1 /* 2 * BPF JIT compiler for ARM64 3 * 4 * Copyright (C) 2014-2016 Zi Shen Lim <zlim.lnx@gmail.com> 5 * 6 * This program is free software; you can redistribute it and/or modify 7 * it under the terms of the GNU General Public License version 2 as 8 * published by the Free Software Foundation. 9 * 10 * This program is distributed in the hope that it will be useful, 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 13 * GNU General Public License for more details. 14 * 15 * You should have received a copy of the GNU General Public License 16 * along with this program. If not, see <http://www.gnu.org/licenses/>. 17 */ 18 19 #define pr_fmt(fmt) "bpf_jit: " fmt 20 21 #include <linux/bpf.h> 22 #include <linux/filter.h> 23 #include <linux/printk.h> 24 #include <linux/skbuff.h> 25 #include <linux/slab.h> 26 27 #include <asm/byteorder.h> 28 #include <asm/cacheflush.h> 29 #include <asm/debug-monitors.h> 30 #include <asm/set_memory.h> 31 32 #include "bpf_jit.h" 33 34 int bpf_jit_enable __read_mostly; 35 36 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0) 37 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1) 38 #define TCALL_CNT (MAX_BPF_JIT_REG + 2) 39 #define TMP_REG_3 (MAX_BPF_JIT_REG + 3) 40 41 /* Map BPF registers to A64 registers */ 42 static const int bpf2a64[] = { 43 /* return value from in-kernel function, and exit value from eBPF */ 44 [BPF_REG_0] = A64_R(7), 45 /* arguments from eBPF program to in-kernel function */ 46 [BPF_REG_1] = A64_R(0), 47 [BPF_REG_2] = A64_R(1), 48 [BPF_REG_3] = A64_R(2), 49 [BPF_REG_4] = A64_R(3), 50 [BPF_REG_5] = A64_R(4), 51 /* callee saved registers that in-kernel function will preserve */ 52 [BPF_REG_6] = A64_R(19), 53 [BPF_REG_7] = A64_R(20), 54 [BPF_REG_8] = A64_R(21), 55 [BPF_REG_9] = A64_R(22), 56 /* read-only frame pointer to access stack */ 57 [BPF_REG_FP] = A64_R(25), 58 /* temporary registers for internal BPF JIT */ 59 [TMP_REG_1] = A64_R(10), 60 [TMP_REG_2] = A64_R(11), 61 [TMP_REG_3] = A64_R(12), 62 /* tail_call_cnt */ 63 [TCALL_CNT] = A64_R(26), 64 /* temporary register for blinding constants */ 65 [BPF_REG_AX] = A64_R(9), 66 }; 67 68 struct jit_ctx { 69 const struct bpf_prog *prog; 70 int idx; 71 int epilogue_offset; 72 int *offset; 73 __le32 *image; 74 u32 stack_size; 75 }; 76 77 static inline void emit(const u32 insn, struct jit_ctx *ctx) 78 { 79 if (ctx->image != NULL) 80 ctx->image[ctx->idx] = cpu_to_le32(insn); 81 82 ctx->idx++; 83 } 84 85 static inline void emit_a64_mov_i64(const int reg, const u64 val, 86 struct jit_ctx *ctx) 87 { 88 u64 tmp = val; 89 int shift = 0; 90 91 emit(A64_MOVZ(1, reg, tmp & 0xffff, shift), ctx); 92 tmp >>= 16; 93 shift += 16; 94 while (tmp) { 95 if (tmp & 0xffff) 96 emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx); 97 tmp >>= 16; 98 shift += 16; 99 } 100 } 101 102 static inline void emit_addr_mov_i64(const int reg, const u64 val, 103 struct jit_ctx *ctx) 104 { 105 u64 tmp = val; 106 int shift = 0; 107 108 emit(A64_MOVZ(1, reg, tmp & 0xffff, shift), ctx); 109 for (;shift < 48;) { 110 tmp >>= 16; 111 shift += 16; 112 emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx); 113 } 114 } 115 116 static inline void emit_a64_mov_i(const int is64, const int reg, 117 const s32 val, struct jit_ctx *ctx) 118 { 119 u16 hi = val >> 16; 120 u16 lo = val & 0xffff; 121 122 if (hi & 0x8000) { 123 if (hi == 0xffff) { 124 emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx); 125 } else { 126 emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx); 127 emit(A64_MOVK(is64, reg, lo, 0), ctx); 128 } 129 } else { 130 emit(A64_MOVZ(is64, reg, lo, 0), ctx); 131 if (hi) 132 emit(A64_MOVK(is64, reg, hi, 16), ctx); 133 } 134 } 135 136 static inline int bpf2a64_offset(int bpf_to, int bpf_from, 137 const struct jit_ctx *ctx) 138 { 139 int to = ctx->offset[bpf_to]; 140 /* -1 to account for the Branch instruction */ 141 int from = ctx->offset[bpf_from] - 1; 142 143 return to - from; 144 } 145 146 static void jit_fill_hole(void *area, unsigned int size) 147 { 148 __le32 *ptr; 149 /* We are guaranteed to have aligned memory. */ 150 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32)) 151 *ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT); 152 } 153 154 static inline int epilogue_offset(const struct jit_ctx *ctx) 155 { 156 int to = ctx->epilogue_offset; 157 int from = ctx->idx; 158 159 return to - from; 160 } 161 162 /* Stack must be multiples of 16B */ 163 #define STACK_ALIGN(sz) (((sz) + 15) & ~15) 164 165 #define PROLOGUE_OFFSET 8 166 167 static int build_prologue(struct jit_ctx *ctx) 168 { 169 const struct bpf_prog *prog = ctx->prog; 170 const u8 r6 = bpf2a64[BPF_REG_6]; 171 const u8 r7 = bpf2a64[BPF_REG_7]; 172 const u8 r8 = bpf2a64[BPF_REG_8]; 173 const u8 r9 = bpf2a64[BPF_REG_9]; 174 const u8 fp = bpf2a64[BPF_REG_FP]; 175 const u8 tcc = bpf2a64[TCALL_CNT]; 176 const int idx0 = ctx->idx; 177 int cur_offset; 178 179 /* 180 * BPF prog stack layout 181 * 182 * high 183 * original A64_SP => 0:+-----+ BPF prologue 184 * |FP/LR| 185 * current A64_FP => -16:+-----+ 186 * | ... | callee saved registers 187 * BPF fp register => -64:+-----+ <= (BPF_FP) 188 * | | 189 * | ... | BPF prog stack 190 * | | 191 * +-----+ <= (BPF_FP - prog->aux->stack_depth) 192 * |RSVD | JIT scratchpad 193 * current A64_SP => +-----+ <= (BPF_FP - ctx->stack_size) 194 * | | 195 * | ... | Function call stack 196 * | | 197 * +-----+ 198 * low 199 * 200 */ 201 202 /* Save FP and LR registers to stay align with ARM64 AAPCS */ 203 emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx); 204 emit(A64_MOV(1, A64_FP, A64_SP), ctx); 205 206 /* Save callee-saved registers */ 207 emit(A64_PUSH(r6, r7, A64_SP), ctx); 208 emit(A64_PUSH(r8, r9, A64_SP), ctx); 209 emit(A64_PUSH(fp, tcc, A64_SP), ctx); 210 211 /* Set up BPF prog stack base register */ 212 emit(A64_MOV(1, fp, A64_SP), ctx); 213 214 /* Initialize tail_call_cnt */ 215 emit(A64_MOVZ(1, tcc, 0, 0), ctx); 216 217 /* 4 byte extra for skb_copy_bits buffer */ 218 ctx->stack_size = prog->aux->stack_depth + 4; 219 ctx->stack_size = STACK_ALIGN(ctx->stack_size); 220 221 /* Set up function call stack */ 222 emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 223 224 cur_offset = ctx->idx - idx0; 225 if (cur_offset != PROLOGUE_OFFSET) { 226 pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n", 227 cur_offset, PROLOGUE_OFFSET); 228 return -1; 229 } 230 return 0; 231 } 232 233 static int out_offset = -1; /* initialized on the first pass of build_body() */ 234 static int emit_bpf_tail_call(struct jit_ctx *ctx) 235 { 236 /* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */ 237 const u8 r2 = bpf2a64[BPF_REG_2]; 238 const u8 r3 = bpf2a64[BPF_REG_3]; 239 240 const u8 tmp = bpf2a64[TMP_REG_1]; 241 const u8 prg = bpf2a64[TMP_REG_2]; 242 const u8 tcc = bpf2a64[TCALL_CNT]; 243 const int idx0 = ctx->idx; 244 #define cur_offset (ctx->idx - idx0) 245 #define jmp_offset (out_offset - (cur_offset)) 246 size_t off; 247 248 /* if (index >= array->map.max_entries) 249 * goto out; 250 */ 251 off = offsetof(struct bpf_array, map.max_entries); 252 emit_a64_mov_i64(tmp, off, ctx); 253 emit(A64_LDR32(tmp, r2, tmp), ctx); 254 emit(A64_CMP(0, r3, tmp), ctx); 255 emit(A64_B_(A64_COND_GE, jmp_offset), ctx); 256 257 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT) 258 * goto out; 259 * tail_call_cnt++; 260 */ 261 emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx); 262 emit(A64_CMP(1, tcc, tmp), ctx); 263 emit(A64_B_(A64_COND_GT, jmp_offset), ctx); 264 emit(A64_ADD_I(1, tcc, tcc, 1), ctx); 265 266 /* prog = array->ptrs[index]; 267 * if (prog == NULL) 268 * goto out; 269 */ 270 off = offsetof(struct bpf_array, ptrs); 271 emit_a64_mov_i64(tmp, off, ctx); 272 emit(A64_ADD(1, tmp, r2, tmp), ctx); 273 emit(A64_LSL(1, prg, r3, 3), ctx); 274 emit(A64_LDR64(prg, tmp, prg), ctx); 275 emit(A64_CBZ(1, prg, jmp_offset), ctx); 276 277 /* goto *(prog->bpf_func + prologue_size); */ 278 off = offsetof(struct bpf_prog, bpf_func); 279 emit_a64_mov_i64(tmp, off, ctx); 280 emit(A64_LDR64(tmp, prg, tmp), ctx); 281 emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx); 282 emit(A64_BR(tmp), ctx); 283 284 /* out: */ 285 if (out_offset == -1) 286 out_offset = cur_offset; 287 if (cur_offset != out_offset) { 288 pr_err_once("tail_call out_offset = %d, expected %d!\n", 289 cur_offset, out_offset); 290 return -1; 291 } 292 return 0; 293 #undef cur_offset 294 #undef jmp_offset 295 } 296 297 static void build_epilogue(struct jit_ctx *ctx) 298 { 299 const u8 r0 = bpf2a64[BPF_REG_0]; 300 const u8 r6 = bpf2a64[BPF_REG_6]; 301 const u8 r7 = bpf2a64[BPF_REG_7]; 302 const u8 r8 = bpf2a64[BPF_REG_8]; 303 const u8 r9 = bpf2a64[BPF_REG_9]; 304 const u8 fp = bpf2a64[BPF_REG_FP]; 305 306 /* We're done with BPF stack */ 307 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 308 309 /* Restore fs (x25) and x26 */ 310 emit(A64_POP(fp, A64_R(26), A64_SP), ctx); 311 312 /* Restore callee-saved register */ 313 emit(A64_POP(r8, r9, A64_SP), ctx); 314 emit(A64_POP(r6, r7, A64_SP), ctx); 315 316 /* Restore FP/LR registers */ 317 emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx); 318 319 /* Set return value */ 320 emit(A64_MOV(1, A64_R(0), r0), ctx); 321 322 emit(A64_RET(A64_LR), ctx); 323 } 324 325 /* JITs an eBPF instruction. 326 * Returns: 327 * 0 - successfully JITed an 8-byte eBPF instruction. 328 * >0 - successfully JITed a 16-byte eBPF instruction. 329 * <0 - failed to JIT. 330 */ 331 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx) 332 { 333 const u8 code = insn->code; 334 const u8 dst = bpf2a64[insn->dst_reg]; 335 const u8 src = bpf2a64[insn->src_reg]; 336 const u8 tmp = bpf2a64[TMP_REG_1]; 337 const u8 tmp2 = bpf2a64[TMP_REG_2]; 338 const u8 tmp3 = bpf2a64[TMP_REG_3]; 339 const s16 off = insn->off; 340 const s32 imm = insn->imm; 341 const int i = insn - ctx->prog->insnsi; 342 const bool is64 = BPF_CLASS(code) == BPF_ALU64; 343 const bool isdw = BPF_SIZE(code) == BPF_DW; 344 u8 jmp_cond; 345 s32 jmp_offset; 346 347 #define check_imm(bits, imm) do { \ 348 if ((((imm) > 0) && ((imm) >> (bits))) || \ 349 (((imm) < 0) && (~(imm) >> (bits)))) { \ 350 pr_info("[%2d] imm=%d(0x%x) out of range\n", \ 351 i, imm, imm); \ 352 return -EINVAL; \ 353 } \ 354 } while (0) 355 #define check_imm19(imm) check_imm(19, imm) 356 #define