1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Linux Socket Filter - Kernel level socket filtering 4 * 5 * Based on the design of the Berkeley Packet Filter. The new 6 * internal format has been designed by PLUMgrid: 7 * 8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 9 * 10 * Authors: 11 * 12 * Jay Schulist <jschlst@samba.org> 13 * Alexei Starovoitov <ast@plumgrid.com> 14 * Daniel Borkmann <dborkman@redhat.com> 15 * 16 * Andi Kleen - Fix a few bad bugs and races. 17 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 18 */ 19 20 #include <uapi/linux/btf.h> 21 #include <linux/filter.h> 22 #include <linux/skbuff.h> 23 #include <linux/static_call.h> 24 #include <linux/vmalloc.h> 25 #include <linux/prandom.h> 26 #include <linux/bpf.h> 27 #include <linux/btf.h> 28 #include <linux/hex.h> 29 #include <linux/objtool.h> 30 #include <linux/overflow.h> 31 #include <linux/rbtree_latch.h> 32 #include <linux/kallsyms.h> 33 #include <linux/rcupdate.h> 34 #include <linux/perf_event.h> 35 #include <linux/extable.h> 36 #include <linux/log2.h> 37 #include <linux/bpf_verifier.h> 38 #include <linux/nodemask.h> 39 #include <linux/nospec.h> 40 #include <linux/bpf_mem_alloc.h> 41 #include <linux/memcontrol.h> 42 #include <linux/execmem.h> 43 #include <crypto/sha2.h> 44 45 #include <asm/barrier.h> 46 #include <linux/unaligned.h> 47 48 /* Registers */ 49 #define BPF_R0 regs[BPF_REG_0] 50 #define BPF_R1 regs[BPF_REG_1] 51 #define BPF_R2 regs[BPF_REG_2] 52 #define BPF_R3 regs[BPF_REG_3] 53 #define BPF_R4 regs[BPF_REG_4] 54 #define BPF_R5 regs[BPF_REG_5] 55 #define BPF_R6 regs[BPF_REG_6] 56 #define BPF_R7 regs[BPF_REG_7] 57 #define BPF_R8 regs[BPF_REG_8] 58 #define BPF_R9 regs[BPF_REG_9] 59 #define BPF_R10 regs[BPF_REG_10] 60 61 /* Named registers */ 62 #define DST regs[insn->dst_reg] 63 #define SRC regs[insn->src_reg] 64 #define FP regs[BPF_REG_FP] 65 #define AX regs[BPF_REG_AX] 66 #define ARG1 regs[BPF_REG_ARG1] 67 #define CTX regs[BPF_REG_CTX] 68 #define OFF insn->off 69 #define IMM insn->imm 70 71 struct bpf_mem_alloc bpf_global_ma; 72 bool bpf_global_ma_set; 73 74 /* No hurry in this branch 75 * 76 * Exported for the bpf jit load helper. 77 */ 78 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 79 { 80 u8 *ptr = NULL; 81 82 if (k >= SKF_NET_OFF) { 83 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 84 } else if (k >= SKF_LL_OFF) { 85 if (unlikely(!skb_mac_header_was_set(skb))) 86 return NULL; 87 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 88 } 89 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 90 return ptr; 91 92 return NULL; 93 } 94 95 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 96 enum page_size_enum { 97 __PAGE_SIZE = PAGE_SIZE 98 }; 99 100 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 101 { 102 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 103 struct bpf_prog_aux *aux; 104 struct bpf_prog *fp; 105 106 size = round_up(size, __PAGE_SIZE); 107 fp = __vmalloc(size, gfp_flags); 108 if (fp == NULL) 109 return NULL; 110 111 aux = kzalloc_obj(*aux, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 112 if (aux == NULL) { 113 vfree(fp); 114 return NULL; 115 } 116 fp->active = __alloc_percpu_gfp(sizeof(u8[BPF_NR_CONTEXTS]), 4, 117 bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 118 if (!fp->active) { 119 vfree(fp); 120 kfree(aux); 121 return NULL; 122 } 123 124 fp->pages = size / PAGE_SIZE; 125 fp->aux = aux; 126 fp->aux->main_prog_aux = aux; 127 fp->aux->prog = fp; 128 fp->jit_requested = ebpf_jit_enabled(); 129 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 130 #ifdef CONFIG_CGROUP_BPF 131 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 132 #endif 133 134 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 135 #ifdef CONFIG_FINEIBT 136 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 137 #endif 138 mutex_init(&fp->aux->used_maps_mutex); 139 mutex_init(&fp->aux->ext_mutex); 140 mutex_init(&fp->aux->dst_mutex); 141 mutex_init(&fp->aux->st_ops_assoc_mutex); 142 143 #ifdef CONFIG_BPF_SYSCALL 144 bpf_prog_stream_init(fp); 145 #endif 146 147 return fp; 148 } 149 150 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 151 { 152 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 153 struct bpf_prog *prog; 154 int cpu; 155 156 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 157 if (!prog) 158 return NULL; 159 160 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 161 if (!prog->stats) { 162 free_percpu(prog->active); 163 kfree(prog->aux); 164 vfree(prog); 165 return NULL; 166 } 167 168 for_each_possible_cpu(cpu) { 169 struct bpf_prog_stats *pstats; 170 171 pstats = per_cpu_ptr(prog->stats, cpu); 172 u64_stats_init(&pstats->syncp); 173 } 174 return prog; 175 } 176 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 177 178 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 179 { 180 if (!prog->aux->nr_linfo || !prog->jit_requested) 181 return 0; 182 183 prog->aux->jited_linfo = kvzalloc_objs(*prog->aux->jited_linfo, 184 prog->aux->nr_linfo, 185 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 186 if (!prog->aux->jited_linfo) 187 return -ENOMEM; 188 189 return 0; 190 } 191 192 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 193 { 194 if (prog->aux->jited_linfo && 195 (!prog->jited || !prog->aux->jited_linfo[0])) { 196 kvfree(prog->aux->jited_linfo); 197 prog->aux->jited_linfo = NULL; 198 } 199 200 kfree(prog->aux->kfunc_tab); 201 prog->aux->kfunc_tab = NULL; 202 } 203 204 /* The jit engine is responsible to provide an array 205 * for insn_off to the jited_off mapping (insn_to_jit_off). 206 * 207 * The idx to this array is the insn_off. Hence, the insn_off 208 * here is relative to the prog itself instead of the main prog. 209 * This array has one entry for each xlated bpf insn. 210 * 211 * jited_off is the byte off to the end of the jited insn. 212 * 213 * Hence, with 214 * insn_start: 215 * The first bpf insn off of the prog. The insn off 216 * here is relative to the main prog. 217 * e.g. if prog is a subprog, insn_start > 0 218 * linfo_idx: 219 * The prog's idx to prog->aux->linfo and jited_linfo 220 * 221 * jited_linfo[linfo_idx] = prog->bpf_func 222 * 223 * For i > linfo_idx, 224 * 225 * jited_linfo[i] = prog->bpf_func + 226 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 227 */ 228 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 229 const u32 *insn_to_jit_off) 230 { 231 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 232 const struct bpf_line_info *linfo; 233 void **jited_linfo; 234 235 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 236 /* Userspace did not provide linfo */ 237 return; 238 239 linfo_idx = prog->aux->linfo_idx; 240 linfo = &prog->aux->linfo[linfo_idx]; 241 insn_start = linfo[0].insn_off; 242 insn_end = insn_start + prog->len; 243 244 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 245 jited_linfo[0] = prog->bpf_func; 246 247 nr_linfo = prog->aux->nr_linfo - linfo_idx; 248 249 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 250 /* The verifier ensures that linfo[i].insn_off is 251 * strictly increasing 252 */ 253 jited_linfo[i] = prog->bpf_func + 254 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 255 } 256 257 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 258 gfp_t gfp_extra_flags) 259 { 260 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 261 struct bpf_prog *fp; 262 u32 pages; 263 264 size = round_up(size, PAGE_SIZE); 265 pages = size / PAGE_SIZE; 266 if (pages <= fp_old->pages) 267 return fp_old; 268 269 fp = __vmalloc(size, gfp_flags); 270 if (fp) { 271 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 272 fp->pages = pages; 273 fp->aux->prog = fp; 274 275 /* We keep fp->aux from fp_old around in the new 276 * reallocated structure. 277 */ 278 fp_old->aux = NULL; 279 fp_old->stats = NULL; 280 fp_old->active = NULL; 281 __bpf_prog_free(fp_old); 282 } 283 284 return fp; 285 } 286 287 void __bpf_prog_free(struct bpf_prog *fp) 288 { 289 if (fp->aux) { 290 mutex_destroy(&fp->aux->used_maps_mutex); 291 mutex_destroy(&fp->aux->dst_mutex); 292 mutex_destroy(&fp->aux->st_ops_assoc_mutex); 293 kfree(fp->aux->poke_tab); 294 kfree(fp->aux); 295 } 296 free_percpu(fp->stats); 297 free_percpu(fp->active); 298 vfree(fp); 299 } 300 301 int bpf_prog_calc_tag(struct bpf_prog *fp) 302 { 303 size_t size = bpf_prog_insn_size(fp); 304 struct bpf_insn *dst; 305 bool was_ld_map; 306 u32 i; 307 308 dst = vmalloc(size); 309 if (!dst) 310 return -ENOMEM; 311 312 /* We need to take out the map fd for the digest calculation 313 * since they are unstable from user space side. 314 */ 315 for (i = 0, was_ld_map = false; i < fp->len; i++) { 316 dst[i] = fp->insnsi[i]; 317 if (!was_ld_map && 318 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 319 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 320 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 321 was_ld_map = true; 322 dst[i].imm = 0; 323 } else if (was_ld_map && 324 dst[i].code == 0 && 325 dst[i].dst_reg == 0 && 326 dst[i].src_reg == 0 && 327 dst[i].off == 0) { 328 was_ld_map = false; 329 dst[i].imm = 0; 330 } else { 331 was_ld_map = false; 332 } 333 } 334 sha256((u8 *)dst, size, fp->digest); 335 vfree(dst); 336 return 0; 337 } 338 339 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 340 s32 end_new, s32 curr, const bool probe_pass) 341 { 342 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 343 s32 delta = end_new - end_old; 344 s64 imm = insn->imm; 345 346 if (curr < pos && curr + imm + 1 >= end_old) 347 imm += delta; 348 else if (curr >= end_new && curr + imm + 1 < end_new) 349 imm -= delta; 350 if (imm < imm_min || imm > imm_max) 351 return -ERANGE; 352 if (!probe_pass) 353 insn->imm = imm; 354 return 0; 355 } 356 357 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 358 s32 end_new, s32 curr, const bool probe_pass) 359 { 360 s64 off_min, off_max, off; 361 s32 delta = end_new - end_old; 362 363 if (insn->code == (BPF_JMP32 | BPF_JA)) { 364 off = insn->imm; 365 off_min = S32_MIN; 366 off_max = S32_MAX; 367 } else { 368 off = insn->off; 369 off_min = S16_MIN; 370 off_max = S16_MAX; 371 } 372 373 if (curr < pos && curr + off + 1 >= end_old) 374 off += delta; 375 else if (curr >= end_new && curr + off + 1 < end_new) 376 off -= delta; 377 if (off < off_min || off > off_max) 378 return -ERANGE; 379 if (!probe_pass) { 380 if (insn->code == (BPF_JMP32 | BPF_JA)) 381 insn->imm = off; 382 else 383 insn->off = off; 384 } 385 return 0; 386 } 387 388 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 389 s32 end_new, const bool probe_pass) 390 { 391 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 392 struct bpf_insn *insn = prog->insnsi; 393 int ret = 0; 394 395 for (i = 0; i < insn_cnt; i++, insn++) { 396 u8 code; 397 398 /* In the probing pass we still operate on the original, 399 * unpatched image in order to check overflows before we 400 * do any other adjustments. Therefore skip the patchlet. 401 */ 402 if (probe_pass && i == pos) { 403 i = end_new; 404 insn = prog->insnsi + end_old; 405 } 406 if (bpf_pseudo_func(insn)) { 407 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 408 end_new, i, probe_pass); 409 if (ret) 410 return ret; 411 continue; 412 } 413 code = insn->code; 414 if ((BPF_CLASS(code) != BPF_JMP && 415 BPF_CLASS(code) != BPF_JMP32) || 416 BPF_OP(code) == BPF_EXIT) 417 continue; 418 /* Adjust offset of jmps if we cross patch boundaries. */ 419 if (BPF_OP(code) == BPF_CALL) { 420 if (insn->src_reg != BPF_PSEUDO_CALL) 421 continue; 422 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 423 end_new, i, probe_pass); 424 } else { 425 ret = bpf_adj_delta_to_off(insn, pos, end_old, 426 end_new, i, probe_pass); 427 } 428 if (ret) 429 break; 430 } 431 432 return ret; 433 } 434 435 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 436 { 437 struct bpf_line_info *linfo; 438 u32 i, nr_linfo; 439 440 nr_linfo = prog->aux->nr_linfo; 441 if (!nr_linfo || !delta) 442 return; 443 444 linfo = prog->aux->linfo; 445 446 for (i = 0; i < nr_linfo; i++) 447 if (off < linfo[i].insn_off) 448 break; 449 450 /* Push all off < linfo[i].insn_off by delta */ 451 for (; i < nr_linfo; i++) 452 linfo[i].insn_off += delta; 453 } 454 455 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 456 const struct bpf_insn *patch, u32 len) 457 { 458 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 459 const u32 cnt_max = S16_MAX; 460 struct bpf_prog *prog_adj; 461 int err; 462 463 /* Since our patchlet doesn't expand the image, we're done. */ 464 if (insn_delta == 0) { 465 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 466 return prog; 467 } 468 469 insn_adj_cnt = prog->len + insn_delta; 470 471 /* Reject anything that would potentially let the insn->off 472 * target overflow when we have excessive program expansions. 