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/slab.h> 25 26 #include <asm/byteorder.h> 27 #include <asm/cacheflush.h> 28 #include <asm/debug-monitors.h> 29 #include <asm/set_memory.h> 30 31 #include "bpf_jit.h" 32 33 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0) 34 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1) 35 #define TCALL_CNT (MAX_BPF_JIT_REG + 2) 36 #define TMP_REG_3 (MAX_BPF_JIT_REG + 3) 37 38 /* Map BPF registers to A64 registers */ 39 static const int bpf2a64[] = { 40 /* return value from in-kernel function, and exit value from eBPF */ 41 [BPF_REG_0] = A64_R(7), 42 /* arguments from eBPF program to in-kernel function */ 43 [BPF_REG_1] = A64_R(0), 44 [BPF_REG_2] = A64_R(1), 45 [BPF_REG_3] = A64_R(2), 46 [BPF_REG_4] = A64_R(3), 47 [BPF_REG_5] = A64_R(4), 48 /* callee saved registers that in-kernel function will preserve */ 49 [BPF_REG_6] = A64_R(19), 50 [BPF_REG_7] = A64_R(20), 51 [BPF_REG_8] = A64_R(21), 52 [BPF_REG_9] = A64_R(22), 53 /* read-only frame pointer to access stack */ 54 [BPF_REG_FP] = A64_R(25), 55 /* temporary registers for internal BPF JIT */ 56 [TMP_REG_1] = A64_R(10), 57 [TMP_REG_2] = A64_R(11), 58 [TMP_REG_3] = A64_R(12), 59 /* tail_call_cnt */ 60 [TCALL_CNT] = A64_R(26), 61 /* temporary register for blinding constants */ 62 [BPF_REG_AX] = A64_R(9), 63 }; 64 65 struct jit_ctx { 66 const struct bpf_prog *prog; 67 int idx; 68 int epilogue_offset; 69 int *offset; 70 __le32 *image; 71 u32 stack_size; 72 }; 73 74 static inline void emit(const u32 insn, struct jit_ctx *ctx) 75 { 76 if (ctx->image != NULL) 77 ctx->image[ctx->idx] = cpu_to_le32(insn); 78 79 ctx->idx++; 80 } 81 82 static inline void emit_a64_mov_i(const int is64, const int reg, 83 const s32 val, struct jit_ctx *ctx) 84 { 85 u16 hi = val >> 16; 86 u16 lo = val & 0xffff; 87 88 if (hi & 0x8000) { 89 if (hi == 0xffff) { 90 emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx); 91 } else { 92 emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx); 93 if (lo != 0xffff) 94 emit(A64_MOVK(is64, reg, lo, 0), ctx); 95 } 96 } else { 97 emit(A64_MOVZ(is64, reg, lo, 0), ctx); 98 if (hi) 99 emit(A64_MOVK(is64, reg, hi, 16), ctx); 100 } 101 } 102 103 static int i64_i16_blocks(const u64 val, bool inverse) 104 { 105 return (((val >> 0) & 0xffff) != (inverse ? 0xffff : 0x0000)) + 106 (((val >> 16) & 0xffff) != (inverse ? 0xffff : 0x0000)) + 107 (((val >> 32) & 0xffff) != (inverse ? 0xffff : 0x0000)) + 108 (((val >> 48) & 0xffff) != (inverse ? 0xffff : 0x0000)); 109 } 110 111 static inline void emit_a64_mov_i64(const int reg, const u64 val, 112 struct jit_ctx *ctx) 113 { 114 u64 nrm_tmp = val, rev_tmp = ~val; 115 bool inverse; 116 int shift; 117 118 if (!(nrm_tmp >> 32)) 119 return emit_a64_mov_i(0, reg, (u32)val, ctx); 120 121 inverse = i64_i16_blocks(nrm_tmp, true) < i64_i16_blocks(nrm_tmp, false); 122 shift = max(round_down((inverse ? (fls64(rev_tmp) - 1) : 123 (fls64(nrm_tmp) - 1)), 16), 0); 124 if (inverse) 125 emit(A64_MOVN(1, reg, (rev_tmp >> shift) & 0xffff, shift), ctx); 126 else 127 emit(A64_MOVZ(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx); 128 shift -= 16; 129 while (shift >= 0) { 130 if (((nrm_tmp >> shift) & 0xffff) != (inverse ? 