1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * BPF JIT compiler for ARM64 4 * 5 * Copyright (C) 2014-2016 Zi Shen Lim <zlim.lnx@gmail.com> 6 */ 7 8 #define pr_fmt(fmt) "bpf_jit: " fmt 9 10 #include <linux/bitfield.h> 11 #include <linux/bpf.h> 12 #include <linux/filter.h> 13 #include <linux/printk.h> 14 #include <linux/slab.h> 15 16 #include <asm/byteorder.h> 17 #include <asm/cacheflush.h> 18 #include <asm/debug-monitors.h> 19 #include <asm/insn.h> 20 #include <asm/set_memory.h> 21 22 #include "bpf_jit.h" 23 24 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0) 25 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1) 26 #define TCALL_CNT (MAX_BPF_JIT_REG + 2) 27 #define TMP_REG_3 (MAX_BPF_JIT_REG + 3) 28 29 /* Map BPF registers to A64 registers */ 30 static const int bpf2a64[] = { 31 /* return value from in-kernel function, and exit value from eBPF */ 32 [BPF_REG_0] = A64_R(7), 33 /* arguments from eBPF program to in-kernel function */ 34 [BPF_REG_1] = A64_R(0), 35 [BPF_REG_2] = A64_R(1), 36 [BPF_REG_3] = A64_R(2), 37 [BPF_REG_4] = A64_R(3), 38 [BPF_REG_5] = A64_R(4), 39 /* callee saved registers that in-kernel function will preserve */ 40 [BPF_REG_6] = A64_R(19), 41 [BPF_REG_7] = A64_R(20), 42 [BPF_REG_8] = A64_R(21), 43 [BPF_REG_9] = A64_R(22), 44 /* read-only frame pointer to access stack */ 45 [BPF_REG_FP] = A64_R(25), 46 /* temporary registers for internal BPF JIT */ 47 [TMP_REG_1] = A64_R(10), 48 [TMP_REG_2] = A64_R(11), 49 [TMP_REG_3] = A64_R(12), 50 /* tail_call_cnt */ 51 [TCALL_CNT] = A64_R(26), 52 /* temporary register for blinding constants */ 53 [BPF_REG_AX] = A64_R(9), 54 }; 55 56 struct jit_ctx { 57 const struct bpf_prog *prog; 58 int idx; 59 int epilogue_offset; 60 int *offset; 61 int exentry_idx; 62 __le32 *image; 63 u32 stack_size; 64 }; 65 66 static inline void emit(const u32 insn, struct jit_ctx *ctx) 67 { 68 if (ctx->image != NULL) 69 ctx->image[ctx->idx] = cpu_to_le32(insn); 70 71 ctx->idx++; 72 } 73 74 static inline void emit_a64_mov_i(const int is64, const int reg, 75 const s32 val, struct jit_ctx *ctx) 76 { 77 u16 hi = val >> 16; 78 u16 lo = val & 0xffff; 79 80 if (hi & 0x8000) { 81 if (hi == 0xffff) { 82 emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx); 83 } else { 84 emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx); 85 if (lo != 0xffff) 86 emit(A64_MOVK(is64, reg, lo, 0), ctx); 87 } 88 } else { 89 emit(A64_MOVZ(is64, reg, lo, 0), ctx); 90 if (hi) 91 emit(A64_MOVK(is64, reg, hi, 16), ctx); 92 } 93 } 94 95 static int i64_i16_blocks(const u64 val, bool inverse) 96 { 97 return (((val >> 0) & 0xffff) != (inverse ? 0xffff : 0x0000)) + 98 (((val >> 16) & 0xffff) != (inverse ? 0xffff : 0x0000)) + 99 (((val >> 32) & 0xffff) != (inverse ? 0xffff : 0x0000)) + 100 (((val >> 48) & 0xffff) != (inverse ? 0xffff : 0x0000)); 101 } 102 103 static inline void emit_a64_mov_i64(const int reg, const u64 val, 104 struct jit_ctx *ctx) 105 { 106 u64 nrm_tmp = val, rev_tmp = ~val; 107 bool inverse; 108 int shift; 109 110 if (!(nrm_tmp >> 32)) 111 return emit_a64_mov_i(0, reg, (u32)val, ctx); 112 113 inverse = i64_i16_blocks(nrm_tmp, true) < i64_i16_blocks(nrm_tmp, false); 114 shift = max(round_down((inverse ? (fls64(rev_tmp) - 1) : 115 (fls64(nrm_tmp) - 1)), 16), 0); 116 if (inverse) 117 emit(A64_MOVN(1, reg, (rev_tmp >> shift) & 0xffff, shift), ctx); 118 else 119 emit(A64_MOVZ(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx); 120 shift -= 16; 121 while (shift >= 0) { 122 if (((nrm_tmp >> shift) & 0xffff) != (inverse ? 0xffff : 0x0000)) 123 emit(A64_MOVK(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx); 124 shift -= 16; 125 } 126 } 127 128 /* 129 * Kernel addresses in the vmalloc space use at most 48 bits, and the 130 * remaining bits are guaranteed to be 0x1. So we can compose the address 131 * with a fixed length movn/movk/movk sequence. 