1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2026 Meta Platforms, Inc. and affiliates. */ 3 #include <linux/bpf.h> 4 #include <linux/btf.h> 5 #include <linux/bpf_verifier.h> 6 #include <linux/filter.h> 7 #include <linux/vmalloc.h> 8 #include <linux/bsearch.h> 9 #include <linux/sort.h> 10 #include <linux/perf_event.h> 11 #include <net/xdp.h> 12 #include "disasm.h" 13 14 #define verbose(env, fmt, args...) bpf_verifier_log_write(env, fmt, ##args) 15 16 static bool is_cmpxchg_insn(const struct bpf_insn *insn) 17 { 18 return BPF_CLASS(insn->code) == BPF_STX && 19 BPF_MODE(insn->code) == BPF_ATOMIC && 20 insn->imm == BPF_CMPXCHG; 21 } 22 23 /* Return the regno defined by the insn, or -1. */ 24 static int insn_def_regno(const struct bpf_insn *insn) 25 { 26 switch (BPF_CLASS(insn->code)) { 27 case BPF_JMP: 28 case BPF_JMP32: 29 case BPF_ST: 30 return -1; 31 case BPF_STX: 32 if (BPF_MODE(insn->code) == BPF_ATOMIC || 33 BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) { 34 if (insn->imm == BPF_CMPXCHG) 35 return BPF_REG_0; 36 else if (insn->imm == BPF_LOAD_ACQ) 37 return insn->dst_reg; 38 else if (insn->imm & BPF_FETCH) 39 return insn->src_reg; 40 } 41 return -1; 42 default: 43 return insn->dst_reg; 44 } 45 } 46 47 /* Return TRUE if INSN has defined any 32-bit value explicitly. */ 48 static bool insn_has_def32(struct bpf_insn *insn) 49 { 50 int dst_reg = insn_def_regno(insn); 51 52 if (dst_reg == -1) 53 return false; 54 55 return !bpf_is_reg64(insn, dst_reg, NULL, DST_OP); 56 } 57 58 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b) 59 { 60 const struct bpf_kfunc_desc *d0 = a; 61 const struct bpf_kfunc_desc *d1 = b; 62 63 if (d0->imm != d1->imm) 64 return d0->imm < d1->imm ? -1 : 1; 65 if (d0->offset != d1->offset) 66 return d0->offset < d1->offset ? -1 : 1; 67 return 0; 68 } 69 70 const struct btf_func_model * 71 bpf_jit_find_kfunc_model(const struct bpf_prog *prog, 72 const struct bpf_insn *insn) 73 { 74 const struct bpf_kfunc_desc desc = { 75 .imm = insn->imm, 76 .offset = insn->off, 77 }; 78 const struct bpf_kfunc_desc *res; 79 struct bpf_kfunc_desc_tab *tab; 80 81 tab = prog->aux->kfunc_tab; 82 res = bsearch(&desc, tab->descs, tab->nr_descs, 83 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off); 84 85 return res ? &res->func_model : NULL; 86 } 87 88 static int set_kfunc_desc_imm(struct bpf_verifier_env *env, struct bpf_kfunc_desc *desc) 89 { 90 unsigned long call_imm; 91 92 if (bpf_jit_supports_far_kfunc_call()) { 93 call_imm = desc->func_id; 94 } else { 95 call_imm = BPF_CALL_IMM(desc->addr); 96 /* Check whether the relative offset overflows desc->imm */ 97 if ((unsigned long)(s32)call_imm != call_imm) { 98 verbose(env, "address of kernel func_id %u is out of range\n", 99 desc->func_id); 100 return -EINVAL; 101 } 102 } 103 desc->imm = call_imm; 104 return 0; 105 } 106 107 static int sort_kfunc_descs_by_imm_off(struct bpf_verifier_env *env) 108 { 109 struct bpf_kfunc_desc_tab *tab; 110 int i, err; 111 112 tab = env->prog->aux->kfunc_tab; 113 if (!tab) 114 return 0; 115 116 for (i = 0; i < tab->nr_descs; i++) { 117 err = set_kfunc_desc_imm(env, &tab->descs[i]); 118 if (err) 119 return err; 120 } 121 122 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]), 123 kfunc_desc_cmp_by_imm_off, NULL); 124 return 0; 125 } 126 127 static int add_kfunc_in_insns(struct bpf_verifier_env *env, 128 struct bpf_insn *insn, int cnt) 129 { 130 int i, ret; 131 132 for (i = 0; i < cnt; i++, insn++) { 133 if (bpf_pseudo_kfunc_call(insn)) { 134 ret = bpf_add_kfunc_call(env, insn->imm, insn->off); 135 if (ret < 0) 136 return ret; 137 } 138 } 139 return 0; 140 } 141 142 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 143 static int get_callee_stack_depth(struct bpf_verifier_env *env, 144 const struct bpf_insn *insn, int idx) 145 { 146 int start = idx + insn->imm + 1, subprog; 147 148 subprog = bpf_find_subprog(env, start); 149 if (verifier_bug_if(subprog < 0, env, "get stack depth: no program at insn %d", start)) 150 return -EFAULT; 151 return env->subprog_info[subprog].stack_depth; 152 } 153 #endif 154 155 /* single env->prog->insni[off] instruction was replaced with the range 156 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying 157 * [0, off) and [off, end) to new locations, so the patched range stays zero 158 */ 159 static void adjust_insn_aux_data(struct bpf_verifier_env *env, 160 struct bpf_prog *new_prog, u32 off, u32 cnt) 161 { 162 struct bpf_insn_aux_data *data = env->insn_aux_data; 163 struct bpf_insn *insn = new_prog->insnsi; 164 u32 old_seen = data[off].seen; 165 u32 prog_len; 166 int i; 167 168 /* aux info at OFF always needs adjustment, no matter fast path 169 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the 170 * original insn at old prog. 171 */ 172 data[off].zext_dst = insn_has_def32(insn + off + cnt - 1); 173 174 if (cnt == 1) 175 return; 176 prog_len = new_prog->len; 177 178 memmove(data + off + cnt - 1, data + off, 179 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1)); 180 memset(data + off, 0, sizeof(struct bpf_insn_aux_data) * (cnt - 1)); 181 for (i = off; i < off + cnt - 1; i++) { 182 /* Expand insni[off]'s seen count to the patched range. */ 183 data[i].seen = old_seen; 184 data[i].zext_dst = insn_has_def32(insn + i); 185 } 186 } 187 188 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len) 189 { 190 int i; 191 192 if (len == 1) 193 return; 194 /* NOTE: fake 'exit' subprog should be updated as well. */ 195 for (i = 0; i <= env->subprog_cnt; i++) { 196 if (env->subprog_info[i].start <= off) 197 continue; 198 env->subprog_info[i].start += len - 1; 199 } 200 } 201 202 static void adjust_insn_arrays(struct bpf_verifier_env *env, u32 off, u32 len) 203 { 204 int i; 205 206 if (len == 1) 207 return; 208 209 for (i = 0; i < env->insn_array_map_cnt; i++) 210 bpf_insn_array_adjust(env->insn_array_maps[i], off, len); 211 } 212 213 static void adjust_insn_arrays_after_remove(struct bpf_verifier_env *env, u32 off, u32 len) 214 { 215 int i; 216 217 for (i = 0; i < env->insn_array_map_cnt; i++) 218 bpf_insn_array_adjust_after_remove(env->insn_array_maps[i], off, len); 219 } 220 221 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len) 222 { 223 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 224 int i, sz = prog->aux->size_poke_tab; 225 struct bpf_jit_poke_descriptor *desc; 226 227 for (i = 0; i < sz; i++) { 228 desc = &tab[i]; 229 if (desc->insn_idx <= off) 230 continue; 231 desc->insn_idx += len - 1; 232 } 233 } 234 235 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off, 236 const struct bpf_insn *patch, u32 len) 237 { 238 struct bpf_prog *new_prog; 239 struct bpf_insn_aux_data *new_data = NULL; 240 241 if (len > 1) { 242 new_data = vrealloc(env->insn_aux_data, 243 array_size(env->prog->len + len - 1, 244 sizeof(struct bpf_insn_aux_data)), 245 GFP_KERNEL_ACCOUNT | __GFP_ZERO); 246 if (!new_data) 247 return NULL; 248 249 env->insn_aux_data = new_data; 250 } 251 252 new_prog = bpf_patch_insn_single(env->prog, off, patch, len); 253 if (IS_ERR(new_prog)) { 254 if (PTR_ERR(new_prog) == -ERANGE) 255 verbose(env, 256 "insn %d cannot be patched due to 16-bit range\n", 257 env->insn_aux_data[off].orig_idx); 258 return NULL; 259 } 260 adjust_insn_aux_data(env, new_prog, off, len); 261 adjust_subprog_starts(env, off, len); 262 adjust_insn_arrays(env, off, len); 263 adjust_poke_descs(new_prog, off, len); 264 return new_prog; 265 } 266 267 /* 268 * For all jmp insns in a given 'prog' that point to 'tgt_idx' insn adjust the 269 * jump offset by 'delta'. 270 */ 271 static int adjust_jmp_off(struct bpf_prog *prog, u32 tgt_idx, u32 delta) 272 { 273 struct bpf_insn *insn = prog->insnsi; 274 u32 insn_cnt = prog->len, i; 275 s32 imm; 276 s16 off; 277 278 for (i = 0; i < insn_cnt; i++, insn++) { 279 u8 code = insn->code; 280 281 if (tgt_idx <= i && i < tgt_idx + delta) 282 continue; 283 284 if ((BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32) || 285 BPF_OP(code) == BPF_CALL || BPF_OP(code) == BPF_EXIT) 286 continue; 287 288 if (insn->code == (BPF_JMP32 | BPF_JA)) { 289 if (i + 1 + insn->imm != tgt_idx) 290 continue; 291 if (check_add_overflow(insn->imm, delta, &imm)) 292 return -ERANGE; 293 insn->imm = imm; 294 } else { 295 if (i + 1 + insn->off != tgt_idx) 296 continue; 297 if (check_add_overflow(insn->off, delta, &off)) 298 return -ERANGE; 299 insn->off = off; 300 } 301 } 302 return 0; 303 } 304 305 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env, 306 u32 off, u32 cnt) 307 { 308 int i, j; 309 310 /* find first prog starting at or after off (first to remove) */ 311 for (i = 0; i < env->subprog_cnt; i++) 312 if (env->subprog_info[i].start >= off) 313 break; 314 /* find first prog starting at or after off + cnt (first to stay) */ 315 for (j = i; j < env->subprog_cnt; j++) 316 if (env->subprog_info[j].start >= off + cnt) 317 break; 318 /* if j doesn't start exactly at off + cnt, we are just removing 319 * the front of previous prog 320 */ 321 if (env->subprog_info[j].start != off + cnt) 322 j--; 323 324 if (j > i) { 325 struct bpf_prog_aux *aux = env->prog->aux; 326 int move; 327 328 /* move fake 'exit' subprog as well */ 329 move = env->subprog_cnt + 1 - j; 330 331 memmove(env->subprog_info + i, 332 env->subprog_info + j, 333 sizeof(*env->subprog_info) * move); 334 env->subprog_cnt -= j - i; 335 336 /* remove func_info */ 337 if (aux->func_info) { 338 move = aux->func_info_cnt - j; 339 340 memmove(aux->func_info + i, 341 aux->func_info + j, 342 sizeof(*aux->func_info) * move); 343 aux->func_info_cnt -= j - i; 344 /* func_info->insn_off is set after all code rewrites, 345 * in adjust_btf_func() - no need to adjust 346 */ 347 } 348 } else { 349 /* convert i from "first prog to remove" to "first to adjust" */ 350 if (env->subprog_info[i].start == off) 351 i++; 352 } 353 354 /* update fake 'exit' subprog as well */ 355 for (; i <= env->subprog_cnt; i++) 356 env->subprog_info[i].start -= cnt; 357 358 return 0; 359 } 360 361 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off, 362 u32 cnt) 363 { 364 struct bpf_prog *prog = env->prog; 365 u32 i, l_off, l_cnt, nr_linfo; 366 struct bpf_line_info *linfo; 367 368 nr_linfo = prog->aux->nr_linfo; 369 if (!nr_linfo) 370 return 0; 371 372 linfo = prog->aux->linfo; 373 374 /* find first line info to remove, count lines to be removed */ 375 for (i = 0; i < nr_linfo; i++) 376 if (linfo[i].insn_off >= off) 377 break; 378 379 l_off = i; 380 l_cnt = 0; 381 for (; i < nr_linfo; i++) 382 if (linfo[i].insn_off < off + cnt) 383 l_cnt++; 384 else 385 break; 386 387 /* First live insn doesn't match first live linfo, it needs to "inherit" 388 * last removed linfo. prog is already modified, so prog->len == off 389 * means no live instructions after (tail of the program was removed). 