0xffff : 0x0000)) 131 emit(A64_MOVK(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx); 132 shift -= 16; 133 } 134 } 135 136 /* 137 * This is an unoptimized 64 immediate emission used for BPF to BPF call 138 * addresses. It will always do a full 64 bit decomposition as otherwise 139 * more complexity in the last extra pass is required since we previously 140 * reserved 4 instructions for the address. 141 */ 142 static inline void emit_addr_mov_i64(const int reg, const u64 val, 143 struct jit_ctx *ctx) 144 { 145 u64 tmp = val; 146 int shift = 0; 147 148 emit(A64_MOVZ(1, reg, tmp & 0xffff, shift), ctx); 149 for (;shift < 48;) { 150 tmp >>= 16; 151 shift += 16; 152 emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx); 153 } 154 } 155 156 static inline int bpf2a64_offset(int bpf_to, int bpf_from, 157 const struct jit_ctx *ctx) 158 { 159 int to = ctx->offset[bpf_to]; 160 /* -1 to account for the Branch instruction */ 161 int from = ctx->offset[bpf_from] - 1; 162 163 return to - from; 164 } 165 166 static void jit_fill_hole(void *area, unsigned int size) 167 { 168 __le32 *ptr; 169 /* We are guaranteed to have aligned memory. */ 170 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32)) 171 *ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT); 172 } 173 174 static inline int epilogue_offset(const struct jit_ctx *ctx) 175 { 176 int to = ctx->epilogue_offset; 177 int from = ctx->idx; 178 179 return to - from; 180 } 181 182 /* Stack must be multiples of 16B */ 183 #define STACK_ALIGN(sz) (((sz) + 15) & ~15) 184 185 /* Tail call offset to jump into */ 186 #define PROLOGUE_OFFSET 7 187 188 static int build_prologue(struct jit_ctx *ctx, bool ebpf_from_cbpf) 189 { 190 const struct bpf_prog *prog = ctx->prog; 191 const u8 r6 = bpf2a64[BPF_REG_6]; 192 const u8 r7 = bpf2a64[BPF_REG_7]; 193 const u8 r8 = bpf2a64[BPF_REG_8]; 194 const u8 r9 = bpf2a64[BPF_REG_9]; 195 const u8 fp = bpf2a64[BPF_REG_FP]; 196 const u8 tcc = bpf2a64[TCALL_CNT]; 197 const int idx0 = ctx->idx; 198 int cur_offset; 199 200 /* 201 * BPF prog stack layout 202 * 203 * high 204 * original A64_SP => 0:+-----+ BPF prologue 205 * |FP/LR| 206 * current A64_FP => -16:+-----+ 207 * | ... | callee saved registers 208 * BPF fp register => -64:+-----+ <= (BPF_FP) 209 * | | 210 * | ... | BPF prog stack 211 * | | 212 * +-----+ <= (BPF_FP - prog->aux->stack_depth) 213 * |RSVD | padding 214 * current A64_SP => +-----+ <= (BPF_FP - ctx->stack_size) 215 * | | 216 * | ... | Function call stack 217 * | | 218 * +-----+ 219 * low 220 * 221 */ 222 223 /* Save FP and LR registers to stay align with ARM64 AAPCS */ 224 emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx); 225 emit(A64_MOV(1, A64_FP, A64_SP), ctx); 226 227 /* Save callee-saved registers */ 228 emit(A64_PUSH(r6, r7, A64_SP), ctx); 229 emit(A64_PUSH(r8, r9, A64_SP), ctx); 230 emit(A64_PUSH(fp, tcc, A64_SP), ctx); 231 232 /* Set up BPF prog stack base register */ 233 emit(A64_MOV(1, fp, A64_SP), ctx); 234 235 if (!ebpf_from_cbpf) { 236 /* Initialize tail_call_cnt */ 237 emit(A64_MOVZ(1, tcc, 0, 0), ctx); 238 239 cur_offset = ctx->idx - idx0; 240 if (cur_offset != PROLOGUE_OFFSET) { 241 pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n", 242 cur_offset, PROLOGUE_OFFSET); 