132 */ 133 static inline void emit_addr_mov_i64(const int reg, const u64 val, 134 struct jit_ctx *ctx) 135 { 136 u64 tmp = val; 137 int shift = 0; 138 139 emit(A64_MOVN(1, reg, ~tmp & 0xffff, shift), ctx); 140 while (shift < 32) { 141 tmp >>= 16; 142 shift += 16; 143 emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx); 144 } 145 } 146 147 static inline int bpf2a64_offset(int bpf_insn, int off, 148 const struct jit_ctx *ctx) 149 { 150 /* BPF JMP offset is relative to the next instruction */ 151 bpf_insn++; 152 /* 153 * Whereas arm64 branch instructions encode the offset 154 * from the branch itself, so we must subtract 1 from the 155 * instruction offset. 156 */ 157 return ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1); 158 } 159 160 static void jit_fill_hole(void *area, unsigned int size) 161 { 162 __le32 *ptr; 163 /* We are guaranteed to have aligned memory. */ 164 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32)) 165 *ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT); 166 } 167 168 static inline int epilogue_offset(const struct jit_ctx *ctx) 169 { 170 int to = ctx->epilogue_offset; 171 int from = ctx->idx; 172 173 return to - from; 174 } 175 176 static bool is_addsub_imm(u32 imm) 177 { 178 /* Either imm12 or shifted imm12. */ 179 return !(imm & ~0xfff) || !(imm & ~0xfff000); 180 } 181 182 /* Tail call offset to jump into */ 183 #if IS_ENABLED(CONFIG_ARM64_BTI_KERNEL) 184 #define PROLOGUE_OFFSET 8 185 #else 186 #define PROLOGUE_OFFSET 7 187 #endif 188 189 static int build_prologue(struct jit_ctx *ctx, bool ebpf_from_cbpf) 190 { 191 const struct bpf_prog *prog = ctx->prog; 192 const u8 r6 = bpf2a64[BPF_REG_6]; 193 const u8 r7 = bpf2a64[BPF_REG_7]; 194 const u8 r8 = bpf2a64[BPF_REG_8]; 195 const u8 r9 = bpf2a64[BPF_REG_9]; 196 const u8 fp = bpf2a64[BPF_REG_FP]; 197 const u8 tcc = bpf2a64[TCALL_CNT]; 198 const int idx0 = ctx->idx; 199 int cur_offset; 200 201 /* 202 * BPF prog stack layout 203 * 204 * high 205 * original A64_SP => 0:+-----+ BPF prologue 206 * |FP/LR| 207 * current A64_FP => -16:+-----+ 208 * | ... | callee saved registers 209 * BPF fp register => -64:+-----+ <= (BPF_FP) 210 * | | 211 * | ... | BPF prog stack 212 * | | 213 * +-----+ <= (BPF_FP - prog->aux->stack_depth) 214 * |RSVD | padding 215 * current A64_SP => +-----+ <= (BPF_FP - ctx->stack_size) 216 * | | 217 * | ... | Function call stack 218 * | | 219 * +-----+ 220 * low 221 * 222 */ 223 224 /* BTI landing pad */ 225 if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) 226 emit(A64_BTI_C, ctx); 227 228 /* Save FP and LR registers to stay align with ARM64 AAPCS */ 229 emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx); 230 emit(A64_MOV(1, A64_FP, A64_SP), ctx); 231 232 /* Save callee-saved registers */ 233 emit(A64_PUSH(r6, r7, A64_SP), ctx); 234 emit(A64_PUSH(r8, r9, A64_SP), ctx); 235 emit(A64_PUSH(fp, tcc, A64_SP), ctx); 236 237 /* Set up BPF prog stack base register */ 238 emit(A64_MOV(1, fp, A64_SP), ctx); 239 240 if (!ebpf_from_cbpf) { 241 /* Initialize tail_call_cnt */ 242 emit(A64_MOVZ(1, tcc, 0, 0), ctx); 243 244 cur_offset = ctx->idx - idx0; 245 if (cur_offset != PROLOGUE_OFFSET) { 246 pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n", 247 cur_offset, PROLOGUE_OFFSET); 248 return -1; 249 } 250 251 /* BTI landing pad for the tail call, done with a BR */ 252 if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) 253 emit(A64_BTI_J, ctx); 254 } 255 256 /* Stack must be multiples of 16B */ 257 ctx->stack_size = round_up(prog->aux->stack_depth, 16); 258 259 /* Set up function call stack */ 260 emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 261 return 0; 262 } 263 264 static int out_offset = -1; /* initialized on the first pass of build_body() */ 265 static int emit_bpf_tail_call(struct jit_ctx *ctx) 266 { 267 /* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */ 268 const u8 r2 = bpf2a64[BPF_REG_2]; 269 const u8 r3 = bpf2a64[BPF_REG_3]; 270 271 const u8 tmp = bpf2a64[TMP_REG_1]; 272 const u8 prg = bpf2a64[TMP_REG_2]; 273 const u8 tcc = bpf2a64[TCALL_CNT]; 274 const int idx0 = ctx->idx; 275 #define cur_offset (ctx->idx - idx0) 276 #define jmp_offset (out_offset - (cur_offset)) 277 size_t off; 278 279 /* if (index >= array->map.max_entries) 280 * goto out; 281 */ 282 off = offsetof(struct