473 * We need to probe here before we do any reallocation where 474 * we afterwards may not fail anymore. 475 */ 476 if (insn_adj_cnt > cnt_max && 477 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 478 return ERR_PTR(err); 479 480 /* Several new instructions need to be inserted. Make room 481 * for them. Likely, there's no need for a new allocation as 482 * last page could have large enough tailroom. 483 */ 484 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 485 GFP_USER); 486 if (!prog_adj) 487 return ERR_PTR(-ENOMEM); 488 489 prog_adj->len = insn_adj_cnt; 490 491 /* Patching happens in 3 steps: 492 * 493 * 1) Move over tail of insnsi from next instruction onwards, 494 * so we can patch the single target insn with one or more 495 * new ones (patching is always from 1 to n insns, n > 0). 496 * 2) Inject new instructions at the target location. 497 * 3) Adjust branch offsets if necessary. 498 */ 499 insn_rest = insn_adj_cnt - off - len; 500 501 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 502 sizeof(*patch) * insn_rest); 503 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 504 505 /* We are guaranteed to not fail at this point, otherwise 506 * the ship has sailed to reverse to the original state. An 507 * overflow cannot happen at this point. 508 */ 509 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 510 511 bpf_adj_linfo(prog_adj, off, insn_delta); 512 513 return prog_adj; 514 } 515 516 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 517 { 518 int err; 519 520 /* Branch offsets can't overflow when program is shrinking, no need 521 * to call bpf_adj_branches(..., true) here 522 */ 523 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 524 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 525 prog->len -= cnt; 526 527 err = bpf_adj_branches(prog, off, off + cnt, off, false); 528 WARN_ON_ONCE(err); 529 return err; 530 } 531 532 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 533 { 534 int i; 535 536 for (i = 0; i < fp->aux->real_func_cnt; i++) 537 bpf_prog_kallsyms_del(fp->aux->func[i]); 538 } 539 540 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 541 { 542 bpf_prog_kallsyms_del_subprogs(fp); 543 bpf_prog_kallsyms_del(fp); 544 } 545 546 #ifdef CONFIG_BPF_JIT 547 /* All BPF JIT sysctl knobs here. */ 548 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 549 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 550 int bpf_jit_harden __read_mostly; 551 long bpf_jit_limit __read_mostly; 552 long bpf_jit_limit_max __read_mostly; 553 554 static void 555 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 556 { 557 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 558 559 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 560 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 561 } 562 563 static void 564 bpf_prog_ksym_set_name(struct bpf_prog *prog) 565 { 566 char *sym = prog->aux->ksym.name; 567 const char *end = sym + KSYM_NAME_LEN; 568 const struct btf_type *type; 569 const char *func_name; 570 571 BUILD_BUG_ON(sizeof("bpf_prog_") + 572 sizeof(prog->tag) * 2 + 573 /* name has been null terminated. 574 * We should need +1 for the '_' preceding 575 * the name. However, the null character 576 * is double counted between the name and the 577 * sizeof("bpf_prog_") above, so we omit 578 * the +1 here. 579 */ 580 sizeof(prog->aux->name) > KSYM_NAME_LEN); 581 582 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 583 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 584 585 /* prog->aux->name will be ignored if full btf name is available */ 586 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 587 type = btf_type_by_id(prog->aux->btf, 588 prog->aux->func_info[prog->aux->func_idx].type_id); 589 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 590 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 591 return; 592 } 593 594 if (prog->aux->name[0]) 595 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 596 else 597 *sym = 0; 598 } 599 600 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 601 { 602 return container_of(n, struct bpf_ksym, tnode)->start; 603 } 604 605 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 606 struct latch_tree_node *b) 607 { 608 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 609 } 610 611 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 612 { 613 unsigned long val = (unsigned long)key; 614 const struct bpf_ksym *ksym; 615 616 ksym = container_of(n, struct bpf_ksym, tnode); 617 618 if (val < ksym->start) 619 return -1; 620 /* Ensure that we detect return addresses as part of the program, when 621 * the final instruction is a call for a program part of the stack 622 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 623 */ 624 if (val > ksym->end) 625 return 1; 626 627 return 0; 628 } 629 630 static const struct latch_tree_ops bpf_tree_ops = { 631 .less = bpf_tree_less, 632 .comp = bpf_tree_comp, 633 }; 634 635 static DEFINE_SPINLOCK(bpf_lock); 636 static LIST_HEAD(bpf_kallsyms); 637 static struct latch_tree_root bpf_tree __cacheline_aligned; 638 639 void bpf_ksym_add(struct bpf_ksym *ksym) 640 { 641 spin_lock_bh(&bpf_lock); 642 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 643 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 644 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 645 spin_unlock_bh(&bpf_lock); 646 } 647 648 static void __bpf_ksym_del(struct bpf_ksym *ksym) 649 { 650 if (list_empty(&ksym->lnode)) 651 return; 652 653 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 654 list_del_rcu(&ksym->lnode); 655 } 656 657 void bpf_ksym_del(struct bpf_ksym *ksym) 658 { 659 spin_lock_bh(&bpf_lock); 660 __bpf_ksym_del(ksym); 661 spin_unlock_bh(&bpf_lock); 662 } 663 664 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 665 { 666 return fp->jited && !bpf_prog_was_classic(fp); 667 } 668 669 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 670 { 671 if (!bpf_prog_kallsyms_candidate(fp) || 672 !bpf_token_capable(fp->aux->token, CAP_BPF)) 673 return; 674 675 bpf_prog_ksym_set_addr(fp); 676 bpf_prog_ksym_set_name(fp); 677 fp->aux->ksym.prog = true; 678 679 bpf_ksym_add(&fp->aux->ksym); 680 681 #ifdef CONFIG_FINEIBT 682 /* 683 * When FineIBT, code in the __cfi_foo() symbols can get executed 684 * and hence unwinder needs help. 685 */ 686 if (cfi_mode != CFI_FINEIBT) 687 return; 688 689 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 690 "__cfi_%s", fp->aux->ksym.name); 691 692 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 693 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 694 695 bpf_ksym_add(&fp->aux->ksym_prefix); 696 #endif 697 } 698 699 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 700 { 701 if (!bpf_prog_kallsyms_candidate(fp)) 702 return; 703 704 bpf_ksym_del(&fp->aux->ksym); 705 #ifdef CONFIG_FINEIBT 706 if (cfi_mode != CFI_FINEIBT) 707 return; 708 bpf_ksym_del(&fp->aux->ksym_prefix); 709 #endif 710 } 711 712 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 713 { 714 struct latch_tree_node *n; 715 716 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 717 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 718 } 719 720 int bpf_address_lookup(unsigned long addr, unsigned long *size, 721 unsigned long *off, char *sym) 722 { 723 struct bpf_ksym *ksym; 724 int ret = 0; 725 726 rcu_read_lock(); 727 ksym = bpf_ksym_find(addr); 728 if (ksym) { 729 unsigned long symbol_start = ksym->start; 730 unsigned long symbol_end = ksym->end; 731 732 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 733 734 if (size) 735 *size = symbol_end - symbol_start; 736 if (off) 737 *off = addr - symbol_start; 738 } 739 rcu_read_unlock(); 740 741 return ret; 742 } 743 744 bool is_bpf_text_address(unsigned long addr) 745 { 746 bool ret; 747 748 rcu_read_lock(); 749 ret = bpf_ksym_find(addr) != NULL; 750 rcu_read_unlock(); 751 752 return ret; 753 } 754 755 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 756 { 757 struct bpf_ksym *ksym; 758 759 WARN_ON_ONCE(!rcu_read_lock_held()); 760 ksym = bpf_ksym_find(addr); 761 762 return ksym && ksym->prog ? 763 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 764 NULL; 765 } 766 767 bool bpf_has_frame_pointer(unsigned long ip) 768 { 769 struct bpf_ksym *ksym; 770 unsigned long offset; 771 772 guard(rcu)(); 773 774 ksym = bpf_ksym_find(ip); 775 if (!ksym || !ksym->fp_start || !ksym->fp_end) 776 return false; 777 778 offset = ip - ksym->start; 779 780 return offset >= ksym->fp_start && offset < ksym->fp_end; 781 } 782 783 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 784 { 785 const struct exception_table_entry *e = NULL; 786 struct bpf_prog *prog; 787 788 rcu_read_lock(); 789 prog = bpf_prog_ksym_find(addr); 790 if (!prog) 791 goto out; 792 if (!prog->aux->num_exentries) 793 goto out; 794 795 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 796 out: 797 rcu_read_unlock(); 798 return e; 799 } 800 801 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 802 char *sym) 803 { 804 struct bpf_ksym *ksym; 805 unsigned int it = 0; 806 int ret = -ERANGE; 807 808 if (!bpf_jit_kallsyms_enabled()) 809 return ret; 810 811 rcu_read_lock(); 812 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 813 if (it++ != symnum) 814 continue; 815 816 strscpy(sym, ksym->name, KSYM_NAME_LEN); 817 818 *value = ksym->start; 819 *type = BPF_SYM_ELF_TYPE; 820 821 ret = 0; 822 break; 823 } 824 rcu_read_unlock(); 825 826 return ret; 827 } 828 829 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 830 struct bpf_jit_poke_descriptor *poke) 831 { 832 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 833 static const u32 poke_tab_max = 1024; 834 u32 slot = prog->aux->size_poke_tab; 835 u32 size = slot + 1; 836 837 if (size > poke_tab_max) 838 return -ENOSPC; 839 if (poke->tailcall_target || poke->tailcall_target_stable || 840 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 841 return -EINVAL; 842 843 switch (poke->reason) { 844 case BPF_POKE_REASON_TAIL_CALL: 845 if (!poke->tail_call.map) 846 return -EINVAL; 847 break; 848 default: 849 return -EINVAL; 850 } 851 852 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 853 if (!tab) 854 return -ENOMEM; 855 856 memcpy(&tab[slot], poke, sizeof(*poke)); 857 prog->aux->size_poke_tab = size; 858 prog->aux->poke_tab = tab; 859 860 return slot; 861 } 862 863 /* 864 * BPF program pack allocator. 865 * 866 * Most BPF programs are pretty small. Allocating a hole page for each 867 * program is sometime a waste. Many small bpf program also adds pressure 868 * to instruction TLB. To solve this issue, we introduce a BPF program pack 869 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 870 * to host BPF programs. 871 */ 872 #define BPF_PROG_CHUNK_SHIFT 6 873 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 874 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 875 876 struct bpf_prog_pack { 877 struct list_head list; 878 void *ptr; 879 bool arch_flush_needed; 880 unsigned long bitmap[]; 881 }; 882 883 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 884 { 885 memset(area, 0, size); 886 } 887 888 DEFINE_STATIC_CALL_NULL(bpf_arch_pred_flush, bpf_arch_pred_flush); 889 890 /* 891 * Enabled once bpf_arch_pred_flush points at a real flush routine. Lets the 892 * pack allocator test "is a predictor flush wired up at all" with a cheap 893 * static branch instead of repeatedly querying the static call target. 