243 return -1; 244 } 245 } 246 247 ctx->stack_size = STACK_ALIGN(prog->aux->stack_depth); 248 249 /* Set up function call stack */ 250 emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 251 return 0; 252 } 253 254 static int out_offset = -1; /* initialized on the first pass of build_body() */ 255 static int emit_bpf_tail_call(struct jit_ctx *ctx) 256 { 257 /* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */ 258 const u8 r2 = bpf2a64[BPF_REG_2]; 259 const u8 r3 = bpf2a64[BPF_REG_3]; 260 261 const u8 tmp = bpf2a64[TMP_REG_1]; 262 const u8 prg = bpf2a64[TMP_REG_2]; 263 const u8 tcc = bpf2a64[TCALL_CNT]; 264 const int idx0 = ctx->idx; 265 #define cur_offset (ctx->idx - idx0) 266 #define jmp_offset (out_offset - (cur_offset)) 267 size_t off; 268 269 /* if (index >= array->map.max_entries) 270 * goto out; 271 */ 272 off = offsetof(struct bpf_array, map.max_entries); 273 emit_a64_mov_i64(tmp, off, ctx); 274 emit(A64_LDR32(tmp, r2, tmp), ctx); 275 emit(A64_MOV(0, r3, r3), ctx); 276 emit(A64_CMP(0, r3, tmp), ctx); 277 emit(A64_B_(A64_COND_CS, jmp_offset), ctx); 278 279 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT) 280 * goto out; 281 * tail_call_cnt++; 282 */ 283 emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx); 284 emit(A64_CMP(1, tcc, tmp), ctx); 285 emit(A64_B_(A64_COND_HI, jmp_offset), ctx); 286 emit(A64_ADD_I(1, tcc, tcc, 1), ctx); 287 288 /* prog = array->ptrs[index]; 289 * if (prog == NULL) 290 * goto out; 291 */ 292 off = offsetof(struct bpf_array, ptrs); 293 emit_a64_mov_i64(tmp, off, ctx); 294 emit(A64_ADD(1, tmp, r2, tmp), ctx); 295 emit(A64_LSL(1, prg, r3, 3), ctx); 296 emit(A64_LDR64(prg, tmp, prg), ctx); 297 emit(A64_CBZ(1, prg, jmp_offset), ctx); 298 299 /* goto *(prog->bpf_func + prologue_offset); */ 300 off = offsetof(struct bpf_prog, bpf_func); 301 emit_a64_mov_i64(tmp, off, ctx); 302 emit(A64_LDR64(tmp, prg, tmp), ctx); 303 emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx); 304 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 305 emit(A64_BR(tmp), ctx); 306 307 /* out: */ 308 if (out_offset == -1) 309 out_offset = cur_offset; 310 if (cur_offset != out_offset) { 311 pr_err_once("tail_call out_offset = %d, expected %d!\n", 312 cur_offset, out_offset); 313 return -1; 314 } 315 return 0; 316 #undef cur_offset 317 #undef jmp_offset 318 } 319 320 static void build_epilogue(struct jit_ctx *ctx) 321 { 322 const u8 r0 = bpf2a64[BPF_REG_0]; 323 const u8 r6 = bpf2a64[BPF_REG_6]; 324 const u8 r7 = bpf2a64[BPF_REG_7]; 325 const u8 r8 = bpf2a64[BPF_REG_8]; 326 const u8 r9 = bpf2a64[BPF_REG_9]; 327 const u8 fp = bpf2a64[BPF_REG_FP]; 328 329 /* We're done with BPF stack */ 330 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx); 331 332 /* Restore fs (x25) and x26 */ 333 emit(A64_POP(fp, A64_R(26), A64_SP), ctx); 334 335 /* Restore callee-saved register */ 336 emit(A64_POP(r8, r9, A64_SP), ctx); 337 emit(A64_POP(r6, r7, A64_SP), ctx); 338 339 /* Restore FP/LR registers */ 340 emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx); 341 342 /* Set return value */ 343 emit(A64_MOV(1, A64_R(0), r0), ctx); 344 345 emit(A64_RET(A64_LR), ctx); 346 } 347 348 /* JITs an eBPF instruction. 349 * Returns: 350 * 0 - successfully JITed an 8-byte eBPF instruction. 