bpf_array, map.max_entries); 283 emit_a64_mov_i64(tmp, off, ctx); 284 emit(A64_LDR32(tmp, r2, tmp), ctx); 285 emit(A64_MOV(0, r3, r3), ctx); 286 emit(A64_CMP(0, r3, tmp), ctx); 287 emit(A64_B_(A64_COND_CS, jmp_offset), ctx); 288 289 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT) 290 * goto out; 291 * tail_call_cnt++; 292 */ 293 emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx); 294 emit(A64_CMP(1, tcc, tmp), ctx); 295 emit(A64_B_(A64_COND_HI, jmp_offset), ctx); 296 emit(A64_ADD_I(1, tcc, tcc, 1), ctx); 297 298 /* prog = array->ptrs[index]; 299 * if (prog == NULL) 300 * goto out; 301 */ 302 off = offsetof(struct bpf_array, ptrs); 303 emit_a64_mov_i64(tmp, off, ctx); 304 emit(A64_ADD(1, tmp, r2, tmp), ctx); 305 emit(A64_LSL(1, prg, r3, 3), ctx); 306 emit(A64_LDR64(prg, tmp, prg), ctx); 307 emit(A64_CBZ(1, prg, jmp_offset), ctx); 308 309 /* goto *(prog->bpf_func + prologue_offset); */ 310 off = offsetof(struct bpf_prog, bpf_func); 311 emit_a64_mov_i64(tmp, off, ctx); 312 emit(A64_LDR64(tmp, prg, tmp), ctx); 313 emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx); 314 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 315 emit(A64_BR(tmp), ctx); 316 317 /* out: */ 318 if (out_offset == -1) 319 out_offset = cur_offset; 320 if (cur_offset != out_offset) { 321 pr_err_once("tail_call out_offset = %d, expected %d!\n", 322 cur_offset, out_offset); 323 return -1; 324 } 325 return 0; 326 #undef cur_offset 327 #undef jmp_offset 328 } 329 330 static void build_epilogue(struct jit_ctx *ctx) 331 { 332 const u8 r0 = bpf2a64[BPF_REG_0]; 333 const u8 r6 = bpf2a64[BPF_REG_6]; 334 const u8 r7 = bpf2a64[BPF_REG_7]; 335 const u8 r8 = bpf2a64[BPF_REG_8]; 336 const u8 r9 = bpf2a64[BPF_REG_9]; 337 const u8 fp = bpf2a64[BPF_REG_FP]; 338 339 /* We're done with BPF stack */ 340 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 341 342 /* Restore fs (x25) and x26 */ 343 emit(A64_POP(fp, A64_R(26), A64_SP), ctx); 344 345 /* Restore callee-saved register */ 346 emit(A64_POP(r8, r9, A64_SP), ctx); 347 emit(A64_POP(r6, r7, A64_SP), ctx); 348 349 /* Restore FP/LR registers */ 350 emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx); 351 352 /* Set return value */ 353 emit(A64_MOV(1, A64_R(0), r0), ctx); 354 355 emit(A64_RET(A64_LR), ctx); 356 } 357 358 #define BPF_FIXUP_OFFSET_MASK GENMASK(26, 0) 359 #define BPF_FIXUP_REG_MASK GENMASK(31, 27) 360 361 int arm64_bpf_fixup_exception(const struct exception_table_entry *ex, 362 struct pt_regs *regs) 363 { 364 off_t offset = FIELD_GET(BPF_FIXUP_OFFSET_MASK, ex->fixup); 365 int dst_reg = FIELD_GET(BPF_FIXUP_REG_MASK, ex->fixup); 366 367 regs->regs[dst_reg] = 0; 368 regs->pc = (unsigned long)&ex->fixup - offset; 369 return 1; 370 } 371 372 /* For accesses to BTF pointers, add an entry to the exception table */ 373 static int add_exception_handler(const struct bpf_insn *insn, 374 struct jit_ctx *ctx, 375 int dst_reg) 376 { 377 off_t offset; 378 unsigned long pc; 379 struct exception_table_entry *ex; 380 381 if (!ctx->image) 382 /* First pass */ 383 return 0; 384 385 if (BPF_MODE(insn->code) != BPF_PROBE_MEM) 386 return 0; 387 388 if (!ctx->prog->aux->extable || 389 WARN_ON_ONCE(ctx->exentry_idx >= ctx->prog->aux->num_exentries)) 390 return -EINVAL; 391 392 ex = &ctx->prog->aux->extable[ctx->exentry_idx]; 393 pc = (unsigned long)&ctx->image[ctx->idx - 1]; 394 395 offset = pc - (long)&ex->insn; 396 if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN)) 397 return -ERANGE; 398 ex->insn = offset; 399 400 /* 401 * Since the extable follows the program, the fixup offset is always 402 * negative and limited to BPF_JIT_REGION_SIZE. Store a positive value 403 * to keep things simple, and put the destination register in the upper 404 * bits. We don't need to worry about buildtime or runtime sort 405 * modifying the upper bits because the table is already sorted, and 406 * isn't part of the main exception table. 