390 */ 391 if (prog->len != off && l_cnt && 392 (i == nr_linfo || linfo[i].insn_off != off + cnt)) { 393 l_cnt--; 394 linfo[--i].insn_off = off + cnt; 395 } 396 397 /* remove the line info which refer to the removed instructions */ 398 if (l_cnt) { 399 memmove(linfo + l_off, linfo + i, 400 sizeof(*linfo) * (nr_linfo - i)); 401 402 prog->aux->nr_linfo -= l_cnt; 403 nr_linfo = prog->aux->nr_linfo; 404 } 405 406 /* pull all linfo[i].insn_off >= off + cnt in by cnt */ 407 for (i = l_off; i < nr_linfo; i++) 408 linfo[i].insn_off -= cnt; 409 410 /* fix up all subprogs (incl. 'exit') which start >= off */ 411 for (i = 0; i <= env->subprog_cnt; i++) 412 if (env->subprog_info[i].linfo_idx > l_off) { 413 /* program may have started in the removed region but 414 * may not be fully removed 415 */ 416 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt) 417 env->subprog_info[i].linfo_idx -= l_cnt; 418 else 419 env->subprog_info[i].linfo_idx = l_off; 420 } 421 422 return 0; 423 } 424 425 /* 426 * Clean up dynamically allocated fields of aux data for instructions [start, ...] 427 */ 428 void bpf_clear_insn_aux_data(struct bpf_verifier_env *env, int start, int len) 429 { 430 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 431 struct bpf_insn *insns = env->prog->insnsi; 432 int end = start + len; 433 int i; 434 435 for (i = start; i < end; i++) { 436 if (aux_data[i].jt) { 437 kvfree(aux_data[i].jt); 438 aux_data[i].jt = NULL; 439 } 440 441 if (bpf_is_ldimm64(&insns[i])) 442 i++; 443 } 444 } 445 446 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt) 447 { 448 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 449 unsigned int orig_prog_len = env->prog->len; 450 int err; 451 452 if (bpf_prog_is_offloaded(env->prog->aux)) 453 bpf_prog_offload_remove_insns(env, off, cnt); 454 455 /* Should be called before bpf_remove_insns, as it uses prog->insnsi */ 456 bpf_clear_insn_aux_data(env, off, cnt); 457 458 err = bpf_remove_insns(env->prog, off, cnt); 459 if (err) 460 return err; 461 462 err = adjust_subprog_starts_after_remove(env, off, cnt); 463 if (err) 464 return err; 465 466 err = bpf_adj_linfo_after_remove(env, off, cnt); 467 if (err) 468 return err; 469 470 adjust_insn_arrays_after_remove(env, off, cnt); 471 472 memmove(aux_data + off, aux_data + off + cnt, 473 sizeof(*aux_data) * (orig_prog_len - off - cnt)); 474 475 return 0; 476 } 477 478 static const struct bpf_insn NOP = BPF_JMP_IMM(BPF_JA, 0, 0, 0); 479 static const struct bpf_insn MAY_GOTO_0 = BPF_RAW_INSN(BPF_JMP | BPF_JCOND, 0, 0, 0, 0); 480 481 bool bpf_insn_is_cond_jump(u8 code) 482 { 483 u8 op; 484 485 op = BPF_OP(code); 486 if (BPF_CLASS(code) == BPF_JMP32) 487 return op != BPF_JA; 488 489 if (BPF_CLASS(code) != BPF_JMP) 490 return false; 491 492 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL; 493 } 494 495 void bpf_opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env) 496 { 497 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 498 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0); 499 struct bpf_insn *insn = env->prog->insnsi; 500 const int insn_cnt = env->prog->len; 501 int i; 502 503 for (i = 0; i < insn_cnt; i++, insn++) { 504 if (!bpf_insn_is_cond_jump(insn->code)) 505 continue; 506 507 if (!aux_data[i + 1].seen) 508 ja.off = insn->off; 509 else if (!aux_data[i + 1 + insn->off].seen) 510 ja.off = 0; 511 else 512 continue; 513 514 if (bpf_prog_is_offloaded(env->prog->aux)) 515 bpf_prog_offload_replace_insn(env, i, &ja); 516 517 memcpy(insn, &ja, sizeof(ja)); 518 } 519 } 520 521 int bpf_opt_remove_dead_code(struct bpf_verifier_env *env) 522 { 523 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 524 int insn_cnt = env->prog->len; 525 int i, err; 526 527 for (i = 0; i < insn_cnt; i++) { 528 int j; 529 530 j = 0; 531 while (i + j < insn_cnt && !aux_data[i + j].seen) 532 j++; 533 if (!j) 534 continue; 535 536 err = verifier_remove_insns(env, i, j); 537 if (err) 538 return err; 539 insn_cnt = env->prog->len; 540 } 541 542 return 0; 543 } 544 545 int bpf_opt_remove_nops(struct bpf_verifier_env *env) 546 { 547 struct bpf_insn *insn = env->prog->insnsi; 548 int insn_cnt = env->prog->len; 549 bool is_may_goto_0, is_ja; 550 int i, err; 551 552 for (i = 0; i < insn_cnt; i++) { 553 is_may_goto_0 = !memcmp(&insn[i], &MAY_GOTO_0, sizeof(MAY_GOTO_0)); 554 is_ja = !memcmp(&insn[i], &NOP, sizeof(NOP)); 555 556 if (!is_may_goto_0 && !is_ja) 557 continue; 558 559 err = verifier_remove_insns(env, i, 1); 560 if (err) 561 return err; 562 insn_cnt--; 563 /* Go back one insn to catch may_goto +1; may_goto +0 sequence */ 564 i -= (is_may_goto_0 && i > 0) ? 2 : 1; 565 } 566 567 return 0; 568 } 569 570 int bpf_opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env, 571 const union bpf_attr *attr) 572 { 573 struct bpf_insn *patch; 574 /* use env->insn_buf as two independent buffers */ 575 struct bpf_insn *zext_patch = env->insn_buf; 576 struct bpf_insn *rnd_hi32_patch = &env->insn_buf[2]; 577 struct bpf_insn_aux_data *aux = env->insn_aux_data; 578 int i, patch_len, delta = 0, len = env->prog->len; 579 struct bpf_insn *insns = env->prog->insnsi; 580 struct bpf_prog *new_prog; 581 bool rnd_hi32; 582 583 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32; 584 zext_patch[1] = BPF_ZEXT_REG(0); 585 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0); 586 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 587 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX); 588 for (i = 0; i < len; i++) { 589 int adj_idx = i + delta; 590 struct bpf_insn insn; 591 int load_reg; 592 593 insn = insns[adj_idx]; 594 load_reg = insn_def_regno(&insn); 595 if (!aux[adj_idx].zext_dst) { 596 u8 code, class; 597 u32 imm_rnd; 598 599 if (!rnd_hi32) 600 continue; 601 602 code = insn.code; 603 class = BPF_CLASS(code); 604 if (load_reg == -1) 605 continue; 606 607 /* NOTE: arg "reg" (the fourth one) is only used for 608 * BPF_STX + SRC_OP, so it is safe to pass NULL 609 * here. 610 */ 611 if (bpf_is_reg64(&insn, load_reg, NULL, DST_OP)) { 612 if (class == BPF_LD && 613 BPF_MODE(code) == BPF_IMM) 614 i++; 615 continue; 616 } 617 618 /* ctx load could be transformed into wider load. */ 619 if (class == BPF_LDX && 620 aux[adj_idx].ptr_type == PTR_TO_CTX) 621 continue; 622 623 imm_rnd = get_random_u32(); 624 rnd_hi32_patch[0] = insn; 625 rnd_hi32_patch[1].imm = imm_rnd; 626 rnd_hi32_patch[3].dst_reg = load_reg; 627 patch = rnd_hi32_patch; 628 patch_len = 4; 629 goto apply_patch_buffer; 630 } 631 632 /* Add in an zero-extend instruction if a) the JIT has requested 633 * it or b) it's a CMPXCHG. 634 * 635 * The latter is because: BPF_CMPXCHG always loads a value into 636 * R0, therefore always zero-extends. However some archs' 637 * equivalent instruction only does this load when the 638 * comparison is successful. This detail of CMPXCHG is 639 * orthogonal to the general zero-extension behaviour of the 640 * CPU, so it's treated independently of bpf_jit_needs_zext. 641 */ 642 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn)) 643 continue; 644 645 /* Zero-extension is done by the caller. */ 646 if (bpf_pseudo_kfunc_call(&insn)) 647 continue; 648 649 if (verifier_bug_if(load_reg == -1, env, 650 "zext_dst is set, but no reg is defined")) 651 return -EFAULT; 652 653 zext_patch[0] = insn; 654 zext_patch[1].dst_reg = load_reg; 655 zext_patch[1].src_reg = load_reg; 656 patch = zext_patch; 657 patch_len = 2; 658 apply_patch_buffer: 659 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len); 660 if (!new_prog) 661 return -ENOMEM; 662 env->prog = new_prog; 663 insns = new_prog->insnsi; 664 aux = env->insn_aux_data; 665 delta += patch_len - 1; 666 } 667 668 return 0; 669 } 670 671 /* convert load instructions that access fields of a context type into a 672 * sequence of instructions that access fields of the underlying structure: 673 * struct __sk_buff -> struct sk_buff 674 * struct bpf_sock_ops -> struct sock 675 */ 676 int bpf_convert_ctx_accesses(struct bpf_verifier_env *env) 677 { 678 struct bpf_subprog_info *subprogs = env->subprog_info; 679 const struct bpf_verifier_ops *ops = env->ops; 680 int i, cnt, size, ctx_field_size, ret, delta = 0, epilogue_cnt = 0; 681 const int insn_cnt = env->prog->len; 682 struct bpf_insn *epilogue_buf = env->epilogue_buf; 683 struct bpf_insn *insn_buf = env->insn_buf; 684 struct bpf_insn *insn; 685 u32 target_size, size_default, off; 686 struct bpf_prog *new_prog; 687 enum bpf_access_type type; 688 bool is_narrower_load; 689 int epilogue_idx = 0; 690 691 if (ops->gen_epilogue) { 692 epilogue_cnt = ops->gen_epilogue(epilogue_buf, env->prog, 693 -(subprogs[0].stack_depth + 8)); 694 if (epilogue_cnt >= INSN_BUF_SIZE) { 695 verifier_bug(env, "epilogue is too long"); 696 return -EFAULT; 697 } else if (epilogue_cnt) { 698 /* Save the ARG_PTR_TO_CTX for the epilogue to use */ 699 cnt = 0; 700 subprogs[0].stack_depth += 8; 701 insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_1, 702 -subprogs[0].stack_depth); 703 insn_buf[cnt++] = env->prog->insnsi[0]; 704 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt); 705 if (!new_prog) 706 return -ENOMEM; 707 env->prog = new_prog; 708 delta += cnt - 1; 709 710 ret = add_kfunc_in_insns(env, epilogue_buf, epilogue_cnt - 1); 711 if (ret < 0) 712 return ret; 713 } 714 } 715 716 if (ops->gen_prologue || env->seen_direct_write) { 717 if (!ops->gen_prologue) { 718 verifier_bug(env, "gen_prologue is null"); 719 return -EFAULT; 720 } 721 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write, 722 env->prog); 723 if (cnt >= INSN_BUF_SIZE) { 724 verifier_bug(env, "prologue is too long"); 725 return -EFAULT; 726 } else if (cnt) { 727 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt); 728 if (!new_prog) 729 return -ENOMEM; 730 731 env->prog = new_prog; 732 delta += cnt - 1; 733 734 ret = add_kfunc_in_insns(env, insn_buf, cnt - 1); 735 if (ret < 0) 736 return ret; 737 } 738 } 739 740 if (delta) 741 WARN_ON(adjust_jmp_off(env->prog, 0, delta)); 742 743 if (bpf_prog_is_offloaded(env->prog->aux)) 744 return 0; 745 746 insn = env->prog->insnsi + delta; 747 748 for (i = 0; i < insn_cnt; i++, insn++) { 749 bpf_convert_ctx_access_t convert_ctx_access; 750 u8 mode; 751 752 if (env->insn_aux_data[i + delta].nospec) { 753 WARN_ON_ONCE(env->insn_aux_data[i + delta].alu_state); 754 struct bpf_insn *patch = insn_buf; 755 756 *patch++ = BPF_ST_NOSPEC(); 757 *patch++ = *insn; 758 cnt = patch - insn_buf; 759 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 760 if (!new_prog) 761 return -ENOMEM; 762 763 delta += cnt - 1; 764 env->prog = new_prog; 765 insn = new_prog->insnsi + i + delta; 766 /* This can not be easily merged with the 767 * nospec_result-case, because an insn may require a 768 * nospec before and after itself. Therefore also do not 769 * 'continue' here but potentially apply further 770 * patching to insn. *insn should equal patch[1] now. 771 */ 772 } 773 774 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) || 775 insn->code == (BPF_LDX | BPF_MEM | BPF_H) || 776 insn->code == (BPF_LDX | BPF_MEM | BPF_W) || 777 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) || 778 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) || 779 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) || 780 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) { 781 type = BPF_READ; 782 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) || 783 insn->code == (BPF_STX | BPF_MEM | BPF_H) || 784 insn->code == (BPF_STX | BPF_MEM | BPF_W) || 785 insn->code == (BPF_STX | BPF_MEM | BPF_DW) || 786 insn->code == (BPF_ST | BPF_MEM | BPF_B) || 787 insn->code == (BPF_ST | BPF_MEM | BPF_H) || 788 insn->code == (BPF_ST | BPF_MEM | BPF_W) || 789 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) { 790 type = BPF_WRITE; 791 } else if ((insn->code == (BPF_STX | BPF_ATOMIC | BPF_B) || 792 insn->code == (BPF_STX | BPF_ATOMIC | BPF_H) || 793 insn->code == (BPF_STX | BPF_ATOMIC | BPF_W) || 794 insn->code == (BPF_STX | BPF_ATOMIC | BPF_DW)) && 795 env->insn_aux_data[i + delta].ptr_type == PTR_TO_ARENA) { 796 insn->code = BPF_STX | BPF_PROBE_ATOMIC | BPF_SIZE(insn->code); 797 env->prog->aux->num_exentries++; 798 continue; 799 } else if (insn->code == (BPF_JMP | BPF_EXIT) && 800 epilogue_cnt && 801 i + delta < subprogs[1].start) { 802 /* Generate epilogue for the main prog */ 803 if (epilogue_idx) { 804 /* jump back to the earlier generated epilogue */ 805 insn_buf[0] = BPF_JMP32_A(epilogue_idx - i - delta - 1); 806 cnt = 1; 807 } else { 808 memcpy(insn_buf, epilogue_buf, 809 epilogue_cnt * sizeof(*epilogue_buf)); 810 cnt = epilogue_cnt; 811 /* epilogue_idx cannot be 0. It must have at 812 * least one ctx ptr saving insn before the 813 * epilogue. 814 */ 815 epilogue_idx = i + delta; 816 } 817 goto patch_insn_buf; 818 } else { 819 continue; 820 } 821 822 if (type == BPF_WRITE && 823 env->insn_aux_data[i + delta].nospec_result) { 824 /* nospec_result is only used to mitigate Spectre v4 and 825 * to limit verification-time for Spectre v1. 826 */ 827 struct bpf_insn *patch = insn_buf; 828 829 *patch++ = *insn; 830 *patch++ = BPF_ST_NOSPEC(); 831 cnt = patch - insn_buf; 832 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 833 if (!new_prog) 834 return -ENOMEM; 835 836 delta += cnt - 1; 837 env->prog = new_prog; 838 insn = new_prog->insnsi + i + delta; 839 continue; 840 } 841 842 switch ((int)env->insn_aux_data[i + delta].ptr_type) { 843 case PTR_TO_CTX: 844 if (!ops->convert_ctx_access) 845 continue; 846 convert_ctx_access = ops->convert_ctx_access; 847 break; 848 case PTR_TO_SOCKET: 849 case PTR_TO_SOCK_COMMON: 850 convert_ctx_access = bpf_sock_convert_ctx_access; 851 break; 852 case PTR_TO_TCP_SOCK: 853 convert_ctx_access = bpf_tcp_sock_convert_ctx_access; 854 break; 855 case PTR_TO_XDP_SOCK: 856 convert_ctx_access = bpf_xdp_sock_convert_ctx_access; 857 break; 858 case PTR_TO_BTF_ID: 859 case PTR_TO_BTF_ID | PTR_UNTRUSTED: 860 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike 861 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot 862 * be said once it is marked PTR_UNTRUSTED, hence we must handle 863 * any faults for loads into such types. BPF_WRITE is disallowed 864 * for this case. 865 */ 866 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED: 867 case PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED: 868 if (type == BPF_READ) { 869 if (BPF_MODE(insn->code) == BPF_MEM) 870 insn->code = BPF_LDX | BPF_PROBE_MEM | 871 BPF_SIZE((insn)->code); 872 else 873 insn->code = BPF_LDX | BPF_PROBE_MEMSX | 874 BPF_SIZE((insn)->code); 875 env->prog->aux->num_exentries++; 876 } 877 continue; 878 case PTR_TO_ARENA: 879 if (BPF_MODE(insn->code) == BPF_MEMSX) { 880 if (!bpf_jit_supports_insn(insn, true)) { 881 verbose(env, "sign extending loads from arena are not supported yet\n"); 882 return -EOPNOTSUPP; 883 } 884 insn->code = BPF_CLASS(insn->code) | BPF_PROBE_MEM32SX | BPF_SIZE(insn->code); 885 } else { 886 insn->code = BPF_CLASS(insn->code) | BPF_PROBE_MEM32 | BPF_SIZE(insn->code); 887 } 888 env->prog->aux->num_exentries++; 889 continue; 890 default: 891 continue; 892 } 893 894 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size; 895 size = BPF_LDST_BYTES(insn); 896 mode = BPF_MODE(insn->code); 897 898 /* If the read access is a narrower load of the field, 899 * convert to a 4/8-byte load, to minimum program type specific 900 * convert_ctx_access changes. If conversion is successful, 901 * we will apply proper mask to the result. 902 */ 903 is_narrower_load = size < ctx_field_size; 904 size_default = bpf_ctx_off_adjust_machine(ctx_field_size); 905 off = insn->off; 906 if (is_narrower_load) { 907 u8 size_code; 908 909 if (type == BPF_WRITE) { 910 verifier_bug(env, "narrow ctx access misconfigured"); 911 return -EFAULT; 912 } 913 914 size_code = BPF_H; 915 if (ctx_field_size == 4) 916 size_code = BPF_W; 917 else if (ctx_field_size == 8) 918 size_code = BPF_DW; 919 920 insn->off = off & ~(size_default - 1); 921 insn->code = BPF_LDX | BPF_MEM | size_code; 922 } 923 924 target_size = 0; 925 cnt = convert_ctx_access(type, insn, insn_buf, env->prog, 926 &target_size); 927 if (cnt == 0 || cnt >= INSN_BUF_SIZE || 928 (ctx_field_size && !target_size)) { 929 verifier_bug(env, "error during ctx access conversion (%d)", cnt); 930 return -EFAULT; 931 } 932 933 if (is_narrower_load && size < target_size) { 934 u8 shift = bpf_ctx_narrow_access_offset( 935 off, size, size_default) * 8; 936 if (shift && cnt + 1 >= INSN_BUF_SIZE) { 937 verifier_bug(env, "narrow ctx load misconfigured"); 938 return -EFAULT; 939 } 940 if (ctx_field_size <= 4) { 941 if (shift) 942 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH, 943 insn->dst_reg, 944 shift); 945 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg, 946 (1 << size * 8) - 1); 947 } else { 948 if (shift) 949 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH, 950 insn->dst_reg, 951 shift); 952 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg, 953 (1ULL << size * 8) - 1); 954 } 955 } 956 if (mode == BPF_MEMSX) 957 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X, 958 insn->dst_reg, insn->dst_reg, 959 size * 8, 0); 960 961 patch_insn_buf: 962 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 963 if (!new_prog) 964 return -ENOMEM; 965 966 delta += cnt - 1; 967 968 /* keep walking new program and skip insns we just inserted */ 969 env->prog = new_prog; 970 insn = new_prog->insnsi + i + delta; 971 } 972 973 return 0; 974 } 975 976 int bpf_jit_subprogs(struct bpf_verifier_env *env) 977 { 978 struct bpf_prog *prog = env->prog, **func, *tmp; 979 int i, j, subprog_start, subprog_end = 0, len, subprog; 980 struct bpf_map *map_ptr; 981 struct bpf_insn *insn; 982 void *old_bpf_func; 983 int err, num_exentries; 984 int old_len, subprog_start_adjustment = 0; 985 986 if (env->subprog_cnt <= 1) 987 return 0; 988 989 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { 990 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn)) 991 continue; 992 993 /* Upon error here we cannot fall back to interpreter but 994 * need a hard reject of the program. Thus -EFAULT is 995 * propagated in any case. 996 */ 997 subprog = bpf_find_subprog(env, i + insn->imm + 1); 998 if (verifier_bug_if(subprog < 0, env, "No program to jit at insn %d", 999 i + insn->imm + 1)) 1000 return -EFAULT; 1001 /* temporarily remember subprog id inside insn instead of 1002 * aux_data, since next loop will split up all insns into funcs 1003 */ 1004 insn->off = subprog; 1005 /* remember original imm in case JIT fails and fallback 1006 * to interpreter will be needed 1007 */ 1008 env->insn_aux_data[i].call_imm = insn->imm; 1009 /* point imm to __bpf_call_base+1 from JITs point of view */ 1010 insn->imm = 1; 1011 if (bpf_pseudo_func(insn)) { 1012 #if defined(MODULES_VADDR) 1013 u64 addr = MODULES_VADDR; 1014 #else 1015 u64 addr = VMALLOC_START; 1016 #endif 1017 /* jit (e.g. x86_64) may emit fewer instructions 1018 * if it learns a u32 imm is the same as a u64 imm. 1019 * Set close enough to possible prog address. 1020 */ 1021 insn[0].imm = (u32)addr; 1022 insn[1].imm = addr >> 32; 1023 } 1024 } 1025 1026 err = bpf_prog_alloc_jited_linfo(prog); 1027 if (err) 1028 goto out_undo_insn; 1029 1030 err = -ENOMEM; 1031 func = kzalloc_objs(prog, env->subprog_cnt); 1032 if (!func) 1033 goto out_undo_insn; 1034 1035 for (i = 0; i < env->subprog_cnt; i++) { 1036 subprog_start = subprog_end; 1037 subprog_end = env->subprog_info[i + 1].start; 1038 1039 len = subprog_end - subprog_start; 1040 /* bpf_prog_run() doesn't call subprogs directly, 1041 * hence main prog stats include the runtime of subprogs. 