351 * >0 - successfully JITed a 16-byte eBPF instruction. 352 * <0 - failed to JIT. 353 */ 354 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx, 355 bool extra_pass) 356 { 357 const u8 code = insn->code; 358 const u8 dst = bpf2a64[insn->dst_reg]; 359 const u8 src = bpf2a64[insn->src_reg]; 360 const u8 tmp = bpf2a64[TMP_REG_1]; 361 const u8 tmp2 = bpf2a64[TMP_REG_2]; 362 const u8 tmp3 = bpf2a64[TMP_REG_3]; 363 const s16 off = insn->off; 364 const s32 imm = insn->imm; 365 const int i = insn - ctx->prog->insnsi; 366 const bool is64 = BPF_CLASS(code) == BPF_ALU64; 367 const bool isdw = BPF_SIZE(code) == BPF_DW; 368 u8 jmp_cond; 369 s32 jmp_offset; 370 371 #define check_imm(bits, imm) do { \ 372 if ((((imm) > 0) && ((imm) >> (bits))) || \ 373 (((imm) < 0) && (~(imm) >> (bits)))) { \ 374 pr_info("[%2d] imm=%d(0x%x) out of range\n", \ 375 i, imm, imm); \ 376 return -EINVAL; \ 377 } \ 378 } while (0) 379 #define check_imm19(imm) check_imm(19, imm) 380 #define check_imm26(imm) check_imm(26, imm) 381 382 switch (code) { 383 /* dst = src */ 384 case BPF_ALU | BPF_MOV | BPF_X: 385 case BPF_ALU64 | BPF_MOV | BPF_X: 386 emit(A64_MOV(is64, dst, src), ctx); 387 break; 388 /* dst = dst OP src */ 389 case BPF_ALU | BPF_ADD | BPF_X: 390 case BPF_ALU64 | BPF_ADD | BPF_X: 391 emit(A64_ADD(is64, dst, dst, src), ctx); 392 break; 393 case BPF_ALU | BPF_SUB | BPF_X: 394 case BPF_ALU64 | BPF_SUB | BPF_X: 395 emit(A64_SUB(is64, dst, dst, src), ctx); 396 break; 397 case BPF_ALU | BPF_AND | BPF_X: 398 case BPF_ALU64 | BPF_AND | BPF_X: 399 emit(A64_AND(is64, dst, dst, src), ctx); 400 break; 401 case BPF_ALU | BPF_OR | BPF_X: 402 case BPF_ALU64 | BPF_OR | BPF_X: 403 emit(A64_ORR(is64, dst, dst, src), ctx); 404 break; 405 case BPF_ALU | BPF_XOR | BPF_X: 406 case BPF_ALU64 | BPF_XOR | BPF_X: 407 emit(A64_EOR(is64, dst, dst, src), ctx); 408 break; 409 case BPF_ALU | BPF_MUL | BPF_X: 410 case BPF_ALU64 | BPF_MUL | BPF_X: 411 emit(A64_MUL(is64, dst, dst, src), ctx); 412 break; 413 case BPF_ALU | BPF_DIV | BPF_X: 414 case BPF_ALU64 | BPF_DIV | BPF_X: 415 case BPF_ALU | BPF_MOD | BPF_X: 416 case BPF_ALU64 | BPF_MOD | BPF_X: 417 switch (BPF_OP(code)) { 418 case BPF_DIV: 419 emit(A64_UDIV(is64, dst, dst, src), ctx); 420 break; 421 case BPF_MOD: 422 emit(A64_UDIV(is64, tmp, dst, src), ctx); 423 emit(A64_MUL(is64, tmp, tmp, src), ctx); 424 emit(A64_SUB(is64, dst, dst, tmp), ctx); 425 break; 426 } 427 break; 428 case BPF_ALU | BPF_LSH | BPF_X: 429 case BPF_ALU64 | BPF_LSH | BPF_X: 430 emit(A64_LSLV(is64, dst, dst, src), ctx); 431 break; 432 case BPF_ALU | BPF_RSH | BPF_X: 433 case BPF_ALU64 | BPF_RSH | BPF_X: 434 emit(A64_LSRV(is64, dst, dst, src), ctx); 435 break; 436 case BPF_ALU | BPF_ARSH | BPF_X: 437 case BPF_ALU64 | BPF_ARSH | BPF_X: 438 emit(A64_ASRV(is64, dst, dst, src), ctx); 439 break; 440 /* dst = -dst */ 441 case BPF_ALU | BPF_NEG: 442 case BPF_ALU64 | BPF_NEG: 443 emit(A64_NEG(is64, dst, dst), ctx); 444 break; 445 /* dst = BSWAP##imm(dst) */ 446 case BPF_ALU | BPF_END | BPF_FROM_LE: 447 case BPF_ALU | BPF_END | BPF_FROM_BE: 448 #ifdef CONFIG_CPU_BIG_ENDIAN 449 if (BPF_SRC(code) == BPF_FROM_BE) 450 goto emit_bswap_uxt; 451 #else /* !CONFIG_CPU_BIG_ENDIAN */ 