407 */ 408 offset = (long)&ex->fixup - (pc + AARCH64_INSN_SIZE); 409 if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset)) 410 return -ERANGE; 411 412 ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) | 413 FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg); 414 415 ctx->exentry_idx++; 416 return 0; 417 } 418 419 /* JITs an eBPF instruction. 420 * Returns: 421 * 0 - successfully JITed an 8-byte eBPF instruction. 422 * >0 - successfully JITed a 16-byte eBPF instruction. 423 * <0 - failed to JIT. 424 */ 425 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx, 426 bool extra_pass) 427 { 428 const u8 code = insn->code; 429 const u8 dst = bpf2a64[insn->dst_reg]; 430 const u8 src = bpf2a64[insn->src_reg]; 431 const u8 tmp = bpf2a64[TMP_REG_1]; 432 const u8 tmp2 = bpf2a64[TMP_REG_2]; 433 const u8 tmp3 = bpf2a64[TMP_REG_3]; 434 const s16 off = insn->off; 435 const s32 imm = insn->imm; 436 const int i = insn - ctx->prog->insnsi; 437 const bool is64 = BPF_CLASS(code) == BPF_ALU64 || 438 BPF_CLASS(code) == BPF_JMP; 439 const bool isdw = BPF_SIZE(code) == BPF_DW; 440 u8 jmp_cond, reg; 441 s32 jmp_offset; 442 u32 a64_insn; 443 int ret; 444 445 #define check_imm(bits, imm) do { \ 446 if ((((imm) > 0) && ((imm) >> (bits))) || \ 447 (((imm) < 0) && (~(imm) >> (bits)))) { \ 448 pr_info("[%2d] imm=%d(0x%x) out of range\n", \ 449 i, imm, imm); \ 450 return -EINVAL; \ 451 } \ 452 } while (0) 453 #define check_imm19(imm) check_imm(19, imm) 454 #define check_imm26(imm) check_imm(26, imm) 455 456 switch (code) { 457 /* dst = src */ 458 case BPF_ALU | BPF_MOV | BPF_X: 459 case BPF_ALU64 | BPF_MOV | BPF_X: 460 emit(A64_MOV(is64, dst, src), ctx); 461 break; 462 /* dst = dst OP src */ 463 case BPF_ALU | BPF_ADD | BPF_X: 464 case BPF_ALU64 | BPF_ADD | BPF_X: 465 emit(A64_ADD(is64, dst, dst, src), ctx); 466 break; 467 case BPF_ALU | BPF_SUB | BPF_X: 468 case BPF_ALU64 | BPF_SUB | BPF_X: 469 emit(A64_SUB(is64, dst, dst, src), ctx); 470 break; 471 case BPF_ALU | BPF_AND | BPF_X: 472 case BPF_ALU64 | BPF_AND | BPF_X: 473 emit(A64_AND(is64, dst, dst, src), ctx); 474 break; 475 case BPF_ALU | BPF_OR | BPF_X: 476 case BPF_ALU64 | BPF_OR | BPF_X: 477 emit(A64_ORR(is64, dst, dst, src), ctx); 478 break; 479 case BPF_ALU | BPF_XOR | BPF_X: 480 case BPF_ALU64 | BPF_XOR | BPF_X: 481 emit(A64_EOR(is64, dst, dst, src), ctx); 482 break; 483 case BPF_ALU | BPF_MUL | BPF_X: 484 case BPF_ALU64 | BPF_MUL | BPF_X: 485 emit(A64_MUL(is64, dst, dst, src), ctx); 486 break; 487 case BPF_ALU | BPF_DIV | BPF_X: 488 case BPF_ALU64 | BPF_DIV | BPF_X: 489 emit(A64_UDIV(is64, dst, dst, src), ctx); 490 break; 491 case BPF_ALU | BPF_MOD | BPF_X: 492 case BPF_ALU64 | BPF_MOD | BPF_X: 493 emit(A64_UDIV(is64, tmp, dst, src), ctx); 494 emit(A64_MSUB(is64, dst, dst, tmp, src), ctx); 495 break; 496 case BPF_ALU | BPF_LSH | BPF_X: 497 case BPF_ALU64 | BPF_LSH | BPF_X: 498 emit(A64_LSLV(is64, dst, dst, src), ctx); 499 break; 500 case BPF_ALU | BPF_RSH | BPF_X: 501 case BPF_ALU64 | BPF_RSH | BPF_X: 502 emit(A64_LSRV(is64, dst, dst, src), ctx); 503 break; 504 case BPF_ALU | BPF_ARSH | BPF_X: 505 case BPF_ALU64 | BPF_ARSH | BPF_X: 506 emit(A64_ASRV(is64, dst, dst, src), ctx); 507 break; 508 /* dst = -dst */ 509 case BPF_ALU | BPF_NEG: 510 case BPF_ALU64 | BPF_NEG: 511 emit(A64_NEG(is64, dst, dst), ctx); 512 break; 513 /* dst = BSWAP##imm(dst) */ 514 case BPF_ALU | BPF_END | BPF_FROM_LE: 515 case BPF_ALU | BPF_END | BPF_FROM_BE: 516 #ifdef CONFIG_CPU_BIG_ENDIAN 517 if (BPF_SRC(code) == BPF_FROM_BE) 518 goto emit_bswap_uxt; 519 #else /* !CONFIG_CPU_BIG_ENDIAN */ 520 if (BPF_SRC(code) == BPF_FROM_LE) 521 goto emit_bswap_uxt; 522 #endif 523 switch (imm) { 524 case 16: 525 emit(A64_REV16(is64, dst, dst), ctx); 526 /* zero-extend 16 bits into 64 bits */ 527 emit(A64_UXTH(is64, dst, dst), ctx); 528 break; 529 case 32: 530 emit(A64_REV32(is64, dst, dst), ctx); 531 /* upper 32 bits already cleared */ 532 break; 533 case 64: 534 emit(A64_REV64(dst, dst), ctx); 535 break; 536 } 537 break; 538 emit_bswap_uxt: 539 switch (imm) { 540 case 16: 541 /* zero-extend 16 bits into 64 bits */ 542 emit(A64_UXTH(is64, dst, dst), ctx); 543 break; 544 case 32: 545 /* zero-extend 32 bits into 64 bits */ 546 emit(A64_UXTW(is64, dst, dst), ctx); 