1042 * subprogs don't have IDs and not reachable via prog_get_next_id 1043 * func[i]->stats will never be accessed and stays NULL 1044 */ 1045 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER); 1046 if (!func[i]) 1047 goto out_free; 1048 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start], 1049 len * sizeof(struct bpf_insn)); 1050 func[i]->type = prog->type; 1051 func[i]->len = len; 1052 if (bpf_prog_calc_tag(func[i])) 1053 goto out_free; 1054 func[i]->is_func = 1; 1055 func[i]->sleepable = prog->sleepable; 1056 func[i]->aux->func_idx = i; 1057 /* Below members will be freed only at prog->aux */ 1058 func[i]->aux->btf = prog->aux->btf; 1059 func[i]->aux->subprog_start = subprog_start + subprog_start_adjustment; 1060 func[i]->aux->func_info = prog->aux->func_info; 1061 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt; 1062 func[i]->aux->poke_tab = prog->aux->poke_tab; 1063 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab; 1064 func[i]->aux->main_prog_aux = prog->aux; 1065 1066 for (j = 0; j < prog->aux->size_poke_tab; j++) { 1067 struct bpf_jit_poke_descriptor *poke; 1068 1069 poke = &prog->aux->poke_tab[j]; 1070 if (poke->insn_idx < subprog_end && 1071 poke->insn_idx >= subprog_start) 1072 poke->aux = func[i]->aux; 1073 } 1074 1075 func[i]->aux->name[0] = 'F'; 1076 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth; 1077 if (env->subprog_info[i].priv_stack_mode == PRIV_STACK_ADAPTIVE) 1078 func[i]->aux->jits_use_priv_stack = true; 1079 1080 func[i]->jit_requested = 1; 1081 func[i]->blinding_requested = prog->blinding_requested; 1082 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab; 1083 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab; 1084 func[i]->aux->linfo = prog->aux->linfo; 1085 func[i]->aux->nr_linfo = prog->aux->nr_linfo; 1086 func[i]->aux->jited_linfo = prog->aux->jited_linfo; 1087 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx; 1088 func[i]->aux->arena = prog->aux->arena; 1089 func[i]->aux->used_maps = env->used_maps; 1090 func[i]->aux->used_map_cnt = env->used_map_cnt; 1091 num_exentries = 0; 1092 insn = func[i]->insnsi; 1093 for (j = 0; j < func[i]->len; j++, insn++) { 1094 if (BPF_CLASS(insn->code) == BPF_LDX && 1095 (BPF_MODE(insn->code) == BPF_PROBE_MEM || 1096 BPF_MODE(insn->code) == BPF_PROBE_MEM32 || 1097 BPF_MODE(insn->code) == BPF_PROBE_MEM32SX || 1098 BPF_MODE(insn->code) == BPF_PROBE_MEMSX)) 1099 num_exentries++; 1100 if ((BPF_CLASS(insn->code) == BPF_STX || 1101 BPF_CLASS(insn->code) == BPF_ST) && 1102 BPF_MODE(insn->code) == BPF_PROBE_MEM32) 1103 num_exentries++; 1104 if (BPF_CLASS(insn->code) == BPF_STX && 1105 BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) 1106 num_exentries++; 1107 } 1108 func[i]->aux->num_exentries = num_exentries; 1109 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable; 1110 func[i]->aux->exception_cb = env->subprog_info[i].is_exception_cb; 1111 func[i]->aux->changes_pkt_data = env->subprog_info[i].changes_pkt_data; 1112 func[i]->aux->might_sleep = env->subprog_info[i].might_sleep; 1113 if (!i) 1114 func[i]->aux->exception_boundary = env->seen_exception; 1115 1116 /* 1117 * To properly pass the absolute subprog start to jit 1118 * all instruction adjustments should be accumulated 1119 */ 1120 old_len = func[i]->len; 1121 func[i] = bpf_int_jit_compile(func[i]); 1122 subprog_start_adjustment += func[i]->len - old_len; 1123 1124 if (!func[i]->jited) { 1125 err = -ENOTSUPP; 1126 goto out_free; 1127 } 1128 cond_resched(); 1129 } 1130 1131 /* at this point all bpf functions were successfully JITed 1132 * now populate all bpf_calls with correct addresses and 1133 * run last pass of JIT 1134 */ 1135 for (i = 0; i < env->subprog_cnt; i++) { 1136 insn = func[i]->insnsi; 1137 for (j = 0; j < func[i]->len; j++, insn++) { 1138 if (bpf_pseudo_func(insn)) { 1139 subprog = insn->off; 1140 insn[0].imm = (u32)(long)func[subprog]->bpf_func; 1141 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32; 1142 continue; 1143 } 1144 if (!bpf_pseudo_call(insn)) 1145 continue; 1146 subprog = insn->off; 1147 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func); 1148 } 1149 1150 /* we use the aux data to keep a list of the start addresses 1151 * of the JITed images for each function in the program 1152 * 1153 * for some architectures, such as powerpc64, the imm field 1154 * might not be large enough to hold the offset of the start 1155 * address of the callee's JITed image from __bpf_call_base 1156 * 1157 * in such cases, we can lookup the start address of a callee 1158 * by using its subprog id, available from the off field of 1159 * the call instruction, as an index for this list 1160 */ 1161 func[i]->aux->func = func; 1162 func[i]->aux->func_cnt = env->subprog_cnt - env->hidden_subprog_cnt; 1163 func[i]->aux->real_func_cnt = env->subprog_cnt; 1164 } 1165 for (i = 0; i < env->subprog_cnt; i++) { 1166 old_bpf_func = func[i]->bpf_func; 1167 tmp = bpf_int_jit_compile(func[i]); 1168 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) { 1169 verbose(env, "JIT doesn't support bpf-to-bpf calls\n"); 1170 err = -ENOTSUPP; 1171 goto out_free; 1172 } 1173 cond_resched(); 1174 } 1175 1176 /* 1177 * Cleanup func[i]->aux fields which aren't required 1178 * or can become invalid in future 1179 */ 1180 for (i = 0; i < env->subprog_cnt; i++) { 1181 func[i]->aux->used_maps = NULL; 1182 func[i]->aux->used_map_cnt = 0; 1183 } 1184 1185 /* finally lock prog and jit images for all functions and 1186 * populate kallsysm. Begin at the first subprogram, since 1187 * bpf_prog_load will add the kallsyms for the main program. 1188 */ 1189 for (i = 1; i < env->subprog_cnt; i++) { 1190 err = bpf_prog_lock_ro(func[i]); 1191 if (err) 1192 goto out_free; 1193 } 1194 1195 for (i = 1; i < env->subprog_cnt; i++) 1196 bpf_prog_kallsyms_add(func[i]); 1197 1198 /* Last step: make now unused interpreter insns from main 1199 * prog consistent for later dump requests, so they can 1200 * later look the same as if they were interpreted only. 1201 */ 1202 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { 1203 if (bpf_pseudo_func(insn)) { 1204 insn[0].imm = env->insn_aux_data[i].call_imm; 1205 insn[1].imm = insn->off; 1206 insn->off = 0; 1207 continue; 1208 } 1209 if (!bpf_pseudo_call(insn)) 1210 continue; 1211 insn->off = env->insn_aux_data[i].call_imm; 1212 subprog = bpf_find_subprog(env, i + insn->off + 1); 1213 insn->imm = subprog; 1214 } 1215 1216 prog->jited = 1; 1217 prog->bpf_func = func[0]->bpf_func; 1218 prog->jited_len = func[0]->jited_len; 1219 prog->aux->extable = func[0]->aux->extable; 1220 prog->aux->num_exentries = func[0]->aux->num_exentries; 1221 prog->aux->func = func; 1222 prog->aux->func_cnt = env->subprog_cnt - env->hidden_subprog_cnt; 1223 prog->aux->real_func_cnt = env->subprog_cnt; 1224 prog->aux->bpf_exception_cb = (void *)func[env->exception_callback_subprog]->bpf_func; 1225 prog->aux->exception_boundary = func[0]->aux->exception_boundary; 1226 bpf_prog_jit_attempt_done(prog); 1227 return 0; 1228 out_free: 1229 /* We failed JIT'ing, so at this point we need to unregister poke 1230 * descriptors from subprogs, so that kernel is not attempting to 1231 * patch it anymore as we're freeing the subprog JIT memory. 1232 */ 1233 for (i = 0; i < prog->aux->size_poke_tab; i++) { 1234 map_ptr = prog->aux->poke_tab[i].tail_call.map; 1235 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux); 1236 } 1237 /* At this point we're guaranteed that poke descriptors are not 1238 * live anymore. We can just unlink its descriptor table as it's 1239 * released with the main prog. 1240 */ 1241 for (i = 0; i < env->subprog_cnt; i++) { 1242 if (!func[i]) 1243 continue; 1244 func[i]->aux->poke_tab = NULL; 1245 bpf_jit_free(func[i]); 1246 } 1247 kfree(func); 1248 out_undo_insn: 1249 /* cleanup main prog to be interpreted */ 1250 prog->jit_requested = 0; 1251 prog->blinding_requested = 0; 1252 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { 1253 if (!bpf_pseudo_call(insn)) 1254 continue; 1255 insn->off = 0; 1256 insn->imm = env->insn_aux_data[i].call_imm; 1257 } 1258 bpf_prog_jit_attempt_done(prog); 1259 return err; 1260 } 1261 1262 int bpf_fixup_call_args(struct bpf_verifier_env *env) 1263 { 1264 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1265 struct bpf_prog *prog = env->prog; 1266 struct bpf_insn *insn = prog->insnsi; 1267 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog); 1268 int i, depth; 1269 #endif 1270 int err = 0; 1271 1272 if (env->prog->jit_requested && 1273 !bpf_prog_is_offloaded(env->prog->aux)) { 1274 err = bpf_jit_subprogs(env); 1275 if (err == 0) 1276 return 0; 1277 if (err == -EFAULT) 1278 return err; 1279 } 1280 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1281 if (has_kfunc_call) { 1282 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n"); 1283 return -EINVAL; 1284 } 1285 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) { 1286 /* When JIT fails the progs with bpf2bpf calls and tail_calls 1287 * have to be rejected, since interpreter doesn't support them yet. 1288 */ 1289 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n"); 1290 return -EINVAL; 1291 } 1292 for (i = 0; i < prog->len; i++, insn++) { 1293 if (bpf_pseudo_func(insn)) { 1294 /* When JIT fails the progs with callback calls 1295 * have to be rejected, since interpreter doesn't support them yet. 1296 */ 1297 verbose(env, "callbacks are not allowed in non-JITed programs\n"); 1298 return -EINVAL; 1299 } 1300 1301 if (!bpf_pseudo_call(insn)) 1302 continue; 1303 depth = get_callee_stack_depth(env, insn, i); 1304 if (depth < 0) 1305 return depth; 1306 bpf_patch_call_args(insn, depth); 1307 } 1308 err = 0; 1309 #endif 1310 return err; 1311 } 1312 1313 1314 /* The function requires that first instruction in 'patch' is insnsi[prog->len - 1] */ 1315 static int add_hidden_subprog(struct bpf_verifier_env *env, struct bpf_insn *patch, int len) 1316 { 1317 struct bpf_subprog_info *info = env->subprog_info; 1318 int cnt = env->subprog_cnt; 1319 struct bpf_prog *prog; 1320 1321 /* We only reserve one slot for hidden subprogs in subprog_info. */ 1322 if (env->hidden_subprog_cnt) { 1323 verifier_bug(env, "only one hidden subprog supported"); 1324 return -EFAULT; 1325 } 1326 /* We're not patching any existing instruction, just appending the new 1327 * ones for the hidden subprog. Hence all of the adjustment operations 1328 * in bpf_patch_insn_data are no-ops. 1329 */ 1330 prog = bpf_patch_insn_data(env, env->prog->len - 1, patch, len); 1331 if (!prog) 1332 return -ENOMEM; 1333 env->prog = prog; 1334 info[cnt + 1].start = info[cnt].start; 1335 info[cnt].start = prog->len - len + 1; 1336 env->subprog_cnt++; 1337 env->hidden_subprog_cnt++; 1338 return 0; 1339 } 1340 1341 /* Do various post-verification rewrites in a single program pass. 1342 * These rewrites simplify JIT and interpreter implementations. 