452 if (BPF_SRC(code) == BPF_FROM_LE) 453 goto emit_bswap_uxt; 454 #endif 455 switch (imm) { 456 case 16: 457 emit(A64_REV16(is64, dst, dst), ctx); 458 /* zero-extend 16 bits into 64 bits */ 459 emit(A64_UXTH(is64, dst, dst), ctx); 460 break; 461 case 32: 462 emit(A64_REV32(is64, dst, dst), ctx); 463 /* upper 32 bits already cleared */ 464 break; 465 case 64: 466 emit(A64_REV64(dst, dst), ctx); 467 break; 468 } 469 break; 470 emit_bswap_uxt: 471 switch (imm) { 472 case 16: 473 /* zero-extend 16 bits into 64 bits */ 474 emit(A64_UXTH(is64, dst, dst), ctx); 475 break; 476 case 32: 477 /* zero-extend 32 bits into 64 bits */ 478 emit(A64_UXTW(is64, dst, dst), ctx); 479 break; 480 case 64: 481 /* nop */ 482 break; 483 } 484 break; 485 /* dst = imm */ 486 case BPF_ALU | BPF_MOV | BPF_K: 487 case BPF_ALU64 | BPF_MOV | BPF_K: 488 emit_a64_mov_i(is64, dst, imm, ctx); 489 break; 490 /* dst = dst OP imm */ 491 case BPF_ALU | BPF_ADD | BPF_K: 492 case BPF_ALU64 | BPF_ADD | BPF_K: 493 emit_a64_mov_i(is64, tmp, imm, ctx); 494 emit(A64_ADD(is64, dst, dst, tmp), ctx); 495 break; 496 case BPF_ALU | BPF_SUB | BPF_K: 497 case BPF_ALU64 | BPF_SUB | BPF_K: 498 emit_a64_mov_i(is64, tmp, imm, ctx); 499 emit(A64_SUB(is64, dst, dst, tmp), ctx); 500 break; 501 case BPF_ALU | BPF_AND | BPF_K: 502 case BPF_ALU64 | BPF_AND | BPF_K: 503 emit_a64_mov_i(is64, tmp, imm, ctx); 504 emit(A64_AND(is64, dst, dst, tmp), ctx); 505 break; 506 case BPF_ALU | BPF_OR | BPF_K: 507 case BPF_ALU64 | BPF_OR | BPF_K: 508 emit_a64_mov_i(is64, tmp, imm, ctx); 509 emit(A64_ORR(is64, dst, dst, tmp), ctx); 510 break; 511 case BPF_ALU | BPF_XOR | BPF_K: 512 case BPF_ALU64 | BPF_XOR | BPF_K: 513 emit_a64_mov_i(is64, tmp, imm, ctx); 514 emit(A64_EOR(is64, dst, dst, tmp), ctx); 515 break; 516 case BPF_ALU | BPF_MUL | BPF_K: 517 case BPF_ALU64 | BPF_MUL | BPF_K: 518 emit_a64_mov_i(is64, tmp, imm, ctx); 519 emit(A64_MUL(is64, dst, dst, tmp), ctx); 520 break; 521 case BPF_ALU | BPF_DIV | BPF_K: 522 case BPF_ALU64 | BPF_DIV | BPF_K: 523 emit_a64_mov_i(is64, tmp, imm, ctx); 524 emit(A64_UDIV(is64, dst, dst, tmp), ctx); 525 break; 526 case BPF_ALU | BPF_MOD | BPF_K: 527 case BPF_ALU64 | BPF_MOD | BPF_K: 528 emit_a64_mov_i(is64, tmp2, imm, ctx); 529 emit(A64_UDIV(is64, tmp, dst, tmp2), ctx); 530 emit(A64_MUL(is64, tmp, tmp, tmp2), ctx); 531 emit(A64_SUB(is64, dst, dst, tmp), ctx); 532 break; 533 case BPF_ALU | BPF_LSH | BPF_K: 534 case BPF_ALU64 | BPF_LSH | BPF_K: 535 emit(A64_LSL(is64, dst, dst, imm), ctx); 536 break; 537 case BPF_ALU | BPF_RSH | BPF_K: 538 case BPF_ALU64 | BPF_RSH | BPF_K: 539 emit(A64_LSR(is64, dst, dst, imm), ctx); 540 break; 541 case BPF_ALU | BPF_ARSH | BPF_K: 542 case BPF_ALU64 | BPF_ARSH | BPF_K: 543 emit(A64_ASR(is64, dst, dst, imm), ctx); 544 break; 545 546 /* JUMP off */ 547 case BPF_JMP | BPF_JA: 548 jmp_offset = bpf2a64_offset(i + off, i, ctx); 549 check_imm26(jmp_offset); 550 emit(A64_B(jmp_offset), ctx); 551 break; 552 /* IF (dst COND src) JUMP off */ 553 case BPF_JMP | BPF_JEQ | BPF_X: 554 case BPF_JMP | BPF_JGT | BPF_X: 555 case BPF_JMP | BPF_JLT | BPF_X: 556 case BPF_JMP | BPF_JGE | BPF_X: 557 case BPF_JMP | BPF_JLE | BPF_X: 558 case BPF_JMP | BPF_JNE | BPF_X: 559 case BPF_JMP | BPF_JSGT | BPF_X: 