894 */ 895 DEFINE_STATIC_KEY_FALSE(bpf_pred_flush_enabled); 896 897 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 898 899 static DEFINE_MUTEX(pack_mutex); 900 static LIST_HEAD(pack_list); 901 902 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 903 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 904 */ 905 #ifdef PMD_SIZE 906 /* PMD_SIZE is really big for some archs. It doesn't make sense to 907 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 908 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 909 * greater than or equal to 2MB. 910 */ 911 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 912 #else 913 #define BPF_PROG_PACK_SIZE PAGE_SIZE 914 #endif 915 916 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 917 918 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 919 { 920 struct bpf_prog_pack *pack; 921 int err; 922 923 pack = kzalloc_flex(*pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)); 924 if (!pack) 925 return NULL; 926 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 927 if (!pack->ptr) 928 goto out; 929 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 930 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 931 932 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 933 pack->arch_flush_needed = true; 934 set_vm_flush_reset_perms(pack->ptr); 935 err = set_memory_rox((unsigned long)pack->ptr, 936 BPF_PROG_PACK_SIZE / PAGE_SIZE); 937 if (err) 938 goto out; 939 list_add_tail(&pack->list, &pack_list); 940 return pack; 941 942 out: 943 bpf_jit_free_exec(pack->ptr); 944 kfree(pack); 945 return NULL; 946 } 947 948 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns, bool was_classic) 949 { 950 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 951 struct bpf_prog_pack *pack, *fallback_pack = NULL; 952 unsigned long pos, fallback_pos = 0; 953 void *ptr = NULL; 954 955 mutex_lock(&pack_mutex); 956 if (size > BPF_PROG_PACK_SIZE) { 957 /* 958 * Allocations larger than a pack get their own pages, and 959 * predictors are not flushed for such allocation. This is only 960 * safe because cBPF programs (the unprivileged attack surface) 961 * are bounded well below a pack size. 962 */ 963 if (was_classic && static_branch_unlikely(&bpf_pred_flush_enabled)) 964 pr_warn_once("BPF: Predictors not flushed for allocations greater than BPF_PROG_PACK_SIZE\n"); 965 size = round_up(size, PAGE_SIZE); 966 ptr = bpf_jit_alloc_exec(size); 967 if (ptr) { 968 int err; 969 970 bpf_fill_ill_insns(ptr, size); 971 set_vm_flush_reset_perms(ptr); 972 err = set_memory_rox((unsigned long)ptr, 973 size / PAGE_SIZE); 974 if (err) { 975 bpf_jit_free_exec(ptr); 976 ptr = NULL; 977 } 978 } 979 goto out; 980 } 981 list_for_each_entry(pack, &pack_list, list) { 982 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 983 nbits, 0); 984 if (pos >= BPF_PROG_CHUNK_COUNT) 985 continue; 986 /* Flush not enabled, use any pack */ 987 if (!static_branch_unlikely(&bpf_pred_flush_enabled)) 988 goto found_free_area; 989 /* 990 * cBPF reuse of a dirty pack triggers a flush, so prefer a 991 * clean pack for cBPF. eBPF never flushes, so steer it to a 992 * dirty pack and keep clean packs free for cBPF. 993 */ 994 if (was_classic ^ pack->arch_flush_needed) 995 goto found_free_area; 996 if (!fallback_pack) { 997 fallback_pack = pack; 998 fallback_pos = pos; 999 } 1000 } 1001 1002 /* No preferred pack found */ 1003 if (fallback_pack) { 1004 pack = fallback_pack; 1005 pos = fallback_pos; 1006 goto found_free_area; 1007 } 1008 1009 pack = alloc_new_pack(bpf_fill_ill_insns); 1010 if (!pack) 1011 goto out; 1012 1013 pos = 0; 1014 1015 found_free_area: 1016 /* Flush only for cBPF as it may contain a crafted gadget */ 1017 if (static_branch_unlikely(&bpf_pred_flush_enabled) && 1018 pack->arch_flush_needed && 1019 was_classic) { 1020 struct bpf_prog_pack *p; 1021 1022 static_call_cond(bpf_arch_pred_flush)(); 1023 list_for_each_entry(p, &pack_list, list) 1024 p->arch_flush_needed = false; 1025 } 1026 bitmap_set(pack->bitmap, pos, nbits); 1027 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 1028 1029 out: 1030 mutex_unlock(&pack_mutex); 1031 return ptr; 1032 } 1033 1034 void bpf_prog_pack_free(void *ptr, u32 size) 1035 { 1036 struct bpf_prog_pack *pack = NULL, *tmp; 1037 unsigned int nbits; 1038 unsigned long pos; 1039 1040 mutex_lock(&pack_mutex); 1041 if (size > BPF_PROG_PACK_SIZE) { 1042 bpf_jit_free_exec(ptr); 1043 goto out; 1044 } 1045 1046 list_for_each_entry(tmp, &pack_list, list) { 1047 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 1048 pack = tmp; 1049 break; 1050 } 1051 } 1052 1053 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1054 goto out; 1055 1056 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1057 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1058 1059 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1060 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1061 1062 bitmap_clear(pack->bitmap, pos, nbits); 1063 1064 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 1065 pack->arch_flush_needed = true; 1066 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1067 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1068 list_del(&pack->list); 1069 bpf_jit_free_exec(pack->ptr); 1070 kfree(pack); 1071 } 1072 out: 1073 mutex_unlock(&pack_mutex); 1074 } 1075 1076 static atomic_long_t bpf_jit_current; 1077 1078 /* Can be overridden by an arch's JIT compiler if it has a custom, 1079 * dedicated BPF backend memory area, or if neither of the two 1080 * below apply. 1081 */ 1082 u64 __weak bpf_jit_alloc_exec_limit(void) 1083 { 1084 #if defined(MODULES_VADDR) 1085 return MODULES_END - MODULES_VADDR; 1086 #else 1087 return VMALLOC_END - VMALLOC_START; 1088 #endif 1089 } 1090 1091 static int __init bpf_jit_charge_init(void) 1092 { 1093 /* Only used as heuristic here to derive limit. */ 1094 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1095 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1096 PAGE_SIZE), LONG_MAX); 1097 return 0; 1098 } 1099 pure_initcall(bpf_jit_charge_init); 1100 1101 int bpf_jit_charge_modmem(u32 size) 1102 { 1103 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1104 if (!bpf_capable()) { 1105 atomic_long_sub(size, &bpf_jit_current); 1106 return -EPERM; 1107 } 1108 } 1109 1110 return 0; 1111 } 1112 1113 void bpf_jit_uncharge_modmem(u32 size) 1114 { 1115 atomic_long_sub(size, &bpf_jit_current); 1116 } 1117 1118 void *__weak bpf_jit_alloc_exec(unsigned long size) 1119 { 1120 return execmem_alloc(EXECMEM_BPF, size); 1121 } 1122 1123 void __weak bpf_jit_free_exec(void *addr) 1124 { 1125 execmem_free(addr); 1126 } 1127 1128 struct bpf_binary_header * 1129 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1130 unsigned int alignment, 1131 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1132 { 1133 struct bpf_binary_header *hdr; 1134 u32 size, hole, start; 1135 1136 WARN_ON_ONCE(!is_power_of_2(alignment) || 1137 alignment > BPF_IMAGE_ALIGNMENT); 1138 1139 /* Most of BPF filters are really small, but if some of them 1140 * fill a page, allow at least 128 extra bytes to insert a 1141 * random section of illegal instructions. 1142 */ 1143 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1144 1145 if (bpf_jit_charge_modmem(size)) 1146 return NULL; 1147 hdr = bpf_jit_alloc_exec(size); 1148 if (!hdr) { 1149 bpf_jit_uncharge_modmem(size); 1150 return NULL; 1151 } 1152 1153 /* Fill space with illegal/arch-dep instructions. */ 1154 bpf_fill_ill_insns(hdr, size); 1155 1156 hdr->size = size; 1157 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1158 PAGE_SIZE - sizeof(*hdr)); 1159 start = get_random_u32_below(hole) & ~(alignment - 1); 1160 1161 /* Leave a random number of instructions before BPF code. */ 1162 *image_ptr = &hdr->image[start]; 1163 1164 return hdr; 1165 } 1166 1167 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1168 { 1169 u32 size = hdr->size; 1170 1171 bpf_jit_free_exec(hdr); 1172 bpf_jit_uncharge_modmem(size); 1173 } 1174 1175 /* Allocate jit binary from bpf_prog_pack allocator. 1176 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1177 * to the memory. To solve this problem, a RW buffer is also allocated at 1178 * as the same time. The JIT engine should calculate offsets based on the 1179 * RO memory address, but write JITed program to the RW buffer. Once the 1180 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1181 * the JITed program to the RO memory. 1182 */ 1183 struct bpf_binary_header * 1184 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1185 unsigned int alignment, 1186 struct bpf_binary_header **rw_header, 1187 u8 **rw_image, 1188 bpf_jit_fill_hole_t bpf_fill_ill_insns, 1189 bool was_classic) 1190 { 1191 struct bpf_binary_header *ro_header; 1192 u32 size, hole, start; 1193 1194 WARN_ON_ONCE(!is_power_of_2(alignment) || 1195 alignment > BPF_IMAGE_ALIGNMENT); 1196 1197 /* add 16 bytes for a random section of illegal instructions */ 1198 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1199 1200 if (bpf_jit_charge_modmem(size)) 1201 return NULL; 1202 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns, was_classic); 1203 if (!ro_header) { 1204 bpf_jit_uncharge_modmem(size); 1205 return NULL; 1206 } 1207 1208 *rw_header = kvmalloc(size, GFP_KERNEL); 1209 if (!*rw_header) { 1210 bpf_prog_pack_free(ro_header, size); 1211 bpf_jit_uncharge_modmem(size); 1212 return NULL; 1213 } 1214 1215 /* Fill space with illegal/arch-dep instructions. */ 1216 bpf_fill_ill_insns(*rw_header, size); 1217 (*rw_header)->size = size; 1218 1219 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1220 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1221 start = get_random_u32_below(hole) & ~(alignment - 1); 1222 1223 *image_ptr = &ro_header->image[start]; 1224 *rw_image = &(*rw_header)->image[start]; 1225 1226 return ro_header; 1227 } 1228 1229 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1230 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1231 struct bpf_binary_header *rw_header) 1232 { 1233 void *ptr; 1234 1235 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1236 1237 kvfree(rw_header); 1238 1239 if (IS_ERR(ptr)) { 1240 bpf_prog_pack_free(ro_header, ro_header->size); 1241 return PTR_ERR(ptr); 1242 } 1243 return 0; 1244 } 1245 1246 /* bpf_jit_binary_pack_free is called in two different scenarios: 1247 * 1) when the program is freed after; 1248 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1249 * For case 2), we need to free both the RO memory and the RW buffer. 1250 * 1251 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1252 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1253 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1254 * bpf_arch_text_copy (when jit fails). 1255 */ 1256 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1257 struct bpf_binary_header *rw_header) 1258 { 1259 u32 size = ro_header->size; 1260 1261 bpf_prog_pack_free(ro_header, size); 1262 kvfree(rw_header); 1263 bpf_jit_uncharge_modmem(size); 1264 } 1265 1266 struct bpf_binary_header * 1267 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1268 { 1269 unsigned long real_start = (unsigned long)fp->bpf_func; 1270 unsigned long addr; 1271 1272 addr = real_start & BPF_PROG_CHUNK_MASK; 1273 return (void *)addr; 1274 } 1275 1276 static inline struct bpf_binary_header * 1277 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1278 { 1279 unsigned long real_start = (unsigned long)fp->bpf_func; 1280 unsigned long addr; 1281 1282 addr = real_start & PAGE_MASK; 1283 return (void *)addr; 1284 } 1285 1286 /* This symbol is only overridden by archs that have different 1287 * requirements than the usual eBPF JITs, f.e. when they only 1288 * implement cBPF JIT, do not set images read-only, etc. 1289 */ 1290 void __weak bpf_jit_free(struct bpf_prog *fp) 1291 { 1292 if (fp->jited) { 1293 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1294 1295 bpf_jit_binary_free(hdr); 1296 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1297 } 1298 1299 bpf_prog_unlock_free(fp); 1300 } 1301 1302 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1303 const struct bpf_insn *insn, bool extra_pass, 1304 u64 *func_addr, bool *func_addr_fixed) 1305 { 1306 s16 off = insn->off; 1307 s32 imm = insn->imm; 1308 u8 *addr; 1309 int err; 1310 1311 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1312 if (!*func_addr_fixed) { 1313 /* Place-holder address till the last pass has collected 1314 * all addresses for JITed subprograms in which case we 1315 * can pick them up from prog->aux. 1316 */ 1317 if (!extra_pass) 1318 addr = NULL; 1319 else if (prog->aux->func && 1320 off >= 0 && off < prog->aux->real_func_cnt) 1321 addr = (u8 *)prog->aux->func[off]->bpf_func; 1322 else 1323 return -EINVAL; 1324 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1325 bpf_jit_supports_far_kfunc_call()) { 1326 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1327 if (err) 1328 return err; 1329 } else { 1330 /* Address of a BPF helper call. Since part of the core 1331 * kernel, it's always at a fixed location. __bpf_call_base 1332 * and the helper with imm relative to it are both in core 1333 * kernel. 1334 */ 1335 addr = (u8 *)__bpf_call_base + imm; 1336 } 1337 1338 *func_addr = (unsigned long)addr; 1339 return 0; 1340 } 1341 1342 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1343 { 1344 if (prog->aux->ksym.prog) 1345 return prog->aux->ksym.name; 1346 return prog->aux->name; 1347 } 1348 1349 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1350 const struct bpf_insn *aux, 1351 struct bpf_insn *to_buff, 1352 bool emit_zext) 1353 { 1354 struct bpf_insn *to = to_buff; 1355 u32 imm_rnd = get_random_u32(); 1356 s16 off; 1357 1358 BUILD_BUG_ON(BPF_REG_PARAMS + 2 != MAX_BPF_JIT_REG); 1359 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1360 1361 /* Constraints on AX register: 1362 * 1363 * AX register is inaccessible from user space. It is mapped in 1364 * all JITs, and used here for constant blinding rewrites. It is 1365 * typically "stateless" meaning its contents are only valid within 1366 * the executed instruction, but not across several instructions. 