check_imm26(imm) check_imm(26, imm) 357 358 switch (code) { 359 /* dst = src */ 360 case BPF_ALU | BPF_MOV | BPF_X: 361 case BPF_ALU64 | BPF_MOV | BPF_X: 362 emit(A64_MOV(is64, dst, src), ctx); 363 break; 364 /* dst = dst OP src */ 365 case BPF_ALU | BPF_ADD | BPF_X: 366 case BPF_ALU64 | BPF_ADD | BPF_X: 367 emit(A64_ADD(is64, dst, dst, src), ctx); 368 break; 369 case BPF_ALU | BPF_SUB | BPF_X: 370 case BPF_ALU64 | BPF_SUB | BPF_X: 371 emit(A64_SUB(is64, dst, dst, src), ctx); 372 break; 373 case BPF_ALU | BPF_AND | BPF_X: 374 case BPF_ALU64 | BPF_AND | BPF_X: 375 emit(A64_AND(is64, dst, dst, src), ctx); 376 break; 377 case BPF_ALU | BPF_OR | BPF_X: 378 case BPF_ALU64 | BPF_OR | BPF_X: 379 emit(A64_ORR(is64, dst, dst, src), ctx); 380 break; 381 case BPF_ALU | BPF_XOR | BPF_X: 382 case BPF_ALU64 | BPF_XOR | BPF_X: 383 emit(A64_EOR(is64, dst, dst, src), ctx); 384 break; 385 case BPF_ALU | BPF_MUL | BPF_X: 386 case BPF_ALU64 | BPF_MUL | BPF_X: 387 emit(A64_MUL(is64, dst, dst, src), ctx); 388 break; 389 case BPF_ALU | BPF_DIV | BPF_X: 390 case BPF_ALU64 | BPF_DIV | BPF_X: 391 case BPF_ALU | BPF_MOD | BPF_X: 392 case BPF_ALU64 | BPF_MOD | BPF_X: 393 { 394 const u8 r0 = bpf2a64[BPF_REG_0]; 395 396 /* if (src == 0) return 0 */ 397 jmp_offset = 3; /* skip ahead to else path */ 398 check_imm19(jmp_offset); 399 emit(A64_CBNZ(is64, src, jmp_offset), ctx); 400 emit(A64_MOVZ(1, r0, 0, 0), ctx); 401 jmp_offset = epilogue_offset(ctx); 402 check_imm26(jmp_offset); 403 emit(A64_B(jmp_offset), ctx); 404 /* else */ 405 switch (BPF_OP(code)) { 406 case BPF_DIV: 407 emit(A64_UDIV(is64, dst, dst, src), ctx); 408 break; 409 case BPF_MOD: 410 emit(A64_UDIV(is64, tmp, dst, src), ctx); 411 emit(A64_MUL(is64, tmp, tmp, src), ctx); 412 emit(A64_SUB(is64, dst, dst, tmp), ctx); 413 break; 414 } 415 break; 416 } 417 case BPF_ALU | BPF_LSH | BPF_X: 418 case BPF_ALU64 | BPF_LSH | BPF_X: 419 emit(A64_LSLV(is64, dst, dst, src), ctx); 420 break; 421 case BPF_ALU | BPF_RSH | BPF_X: 422 case BPF_ALU64 | BPF_RSH | BPF_X: 423 emit(A64_LSRV(is64, dst, dst, src), ctx); 424 break; 425 case BPF_ALU | BPF_ARSH | BPF_X: 426 case BPF_ALU64 | BPF_ARSH | BPF_X: 427 emit(A64_ASRV(is64, dst, dst, src), ctx); 428 break; 429 /* dst = -dst */ 430 case BPF_ALU | BPF_NEG: 431 case BPF_ALU64 | BPF_NEG: 432 emit(A64_NEG(is64, dst, dst), ctx); 433 break; 434 /* dst = BSWAP##imm(dst) */ 435 case BPF_ALU | BPF_END | BPF_FROM_LE: 436 case BPF_ALU | BPF_END | BPF_FROM_BE: 437 #ifdef CONFIG_CPU_BIG_ENDIAN 438 if (BPF_SRC(code) == BPF_FROM_BE) 439 goto emit_bswap_uxt; 440 #else /* !CONFIG_CPU_BIG_ENDIAN */ 441 if (BPF_SRC(code) == BPF_FROM_LE) 442 goto emit_bswap_uxt; 443 #endif 444 switch (imm) { 445 case 16: 446 emit(A64_REV16(is64, dst, dst), ctx); 447 /* zero-extend 16 bits into 64 bits */ 448 emit(A64_UXTH(is64, dst, dst), ctx); 449 break; 450 case 32: 451 emit(A64_REV32(is64, dst, dst), ctx); 452 /* upper 32 bits already cleared */ 453 break; 454 case 64: 455 emit(A64_REV64(dst, dst), ctx); 456 break; 457 } 458 break; 459 emit_bswap_uxt: 460 switch (imm) { 461 case 16: 462 /* zero-extend 16 bits into 64 bits */ 463 emit(A64_UXTH(is64, dst, dst), ctx); 464 break; 465 case 32: 466 /* zero-extend 32 bits into 64 bits */ 467 emit(A64_UXTW(is64, dst, dst), ctx); 468 break; 469 case 64: 470 /* nop */ 471 break; 472 } 473 break; 474 /* dst = imm */ 475 case BPF_ALU | BPF_MOV | BPF_K: 476 case BPF_ALU64 | BPF_MOV | BPF_K: 477 