547 break; 548 case 64: 549 /* nop */ 550 break; 551 } 552 break; 553 /* dst = imm */ 554 case BPF_ALU | BPF_MOV | BPF_K: 555 case BPF_ALU64 | BPF_MOV | BPF_K: 556 emit_a64_mov_i(is64, dst, imm, ctx); 557 break; 558 /* dst = dst OP imm */ 559 case BPF_ALU | BPF_ADD | BPF_K: 560 case BPF_ALU64 | BPF_ADD | BPF_K: 561 if (is_addsub_imm(imm)) { 562 emit(A64_ADD_I(is64, dst, dst, imm), ctx); 563 } else if (is_addsub_imm(-imm)) { 564 emit(A64_SUB_I(is64, dst, dst, -imm), ctx); 565 } else { 566 emit_a64_mov_i(is64, tmp, imm, ctx); 567 emit(A64_ADD(is64, dst, dst, tmp), ctx); 568 } 569 break; 570 case BPF_ALU | BPF_SUB | BPF_K: 571 case BPF_ALU64 | BPF_SUB | BPF_K: 572 if (is_addsub_imm(imm)) { 573 emit(A64_SUB_I(is64, dst, dst, imm), ctx); 574 } else if (is_addsub_imm(-imm)) { 575 emit(A64_ADD_I(is64, dst, dst, -imm), ctx); 576 } else { 577 emit_a64_mov_i(is64, tmp, imm, ctx); 578 emit(A64_SUB(is64, dst, dst, tmp), ctx); 579 } 580 break; 581 case BPF_ALU | BPF_AND | BPF_K: 582 case BPF_ALU64 | BPF_AND | BPF_K: 583 a64_insn = A64_AND_I(is64, dst, dst, imm); 584 if (a64_insn != AARCH64_BREAK_FAULT) { 585 emit(a64_insn, ctx); 586 } else { 587 emit_a64_mov_i(is64, tmp, imm, ctx); 588 emit(A64_AND(is64, dst, dst, tmp), ctx); 589 } 590 break; 591 case BPF_ALU | BPF_OR | BPF_K: 592 case BPF_ALU64 | BPF_OR | BPF_K: 593 a64_insn = A64_ORR_I(is64, dst, dst, imm); 594 if (a64_insn != AARCH64_BREAK_FAULT) { 595 emit(a64_insn, ctx); 596 } else { 597 emit_a64_mov_i(is64, tmp, imm, ctx); 598 emit(A64_ORR(is64, dst, dst, tmp), ctx); 599 } 600 break; 601 case BPF_ALU | BPF_XOR | BPF_K: 602 case BPF_ALU64 | BPF_XOR | BPF_K: 603 a64_insn = A64_EOR_I(is64, dst, dst, imm); 604 if (a64_insn != AARCH64_BREAK_FAULT) { 605 emit(a64_insn, ctx); 606 } else { 607 emit_a64_mov_i(is64, tmp, imm, ctx); 608 emit(A64_EOR(is64, dst, dst, tmp), ctx); 609 } 610 break; 611 case BPF_ALU | BPF_MUL | BPF_K: 612 case BPF_ALU64 | BPF_MUL | BPF_K: 613 emit_a64_mov_i(is64, tmp, imm, ctx); 614 emit(A64_MUL(is64, dst, dst, tmp), ctx); 615 break; 616 case BPF_ALU | BPF_DIV | BPF_K: 617 case BPF_ALU64 | BPF_DIV | BPF_K: 618 emit_a64_mov_i(is64, tmp, imm, ctx); 619 emit(A64_UDIV(is64, dst, dst, tmp), ctx); 620 break; 621 case BPF_ALU | BPF_MOD | BPF_K: 622 case BPF_ALU64 | BPF_MOD | BPF_K: 623 emit_a64_mov_i(is64, tmp2, imm, ctx); 624 emit(A64_UDIV(is64, tmp, dst, tmp2), ctx); 625 emit(A64_MSUB(is64, dst, dst, tmp, tmp2), ctx); 626 break; 627 case BPF_ALU | BPF_LSH | BPF_K: 628 case BPF_ALU64 | BPF_LSH | BPF_K: 629 emit(A64_LSL(is64, dst, dst, imm), ctx); 630 break; 631 case BPF_ALU | BPF_RSH | BPF_K: 632 case BPF_ALU64 | BPF_RSH | BPF_K: 633 emit(A64_LSR(is64, dst, dst, imm), ctx); 634 break; 635 case BPF_ALU | BPF_ARSH | BPF_K: 636 case BPF_ALU64 | BPF_ARSH | BPF_K: 637 emit(A64_ASR(is64, dst, dst, imm), ctx); 638 break; 639 640 /* JUMP off */ 641 case BPF_JMP | BPF_JA: 642 jmp_offset = bpf2a64_offset(i, off, ctx); 643 check_imm26(jmp_offset); 644 emit(A64_B(jmp_offset), ctx); 645 break; 646 /* IF (dst COND src) JUMP off */ 647 case BPF_JMP | BPF_JEQ | BPF_X: 648 case BPF_JMP | BPF_JGT | BPF_X: 649 case BPF_JMP | BPF_JLT | BPF_X: 650 case BPF_JMP | BPF_JGE | BPF_X: 651 case BPF_JMP | BPF_JLE | BPF_X: 652 case BPF_JMP | BPF_JNE | BPF_X: 653 case BPF_JMP | BPF_JSGT | BPF_X: 654 case BPF_JMP | BPF_JSLT | BPF_X: 655 case BPF_JMP | BPF_JSGE | BPF_X: 656 case BPF_JMP | BPF_JSLE | BPF_X: 657 case BPF_JMP32 | BPF_JEQ | BPF_X: 658 case BPF_JMP32 | BPF_JGT | BPF_X: 659 case BPF_JMP32 | BPF_JLT | BPF_X: 660 case BPF_JMP32 | BPF_JGE | BPF_X: 661 case BPF_JMP32 | BPF_JLE | BPF_X: 662 case BPF_JMP32 | BPF_JNE | BPF_X: 663 case BPF_JMP32 | BPF_JSGT | BPF_X: 664 case BPF_JMP32 | BPF_JSLT | BPF_X: 665 case BPF_JMP32 | BPF_JSGE | BPF_X: 666 case BPF_JMP32 | BPF_JSLE | BPF_X: 667 emit(A64_CMP(is64, dst, src), ctx); 668 emit_cond_jmp: 669 jmp_offset = bpf2a64_offset(i, off, ctx); 670 check_imm19(jmp_offset); 671 switch (BPF_OP(code)) { 672 case BPF_JEQ: 673 jmp_cond = A64_COND_EQ; 674 break; 675 case BPF_JGT: 676 jmp_cond = A64_COND_HI; 677 break; 678 case BPF_JLT: 