560 case BPF_JMP | BPF_JSLT | BPF_X: 561 case BPF_JMP | BPF_JSGE | BPF_X: 562 case BPF_JMP | BPF_JSLE | BPF_X: 563 emit(A64_CMP(1, dst, src), ctx); 564 emit_cond_jmp: 565 jmp_offset = bpf2a64_offset(i + off, i, ctx); 566 check_imm19(jmp_offset); 567 switch (BPF_OP(code)) { 568 case BPF_JEQ: 569 jmp_cond = A64_COND_EQ; 570 break; 571 case BPF_JGT: 572 jmp_cond = A64_COND_HI; 573 break; 574 case BPF_JLT: 575 jmp_cond = A64_COND_CC; 576 break; 577 case BPF_JGE: 578 jmp_cond = A64_COND_CS; 579 break; 580 case BPF_JLE: 581 jmp_cond = A64_COND_LS; 582 break; 583 case BPF_JSET: 584 case BPF_JNE: 585 jmp_cond = A64_COND_NE; 586 break; 587 case BPF_JSGT: 588 jmp_cond = A64_COND_GT; 589 break; 590 case BPF_JSLT: 591 jmp_cond = A64_COND_LT; 592 break; 593 case BPF_JSGE: 594 jmp_cond = A64_COND_GE; 595 break; 596 case BPF_JSLE: 597 jmp_cond = A64_COND_LE; 598 break; 599 default: 600 return -EFAULT; 601 } 602 emit(A64_B_(jmp_cond, jmp_offset), ctx); 603 break; 604 case BPF_JMP | BPF_JSET | BPF_X: 605 emit(A64_TST(1, dst, src), ctx); 606 goto emit_cond_jmp; 607 /* IF (dst COND imm) JUMP off */ 608 case BPF_JMP | BPF_JEQ | BPF_K: 609 case BPF_JMP | BPF_JGT | BPF_K: 610 case BPF_JMP | BPF_JLT | BPF_K: 611 case BPF_JMP | BPF_JGE | BPF_K: 612 case BPF_JMP | BPF_JLE | BPF_K: 613 case BPF_JMP | BPF_JNE | BPF_K: 614 case BPF_JMP | BPF_JSGT | BPF_K: 615 case BPF_JMP | BPF_JSLT | BPF_K: 616 case BPF_JMP | BPF_JSGE | BPF_K: 617 case BPF_JMP | BPF_JSLE | BPF_K: 618 emit_a64_mov_i(1, tmp, imm, ctx); 619 emit(A64_CMP(1, dst, tmp), ctx); 620 goto emit_cond_jmp; 621 case BPF_JMP | BPF_JSET | BPF_K: 622 emit_a64_mov_i(1, tmp, imm, ctx); 623 emit(A64_TST(1, dst, tmp), ctx); 624 goto emit_cond_jmp; 625 /* function call */ 626 case BPF_JMP | BPF_CALL: 627 { 628 const u8 r0 = bpf2a64[BPF_REG_0]; 629 bool func_addr_fixed; 630 u64 func_addr; 631 int ret; 632 633 ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass, 634 &func_addr, &func_addr_fixed); 635 if (ret < 0) 636 return ret; 637 if (func_addr_fixed) 638 /* We can use optimized emission here. */ 639 emit_a64_mov_i64(tmp, func_addr, ctx); 640 else 641 emit_addr_mov_i64(tmp, func_addr, ctx); 642 emit(A64_BLR(tmp), ctx); 643 emit(A64_MOV(1, r0, A64_R(0)), ctx); 644 break; 645 } 646 /* tail call */ 647 case BPF_JMP | BPF_TAIL_CALL: 648 if (emit_bpf_tail_call(ctx)) 649 return -EFAULT; 650 break; 651 /* function return */ 652 case BPF_JMP | BPF_EXIT: 653 /* Optimization: when last instruction is EXIT, 654 simply fallthrough to epilogue. */ 655 if (i == ctx->prog->len - 1) 656 break; 657 jmp_offset = epilogue_offset(ctx); 658 check_imm26(jmp_offset); 659 emit(A64_B(jmp_offset), ctx); 660 break; 661 662 /* dst = imm64 */ 663 case BPF_LD | BPF_IMM | BPF_DW: 664 { 665 const struct bpf_insn insn1 = insn[1]; 666 u64 imm64; 667 668 imm64 = (u64)insn1.imm << 32 | (u32)imm; 669 emit_a64_mov_i64(dst, imm64, ctx); 670 671 return 1; 672 } 673 674 /* LDX: dst = *(size *)(src + off) */ 675 case BPF_LDX | BPF_MEM | BPF_W: 676 case BPF_LDX | BPF_MEM | BPF_H: 677 case BPF_LDX | BPF_MEM | BPF_B: 678 case BPF_LDX | BPF_MEM | BPF_DW: 679 emit_a64_mov_i(1, tmp, off, ctx); 680 switch (BPF_SIZE(code)) { 681 case BPF_W: 682 emit(A64_LDR32(dst, src, tmp), ctx); 683 break; 684 