 308 309 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used) 310 { 311 u8 *prog = *pprog; 312 313 if (callee_regs_used[3]) 314 EMIT2(0x41, 0x5F); /* pop r15 */ 315 if (callee_regs_used[2]) 316 EMIT2(0x41, 0x5E); /* pop r14 */ 317 if (callee_regs_used[1]) 318 EMIT2(0x41, 0x5D); /* pop r13 */ 319 if (callee_regs_used[0]) 320 EMIT1(0x5B); /* pop rbx */ 321 *pprog = prog; 322 } 323 324 static void emit_nops(u8 **pprog, int len) 325 { 326 u8 *prog = *pprog; 327 int i, noplen; 328 329 while (len > 0) { 330 noplen = len; 331 332 if (noplen > ASM_NOP_MAX) 333 noplen = ASM_NOP_MAX; 334 335 for (i = 0; i < noplen; i++) 336 EMIT1(x86_nops[noplen][i]); 337 len -= noplen; 338 } 339 340 *pprog = prog; 341 } 342 343 /* 344 * Emit the various CFI preambles, see asm/cfi.h and the comments about FineIBT 345 * in arch/x86/kernel/alternative.c 346 */ 347 348 static void emit_fineibt(u8 **pprog, u32 hash) 349 { 350 u8 *prog = *pprog; 351 352 EMIT_ENDBR(); 353 EMIT3_off32(0x41, 0x81, 0xea, hash); /* subl $hash, %r10d */ 354 EMIT2(0x74, 0x07); /* jz.d8 +7 */ 355 EMIT2(0x0f, 0x0b); /* ud2 */ 356 EMIT1(0x90); /* nop */ 357 EMIT_ENDBR_POISON(); 358 359 *pprog = prog; 360 } 361 362 static void emit_kcfi(u8 **pprog, u32 hash) 363 { 364 u8 *prog = *pprog; 365 366 EMIT1_off32(0xb8, hash); /* movl $hash, %eax */ 367 #ifdef CONFIG_CALL_PADDING 368 EMIT1(0x90); 369 EMIT1(0x90); 370 EMIT1(0x90); 371 EMIT1(0x90); 372 EMIT1(0x90); 373 EMIT1(0x90); 374 EMIT1(0x90); 375 EMIT1(0x90); 376 EMIT1(0x90); 377 EMIT1(0x90); 378 EMIT1(0x90); 379 #endif 380 EMIT_ENDBR(); 381 382 *pprog = prog; 383 } 384 385 static void emit_cfi(u8 **pprog, u32 hash) 386 { 387 u8 *prog = *pprog; 388 389 switch (cfi_mode) { 390 case CFI_FINEIBT: 391 emit_fineibt(&prog, hash); 392 break; 393 394 case CFI_KCFI: 395 emit_kcfi(&prog, hash); 396 break; 397 398 default: 399 EMIT_ENDBR(); 400 break; 401 } 402 403 *pprog = prog; 404 } 405 406 /* 407 * Emit x86-64 prologue code for BPF program. 408 * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes 409 * while jumping to another program 410 */ 411 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf, 412 bool tail_call_reachable, bool is_subprog, 413 bool is_exception_cb) 414 { 415 u8 *prog = *pprog; 416 417 emit_cfi(&prog, is_subprog ? cfi_bpf_subprog_hash : cfi_bpf_hash); 418 /* BPF trampoline can be made to work without these nops, 419 * but let's waste 5 bytes for now and optimize later 420 */ 421 emit_nops(&prog, X86_PATCH_SIZE); 422 if (!ebpf_from_cbpf) { 423 if (tail_call_reachable && !is_subprog) 424 /* When it's the entry of the whole tailcall context, 425 * zeroing rax means initialising tail_call_cnt. 426 */ 427 EMIT2(0x31, 0xC0); /* xor eax, eax */ 428 else 429 /* Keep the same instruction layout. */ 430 EMIT2(0x66, 0x90); /* nop2 */ 431 } 432 /* Exception callback receives FP as third parameter */ 433 if (is_exception_cb) { 434 EMIT3(0x48, 0x89, 0xF4); /* mov rsp, rsi */ 435 EMIT3(0x48, 0x89, 0xD5); /* mov rbp, rdx */ 436 /* The main frame must have exception_boundary as true, so we 437 * first restore those callee-saved regs from stack, before 438 * reusing the stack frame. 439 */ 440 pop_callee_regs(&prog, all_callee_regs_used); 441 pop_r12(&prog); 442 /* Reset the stack frame. */ 443 EMIT3(0x48, 0x89, 0xEC); /* mov rsp, rbp */ 444 } else { 445 EMIT1(0x55); /* push rbp */ 446 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */ 447 } 448 449 /* X86_TAIL_CALL_OFFSET is here */ 450 EMIT_ENDBR(); 451 452 /* sub rsp, rounded_stack_depth */ 453 if (stack_depth) 454 EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8)); 455 if (tail_call_reachable) 456 EMIT1(0x50); /* push rax */ 457 *pprog = prog; 458 } 459 460 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode) 461 { 462 u8 *prog = *pprog; 463 s64 offset; 464 465 offset = func - (ip + X86_PATCH_SIZE); 466 if (!is_simm32(offset)) { 467 pr_err("Target call %p is out of range\n", func); 468 return -ERANGE; 469 } 470 EMIT1_off32(opcode, offset); 471 *pprog = prog; 472 return 0; 473 } 474 475 static int emit_call(u8 **pprog, void *func, void *ip) 476 { 477 return emit_patch(pprog, func, ip, 0xE8); 478 } 479 480 static int emit_rsb_call(u8 **pprog, void *func, void *ip) 481 { 482 OPTIMIZER_HIDE_VAR(func); 483 x86_call_depth_emit_accounting(pprog, func); 484 return emit_patch(pprog, func, ip, 0xE8); 485 } 486 487 static int emit_jump(u8 **pprog, void *func, void *ip) 488 { 489 return emit_patch(pprog, func, ip, 0xE9); 490 } 491 492 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 493 void *old_addr, void *new_addr) 494 { 495 const u8 *nop_insn = x86_nops[5]; 496 u8 old_insn[X86_PATCH_SIZE]; 497 u8 new_insn[X86_PATCH_SIZE]; 498 u8 *prog; 499 int ret; 500 501 memcpy(old_insn, nop_insn, X86_PATCH_SIZE); 502 if (old_addr) { 503 prog = old_insn; 504 ret = t == BPF_MOD_CALL ? 505 emit_call(&prog, old_addr, ip) : 506 emit_jump(&prog, old_addr, ip); 507 if (ret) 508 return ret; 509 } 510 511 memcpy(new_insn, nop_insn, X86_PATCH_SIZE); 512 if (new_addr) { 513 prog = new_insn; 514 ret = t == BPF_MOD_CALL ? 515 emit_call(&prog, new_addr, ip) : 516 emit_jump(&prog, new_addr, ip); 517 if (ret) 518 return ret; 519 } 520 521 ret = -EBUSY; 522 mutex_lock(&text_mutex); 523 if (memcmp(ip, old_insn, X86_PATCH_SIZE)) 524 goto out; 525 ret = 1; 526 if (memcmp(ip, new_insn, X86_PATCH_SIZE)) { 527 text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL); 528 ret = 0; 529 } 530 out: 531 mutex_unlock(&text_mutex); 532 return ret; 533 } 534 535 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 536 void *old_addr, void *new_addr) 537 { 538 if (!is_kernel_text((long)ip) && 539 !is_bpf_text_address((long)ip)) 540 /* BPF poking in modules is not supported */ 541 return -EINVAL; 542 543 /* 544 * See emit_prologue(), for IBT builds the trampoline hook is preceded 545 * with an ENDBR instruction. 546 */ 547 if (is_endbr(*(u32 *)ip)) 548 ip += ENDBR_INSN_SIZE; 549 550 return __bpf_arch_text_poke(ip, t, old_addr, new_addr); 551 } 552 553 #define EMIT_LFENCE() EMIT3(0x0F, 0xAE, 0xE8) 554 555 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip) 556 { 557 u8 *prog = *pprog; 558 559 if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) { 560 EMIT_LFENCE(); 561 EMIT2(0xFF, 0xE0 + reg); 562 } else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) { 563 OPTIMIZER_HIDE_VAR(reg); 564 if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH)) 565 emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip); 566 else 567 emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip); 568 } else { 569 EMIT2(0xFF, 0xE0 + reg); /* jmp *%\reg */ 570 if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) || IS_ENABLED(CONFIG_MITIGATION_SLS)) 571 EMIT1(0xCC); /* int3 */ 572 } 573 574 *pprog = prog; 575 } 576 577 static void emit_return(u8 **pprog, u8 *ip) 578 { 579 u8 *prog = *pprog; 580 581 if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) { 582 emit_jump(&prog, x86_return_thunk, ip); 583 } else { 584 EMIT1(0xC3); /* ret */ 585 if (IS_ENABLED(CONFIG_MITIGATION_SLS)) 586 EMIT1(0xCC); /* int3 */ 587 } 588 589 *pprog = prog; 590 } 591 592 /* 593 * Generate the following code: 594 * 595 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ... 596 * if (index >= array->map.max_entries) 597 * goto out; 598 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT) 599 * goto out; 600 * prog = array->ptrs[index]; 601 * if (prog == NULL) 602 * goto out; 603 * goto *(prog->bpf_func + prologue_size); 604 * out: 605 */ 606 static void emit_bpf_tail_call_indirect(struct bpf_prog *bpf_prog, 607 u8 **pprog, bool *callee_regs_used, 608 u32 stack_depth, u8 *ip, 609 struct jit_context *ctx) 610 { 611 int tcc_off = -4 - round_up(stack_depth, 8); 612 u8 *prog = *pprog, *start = *pprog; 613 int offset; 614 615 /* 616 * rdi - pointer to ctx 617 * rsi - pointer to bpf_array 618 * rdx - index in bpf_array 619 */ 620 621 /* 622 * if (index >= array->map.max_entries) 623 * goto out; 624 */ 625 EMIT2(0x89, 0xD2); /* mov edx, edx */ 626 EMIT3(0x39, 0x56, /* cmp dword ptr [rsi + 16], edx */ 627 offsetof(struct bpf_array, map.max_entries)); 628 629 offset = ctx->tail_call_indirect_label - (prog + 2 - start); 630 EMIT2(X86_JBE, offset); /* jbe out */ 631 632 /* 633 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT) 634 * goto out; 635 */ 636 EMIT2_off32(0x8B, 0x85, tcc_off); /* mov eax, dword ptr [rbp - tcc_off] */ 637 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */ 638 639 offset = ctx->tail_call_indirect_label - (prog + 2 - start); 640 EMIT2(X86_JAE, offset); /* jae out */ 641 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */ 642 EMIT2_off32(0x89, 0x85, tcc_off); /* mov dword ptr [rbp - tcc_off], eax */ 643 644 /* prog = array->ptrs[index]; */ 645 EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6, /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */ 646 offsetof(struct bpf_array, ptrs)); 647 648 /* 649 * if (prog == NULL) 650 * goto out; 651 */ 652 EMIT3(0x48, 0x85, 0xC9); /* test rcx,rcx */ 653 654 offset = ctx->tail_call_indirect_label - (prog + 2 - start); 655 EMIT2(X86_JE, offset); /* je out */ 656 657 if (bpf_prog->aux->exception_boundary) { 658 pop_callee_regs(&prog, all_callee_regs_used); 659 pop_r12(&prog); 660 } else { 661 pop_callee_regs(&prog, callee_regs_used); 662 if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena)) 663 pop_r12(&prog); 664 } 665 666 EMIT1(0x58); /* pop rax */ 667 if (stack_depth) 668 EMIT3_off32(0x48, 0x81, 0xC4, /* add rsp, sd */ 669 round_up(stack_depth, 8)); 670 671 /* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */ 672 EMIT4(0x48, 0x8B, 0x49, /* mov rcx, qword ptr [rcx + 32] */ 673 offsetof(struct bpf_prog, bpf_func)); 674 EMIT4(0x48, 0x83, 0xC1, /* add rcx, X86_TAIL_CALL_OFFSET */ 675 X86_TAIL_CALL_OFFSET); 676 /* 677 * Now we're ready to jump into next BPF program 678 * rdi == ctx (1st arg) 679 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET 680 */ 681 emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start)); 682 683 /* out: */ 684 ctx->tail_call_indirect_label = prog - start; 685 *pprog = prog; 686 } 687 688 static void emit_bpf_tail_call_direct(struct bpf_prog *bpf_prog, 689 struct bpf_jit_poke_descriptor *poke, 690 u8 **pprog, u8 *ip, 691 bool *callee_regs_used, u32 stack_depth, 692 struct jit_context *ctx) 693 { 694 int tcc_off = -4 - round_up(stack_depth, 8); 695 u8 *prog = *pprog, *start = *pprog; 696 int offset; 697 698 /* 699 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT) 700 * goto out; 701 */ 702 EMIT2_off32(0x8B, 0x85, tcc_off); /* mov eax, dword ptr [rbp - tcc_off] */ 703 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */ 704 705 offset = ctx->tail_call_direct_label - (prog + 2 - start); 706 EMIT2(X86_JAE, offset); /* jae out */ 707 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */ 708 EMIT2_off32(0x89, 0x85, tcc_off); /* mov dword ptr [rbp - tcc_off], eax */ 709 710 poke->tailcall_bypass = ip + (prog - start); 711 poke->adj_off = X86_TAIL_CALL_OFFSET; 712 poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE; 713 poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE; 714 715 emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE, 716 poke->tailcall_bypass); 717 718 if (bpf_prog->aux->exception_boundary) { 719 pop_callee_regs(&prog, all_callee_regs_used); 720 pop_r12(&prog); 721 } else { 722 pop_callee_regs(&prog, callee_regs_used); 723 if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena)) 724 pop_r12(&prog); 725 } 726 727 EMIT1(0x58); /* pop rax */ 728 if (stack_depth) 729 EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8)); 730 731 emit_nops(&prog, X86_PATCH_SIZE); 732 733 /* out: */ 734 ctx->tail_call_direct_label = prog - start; 735 736 *pprog = prog; 737 } 738 739 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog) 740 { 741 struct bpf_jit_poke_descriptor *poke; 742 struct bpf_array *array; 743 struct bpf_prog *target; 744 int i, ret; 745 746 for (i = 0; i < prog->aux->size_poke_tab; i++) { 747 poke = &prog->aux->poke_tab[i]; 748 if (poke->aux && poke->aux != prog->aux) 749 continue; 750 751 WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable)); 752 753 if (poke->reason != BPF_POKE_REASON_TAIL_CALL) 754 continue; 755 756 array = container_of(poke->tail_call.map, struct bpf_array, map); 757 mutex_lock(&array->aux->poke_mutex); 758 target = array->ptrs[poke->tail_call.key]; 759 if (target) { 760 ret = __bpf_arch_text_poke(poke->tailcall_target, 761 BPF_MOD_JUMP, NULL, 762 (u8 *)target->bpf_func + 763 poke->adj_off); 764 BUG_ON(ret < 0); 765 ret = __bpf_arch_text_poke(poke->tailcall_bypass, 766 BPF_MOD_JUMP, 767 (u8 *)poke->tailcall_target + 768 X86_PATCH_SIZE, NULL); 769 BUG_ON(ret < 0); 770 } 771 WRITE_ONCE(poke->tailcall_target_stable, true); 772 mutex_unlock(&array->aux->poke_mutex); 773 } 774 } 775 776 static void emit_mov_imm32(u8 **pprog, bool sign_propagate, 777 u32 dst_reg, const u32 imm32) 778 { 779 u8 *prog = *pprog; 780 u8 b1, b2, b3; 781 782 /* 783 * Optimization: if imm32 is positive, use 'mov %eax, imm32' 784 * (which zero-extends imm32) to save 2 bytes. 