1343 */ 1344 int bpf_do_misc_fixups(struct bpf_verifier_env *env) 1345 { 1346 struct bpf_prog *prog = env->prog; 1347 enum bpf_attach_type eatype = prog->expected_attach_type; 1348 enum bpf_prog_type prog_type = resolve_prog_type(prog); 1349 struct bpf_insn *insn = prog->insnsi; 1350 const struct bpf_func_proto *fn; 1351 const int insn_cnt = prog->len; 1352 const struct bpf_map_ops *ops; 1353 struct bpf_insn_aux_data *aux; 1354 struct bpf_insn *insn_buf = env->insn_buf; 1355 struct bpf_prog *new_prog; 1356 struct bpf_map *map_ptr; 1357 int i, ret, cnt, delta = 0, cur_subprog = 0; 1358 struct bpf_subprog_info *subprogs = env->subprog_info; 1359 u16 stack_depth = subprogs[cur_subprog].stack_depth; 1360 u16 stack_depth_extra = 0; 1361 1362 if (env->seen_exception && !env->exception_callback_subprog) { 1363 struct bpf_insn *patch = insn_buf; 1364 1365 *patch++ = env->prog->insnsi[insn_cnt - 1]; 1366 *patch++ = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1); 1367 *patch++ = BPF_EXIT_INSN(); 1368 ret = add_hidden_subprog(env, insn_buf, patch - insn_buf); 1369 if (ret < 0) 1370 return ret; 1371 prog = env->prog; 1372 insn = prog->insnsi; 1373 1374 env->exception_callback_subprog = env->subprog_cnt - 1; 1375 /* Don't update insn_cnt, as add_hidden_subprog always appends insns */ 1376 bpf_mark_subprog_exc_cb(env, env->exception_callback_subprog); 1377 } 1378 1379 for (i = 0; i < insn_cnt;) { 1380 if (insn->code == (BPF_ALU64 | BPF_MOV | BPF_X) && insn->imm) { 1381 if ((insn->off == BPF_ADDR_SPACE_CAST && insn->imm == 1) || 1382 (((struct bpf_map *)env->prog->aux->arena)->map_flags & BPF_F_NO_USER_CONV)) { 1383 /* convert to 32-bit mov that clears upper 32-bit */ 1384 insn->code = BPF_ALU | BPF_MOV | BPF_X; 1385 /* clear off and imm, so it's a normal 'wX = wY' from JIT pov */ 1386 insn->off = 0; 1387 insn->imm = 0; 1388 } /* cast from as(0) to as(1) should be handled by JIT */ 1389 goto next_insn; 1390 } 1391 1392 if (env->insn_aux_data[i + delta].needs_zext) 1393 /* Convert BPF_CLASS(insn->code) == BPF_ALU64 to 32-bit ALU */ 1394 insn->code = BPF_ALU | BPF_OP(insn->code) | BPF_SRC(insn->code); 1395 1396 /* Make sdiv/smod divide-by-minus-one exceptions impossible. */ 1397 if ((insn->code == (BPF_ALU64 | BPF_MOD | BPF_K) || 1398 insn->code == (BPF_ALU64 | BPF_DIV | BPF_K) || 1399 insn->code == (BPF_ALU | BPF_MOD | BPF_K) || 1400 insn->code == (BPF_ALU | BPF_DIV | BPF_K)) && 1401 insn->off == 1 && insn->imm == -1) { 1402 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; 1403 bool isdiv = BPF_OP(insn->code) == BPF_DIV; 1404 struct bpf_insn *patch = insn_buf; 1405 1406 if (isdiv) 1407 *patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) | 1408 BPF_NEG | BPF_K, insn->dst_reg, 1409 0, 0, 0); 1410 else 1411 *patch++ = BPF_MOV32_IMM(insn->dst_reg, 0); 1412 1413 cnt = patch - insn_buf; 1414 1415 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1416 if (!new_prog) 1417 return -ENOMEM; 1418 1419 delta += cnt - 1; 1420 env->prog = prog = new_prog; 1421 insn = new_prog->insnsi + i + delta; 1422 goto next_insn; 1423 } 1424 1425 /* Make divide-by-zero and divide-by-minus-one exceptions impossible. */ 1426 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) || 1427 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) || 1428 insn->code == (BPF_ALU | BPF_MOD | BPF_X) || 1429 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) { 1430 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; 1431 bool isdiv = BPF_OP(insn->code) == BPF_DIV; 1432 bool is_sdiv = isdiv && insn->off == 1; 1433 bool is_smod = !isdiv && insn->off == 1; 1434 struct bpf_insn *patch = insn_buf; 1435 1436 if (is_sdiv) { 1437 /* [R,W]x sdiv 0 -> 0 1438 * LLONG_MIN sdiv -1 -> LLONG_MIN 1439 * INT_MIN sdiv -1 -> INT_MIN 1440 */ 1441 *patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg); 1442 *patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) | 1443 BPF_ADD | BPF_K, BPF_REG_AX, 1444 0, 0, 1); 1445 *patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 1446 BPF_JGT | BPF_K, BPF_REG_AX, 1447 0, 4, 1); 1448 *patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 1449 BPF_JEQ | BPF_K, BPF_REG_AX, 1450 0, 1, 0); 1451 *patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) | 1452 BPF_MOV | BPF_K, insn->dst_reg, 1453 0, 0, 0); 1454 /* BPF_NEG(LLONG_MIN) == -LLONG_MIN == LLONG_MIN */ 1455 *patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) | 1456 BPF_NEG | BPF_K, insn->dst_reg, 1457 0, 0, 0); 1458 *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1); 1459 *patch++ = *insn; 1460 cnt = patch - insn_buf; 1461 } else if (is_smod) { 1462 /* [R,W]x mod 0 -> [R,W]x */ 1463 /* [R,W]x mod -1 -> 0 */ 1464 *patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg); 1465 *patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) | 1466 BPF_ADD | BPF_K, BPF_REG_AX, 1467 0, 0, 1); 1468 *patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 1469 BPF_JGT | BPF_K, BPF_REG_AX, 1470 0, 3, 1); 1471 *patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 1472 BPF_JEQ | BPF_K, BPF_REG_AX, 1473 0, 3 + (is64 ? 0 : 1), 1); 1474 *patch++ = BPF_MOV32_IMM(insn->dst_reg, 0); 1475 *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1); 1476 *patch++ = *insn; 1477 1478 if (!is64) { 1479 *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1); 1480 *patch++ = BPF_MOV32_REG(insn->dst_reg, insn->dst_reg); 1481 } 1482 cnt = patch - insn_buf; 1483 } else if (isdiv) { 1484 /* [R,W]x div 0 -> 0 */ 1485 *patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 1486 BPF_JNE | BPF_K, insn->src_reg, 1487 0, 2, 0); 1488 *patch++ = BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg); 1489 *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1); 1490 *patch++ = *insn; 1491 cnt = patch - insn_buf; 1492 } else { 1493 /* [R,W]x mod 0 -> [R,W]x */ 1494 *patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 1495 BPF_JEQ | BPF_K, insn->src_reg, 1496 0, 1 + (is64 ? 0 : 1), 0); 1497 *patch++ = *insn; 1498 1499 if (!is64) { 1500 *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1); 1501 *patch++ = BPF_MOV32_REG(insn->dst_reg, insn->dst_reg); 1502 } 1503 cnt = patch - insn_buf; 1504 } 1505 1506 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1507 if (!new_prog) 1508 return -ENOMEM; 1509 1510 delta += cnt - 1; 1511 env->prog = prog = new_prog; 1512 insn = new_prog->insnsi + i + delta; 1513 goto next_insn; 1514 } 1515 1516 /* Make it impossible to de-reference a userspace address */ 1517 if (BPF_CLASS(insn->code) == BPF_LDX && 1518 (BPF_MODE(insn->code) == BPF_PROBE_MEM || 1519 BPF_MODE(insn->code) == BPF_PROBE_MEMSX)) { 1520 struct bpf_insn *patch = insn_buf; 1521 u64 uaddress_limit = bpf_arch_uaddress_limit(); 1522 1523 if (!uaddress_limit) 1524 goto next_insn; 1525 1526 *patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg); 1527 if (insn->off) 1528 *patch++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_AX, insn->off); 1529 *patch++ = BPF_ALU64_IMM(BPF_RSH, BPF_REG_AX, 32); 1530 *patch++ = BPF_JMP_IMM(BPF_JLE, BPF_REG_AX, uaddress_limit >> 32, 2); 1531 *patch++ = *insn; 1532 *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1); 1533 *patch++ = BPF_MOV64_IMM(insn->dst_reg, 0); 1534 1535 cnt = patch - insn_buf; 1536 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1537 if (!new_prog) 1538 return -ENOMEM; 1539 1540 delta += cnt - 1; 1541 env->prog = prog = new_prog; 1542 insn = new_prog->insnsi + i + delta; 1543 goto next_insn; 1544 } 1545 1546 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */ 1547 if (BPF_CLASS(insn->code) == BPF_LD && 1548 (BPF_MODE(insn->code) == BPF_ABS || 1549 BPF_MODE(insn->code) == BPF_IND)) { 1550 cnt = env->ops->gen_ld_abs(insn, insn_buf); 1551 if (cnt == 0 || cnt >= INSN_BUF_SIZE) { 1552 verifier_bug(env, "%d insns generated for ld_abs", cnt); 1553 return -EFAULT; 1554 } 1555 1556 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1557 if (!new_prog) 1558 return -ENOMEM; 1559 1560 delta += cnt - 1; 1561 env->prog = prog = new_prog; 1562 insn = new_prog->insnsi + i + delta; 1563 goto next_insn; 1564 } 1565 1566 /* Rewrite pointer arithmetic to mitigate speculation attacks. */ 1567 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) || 1568 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) { 1569 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X; 1570 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X; 1571 struct bpf_insn *patch = insn_buf; 1572 bool issrc, isneg, isimm; 1573 u32 off_reg; 1574 1575 aux = &env->insn_aux_data[i + delta]; 1576 if (!aux->alu_state || 1577 aux->alu_state == BPF_ALU_NON_POINTER) 1578 goto next_insn; 1579 1580 isneg = aux->alu_state & BPF_ALU_NEG_VALUE; 1581 issrc = (aux->alu_state & BPF_ALU_SANITIZE) == 1582 BPF_ALU_SANITIZE_SRC; 1583 isimm = aux->alu_state & BPF_ALU_IMMEDIATE; 1584 1585 off_reg = issrc ? insn->src_reg : insn->dst_reg; 1586 if (isimm) { 1587 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit); 1588 } else { 1589 if (isneg) 1590 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1); 1591 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit); 1592 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg); 1593 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg); 1594 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0); 1595 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63); 1596 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg); 1597 } 1598 if (!issrc) 1599 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg); 1600 insn->src_reg = BPF_REG_AX; 1601 if (isneg) 1602 insn->code = insn->code == code_add ? 