emit_a64_mov_i(is64, dst, imm, ctx); 478 break; 479 /* dst = dst OP imm */ 480 case BPF_ALU | BPF_ADD | BPF_K: 481 case BPF_ALU64 | BPF_ADD | BPF_K: 482 emit_a64_mov_i(is64, tmp, imm, ctx); 483 emit(A64_ADD(is64, dst, dst, tmp), ctx); 484 break; 485 case BPF_ALU | BPF_SUB | BPF_K: 486 case BPF_ALU64 | BPF_SUB | BPF_K: 487 emit_a64_mov_i(is64, tmp, imm, ctx); 488 emit(A64_SUB(is64, dst, dst, tmp), ctx); 489 break; 490 case BPF_ALU | BPF_AND | BPF_K: 491 case BPF_ALU64 | BPF_AND | BPF_K: 492 emit_a64_mov_i(is64, tmp, imm, ctx); 493 emit(A64_AND(is64, dst, dst, tmp), ctx); 494 break; 495 case BPF_ALU | BPF_OR | BPF_K: 496 case BPF_ALU64 | BPF_OR | BPF_K: 497 emit_a64_mov_i(is64, tmp, imm, ctx); 498 emit(A64_ORR(is64, dst, dst, tmp), ctx); 499 break; 500 case BPF_ALU | BPF_XOR | BPF_K: 501 case BPF_ALU64 | BPF_XOR | BPF_K: 502 emit_a64_mov_i(is64, tmp, imm, ctx); 503 emit(A64_EOR(is64, dst, dst, tmp), ctx); 504 break; 505 case BPF_ALU | BPF_MUL | BPF_K: 506 case BPF_ALU64 | BPF_MUL | BPF_K: 507 emit_a64_mov_i(is64, tmp, imm, ctx); 508 emit(A64_MUL(is64, dst, dst, tmp), ctx); 509 break; 510 case BPF_ALU | BPF_DIV | BPF_K: 511 case BPF_ALU64 | BPF_DIV | BPF_K: 512 emit_a64_mov_i(is64, tmp, imm, ctx); 513 emit(A64_UDIV(is64, dst, dst, tmp), ctx); 514 break; 515 case BPF_ALU | BPF_MOD | BPF_K: 516 case BPF_ALU64 | BPF_MOD | BPF_K: 517 emit_a64_mov_i(is64, tmp2, imm, ctx); 518 emit(A64_UDIV(is64, tmp, dst, tmp2), ctx); 519 emit(A64_MUL(is64, tmp, tmp, tmp2), ctx); 520 emit(A64_SUB(is64, dst, dst, tmp), ctx); 521 break; 522 case BPF_ALU | BPF_LSH | BPF_K: 523 case BPF_ALU64 | BPF_LSH | BPF_K: 524 emit(A64_LSL(is64, dst, dst, imm), ctx); 525 break; 526 case BPF_ALU | BPF_RSH | BPF_K: 527 case BPF_ALU64 | BPF_RSH | BPF_K: 528 emit(A64_LSR(is64, dst, dst, imm), ctx); 529 break; 530 case BPF_ALU | BPF_ARSH | BPF_K: 531 case BPF_ALU64 | BPF_ARSH | BPF_K: 532 emit(A64_ASR(is64, dst, dst, imm), ctx); 533 break; 534 535 /* JUMP off */ 536 case BPF_JMP | BPF_JA: 537 jmp_offset = bpf2a64_offset(i + off, i, ctx); 538 check_imm26(jmp_offset); 539 emit(A64_B(jmp_offset), ctx); 540 break; 541 /* IF (dst COND src) JUMP off */ 542 case BPF_JMP | BPF_JEQ | BPF_X: 543 case BPF_JMP | BPF_JGT | BPF_X: 544 case BPF_JMP | BPF_JLT | BPF_X: 545 case BPF_JMP | BPF_JGE | BPF_X: 546 case BPF_JMP | BPF_JLE | BPF_X: 547 case BPF_JMP | BPF_JNE | BPF_X: 548 case BPF_JMP | BPF_JSGT | BPF_X: 549 case BPF_JMP | BPF_JSLT | BPF_X: 550 case BPF_JMP | BPF_JSGE | BPF_X: 551 case BPF_JMP | BPF_JSLE | BPF_X: 552 emit(A64_CMP(1, dst, src), ctx); 553 emit_cond_jmp: 554 jmp_offset = bpf2a64_offset(i + off, i, ctx); 555 check_imm19(jmp_offset); 556 switch (BPF_OP(code)) { 557 case BPF_JEQ: 558 jmp_cond = A64_COND_EQ; 559 break; 560 case BPF_JGT: 561 jmp_cond = A64_COND_HI; 562 break; 563 case BPF_JLT: 564 jmp_cond = A64_COND_CC; 565 break; 566 case BPF_JGE: 567 jmp_cond = A64_COND_CS; 568 break; 569 case BPF_JLE: 570 jmp_cond = A64_COND_LS; 571 break; 572 case BPF_JSET: 573 case BPF_JNE: 574 jmp_cond = A64_COND_NE; 575 break; 576 case BPF_JSGT: 577 jmp_cond = A64_COND_GT; 578 break; 579 case BPF_JSLT: 580 jmp_cond = A64_COND_LT; 581 break; 582 case BPF_JSGE: 583 jmp_cond = A64_COND_GE; 584 break; 585 case BPF_JSLE: 586 jmp_cond = A64_COND_LE; 587 break; 588 default: 589 return -EFAULT; 590 } 591 emit(A64_B_(jmp_cond, jmp_offset), ctx); 592 break; 593 case BPF_JMP | BPF_JSET | BPF_X: 594 emit(A64_TST(1, dst, src), ctx); 595 goto