785 */ 786 if (sign_propagate && (s32)imm32 < 0) { 787 /* 'mov %rax, imm32' sign extends imm32 */ 788 b1 = add_1mod(0x48, dst_reg); 789 b2 = 0xC7; 790 b3 = 0xC0; 791 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32); 792 goto done; 793 } 794 795 /* 796 * Optimization: if imm32 is zero, use 'xor %eax, %eax' 797 * to save 3 bytes. 798 */ 799 if (imm32 == 0) { 800 if (is_ereg(dst_reg)) 801 EMIT1(add_2mod(0x40, dst_reg, dst_reg)); 802 b2 = 0x31; /* xor */ 803 b3 = 0xC0; 804 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg)); 805 goto done; 806 } 807 808 /* mov %eax, imm32 */ 809 if (is_ereg(dst_reg)) 810 EMIT1(add_1mod(0x40, dst_reg)); 811 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32); 812 done: 813 *pprog = prog; 814 } 815 816 static void emit_mov_imm64(u8 **pprog, u32 dst_reg, 817 const u32 imm32_hi, const u32 imm32_lo) 818 { 819 u8 *prog = *pprog; 820 821 if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) { 822 /* 823 * For emitting plain u32, where sign bit must not be 824 * propagated LLVM tends to load imm64 over mov32 825 * directly, so save couple of bytes by just doing 826 * 'mov %eax, imm32' instead. 827 */ 828 emit_mov_imm32(&prog, false, dst_reg, imm32_lo); 829 } else { 830 /* movabsq rax, imm64 */ 831 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg)); 832 EMIT(imm32_lo, 4); 833 EMIT(imm32_hi, 4); 834 } 835 836 *pprog = prog; 837 } 838 839 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg) 840 { 841 u8 *prog = *pprog; 842 843 if (is64) { 844 /* mov dst, src */ 845 EMIT_mov(dst_reg, src_reg); 846 } else { 847 /* mov32 dst, src */ 848 if (is_ereg(dst_reg) || is_ereg(src_reg)) 849 EMIT1(add_2mod(0x40, dst_reg, src_reg)); 850 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg)); 851 } 852 853 *pprog = prog; 854 } 855 856 static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg, 857 u32 src_reg) 858 { 859 u8 *prog = *pprog; 860 861 if (is64) { 862 /* movs[b,w,l]q dst, src */ 863 if (num_bits == 8) 864 EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe, 865 add_2reg(0xC0, src_reg, dst_reg)); 866 else if (num_bits == 16) 867 EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf, 868 add_2reg(0xC0, src_reg, dst_reg)); 869 else if (num_bits == 32) 870 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63, 871 add_2reg(0xC0, src_reg, dst_reg)); 872 } else { 873 /* movs[b,w]l dst, src */ 874 if (num_bits == 8) { 875 EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe, 876 add_2reg(0xC0, src_reg, dst_reg)); 877 } else if (num_bits == 16) { 878 if (is_ereg(dst_reg) || is_ereg(src_reg)) 879 EMIT1(add_2mod(0x40, src_reg, dst_reg)); 880 EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf, 881 add_2reg(0xC0, src_reg, dst_reg)); 882 } 883 } 884 885 *pprog = prog; 886 } 887 888 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */ 889 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off) 890 { 891 u8 *prog = *pprog; 892 893 if (is_imm8(off)) { 894 /* 1-byte signed displacement. 895 * 896 * If off == 0 we could skip this and save one extra byte, but 897 * special case of x86 R13 which always needs an offset is not 898 * worth the hassle 899 */ 900 EMIT2(add_2reg(0x40, ptr_reg, val_reg), off); 901 } else { 902 /* 4-byte signed displacement */ 903 EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off); 904 } 905 *pprog = prog; 906 } 907 908 static void emit_insn_suffix_SIB(u8 **pprog, u32 ptr_reg, u32 val_reg, u32 index_reg, int off) 909 { 910 u8 *prog = *pprog; 911 912 if (is_imm8(off)) { 913 EMIT3(add_2reg(0x44, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off); 914 } else { 915 EMIT2_off32(add_2reg(0x84, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off); 916 } 917 *pprog = prog; 918 } 919 920 /* 921 * Emit a REX byte if it will be necessary to address these registers 922 */ 923 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64) 924 { 925 u8 *prog = *pprog; 926 927 if (is64) 928 EMIT1(add_2mod(0x48, dst_reg, src_reg)); 929 else if (is_ereg(dst_reg) || is_ereg(src_reg)) 930 EMIT1(add_2mod(0x40, dst_reg, src_reg)); 931 *pprog = prog; 932 } 933 934 /* 935 * Similar version of maybe_emit_mod() for a single register 936 */ 937 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64) 938 { 939 u8 *prog = *pprog; 940 941 if (is64) 942 EMIT1(add_1mod(0x48, reg)); 943 else if (is_ereg(reg)) 944 EMIT1(add_1mod(0x40, reg)); 945 *pprog = prog; 946 } 947 948 /* LDX: dst_reg = *(u8*)(src_reg + off) */ 949 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off) 950 { 951 u8 *prog = *pprog; 952 953 switch (size) { 954 case BPF_B: 955 /* Emit 'movzx rax, byte ptr [rax + off]' */ 956 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6); 957 break; 958 case BPF_H: 959 /* Emit 'movzx rax, word ptr [rax + off]' */ 960 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7); 961 break; 962 case BPF_W: 963 /* Emit 'mov eax, dword ptr [rax+0x14]' */ 964 if (is_ereg(dst_reg) || is_ereg(src_reg)) 965 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B); 966 else 967 EMIT1(0x8B); 968 break; 969 case BPF_DW: 970 /* Emit 'mov rax, qword ptr [rax+0x14]' */ 971 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B); 972 break; 973 } 974 emit_insn_suffix(&prog, src_reg, dst_reg, off); 975 *pprog = prog; 976 } 977 978 /* LDSX: dst_reg = *(s8*)(src_reg + off) */ 979 static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off) 980 { 981 u8 *prog = *pprog; 982 983 switch (size) { 984 case BPF_B: 985 /* Emit 'movsx rax, byte ptr [rax + off]' */ 986 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE); 987 break; 988 case BPF_H: 989 /* Emit 'movsx rax, word ptr [rax + off]' */ 990 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF); 991 break; 992 case BPF_W: 993 /* Emit 'movsx rax, dword ptr [rax+0x14]' */ 994 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63); 995 break; 996 } 997 emit_insn_suffix(&prog, src_reg, dst_reg, off); 998 *pprog = prog; 999 } 1000 1001 static void emit_ldx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off) 1002 { 1003 u8 *prog = *pprog; 1004 1005 switch (size) { 1006 case BPF_B: 1007 /* movzx rax, byte ptr [rax + r12 + off] */ 1008 EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB6); 1009 break; 1010 case BPF_H: 1011 /* movzx rax, word ptr [rax + r12 + off] */ 1012 EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB7); 1013 break; 1014 case BPF_W: 1015 /* mov eax, dword ptr [rax + r12 + off] */ 1016 EMIT2(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x8B); 1017 break; 1018 case BPF_DW: 1019 /* mov rax, qword ptr [rax + r12 + off] */ 1020 EMIT2(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x8B); 1021 break; 1022 } 1023 emit_insn_suffix_SIB(&prog, src_reg, dst_reg, index_reg, off); 1024 *pprog = prog; 1025 } 1026 1027 static void emit_ldx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off) 1028 { 1029 emit_ldx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off); 1030 } 1031 1032 /* STX: *(u8*)(dst_reg + off) = src_reg */ 1033 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off) 1034 { 1035 u8 *prog = *pprog; 1036 1037 switch (size) { 1038 case BPF_B: 1039 /* Emit 'mov byte ptr [rax + off], al' */ 1040 if (is_ereg(dst_reg) || is_ereg_8l(src_reg)) 1041 /* Add extra byte for eregs or SIL,DIL,BPL in src_reg */ 1042 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88); 1043 else 1044 EMIT1(0x88); 1045 break; 1046 case BPF_H: 1047 if (is_ereg(dst_reg) || is_ereg(src_reg)) 1048 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89); 1049 else 1050 EMIT2(0x66, 0x89); 1051 break; 1052 case BPF_W: 1053 if (is_ereg(dst_reg) || is_ereg(src_reg)) 1054 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89); 1055 else 1056 EMIT1(0x89); 1057 break; 1058 case BPF_DW: 1059 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89); 1060 break; 1061 } 1062 emit_insn_suffix(&prog, dst_reg, src_reg, off); 1063 *pprog = prog; 1064 } 1065 1066 /* STX: *(u8*)(dst_reg + index_reg + off) = src_reg */ 1067 static void emit_stx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off) 1068 { 1069 u8 *prog = *pprog; 1070 1071 switch (size) { 1072 case BPF_B: 1073 /* mov byte ptr [rax + r12 + off], al */ 1074 EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x88); 1075 break; 1076 case BPF_H: 1077 /* mov word ptr [rax + r12 + off], ax */ 1078 EMIT3(0x66, add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89); 1079 break; 1080 case BPF_W: 1081 /* mov dword ptr [rax + r12 + 1], eax */ 1082 EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89); 1083 break; 1084 case BPF_DW: 1085 /* mov qword ptr [rax + r12 + 1], rax */ 1086 EMIT2(add_3mod(0x48, dst_reg, src_reg, index_reg), 0x89); 1087 break; 1088 } 1089 emit_insn_suffix_SIB(&prog, dst_reg, src_reg, index_reg, off); 1090 *pprog = prog; 1091 } 1092 1093 static void emit_stx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off) 1094 { 1095 emit_stx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off); 1096 } 1097 1098 /* ST: *(u8*)(dst_reg + index_reg + off) = imm32 */ 1099 static void emit_st_index(u8 **pprog, u32 size, u32 dst_reg, u32 index_reg, int off, int imm) 1100 { 1101 u8 *prog = *pprog; 1102 1103 switch (size) { 1104 case BPF_B: 1105 /* mov byte ptr [rax + r12 + off], imm8 */ 1106 EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC6); 1107 break; 1108 case BPF_H: 1109 /* mov word ptr [rax + r12 + off], imm16 */ 1110 EMIT3(0x66, add_3mod(0x40, dst_reg, 0, index_reg), 0xC7); 1111 break; 1112 case BPF_W: 1113 /* mov dword ptr [rax + r12 + 1], imm32 */ 1114 EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC7); 1115 break; 1116 case BPF_DW: 1117 /* mov qword ptr [rax + r12 + 1], imm32 */ 1118 EMIT2(add_3mod(0x48, dst_reg, 0, index_reg), 0xC7); 1119 break; 1120 } 1121 emit_insn_suffix_SIB(&prog, dst_reg, 0, index_reg, off); 1122 EMIT(imm, bpf_size_to_x86_bytes(size)); 1123 *pprog = prog; 1124 } 1125 1126 static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm) 1127 { 1128 emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm); 1129 } 1130 1131 static int emit_atomic(u8 **pprog, u8 atomic_op, 1132 u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size) 1133 { 1134 u8 *prog = *pprog; 1135 1136 EMIT1(0xF0); /* lock prefix */ 1137 1138 maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW); 1139 1140 /* emit opcode */ 1141 switch (atomic_op) { 1142 case BPF_ADD: 1143 case BPF_AND: 1144 case BPF_OR: 1145 case BPF_XOR: 1146 /* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */ 1147 EMIT1(simple_alu_opcodes[atomic_op]); 1148 break; 1149 case BPF_ADD | BPF_FETCH: 1150 /* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */ 1151 EMIT2(0x0F, 0xC1); 1152 break; 1153 case BPF_XCHG: 1154 /* src_reg = atomic_xchg(dst_reg + off, src_reg); */ 1155 EMIT1(0x87); 1156 break; 1157 case BPF_CMPXCHG: 1158 /* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */ 1159 EMIT2(0x0F, 0xB1); 1160 break; 1161 default: 1162 pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op); 1163 return -EFAULT; 1164 } 1165 1166 emit_insn_suffix(&prog, dst_reg, src_reg, off); 1167 1168 *pprog = prog; 1169 return 0; 1170 } 1171 1172 #define DONT_CLEAR 1 1173 1174 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs) 1175 { 1176 u32 reg = x->fixup >> 8; 1177 1178 /* jump over faulting load and clear dest register */ 1179 if (reg != DONT_CLEAR) 1180 *(unsigned long *)((void *)regs + reg) = 0; 1181 regs->ip += x->fixup & 0xff; 1182 return true; 1183 } 1184 1185 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt, 1186 bool *regs_used, bool *tail_call_seen) 1187 { 1188 int i; 1189 1190 for (i = 1; i <= insn_cnt; i++, insn++) { 1191 if (insn->code == (BPF_JMP | BPF_TAIL_CALL)) 1192 *tail_call_seen = true; 1193 if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6) 1194 regs_used[0] = true; 1195 if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7) 1196 regs_used[1] = true; 1197 if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8) 1198 regs_used[2] = true; 1199 if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9) 1200 regs_used[3] = true; 1201 } 1202 } 1203 1204 /* emit the 3-byte VEX prefix 1205 * 1206 * r: same as rex.r, extra bit for ModRM reg field 1207 * x: same as rex.x, extra bit for SIB index field 1208 * b: same as rex.b, extra bit for ModRM r/m, or SIB base 1209 * m: opcode map select, encoding escape bytes e.g. 0x0f38 1210 * w: same as rex.w (32 bit or 64 bit) or opcode specific 1211 * src_reg2: additional source reg (encoded as BPF reg) 1212 * l: vector length (128 bit or 256 bit) or reserved 1213 * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3) 1214 */ 1215 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m, 1216 bool w, u8 src_reg2, bool l, u8 pp) 1217 { 1218 u8 *prog = *pprog; 1219 const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */ 1220 u8 b1, b2; 1221 u8 vvvv = reg2hex[src_reg2]; 1222 1223 /* reg2hex gives only the lower 3 bit of vvvv */ 1224 if (is_ereg(src_reg2)) 1225 vvvv |= 1 << 3; 1226 1227 /* 1228 * 2nd byte of 3-byte VEX prefix 1229 * ~ means bit inverted encoding 1230 * 1231 * 7 0 1232 * +---+---+---+---+---+---+---+---+ 1233 * |~R |~X |~B | m | 1234 * +---+---+---+---+---+---+---+---+ 1235 */ 1236 b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f); 1237 /* 1238 * 3rd byte of 3-byte VEX prefix 1239 * 1240 * 7 0 1241 * +---+---+---+---+---+---+---+---+ 1242 * | W | ~vvvv | L | pp | 1243 * +---+---+---+---+---+---+---+---+ 1244 */ 1245 b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3); 1246 1247 EMIT3(b0, b1, b2); 1248 *pprog = prog; 1249 } 1250 1251 /* emit BMI2 shift instruction */ 1252 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op) 1253 { 1254 u8 *prog = *pprog; 1255 bool r = is_ereg(dst_reg); 1256 u8 m = 2; /* escape code 0f38 */ 1257 1258 emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op); 1259 EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg)); 1260 *pprog = prog; 1261 } 1262 1263 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp))) 1264 1265 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */ 1266 #define RESTORE_TAIL_CALL_CNT(stack) \ 1267 EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8) 1268 1269 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image, 1270 int oldproglen, struct jit_context *ctx, bool jmp_padding) 1271 { 1272 bool tail_call_reachable = bpf_prog->aux->tail_call_reachable; 1273 struct bpf_insn *insn = bpf_prog->insnsi; 1274 bool callee_regs_used[4] = {}; 1275 int insn_cnt = bpf_prog->len; 1276 bool tail_call_seen = false; 1277 bool seen_exit = false; 1278 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY]; 1279 u64 arena_vm_start, user_vm_start; 1280 int i, excnt = 0; 1281 int ilen, proglen = 0; 1282 u8 *prog = temp; 1283 int err; 1284 1285 arena_vm_start = bpf_arena_get_kern_vm_start(bpf_prog->aux->arena); 1286 user_vm_start = bpf_arena_get_user_vm_start(bpf_prog->aux->arena); 1287 1288 detect_reg_usage(insn, insn_cnt, callee_regs_used, 1289 &tail_call_seen); 1290 1291 /* tail call's presence in current prog implies it is reachable */ 1292 tail_call_reachable |= tail_call_seen; 1293 1294 emit_prologue(&prog, bpf_prog->aux->stack_depth, 1295 bpf_prog_was_classic(bpf_prog), tail_call_reachable, 1296 bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb); 1297 /* Exception callback will clobber callee regs for its own use, and 1298 * restore the original callee regs from main prog's stack frame. 1299 */ 1300 if (bpf_prog->aux->exception_boundary) { 1301 /* We also need to save r12, which is not mapped to any BPF 1302 * register, as we throw after entry into the kernel, which may 1303 * overwrite r12. 1304 */ 1305 push_r12(&prog); 1306 push_callee_regs(&prog, all_callee_regs_used); 1307 } else { 1308 if (arena_vm_start) 1309 push_r12(&prog); 1310 push_callee_regs(&prog, callee_regs_used); 1311 } 1312 if (arena_vm_start) 1313 emit_mov_imm64(&prog, X86_REG_R12, 1314 arena_vm_start >> 32, (u32) arena_vm_start); 1315 1316 ilen = prog - temp; 1317 if (rw_image) 1318 memcpy(rw_image + proglen, temp, ilen); 1319 proglen += ilen; 1320 addrs[0] = proglen; 1321 prog = temp; 1322 1323 for (i = 1; i <= insn_cnt; i++, insn++) { 1324 const s32 imm32 = insn->imm; 1325 u32 dst_reg = insn->dst_reg; 1326 u32 src_reg = insn->src_reg; 1327 u8 b2 = 0, b3 = 0; 1328 u8 *start_of_ldx; 1329 s64 jmp_offset; 1330 s16 insn_off; 1331 u8 jmp_cond; 1332 u8 *func; 1333 int nops; 1334 1335 switch (insn->code) { 1336 /* ALU */ 1337 case BPF_ALU | BPF_ADD | BPF_X: 1338 case BPF_ALU | BPF_SUB | BPF_X: 1339 case BPF_ALU | BPF_AND | BPF_X: 1340 case BPF_ALU | BPF_OR | BPF_X: 1341 case BPF_ALU | BPF_XOR | BPF_X: 1342 case BPF_ALU64 | BPF_ADD | BPF_X: 1343 case BPF_ALU64 | BPF_SUB | BPF_X: 1344 case BPF_ALU64 | BPF_AND | BPF_X: 1345 case BPF_ALU64 | BPF_OR | BPF_X: 1346 case BPF_ALU64 | BPF_XOR | BPF_X: 1347 maybe_emit_mod(&prog, dst_reg, src_reg, 1348 BPF_CLASS(insn->code) == BPF_ALU64); 1349 b2 = simple_alu_opcodes[BPF_OP(insn->code)]; 1350 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg)); 1351 break; 1352 1353 case BPF_ALU64 | BPF_MOV | BPF_X: 1354 if (insn->off == BPF_ADDR_SPACE_CAST && 1355 insn->imm == 1U << 16) { 1356 if (dst_reg != src_reg) 1357 /* 32-bit mov */ 1358 emit_mov_reg(&prog, false, dst_reg, src_reg); 1359 /* shl dst_reg, 32 */ 1360 maybe_emit_1mod(&prog, dst_reg, true); 1361 EMIT3(0xC1, add_1reg(0xE0, dst_reg), 32); 1362 1363 /* or dst_reg, user_vm_start */ 1364 maybe_emit_1mod(&prog, dst_reg, true); 1365 if (is_axreg(dst_reg)) 1366 EMIT1_off32(0x0D, user_vm_start >> 32); 1367 else 1368 EMIT2_off32(0x81, add_1reg(0xC8, dst_reg), user_vm_start >> 32); 1369 1370 /* rol dst_reg, 32 */ 1371 maybe_emit_1mod(&prog, dst_reg, true); 1372 EMIT3(0xC1, add_1reg(0xC0, dst_reg), 32); 1373 1374 /* xor r11, r11 */ 1375 EMIT3(0x4D, 0x31, 0xDB); 1376 1377 /* test dst_reg32, dst_reg32; check if lower 32-bit are zero */ 1378 maybe_emit_mod(&prog, dst_reg, dst_reg, false); 1379 EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg)); 1380 1381 /* cmove r11, dst_reg; if so, set dst_reg to zero */ 1382 /* WARNING: Intel swapped src/dst register encoding in CMOVcc !!! */ 1383 maybe_emit_mod(&prog, AUX_REG, dst_reg, true); 1384 EMIT3(0x0F, 0x44, add_2reg(0xC0, AUX_REG, dst_reg)); 1385 break; 1386 } 1387 fallthrough; 1388 case BPF_ALU | BPF_MOV | BPF_X: 1389 if (insn->off == 0) 1390 emit_mov_reg(&prog, 1391 BPF_CLASS(insn->code) == BPF_ALU64, 1392 dst_reg, src_reg); 1393 else 1394 emit_movsx_reg(&prog, insn->off, 1395 BPF_CLASS(insn->code) == BPF_ALU64, 1396 dst_reg, src_reg); 1397 break; 1398 1399 /* neg dst */ 1400 case BPF_ALU | BPF_NEG: 1401 case BPF_ALU64 | BPF_NEG: 1402 maybe_emit_1mod(&prog, dst_reg, 1403 BPF_CLASS(insn->code) == BPF_ALU64); 1404 EMIT2(0xF7, add_1reg(0xD8, dst_reg)); 1405 break; 1406 1407 case BPF_ALU | BPF_ADD | BPF_K: 1408 case BPF_ALU | BPF_SUB | BPF_K: 1409 case BPF_ALU | BPF_AND | BPF_K: 1410 case BPF_ALU | BPF_OR | BPF_K: 1411 case BPF_ALU | BPF_XOR | BPF_K: 1412 case BPF_ALU64 | BPF_ADD | BPF_K: 1413 case BPF_ALU64 | BPF_SUB | BPF_K: 1414 case BPF_ALU64 | BPF_AND | BPF_K: 1415 case BPF_ALU64 | BPF_OR | BPF_K: 1416 case BPF_ALU64 | BPF_XOR | BPF_K: 1417 maybe_emit_1mod(&prog, dst_reg, 1418 BPF_CLASS(insn->code) == BPF_ALU64); 1419 1420 /* 1421 * b3 holds 'normal' opcode, b2 short form only valid 1422 * in case dst is eax/rax. 1423 */ 1424 switch (BPF_OP(insn->code)) { 1425 case BPF_ADD: 1426 b3 = 0xC0; 1427 b2 = 0x05; 1428 break; 1429 case BPF_SUB: 1430 b3 = 0xE8; 1431 b2 = 0x2D; 1432 break; 1433 case BPF_AND: 1434 b3 = 0xE0; 1435 b2 = 0x25; 1436 break; 1437 case BPF_OR: 1438 b3 = 0xC8; 1439 b2 = 0x0D; 1440 break; 1441 case BPF_XOR: 1442 b3 = 0xF0; 1443 b2 = 0x35; 1444 break; 1445 } 1446 1447 if (is_imm8(imm32)) 1448 EMIT3(0x83, add_1reg(b3, dst_reg), imm32); 1449 else if (is_axreg(dst_reg)) 1450 EMIT1_off32(b2, imm32); 1451 else 1452 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32); 1453 break; 1454 1455 case BPF_ALU64 | BPF_MOV | BPF_K: 1456 case BPF_ALU | BPF_MOV | BPF_K: 1457 emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64, 1458 dst_reg, imm32); 1459 break; 1460 1461 case BPF_LD | BPF_IMM | BPF_DW: 1462 emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm); 1463 insn++; 1464 i++; 1465 break; 1466 1467 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */ 1468 case BPF_ALU | BPF_MOD | BPF_X: 1469 case BPF_ALU | BPF_DIV | BPF_X: 1470 case BPF_ALU | BPF_MOD | BPF_K: 1471 case BPF_ALU | BPF_DIV | BPF_K: 1472 case BPF_ALU64 | BPF_MOD | BPF_X: 1473 case BPF_ALU64 | BPF_DIV | BPF_X: 1474 case BPF_ALU64 | BPF_MOD | BPF_K: 1475 case BPF_ALU64 | BPF_DIV | BPF_K: { 1476 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; 1477 1478 if (dst_reg != BPF_REG_0) 1479 EMIT1(0x50); /* push rax */ 1480 if (dst_reg != BPF_REG_3) 1481 EMIT1(0x52); /* push rdx */ 1482 1483 if (BPF_SRC(insn->code) == BPF_X) { 1484 if (src_reg == BPF_REG_0 || 1485 src_reg == BPF_REG_3) { 1486 /* mov r11, src_reg */ 1487 EMIT_mov(AUX_REG, src_reg); 1488 src_reg = AUX_REG; 1489 } 1490 } else { 1491 /* mov r11, imm32 */ 1492 EMIT3_off32(0x49, 0xC7, 0xC3, imm32); 1493 src_reg = AUX_REG; 1494 } 1495 1496 if (dst_reg != BPF_REG_0) 1497 /* mov rax, dst_reg */ 1498 emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg); 1499 1500 if (insn->off == 0) { 1501 /* 1502 * xor edx, edx 1503 * equivalent to 'xor rdx, rdx', but one byte less 1504 */ 1505 EMIT2(0x31, 0xd2); 1506 1507 /* div src_reg */ 1508 maybe_emit_1mod(&prog, src_reg, is64); 1509 EMIT2(0xF7, add_1reg(0xF0, src_reg)); 1510 } else { 1511 if (BPF_CLASS(insn->code) == BPF_ALU) 1512 EMIT1(0x99); /* cdq */ 1513 else 1514 EMIT2(0x48, 0x99); /* cqo */ 1515 1516 /* idiv src_reg */ 1517 maybe_emit_1mod(&prog, src_reg, is64); 1518 EMIT2(0xF7, add_1reg(0xF8, src_reg)); 1519 } 1520 1521 if (BPF_OP(insn->code) == BPF_MOD && 1522 dst_reg != BPF_REG_3) 1523 /* mov dst_reg, rdx */ 1524 emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3); 1525 else if (BPF_OP(insn->code) == BPF_DIV && 1526 dst_reg != BPF_REG_0) 1527 /* mov dst_reg, rax */ 1528 emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0); 1529 1530 if (dst_reg != BPF_REG_3) 1531 EMIT1(0x5A); /* pop rdx */ 1532 if (dst_reg != BPF_REG_0) 1533 EMIT1(0x58); /* pop rax */ 1534 break; 1535 } 1536 1537 case BPF_ALU | BPF_MUL | BPF_K: 1538 case BPF_ALU64 | BPF_MUL | BPF_K: 1539 maybe_emit_mod(&prog, dst_reg, dst_reg, 1540 BPF_CLASS(insn->code) == BPF_ALU64); 1541 1542 if (is_imm8(imm32)) 1543 /* imul dst_reg, dst_reg, imm8 */ 1544 EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg), 1545 imm32); 1546 else 1547 /* imul dst_reg, dst_reg, imm32 */ 1548 EMIT2_off32(0x69, 1549 add_2reg(0xC0, dst_reg, dst_reg), 1550 imm32); 1551 break; 1552 1553 case BPF_ALU | BPF_MUL | BPF_X: 1554 case BPF_ALU64 | BPF_MUL | BPF_X: 1555 maybe_emit_mod(&prog, src_reg, dst_reg, 1556 BPF_CLASS(insn->code) == BPF_ALU64); 1557 1558 /* imul dst_reg, src_reg */ 1559 EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg)); 1560 break; 1561 1562 /* Shifts */ 1563 case BPF_ALU | BPF_LSH | BPF_K: 1564 case BPF_ALU | BPF_RSH | BPF_K: 1565 case BPF_ALU | BPF_ARSH | BPF_K: 1566 case BPF_ALU64 | BPF_LSH | BPF_K: 1567 case BPF_ALU64 | BPF_RSH | BPF_K: 1568 case BPF_ALU64 | BPF_ARSH | BPF_K: 1569 maybe_emit_1mod(&prog, dst_reg, 1570 BPF_CLASS(insn->code) == BPF_ALU64); 1571 1572 b3 = simple_alu_opcodes[BPF_OP(insn->code)]; 1573 if (imm32 == 1) 1574 EMIT2(0xD1, add_1reg(b3, dst_reg)); 1575 else 1576 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32); 1577 break; 1578 1579 case BPF_ALU | BPF_LSH | BPF_X: 1580 case BPF_ALU | BPF_RSH | BPF_X: 1581 case BPF_ALU | BPF_ARSH | BPF_X: 1582 case BPF_ALU64 | BPF_LSH | BPF_X: 1583 case BPF_ALU64 | BPF_RSH | BPF_X: 1584 case BPF_ALU64 | BPF_ARSH | BPF_X: 1585 /* BMI2 shifts aren't better when shift count is already in rcx */ 1586 if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) { 1587 /* shrx/sarx/shlx dst_reg, dst_reg, src_reg */ 1588 bool w = (BPF_CLASS(insn->code) == BPF_ALU64); 1589 u8 op; 1590 1591 switch (BPF_OP(insn->code)) { 1592 case BPF_LSH: 1593 op = 1; /* prefix 0x66 */ 1594 break; 1595 case BPF_RSH: 1596 op = 3; /* prefix 0xf2 */ 1597 break; 1598 case BPF_ARSH: 1599 op = 2; /* prefix 0xf3 */ 1600 break; 1601 } 1602 1603 emit_shiftx(&prog, dst_reg, src_reg, w, op); 1604 1605 break; 1606 } 1607 1608 if (src_reg != BPF_REG_4) { /* common case */ 1609 /* Check for bad case when dst_reg == rcx */ 1610 if (dst_reg == BPF_REG_4) { 1611 /* mov r11, dst_reg */ 1612 EMIT_mov(AUX_REG, dst_reg); 1613 dst_reg = AUX_REG; 1614 } else { 1615 EMIT1(0x51); /* push rcx */ 1616 } 1617 /* mov rcx, src_reg */ 1618 EMIT_mov(BPF_REG_4, src_reg); 1619 } 1620 1621 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */ 1622 maybe_emit_1mod(&prog, dst_reg, 1623 BPF_CLASS(insn->code) == BPF_ALU64); 1624 1625 b3 = simple_alu_opcodes[BPF_OP(insn->code)]; 1626 EMIT2(0xD3, add_1reg(b3, dst_reg)); 1627 1628 if (src_reg != BPF_REG_4) { 1629 if (insn->dst_reg == BPF_REG_4) 1630 /* mov dst_reg, r11 */ 1631 EMIT_mov(insn->dst_reg, AUX_REG); 1632 else 1633 EMIT1(0x59); /* pop rcx */ 1634 } 1635 1636 break; 1637 1638 case BPF_ALU | BPF_END | BPF_FROM_BE: 1639 case BPF_ALU64 | BPF_END | BPF_FROM_LE: 1640 switch (imm32) { 1641 case 16: 1642 /* Emit 'ror %ax, 8' to swap lower 2 bytes */ 1643 EMIT1(0x66); 1644 if (is_ereg(dst_reg)) 1645 EMIT1(0x41); 1646 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8); 1647 1648 /* Emit 'movzwl eax, ax' */ 1649 if (is_ereg(dst_reg)) 1650 EMIT3(0x45, 0x0F, 0xB7); 1651 else 1652 EMIT2(0x0F, 0xB7); 1653 EMIT1(add_2reg(0xC0, dst_reg, dst_reg)); 1654 break; 1655 case 32: 1656 /* Emit 'bswap eax' to swap lower 4 bytes */ 1657 if (is_ereg(dst_reg)) 1658 EMIT2(0x41, 0x0F); 1659 else 1660 EMIT1(0x0F); 1661 EMIT1(add_1reg(0xC8, dst_reg)); 1662 break; 1663 case 64: 1664 /* Emit 'bswap rax' to swap 8 bytes */ 1665 EMIT3(add_1mod(0x48, dst_reg), 0x0F, 1666 add_1reg(0xC8, dst_reg)); 1667 break; 1668 } 1669 break; 1670 1671 case BPF_ALU | BPF_END | BPF_FROM_LE: 1672 switch (imm32) { 1673 case 16: 1674 /* 1675 * Emit 'movzwl eax, ax' to zero extend 16-bit 1676 * into 64 bit 1677 */ 1678 if (is_ereg(dst_reg)) 1679 EMIT3(0x45, 0x0F, 0xB7); 1680 else 1681 EMIT2(0x0F, 0xB7); 1682 EMIT1(add_2reg(0xC0, dst_reg, dst_reg)); 1683 break; 1684 case 32: 1685 /* Emit 'mov eax, eax' to clear upper 32-bits */ 1686 if (is_ereg(dst_reg)) 1687 EMIT1(0x45); 1688 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg)); 1689 break; 1690 case 64: 1691 /* nop */ 1692 break; 1693 } 1694 break; 1695 1696 /* speculation barrier */ 1697 case BPF_ST | BPF_NOSPEC: 1698 EMIT_LFENCE(); 1699 break; 1700 1701 /* ST: *(u8*)(dst_reg + off) = imm */ 1702 case BPF_ST | BPF_MEM | BPF_B: 1703 if (is_ereg(dst_reg)) 1704 EMIT2(0x41, 0xC6); 1705 else 1706 EMIT1(0xC6); 1707 goto st; 1708 case BPF_ST | BPF_MEM | BPF_H: 1709 if (is_ereg(dst_reg)) 1710 EMIT3(0x66, 0x41, 0xC7); 1711 else 1712 EMIT2(0x66, 0xC7); 1713 goto st; 1714 case BPF_ST | BPF_MEM | BPF_W: 1715 if (is_ereg(dst_reg)) 1716 EMIT2(0x41, 0xC7); 1717 else 1718 EMIT1(0xC7); 1719 goto st; 1720 case BPF_ST | BPF_MEM | BPF_DW: 1721 EMIT2(add_1mod(0x48, dst_reg), 0xC7); 1722 1723 st: if (is_imm8(insn->off)) 1724 EMIT2(add_1reg(0x40, dst_reg), insn->off); 1725 else 1726 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off); 1727 1728 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code))); 1729 break; 1730 1731 /* STX: *(u8*)(dst_reg + off) = src_reg */ 1732 case BPF_STX | BPF_MEM | BPF_B: 1733 case BPF_STX | BPF_MEM | BPF_H: 1734 case BPF_STX | BPF_MEM | BPF_W: 1735 case BPF_STX | BPF_MEM | BPF_DW: 1736 emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off); 1737 break; 1738 1739 case BPF_ST | BPF_PROBE_MEM32 | BPF_B: 1740 case BPF_ST | BPF_PROBE_MEM32 | BPF_H: 1741 case BPF_ST | BPF_PROBE_MEM32 | BPF_W: 1742 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW: 1743 start_of_ldx = prog; 1744 emit_st_r12(&prog, BPF_SIZE(insn->code), dst_reg, insn->off, insn->imm); 1745 goto populate_extable; 1746 1747 /* LDX: dst_reg = *(u8*)(src_reg + r12 + off) */ 1748 case BPF_LDX | BPF_PROBE_MEM32 | BPF_B: 1749 case BPF_LDX | BPF_PROBE_MEM32 | BPF_H: 1750 case BPF_LDX | BPF_PROBE_MEM32 | BPF_W: 1751 case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW: 1752 case BPF_STX | BPF_PROBE_MEM32 | BPF_B: 1753 case BPF_STX | BPF_PROBE_MEM32 | BPF_H: 1754 case BPF_STX | BPF_PROBE_MEM32 | BPF_W: 1755 case BPF_STX | BPF_PROBE_MEM32 | BPF_DW: 1756 start_of_ldx = prog; 1757 if (BPF_CLASS(insn->code) == BPF_LDX) 1758 emit_ldx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off); 1759 else 1760 emit_stx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off); 1761 populate_extable: 1762 { 1763 struct exception_table_entry *ex; 1764 u8 *_insn = image + proglen + (start_of_ldx - temp); 1765 s64 delta; 1766 1767 if (!bpf_prog->aux->extable) 1768 break; 1769 1770 if (excnt >= bpf_prog->aux->num_exentries) { 1771 pr_err("mem32 extable bug\n"); 1772 return -EFAULT; 1773 } 1774 ex = &bpf_prog->aux->extable[excnt++]; 1775 1776 delta = _insn - (u8 *)&ex->insn; 1777 /* switch ex to rw buffer for writes */ 1778 ex = (void *)rw_image + ((void *)ex - (void *)image); 1779 1780 ex->insn = delta; 1781 1782 ex->data = EX_TYPE_BPF; 1783 1784 ex->fixup = (prog - start_of_ldx) | 1785 ((BPF_CLASS(insn->code) == BPF_LDX ? reg2pt_regs[dst_reg] : DONT_CLEAR) << 8); 1786 } 1787 break; 1788 1789 /* LDX: dst_reg = *(u8*)(src_reg + off) */ 1790 case BPF_LDX | BPF_MEM | BPF_B: 1791 case BPF_LDX | BPF_PROBE_MEM | BPF_B: 1792 case BPF_LDX | BPF_MEM | BPF_H: 1793 case BPF_LDX | BPF_PROBE_MEM | BPF_H: 1794 case BPF_LDX | BPF_MEM | BPF_W: 1795 case BPF_LDX | BPF_PROBE_MEM | BPF_W: 1796 case BPF_LDX | BPF_MEM | BPF_DW: 1797 case BPF_LDX | BPF_PROBE_MEM | BPF_DW: 1798 /* LDXS: dst_reg = *(s8*)(src_reg + off) */ 1799 case BPF_LDX | BPF_MEMSX | BPF_B: 1800 case BPF_LDX | BPF_MEMSX | BPF_H: 1801 case BPF_LDX | BPF_MEMSX | BPF_W: 1802 case BPF_LDX | BPF_PROBE_MEMSX | BPF_B: 1803 case BPF_LDX | BPF_PROBE_MEMSX | BPF_H: 1804 case BPF_LDX | BPF_PROBE_MEMSX | BPF_W: 1805 insn_off = insn->off; 1806 1807 if (BPF_MODE(insn->code) == BPF_PROBE_MEM || 1808 BPF_MODE(insn->code) == BPF_PROBE_MEMSX) { 1809 /* Conservatively check that src_reg + insn->off is a kernel address: 1810 * src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE 1811 * src_reg is used as scratch for src_reg += insn->off and restored 1812 * after emit_ldx if necessary 1813 */ 1814 1815 u64 limit = TASK_SIZE_MAX + PAGE_SIZE; 1816 u8 *end_of_jmp; 1817 1818 /* At end of these emitted checks, insn->off will have been added 1819 * to src_reg, so no need to do relative load with insn->off offset 1820 */ 1821 insn_off = 0; 1822 1823 /* movabsq r11, limit */ 1824 EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG)); 1825 EMIT((u32)limit, 4); 1826 EMIT(limit >> 32, 4); 1827 1828 if (insn->off) { 1829 /* add src_reg, insn->off */ 1830 maybe_emit_1mod(&prog, src_reg, true); 1831 EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off); 1832 } 1833 1834 /* cmp src_reg, r11 */ 1835 maybe_emit_mod(&prog, src_reg, AUX_REG, true); 1836 EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG)); 1837 1838 /* if unsigned '>=', goto load */ 1839 EMIT2(X86_JAE, 0); 1840 end_of_jmp = prog; 1841 1842 /* xor dst_reg, dst_reg */ 1843 emit_mov_imm32(&prog, false, dst_reg, 0); 1844 /* jmp byte_after_ldx */ 1845 EMIT2(0xEB, 0); 1846 1847 /* populate jmp_offset for JAE above to jump to start_of_ldx */ 1848 start_of_ldx = prog; 1849 end_of_jmp[-1] = start_of_ldx - end_of_jmp; 1850 } 1851 if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX || 1852 BPF_MODE(insn->code) == BPF_MEMSX) 1853 emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off); 1854 else 1855 emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off); 1856 if (BPF_MODE(insn->code) == BPF_PROBE_MEM || 1857 BPF_MODE(insn->code) == BPF_PROBE_MEMSX) { 1858 struct exception_table_entry *ex; 1859 u8 *_insn = image + proglen + (start_of_ldx - temp); 1860 s64 delta; 1861 1862 /* populate jmp_offset for JMP above */ 1863 start_of_ldx[-1] = prog - start_of_ldx; 1864 1865 if (insn->off && src_reg != dst_reg) { 1866 /* sub src_reg, insn->off 1867 * Restore src_reg after "add src_reg, insn->off" in prev 1868 * if statement. But if src_reg == dst_reg, emit_ldx 1869 * above already clobbered src_reg, so no need to restore. 1870 * If add src_reg, insn->off was unnecessary, no need to 1871 * restore either. 1872 */ 1873 maybe_emit_1mod(&prog, src_reg, true); 1874 EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off); 1875 } 1876 1877 if (!bpf_prog->aux->extable) 1878 break; 1879 1880 if (excnt >= bpf_prog->aux->num_exentries) { 1881 pr_err("ex gen bug\n"); 1882 return -EFAULT; 1883 } 1884 ex = &bpf_prog->aux->extable[excnt++]; 1885 1886 delta = _insn - (u8 *)&ex->insn; 1887 if (!is_simm32(delta)) { 1888 pr_err("extable->insn doesn't fit into 32-bit\n"); 1889 return -EFAULT; 1890 } 1891 /* switch ex to rw buffer for writes */ 1892 ex = (void *)rw_image + ((void *)ex - (void *)image); 1893 1894 ex->insn = delta; 1895 1896 ex->data = EX_TYPE_BPF; 1897 1898 if (dst_reg > BPF_REG_9) { 1899 pr_err("verifier error\n"); 1900 return -EFAULT; 1901 } 1902 /* 1903 * Compute size of x86 insn and its target dest x86 register. 1904 * ex_handler_bpf() will use lower 8 bits to adjust 1905 * pt_regs->ip to jump over this x86 instruction 1906 * and upper bits to figure out which pt_regs to zero out. 1907 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]" 1908 * of 4 bytes will be ignored and rbx will be zero inited. 1909 */ 1910 ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8); 1911 } 1912 break; 1913 1914 case BPF_STX | BPF_ATOMIC | BPF_W: 1915 case BPF_STX | BPF_ATOMIC | BPF_DW: 1916 if (insn->imm == (BPF_AND | BPF_FETCH) || 1917 insn->imm == (BPF_OR | BPF_FETCH) || 1918 insn->imm == (BPF_XOR | BPF_FETCH)) { 1919 bool is64 = BPF_SIZE(insn->code) == BPF_DW; 1920 u32 real_src_reg = src_reg; 1921 u32 real_dst_reg = dst_reg; 1922 u8 *branch_target; 1923 1924 /* 1925 * Can't be implemented with a single x86 insn. 1926 * Need to do a CMPXCHG loop. 1927 */ 1928 1929 /* Will need RAX as a CMPXCHG operand so save R0 */ 1930 emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0); 1931 if (src_reg == BPF_REG_0) 1932 real_src_reg = BPF_REG_AX; 1933 if (dst_reg == BPF_REG_0) 1934 real_dst_reg = BPF_REG_AX; 1935 1936 branch_target = prog; 1937 /* Load old value */ 1938 emit_ldx(&prog, BPF_SIZE(insn->code), 1939 BPF_REG_0, real_dst_reg, insn->off); 1940 /* 1941 * Perform the (commutative) operation locally, 1942 * put the result in the AUX_REG. 1943 */ 1944 emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0); 1945 maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64); 1946 EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)], 1947 add_2reg(0xC0, AUX_REG, real_src_reg)); 1948 /* Attempt to swap in new value */ 1949 err = emit_atomic(&prog, BPF_CMPXCHG, 1950 real_dst_reg, AUX_REG, 1951 insn->off, 1952 BPF_SIZE(insn->code)); 1953 if (WARN_ON(err)) 1954 return err; 1955 /* 1956 * ZF tells us whether we won the race. If it's 1957 * cleared we need to try again. 