1603 code_sub : code_add; 1604 *patch++ = *insn; 1605 if (issrc && isneg && !isimm) 1606 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1); 1607 cnt = patch - insn_buf; 1608 1609 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1610 if (!new_prog) 1611 return -ENOMEM; 1612 1613 delta += cnt - 1; 1614 env->prog = prog = new_prog; 1615 insn = new_prog->insnsi + i + delta; 1616 goto next_insn; 1617 } 1618 1619 if (bpf_is_may_goto_insn(insn) && bpf_jit_supports_timed_may_goto()) { 1620 int stack_off_cnt = -stack_depth - 16; 1621 1622 /* 1623 * Two 8 byte slots, depth-16 stores the count, and 1624 * depth-8 stores the start timestamp of the loop. 1625 * 1626 * The starting value of count is BPF_MAX_TIMED_LOOPS 1627 * (0xffff). Every iteration loads it and subs it by 1, 1628 * until the value becomes 0 in AX (thus, 1 in stack), 1629 * after which we call arch_bpf_timed_may_goto, which 1630 * either sets AX to 0xffff to keep looping, or to 0 1631 * upon timeout. AX is then stored into the stack. In 1632 * the next iteration, we either see 0 and break out, or 1633 * continue iterating until the next time value is 0 1634 * after subtraction, rinse and repeat. 1635 */ 1636 stack_depth_extra = 16; 1637 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_AX, BPF_REG_10, stack_off_cnt); 1638 if (insn->off >= 0) 1639 insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off + 5); 1640 else 1641 insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off - 1); 1642 insn_buf[2] = BPF_ALU64_IMM(BPF_SUB, BPF_REG_AX, 1); 1643 insn_buf[3] = BPF_JMP_IMM(BPF_JNE, BPF_REG_AX, 0, 2); 1644 /* 1645 * AX is used as an argument to pass in stack_off_cnt 1646 * (to add to r10/fp), and also as the return value of 1647 * the call to arch_bpf_timed_may_goto. 1648 */ 1649 insn_buf[4] = BPF_MOV64_IMM(BPF_REG_AX, stack_off_cnt); 1650 insn_buf[5] = BPF_EMIT_CALL(arch_bpf_timed_may_goto); 1651 insn_buf[6] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_AX, stack_off_cnt); 1652 cnt = 7; 1653 1654 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1655 if (!new_prog) 1656 return -ENOMEM; 1657 1658 delta += cnt - 1; 1659 env->prog = prog = new_prog; 1660 insn = new_prog->insnsi + i + delta; 1661 goto next_insn; 1662 } else if (bpf_is_may_goto_insn(insn)) { 1663 int stack_off = -stack_depth - 8; 1664 1665 stack_depth_extra = 8; 1666 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_AX, BPF_REG_10, stack_off); 1667 if (insn->off >= 0) 1668 insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off + 2); 1669 else 1670 insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off - 1); 1671 insn_buf[2] = BPF_ALU64_IMM(BPF_SUB, BPF_REG_AX, 1); 1672 insn_buf[3] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_AX, stack_off); 1673 cnt = 4; 1674 1675 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1676 if (!new_prog) 1677 return -ENOMEM; 1678 1679 delta += cnt - 1; 1680 env->prog = prog = new_prog; 1681 insn = new_prog->insnsi + i + delta; 1682 goto next_insn; 1683 } 1684 1685 if (insn->code != (BPF_JMP | BPF_CALL)) 1686 goto next_insn; 1687 if (insn->src_reg == BPF_PSEUDO_CALL) 1688 goto next_insn; 1689 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) { 1690 ret = bpf_fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt); 1691 if (ret) 1692 return ret; 1693 if (cnt == 0) 1694 goto next_insn; 1695 1696 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1697 if (!new_prog) 1698 return -ENOMEM; 1699 1700 delta += cnt - 1; 1701 env->prog = prog = new_prog; 1702 insn = new_prog->insnsi + i + delta; 1703 goto next_insn; 1704 } 1705 1706 /* Skip inlining the helper call if the JIT does it. */ 1707 if (bpf_jit_inlines_helper_call(insn->imm)) 1708 goto next_insn; 1709 1710 if (insn->imm == BPF_FUNC_get_route_realm) 1711 prog->dst_needed = 1; 1712 if (insn->imm == BPF_FUNC_get_prandom_u32) 1713 bpf_user_rnd_init_once(); 1714 if (insn->imm == BPF_FUNC_override_return) 1715 prog->kprobe_override = 1; 1716 if (insn->imm == BPF_FUNC_tail_call) { 1717 /* If we tail call into other programs, we 1718 * cannot make any assumptions since they can 1719 * be replaced dynamically during runtime in 1720 * the program array. 1721 */ 1722 prog->cb_access = 1; 1723 if (!bpf_allow_tail_call_in_subprogs(env)) 1724 prog->aux->stack_depth = MAX_BPF_STACK; 1725 prog->aux->max_pkt_offset = MAX_PACKET_OFF; 1726 1727 /* mark bpf_tail_call as different opcode to avoid 1728 * conditional branch in the interpreter for every normal 1729 * call and to prevent accidental JITing by JIT compiler 1730 * that doesn't support bpf_tail_call yet 1731 */ 1732 insn->imm = 0; 1733 insn->code = BPF_JMP | BPF_TAIL_CALL; 1734 1735 aux = &env->insn_aux_data[i + delta]; 1736 if (env->bpf_capable && !prog->blinding_requested && 1737 prog->jit_requested && 1738 !bpf_map_key_poisoned(aux) && 1739 !bpf_map_ptr_poisoned(aux) && 1740 !bpf_map_ptr_unpriv(aux)) { 1741 struct bpf_jit_poke_descriptor desc = { 1742 .reason = BPF_POKE_REASON_TAIL_CALL, 1743 .tail_call.map = aux->map_ptr_state.map_ptr, 1744 .tail_call.key = bpf_map_key_immediate(aux), 1745 .insn_idx = i + delta, 1746 }; 1747 1748 ret = bpf_jit_add_poke_descriptor(prog, &desc); 1749 if (ret < 0) { 1750 verbose(env, "adding tail call poke descriptor failed\n"); 1751 return ret; 1752 } 1753 1754 insn->imm = ret + 1; 1755 goto next_insn; 1756 } 1757 1758 if (!bpf_map_ptr_unpriv(aux)) 1759 goto next_insn; 1760 1761 /* instead of changing every JIT dealing with tail_call 1762 * emit two extra insns: 1763 * if (index >= max_entries) goto out; 1764 * index &= array->index_mask; 1765 * to avoid out-of-bounds cpu speculation 1766 */ 1767 if (bpf_map_ptr_poisoned(aux)) { 1768 verbose(env, "tail_call abusing map_ptr\n"); 1769 return -EINVAL; 1770 } 1771 1772 map_ptr = aux->map_ptr_state.map_ptr; 1773 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3, 1774 map_ptr->max_entries, 2); 1775 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3, 1776 container_of(map_ptr, 1777 struct bpf_array, 1778 map)->index_mask); 1779 insn_buf[2] = *insn; 1780 cnt = 3; 1781 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1782 if (!new_prog) 1783 return -ENOMEM; 1784 1785 delta += cnt - 1; 1786 env->prog = prog = new_prog; 1787 insn = new_prog->insnsi + i + delta; 1788 goto next_insn; 1789 } 1790 1791 if (insn->imm == BPF_FUNC_timer_set_callback) { 1792 /* The verifier will process callback_fn as many times as necessary 1793 * with different maps and the register states prepared by 1794 * set_timer_callback_state will be accurate. 1795 * 1796 * The following use case is valid: 1797 * map1 is shared by prog1, prog2, prog3. 1798 * prog1 calls bpf_timer_init for some map1 elements 1799 * prog2 calls bpf_timer_set_callback for some map1 elements. 1800 * Those that were not bpf_timer_init-ed will return -EINVAL. 1801 * prog3 calls bpf_timer_start for some map1 elements. 1802 * Those that were not both bpf_timer_init-ed and 1803 * bpf_timer_set_callback-ed will return -EINVAL. 1804 */ 1805 struct bpf_insn ld_addrs[2] = { 1806 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux), 1807 }; 1808 1809 insn_buf[0] = ld_addrs[0]; 1810 insn_buf[1] = ld_addrs[1]; 1811 insn_buf[2] = *insn; 1812 cnt = 3; 1813 1814 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1815 if (!new_prog) 1816 return -ENOMEM; 1817 1818 delta += cnt - 1; 1819 env->prog = prog = new_prog; 1820 insn = new_prog->insnsi + i + delta; 1821 goto patch_call_imm; 1822 } 1823 1824 /* bpf_per_cpu_ptr() and bpf_this_cpu_ptr() */ 1825 if (env->insn_aux_data[i + delta].call_with_percpu_alloc_ptr) { 1826 /* patch with 'r1 = *(u64 *)(r1 + 0)' since for percpu data, 1827 * bpf_mem_alloc() returns a ptr to the percpu data ptr. 1828 */ 1829 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_1, BPF_REG_1, 0); 1830 insn_buf[1] = *insn; 1831 cnt = 2; 1832 1833 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1834 if (!new_prog) 1835 return -ENOMEM; 1836 1837 delta += cnt - 1; 1838 env->prog = prog = new_prog; 1839 insn = new_prog->insnsi + i + delta; 1840 goto patch_call_imm; 1841 } 1842 1843 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup 1844 * and other inlining handlers are currently limited to 64 bit 1845 * only. 1846 */ 1847 if (prog->jit_requested && BITS_PER_LONG == 64 && 1848 (insn->imm == BPF_FUNC_map_lookup_elem || 1849 insn->imm == BPF_FUNC_map_update_elem || 1850 insn->imm == BPF_FUNC_map_delete_elem || 1851 insn->imm == BPF_FUNC_map_push_elem || 1852 insn->imm == BPF_FUNC_map_pop_elem || 1853 insn->imm == BPF_FUNC_map_peek_elem || 1854 insn->imm == BPF_FUNC_redirect_map || 1855 insn->imm == BPF_FUNC_for_each_map_elem || 1856 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) { 1857 aux = &env->insn_aux_data[i + delta]; 1858 if (bpf_map_ptr_poisoned(aux)) 1859 goto patch_call_imm; 1860 1861 map_ptr = aux->map_ptr_state.map_ptr; 1862 ops = map_ptr->ops; 1863 if (insn->imm == BPF_FUNC_map_lookup_elem && 1864 ops->map_gen_lookup) { 1865 cnt = ops->map_gen_lookup(map_ptr, insn_buf); 1866 if (cnt == -EOPNOTSUPP) 1867 goto patch_map_ops_generic; 1868 if (cnt <= 0 || cnt >= INSN_BUF_SIZE) { 1869 verifier_bug(env, "%d insns generated for map lookup", cnt); 1870 return -EFAULT; 1871 } 1872 1873 new_prog = bpf_patch_insn_data(env, i + delta, 1874 insn_buf, cnt); 1875 if (!new_prog) 1876 return -ENOMEM; 1877 1878 delta += cnt - 1; 1879 env->prog = prog = new_prog; 1880 insn = new_prog->insnsi + i + delta; 1881 goto next_insn; 1882 } 1883 1884 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem, 1885 (void *(*)(struct bpf_map *map, void *key))NULL)); 1886 BUILD_BUG_ON(!__same_type(ops->map_delete_elem, 1887 (long (*)(struct bpf_map *map, void *key))NULL)); 1888 BUILD_BUG_ON(!__same_type(ops->map_update_elem, 1889 (long (*)(struct bpf_map *map, void *key, void *value, 1890 u64 flags))NULL)); 1891 BUILD_BUG_ON(!__same_type(ops->map_push_elem, 1892 (long (*)(struct bpf_map *map, void *value, 1893 u64 flags))NULL)); 1894 BUILD_BUG_ON(!__same_type(ops->map_pop_elem, 1895 (long (*)(struct bpf_map *map, void *value))NULL)); 1896 BUILD_BUG_ON(!