case BPF_H: 685 emit(A64_LDRH(dst, src, tmp), ctx); 686 break; 687 case BPF_B: 688 emit(A64_LDRB(dst, src, tmp), ctx); 689 break; 690 case BPF_DW: 691 emit(A64_LDR64(dst, src, tmp), ctx); 692 break; 693 } 694 break; 695 696 /* ST: *(size *)(dst + off) = imm */ 697 case BPF_ST | BPF_MEM | BPF_W: 698 case BPF_ST | BPF_MEM | BPF_H: 699 case BPF_ST | BPF_MEM | BPF_B: 700 case BPF_ST | BPF_MEM | BPF_DW: 701 /* Load imm to a register then store it */ 702 emit_a64_mov_i(1, tmp2, off, ctx); 703 emit_a64_mov_i(1, tmp, imm, ctx); 704 switch (BPF_SIZE(code)) { 705 case BPF_W: 706 emit(A64_STR32(tmp, dst, tmp2), ctx); 707 break; 708 case BPF_H: 709 emit(A64_STRH(tmp, dst, tmp2), ctx); 710 break; 711 case BPF_B: 712 emit(A64_STRB(tmp, dst, tmp2), ctx); 713 break; 714 case BPF_DW: 715 emit(A64_STR64(tmp, dst, tmp2), ctx); 716 break; 717 } 718 break; 719 720 /* STX: *(size *)(dst + off) = src */ 721 case BPF_STX | BPF_MEM | BPF_W: 722 case BPF_STX | BPF_MEM | BPF_H: 723 case BPF_STX | BPF_MEM | BPF_B: 724 case BPF_STX | BPF_MEM | BPF_DW: 725 emit_a64_mov_i(1, tmp, off, ctx); 726 switch (BPF_SIZE(code)) { 727 case BPF_W: 728 emit(A64_STR32(src, dst, tmp), ctx); 729 break; 730 case BPF_H: 731 emit(A64_STRH(src, dst, tmp), ctx); 732 break; 733 case BPF_B: 734 emit(A64_STRB(src, dst, tmp), ctx); 735 break; 736 case BPF_DW: 737 emit(A64_STR64(src, dst, tmp), ctx); 738 break; 739 } 740 break; 741 /* STX XADD: lock *(u32 *)(dst + off) += src */ 742 case BPF_STX | BPF_XADD | BPF_W: 743 /* STX XADD: lock *(u64 *)(dst + off) += src */ 744 case BPF_STX | BPF_XADD | BPF_DW: 745 emit_a64_mov_i(1, tmp, off, ctx); 746 emit(A64_ADD(1, tmp, tmp, dst), ctx); 747 emit(A64_PRFM(tmp, PST, L1, STRM), ctx); 748 emit(A64_LDXR(isdw, tmp2, tmp), ctx); 749 emit(A64_ADD(isdw, tmp2, tmp2, src), ctx); 750 emit(A64_STXR(isdw, tmp2, tmp, tmp3), ctx); 751 jmp_offset = -3; 752 check_imm19(jmp_offset); 753 emit(A64_CBNZ(0, tmp3, jmp_offset), ctx); 754 break; 755 756 default: 757 pr_err_once("unknown opcode %02x\n", code); 758 return -EINVAL; 759 } 760 761 return 0; 762 } 763 764 static int build_body(struct jit_ctx *ctx, bool extra_pass) 765 { 766 const struct bpf_prog *prog = ctx->prog; 767 int i; 768 769 for (i = 0; i < prog->len; i++) { 770 const struct bpf_insn *insn = &prog->insnsi[i]; 771 int ret; 772 773 ret = build_insn(insn, ctx, extra_pass); 774 if (ret > 0) { 775 i++; 776 if (ctx->image == NULL) 777 ctx->offset[i] = ctx->idx; 778 continue; 779 } 780 if (ctx->image == NULL) 781 ctx->offset[i] = ctx->idx; 782 if (ret) 783 return ret; 784 } 785 786 return 0; 787 } 788 789 static int validate_code(struct jit_ctx *ctx) 790 { 791 int i; 792 793 for (i = 0; i < ctx->idx; i++) { 794 u32 a64_insn = le32_to_cpu(ctx->image[i]); 795 796 if (a64_insn == AARCH64_BREAK_FAULT) 797 return -1; 798 } 799 800 return 0; 801 } 802 803 static inline void bpf_flush_icache(void *start, void *end) 804 { 805 flush_icache_range((unsigned long)start, (unsigned long)end); 806 } 807 808 struct arm64_jit_data { 809 struct bpf_binary_header *header; 810 u8 *image; 811 struct jit_ctx ctx; 812 }; 813 814 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) 815 { 816 struct bpf_prog *tmp, *orig_prog = prog; 817 struct bpf_binary_header *header; 818 struct arm64_jit_data *jit_data; 819 bool was_classic = bpf_prog_was_classic(prog); 820 bool tmp_blinded = false; 821 bool extra_pass = false; 822 struct jit_ctx ctx; 823 int image_size; 824 u8 *image_ptr; 825 826 if (!prog->jit_requested) 827 return orig_prog; 828 829 tmp = bpf_jit_blind_constants(prog); 830 /* If blinding was requested and we failed during blinding, 831 * we must fall back to the interpreter. 