1958 */ 1959 EMIT2(X86_JNE, -(prog - branch_target) - 2); 1960 /* Return the pre-modification value */ 1961 emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0); 1962 /* Restore R0 after clobbering RAX */ 1963 emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX); 1964 break; 1965 } 1966 1967 err = emit_atomic(&prog, insn->imm, dst_reg, src_reg, 1968 insn->off, BPF_SIZE(insn->code)); 1969 if (err) 1970 return err; 1971 break; 1972 1973 /* call */ 1974 case BPF_JMP | BPF_CALL: { 1975 int offs; 1976 1977 func = (u8 *) __bpf_call_base + imm32; 1978 if (tail_call_reachable) { 1979 RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth); 1980 if (!imm32) 1981 return -EINVAL; 1982 offs = 7 + x86_call_depth_emit_accounting(&prog, func); 1983 } else { 1984 if (!imm32) 1985 return -EINVAL; 1986 offs = x86_call_depth_emit_accounting(&prog, func); 1987 } 1988 if (emit_call(&prog, func, image + addrs[i - 1] + offs)) 1989 return -EINVAL; 1990 break; 1991 } 1992 1993 case BPF_JMP | BPF_TAIL_CALL: 1994 if (imm32) 1995 emit_bpf_tail_call_direct(bpf_prog, 1996 &bpf_prog->aux->poke_tab[imm32 - 1], 1997 &prog, image + addrs[i - 1], 1998 callee_regs_used, 1999 bpf_prog->aux->stack_depth, 2000 ctx); 2001 else 2002 emit_bpf_tail_call_indirect(bpf_prog, 2003 &prog, 2004 callee_regs_used, 2005 bpf_prog->aux->stack_depth, 2006 image + addrs[i - 1], 2007 ctx); 2008 break; 2009 2010 /* cond jump */ 2011 case BPF_JMP | BPF_JEQ | BPF_X: 2012 case BPF_JMP | BPF_JNE | BPF_X: 2013 case BPF_JMP | BPF_JGT | BPF_X: 2014 case BPF_JMP | BPF_JLT | BPF_X: 2015 case BPF_JMP | BPF_JGE | BPF_X: 2016 case BPF_JMP | BPF_JLE | BPF_X: 2017 case BPF_JMP | BPF_JSGT | BPF_X: 2018 case BPF_JMP | BPF_JSLT | BPF_X: 2019 case BPF_JMP | BPF_JSGE | BPF_X: 2020 case BPF_JMP | BPF_JSLE | BPF_X: 2021 case BPF_JMP32 | BPF_JEQ | BPF_X: 2022 case BPF_JMP32 | BPF_JNE | BPF_X: 2023 case BPF_JMP32 | BPF_JGT | BPF_X: 2024 case BPF_JMP32 | BPF_JLT | BPF_X: 2025 case BPF_JMP32 | BPF_JGE | BPF_X: 2026 case BPF_JMP32 | BPF_JLE | BPF_X: 2027 case BPF_JMP32 | BPF_JSGT | BPF_X: 2028 case BPF_JMP32 | BPF_JSLT | BPF_X: 2029 case BPF_JMP32 | BPF_JSGE | BPF_X: 2030 case BPF_JMP32 | BPF_JSLE | BPF_X: 2031 /* cmp dst_reg, src_reg */ 2032 maybe_emit_mod(&prog, dst_reg, src_reg, 2033 BPF_CLASS(insn->code) == BPF_JMP); 2034 EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg)); 2035 goto emit_cond_jmp; 2036 2037 case BPF_JMP | BPF_JSET | BPF_X: 2038 case BPF_JMP32 | BPF_JSET | BPF_X: 2039 /* test dst_reg, src_reg */ 2040 maybe_emit_mod(&prog, dst_reg, src_reg, 2041 BPF_CLASS(insn->code) == BPF_JMP); 2042 EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg)); 2043 goto emit_cond_jmp; 2044 2045 case BPF_JMP | BPF_JSET | BPF_K: 2046 case BPF_JMP32 | BPF_JSET | BPF_K: 2047 /* test dst_reg, imm32 */ 2048 maybe_emit_1mod(&prog, dst_reg, 2049 BPF_CLASS(insn->code) == BPF_JMP); 2050 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32); 2051 goto emit_cond_jmp; 2052 2053 case BPF_JMP | BPF_JEQ | BPF_K: 2054 case BPF_JMP | BPF_JNE | BPF_K: 2055 case BPF_JMP | BPF_JGT | BPF_K: 2056 case BPF_JMP | BPF_JLT | BPF_K: 2057 case BPF_JMP | BPF_JGE | BPF_K: 2058 case BPF_JMP | BPF_JLE | BPF_K: 2059 case BPF_JMP | BPF_JSGT | BPF_K: 2060 case BPF_JMP | BPF_JSLT | BPF_K: 2061 case BPF_JMP | BPF_JSGE | BPF_K: 2062 case BPF_JMP | BPF_JSLE | BPF_K: 2063 case BPF_JMP32 | BPF_JEQ | BPF_K: 2064 case BPF_JMP32 | BPF_JNE | BPF_K: 2065 case BPF_JMP32 | BPF_JGT | BPF_K: 2066 case BPF_JMP32 | BPF_JLT | BPF_K: 2067 case BPF_JMP32 | BPF_JGE | BPF_K: 2068 case BPF_JMP32 | BPF_JLE | BPF_K: 2069 case BPF_JMP32 | BPF_JSGT | BPF_K: 2070 case BPF_JMP32 | BPF_JSLT | BPF_K: 2071 case BPF_JMP32 | BPF_JSGE | BPF_K: 2072 case BPF_JMP32 | BPF_JSLE | BPF_K: 2073 /* test dst_reg, dst_reg to save one extra byte */ 2074 if (imm32 == 0) { 2075 maybe_emit_mod(&prog, dst_reg, dst_reg, 2076 BPF_CLASS(insn->code) == BPF_JMP); 2077 EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg)); 2078 goto emit_cond_jmp; 2079 } 2080 2081 /* cmp dst_reg, imm8/32 */ 2082 maybe_emit_1mod(&prog, dst_reg, 2083 BPF_CLASS(insn->code) == BPF_JMP); 2084 2085 if (is_imm8(imm32)) 2086 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32); 2087 else 2088 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32); 2089 2090 emit_cond_jmp: /* Convert BPF opcode to x86 */ 2091 switch (BPF_OP(insn->code)) { 2092 case BPF_JEQ: 2093 jmp_cond = X86_JE; 2094 break; 2095 case BPF_JSET: 2096 case BPF_JNE: 2097 jmp_cond = X86_JNE; 2098 break; 2099 case BPF_JGT: 2100 /* GT is unsigned '>', JA in x86 */ 2101 jmp_cond = X86_JA; 2102 break; 2103 case BPF_JLT: 2104 /* LT is unsigned '<', JB in x86 */ 2105 jmp_cond = X86_JB; 2106 break; 2107 case BPF_JGE: 2108 /* GE is unsigned '>=', JAE in x86 */ 2109 jmp_cond = X86_JAE; 2110 break; 2111 case BPF_JLE: 2112 /* LE is unsigned '<=', JBE in x86 */ 2113 jmp_cond = X86_JBE; 2114 break; 2115 case BPF_JSGT: 2116 /* Signed '>', GT in x86 */ 2117 jmp_cond = X86_JG; 2118 break; 2119 case BPF_JSLT: 2120 /* Signed '<', LT in x86 */ 2121 jmp_cond = X86_JL; 2122 break; 2123 case BPF_JSGE: 2124 /* Signed '>=', GE in x86 */ 2125 jmp_cond = X86_JGE; 2126 break; 2127 case BPF_JSLE: 2128 /* Signed '<=', LE in x86 */ 2129 jmp_cond = X86_JLE; 2130 break; 2131 default: /* to silence GCC warning */ 2132 return -EFAULT; 2133 } 2134 jmp_offset = addrs[i + insn->off] - addrs[i]; 2135 if (is_imm8(jmp_offset)) { 2136 if (jmp_padding) { 2137 /* To keep the jmp_offset valid, the extra bytes are 2138 * padded before the jump insn, so we subtract the 2139 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF. 2140 * 2141 * If the previous pass already emits an imm8 2142 * jmp_cond, then this BPF insn won't shrink, so 2143 * "nops" is 0. 2144 * 2145 * On the other hand, if the previous pass emits an 2146 * imm32 jmp_cond, the extra 4 bytes(*) is padded to 2147 * keep the image from shrinking further. 2148 * 2149 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond 2150 * is 2 bytes, so the size difference is 4 bytes. 2151 */ 2152 nops = INSN_SZ_DIFF - 2; 2153 if (nops != 0 && nops != 4) { 2154 pr_err("unexpected jmp_cond padding: %d bytes\n", 2155 nops); 2156 return -EFAULT; 2157 } 2158 emit_nops(&prog, nops); 2159 } 2160 EMIT2(jmp_cond, jmp_offset); 2161 } else if (is_simm32(jmp_offset)) { 2162 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset); 2163 } else { 2164 pr_err("cond_jmp gen bug %llx\n", jmp_offset); 2165 return -EFAULT; 2166 } 2167 2168 break; 2169 2170 case BPF_JMP | BPF_JA: 2171 case BPF_JMP32 | BPF_JA: 2172 if (BPF_CLASS(insn->code) == BPF_JMP) { 2173 if (insn->off == -1) 2174 /* -1 jmp instructions will always jump 2175 * backwards two bytes. Explicitly handling 2176 * this case avoids wasting too many passes 2177 * when there are long sequences of replaced 2178 * dead code. 2179 */ 2180 jmp_offset = -2; 2181 else 2182 jmp_offset = addrs[i + insn->off] - addrs[i]; 2183 } else { 2184 if (insn->imm == -1) 2185 jmp_offset = -2; 2186 else 2187 jmp_offset = addrs[i + insn->imm] - addrs[i]; 2188 } 2189 2190 if (!jmp_offset) { 2191 /* 2192 * If jmp_padding is enabled, the extra nops will 2193 * be inserted. Otherwise, optimize out nop jumps. 2194 */ 2195 if (jmp_padding) { 2196 /* There are 3 possible conditions. 2197 * (1) This BPF_JA is already optimized out in 2198 * the previous run, so there is no need 2199 * to pad any extra byte (0 byte). 2200 * (2) The previous pass emits an imm8 jmp, 2201 * so we pad 2 bytes to match the previous 2202 * insn size. 2203 * (3) Similarly, the previous pass emits an 2204 * imm32 jmp, and 5 bytes is padded. 2205 */ 2206 nops = INSN_SZ_DIFF; 2207 if (nops != 0 && nops != 2 && nops != 5) { 2208 pr_err("unexpected nop jump padding: %d bytes\n", 2209 nops); 2210 return -EFAULT; 2211 } 2212 emit_nops(&prog, nops); 2213 } 2214 break; 2215 } 2216 emit_jmp: 2217 if (is_imm8(jmp_offset)) { 2218 if (jmp_padding) { 2219 /* To avoid breaking jmp_offset, the extra bytes 2220 * are padded before the actual jmp insn, so 2221 * 2 bytes is subtracted from INSN_SZ_DIFF. 2222 * 2223 * If the previous pass already emits an imm8 2224 * jmp, there is nothing to pad (0 byte). 2225 * 2226 * If it emits an imm32 jmp (5 bytes) previously 2227 * and now an imm8 jmp (2 bytes), then we pad 2228 * (5 - 2 = 3) bytes to stop the image from 2229 * shrinking further. 2230 */ 2231 nops = INSN_SZ_DIFF - 2; 2232 if (nops != 0 && nops != 3) { 2233 pr_err("unexpected jump padding: %d bytes\n", 2234 nops); 2235 return -EFAULT; 2236 } 2237 emit_nops(&prog, INSN_SZ_DIFF - 2); 2238 } 2239 EMIT2(0xEB, jmp_offset); 2240 } else if (is_simm32(jmp_offset)) { 2241 EMIT1_off32(0xE9, jmp_offset); 2242 } else { 2243 pr_err("jmp gen bug %llx\n", jmp_offset); 2244 return -EFAULT; 2245 } 2246 break; 2247 2248 case BPF_JMP | BPF_EXIT: 2249 if (seen_exit) { 2250 jmp_offset = ctx->cleanup_addr - addrs[i]; 2251 goto emit_jmp; 2252 } 2253 seen_exit = true; 2254 /* Update cleanup_addr */ 2255 ctx->cleanup_addr = proglen; 2256 if (bpf_prog->aux->exception_boundary) { 2257 pop_callee_regs(&prog, all_callee_regs_used); 2258 pop_r12(&prog); 2259 } else { 2260 pop_callee_regs(&prog, callee_regs_used); 2261 if (arena_vm_start) 2262 pop_r12(&prog); 2263 } 2264 EMIT1(0xC9); /* leave */ 2265 emit_return(&prog, image + addrs[i - 1] + (prog - temp)); 2266 break; 2267 2268 default: 2269 /* 2270 * By design x86-64 JIT should support all BPF instructions. 2271 * This error will be seen if new instruction was added 2272 * to the interpreter, but not to the JIT, or if there is 2273 * junk in bpf_prog. 2274 */ 2275 pr_err("bpf_jit: unknown opcode %02x\n", insn->code); 2276 return -EINVAL; 2277 } 2278 2279 ilen = prog - temp; 2280 if (ilen > BPF_MAX_INSN_SIZE) { 2281 pr_err("bpf_jit: fatal insn size error\n"); 2282 return -EFAULT; 2283 } 2284 2285 if (image) { 2286 /* 2287 * When populating the image, assert that: 2288 * 2289 * i) We do not write beyond the allocated space, and 2290 * ii) addrs[i] did not change from the prior run, in order 2291 * to validate assumptions made for computing branch 2292 * displacements. 2293 */ 2294 if (unlikely(proglen + ilen > oldproglen || 2295 proglen + ilen != addrs[i])) { 2296 pr_err("bpf_jit: fatal error\n"); 2297 return -EFAULT; 2298 } 2299 memcpy(rw_image + proglen, temp, ilen); 2300 } 2301 proglen += ilen; 2302 addrs[i] = proglen; 2303 prog = temp; 2304 } 2305 2306 if (image && excnt != bpf_prog->aux->num_exentries) { 2307 pr_err("extable is not populated\n"); 2308 return -EFAULT; 2309 } 2310 return proglen; 2311 } 2312 2313 static void clean_stack_garbage(const struct btf_func_model *m, 2314 u8 **pprog, int nr_stack_slots, 2315 int stack_size) 2316 { 2317 int arg_size, off; 2318 u8 *prog; 2319 2320 /* Generally speaking, the compiler will pass the arguments 2321 * on-stack with "push" instruction, which will take 8-byte 2322 * on the stack. In this case, there won't be garbage values 2323 * while we copy the arguments from origin stack frame to current 2324 * in BPF_DW. 2325 * 2326 * However, sometimes the compiler will only allocate 4-byte on 2327 * the stack for the arguments. For now, this case will only 2328 * happen if there is only one argument on-stack and its size 2329 * not more than 4 byte. In this case, there will be garbage 2330 * values on the upper 4-byte where we store the argument on 2331 * current stack frame. 