__same_type(ops->map_peek_elem, 1897 (long (*)(struct bpf_map *map, void *value))NULL)); 1898 BUILD_BUG_ON(!__same_type(ops->map_redirect, 1899 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL)); 1900 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback, 1901 (long (*)(struct bpf_map *map, 1902 bpf_callback_t callback_fn, 1903 void *callback_ctx, 1904 u64 flags))NULL)); 1905 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem, 1906 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL)); 1907 1908 patch_map_ops_generic: 1909 switch (insn->imm) { 1910 case BPF_FUNC_map_lookup_elem: 1911 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem); 1912 goto next_insn; 1913 case BPF_FUNC_map_update_elem: 1914 insn->imm = BPF_CALL_IMM(ops->map_update_elem); 1915 goto next_insn; 1916 case BPF_FUNC_map_delete_elem: 1917 insn->imm = BPF_CALL_IMM(ops->map_delete_elem); 1918 goto next_insn; 1919 case BPF_FUNC_map_push_elem: 1920 insn->imm = BPF_CALL_IMM(ops->map_push_elem); 1921 goto next_insn; 1922 case BPF_FUNC_map_pop_elem: 1923 insn->imm = BPF_CALL_IMM(ops->map_pop_elem); 1924 goto next_insn; 1925 case BPF_FUNC_map_peek_elem: 1926 insn->imm = BPF_CALL_IMM(ops->map_peek_elem); 1927 goto next_insn; 1928 case BPF_FUNC_redirect_map: 1929 insn->imm = BPF_CALL_IMM(ops->map_redirect); 1930 goto next_insn; 1931 case BPF_FUNC_for_each_map_elem: 1932 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback); 1933 goto next_insn; 1934 case BPF_FUNC_map_lookup_percpu_elem: 1935 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem); 1936 goto next_insn; 1937 } 1938 1939 goto patch_call_imm; 1940 } 1941 1942 /* Implement bpf_jiffies64 inline. */ 1943 if (prog->jit_requested && BITS_PER_LONG == 64 && 1944 insn->imm == BPF_FUNC_jiffies64) { 1945 struct bpf_insn ld_jiffies_addr[2] = { 1946 BPF_LD_IMM64(BPF_REG_0, 1947 (unsigned long)&jiffies), 1948 }; 1949 1950 insn_buf[0] = ld_jiffies_addr[0]; 1951 insn_buf[1] = ld_jiffies_addr[1]; 1952 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, 1953 BPF_REG_0, 0); 1954 cnt = 3; 1955 1956 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1957 cnt); 1958 if (!new_prog) 1959 return -ENOMEM; 1960 1961 delta += cnt - 1; 1962 env->prog = prog = new_prog; 1963 insn = new_prog->insnsi + i + delta; 1964 goto next_insn; 1965 } 1966 1967 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML) 1968 /* Implement bpf_get_smp_processor_id() inline. */ 1969 if (insn->imm == BPF_FUNC_get_smp_processor_id && 1970 bpf_verifier_inlines_helper_call(env, insn->imm)) { 1971 /* BPF_FUNC_get_smp_processor_id inlining is an 1972 * optimization, so if cpu_number is ever 1973 * changed in some incompatible and hard to support 1974 * way, it's fine to back out this inlining logic 1975 */ 1976 #ifdef CONFIG_SMP 1977 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, (u32)(unsigned long)&cpu_number); 1978 insn_buf[1] = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0); 1979 insn_buf[2] = BPF_LDX_MEM(BPF_W, BPF_REG_0, BPF_REG_0, 0); 1980 cnt = 3; 1981 #else 1982 insn_buf[0] = BPF_ALU32_REG(BPF_XOR, BPF_REG_0, BPF_REG_0); 1983 cnt = 1; 1984 #endif 1985 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 1986 if (!new_prog) 1987 return -ENOMEM; 1988 1989 delta += cnt - 1; 1990 env->prog = prog = new_prog; 1991 insn = new_prog->insnsi + i + delta; 1992 goto next_insn; 1993 } 1994 1995 /* Implement bpf_get_current_task() and bpf_get_current_task_btf() inline. */ 1996 if ((insn->imm == BPF_FUNC_get_current_task || insn->imm == BPF_FUNC_get_current_task_btf) && 1997 bpf_verifier_inlines_helper_call(env, insn->imm)) { 1998 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, (u32)(unsigned long)¤t_task); 1999 insn_buf[1] = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0); 2000 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_0, 0); 2001 cnt = 3; 2002 2003 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 2004 if (!new_prog) 2005 return -ENOMEM; 2006 2007 delta += cnt - 1; 2008 env->prog = prog = new_prog; 2009 insn = new_prog->insnsi + i + delta; 2010 goto next_insn; 2011 } 2012 #endif 2013 /* Implement bpf_get_func_arg inline. */ 2014 if (prog_type == BPF_PROG_TYPE_TRACING && 2015 insn->imm == BPF_FUNC_get_func_arg) { 2016 if (eatype == BPF_TRACE_RAW_TP) { 2017 int nr_args = btf_type_vlen(prog->aux->attach_func_proto); 2018 2019 /* skip 'void *__data' in btf_trace_##name() and save to reg0 */ 2020 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, nr_args - 1); 2021 cnt = 1; 2022 } else { 2023 /* Load nr_args from ctx - 8 */ 2024 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8); 2025 insn_buf[1] = BPF_ALU64_IMM(BPF_AND, BPF_REG_0, 0xFF); 2026 cnt = 2; 2027 } 2028 insn_buf[cnt++] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6); 2029 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3); 2030 insn_buf[cnt++] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1); 2031 insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0); 2032 insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0); 2033 insn_buf[cnt++] = BPF_MOV64_IMM(BPF_REG_0, 0); 2034 insn_buf[cnt++] = BPF_JMP_A(1); 2035 insn_buf[cnt++] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL); 2036 2037 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 2038 if (!new_prog) 2039 return -ENOMEM; 2040 2041 delta += cnt - 1; 2042 env->prog = prog = new_prog; 2043 insn = new_prog->insnsi + i + delta; 2044 goto next_insn; 2045 } 2046 2047 /* Implement bpf_get_func_ret inline. */ 2048 if (prog_type == BPF_PROG_TYPE_TRACING && 2049 insn->imm == BPF_FUNC_get_func_ret) { 2050 if (eatype == BPF_TRACE_FEXIT || 2051 eatype == BPF_TRACE_FSESSION || 2052 eatype == BPF_MODIFY_RETURN) { 2053 /* Load nr_args from ctx - 8 */ 2054 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8); 2055 insn_buf[1] = BPF_ALU64_IMM(BPF_AND, BPF_REG_0, 0xFF); 2056 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3); 2057 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1); 2058 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0); 2059 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0); 2060 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0); 2061 cnt = 7; 2062 } else { 2063 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP); 2064 cnt = 1; 2065 } 2066 2067 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 2068 if (!new_prog) 2069 return -ENOMEM; 2070 2071 delta += cnt - 1; 2072 env->prog = prog = new_prog; 2073 insn = new_prog->insnsi + i + delta; 2074 goto next_insn; 2075 } 2076 2077 /* Implement get_func_arg_cnt inline. */ 2078 if (prog_type == BPF_PROG_TYPE_TRACING && 2079 insn->imm == BPF_FUNC_get_func_arg_cnt) { 2080 if (eatype == BPF_TRACE_RAW_TP) { 2081 int nr_args = btf_type_vlen(prog->aux->attach_func_proto); 2082 2083 /* skip 'void *__data' in btf_trace_##name() and save to reg0 */ 2084 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, nr_args - 1); 2085 cnt = 1; 2086 } else { 2087 /* Load nr_args from ctx - 8 */ 2088 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8); 2089 insn_buf[1] = BPF_ALU64_IMM(BPF_AND, BPF_REG_0, 0xFF); 2090 cnt = 2; 2091 } 2092 2093 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 2094 if (!new_prog) 2095 return -ENOMEM; 2096 2097 delta += cnt - 1; 2098 env->prog = prog = new_prog; 2099 insn = new_prog->insnsi + i + delta; 2100 goto next_insn; 2101 } 2102 2103 /* Implement bpf_get_func_ip inline. */ 2104 if (prog_type == BPF_PROG_TYPE_TRACING && 2105 insn->imm == BPF_FUNC_get_func_ip) { 2106 /* Load IP address from ctx - 16 */ 2107 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16); 2108 2109 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1); 2110 if (!new_prog) 2111 return -ENOMEM; 2112 2113 env->prog = prog = new_prog; 2114 insn = new_prog->insnsi + i + delta; 2115 goto next_insn; 2116 } 2117 2118 /* Implement bpf_get_branch_snapshot inline. */ 2119 if (IS_ENABLED(CONFIG_PERF_EVENTS) && 2120 prog->jit_requested && BITS_PER_LONG == 64 && 2121 insn->imm == BPF_FUNC_get_branch_snapshot) { 2122 /* We are dealing with the following func protos: 2123 * u64 bpf_get_branch_snapshot(void *buf, u32 size, u64 flags); 2124 * int perf_snapshot_branch_stack(struct perf_branch_entry *entries, u32 cnt); 2125 */ 2126 const u32 br_entry_size = sizeof(struct perf_branch_entry); 2127 2128 /* struct perf_branch_entry is part of UAPI and is 2129 * used as an array element, so extremely unlikely to 2130 * ever grow or shrink 2131 */ 2132 BUILD_BUG_ON(br_entry_size != 24); 2133 2134 /* if (unlikely(flags)) return -EINVAL */ 2135 insn_buf[0] = BPF_JMP_IMM(BPF_JNE, BPF_REG_3, 0, 7); 2136 2137 /* Transform size (bytes) into number of entries (cnt = size / 24). 2138 * But to avoid expensive division instruction, we implement 2139 * divide-by-3 through multiplication, followed by further 2140 * division by 8 through 3-bit right shift. 2141 * Refer to book "Hacker's Delight, 2nd ed." by Henry S. Warren, Jr., 2142 * p. 227, chapter "Unsigned Division by 3" for details and proofs. 2143 * 2144 * N / 3 <=> M * N / 2^33, where M = (2^33 + 1) / 3 = 0xaaaaaaab. 