832 */ 833 if (IS_ERR(tmp)) 834 return orig_prog; 835 if (tmp != prog) { 836 tmp_blinded = true; 837 prog = tmp; 838 } 839 840 jit_data = prog->aux->jit_data; 841 if (!jit_data) { 842 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL); 843 if (!jit_data) { 844 prog = orig_prog; 845 goto out; 846 } 847 prog->aux->jit_data = jit_data; 848 } 849 if (jit_data->ctx.offset) { 850 ctx = jit_data->ctx; 851 image_ptr = jit_data->image; 852 header = jit_data->header; 853 extra_pass = true; 854 image_size = sizeof(u32) * ctx.idx; 855 goto skip_init_ctx; 856 } 857 memset(&ctx, 0, sizeof(ctx)); 858 ctx.prog = prog; 859 860 ctx.offset = kcalloc(prog->len, sizeof(int), GFP_KERNEL); 861 if (ctx.offset == NULL) { 862 prog = orig_prog; 863 goto out_off; 864 } 865 866 /* 1. Initial fake pass to compute ctx->idx. */ 867 868 /* Fake pass to fill in ctx->offset. */ 869 if (build_body(&ctx, extra_pass)) { 870 prog = orig_prog; 871 goto out_off; 872 } 873 874 if (build_prologue(&ctx, was_classic)) { 875 prog = orig_prog; 876 goto out_off; 877 } 878 879 ctx.epilogue_offset = ctx.idx; 880 build_epilogue(&ctx); 881 882 /* Now we know the actual image size. */ 883 image_size = sizeof(u32) * ctx.idx; 884 header = bpf_jit_binary_alloc(image_size, &image_ptr, 885 sizeof(u32), jit_fill_hole); 886 if (header == NULL) { 887 prog = orig_prog; 888 goto out_off; 889 } 890 891 /* 2. Now, the actual pass. */ 892 893 ctx.image = (__le32 *)image_ptr; 894 skip_init_ctx: 895 ctx.idx = 0; 896 897 build_prologue(&ctx, was_classic); 898 899 if (build_body(&ctx, extra_pass)) { 900 bpf_jit_binary_free(header); 901 prog = orig_prog; 902 goto out_off; 903 } 904 905 build_epilogue(&ctx); 906 907 /* 3. Extra pass to validate JITed code. */ 908 if (validate_code(&ctx)) { 909 bpf_jit_binary_free(header); 910 prog = orig_prog; 911 goto out_off; 912 } 913 914 /* And we're done. */ 915 if (bpf_jit_enable > 1) 916 bpf_jit_dump(prog->len, image_size, 2, ctx.image); 917 918 bpf_flush_icache(header, ctx.image + ctx.idx); 919 920 if (!prog->is_func || extra_pass) { 921 if (extra_pass && ctx.idx != jit_data->ctx.idx) { 922 pr_err_once("multi-func JIT bug %d != %d\n", 923 ctx.idx, jit_data->ctx.idx); 924 bpf_jit_binary_free(header); 925 prog->bpf_func = NULL; 926 prog->jited = 0; 927 goto out_off; 928 } 929 bpf_jit_binary_lock_ro(header); 930 } else { 931 jit_data->ctx = ctx; 932 jit_data->image = image_ptr; 933 jit_data->header = header; 934 } 935 prog->bpf_func = (void *)ctx.image; 936 prog->jited = 1; 937 prog->jited_len = image_size; 938 939 if (!prog->is_func || extra_pass) { 940 out_off: 941 kfree(ctx.offset); 942 kfree(jit_data); 943 prog->aux->jit_data = NULL; 944 } 945 out: 946 if (tmp_blinded) 947 bpf_jit_prog_release_other(prog, prog == orig_prog ? 948 tmp : orig_prog); 949 return prog; 950 } 951