679 jmp_cond = A64_COND_CC; 680 break; 681 case BPF_JGE: 682 jmp_cond = A64_COND_CS; 683 break; 684 case BPF_JLE: 685 jmp_cond = A64_COND_LS; 686 break; 687 case BPF_JSET: 688 case BPF_JNE: 689 jmp_cond = A64_COND_NE; 690 break; 691 case BPF_JSGT: 692 jmp_cond = A64_COND_GT; 693 break; 694 case BPF_JSLT: 695 jmp_cond = A64_COND_LT; 696 break; 697 case BPF_JSGE: 698 jmp_cond = A64_COND_GE; 699 break; 700 case BPF_JSLE: 701 jmp_cond = A64_COND_LE; 702 break; 703 default: 704 return -EFAULT; 705 } 706 emit(A64_B_(jmp_cond, jmp_offset), ctx); 707 break; 708 case BPF_JMP | BPF_JSET | BPF_X: 709 case BPF_JMP32 | BPF_JSET | BPF_X: 710 emit(A64_TST(is64, dst, src), ctx); 711 goto emit_cond_jmp; 712 /* IF (dst COND imm) JUMP off */ 713 case BPF_JMP | BPF_JEQ | BPF_K: 714 case BPF_JMP | BPF_JGT | BPF_K: 715 case BPF_JMP | BPF_JLT | BPF_K: 716 case BPF_JMP | BPF_JGE | BPF_K: 717 case BPF_JMP | BPF_JLE | BPF_K: 718 case BPF_JMP | BPF_JNE | BPF_K: 719 case BPF_JMP | BPF_JSGT | BPF_K: 720 case BPF_JMP | BPF_JSLT | BPF_K: 721 case BPF_JMP | BPF_JSGE | BPF_K: 722 case BPF_JMP | BPF_JSLE | BPF_K: 723 case BPF_JMP32 | BPF_JEQ | BPF_K: 724 case BPF_JMP32 | BPF_JGT | BPF_K: 725 case BPF_JMP32 | BPF_JLT | BPF_K: 726 case BPF_JMP32 | BPF_JGE | BPF_K: 727 case BPF_JMP32 | BPF_JLE | BPF_K: 728 case BPF_JMP32 | BPF_JNE | BPF_K: 729 case BPF_JMP32 | BPF_JSGT | BPF_K: 730 case BPF_JMP32 | BPF_JSLT | BPF_K: 731 case BPF_JMP32 | BPF_JSGE | BPF_K: 732 case BPF_JMP32 | BPF_JSLE | BPF_K: 733 if (is_addsub_imm(imm)) { 734 emit(A64_CMP_I(is64, dst, imm), ctx); 735 } else if (is_addsub_imm(-imm)) { 736 emit(A64_CMN_I(is64, dst, -imm), ctx); 737 } else { 738 emit_a64_mov_i(is64, tmp, imm, ctx); 739 emit(A64_CMP(is64, dst, tmp), ctx); 740 } 741 goto emit_cond_jmp; 742 case BPF_JMP | BPF_JSET | BPF_K: 743 case BPF_JMP32 | BPF_JSET | BPF_K: 744 a64_insn = A64_TST_I(is64, dst, imm); 745 if (a64_insn != AARCH64_BREAK_FAULT) { 746 emit(a64_insn, ctx); 747 } else { 748 emit_a64_mov_i(is64, tmp, imm, ctx); 749 emit(A64_TST(is64, dst, tmp), ctx); 750 } 751 goto emit_cond_jmp; 752 /* function call */ 753 case BPF_JMP | BPF_CALL: 754 { 755 const u8 r0 = bpf2a64[BPF_REG_0]; 756 bool func_addr_fixed; 757 u64 func_addr; 758 759 ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass, 760 &func_addr, &func_addr_fixed); 761 if (ret < 0) 762 return ret; 763 emit_addr_mov_i64(tmp, func_addr, ctx); 764 emit(A64_BLR(tmp), ctx); 765 emit(A64_MOV(1, r0, A64_R(0)), ctx); 766 break; 767 } 768 /* tail call */ 769 case BPF_JMP | BPF_TAIL_CALL: 770 if (emit_bpf_tail_call(ctx)) 771 return -EFAULT; 772 break; 773 /* function return */ 774 case BPF_JMP | BPF_EXIT: 775 /* Optimization: when last instruction is EXIT, 776 simply fallthrough to epilogue. */ 777 if (i == ctx->prog->len - 1) 778 break; 779 jmp_offset = epilogue_offset(ctx); 780 check_imm26(jmp_offset); 781 emit(A64_B(jmp_offset), ctx); 782 break; 783 784 /* dst = imm64 */ 785 case BPF_LD | BPF_IMM | BPF_DW: 786 { 787 const struct bpf_insn insn1 = insn[1]; 788 u64 imm64; 789 790 imm64 = (u64)insn1.imm << 32 | (u32)imm; 791 emit_a64_mov_i64(dst, imm64, ctx); 792 793 return 1; 794 } 795 796 /* LDX: dst = *(size *)(src + off) */ 797 case BPF_LDX | BPF_MEM | BPF_W: 798 case BPF_LDX | BPF_MEM | BPF_H: 799 case BPF_LDX | BPF_MEM | BPF_B: 800 case BPF_LDX | BPF_MEM | BPF_DW: 801 case BPF_LDX | BPF_PROBE_MEM | BPF_DW: 802 case BPF_LDX | BPF_PROBE_MEM | BPF_W: 803 case BPF_LDX | BPF_PROBE_MEM | BPF_H: 804 case BPF_LDX | BPF_PROBE_MEM | BPF_B: 805 emit_a64_mov_i(1, tmp, off, ctx); 806 switch (BPF_SIZE(code)) { 807 case BPF_W: 808 emit(A64_LDR32(dst, src, tmp), ctx); 809 break; 810 case BPF_H: 811 emit(A64_LDRH(dst, src, tmp), ctx); 812 break; 813 case BPF_B: 814 emit(A64_LDRB(dst, src, tmp), ctx); 815 break; 816 case BPF_DW: 817 emit(A64_LDR64(dst, src, tmp), ctx); 818 break; 819 } 820 821 ret = add_exception_handler(insn, ctx, dst); 822 if (ret) 823 return ret; 824 break; 825 826 /* speculation barrier */ 827 case BPF_ST | BPF_NOSPEC: 828 /* 829 * Nothing required here. 830 * 831 * In case of arm64, we rely on the firmware mitigation of 832 * Speculative Store Bypass as controlled via the ssbd kernel 833 * parameter. Whenever the mitigation is enabled, it works 834 * for all of the kernel code with no need to provide any 835 * additional instructions. 