1367 * There are a few exceptions however which are further detailed 1368 * below. 1369 * 1370 * Constant blinding is only used by JITs, not in the interpreter. 1371 * The interpreter uses AX in some occasions as a local temporary 1372 * register e.g. in DIV or MOD instructions. 1373 * 1374 * In restricted circumstances, the verifier can also use the AX 1375 * register for rewrites as long as they do not interfere with 1376 * the above cases! 1377 */ 1378 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1379 goto out; 1380 1381 if (from->imm == 0 && 1382 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1383 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1384 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1385 goto out; 1386 } 1387 1388 switch (from->code) { 1389 case BPF_ALU | BPF_ADD | BPF_K: 1390 case BPF_ALU | BPF_SUB | BPF_K: 1391 case BPF_ALU | BPF_AND | BPF_K: 1392 case BPF_ALU | BPF_OR | BPF_K: 1393 case BPF_ALU | BPF_XOR | BPF_K: 1394 case BPF_ALU | BPF_MUL | BPF_K: 1395 case BPF_ALU | BPF_MOV | BPF_K: 1396 case BPF_ALU | BPF_DIV | BPF_K: 1397 case BPF_ALU | BPF_MOD | BPF_K: 1398 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1399 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1400 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1401 break; 1402 1403 case BPF_ALU64 | BPF_ADD | BPF_K: 1404 case BPF_ALU64 | BPF_SUB | BPF_K: 1405 case BPF_ALU64 | BPF_AND | BPF_K: 1406 case BPF_ALU64 | BPF_OR | BPF_K: 1407 case BPF_ALU64 | BPF_XOR | BPF_K: 1408 case BPF_ALU64 | BPF_MUL | BPF_K: 1409 case BPF_ALU64 | BPF_MOV | BPF_K: 1410 case BPF_ALU64 | BPF_DIV | BPF_K: 1411 case BPF_ALU64 | BPF_MOD | BPF_K: 1412 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1413 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1414 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1415 break; 1416 1417 case BPF_JMP | BPF_JEQ | BPF_K: 1418 case BPF_JMP | BPF_JNE | BPF_K: 1419 case BPF_JMP | BPF_JGT | BPF_K: 1420 case BPF_JMP | BPF_JLT | BPF_K: 1421 case BPF_JMP | BPF_JGE | BPF_K: 1422 case BPF_JMP | BPF_JLE | BPF_K: 1423 case BPF_JMP | BPF_JSGT | BPF_K: 1424 case BPF_JMP | BPF_JSLT | BPF_K: 1425 case BPF_JMP | BPF_JSGE | BPF_K: 1426 case BPF_JMP | BPF_JSLE | BPF_K: 1427 case BPF_JMP | BPF_JSET | BPF_K: 1428 /* Accommodate for extra offset in case of a backjump. */ 1429 off = from->off; 1430 if (off < 0) 1431 off -= 2; 1432 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1433 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1434 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1435 break; 1436 1437 case BPF_JMP32 | BPF_JEQ | BPF_K: 1438 case BPF_JMP32 | BPF_JNE | BPF_K: 1439 case BPF_JMP32 | BPF_JGT | BPF_K: 1440 case BPF_JMP32 | BPF_JLT | BPF_K: 1441 case BPF_JMP32 | BPF_JGE | BPF_K: 1442 case BPF_JMP32 | BPF_JLE | BPF_K: 1443 case BPF_JMP32 | BPF_JSGT | BPF_K: 1444 case BPF_JMP32 | BPF_JSLT | BPF_K: 1445 case BPF_JMP32 | BPF_JSGE | BPF_K: 1446 case BPF_JMP32 | BPF_JSLE | BPF_K: 1447 case BPF_JMP32 | BPF_JSET | BPF_K: 1448 /* Accommodate for extra offset in case of a backjump. */ 1449 off = from->off; 1450 if (off < 0) 1451 off -= 2; 1452 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1453 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1454 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1455 off); 1456 break; 1457 1458 case BPF_LD | BPF_IMM | BPF_DW: 1459 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1460 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1461 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1462 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1463 break; 1464 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1465 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1466 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1467 if (emit_zext) 1468 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1469 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1470 break; 1471 1472 case BPF_ST | BPF_MEM | BPF_DW: 1473 case BPF_ST | BPF_MEM | BPF_W: 1474 case BPF_ST | BPF_MEM | BPF_H: 1475 case BPF_ST | BPF_MEM | BPF_B: 1476 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1477 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1478 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1479 break; 1480 1481 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW: 1482 case BPF_ST | BPF_PROBE_MEM32 | BPF_W: 1483 case BPF_ST | BPF_PROBE_MEM32 | BPF_H: 1484 case BPF_ST | BPF_PROBE_MEM32 | BPF_B: 1485 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ 1486 from->imm); 1487 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1488 /* 1489 * Cannot use BPF_STX_MEM() macro here as it 1490 * hardcodes BPF_MEM mode, losing PROBE_MEM32 1491 * and breaking arena addressing in the JIT. 1492 */ 1493 *to++ = (struct bpf_insn) { 1494 .code = BPF_STX | BPF_PROBE_MEM32 | 1495 BPF_SIZE(from->code), 1496 .dst_reg = from->dst_reg, 1497 .src_reg = BPF_REG_AX, 1498 .off = from->off, 1499 }; 1500 break; 1501 } 1502 out: 1503 return to - to_buff; 1504 } 1505 1506 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1507 gfp_t gfp_extra_flags) 1508 { 1509 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1510 struct bpf_prog *fp; 1511 1512 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1513 if (fp != NULL) { 1514 /* aux->prog still points to the fp_other one, so 1515 * when promoting the clone to the real program, 1516 * this still needs to be adapted. 1517 */ 1518 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1519 } 1520 1521 return fp; 1522 } 1523 1524 static void bpf_prog_clone_free(struct bpf_prog *fp) 1525 { 1526 /* aux was stolen by the other clone, so we cannot free 1527 * it from this path! It will be freed eventually by the 1528 * other program on release. 1529 * 1530 * At this point, we don't need a deferred release since 1531 * clone is guaranteed to not be locked. 1532 */ 1533 fp->aux = NULL; 1534 fp->stats = NULL; 1535 fp->active = NULL; 1536 __bpf_prog_free(fp); 1537 } 1538 1539 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1540 { 1541 /* We have to repoint aux->prog to self, as we don't 1542 * know whether fp here is the clone or the original. 1543 */ 1544 fp->aux->prog = fp; 1545 if (fp->aux->offload) 1546 fp->aux->offload->prog = fp; 1547 bpf_prog_clone_free(fp_other); 1548 } 1549 1550 /* 1551 * Now this function is used only to blind the main prog and must be invoked only when 1552 * bpf_prog_need_blind() returns true. 1553 */ 1554 struct bpf_prog *bpf_jit_blind_constants(struct bpf_verifier_env *env, struct bpf_prog *prog) 1555 { 1556 struct bpf_insn insn_buff[16], aux[2]; 1557 struct bpf_prog *clone, *tmp; 1558 int insn_delta, insn_cnt; 1559 struct bpf_insn *insn; 1560 int i, rewritten; 1561 1562 if (WARN_ON_ONCE(env && env->prog != prog)) 1563 return ERR_PTR(-EINVAL); 1564 1565 clone = bpf_prog_clone_create(prog, GFP_USER); 1566 if (!clone) 1567 return ERR_PTR(-ENOMEM); 1568 1569 /* make sure bpf_patch_insn_data() patches the correct prog */ 1570 if (env) 1571 env->prog = clone; 1572 1573 insn_cnt = clone->len; 1574 insn = clone->insnsi; 1575 1576 for (i = 0; i < insn_cnt; i++, insn++) { 1577 if (bpf_pseudo_func(insn)) { 1578 /* ld_imm64 with an address of bpf subprog is not 1579 * a user controlled constant. Don't randomize it, 1580 * since it will conflict with jit_subprogs() logic. 1581 */ 1582 insn++; 1583 i++; 1584 continue; 1585 } 1586 1587 /* We temporarily need to hold the original ld64 insn 1588 * so that we can still access the first part in the 1589 * second blinding run. 1590 */ 1591 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1592 insn[1].code == 0) 1593 memcpy(aux, insn, sizeof(aux)); 1594 1595 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1596 clone->aux->verifier_zext); 1597 if (!rewritten) 1598 continue; 1599 1600 if (env) 1601 tmp = bpf_patch_insn_data(env, i, insn_buff, rewritten); 1602 else 1603 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1604 1605 if (IS_ERR_OR_NULL(tmp)) { 1606 if (env) 1607 /* restore the original prog */ 1608 env->prog = prog; 1609 /* Patching may have repointed aux->prog during 1610 * realloc from the original one, so we need to 1611 * fix it up here on error. 1612 */ 1613 bpf_jit_prog_release_other(prog, clone); 1614 return IS_ERR(tmp) ? tmp : ERR_PTR(-ENOMEM); 1615 } 1616 1617 clone = tmp; 1618 insn_delta = rewritten - 1; 1619 1620 if (env) 1621 env->prog = clone; 1622 1623 /* Walk new program and skip insns we just inserted. */ 1624 insn = clone->insnsi + i + insn_delta; 1625 insn_cnt += insn_delta; 1626 i += insn_delta; 1627 } 1628 1629 clone->blinded = 1; 1630 return clone; 1631 } 1632 1633 bool bpf_insn_is_indirect_target(const struct bpf_verifier_env *env, const struct bpf_prog *prog, 1634 int insn_idx) 1635 { 1636 if (!env) 1637 return false; 1638 insn_idx += prog->aux->subprog_start; 1639 return env->insn_aux_data[insn_idx].indirect_target; 1640 } 1641 1642 u16 bpf_out_stack_arg_cnt(const struct bpf_verifier_env *env, const struct bpf_prog *prog) 1643 { 1644 const struct bpf_subprog_info *sub; 1645 1646 if (!env) 1647 return 0; 1648 sub = &env->subprog_info[prog->aux->func_idx]; 1649 return sub->stack_arg_cnt - bpf_in_stack_arg_cnt(sub); 1650 } 1651 #endif /* CONFIG_BPF_JIT */ 1652 1653 /* Base function for offset calculation. Needs to go into .text section, 1654 * therefore keeping it non-static as well; will also be used by JITs 1655 * anyway later on, so do not let the compiler omit it. This also needs 1656 * to go into kallsyms for correlation from e.g. bpftool, so naming 1657 * must not change. 1658 */ 1659 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1660 { 1661 return 0; 1662 } 1663 EXPORT_SYMBOL_GPL(__bpf_call_base); 1664 1665 /* All UAPI available opcodes. */ 1666 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1667 /* 32 bit ALU operations. */ \ 1668 /* Register based. */ \ 1669 INSN_3(ALU, ADD, X), \ 1670 INSN_3(ALU, SUB, X), \ 1671 INSN_3(ALU, AND, X), \ 1672 INSN_3(ALU, OR, X), \ 1673 INSN_3(ALU, LSH, X), \ 1674 INSN_3(ALU, RSH, X), \ 1675 INSN_3(ALU, XOR, X), \ 1676 INSN_3(ALU, MUL, X), \ 1677 INSN_3(ALU, MOV, X), \ 1678 INSN_3(ALU, ARSH, X), \ 1679 INSN_3(ALU, DIV, X), \ 1680 INSN_3(ALU, MOD, X), \ 1681 INSN_2(ALU, NEG), \ 1682 INSN_3(ALU, END, TO_BE), \ 1683 INSN_3(ALU, END, TO_LE), \ 1684 /* Immediate based. */ \ 1685 INSN_3(ALU, ADD, K), \ 1686 INSN_3(ALU, SUB, K), \ 1687 INSN_3(ALU, AND, K), \ 1688 INSN_3(ALU, OR, K), \ 1689 INSN_3(ALU, LSH, K), \ 1690 INSN_3(ALU, RSH, K), \ 1691 INSN_3(ALU, XOR, K), \ 1692 INSN_3(ALU, MUL, K), \ 1693 INSN_3(ALU, MOV, K), \ 1694 INSN_3(ALU, ARSH, K), \ 1695 INSN_3(ALU, DIV, K), \ 1696 INSN_3(ALU, MOD, K), \ 1697 /* 64 bit ALU operations. */ \ 1698 /* Register based. */ \ 1699 INSN_3(ALU64, ADD, X), \ 1700 INSN_3(ALU64, SUB, X), \ 1701 INSN_3(ALU64, AND, X), \ 1702 INSN_3(ALU64, OR, X), \ 1703 INSN_3(ALU64, LSH, X), \ 1704 INSN_3(ALU64, RSH, X), \ 1705 INSN_3(ALU64, XOR, X), \ 1706 INSN_3(ALU64, MUL, X), \ 1707 INSN_3(ALU64, MOV, X), \ 1708 INSN_3(ALU64, ARSH, X), \ 1709 INSN_3(ALU64, DIV, X), \ 1710 INSN_3(ALU64, MOD, X), \ 1711 INSN_2(ALU64, NEG), \ 1712 INSN_3(ALU64, END, TO_LE), \ 1713 /* Immediate based. */ \ 1714 INSN_3(ALU64, ADD, K), \ 1715 INSN_3(ALU64, SUB, K), \ 1716 INSN_3(ALU64, AND, K), \ 1717 INSN_3(ALU64, OR, K), \ 1718 INSN_3(ALU64, LSH, K), \ 1719 INSN_3(ALU64, RSH, K), \ 1720 INSN_3(ALU64, XOR, K), \ 1721 INSN_3(ALU64, MUL, K), \ 1722 INSN_3(ALU64, MOV, K), \ 1723 INSN_3(ALU64, ARSH, K), \ 1724 INSN_3(ALU64, DIV, K), \ 1725 INSN_3(ALU64, MOD, K), \ 1726 /* Call instruction. */ \ 1727 INSN_2(JMP, CALL), \ 1728 /* Exit instruction. */ \ 1729 INSN_2(JMP, EXIT), \ 1730 /* 32-bit Jump instructions. */ \ 1731 /* Register based. */ \ 1732 INSN_3(JMP32, JEQ, X), \ 1733 INSN_3(JMP32, JNE, X), \ 1734 INSN_3(JMP32, JGT, X), \ 1735 INSN_3(JMP32, JLT, X), \ 1736 INSN_3(JMP32, JGE, X), \ 1737 INSN_3(JMP32, JLE, X), \ 1738 INSN_3(JMP32, JSGT, X), \ 1739 INSN_3(JMP32, JSLT, X), \ 1740 INSN_3(JMP32, JSGE, X), \ 1741 INSN_3(JMP32, JSLE, X), \ 1742 INSN_3(JMP32, JSET, X), \ 1743 /* Immediate based. */ \ 1744 INSN_3(JMP32, JEQ, K), \ 1745 INSN_3(JMP32, JNE, K), \ 1746 INSN_3(JMP32, JGT, K), \ 1747 INSN_3(JMP32, JLT, K), \ 1748 INSN_3(JMP32, JGE, K), \ 1749 INSN_3(JMP32, JLE, K), \ 1750 INSN_3(JMP32, JSGT, K), \ 1751 INSN_3(JMP32, JSLT, K), \ 1752 INSN_3(JMP32, JSGE, K), \ 1753 INSN_3(JMP32, JSLE, K), \ 1754 INSN_3(JMP32, JSET, K), \ 1755 /* Jump instructions. */ \ 1756 /* Register based. */ \ 1757 INSN_3(JMP, JEQ, X), \ 1758 INSN_3(JMP, JNE, X), \ 1759 INSN_3(JMP, JGT, X), \ 1760 INSN_3(JMP, JLT, X), \ 1761 INSN_3(JMP, JGE, X), \ 1762 INSN_3(JMP, JLE, X), \ 1763 INSN_3(JMP, JSGT, X), \ 1764 INSN_3(JMP, JSLT, X), \ 1765 INSN_3(JMP, JSGE, X), \ 1766 INSN_3(JMP, JSLE, X), \ 1767 INSN_3(JMP, JSET, X), \ 1768 /* Immediate based. */ \ 1769 INSN_3(JMP, JEQ, K), \ 1770 INSN_3(JMP, JNE, K), \ 1771 INSN_3(JMP, JGT, K), \ 1772 INSN_3(JMP, JLT, K), \ 1773 INSN_3(JMP, JGE, K), \ 1774 INSN_3(JMP, JLE, K), \ 1775 INSN_3(JMP, JSGT, K), \ 1776 INSN_3(JMP, JSLT, K), \ 1777 INSN_3(JMP, JSGE, K), \ 1778 INSN_3(JMP, JSLE, K), \ 1779 INSN_3(JMP, JSET, K), \ 1780 INSN_2(JMP, JA), \ 1781 INSN_2(JMP32, JA), \ 1782 /* Atomic operations. */ \ 1783 INSN_3(STX, ATOMIC, B), \ 1784 INSN_3(STX, ATOMIC, H), \ 1785 INSN_3(STX, ATOMIC, W), \ 1786 INSN_3(STX, ATOMIC, DW), \ 1787 /* Store instructions. */ \ 1788 /* Register based. */ \ 1789 INSN_3(STX, MEM, B), \ 1790 INSN_3(STX, MEM, H), \ 1791 INSN_3(STX, MEM, W), \ 1792 INSN_3(STX, MEM, DW), \ 1793 /* Immediate based. */ \ 1794 INSN_3(ST, MEM, B), \ 1795 INSN_3(ST, MEM, H), \ 1796 INSN_3(ST, MEM, W), \ 1797 INSN_3(ST, MEM, DW), \ 1798 /* Load instructions. */ \ 1799 /* Register based. */ \ 1800 INSN_3(LDX, MEM, B), \ 1801 INSN_3(LDX, MEM, H), \ 1802 INSN_3(LDX, MEM, W), \ 1803 INSN_3(LDX, MEM, DW), \ 1804 INSN_3(LDX, MEMSX, B), \ 1805 INSN_3(LDX, MEMSX, H), \ 1806 INSN_3(LDX, MEMSX, W), \ 1807 /* Immediate based. */ \ 1808 INSN_3(LD, IMM, DW) 1809 1810 bool bpf_opcode_in_insntable(u8 code) 1811 { 1812 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1813 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1814 static const bool public_insntable[256] = { 1815 [0 ... 