emit_cond_jmp; 596 /* IF (dst COND imm) JUMP off */ 597 case BPF_JMP | BPF_JEQ | BPF_K: 598 case BPF_JMP | BPF_JGT | BPF_K: 599 case BPF_JMP | BPF_JLT | BPF_K: 600 case BPF_JMP | BPF_JGE | BPF_K: 601 case BPF_JMP | BPF_JLE | BPF_K: 602 case BPF_JMP | BPF_JNE | BPF_K: 603 case BPF_JMP | BPF_JSGT | BPF_K: 604 case BPF_JMP | BPF_JSLT | BPF_K: 605 case BPF_JMP | BPF_JSGE | BPF_K: 606 case BPF_JMP | BPF_JSLE | BPF_K: 607 emit_a64_mov_i(1, tmp, imm, ctx); 608 emit(A64_CMP(1, dst, tmp), ctx); 609 goto emit_cond_jmp; 610 case BPF_JMP | BPF_JSET | BPF_K: 611 emit_a64_mov_i(1, tmp, imm, ctx); 612 emit(A64_TST(1, dst, tmp), ctx); 613 goto emit_cond_jmp; 614 /* function call */ 615 case BPF_JMP | BPF_CALL: 616 { 617 const u8 r0 = bpf2a64[BPF_REG_0]; 618 const u64 func = (u64)__bpf_call_base + imm; 619 620 if (ctx->prog->is_func) 621 emit_addr_mov_i64(tmp, func, ctx); 622 else 623 emit_a64_mov_i64(tmp, func, ctx); 624 emit(A64_BLR(tmp), ctx); 625 emit(A64_MOV(1, r0, A64_R(0)), ctx); 626 break; 627 } 628 /* tail call */ 629 case BPF_JMP | BPF_TAIL_CALL: 630 if (emit_bpf_tail_call(ctx)) 631 return -EFAULT; 632 break; 633 /* function return */ 634 case BPF_JMP | BPF_EXIT: 635 /* Optimization: when last instruction is EXIT, 636 simply fallthrough to epilogue. */ 637 if (i == ctx->prog->len - 1) 638 break; 639 jmp_offset = epilogue_offset(ctx); 640 check_imm26(jmp_offset); 641 emit(A64_B(jmp_offset), ctx); 642 break; 643 644 /* dst = imm64 */ 645 case BPF_LD | BPF_IMM | BPF_DW: 646 { 647 const struct bpf_insn insn1 = insn[1]; 648 u64 imm64; 649 650 imm64 = (u64)insn1.imm << 32 | (u32)imm; 651 emit_a64_mov_i64(dst, imm64, ctx); 652 653 return 1; 654 } 655 656 /* LDX: dst = *(size *)(src + off) */ 657 case BPF_LDX | BPF_MEM | BPF_W: 658 case BPF_LDX | BPF_MEM | BPF_H: 659 case BPF_LDX | BPF_MEM | BPF_B: 660 case BPF_LDX | BPF_MEM | BPF_DW: 661 emit_a64_mov_i(1, tmp, off, ctx); 662 switch (BPF_SIZE(code)) { 663 case BPF_W: 664 emit(A64_LDR32(dst, src, tmp), ctx); 665 break; 666 case BPF_H: 667 emit(A64_LDRH(dst, src, tmp), ctx); 668 break; 669 case BPF_B: 670 emit(A64_LDRB(dst, src, tmp), ctx); 671 break; 672 case BPF_DW: 673 emit(A64_LDR64(dst, src, tmp), ctx); 674 break; 675 } 676 break; 677 678 /* ST: *(size *)(dst + off) = imm */ 679 case BPF_ST | BPF_MEM | BPF_W: 680 case BPF_ST | BPF_MEM | BPF_H: 681 case BPF_ST | BPF_MEM | BPF_B: 682 case BPF_ST | BPF_MEM | BPF_DW: 683 /* Load imm to a register then store it */ 684 emit_a64_mov_i(1, tmp2, off, ctx); 685 emit_a64_mov_i(1, tmp, imm, ctx); 686 switch (BPF_SIZE(code)) { 687 case BPF_W: 688 emit(A64_STR32(tmp, dst, tmp2), ctx); 689 break; 690 case BPF_H: 691 emit(A64_STRH(tmp, dst, tmp2), ctx); 692 break; 693 case BPF_B: 694 emit(A64_STRB(tmp, dst, tmp2), ctx); 695 break; 696 case BPF_DW: 697 emit(A64_STR64(tmp, dst, tmp2), ctx); 698 break; 699 } 700 break; 701 702 /* STX: *(size *)(dst + off) = src */ 703 case BPF_STX | BPF_MEM | BPF_W: 704 case BPF_STX | BPF_MEM | BPF_H: 705 case BPF_STX | BPF_MEM | BPF_B: 706 case BPF_STX | BPF_MEM | BPF_DW: 707 emit_a64_mov_i(1, tmp, off, ctx); 708 switch (BPF_SIZE(code)) { 709 case BPF_W: 710 emit(A64_STR32(src, dst, tmp), ctx); 711 break; 712 case BPF_H: 713 emit(A64_STRH(src, dst, tmp), ctx); 714 break; 715 case BPF_B: 716 emit(A64_STRB(src, dst, tmp), ctx); 717 break; 718 case BPF_DW: 719 emit(A64_STR64(src, dst, tmp), ctx); 720 break; 721 } 722 break; 723 /* STX XADD: lock *(u32 *)(dst + off) += src */ 724 case BPF_STX | BPF_XADD | BPF_W: 725 /* STX XADD: lock *(u64 *)(dst + off) += src */ 726 case BPF_STX | BPF_XADD | BPF_DW: 727 emit_a64_mov_i(1, tmp, off, ctx); 728 emit(A64_ADD(1, tmp, tmp, dst), ctx); 729 emit(A64_PRFM(tmp, PST, L1, STRM), ctx); 730 emit(A64_LDXR(isdw, tmp2, tmp), ctx); 731 emit(A64_ADD(isdw, tmp2, tmp2, src), ctx); 732 emit(A64_STXR(isdw, tmp2, tmp, tmp3), ctx); 733 jmp_offset = -3; 734 check_imm19(jmp_offset); 735 emit(A64_CBNZ(0, tmp3, jmp_offset), ctx); 736 break; 737 738 /* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */ 739 case BPF_LD | BPF_ABS | BPF_W: 740 case BPF_LD | BPF_ABS | BPF_H: 741 case BPF_LD | BPF_ABS | BPF_B: 742 /* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */ 743 case BPF_LD | BPF_IND | BPF_W: 744 case BPF_LD | BPF_IND | BPF_H: 745 case BPF_LD | BPF_IND | BPF_B: 746 { 747 const u8 r0 = bpf2a64[BPF_REG_0]; /* r0 = return value */ 748 const u8 r6 = bpf2a64[BPF_REG_6]; /* r6 = pointer to sk_buff */ 749 const u8 fp = bpf2a64[BPF_REG_FP]; 750 const u8 r1 = bpf2a64[BPF_REG_1]; /* r1: struct sk_buff *skb */ 751 const u8 r2 = bpf2a64[BPF_REG_2]; /* r2: int k */ 752 const u8 r3 = bpf2a64[BPF_REG_3]; /* r3: unsigned int size */ 753 const u8 r4 = bpf2a64[BPF_REG_4]; /* r4: void *buffer */ 754 const u8 r5 = bpf2a64[BPF_REG_5]; /* r5: void *(*func)(...) */ 755 int size; 756 757 emit(A64_MOV(1, r1, r6), ctx); 758 emit_a64_mov_i(0, r2, imm, ctx); 759 if (BPF_MODE(code) == BPF_IND) 760 emit(A64_ADD(0, r2, r2, src), ctx); 761 switch (BPF_SIZE(code)) { 762 case BPF_W: 763 size = 4; 764 break; 765 case BPF_H: 766 size = 2; 767 break; 768 case BPF_B: 769 size = 1; 770 break; 771 default: 772 return -EINVAL; 773 } 774 emit_a64_mov_i64(r3, size, ctx); 775 emit(A64_SUB_I(1, r4, fp, ctx->stack_size), ctx); 776 emit_a64_mov_i64(r5, (unsigned long)bpf_load_pointer, ctx); 777 emit(A64_BLR(r5), ctx); 778 emit(A64_MOV(1, r0, A64_R(0)), ctx); 779 780 jmp_offset = epilogue_offset(ctx); 781 check_imm19(jmp_offset); 782 emit(A64_CBZ(1, r0, jmp_offset), ctx); 783 emit(A64_MOV(1, r5, r0), ctx); 784 switch (BPF_SIZE(code)) { 785 case BPF_W: 786 emit(A64_LDR32(r0, r5, A64_ZR), ctx); 787 #ifndef CONFIG_CPU_BIG_ENDIAN 788 emit(A64_REV32(0, r0, r0), ctx); 789 #endif 790 break; 791 case BPF_H: 792 emit(A64_LDRH(r0, r5, A64_ZR), ctx); 793 #ifndef CONFIG_CPU_BIG_ENDIAN 794 emit(A64_REV16(0, r0, r0), ctx); 795 #endif 796 break; 797 case BPF_B: 798 emit(A64_LDRB(r0, r5, A64_ZR), ctx); 799 break; 800 } 801 break; 802 } 803 default: 804 pr_err_once("unknown opcode %02x\n", code); 805 return -EINVAL; 806 } 807 808 return 0; 809 } 810 811 static int build_body(struct jit_ctx *ctx) 812 { 813 const struct bpf_prog *prog = ctx->prog; 814 int i; 815 816 for (i = 0; i < prog->len; i++) { 817 const struct bpf_insn *insn = &prog->insnsi[i]; 818 int ret; 819 820 ret = build_insn(insn, ctx); 821 if (ret > 0) { 822 i++; 823 if (ctx->image == NULL) 824 ctx->offset[i] = ctx->idx; 825 continue; 826 } 827 if (ctx->image == NULL) 828 ctx->offset[i] = ctx->idx; 829 if (ret) 830 return ret; 831 } 832 833 return 0; 834 } 835 836 static int validate_code(struct jit_ctx *ctx) 837 { 838 int i; 839 840 for (i = 0; i < ctx->idx; i++) { 841 u32 a64_insn = le32_to_cpu(ctx->image[i]); 842 843 if (a64_insn == AARCH64_BREAK_FAULT) 844 return -1; 845 } 846 847 return 0; 848 } 849 850 static inline void bpf_flush_icache(void *start, void *end) 851 { 852 flush_icache_range((unsigned