836 */ 837 break; 838 839 /* ST: *(size *)(dst + off) = imm */ 840 case BPF_ST | BPF_MEM | BPF_W: 841 case BPF_ST | BPF_MEM | BPF_H: 842 case BPF_ST | BPF_MEM | BPF_B: 843 case BPF_ST | BPF_MEM | BPF_DW: 844 /* Load imm to a register then store it */ 845 emit_a64_mov_i(1, tmp2, off, ctx); 846 emit_a64_mov_i(1, tmp, imm, ctx); 847 switch (BPF_SIZE(code)) { 848 case BPF_W: 849 emit(A64_STR32(tmp, dst, tmp2), ctx); 850 break; 851 case BPF_H: 852 emit(A64_STRH(tmp, dst, tmp2), ctx); 853 break; 854 case BPF_B: 855 emit(A64_STRB(tmp, dst, tmp2), ctx); 856 break; 857 case BPF_DW: 858 emit(A64_STR64(tmp, dst, tmp2), ctx); 859 break; 860 } 861 break; 862 863 /* STX: *(size *)(dst + off) = src */ 864 case BPF_STX | BPF_MEM | BPF_W: 865 case BPF_STX | BPF_MEM | BPF_H: 866 case BPF_STX | BPF_MEM | BPF_B: 867 case BPF_STX | BPF_MEM | BPF_DW: 868 emit_a64_mov_i(1, tmp, off, ctx); 869 switch (BPF_SIZE(code)) { 870 case BPF_W: 871 emit(A64_STR32(src, dst, tmp), ctx); 872 break; 873 case BPF_H: 874 emit(A64_STRH(src, dst, tmp), ctx); 875 break; 876 case BPF_B: 877 emit(A64_STRB(src, dst, tmp), ctx); 878 break; 879 case BPF_DW: 880 emit(A64_STR64(src, dst, tmp), ctx); 881 break; 882 } 883 break; 884 885 case BPF_STX | BPF_ATOMIC | BPF_W: 886 case BPF_STX | BPF_ATOMIC | BPF_DW: 887 if (insn->imm != BPF_ADD) { 888 pr_err_once("unknown atomic op code %02x\n", insn->imm); 889 return -EINVAL; 890 } 891 892 /* STX XADD: lock *(u32 *)(dst + off) += src 893 * and 894 * STX XADD: lock *(u64 *)(dst + off) += src 895 */ 896 897 if (!off) { 898 reg = dst; 899 } else { 900 emit_a64_mov_i(1, tmp, off, ctx); 901 emit(A64_ADD(1, tmp, tmp, dst), ctx); 902 reg = tmp; 903 } 904 if (cpus_have_cap(ARM64_HAS_LSE_ATOMICS)) { 905 emit(A64_STADD(isdw, reg, src), ctx); 906 } else { 907 emit(A64_LDXR(isdw, tmp2, reg), ctx); 908 emit(A64_ADD(isdw, tmp2, tmp2, src), ctx); 909 emit(A64_STXR(isdw, tmp2, reg, tmp3), ctx); 910 jmp_offset = -3; 911 check_imm19(jmp_offset); 912 emit(A64_CBNZ(0, tmp3, jmp_offset), ctx); 913 } 914 break; 915 916 default: 917 pr_err_once("unknown opcode %02x\n", code); 918 return -EINVAL; 919 } 920 921 return 0; 922 } 923 924 static int build_body(struct jit_ctx *ctx, bool extra_pass) 925 { 926 const struct bpf_prog *prog = ctx->prog; 927 int i; 928 929 /* 930 * - offset[0] offset of the end of prologue, 931 * start of the 1st instruction. 932 * - offset[1] - offset of the end of 1st instruction, 933 * start of the 2nd instruction 934 * [....] 935 * - offset[3] - offset of the end of 3rd instruction, 936 * start of 4th instruction 937 */ 938 for (i = 0; i < prog->len; i++) { 939 const struct bpf_insn *insn = &prog->insnsi[i]; 940 int ret; 941 942 if (ctx->image == NULL) 943 ctx->offset[i] = ctx->idx; 944 ret = build_insn(insn, ctx, extra_pass); 945 if (ret > 0) { 946 i++; 947 if (ctx->image == NULL) 948 ctx->offset[i] = ctx->idx; 949 continue; 950 } 951 if (ret) 952 return ret; 953 } 954 /* 955 * offset is allocated with prog->len + 1 so fill in 956 * the last element with the offset after the last 957 * instruction (end of program) 958 */ 959 if (ctx->image == NULL) 960 ctx->offset[i] = ctx->idx; 961 962 return 0; 963 } 964 965 static int validate_code(struct jit_ctx *ctx) 966 { 967 int i; 968 969 for (i = 0; i < ctx->idx; i++) { 970 u32 a64_insn = le32_to_cpu(ctx->image[i]); 971 972 if (a64_insn == AARCH64_BREAK_FAULT) 973 return -1; 974 } 975 976 if (WARN_ON_ONCE(ctx->exentry_idx != ctx->prog->aux->num_exentries)) 977 return -1; 978 979 return 0; 980 } 981 982 static inline void bpf_flush_icache(void *start, void *end) 983 { 984 flush_icache_range((unsigned long)start, (unsigned long)end); 985 } 986 987 struct arm64_jit_data { 988 struct bpf_binary_header *header; 989 u8 *image; 990 struct jit_ctx ctx; 991 }; 992 993 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) 994 { 995 int image_size, prog_size, extable_size; 996 struct bpf_prog *tmp, *orig_prog = prog; 997 struct bpf_binary_header *header; 998 struct arm64_jit_data *jit_data; 999 bool was_classic = bpf_prog_was_classic(prog); 1000 bool tmp_blinded = false; 1001 bool extra_pass = false; 1002 struct jit_ctx ctx; 1003 u8 *image_ptr; 1004 1005 if (!prog->jit_requested) 1006 return orig_prog; 1007 1008 tmp = bpf_jit_blind_constants(prog); 1009 /* If blinding was requested and we failed during blinding, 1010 * we must fall back to the interpreter. 