2332 * 2333 * arguments on origin stack: 2334 * 2335 * stack_arg_1(4-byte) xxx(4-byte) 2336 * 2337 * what we copy: 2338 * 2339 * stack_arg_1(8-byte): stack_arg_1(origin) xxx 2340 * 2341 * and the xxx is the garbage values which we should clean here. 2342 */ 2343 if (nr_stack_slots != 1) 2344 return; 2345 2346 /* the size of the last argument */ 2347 arg_size = m->arg_size[m->nr_args - 1]; 2348 if (arg_size <= 4) { 2349 off = -(stack_size - 4); 2350 prog = *pprog; 2351 /* mov DWORD PTR [rbp + off], 0 */ 2352 if (!is_imm8(off)) 2353 EMIT2_off32(0xC7, 0x85, off); 2354 else 2355 EMIT3(0xC7, 0x45, off); 2356 EMIT(0, 4); 2357 *pprog = prog; 2358 } 2359 } 2360 2361 /* get the count of the regs that are used to pass arguments */ 2362 static int get_nr_used_regs(const struct btf_func_model *m) 2363 { 2364 int i, arg_regs, nr_used_regs = 0; 2365 2366 for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) { 2367 arg_regs = (m->arg_size[i] + 7) / 8; 2368 if (nr_used_regs + arg_regs <= 6) 2369 nr_used_regs += arg_regs; 2370 2371 if (nr_used_regs >= 6) 2372 break; 2373 } 2374 2375 return nr_used_regs; 2376 } 2377 2378 static void save_args(const struct btf_func_model *m, u8 **prog, 2379 int stack_size, bool for_call_origin) 2380 { 2381 int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0; 2382 int i, j; 2383 2384 /* Store function arguments to stack. 2385 * For a function that accepts two pointers the sequence will be: 2386 * mov QWORD PTR [rbp-0x10],rdi 2387 * mov QWORD PTR [rbp-0x8],rsi 2388 */ 2389 for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) { 2390 arg_regs = (m->arg_size[i] + 7) / 8; 2391 2392 /* According to the research of Yonghong, struct members 2393 * should be all in register or all on the stack. 2394 * Meanwhile, the compiler will pass the argument on regs 2395 * if the remaining regs can hold the argument. 2396 * 2397 * Disorder of the args can happen. For example: 2398 * 2399 * struct foo_struct { 2400 * long a; 2401 * int b; 2402 * }; 2403 * int foo(char, char, char, char, char, struct foo_struct, 2404 * char); 2405 * 2406 * the arg1-5,arg7 will be passed by regs, and arg6 will 2407 * by stack. 2408 */ 2409 if (nr_regs + arg_regs > 6) { 2410 /* copy function arguments from origin stack frame 2411 * into current stack frame. 2412 * 2413 * The starting address of the arguments on-stack 2414 * is: 2415 * rbp + 8(push rbp) + 2416 * 8(return addr of origin call) + 2417 * 8(return addr of the caller) 2418 * which means: rbp + 24 2419 */ 2420 for (j = 0; j < arg_regs; j++) { 2421 emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 2422 nr_stack_slots * 8 + 0x18); 2423 emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0, 2424 -stack_size); 2425 2426 if (!nr_stack_slots) 2427 first_off = stack_size; 2428 stack_size -= 8; 2429 nr_stack_slots++; 2430 } 2431 } else { 2432 /* Only copy the arguments on-stack to current 2433 * 'stack_size' and ignore the regs, used to 2434 * prepare the arguments on-stack for origin call. 2435 */ 2436 if (for_call_origin) { 2437 nr_regs += arg_regs; 2438 continue; 2439 } 2440 2441 /* copy the arguments from regs into stack */ 2442 for (j = 0; j < arg_regs; j++) { 2443 emit_stx(prog, BPF_DW, BPF_REG_FP, 2444 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs, 2445 -stack_size); 2446 stack_size -= 8; 2447 nr_regs++; 2448 } 2449 } 2450 } 2451 2452 clean_stack_garbage(m, prog, nr_stack_slots, first_off); 2453 } 2454 2455 static void restore_regs(const struct btf_func_model *m, u8 **prog, 2456 int stack_size) 2457 { 2458 int i, j, arg_regs, nr_regs = 0; 2459 2460 /* Restore function arguments from stack. 2461 * For a function that accepts two pointers the sequence will be: 2462 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10] 2463 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8] 2464 * 2465 * The logic here is similar to what we do in save_args() 2466 */ 2467 for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) { 2468 arg_regs = (m->arg_size[i] + 7) / 8; 2469 if (nr_regs + arg_regs <= 6) { 2470 for (j = 0; j < arg_regs; j++) { 2471 emit_ldx(prog, BPF_DW, 2472 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs, 2473 BPF_REG_FP, 2474 -stack_size); 2475 stack_size -= 8; 2476 nr_regs++; 2477 } 2478 } else { 2479 stack_size -= 8 * arg_regs; 2480 } 2481 2482 if (nr_regs >= 6) 2483 break; 2484 } 2485 } 2486 2487 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog, 2488 struct bpf_tramp_link *l, int stack_size, 2489 int run_ctx_off, bool save_ret, 2490 void *image, void *rw_image) 2491 { 2492 u8 *prog = *pprog; 2493 u8 *jmp_insn; 2494 int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie); 2495 struct bpf_prog *p = l->link.prog; 2496 u64 cookie = l->cookie; 2497 2498 /* mov rdi, cookie */ 2499 emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie); 2500 2501 /* Prepare struct bpf_tramp_run_ctx. 2502 * 2503 * bpf_tramp_run_ctx is already preserved by 2504 * arch_prepare_bpf_trampoline(). 2505 * 2506 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi 2507 */ 2508 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off); 2509 2510 /* arg1: mov rdi, progs[i] */ 2511 emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p); 2512 /* arg2: lea rsi, [rbp - ctx_cookie_off] */ 2513 if (!is_imm8(-run_ctx_off)) 2514 EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off); 2515 else 2516 EMIT4(0x48, 0x8D, 0x75, -run_ctx_off); 2517 2518 if (emit_rsb_call(&prog, bpf_trampoline_enter(p), image + (prog - (u8 *)rw_image))) 2519 return -EINVAL; 2520 /* remember prog start time returned by __bpf_prog_enter */ 2521 emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0); 2522 2523 /* if (__bpf_prog_enter*(prog) == 0) 2524 * goto skip_exec_of_prog; 2525 */ 2526 EMIT3(0x48, 0x85, 0xC0); /* test rax,rax */ 2527 /* emit 2 nops that will be replaced with JE insn */ 2528 jmp_insn = prog; 2529 emit_nops(&prog, 2); 2530 2531 /* arg1: lea rdi, [rbp - stack_size] */ 2532 if (!is_imm8(-stack_size)) 2533 EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size); 2534 else 2535 EMIT4(0x48, 0x8D, 0x7D, -stack_size); 2536 /* arg2: progs[i]->insnsi for interpreter */ 2537 if (!p->jited) 2538 emit_mov_imm64(&prog, BPF_REG_2, 2539 (long) p->insnsi >> 32, 2540 (u32) (long) p->insnsi); 2541 /* call JITed bpf program or interpreter */ 2542 if (emit_rsb_call(&prog, p->bpf_func, image + (prog - (u8 *)rw_image))) 2543 return -EINVAL; 2544 2545 /* 2546 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return 2547 * of the previous call which is then passed on the stack to 2548 * the next BPF program. 2549 * 2550 * BPF_TRAMP_FENTRY trampoline may need to return the return 2551 * value of BPF_PROG_TYPE_STRUCT_OPS prog. 2552 */ 2553 if (save_ret) 2554 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8); 2555 2556 /* replace 2 nops with JE insn, since jmp target is known */ 2557 jmp_insn[0] = X86_JE; 2558 jmp_insn[1] = prog - jmp_insn - 2; 2559 2560 /* arg1: mov rdi, progs[i] */ 2561 emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p); 2562 /* arg2: mov rsi, rbx <- start time in nsec */ 2563 emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6); 2564 /* arg3: lea rdx, [rbp - run_ctx_off] */ 2565 if (!is_imm8(-run_ctx_off)) 2566 EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off); 2567 else 2568 EMIT4(0x48, 0x8D, 0x55, -run_ctx_off); 2569 if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image))) 2570 return -EINVAL; 2571 2572 *pprog = prog; 2573 return 0; 2574 } 2575 2576 static void emit_align(u8 **pprog, u32 align) 2577 { 2578 u8 *target, *prog = *pprog; 2579 2580 target = PTR_ALIGN(prog, align); 2581 if (target != prog) 2582 emit_nops(&prog, target - prog); 2583 2584 *pprog = prog; 2585 } 2586 2587 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond) 2588 { 2589 u8 *prog = *pprog; 2590 s64 offset; 2591 2592 offset = func - (ip + 2 + 4); 2593 if (!is_simm32(offset)) { 2594 pr_err("Target %p is out of range\n", func); 2595 return -EINVAL; 2596 } 2597 EMIT2_off32(0x0F, jmp_cond + 0x10, offset); 2598 *pprog = prog; 2599 return 0; 2600 } 2601 2602 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog, 2603 struct bpf_tramp_links *tl, int stack_size, 2604 int run_ctx_off, bool save_ret, 2605 void *image, void *rw_image) 2606 { 2607 int i; 2608 u8 *prog = *pprog; 2609 2610 for (i = 0; i < tl->nr_links; i++) { 2611 if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, 2612 run_ctx_off, save_ret, image, rw_image)) 2613 return -EINVAL; 2614 } 2615 *pprog = prog; 2616 return 0; 2617 } 2618 2619 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog, 2620 struct bpf_tramp_links *tl, int stack_size, 2621 int run_ctx_off, u8 **branches, 2622 void *image, void *rw_image) 2623 { 2624 u8 *prog = *pprog; 2625 int i; 2626 2627 /* The first fmod_ret program will receive a garbage return value. 2628 * Set this to 0 to avoid confusing the program. 2629 */ 2630 emit_mov_imm32(&prog, false, BPF_REG_0, 0); 2631 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8); 2632 for (i = 0; i < tl->nr_links; i++) { 2633 if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true, 2634 image, rw_image)) 2635 return -EINVAL; 2636 2637 /* mod_ret prog stored return value into [rbp - 8]. Emit: 2638 * if (*(u64 *)(rbp - 8) != 0) 2639 * goto do_fexit; 2640 */ 2641 /* cmp QWORD PTR [rbp - 0x8], 0x0 */ 2642 EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00); 2643 2644 /* Save the location of the branch and Generate 6 nops 2645 * (4 bytes for an offset and 2 bytes for the jump) These nops 2646 * are replaced with a conditional jump once do_fexit (i.e. the 2647 * start of the fexit invocation) is finalized. 2648 */ 2649 branches[i] = prog; 2650 emit_nops(&prog, 4 + 2); 2651 } 2652 2653 *pprog = prog; 2654 return 0; 2655 } 2656 2657 /* Example: 2658 * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev); 2659 * its 'struct btf_func_model' will be nr_args=2 2660 * The assembly code when eth_type_trans is executing after trampoline: 2661 * 2662 * push rbp 2663 * mov rbp, rsp 2664 * sub rsp, 16 // space for skb and dev 2665 * push rbx // temp regs to pass start time 2666 * mov qword ptr [rbp - 16], rdi // save skb pointer to stack 2667 * mov qword ptr [rbp - 8], rsi // save dev pointer to stack 2668 * call __bpf_prog_enter // rcu_read_lock and preempt_disable 2669 * mov rbx, rax // remember start time in bpf stats are enabled 2670 * lea rdi, [rbp - 16] // R1==ctx of bpf prog 2671 * call addr_of_jited_FENTRY_prog 2672 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off 2673 * mov rsi, rbx // prog start time 2674 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math 2675 * mov rdi, qword ptr [rbp - 16] // restore skb pointer from stack 2676 * mov rsi, qword ptr [rbp - 8] // restore dev pointer from stack 2677 * pop rbx 2678 * leave 2679 * ret 2680 * 2681 * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be 2682 * replaced with 'call generated_bpf_trampoline'. When it returns 2683 * eth_type_trans will continue executing with original skb and dev pointers. 