long)start, (unsigned long)end); 853 } 854 855 struct arm64_jit_data { 856 struct bpf_binary_header *header; 857 u8 *image; 858 struct jit_ctx ctx; 859 }; 860 861 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) 862 { 863 struct bpf_prog *tmp, *orig_prog = prog; 864 struct bpf_binary_header *header; 865 struct arm64_jit_data *jit_data; 866 bool tmp_blinded = false; 867 bool extra_pass = false; 868 struct jit_ctx ctx; 869 int image_size; 870 u8 *image_ptr; 871 872 if (!prog->jit_requested) 873 return orig_prog; 874 875 tmp = bpf_jit_blind_constants(prog); 876 /* If blinding was requested and we failed during blinding, 877 * we must fall back to the interpreter. 878 */ 879 if (IS_ERR(tmp)) 880 return orig_prog; 881 if (tmp != prog) { 882 tmp_blinded = true; 883 prog = tmp; 884 } 885 886 jit_data = prog->aux->jit_data; 887 if (!jit_data) { 888 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL); 889 if (!jit_data) { 890 prog = orig_prog; 891 goto out; 892 } 893 prog->aux->jit_data = jit_data; 894 } 895 if (jit_data->ctx.offset) { 896 ctx = jit_data->ctx; 897 image_ptr = jit_data->image; 898 header = jit_data->header; 899 extra_pass = true; 900 image_size = sizeof(u32) * ctx.idx; 901 goto skip_init_ctx; 902 } 903 memset(&ctx, 0, sizeof(ctx)); 904 ctx.prog = prog; 905 906 ctx.offset = kcalloc(prog->len, sizeof(int), GFP_KERNEL); 907 if (ctx.offset == NULL) { 908 prog = orig_prog; 909 goto out_off; 910 } 911 912 /* 1. Initial fake pass to compute ctx->idx. */ 913 914 /* Fake pass to fill in ctx->offset. */ 915 if (build_body(&ctx)) { 916 prog = orig_prog; 917 goto out_off; 918 } 919 920 if (build_prologue(&ctx)) { 921 prog = orig_prog; 922 goto out_off; 923 } 924 925 ctx.epilogue_offset = ctx.idx; 926 build_epilogue(&ctx); 927 928 /* Now we know the actual image size. */ 929 image_size = sizeof(u32) * ctx.idx; 930 header = bpf_jit_binary_alloc(image_size, &image_ptr, 931 sizeof(u32), jit_fill_hole); 932 if (header == NULL) { 933 prog = orig_prog; 934 goto out_off; 935 } 936 937 /* 2. Now, the actual pass. */ 938 939 ctx.image = (__le32 *)image_ptr; 940 skip_init_ctx: 941 ctx.idx = 0; 942 943 build_prologue(&ctx); 944 945 if (build_body(&ctx)) { 946 bpf_jit_binary_free(header); 947 prog = orig_prog; 948 goto out_off; 949 } 950 951 build_epilogue(&ctx); 952 953 /* 3. Extra pass to validate JITed code. */ 954 if (validate_code(&ctx)) { 955 bpf_jit_binary_free(header); 956 prog = orig_prog; 957 goto out_off; 958 } 959 960 /* And we're done. */ 961 if (bpf_jit_enable > 1) 962 bpf_jit_dump(prog->len, image_size, 2, ctx.image); 963 964 bpf_flush_icache(header, ctx.image + ctx.idx); 965 966 if (!prog->is_func || extra_pass) { 967 if (extra_pass && ctx.idx != jit_data->ctx.idx) { 968 pr_err_once("multi-func JIT bug %d != %d\n", 969 ctx.idx, jit_data->ctx.idx); 970 bpf_jit_binary_free(header); 971 prog->bpf_func = NULL; 972 prog->jited = 0; 973 goto out_off; 974 } 975 bpf_jit_binary_lock_ro(header); 976 } else { 977 jit_data->ctx = ctx; 978 jit_data->image = image_ptr; 979 jit_data->header = header; 980 } 981 prog->bpf_func = (void *)ctx.image; 982 prog->jited = 1; 983 prog->jited_len = image_size; 984 985 if (!prog->is_func || extra_pass) { 986 out_off: 987 kfree(ctx.offset); 988 kfree(jit_data); 989 prog->aux->jit_data = NULL; 990 } 991 out: 992 if (tmp_blinded) 993 bpf_jit_prog_release_other(prog, prog == orig_prog ? 994 tmp : orig_prog); 995 return prog; 996 } 997