2145 */ 2146 insn_buf[1] = BPF_MOV32_IMM(BPF_REG_0, 0xaaaaaaab); 2147 insn_buf[2] = BPF_ALU64_REG(BPF_MUL, BPF_REG_2, BPF_REG_0); 2148 insn_buf[3] = BPF_ALU64_IMM(BPF_RSH, BPF_REG_2, 36); 2149 2150 /* call perf_snapshot_branch_stack implementation */ 2151 insn_buf[4] = BPF_EMIT_CALL(static_call_query(perf_snapshot_branch_stack)); 2152 /* if (entry_cnt == 0) return -ENOENT */ 2153 insn_buf[5] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 4); 2154 /* return entry_cnt * sizeof(struct perf_branch_entry) */ 2155 insn_buf[6] = BPF_ALU32_IMM(BPF_MUL, BPF_REG_0, br_entry_size); 2156 insn_buf[7] = BPF_JMP_A(3); 2157 /* return -EINVAL; */ 2158 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL); 2159 insn_buf[9] = BPF_JMP_A(1); 2160 /* return -ENOENT; */ 2161 insn_buf[10] = BPF_MOV64_IMM(BPF_REG_0, -ENOENT); 2162 cnt = 11; 2163 2164 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 2165 if (!new_prog) 2166 return -ENOMEM; 2167 2168 delta += cnt - 1; 2169 env->prog = prog = new_prog; 2170 insn = new_prog->insnsi + i + delta; 2171 goto next_insn; 2172 } 2173 2174 /* Implement bpf_kptr_xchg inline */ 2175 if (prog->jit_requested && BITS_PER_LONG == 64 && 2176 insn->imm == BPF_FUNC_kptr_xchg && 2177 bpf_jit_supports_ptr_xchg()) { 2178 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_2); 2179 insn_buf[1] = BPF_ATOMIC_OP(BPF_DW, BPF_XCHG, BPF_REG_1, BPF_REG_0, 0); 2180 cnt = 2; 2181 2182 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 2183 if (!new_prog) 2184 return -ENOMEM; 2185 2186 delta += cnt - 1; 2187 env->prog = prog = new_prog; 2188 insn = new_prog->insnsi + i + delta; 2189 goto next_insn; 2190 } 2191 patch_call_imm: 2192 fn = env->ops->get_func_proto(insn->imm, env->prog); 2193 /* all functions that have prototype and verifier allowed 2194 * programs to call them, must be real in-kernel functions 2195 */ 2196 if (!fn->func) { 2197 verifier_bug(env, 2198 "not inlined functions %s#%d is missing func", 2199 func_id_name(insn->imm), insn->imm); 2200 return -EFAULT; 2201 } 2202 insn->imm = fn->func - __bpf_call_base; 2203 next_insn: 2204 if (subprogs[cur_subprog + 1].start == i + delta + 1) { 2205 subprogs[cur_subprog].stack_depth += stack_depth_extra; 2206 subprogs[cur_subprog].stack_extra = stack_depth_extra; 2207 2208 stack_depth = subprogs[cur_subprog].stack_depth; 2209 if (stack_depth > MAX_BPF_STACK && !prog->jit_requested) { 2210 verbose(env, "stack size %d(extra %d) is too large\n", 2211 stack_depth, stack_depth_extra); 2212 return -EINVAL; 2213 } 2214 cur_subprog++; 2215 stack_depth = subprogs[cur_subprog].stack_depth; 2216 stack_depth_extra = 0; 2217 } 2218 i++; 2219 insn++; 2220 } 2221 2222 env->prog->aux->stack_depth = subprogs[0].stack_depth; 2223 for (i = 0; i < env->subprog_cnt; i++) { 2224 int delta = bpf_jit_supports_timed_may_goto() ? 2 : 1; 2225 int subprog_start = subprogs[i].start; 2226 int stack_slots = subprogs[i].stack_extra / 8; 2227 int slots = delta, cnt = 0; 2228 2229 if (!stack_slots) 2230 continue; 2231 /* We need two slots in case timed may_goto is supported. */ 2232 if (stack_slots > slots) { 2233 verifier_bug(env, "stack_slots supports may_goto only"); 2234 return -EFAULT; 2235 } 2236 2237 stack_depth = subprogs[i].stack_depth; 2238 if (bpf_jit_supports_timed_may_goto()) { 2239 insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth, 2240 BPF_MAX_TIMED_LOOPS); 2241 insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth + 8, 0); 2242 } else { 2243 /* Add ST insn to subprog prologue to init extra stack */ 2244 insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth, 2245 BPF_MAX_LOOPS); 2246 } 2247 /* Copy first actual insn to preserve it */ 2248 insn_buf[cnt++] = env->prog->insnsi[subprog_start]; 2249 2250 new_prog = bpf_patch_insn_data(env, subprog_start, insn_buf, cnt); 2251 if (!new_prog) 2252 return -ENOMEM; 2253 env->prog = prog = new_prog; 2254 /* 2255 * If may_goto is a first insn of a prog there could be a jmp 2256 * insn that points to it, hence adjust all such jmps to point 2257 * to insn after BPF_ST that inits may_goto count. 2258 * Adjustment will succeed because bpf_patch_insn_data() didn't fail. 2259 */ 2260 WARN_ON(adjust_jmp_off(env->prog, subprog_start, delta)); 2261 } 2262 2263 /* Since poke tab is now finalized, publish aux to tracker. */ 2264 for (i = 0; i < prog->aux->size_poke_tab; i++) { 2265 map_ptr = prog->aux->poke_tab[i].tail_call.map; 2266 if (!map_ptr->ops->map_poke_track || 2267 !map_ptr->ops->map_poke_untrack || 2268 !map_ptr->ops->map_poke_run) { 2269 verifier_bug(env, "poke tab is misconfigured"); 2270 return -EFAULT; 2271 } 2272 2273 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux); 2274 if (ret < 0) { 2275 verbose(env, "tracking tail call prog failed\n"); 2276 return ret; 2277 } 2278 } 2279 2280 ret = sort_kfunc_descs_by_imm_off(env); 2281 if (ret) 2282 return ret; 2283 2284 return 0; 2285 } 2286 2287 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env, 2288 int position, 2289 s32 stack_base, 2290 u32 callback_subprogno, 2291 u32 *total_cnt) 2292 { 2293 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE; 2294 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE; 2295 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE; 2296 int reg_loop_max = BPF_REG_6; 2297 int reg_loop_cnt = BPF_REG_7; 2298 int reg_loop_ctx = BPF_REG_8; 2299 2300 struct bpf_insn *insn_buf = env->insn_buf; 2301 struct bpf_prog *new_prog; 2302 u32 callback_start; 2303 u32 call_insn_offset; 2304 s32 callback_offset; 2305 u32 cnt = 0; 2306 2307 /* This represents an inlined version of bpf_iter.c:bpf_loop, 2308 * be careful to modify this code in sync. 2309 */ 2310 2311 /* Return error and jump to the end of the patch if 2312 * expected number of iterations is too big. 2313 */ 2314 insn_buf[cnt++] = BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2); 2315 insn_buf[cnt++] = BPF_MOV32_IMM(BPF_REG_0, -E2BIG); 2316 insn_buf[cnt++] = BPF_JMP_IMM(BPF_JA, 0, 0, 16); 2317 /* spill R6, R7, R8 to use these as loop vars */ 2318 insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset); 2319 insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset); 2320 insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset); 2321 /* initialize loop vars */ 2322 insn_buf[cnt++] = BPF_MOV64_REG(reg_loop_max, BPF_REG_1); 2323 insn_buf[cnt++] = BPF_MOV32_IMM(reg_loop_cnt, 0); 2324 insn_buf[cnt++] = BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3); 2325 /* loop header, 2326 * if reg_loop_cnt >= reg_loop_max skip the loop body 2327 */ 2328 insn_buf[cnt++] = BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5); 2329 /* callback call, 2330 * correct callback offset would be set after patching 2331 */ 2332 insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt); 2333 insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx); 2334 insn_buf[cnt++] = BPF_CALL_REL(0); 2335 /* increment loop counter */ 2336 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1); 2337 /* jump to loop header if callback returned 0 */ 2338 insn_buf[cnt++] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6); 2339 /* return value of bpf_loop, 2340 * set R0 to the number of iterations 2341 */ 2342 insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt); 2343 /* restore original values of R6, R7, R8 */ 2344 insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset); 2345 insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset); 2346 insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset); 2347 2348 *total_cnt = cnt; 2349 new_prog = bpf_patch_insn_data(env, position, insn_buf, cnt); 2350 if (!new_prog) 2351 return new_prog; 2352 2353 /* callback start is known only after patching */ 2354 callback_start = env->subprog_info[callback_subprogno].start; 2355 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */ 2356 call_insn_offset = position + 12; 2357 callback_offset = callback_start - call_insn_offset - 1; 2358 new_prog->insnsi[call_insn_offset].imm = callback_offset; 2359 2360 return new_prog; 2361 } 2362 2363 static bool is_bpf_loop_call(struct bpf_insn *insn) 2364 { 2365 return insn->code == (BPF_JMP | BPF_CALL) && 2366 insn->src_reg == 0 && 2367 insn->imm == BPF_FUNC_loop; 2368 } 2369 2370 /* For all sub-programs in the program (including main) check 2371 * insn_aux_data to see if there are bpf_loop calls that require 2372 * inlining. If such calls are found the calls are replaced with a 2373 * sequence of instructions produced by `inline_bpf_loop` function and 2374 * subprog stack_depth is increased by the size of 3 registers. 2375 * This stack space is used to spill values of the R6, R7, R8. These 2376 * registers are used to store the loop bound, counter and context 2377 * variables. 2378 */ 2379 int bpf_optimize_bpf_loop(struct bpf_verifier_env *env) 2380 { 2381 struct bpf_subprog_info *subprogs = env->subprog_info; 2382 int i, cur_subprog = 0, cnt, delta = 0; 2383 struct bpf_insn *insn = env->prog->insnsi; 2384 int insn_cnt = env->prog->len; 2385 u16 stack_depth = subprogs[cur_subprog].stack_depth; 2386 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth; 2387 u16 stack_depth_extra = 0; 2388 2389 for (i = 0; i < insn_cnt; i++, insn++) { 2390 struct bpf_loop_inline_state *inline_state = 2391 &env->insn_aux_data[i + delta].loop_inline_state; 2392 2393 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) { 2394 struct bpf_prog *new_prog; 2395 2396 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup; 2397 new_prog = inline_bpf_loop(env, 2398 i + delta, 2399 -(stack_depth + stack_depth_extra), 2400 inline_state->callback_subprogno, 2401 &cnt); 2402 if (!new_prog) 2403 return -ENOMEM; 2404 2405 delta += cnt - 1; 2406 env->prog = new_prog; 2407 insn = new_prog->insnsi + i + delta; 2408 } 2409 2410 if (subprogs[cur_subprog + 1].start == i + delta + 1) { 2411 subprogs[cur_subprog].stack_depth += stack_depth_extra; 2412 cur_subprog++; 2413 stack_depth = subprogs[cur_subprog].stack_depth; 2414 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth; 2415 stack_depth_extra = 0; 2416 } 2417 } 2418 2419 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth; 2420 2421 return 0; 2422 } 2423 2424 /* Remove unnecessary spill/fill pairs, members of fastcall pattern, 2425 * adjust subprograms stack depth when possible. 2426 */ 2427 int bpf_remove_fastcall_spills_fills(struct bpf_verifier_env *env) 2428 { 2429 struct bpf_subprog_info *subprog = env->subprog_info; 2430 struct bpf_insn_aux_data *aux = env->insn_aux_data; 2431 struct bpf_insn *insn = env->prog->insnsi; 2432 int insn_cnt = env->prog->len; 2433 u32 spills_num; 2434 bool modified = false; 2435 int i, j; 2436 2437 for (i = 0; i < insn_cnt; i++, insn++) { 2438 if (aux[i].fastcall_spills_num > 0) { 2439 spills_num = aux[i].fastcall_spills_num; 2440 /* NOPs would be removed by opt_remove_nops() */ 2441 for (j = 1; j <= spills_num; ++j) { 2442 *(insn - j) = NOP; 2443 *(insn + j) = NOP; 2444 } 2445 modified = true; 2446 } 2447 if ((subprog + 1)->start == i + 1) { 2448 if (modified && !subprog->keep_fastcall_stack) 2449 subprog->stack_depth = -subprog->fastcall_stack_off; 2450 subprog++; 2451 modified = false; 2452 } 2453 } 2454 2455 return 0; 2456 } 2457 2458