2684 * 2685 * The assembly code when eth_type_trans is called from trampoline: 2686 * 2687 * push rbp 2688 * mov rbp, rsp 2689 * sub rsp, 24 // space for skb, dev, return value 2690 * push rbx // temp regs to pass start time 2691 * mov qword ptr [rbp - 24], rdi // save skb pointer to stack 2692 * mov qword ptr [rbp - 16], rsi // save dev pointer to stack 2693 * call __bpf_prog_enter // rcu_read_lock and preempt_disable 2694 * mov rbx, rax // remember start time if bpf stats are enabled 2695 * lea rdi, [rbp - 24] // R1==ctx of bpf prog 2696 * call addr_of_jited_FENTRY_prog // bpf prog can access skb and dev 2697 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off 2698 * mov rsi, rbx // prog start time 2699 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math 2700 * mov rdi, qword ptr [rbp - 24] // restore skb pointer from stack 2701 * mov rsi, qword ptr [rbp - 16] // restore dev pointer from stack 2702 * call eth_type_trans+5 // execute body of eth_type_trans 2703 * mov qword ptr [rbp - 8], rax // save return value 2704 * call __bpf_prog_enter // rcu_read_lock and preempt_disable 2705 * mov rbx, rax // remember start time in bpf stats are enabled 2706 * lea rdi, [rbp - 24] // R1==ctx of bpf prog 2707 * call addr_of_jited_FEXIT_prog // bpf prog can access skb, dev, return value 2708 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off 2709 * mov rsi, rbx // prog start time 2710 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math 2711 * mov rax, qword ptr [rbp - 8] // restore eth_type_trans's return value 2712 * pop rbx 2713 * leave 2714 * add rsp, 8 // skip eth_type_trans's frame 2715 * ret // return to its caller 2716 */ 2717 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image, 2718 void *rw_image_end, void *image, 2719 const struct btf_func_model *m, u32 flags, 2720 struct bpf_tramp_links *tlinks, 2721 void *func_addr) 2722 { 2723 int i, ret, nr_regs = m->nr_args, stack_size = 0; 2724 int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off; 2725 struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY]; 2726 struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT]; 2727 struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN]; 2728 void *orig_call = func_addr; 2729 u8 **branches = NULL; 2730 u8 *prog; 2731 bool save_ret; 2732 2733 /* 2734 * F_INDIRECT is only compatible with F_RET_FENTRY_RET, it is 2735 * explicitly incompatible with F_CALL_ORIG | F_SKIP_FRAME | F_IP_ARG 2736 * because @func_addr. 2737 */ 2738 WARN_ON_ONCE((flags & BPF_TRAMP_F_INDIRECT) && 2739 (flags & ~(BPF_TRAMP_F_INDIRECT | BPF_TRAMP_F_RET_FENTRY_RET))); 2740 2741 /* extra registers for struct arguments */ 2742 for (i = 0; i < m->nr_args; i++) { 2743 if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) 2744 nr_regs += (m->arg_size[i] + 7) / 8 - 1; 2745 } 2746 2747 /* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6 2748 * are passed through regs, the remains are through stack. 2749 */ 2750 if (nr_regs > MAX_BPF_FUNC_ARGS) 2751 return -ENOTSUPP; 2752 2753 /* Generated trampoline stack layout: 2754 * 2755 * RBP + 8 [ return address ] 2756 * RBP + 0 [ RBP ] 2757 * 2758 * RBP - 8 [ return value ] BPF_TRAMP_F_CALL_ORIG or 2759 * BPF_TRAMP_F_RET_FENTRY_RET flags 2760 * 2761 * [ reg_argN ] always 2762 * [ ... ] 2763 * RBP - regs_off [ reg_arg1 ] program's ctx pointer 2764 * 2765 * RBP - nregs_off [ regs count ] always 2766 * 2767 * RBP - ip_off [ traced function ] BPF_TRAMP_F_IP_ARG flag 2768 * 2769 * RBP - rbx_off [ rbx value ] always 2770 * 2771 * RBP - run_ctx_off [ bpf_tramp_run_ctx ] 2772 * 2773 * [ stack_argN ] BPF_TRAMP_F_CALL_ORIG 2774 * [ ... ] 2775 * [ stack_arg2 ] 2776 * RBP - arg_stack_off [ stack_arg1 ] 2777 * RSP [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX 2778 */ 2779 2780 /* room for return value of orig_call or fentry prog */ 2781 save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET); 2782 if (save_ret) 2783 stack_size += 8; 2784 2785 stack_size += nr_regs * 8; 2786 regs_off = stack_size; 2787 2788 /* regs count */ 2789 stack_size += 8; 2790 nregs_off = stack_size; 2791 2792 if (flags & BPF_TRAMP_F_IP_ARG) 2793 stack_size += 8; /* room for IP address argument */ 2794 2795 ip_off = stack_size; 2796 2797 stack_size += 8; 2798 rbx_off = stack_size; 2799 2800 stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7; 2801 run_ctx_off = stack_size; 2802 2803 if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) { 2804 /* the space that used to pass arguments on-stack */ 2805 stack_size += (nr_regs - get_nr_used_regs(m)) * 8; 2806 /* make sure the stack pointer is 16-byte aligned if we 2807 * need pass arguments on stack, which means 2808 * [stack_size + 8(rbp) + 8(rip) + 8(origin rip)] 2809 * should be 16-byte aligned. Following code depend on 2810 * that stack_size is already 8-byte aligned. 2811 */ 2812 stack_size += (stack_size % 16) ? 0 : 8; 2813 } 2814 2815 arg_stack_off = stack_size; 2816 2817 if (flags & BPF_TRAMP_F_SKIP_FRAME) { 2818 /* skip patched call instruction and point orig_call to actual 2819 * body of the kernel function. 2820 */ 2821 if (is_endbr(*(u32 *)orig_call)) 2822 orig_call += ENDBR_INSN_SIZE; 2823 orig_call += X86_PATCH_SIZE; 2824 } 2825 2826 prog = rw_image; 2827 2828 if (flags & BPF_TRAMP_F_INDIRECT) { 2829 /* 2830 * Indirect call for bpf_struct_ops 2831 */ 2832 emit_cfi(&prog, cfi_get_func_hash(func_addr)); 2833 } else { 2834 /* 2835 * Direct-call fentry stub, as such it needs accounting for the 2836 * __fentry__ call. 2837 */ 2838 x86_call_depth_emit_accounting(&prog, NULL); 2839 } 2840 EMIT1(0x55); /* push rbp */ 2841 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */ 2842 if (!is_imm8(stack_size)) { 2843 /* sub rsp, stack_size */ 2844 EMIT3_off32(0x48, 0x81, 0xEC, stack_size); 2845 } else { 2846 /* sub rsp, stack_size */ 2847 EMIT4(0x48, 0x83, 0xEC, stack_size); 2848 } 2849 if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) 2850 EMIT1(0x50); /* push rax */ 2851 /* mov QWORD PTR [rbp - rbx_off], rbx */ 2852 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off); 2853 2854 /* Store number of argument registers of the traced function: 2855 * mov rax, nr_regs 2856 * mov QWORD PTR [rbp - nregs_off], rax 2857 */ 2858 emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs); 2859 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off); 2860 2861 if (flags & BPF_TRAMP_F_IP_ARG) { 2862 /* Store IP address of the traced function: 2863 * movabsq rax, func_addr 2864 * mov QWORD PTR [rbp - ip_off], rax 2865 */ 2866 emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr); 2867 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off); 2868 } 2869 2870 save_args(m, &prog, regs_off, false); 2871 2872 if (flags & BPF_TRAMP_F_CALL_ORIG) { 2873 /* arg1: mov rdi, im */ 2874 emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im); 2875 if (emit_rsb_call(&prog, __bpf_tramp_enter, 2876 image + (prog - (u8 *)rw_image))) { 2877 ret = -EINVAL; 2878 goto cleanup; 2879 } 2880 } 2881 2882 if (fentry->nr_links) { 2883 if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off, 2884 flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image)) 2885 return -EINVAL; 2886 } 2887 2888 if (fmod_ret->nr_links) { 2889 branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *), 2890 GFP_KERNEL); 2891 if (!branches) 2892 return -ENOMEM; 2893 2894 if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off, 2895 run_ctx_off, branches, image, rw_image)) { 2896 ret = -EINVAL; 2897 goto cleanup; 2898 } 2899 } 2900 2901 if (flags & BPF_TRAMP_F_CALL_ORIG) { 2902 restore_regs(m, &prog, regs_off); 2903 save_args(m, &prog, arg_stack_off, true); 2904 2905 if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) { 2906 /* Before calling the original function, restore the 2907 * tail_call_cnt from stack to rax. 2908 */ 2909 RESTORE_TAIL_CALL_CNT(stack_size); 2910 } 2911 2912 if (flags & BPF_TRAMP_F_ORIG_STACK) { 2913 emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8); 2914 EMIT2(0xff, 0xd3); /* call *rbx */ 2915 } else { 2916 /* call original function */ 2917 if (emit_rsb_call(&prog, orig_call, image + (prog - (u8 *)rw_image))) { 2918 ret = -EINVAL; 2919 goto cleanup; 2920 } 2921 } 2922 /* remember return value in a stack for bpf prog to access */ 2923 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8); 2924 im->ip_after_call = image + (prog - (u8 *)rw_image); 2925 emit_nops(&prog, X86_PATCH_SIZE); 2926 } 2927 2928 if (fmod_ret->nr_links) { 2929 /* From Intel 64 and IA-32 Architectures Optimization 2930 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler 2931 * Coding Rule 11: All branch targets should be 16-byte 2932 * aligned. 2933 */ 2934 emit_align(&prog, 16); 2935 /* Update the branches saved in invoke_bpf_mod_ret with the 2936 * aligned address of do_fexit. 2937 */ 2938 for (i = 0; i < fmod_ret->nr_links; i++) { 2939 emit_cond_near_jump(&branches[i], image + (prog - (u8 *)rw_image), 2940 image + (branches[i] - (u8 *)rw_image), X86_JNE); 2941 } 2942 } 2943 2944 if (fexit->nr_links) { 2945 if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, 2946 false, image, rw_image)) { 2947 ret = -EINVAL; 2948 goto cleanup; 2949 } 2950 } 2951 2952 if (flags & BPF_TRAMP_F_RESTORE_REGS) 2953 restore_regs(m, &prog, regs_off); 2954 2955 /* This needs to be done regardless. If there were fmod_ret programs, 2956 * the return value is only updated on the stack and still needs to be 2957 * restored to R0. 2958 */ 2959 if (flags & BPF_TRAMP_F_CALL_ORIG) { 2960 im->ip_epilogue = image + (prog - (u8 *)rw_image); 2961 /* arg1: mov rdi, im */ 2962 emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im); 2963 if (emit_rsb_call(&prog, __bpf_tramp_exit, image + (prog - (u8 *)rw_image))) { 2964 ret = -EINVAL; 2965 goto cleanup; 2966 } 2967 } else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) { 2968 /* Before running the original function, restore the 2969 * tail_call_cnt from stack to rax. 2970 */ 2971 RESTORE_TAIL_CALL_CNT(stack_size); 2972 } 2973 2974 /* restore return value of orig_call or fentry prog back into RAX */ 2975 if (save_ret) 2976 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8); 2977 2978 emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off); 2979 EMIT1(0xC9); /* leave */ 2980 if (flags & BPF_TRAMP_F_SKIP_FRAME) { 2981 /* skip our return address and return to parent */ 2982 EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */ 2983 } 2984 emit_return(&prog, image + (prog - (u8 *)rw_image)); 2985 /* Make sure the trampoline generation logic doesn't overflow */ 2986 if (WARN_ON_ONCE(prog > (u8 *)rw_image_end - BPF_INSN_SAFETY)) { 2987 ret = -EFAULT; 2988 goto cleanup; 2989 } 2990 ret = prog - (u8 *)rw_image + BPF_INSN_SAFETY; 2991 2992 cleanup: 2993 kfree(branches); 2994 return ret; 2995 } 2996 2997 void *arch_alloc_bpf_trampoline(unsigned int size) 2998 { 2999 return bpf_prog_pack_alloc(size, jit_fill_hole); 3000 } 3001 3002 void arch_free_bpf_trampoline(void *image, unsigned int size) 3003 { 3004 bpf_prog_pack_free(image, size); 3005 } 3006 3007 void arch_protect_bpf_trampoline(void *image, unsigned int size) 3008 { 3009 } 3010 3011 void arch_unprotect_bpf_trampoline(void *image, unsigned int size) 3012 { 3013 } 3014 3015 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end, 3016 const struct btf_func_model *m, u32 flags, 3017 struct bpf_tramp_links *tlinks, 3018 void *func_addr) 3019 { 3020 void *rw_image, *tmp; 3021 int ret; 3022 u32 size = image_end - image; 3023 3024 /* rw_image doesn't need to be in module memory range, so we can 3025 * use kvmalloc. 3026 */ 3027 rw_image = kvmalloc(size, GFP_KERNEL); 3028 if (!rw_image) 3029 return -ENOMEM; 3030 3031 ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m, 3032 flags, tlinks, func_addr); 3033 if (ret < 0) 3034 goto out; 3035 3036 tmp = bpf_arch_text_copy(image, rw_image, size); 3037 if (IS_ERR(tmp)) 3038 ret = PTR_ERR(tmp); 3039 out: 3040 kvfree(rw_image); 3041 return ret; 3042 } 3043 3044 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags, 3045 struct bpf_tramp_links *tlinks, void *func_addr) 3046 { 3047 struct bpf_tramp_image im; 3048 void *image; 3049 int ret; 3050 3051 /* Allocate a temporary buffer for __arch_prepare_bpf_trampoline(). 3052 * This will NOT cause fragmentation in direct map, as we do not 3053 * call set_memory_*() on this buffer. 3054 * 3055 * We cannot use kvmalloc here, because we need image to be in 3056 * module memory range. 