255] = false, 1816 /* Now overwrite non-defaults ... */ 1817 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1818 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1819 [BPF_LD | BPF_ABS | BPF_B] = true, 1820 [BPF_LD | BPF_ABS | BPF_H] = true, 1821 [BPF_LD | BPF_ABS | BPF_W] = true, 1822 [BPF_LD | BPF_IND | BPF_B] = true, 1823 [BPF_LD | BPF_IND | BPF_H] = true, 1824 [BPF_LD | BPF_IND | BPF_W] = true, 1825 [BPF_JMP | BPF_JA | BPF_X] = true, 1826 [BPF_JMP | BPF_JCOND] = true, 1827 }; 1828 #undef BPF_INSN_3_TBL 1829 #undef BPF_INSN_2_TBL 1830 return public_insntable[code]; 1831 } 1832 1833 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1834 /* Absolute value of s32 without undefined behavior for S32_MIN */ 1835 static u32 abs_s32(s32 x) 1836 { 1837 return x >= 0 ? (u32)x : -(u32)x; 1838 } 1839 1840 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1841 const struct bpf_insn *insn); 1842 1843 /** 1844 * ___bpf_prog_run - run eBPF program on a given context 1845 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1846 * @insn: is the array of eBPF instructions 1847 * 1848 * Decode and execute eBPF instructions. 1849 * 1850 * Return: whatever value is in %BPF_R0 at program exit 1851 */ 1852 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1853 { 1854 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1855 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1856 static const void * const jumptable[256] __annotate_jump_table = { 1857 [0 ... 255] = &&default_label, 1858 /* Now overwrite non-defaults ... */ 1859 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1860 /* Non-UAPI available opcodes. */ 1861 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1862 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1863 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1864 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1865 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1866 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1867 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1868 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1869 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1870 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1871 }; 1872 #undef BPF_INSN_3_LBL 1873 #undef BPF_INSN_2_LBL 1874 u32 tail_call_cnt = 0; 1875 1876 #define CONT ({ insn++; goto select_insn; }) 1877 #define CONT_JMP ({ insn++; goto select_insn; }) 1878 1879 select_insn: 1880 goto *jumptable[insn->code]; 1881 1882 /* Explicitly mask the register-based shift amounts with 63 or 31 1883 * to avoid undefined behavior. Normally this won't affect the 1884 * generated code, for example, in case of native 64 bit archs such 1885 * as x86-64 or arm64, the compiler is optimizing the AND away for 1886 * the interpreter. In case of JITs, each of the JIT backends compiles 1887 * the BPF shift operations to machine instructions which produce 1888 * implementation-defined results in such a case; the resulting 1889 * contents of the register may be arbitrary, but program behaviour 1890 * as a whole remains defined. In other words, in case of JIT backends, 1891 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1892 */ 1893 /* ALU (shifts) */ 1894 #define SHT(OPCODE, OP) \ 1895 ALU64_##OPCODE##_X: \ 1896 DST = DST OP (SRC & 63); \ 1897 CONT; \ 1898 ALU_##OPCODE##_X: \ 1899 DST = (u32) DST OP ((u32) SRC & 31); \ 1900 CONT; \ 1901 ALU64_##OPCODE##_K: \ 1902 DST = DST OP IMM; \ 1903 CONT; \ 1904 ALU_##OPCODE##_K: \ 1905 DST = (u32) DST OP (u32) IMM; \ 1906 CONT; 1907 /* ALU (rest) */ 1908 #define ALU(OPCODE, OP) \ 1909 ALU64_##OPCODE##_X: \ 1910 DST = DST OP SRC; \ 1911 CONT; \ 1912 ALU_##OPCODE##_X: \ 1913 DST = (u32) DST OP (u32) SRC; \ 1914 CONT; \ 1915 ALU64_##OPCODE##_K: \ 1916 DST = DST OP IMM; \ 1917 CONT; \ 1918 ALU_##OPCODE##_K: \ 1919 DST = (u32) DST OP (u32) IMM; \ 1920 CONT; 1921 ALU(ADD, +) 1922 ALU(SUB, -) 1923 ALU(AND, &) 1924 ALU(OR, |) 1925 ALU(XOR, ^) 1926 ALU(MUL, *) 1927 SHT(LSH, <<) 1928 SHT(RSH, >>) 1929 #undef SHT 1930 #undef ALU 1931 ALU_NEG: 1932 DST = (u32) -DST; 1933 CONT; 1934 ALU64_NEG: 1935 DST = -DST; 1936 CONT; 1937 ALU_MOV_X: 1938 switch (OFF) { 1939 case 0: 1940 DST = (u32) SRC; 1941 break; 1942 case 8: 1943 DST = (u32)(s8) SRC; 1944 break; 1945 case 16: 1946 DST = (u32)(s16) SRC; 1947 break; 1948 } 1949 CONT; 1950 ALU_MOV_K: 1951 DST = (u32) IMM; 1952 CONT; 1953 ALU64_MOV_X: 1954 switch (OFF) { 1955 case 0: 1956 DST = SRC; 1957 break; 1958 case 8: 1959 DST = (s8) SRC; 1960 break; 1961 case 16: 1962 DST = (s16) SRC; 1963 break; 1964 case 32: 1965 DST = (s32) SRC; 1966 break; 1967 } 1968 CONT; 1969 ALU64_MOV_K: 1970 DST = IMM; 1971 CONT; 1972 LD_IMM_DW: 1973 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1974 insn++; 1975 CONT; 1976 ALU_ARSH_X: 1977 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1978 CONT; 1979 ALU_ARSH_K: 1980 DST = (u64) (u32) (((s32) DST) >> IMM); 1981 CONT; 1982 ALU64_ARSH_X: 1983 (*(s64 *) &DST) >>= (SRC & 63); 1984 CONT; 1985 ALU64_ARSH_K: 1986 (*(s64 *) &DST) >>= IMM; 1987 CONT; 1988 ALU64_MOD_X: 1989 switch (OFF) { 1990 case 0: 1991 div64_u64_rem(DST, SRC, &AX); 1992 DST = AX; 1993 break; 1994 case 1: 1995 AX = div64_s64(DST, SRC); 1996 DST = DST - AX * SRC; 1997 break; 1998 } 1999 CONT; 2000 ALU_MOD_X: 2001 switch (OFF) { 2002 case 0: 2003 AX = (u32) DST; 2004 DST = do_div(AX, (u32) SRC); 2005 break; 2006 case 1: 2007 AX = abs_s32((s32)DST); 2008 AX = do_div(AX, abs_s32((s32)SRC)); 2009 if ((s32)DST < 0) 2010 DST = (u32)-AX; 2011 else 2012 DST = (u32)AX; 2013 break; 2014 } 2015 CONT; 2016 ALU64_MOD_K: 2017 switch (OFF) { 2018 case 0: 2019 div64_u64_rem(DST, IMM, &AX); 2020 DST = AX; 2021 break; 2022 case 1: 2023 AX = div64_s64(DST, IMM); 2024 DST = DST - AX * IMM; 2025 break; 2026 } 2027 CONT; 2028 ALU_MOD_K: 2029 switch (OFF) { 2030 case 0: 2031 AX = (u32) DST; 2032 DST = do_div(AX, (u32) IMM); 2033 break; 2034 case 1: 2035 AX = abs_s32((s32)DST); 2036 AX = do_div(AX, abs_s32((s32)IMM)); 2037 if ((s32)DST < 0) 2038 DST = (u32)-AX; 2039 else 2040 DST = (u32)AX; 2041 break; 2042 } 2043 CONT; 2044 ALU64_DIV_X: 2045 switch (OFF) { 2046 case 0: 2047 DST = div64_u64(DST, SRC); 2048 break; 2049 case 1: 2050 DST = div64_s64(DST, SRC); 2051 break; 2052 } 2053 CONT; 2054 ALU_DIV_X: 2055 switch (OFF) { 2056 case 0: 2057 AX = (u32) DST; 2058 do_div(AX, (u32) SRC); 2059 DST = (u32) AX; 2060 break; 2061 case 1: 2062 AX = abs_s32((s32)DST); 2063 do_div(AX, abs_s32((s32)SRC)); 2064 if (((s32)DST < 0) == ((s32)SRC < 0)) 2065 DST = (u32)AX; 2066 else 2067 DST = (u32)-AX; 2068 break; 2069 } 2070 CONT; 2071 ALU64_DIV_K: 2072 switch (OFF) { 2073 case 0: 2074 DST = div64_u64(DST, IMM); 2075 break; 2076 case 1: 2077 DST = div64_s64(DST, IMM); 2078 break; 2079 } 2080 CONT; 2081 ALU_DIV_K: 2082 switch (OFF) { 2083 case 0: 2084 AX = (u32) DST; 2085 do_div(AX, (u32) IMM); 2086 DST = (u32) AX; 2087 break; 2088 case 1: 2089 AX = abs_s32((s32)DST); 2090 do_div(AX, abs_s32((s32)IMM)); 2091 if (((s32)DST < 0) == ((s32)IMM < 0)) 2092 DST = (u32)AX; 2093 else 2094 DST = (u32)-AX; 2095 break; 2096 } 2097 CONT; 2098 ALU_END_TO_BE: 2099 switch (IMM) { 2100 case 16: 2101 DST = (__force u16) cpu_to_be16(DST); 2102 break; 2103 case 32: 2104 DST = (__force u32) cpu_to_be32(DST); 2105 break; 2106 case 64: 2107 DST = (__force u64) cpu_to_be64(DST); 2108 break; 2109 } 2110 CONT; 2111 ALU_END_TO_LE: 2112 switch (IMM) { 2113 case 16: 2114 DST = (__force u16) cpu_to_le16(DST); 2115 break; 2116 case 32: 2117 DST = (__force u32) cpu_to_le32(DST); 2118 break; 2119 case 64: 2120 DST = (__force u64) cpu_to_le64(DST); 2121 break; 2122 } 2123 CONT; 2124 ALU64_END_TO_LE: 2125 switch (IMM) { 2126 case 16: 2127 DST = (__force u16) __swab16(DST); 2128 break; 2129 case 32: 2130 DST = (__force u32) __swab32(DST); 2131 break; 2132 case 64: 2133 DST = (__force u64) __swab64(DST); 2134 break; 2135 } 2136 CONT; 2137 2138 /* CALL */ 2139 JMP_CALL: 2140 /* Function call scratches BPF_R1-BPF_R5 registers, 2141 * preserves BPF_R6-BPF_R9, and stores return value 2142 * into BPF_R0. 2143 */ 2144 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2145 BPF_R4, BPF_R5); 2146 CONT; 2147 2148 JMP_CALL_ARGS: 2149 BPF_R0 = interpreters_args[insn->off](BPF_R1, BPF_R2, BPF_R3, 2150 BPF_R4, BPF_R5, 2151 insn + insn->imm + 1); 2152 CONT; 2153 2154 JMP_TAIL_CALL: { 2155 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2156 struct bpf_array *array = container_of(map, struct bpf_array, map); 2157 struct bpf_prog *prog; 2158 u32 index = BPF_R3; 2159 2160 if (unlikely(index >= array->map.max_entries)) 2161 goto out; 2162 2163 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2164 goto out; 2165 2166 prog = READ_ONCE(array->ptrs[index]); 2167 if (!prog) 2168 goto out; 2169 2170 tail_call_cnt++; 2171 2172 /* ARG1 at this point is guaranteed to point to CTX from 2173 * the verifier side due to the fact that the tail call is 2174 * handled like a helper, that is, bpf_tail_call_proto, 2175 * where arg1_type is ARG_PTR_TO_CTX. 2176 */ 2177 insn = prog->insnsi; 2178 goto select_insn; 2179 out: 2180 CONT; 2181 } 2182 JMP_JA: 2183 insn += insn->off; 2184 CONT; 2185 JMP32_JA: 2186 insn += insn->imm; 2187 CONT; 2188 JMP_EXIT: 2189 return BPF_R0; 2190 /* JMP */ 2191 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2192 JMP_##OPCODE##_X: \ 2193 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2194 insn += insn->off; \ 2195 CONT_JMP; \ 2196 } \ 2197 CONT; \ 2198 JMP32_##OPCODE##_X: \ 2199 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2200 insn += insn->off; \ 2201 CONT_JMP; \ 2202 } \ 2203 CONT; \ 2204 JMP_##OPCODE##_K: \ 2205 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2206 insn += insn->off; \ 2207 CONT_JMP; \ 2208 } \ 2209 CONT; \ 2210 JMP32_##OPCODE##_K: \ 2211 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2212 insn += insn->off; \ 2213 CONT_JMP; \ 2214 } \ 2215 CONT; 2216 COND_JMP(u, JEQ, ==) 2217 COND_JMP(u, JNE, !