1011 */ 1012 if (IS_ERR(tmp)) 1013 return orig_prog; 1014 if (tmp != prog) { 1015 tmp_blinded = true; 1016 prog = tmp; 1017 } 1018 1019 jit_data = prog->aux->jit_data; 1020 if (!jit_data) { 1021 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL); 1022 if (!jit_data) { 1023 prog = orig_prog; 1024 goto out; 1025 } 1026 prog->aux->jit_data = jit_data; 1027 } 1028 if (jit_data->ctx.offset) { 1029 ctx = jit_data->ctx; 1030 image_ptr = jit_data->image; 1031 header = jit_data->header; 1032 extra_pass = true; 1033 prog_size = sizeof(u32) * ctx.idx; 1034 goto skip_init_ctx; 1035 } 1036 memset(&ctx, 0, sizeof(ctx)); 1037 ctx.prog = prog; 1038 1039 ctx.offset = kcalloc(prog->len + 1, sizeof(int), GFP_KERNEL); 1040 if (ctx.offset == NULL) { 1041 prog = orig_prog; 1042 goto out_off; 1043 } 1044 1045 /* 1. Initial fake pass to compute ctx->idx. */ 1046 1047 /* Fake pass to fill in ctx->offset. */ 1048 if (build_body(&ctx, extra_pass)) { 1049 prog = orig_prog; 1050 goto out_off; 1051 } 1052 1053 if (build_prologue(&ctx, was_classic)) { 1054 prog = orig_prog; 1055 goto out_off; 1056 } 1057 1058 ctx.epilogue_offset = ctx.idx; 1059 build_epilogue(&ctx); 1060 1061 extable_size = prog->aux->num_exentries * 1062 sizeof(struct exception_table_entry); 1063 1064 /* Now we know the actual image size. */ 1065 prog_size = sizeof(u32) * ctx.idx; 1066 image_size = prog_size + extable_size; 1067 header = bpf_jit_binary_alloc(image_size, &image_ptr, 1068 sizeof(u32), jit_fill_hole); 1069 if (header == NULL) { 1070 prog = orig_prog; 1071 goto out_off; 1072 } 1073 1074 /* 2. Now, the actual pass. */ 1075 1076 ctx.image = (__le32 *)image_ptr; 1077 if (extable_size) 1078 prog->aux->extable = (void *)image_ptr + prog_size; 1079 skip_init_ctx: 1080 ctx.idx = 0; 1081 ctx.exentry_idx = 0; 1082 1083 build_prologue(&ctx, was_classic); 1084 1085 if (build_body(&ctx, extra_pass)) { 1086 bpf_jit_binary_free(header); 1087 prog = orig_prog; 1088 goto out_off; 1089 } 1090 1091 build_epilogue(&ctx); 1092 1093 /* 3. Extra pass to validate JITed code. */ 1094 if (validate_code(&ctx)) { 1095 bpf_jit_binary_free(header); 1096 prog = orig_prog; 1097 goto out_off; 1098 } 1099 1100 /* And we're done. */ 1101 if (bpf_jit_enable > 1) 1102 bpf_jit_dump(prog->len, prog_size, 2, ctx.image); 1103 1104 bpf_flush_icache(header, ctx.image + ctx.idx); 1105 1106 if (!prog->is_func || extra_pass) { 1107 if (extra_pass && ctx.idx != jit_data->ctx.idx) { 1108 pr_err_once("multi-func JIT bug %d != %d\n", 1109 ctx.idx, jit_data->ctx.idx); 1110 bpf_jit_binary_free(header); 1111 prog->bpf_func = NULL; 1112 prog->jited = 0; 1113 goto out_off; 1114 } 1115 bpf_jit_binary_lock_ro(header); 1116 } else { 1117 jit_data->ctx = ctx; 1118 jit_data->image = image_ptr; 1119 jit_data->header = header; 1120 } 1121 prog->bpf_func = (void *)ctx.image; 1122 prog->jited = 1; 1123 prog->jited_len = prog_size; 1124 1125 if (!prog->is_func || extra_pass) { 1126 bpf_prog_fill_jited_linfo(prog, ctx.offset + 1); 1127 out_off: 1128 kfree(ctx.offset); 1129 kfree(jit_data); 1130 prog->aux->jit_data = NULL; 1131 } 1132 out: 1133 if (tmp_blinded) 1134 bpf_jit_prog_release_other(prog, prog == orig_prog ? 1135 tmp : orig_prog); 1136 return prog; 1137 } 1138 1139 void *bpf_jit_alloc_exec(unsigned long size) 1140 { 1141 return __vmalloc_node_range(size, PAGE_SIZE, BPF_JIT_REGION_START, 1142 BPF_JIT_REGION_END, GFP_KERNEL, 1143 PAGE_KERNEL, 0, NUMA_NO_NODE, 1144 __builtin_return_address(0)); 1145 } 1146 1147 void bpf_jit_free_exec(void *addr) 1148 { 1149 return vfree(addr); 1150 } 1151