3057 */ 3058 image = bpf_jit_alloc_exec(PAGE_SIZE); 3059 if (!image) 3060 return -ENOMEM; 3061 3062 ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image, 3063 m, flags, tlinks, func_addr); 3064 bpf_jit_free_exec(image); 3065 return ret; 3066 } 3067 3068 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf) 3069 { 3070 u8 *jg_reloc, *prog = *pprog; 3071 int pivot, err, jg_bytes = 1; 3072 s64 jg_offset; 3073 3074 if (a == b) { 3075 /* Leaf node of recursion, i.e. not a range of indices 3076 * anymore. 3077 */ 3078 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */ 3079 if (!is_simm32(progs[a])) 3080 return -1; 3081 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), 3082 progs[a]); 3083 err = emit_cond_near_jump(&prog, /* je func */ 3084 (void *)progs[a], image + (prog - buf), 3085 X86_JE); 3086 if (err) 3087 return err; 3088 3089 emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf)); 3090 3091 *pprog = prog; 3092 return 0; 3093 } 3094 3095 /* Not a leaf node, so we pivot, and recursively descend into 3096 * the lower and upper ranges. 3097 */ 3098 pivot = (b - a) / 2; 3099 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */ 3100 if (!is_simm32(progs[a + pivot])) 3101 return -1; 3102 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]); 3103 3104 if (pivot > 2) { /* jg upper_part */ 3105 /* Require near jump. */ 3106 jg_bytes = 4; 3107 EMIT2_off32(0x0F, X86_JG + 0x10, 0); 3108 } else { 3109 EMIT2(X86_JG, 0); 3110 } 3111 jg_reloc = prog; 3112 3113 err = emit_bpf_dispatcher(&prog, a, a + pivot, /* emit lower_part */ 3114 progs, image, buf); 3115 if (err) 3116 return err; 3117 3118 /* From Intel 64 and IA-32 Architectures Optimization 3119 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler 3120 * Coding Rule 11: All branch targets should be 16-byte 3121 * aligned. 3122 */ 3123 emit_align(&prog, 16); 3124 jg_offset = prog - jg_reloc; 3125 emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes); 3126 3127 err = emit_bpf_dispatcher(&prog, a + pivot + 1, /* emit upper_part */ 3128 b, progs, image, buf); 3129 if (err) 3130 return err; 3131 3132 *pprog = prog; 3133 return 0; 3134 } 3135 3136 static int cmp_ips(const void *a, const void *b) 3137 { 3138 const s64 *ipa = a; 3139 const s64 *ipb = b; 3140 3141 if (*ipa > *ipb) 3142 return 1; 3143 if (*ipa < *ipb) 3144 return -1; 3145 return 0; 3146 } 3147 3148 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs) 3149 { 3150 u8 *prog = buf; 3151 3152 sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL); 3153 return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf); 3154 } 3155 3156 struct x64_jit_data { 3157 struct bpf_binary_header *rw_header; 3158 struct bpf_binary_header *header; 3159 int *addrs; 3160 u8 *image; 3161 int proglen; 3162 struct jit_context ctx; 3163 }; 3164 3165 #define MAX_PASSES 20 3166 #define PADDING_PASSES (MAX_PASSES - 5) 3167 3168 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) 3169 { 3170 struct bpf_binary_header *rw_header = NULL; 3171 struct bpf_binary_header *header = NULL; 3172 struct bpf_prog *tmp, *orig_prog = prog; 3173 struct x64_jit_data *jit_data; 3174 int proglen, oldproglen = 0; 3175 struct jit_context ctx = {}; 3176 bool tmp_blinded = false; 3177 bool extra_pass = false; 3178 bool padding = false; 3179 u8 *rw_image = NULL; 3180 u8 *image = NULL; 3181 int *addrs; 3182 int pass; 3183 int i; 3184 3185 if (!prog->jit_requested) 3186 return orig_prog; 3187 3188 tmp = bpf_jit_blind_constants(prog); 3189 /* 3190 * If blinding was requested and we failed during blinding, 3191 * we must fall back to the interpreter. 3192 */ 3193 if (IS_ERR(tmp)) 3194 return orig_prog; 3195 if (tmp != prog) { 3196 tmp_blinded = true; 3197 prog = tmp; 3198 } 3199 3200 jit_data = prog->aux->jit_data; 3201 if (!jit_data) { 3202 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL); 3203 if (!jit_data) { 3204 prog = orig_prog; 3205 goto out; 3206 } 3207 prog->aux->jit_data = jit_data; 3208 } 3209 addrs = jit_data->addrs; 3210 if (addrs) { 3211 ctx = jit_data->ctx; 3212 oldproglen = jit_data->proglen; 3213 image = jit_data->image; 3214 header = jit_data->header; 3215 rw_header = jit_data->rw_header; 3216 rw_image = (void *)rw_header + ((void *)image - (void *)header); 3217 extra_pass = true; 3218 padding = true; 3219 goto skip_init_addrs; 3220 } 3221 addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL); 3222 if (!addrs) { 3223 prog = orig_prog; 3224 goto out_addrs; 3225 } 3226 3227 /* 3228 * Before first pass, make a rough estimation of addrs[] 3229 * each BPF instruction is translated to less than 64 bytes 3230 */ 3231 for (proglen = 0, i = 0; i <= prog->len; i++) { 3232 proglen += 64; 3233 addrs[i] = proglen; 3234 } 3235 ctx.cleanup_addr = proglen; 3236 skip_init_addrs: 3237 3238 /* 3239 * JITed image shrinks with every pass and the loop iterates 3240 * until the image stops shrinking. Very large BPF programs 3241 * may converge on the last pass. In such case do one more 3242 * pass to emit the final image. 3243 */ 3244 for (pass = 0; pass < MAX_PASSES || image; pass++) { 3245 if (!padding && pass >= PADDING_PASSES) 3246 padding = true; 3247 proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding); 3248 if (proglen <= 0) { 3249 out_image: 3250 image = NULL; 3251 if (header) { 3252 bpf_arch_text_copy(&header->size, &rw_header->size, 3253 sizeof(rw_header->size)); 3254 bpf_jit_binary_pack_free(header, rw_header); 3255 } 3256 /* Fall back to interpreter mode */ 3257 prog = orig_prog; 3258 if (extra_pass) { 3259 prog->bpf_func = NULL; 3260 prog->jited = 0; 3261 prog->jited_len = 0; 3262 } 3263 goto out_addrs; 3264 } 3265 if (image) { 3266 if (proglen != oldproglen) { 3267 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n", 3268 proglen, oldproglen); 3269 goto out_image; 3270 } 3271 break; 3272 } 3273 if (proglen == oldproglen) { 3274 /* 3275 * The number of entries in extable is the number of BPF_LDX 3276 * insns that access kernel memory via "pointer to BTF type". 3277 * The verifier changed their opcode from LDX|MEM|size 3278 * to LDX|PROBE_MEM|size to make JITing easier. 3279 */ 3280 u32 align = __alignof__(struct exception_table_entry); 3281 u32 extable_size = prog->aux->num_exentries * 3282 sizeof(struct exception_table_entry); 3283 3284 /* allocate module memory for x86 insns and extable */ 3285 header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size, 3286 &image, align, &rw_header, &rw_image, 3287 jit_fill_hole); 3288 if (!header) { 3289 prog = orig_prog; 3290 goto out_addrs; 3291 } 3292 prog->aux->extable = (void *) image + roundup(proglen, align); 3293 } 3294 oldproglen = proglen; 3295 cond_resched(); 3296 } 3297 3298 if (bpf_jit_enable > 1) 3299 bpf_jit_dump(prog->len, proglen, pass + 1, rw_image); 3300 3301 if (image) { 3302 if (!prog->is_func || extra_pass) { 3303 /* 3304 * bpf_jit_binary_pack_finalize fails in two scenarios: 3305 * 1) header is not pointing to proper module memory; 3306 * 2) the arch doesn't support bpf_arch_text_copy(). 3307 * 3308 * Both cases are serious bugs and justify WARN_ON. 3309 */ 3310 if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) { 3311 /* header has been freed */ 3312 header = NULL; 3313 goto out_image; 3314 } 3315 3316 bpf_tail_call_direct_fixup(prog); 3317 } else { 3318 jit_data->addrs = addrs; 3319 jit_data->ctx = ctx; 3320 jit_data->proglen = proglen; 3321 jit_data->image = image; 3322 jit_data->header = header; 3323 jit_data->rw_header = rw_header; 3324 } 3325 /* 3326 * ctx.prog_offset is used when CFI preambles put code *before* 3327 * the function. See emit_cfi(). For FineIBT specifically this code 3328 * can also be executed and bpf_prog_kallsyms_add() will 3329 * generate an additional symbol to cover this, hence also 3330 * decrement proglen. 3331 */ 3332 prog->bpf_func = (void *)image + cfi_get_offset(); 3333 prog->jited = 1; 3334 prog->jited_len = proglen - cfi_get_offset(); 3335 } else { 3336 prog = orig_prog; 3337 } 3338 3339 if (!image || !prog->is_func || extra_pass) { 3340 if (image) 3341 bpf_prog_fill_jited_linfo(prog, addrs + 1); 3342 out_addrs: 3343 kvfree(addrs); 3344 kfree(jit_data); 3345 prog->aux->jit_data = NULL; 3346 } 3347 out: 3348 if (tmp_blinded) 3349 bpf_jit_prog_release_other(prog, prog == orig_prog ? 3350 tmp : orig_prog); 3351 return prog; 3352 } 3353 3354 bool bpf_jit_supports_kfunc_call(void) 3355 { 3356 return true; 3357 } 3358 3359 void *bpf_arch_text_copy(void *dst, void *src, size_t len) 3360 { 3361 if (text_poke_copy(dst, src, len) == NULL) 3362 return ERR_PTR(-EINVAL); 3363 return dst; 3364 } 3365 3366 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */ 3367 bool bpf_jit_supports_subprog_tailcalls(void) 3368 { 3369 return true; 3370 } 3371 3372 void bpf_jit_free(struct bpf_prog *prog) 3373 { 3374 if (prog->jited) { 3375 struct x64_jit_data *jit_data = prog->aux->jit_data; 3376 struct bpf_binary_header *hdr; 3377 3378 /* 3379 * If we fail the final pass of JIT (from jit_subprogs), 3380 * the program may not be finalized yet. Call finalize here 3381 * before freeing it. 3382 */ 3383 if (jit_data) { 3384 bpf_jit_binary_pack_finalize(prog, jit_data->header, 3385 jit_data->rw_header); 3386 kvfree(jit_data->addrs); 3387 kfree(jit_data); 3388 } 3389 prog->bpf_func = (void *)prog->bpf_func - cfi_get_offset(); 3390 hdr = bpf_jit_binary_pack_hdr(prog); 3391 bpf_jit_binary_pack_free(hdr, NULL); 3392 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog)); 3393 } 3394 3395 bpf_prog_unlock_free(prog); 3396 } 3397 3398 bool bpf_jit_supports_exceptions(void) 3399 { 3400 /* We unwind through both kernel frames (starting from within bpf_throw 3401 * call) and BPF frames. Therefore we require ORC unwinder to be enabled 3402 * to walk kernel frames and reach BPF frames in the stack trace. 3403 */ 3404 return IS_ENABLED(CONFIG_UNWINDER_ORC); 3405 } 3406 3407 void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3408 { 3409 #if defined(CONFIG_UNWINDER_ORC) 3410 struct unwind_state state; 3411 unsigned long addr; 3412 3413 for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state); 3414 unwind_next_frame(&state)) { 3415 addr = unwind_get_return_address(&state); 3416 if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp)) 3417 break; 3418 } 3419 return; 3420 #endif 3421 WARN(1, "verification of programs using bpf_throw should have failed\n"); 3422 } 3423 3424 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke, 3425 struct bpf_prog *new, struct bpf_prog *old) 3426 { 3427 u8 *old_addr, *new_addr, *old_bypass_addr; 3428 int ret; 3429 3430 old_bypass_addr = old ? NULL : poke->bypass_addr; 3431 old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL; 3432 new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL; 3433 3434 /* 3435 * On program loading or teardown, the program's kallsym entry 3436 * might not be in place, so we use __bpf_arch_text_poke to skip 3437 * the kallsyms check. 3438 */ 3439 if (new) { 3440 ret = __bpf_arch_text_poke(poke->tailcall_target, 3441 BPF_MOD_JUMP, 3442 old_addr, new_addr); 3443 BUG_ON(ret < 0); 3444 if (!old) { 3445 ret = __bpf_arch_text_poke(poke->tailcall_bypass, 3446 BPF_MOD_JUMP, 3447 poke->bypass_addr, 3448 NULL); 3449 BUG_ON(ret < 0); 3450 } 3451 } else { 3452 ret = __bpf_arch_text_poke(poke->tailcall_bypass, 3453 BPF_MOD_JUMP, 3454 old_bypass_addr, 3455 poke->bypass_addr); 3456 BUG_ON(ret < 0); 3457 /* let other CPUs finish the execution of program 3458 * so that it will not possible to expose them 3459 * to invalid nop, stack unwind, nop state 3460 */ 3461 if (!ret) 3462 synchronize_rcu(); 3463 ret = __bpf_arch_text_poke(poke->tailcall_target, 3464 BPF_MOD_JUMP, 3465 old_addr, NULL); 3466 BUG_ON(ret < 0); 3467 } 3468 } 3469 3470 bool bpf_jit_supports_arena(void) 3471 { 3472 return true; 3473 } 3474 3475 bool bpf_jit_supports_ptr_xchg(void) 3476 { 3477 return true; 3478 } 3479