=) 2218 COND_JMP(u, JGT, >) 2219 COND_JMP(u, JLT, <) 2220 COND_JMP(u, JGE, >=) 2221 COND_JMP(u, JLE, <=) 2222 COND_JMP(u, JSET, &) 2223 COND_JMP(s, JSGT, >) 2224 COND_JMP(s, JSLT, <) 2225 COND_JMP(s, JSGE, >=) 2226 COND_JMP(s, JSLE, <=) 2227 #undef COND_JMP 2228 /* ST, STX and LDX*/ 2229 ST_NOSPEC: 2230 /* Speculation barrier for mitigating Speculative Store Bypass, 2231 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2232 * rely on the firmware mitigation as controlled via the ssbd 2233 * kernel parameter. Whenever the mitigation is enabled, it 2234 * works for all of the kernel code with no need to provide any 2235 * additional instructions here. In case of x86, we use 'lfence' 2236 * insn for mitigation. We reuse preexisting logic from Spectre 2237 * v1 mitigation that happens to produce the required code on 2238 * x86 for v4 as well. 2239 */ 2240 barrier_nospec(); 2241 CONT; 2242 #define LDST(SIZEOP, SIZE) \ 2243 STX_MEM_##SIZEOP: \ 2244 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2245 CONT; \ 2246 ST_MEM_##SIZEOP: \ 2247 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2248 CONT; \ 2249 LDX_MEM_##SIZEOP: \ 2250 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2251 CONT; \ 2252 LDX_PROBE_MEM_##SIZEOP: \ 2253 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2254 (const void *)(long) (SRC + insn->off)); \ 2255 DST = *((SIZE *)&DST); \ 2256 CONT; 2257 2258 LDST(B, u8) 2259 LDST(H, u16) 2260 LDST(W, u32) 2261 LDST(DW, u64) 2262 #undef LDST 2263 2264 #define LDSX(SIZEOP, SIZE) \ 2265 LDX_MEMSX_##SIZEOP: \ 2266 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2267 CONT; \ 2268 LDX_PROBE_MEMSX_##SIZEOP: \ 2269 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2270 (const void *)(long) (SRC + insn->off)); \ 2271 DST = *((SIZE *)&DST); \ 2272 CONT; 2273 2274 LDSX(B, s8) 2275 LDSX(H, s16) 2276 LDSX(W, s32) 2277 #undef LDSX 2278 2279 #define ATOMIC_ALU_OP(BOP, KOP) \ 2280 case BOP: \ 2281 if (BPF_SIZE(insn->code) == BPF_W) \ 2282 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2283 (DST + insn->off)); \ 2284 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2285 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2286 (DST + insn->off)); \ 2287 else \ 2288 goto default_label; \ 2289 break; \ 2290 case BOP | BPF_FETCH: \ 2291 if (BPF_SIZE(insn->code) == BPF_W) \ 2292 SRC = (u32) atomic_fetch_##KOP( \ 2293 (u32) SRC, \ 2294 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2295 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2296 SRC = (u64) atomic64_fetch_##KOP( \ 2297 (u64) SRC, \ 2298 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2299 else \ 2300 goto default_label; \ 2301 break; 2302 2303 STX_ATOMIC_DW: 2304 STX_ATOMIC_W: 2305 STX_ATOMIC_H: 2306 STX_ATOMIC_B: 2307 switch (IMM) { 2308 /* Atomic read-modify-write instructions support only W and DW 2309 * size modifiers. 2310 */ 2311 ATOMIC_ALU_OP(BPF_ADD, add) 2312 ATOMIC_ALU_OP(BPF_AND, and) 2313 ATOMIC_ALU_OP(BPF_OR, or) 2314 ATOMIC_ALU_OP(BPF_XOR, xor) 2315 #undef ATOMIC_ALU_OP 2316 2317 case BPF_XCHG: 2318 if (BPF_SIZE(insn->code) == BPF_W) 2319 SRC = (u32) atomic_xchg( 2320 (atomic_t *)(unsigned long) (DST + insn->off), 2321 (u32) SRC); 2322 else if (BPF_SIZE(insn->code) == BPF_DW) 2323 SRC = (u64) atomic64_xchg( 2324 (atomic64_t *)(unsigned long) (DST + insn->off), 2325 (u64) SRC); 2326 else 2327 goto default_label; 2328 break; 2329 case BPF_CMPXCHG: 2330 if (BPF_SIZE(insn->code) == BPF_W) 2331 BPF_R0 = (u32) atomic_cmpxchg( 2332 (atomic_t *)(unsigned long) (DST + insn->off), 2333 (u32) BPF_R0, (u32) SRC); 2334 else if (BPF_SIZE(insn->code) == BPF_DW) 2335 BPF_R0 = (u64) atomic64_cmpxchg( 2336 (atomic64_t *)(unsigned long) (DST + insn->off), 2337 (u64) BPF_R0, (u64) SRC); 2338 else 2339 goto default_label; 2340 break; 2341 /* Atomic load and store instructions support all size 2342 * modifiers. 2343 */ 2344 case BPF_LOAD_ACQ: 2345 switch (BPF_SIZE(insn->code)) { 2346 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2347 case BPF_##SIZEOP: \ 2348 DST = (SIZE)smp_load_acquire( \ 2349 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2350 break; 2351 LOAD_ACQUIRE(B, u8) 2352 LOAD_ACQUIRE(H, u16) 2353 LOAD_ACQUIRE(W, u32) 2354 #ifdef CONFIG_64BIT 2355 LOAD_ACQUIRE(DW, u64) 2356 #endif 2357 #undef LOAD_ACQUIRE 2358 default: 2359 goto default_label; 2360 } 2361 break; 2362 case BPF_STORE_REL: 2363 switch (BPF_SIZE(insn->code)) { 2364 #define STORE_RELEASE(SIZEOP, SIZE) \ 2365 case BPF_##SIZEOP: \ 2366 smp_store_release( \ 2367 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2368 break; 2369 STORE_RELEASE(B, u8) 2370 STORE_RELEASE(H, u16) 2371 STORE_RELEASE(W, u32) 2372 #ifdef CONFIG_64BIT 2373 STORE_RELEASE(DW, u64) 2374 #endif 2375 #undef STORE_RELEASE 2376 default: 2377 goto default_label; 2378 } 2379 break; 2380 2381 default: 2382 goto default_label; 2383 } 2384 CONT; 2385 2386 default_label: 2387 /* If we ever reach this, we have a bug somewhere. Die hard here 2388 * instead of just returning 0; we could be somewhere in a subprog, 2389 * so execution could continue otherwise which we do /not/ want. 2390 * 2391 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2392 */ 2393 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2394 insn->code, insn->imm); 2395 BUG_ON(1); 2396 return 0; 2397 } 2398 2399 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2400 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2401 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2402 { \ 2403 u64 stack[stack_size / sizeof(u64)]; \ 2404 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2405 \ 2406 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2407 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2408 ARG1 = (u64) (unsigned long) ctx; \ 2409 return ___bpf_prog_run(regs, insn); \ 2410 } 2411 2412 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2413 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2414 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2415 const struct bpf_insn *insn) \ 2416 { \ 2417 u64 stack[stack_size / sizeof(u64)]; \ 2418 u64 regs[MAX_BPF_EXT_REG]; \ 2419 \ 2420 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2421 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2422 BPF_R1 = r1; \ 2423 BPF_R2 = r2; \ 2424 BPF_R3 = r3; \ 2425 BPF_R4 = r4; \ 2426 BPF_R5 = r5; \ 2427 return ___bpf_prog_run(regs, insn); \ 2428 } 2429 2430 #define EVAL1(FN, X) FN(X) 2431 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2432 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2433 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2434 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2435 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2436 2437 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2438 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2439 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2440 2441 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2442 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2443 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2444 2445 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2446 2447 static unsigned int (*interpreters[])(const void *ctx, 2448 const struct bpf_insn *insn) = { 2449 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2450 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2451 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2452 }; 2453 #undef PROG_NAME_LIST 2454 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2455 static __maybe_unused 2456 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2457 const struct bpf_insn *insn) = { 2458 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2459 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2460 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2461 }; 2462 #undef PROG_NAME_LIST 2463 2464 #ifdef CONFIG_BPF_SYSCALL 2465 int bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2466 { 2467 stack_depth = max_t(u32, stack_depth, 1); 2468 /* Prevent out-of-bounds read to interpreters_args */ 2469 if (stack_depth > MAX_BPF_STACK) 2470 return -EINVAL; 2471 insn->off = (round_up(stack_depth, 32) / 32) - 1; 2472 insn->code = BPF_JMP | BPF_CALL_ARGS; 2473 return 0; 2474 } 2475 2476 s32 bpf_call_args_imm(s16 idx) 2477 { 2478 if (WARN_ON_ONCE(idx < 0 || idx >= ARRAY_SIZE(interpreters_args))) 2479 return 0; 2480 return BPF_CALL_IMM(interpreters_args[idx]); 2481 } 2482 #endif 2483 #endif 2484 2485 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2486 const struct bpf_insn *insn) 2487 { 2488 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2489 * is not working properly, so warn about it! 2490 */ 2491 WARN_ON_ONCE(1); 2492 return 0; 2493 } 2494 2495 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2496 const struct bpf_prog *fp) 2497 { 2498 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2499 struct bpf_prog_aux *aux = fp->aux; 2500 enum bpf_cgroup_storage_type i; 2501 bool ret = false; 2502 u64 cookie; 2503 2504 if (fp->kprobe_override) 2505 return ret; 2506 2507 spin_lock(&map->owner_lock); 2508 /* There's no owner yet where we could check for compatibility. */ 2509 if (!map->owner) { 2510 map->owner = bpf_map_owner_alloc(map); 2511 if (!map->owner) 2512 goto err; 2513 map->owner->type = prog_type; 2514 map->owner->jited = fp->jited; 2515 map->owner->xdp_has_frags = aux->xdp_has_frags; 2516 map->owner->sleepable = fp->sleepable; 2517 map->owner->expected_attach_type = fp->expected_attach_type; 2518 map->owner->attach_func_proto = aux->attach_func_proto; 2519 for_each_cgroup_storage_type(i) { 2520 map->owner->storage_cookie[i] = 2521 aux->cgroup_storage[i] ? 2522 aux->cgroup_storage[i]->cookie : 0; 2523 } 2524 ret = true; 2525 } else { 2526 ret = map->owner->type == prog_type && 2527 map->owner->jited == fp->jited && 2528 map->owner->xdp_has_frags == aux->xdp_has_frags && 2529 map->owner->sleepable == fp->sleepable; 2530 if (ret && 2531 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2532 map->owner->expected_attach_type != fp->expected_attach_type) 2533 ret = false; 2534 for_each_cgroup_storage_type(i) { 2535 if (!ret) 2536 break; 2537 cookie = aux->cgroup_storage[i] ? 2538 aux->cgroup_storage[i]->cookie : 0; 2539 ret = map->owner->storage_cookie[i] == cookie || 2540 (!cookie && !aux->tail_call_reachable); 2541 } 2542 if (ret && 2543 map->owner->attach_func_proto != aux->attach_func_proto) { 2544 switch (prog_type) { 2545 case BPF_PROG_TYPE_TRACING: 2546 case BPF_PROG_TYPE_LSM: 2547 case BPF_PROG_TYPE_EXT: 2548 case BPF_PROG_TYPE_STRUCT_OPS: 2549 ret = false; 2550 break; 2551 default: 2552 break; 2553 } 2554 } 2555 } 2556 err: 2557 spin_unlock(&map->owner_lock); 2558 return ret; 2559 } 2560 2561 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2562 { 2563 /* XDP programs inserted into maps are not guaranteed to run on 2564 * a particular netdev (and can run outside driver context entirely 2565 * in the case of devmap and cpumap). Until device checks 2566 * are implemented, prohibit adding dev-bound programs to program maps. 2567 */ 2568 if (bpf_prog_is_dev_bound(fp->aux)) 2569 return false; 2570 2571 return __bpf_prog_map_compatible(map, fp); 2572 } 2573 2574 static int bpf_check_tail_call(const struct bpf_prog *fp) 2575 { 2576 struct bpf_prog_aux *aux = fp->aux; 2577 int i, ret = 0; 2578 2579 mutex_lock(&aux->used_maps_mutex); 2580 for (i = 0; i < aux->used_map_cnt; i++) { 2581 struct bpf_map *map = aux->used_maps[i]; 2582 2583 if (!map_type_contains_progs(map)) 2584 continue; 2585 2586 if (!__bpf_prog_map_compatible(map, fp)) { 2587 ret = -EINVAL; 2588 goto out; 2589 } 2590 } 2591 2592 out: 2593 mutex_unlock(&aux->used_maps_mutex); 2594 return ret; 2595 } 2596 2597 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2598 { 2599 bool select_interpreter = false; 2600 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2601 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2602 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2603 2604 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2605 * But for non-JITed programs, we don't need bpf_func, so no bounds 2606 * check needed. 2607 */ 2608 if (idx < ARRAY_SIZE(interpreters)) { 2609 fp->bpf_func = interpreters[idx]; 2610 select_interpreter = true; 2611 } else { 2612 fp->bpf_func = __bpf_prog_ret0_warn; 2613 } 2614 #else 2615 fp->bpf_func = __bpf_prog_ret0_warn; 2616 #endif 2617 return select_interpreter; 2618 } 2619 2620 static struct bpf_prog *bpf_prog_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 2621 { 2622 #ifdef CONFIG_BPF_JIT 2623 struct bpf_prog *orig_prog; 2624 struct bpf_insn_aux_data *orig_insn_aux; 2625 2626 if (!bpf_prog_need_blind(prog)) 2627 return bpf_int_jit_compile(env, prog); 2628 2629 if (env) { 2630 /* 2631 * If env is not NULL, we are called from the end of bpf_check(), at this 2632 * point, only insn_aux_data is used after failure, so it should be restored 2633 * on failure. 2634 */ 2635 orig_insn_aux = bpf_dup_insn_aux_data(env); 2636 if (!orig_insn_aux) 2637 return prog; 2638 } 2639 2640 orig_prog = prog; 2641 prog = bpf_jit_blind_constants(env, prog); 2642 /* 2643 * If blinding was requested and we failed during blinding, we must fall 2644 * back to the interpreter. 2645 */ 2646 if (IS_ERR(prog)) 2647 goto out_restore; 2648 2649 prog = bpf_int_jit_compile(env, prog); 2650 if (prog->jited) { 2651 bpf_jit_prog_release_other(prog, orig_prog); 2652 if (env) 2653 vfree(orig_insn_aux); 2654 return prog; 2655 } 2656 2657 bpf_jit_prog_release_other(orig_prog, prog); 2658 2659 out_restore: 2660 prog = orig_prog; 2661 if (env) 2662 bpf_restore_insn_aux_data(env, orig_insn_aux); 2663 #endif 2664 return prog; 2665 } 2666 2667 struct bpf_prog *__bpf_prog_select_runtime(struct bpf_verifier_env *env, struct bpf_prog *fp, 2668 int *err) 2669 { 2670 /* In case of BPF to BPF calls, verifier did all the prep 2671 * work with regards to JITing, etc. 2672 */ 2673 bool jit_needed = false; 2674 2675 if (fp->bpf_func) 2676 goto finalize; 2677 2678 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2679 bpf_prog_has_kfunc_call(fp)) 2680 jit_needed = true; 2681 2682 if (!bpf_prog_select_interpreter(fp)) 2683 jit_needed = true; 2684 2685 /* eBPF JITs can rewrite the program in case constant 2686 * blinding is active. However, in case of error during 2687 * blinding, bpf_int_jit_compile() must always return a 2688 * valid program, which in this case would simply not 2689 * be JITed, but falls back to the interpreter. 2690 */ 2691 if (!bpf_prog_is_offloaded(fp->aux)) { 2692 *err = bpf_prog_alloc_jited_linfo(fp); 2693 if (*err) 2694 return fp; 2695 2696 fp = bpf_prog_jit_compile(env, fp); 2697 bpf_prog_jit_attempt_done(fp); 2698 if (!fp->jited && jit_needed) { 2699 *err = -ENOTSUPP; 2700 return fp; 2701 } 2702 } else { 2703 *err = bpf_prog_offload_compile(fp); 2704 if (*err) 2705 return fp; 2706 } 2707 2708 finalize: 2709 *err = bpf_prog_lock_ro(fp); 2710 if (*err) 2711 return fp; 2712 2713 /* The tail call compatibility check can only be done at 2714 * this late stage as we need to determine, if we deal 2715 * with JITed or non JITed program concatenations and not 2716 * all eBPF JITs might immediately support all features. 2717 */ 2718 *err = bpf_check_tail_call(fp); 2719 2720 return fp; 2721 } 2722 2723 /** 2724 * bpf_prog_select_runtime - select exec runtime for BPF program 2725 * @fp: bpf_prog populated with BPF program 2726 * @err: pointer to error variable 2727 * 2728 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2729 * The BPF program will be executed via bpf_prog_run() function. 2730 * 2731 * Return: the &fp argument along with &err set to 0 for success or 2732 * a negative errno code on failure 2733 */ 2734 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2735 { 2736 return __bpf_prog_select_runtime(NULL, fp, err); 2737 } 2738 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2739 2740 static unsigned int __bpf_prog_ret1(const void *ctx, 2741 const struct bpf_insn *insn) 2742 { 2743 return 1; 2744 } 2745 2746 static struct bpf_prog_dummy { 2747 struct bpf_prog prog; 2748 } dummy_bpf_prog = { 2749 .prog = { 2750 .bpf_func = __bpf_prog_ret1, 2751 }, 2752 }; 2753 2754 struct bpf_prog_array bpf_empty_prog_array = { 2755 .items = { 2756 { .prog = NULL }, 2757 }, 2758 }; 2759 EXPORT_SYMBOL(bpf_empty_prog_array); 2760 2761 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2762 { 2763 struct bpf_prog_array *p; 2764 2765 if (prog_cnt) 2766 p = kzalloc_flex(*p, items, prog_cnt + 1, flags); 2767 else 2768 p = &bpf_empty_prog_array; 2769 2770 return p; 2771 } 2772 2773 void bpf_prog_array_free(struct bpf_prog_array *progs) 2774 { 2775 if (!progs || progs == &bpf_empty_prog_array) 2776 return; 2777 kfree_rcu(progs, rcu); 2778 } 2779 2780 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2781 { 2782 struct bpf_prog_array *progs; 2783 2784 /* 2785 * RCU Tasks Trace grace period implies RCU grace period, there is no 2786 * need to call kfree_rcu(), just call kfree() directly. 2787 */ 2788 progs = container_of(rcu, struct bpf_prog_array, rcu); 2789 kfree(progs); 2790 } 2791 2792 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2793 { 2794 if (!progs || progs == &bpf_empty_prog_array) 2795 return; 2796 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2797 } 2798 2799 int bpf_prog_array_length(struct bpf_prog_array *array) 2800 { 2801 struct bpf_prog_array_item *item; 2802 u32 cnt = 0; 2803 2804 for (item = array->items; item->prog; item++) 2805 if (item->prog != &dummy_bpf_prog.prog) 2806 cnt++; 2807 return cnt; 2808 } 2809 2810 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2811 { 2812 struct bpf_prog_array_item *item; 2813 2814 for (item = array->items; item->prog; item++) 2815 if (item->prog != &dummy_bpf_prog.prog) 2816 return false; 2817 return true; 2818 } 2819 2820 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2821 u32 *prog_ids, 2822 u32 request_cnt) 2823 { 2824 struct bpf_prog_array_item *item; 2825 int i = 0; 2826 2827 for (item = array->items; item->prog; item++) { 2828 if (item->prog == &dummy_bpf_prog.prog) 2829 continue; 2830 prog_ids[i] = item->prog->aux->id; 2831 if (++i == request_cnt) { 2832 item++; 2833 break; 2834 } 2835 } 2836 2837 return !!(item->prog); 2838 } 2839 2840 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2841 __u32 __user *prog_ids, u32 cnt) 2842 { 2843 unsigned long err = 0; 2844 bool nospc; 2845 u32 *ids; 2846 2847 /* users of this function are doing: 2848 * cnt = bpf_prog_array_length(); 2849 * if (cnt > 0) 2850 * bpf_prog_array_copy_to_user(..., cnt); 2851 * so below kcalloc doesn't need extra cnt > 0 check. 2852 */ 2853 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2854 if (!ids) 2855 return -ENOMEM; 2856 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2857 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2858 kfree(ids); 2859 if (err) 2860 return -EFAULT; 2861 if (nospc) 2862 return -ENOSPC; 2863 return 0; 2864 } 2865 2866 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2867 struct bpf_prog *old_prog) 2868 { 2869 struct bpf_prog_array_item *item; 2870 2871 for (item = array->items; item->prog; item++) 2872 if (item->prog == old_prog) { 2873 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2874 break; 2875 } 2876 } 2877 2878 /** 2879 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2880 * index into the program array with 2881 * a dummy no-op program. 2882 * @array: a bpf_prog_array 2883 * @index: the index of the program to replace 2884 * 2885 * Skips over dummy programs, by not counting them, when calculating 2886 * the position of the program to replace. 2887 * 2888 * Return: 2889 * * 0 - Success 2890 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2891 * * -ENOENT - Index out of range 2892 */ 2893 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2894 { 2895 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2896 } 2897 2898 /** 2899 * bpf_prog_array_update_at() - Updates the program at the given index 2900 * into the program array. 2901 * @array: a bpf_prog_array 2902 * @index: the index of the program to update 2903 * @prog: the program to insert into the array 2904 * 2905 * Skips over dummy programs, by not counting them, when calculating 2906 * the position of the program to update. 2907 * 2908 * Return: 2909 * * 0 - Success 2910 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2911 * * -ENOENT - Index out of range 2912 */ 2913 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2914 struct bpf_prog *prog) 2915 { 2916 struct bpf_prog_array_item *item; 2917 2918 if (unlikely(index < 0)) 2919 return -EINVAL; 2920 2921 for (item = array->items; item->prog; item++) { 2922 if (item->prog == &dummy_bpf_prog.prog) 2923 continue; 2924 if (!index) { 2925 WRITE_ONCE(item->prog, prog); 2926 return 0; 2927 } 2928 index--; 2929 } 2930 return -ENOENT; 2931 } 2932 2933 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2934 struct bpf_prog *exclude_prog, 2935 struct bpf_prog *include_prog, 2936 u64 bpf_cookie, 2937 struct bpf_prog_array **new_array) 2938 { 2939 int new_prog_cnt, carry_prog_cnt = 0; 2940 struct bpf_prog_array_item *existing, *new; 2941 struct bpf_prog_array *array; 2942 bool found_exclude = false; 2943 2944 /* Figure out how many existing progs we need to carry over to 2945 * the new array. 2946 */ 2947 if (old_array) { 2948 existing = old_array->items; 2949 for (; existing->prog; existing++) { 2950 if (existing->prog == exclude_prog) { 2951 found_exclude = true; 2952 continue; 2953 } 2954 if (existing->prog != &dummy_bpf_prog.prog) 2955 carry_prog_cnt++; 2956 if (existing->prog == include_prog) 2957 return -EEXIST; 2958 } 2959 } 2960 2961 if (exclude_prog && !found_exclude) 2962 return -ENOENT; 2963 2964 /* How many progs (not NULL) will be in the new array? */ 2965 new_prog_cnt = carry_prog_cnt; 2966 if (include_prog) 2967 new_prog_cnt += 1; 2968 2969 /* Do we have any prog (not NULL) in the new array? */ 2970 if (!new_prog_cnt) { 2971 *new_array = NULL; 2972 return 0; 2973 } 2974 2975 /* +1 as the end of prog_array is marked with NULL */ 2976 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2977 if (!array) 2978 return -ENOMEM; 2979 new = array->items; 2980 2981 /* Fill in the new prog array */ 2982 if (carry_prog_cnt) { 2983 existing = old_array->items; 2984 for (; existing->prog; existing++) { 2985 if (existing->prog == exclude_prog || 2986 existing->prog == &dummy_bpf_prog.prog) 2987 continue; 2988 2989 new->prog = existing->prog; 2990 new->bpf_cookie = existing->bpf_cookie; 2991 new++; 2992 } 2993 } 2994 if (include_prog) { 2995 new->prog = include_prog; 2996 new->bpf_cookie = bpf_cookie; 2997 new++; 2998 } 2999 new->prog = NULL; 3000 *new_array = array; 3001 return 0; 3002 } 3003 3004 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 3005 u32 *prog_ids, u32 request_cnt, 3006 u32 *prog_cnt) 3007 { 3008 u32 cnt = 0; 3009 3010 if (array) 3011 cnt = bpf_prog_array_length(array); 3012 3013 *prog_cnt = cnt; 3014 3015 /* return early if user requested only program count or nothing to copy */ 3016 if (!request_cnt || !cnt) 3017 return 0; 3018 3019 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 3020 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 3021 : 0; 3022 } 3023 3024 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 3025 struct bpf_map **used_maps, u32 len) 3026 { 3027 struct bpf_map *map; 3028 bool sleepable; 3029 u32 i; 3030 3031 sleepable = aux->prog->sleepable; 3032 for (i = 0; i < len; i++) { 3033 map = used_maps[i]; 3034 if (map->ops->map_poke_untrack) 3035 map->ops->map_poke_untrack(map, aux); 3036 if (sleepable) 3037 atomic64_dec(&map->sleepable_refcnt); 3038 bpf_map_put(map); 3039 } 3040 } 3041 3042 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 3043 { 3044 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 3045 kfree(aux->used_maps); 3046 } 3047 3048 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 3049 { 3050 #ifdef CONFIG_BPF_SYSCALL 3051 struct btf_mod_pair *btf_mod; 3052 u32 i; 3053 3054 for (i = 0; i < len; i++) { 3055 btf_mod = &used_btfs[i]; 3056 if (btf_mod->module) 3057 module_put(btf_mod->module); 3058 btf_put(btf_mod->btf); 3059 } 3060 #endif 3061 } 3062 3063 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 3064 { 3065 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 3066 kfree(aux->used_btfs); 3067 } 3068 3069 static void bpf_prog_free_deferred(struct work_struct *work) 3070 { 3071 struct bpf_prog_aux *aux; 3072 int i; 3073 3074 aux = container_of(work, struct bpf_prog_aux, work); 3075 #ifdef CONFIG_BPF_SYSCALL 3076 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 3077 bpf_prog_stream_free(aux->prog); 3078 #endif 3079 #ifdef CONFIG_CGROUP_BPF 3080 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 3081 bpf_cgroup_atype_put(aux->cgroup_atype); 3082 #endif 3083 bpf_free_used_maps(aux); 3084 bpf_free_used_btfs(aux); 3085 bpf_prog_disassoc_struct_ops(aux->prog); 3086 if (bpf_prog_is_dev_bound(aux)) 3087 bpf_prog_dev_bound_destroy(aux->prog); 3088 #ifdef CONFIG_PERF_EVENTS 3089 if (aux->prog->has_callchain_buf) 3090 put_callchain_buffers(); 3091 #endif 3092 if (aux->dst_trampoline) 3093 bpf_trampoline_put(aux->dst_trampoline); 3094 for (i = 0; i < aux->real_func_cnt; i++) { 3095 /* We can just unlink the subprog poke descriptor table as 3096 * it was originally linked to the main program and is also 3097 * released along with it. 3098 */ 3099 aux->func[i]->aux->poke_tab = NULL; 3100 bpf_jit_free(aux->func[i]); 3101 } 3102 if (aux->real_func_cnt) { 3103 kfree(aux->func); 3104 bpf_prog_unlock_free(aux->prog); 3105 } else { 3106 bpf_jit_free(aux->prog); 3107 } 3108 } 3109 3110 void bpf_prog_free(struct bpf_prog *fp) 3111 { 3112 struct bpf_prog_aux *aux = fp->aux; 3113 3114 if (aux->dst_prog) 3115 bpf_prog_put(aux->dst_prog); 3116 bpf_token_put(aux->token); 3117 INIT_WORK(&aux->work, bpf_prog_free_deferred); 3118 schedule_work(&aux->work); 3119 } 3120 EXPORT_SYMBOL_GPL(bpf_prog_free); 3121 3122 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 3123 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 3124 3125 void bpf_user_rnd_init_once(void) 3126 { 3127 prandom_init_once(&bpf_user_rnd_state); 3128 } 3129 3130 BPF_CALL_0(bpf_user_rnd_u32) 3131 { 3132 /* Should someone ever have the rather unwise idea to use some 3133 * of the registers passed into this function, then note that 3134 * this function is called from native eBPF and classic-to-eBPF 3135 * transformations. Register assignments from both sides are 3136 * different, f.e. classic always sets fn(ctx, A, X) here. 3137 */ 3138 struct rnd_state *state; 3139 u32 res; 3140 3141 state = &get_cpu_var(bpf_user_rnd_state); 3142 res = prandom_u32_state(state); 3143 put_cpu_var(bpf_user_rnd_state); 3144 3145 return res; 3146 } 3147 3148 BPF_CALL_0(bpf_get_raw_cpu_id) 3149 { 3150 return raw_smp_processor_id(); 3151 } 3152 3153 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 3154 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 3155 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 3156 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 3157 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 3158 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 3159 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 3160 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 3161 const struct bpf_func_proto bpf_spin_lock_proto __weak; 3162 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 3163 const struct bpf_func_proto bpf_jiffies64_proto __weak; 3164 3165 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3166 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3167 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3168 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3169 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3170 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3171 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3172 3173 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3174 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3175 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3176 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3177 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3178 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3179 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3180 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3181 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3182 const struct bpf_func_proto bpf_set_retval_proto __weak; 3183 const struct bpf_func_proto bpf_get_retval_proto __weak; 3184 3185 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3186 { 3187 return NULL; 3188 } 3189 3190 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3191 { 3192 return NULL; 3193 } 3194 3195 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3196 { 3197 return NULL; 3198 } 3199 3200 u64 __weak 3201 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3202 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3203 { 3204 return -ENOTSUPP; 3205 } 3206 EXPORT_SYMBOL_GPL(bpf_event_output); 3207 3208 /* Always built-in helper functions. */ 3209 const struct bpf_func_proto bpf_tail_call_proto = { 3210 /* func is unused for tail_call, we set it to pass the 3211 * get_helper_proto check 3212 */ 3213 .func = BPF_PTR_POISON, 3214 .gpl_only = false, 3215 .ret_type = RET_VOID, 3216 .arg1_type = ARG_PTR_TO_CTX, 3217 .arg2_type = ARG_CONST_MAP_PTR, 3218 .arg3_type = ARG_ANYTHING, 3219 }; 3220 3221 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3222 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3223 * eBPF and implicitly also cBPF can get JITed! 3224 */ 3225 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 3226 { 3227 return prog; 3228 } 3229 3230 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3231 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3232 */ 3233 void __weak bpf_jit_compile(struct bpf_prog *prog) 3234 { 3235 } 3236 3237 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3238 { 3239 return false; 3240 } 3241 3242 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3243 * analysis code and wants explicit zero extension inserted by verifier. 3244 * Otherwise, return FALSE. 3245 * 3246 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3247 * you don't override this. JITs that don't want these extra insns can detect 3248 * them using insn_is_zext. 3249 */ 3250 bool __weak bpf_jit_needs_zext(void) 3251 { 3252 return false; 3253 } 3254 3255 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3256 * all archs. The value returned must not change at runtime as there is 3257 * currently no support for reloading programs that were loaded without 3258 * mitigations. 3259 */ 3260 bool __weak bpf_jit_bypass_spec_v1(void) 3261 { 3262 return false; 3263 } 3264 3265 bool __weak bpf_jit_bypass_spec_v4(void) 3266 { 3267 return false; 3268 } 3269 3270 /* Return true if the JIT inlines the call to the helper corresponding to 3271 * the imm. 3272 * 3273 * The verifier will not patch the insn->imm for the call to the helper if 3274 * this returns true. 3275 */ 3276 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3277 { 3278 return false; 3279 } 3280 3281 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3282 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3283 { 3284 return false; 3285 } 3286 3287 bool __weak bpf_jit_supports_percpu_insn(void) 3288 { 3289 return false; 3290 } 3291 3292 bool __weak bpf_jit_supports_kfunc_call(void) 3293 { 3294 return false; 3295 } 3296 3297 bool __weak bpf_jit_supports_stack_args(void) 3298 { 3299 return false; 3300 } 3301 3302 bool __weak bpf_jit_supports_far_kfunc_call(void) 3303 { 3304 return false; 3305 } 3306 3307 bool __weak bpf_jit_supports_arena(void) 3308 { 3309 return false; 3310 } 3311 3312 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3313 { 3314 return false; 3315 } 3316 3317 bool __weak bpf_jit_supports_fsession(void) 3318 { 3319 return false; 3320 } 3321 3322 u64 __weak bpf_arch_uaddress_limit(void) 3323 { 3324 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3325 return TASK_SIZE; 3326 #else 3327 return 0; 3328 #endif 3329 } 3330 3331 /* Return TRUE if the JIT backend satisfies the following two conditions: 3332 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3333 * 2) Under the specific arch, the implementation of xchg() is the same 3334 * as atomic_xchg() on pointer-sized words. 3335 */ 3336 bool __weak bpf_jit_supports_ptr_xchg(void) 3337 { 3338 return false; 3339 } 3340 3341 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3342 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3343 */ 3344 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3345 int len) 3346 { 3347 return -EFAULT; 3348 } 3349 3350 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3351 enum bpf_text_poke_type new_t, void *old_addr, 3352 void *new_addr) 3353 { 3354 return -ENOTSUPP; 3355 } 3356 3357 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3358 { 3359 return ERR_PTR(-ENOTSUPP); 3360 } 3361 3362 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3363 { 3364 return -ENOTSUPP; 3365 } 3366 3367 bool __weak bpf_jit_supports_exceptions(void) 3368 { 3369 return false; 3370 } 3371 3372 bool __weak bpf_jit_supports_private_stack(void) 3373 { 3374 return false; 3375 } 3376 3377 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3378 { 3379 } 3380 3381 bool __weak bpf_jit_supports_timed_may_goto(void) 3382 { 3383 return false; 3384 } 3385 3386 u64 __weak arch_bpf_timed_may_goto(void) 3387 { 3388 return 0; 3389 } 3390 3391 static noinline void bpf_prog_report_may_goto_violation(void) 3392 { 3393 #ifdef CONFIG_BPF_SYSCALL 3394 struct bpf_stream_stage ss; 3395 struct bpf_prog *prog; 3396 3397 prog = bpf_prog_find_from_stack(); 3398 if (!prog) 3399 return; 3400 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3401 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3402 bpf_stream_dump_stack(ss); 3403 })); 3404 #endif 3405 } 3406 3407 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3408 { 3409 u64 time = ktime_get_mono_fast_ns(); 3410 3411 /* Populate the timestamp for this stack frame, and refresh count. */ 3412 if (!p->timestamp) { 3413 p->timestamp = time; 3414 return BPF_MAX_TIMED_LOOPS; 3415 } 3416 /* Check if we've exhausted our time slice, and zero count. */ 3417 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3418 bpf_prog_report_may_goto_violation(); 3419 return 0; 3420 } 3421 /* Refresh the count for the stack frame. */ 3422 return BPF_MAX_TIMED_LOOPS; 3423 } 3424 3425 /* for configs without MMU or 32-bit */ 3426 __weak const struct bpf_map_ops arena_map_ops; 3427 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3428 { 3429 return 0; 3430 } 3431 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3432 { 3433 return 0; 3434 } 3435 3436 #ifdef CONFIG_BPF_SYSCALL 3437 __weak bool bpf_arena_handle_page_fault(unsigned long addr, bool is_write, 3438 unsigned long fault_ip) 3439 { 3440 return false; 3441 } 3442 3443 static int __init bpf_global_ma_init(void) 3444 { 3445 int ret; 3446 3447 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3448 bpf_global_ma_set = !ret; 3449 return ret; 3450 } 3451 late_initcall(bpf_global_ma_init); 3452 #endif 3453 3454 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3455 EXPORT_SYMBOL(bpf_stats_enabled_key); 3456 3457 /* All definitions of tracepoints related to BPF. */ 3458 #define CREATE_TRACE_POINTS 3459 #include <linux/bpf_trace.h> 3460 3461 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3462 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3463 3464 #ifdef CONFIG_BPF_SYSCALL 3465 3466 void bpf_get_linfo_file_line(struct btf *btf, const struct bpf_line_info *linfo, 3467 const char **filep, const char **linep, int *nump) 3468 { 3469 /* Get base component of the file path. */ 3470 if (filep) { 3471 *filep = btf_name_by_offset(btf, linfo->file_name_off); 3472 *filep = kbasename(*filep); 3473 } 3474 3475 /* Obtain the source line, and strip whitespace in prefix. */ 3476 if (linep) { 3477 *linep = btf_name_by_offset(btf, linfo->line_off); 3478 while (isspace(**linep)) 3479 *linep += 1; 3480 } 3481 3482 if (nump) 3483 *nump = BPF_LINE_INFO_LINE_NUM(linfo->line_col); 3484 } 3485 3486 const struct bpf_line_info *bpf_find_linfo(const struct bpf_prog *prog, u32 insn_off) 3487 { 3488 const struct bpf_line_info *linfo; 3489 u32 nr_linfo; 3490 int l, r, m; 3491 3492 nr_linfo = prog->aux->nr_linfo; 3493 if (!nr_linfo || insn_off >= prog->len) 3494 return NULL; 3495 3496 linfo = prog->aux->linfo; 3497 /* Loop invariant: linfo[l].insn_off <= insns_off. 3498 * linfo[0].insn_off == 0 which always satisfies above condition. 3499 * Binary search is searching for rightmost linfo entry that satisfies 3500 * the above invariant, giving us the desired record that covers given 3501 * instruction offset. 3502 */ 3503 l = 0; 3504 r = nr_linfo - 1; 3505 while (l < r) { 3506 /* (r - l + 1) / 2 means we break a tie to the right, so if: 3507 * l=1, r=2, linfo[l].insn_off <= insn_off, linfo[r].insn_off > insn_off, 3508 * then m=2, we see that linfo[m].insn_off > insn_off, and so 3509 * r becomes 1 and we exit the loop with correct l==1. 3510 * If the tie was broken to the left, m=1 would end us up in 3511 * an endless loop where l and m stay at 1 and r stays at 2. 3512 */ 3513 m = l + (r - l + 1) / 2; 3514 if (linfo[m].insn_off <= insn_off) 3515 l = m; 3516 else 3517 r = m - 1; 3518 } 3519 3520 return &linfo[l]; 3521 } 3522 3523 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3524 const char **linep, int *nump) 3525 { 3526 int idx = -1, insn_start, insn_end, len; 3527 struct bpf_line_info *linfo; 3528 void **jited_linfo; 3529 struct btf *btf; 3530 int nr_linfo; 3531 3532 btf = prog->aux->btf; 3533 linfo = prog->aux->linfo; 3534 jited_linfo = prog->aux->jited_linfo; 3535 3536 if (!btf || !linfo || !jited_linfo) 3537 return -EINVAL; 3538 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3539 3540 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3541 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3542 3543 insn_start = linfo[0].insn_off; 3544 insn_end = insn_start + len; 3545 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3546 3547 for (int i = 0; i < nr_linfo && 3548 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3549 if (jited_linfo[i] >= (void *)ip) 3550 break; 3551 idx = i; 3552 } 3553 3554 if (idx == -1) 3555 return -ENOENT; 3556 3557 bpf_get_linfo_file_line(btf, &linfo[idx], filep, linep, nump); 3558 return 0; 3559 } 3560 3561 struct walk_stack_ctx { 3562 struct bpf_prog *prog; 3563 }; 3564 3565 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3566 { 3567 struct walk_stack_ctx *ctxp = cookie; 3568 struct bpf_prog *prog; 3569 3570 /* 3571 * The RCU read lock is held to safely traverse the latch tree, but we 3572 * don't need its protection when accessing the prog, since it has an 3573 * active stack frame on the current stack trace, and won't disappear. 3574 */ 3575 rcu_read_lock(); 3576 prog = bpf_prog_ksym_find(ip); 3577 rcu_read_unlock(); 3578 if (!prog) 3579 return true; 3580 /* Make sure we return the main prog if we found a subprog */ 3581 ctxp->prog = prog->aux->main_prog_aux->prog; 3582 return false; 3583 } 3584 3585 struct bpf_prog *bpf_prog_find_from_stack(void) 3586 { 3587 struct walk_stack_ctx ctx = {}; 3588 3589 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3590 return ctx.prog; 3591 } 3592 3593 #endif 3594