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/vmalloc.h> 24 #include <linux/random.h> 25 #include <linux/moduleloader.h> 26 #include <linux/bpf.h> 27 #include <linux/btf.h> 28 #include <linux/objtool.h> 29 #include <linux/rbtree_latch.h> 30 #include <linux/kallsyms.h> 31 #include <linux/rcupdate.h> 32 #include <linux/perf_event.h> 33 #include <linux/extable.h> 34 #include <linux/log2.h> 35 #include <asm/unaligned.h> 36 37 /* Registers */ 38 #define BPF_R0 regs[BPF_REG_0] 39 #define BPF_R1 regs[BPF_REG_1] 40 #define BPF_R2 regs[BPF_REG_2] 41 #define BPF_R3 regs[BPF_REG_3] 42 #define BPF_R4 regs[BPF_REG_4] 43 #define BPF_R5 regs[BPF_REG_5] 44 #define BPF_R6 regs[BPF_REG_6] 45 #define BPF_R7 regs[BPF_REG_7] 46 #define BPF_R8 regs[BPF_REG_8] 47 #define BPF_R9 regs[BPF_REG_9] 48 #define BPF_R10 regs[BPF_REG_10] 49 50 /* Named registers */ 51 #define DST regs[insn->dst_reg] 52 #define SRC regs[insn->src_reg] 53 #define FP regs[BPF_REG_FP] 54 #define AX regs[BPF_REG_AX] 55 #define ARG1 regs[BPF_REG_ARG1] 56 #define CTX regs[BPF_REG_CTX] 57 #define IMM insn->imm 58 59 /* No hurry in this branch 60 * 61 * Exported for the bpf jit load helper. 62 */ 63 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 64 { 65 u8 *ptr = NULL; 66 67 if (k >= SKF_NET_OFF) 68 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 69 else if (k >= SKF_LL_OFF) 70 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 71 72 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 73 return ptr; 74 75 return NULL; 76 } 77 78 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 79 { 80 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 81 struct bpf_prog_aux *aux; 82 struct bpf_prog *fp; 83 84 size = round_up(size, PAGE_SIZE); 85 fp = __vmalloc(size, gfp_flags); 86 if (fp == NULL) 87 return NULL; 88 89 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags); 90 if (aux == NULL) { 91 vfree(fp); 92 return NULL; 93 } 94 95 fp->pages = size / PAGE_SIZE; 96 fp->aux = aux; 97 fp->aux->prog = fp; 98 fp->jit_requested = ebpf_jit_enabled(); 99 100 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 101 mutex_init(&fp->aux->used_maps_mutex); 102 mutex_init(&fp->aux->dst_mutex); 103 104 return fp; 105 } 106 107 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 108 { 109 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 110 struct bpf_prog *prog; 111 int cpu; 112 113 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 114 if (!prog) 115 return NULL; 116 117 prog->aux->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 118 if (!prog->aux->stats) { 119 kfree(prog->aux); 120 vfree(prog); 121 return NULL; 122 } 123 124 for_each_possible_cpu(cpu) { 125 struct bpf_prog_stats *pstats; 126 127 pstats = per_cpu_ptr(prog->aux->stats, cpu); 128 u64_stats_init(&pstats->syncp); 129 } 130 return prog; 131 } 132 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 133 134 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 135 { 136 if (!prog->aux->nr_linfo || !prog->jit_requested) 137 return 0; 138 139 prog->aux->jited_linfo = kcalloc(prog->aux->nr_linfo, 140 sizeof(*prog->aux->jited_linfo), 141 GFP_KERNEL_ACCOUNT | __GFP_NOWARN); 142 if (!prog->aux->jited_linfo) 143 return -ENOMEM; 144 145 return 0; 146 } 147 148 void bpf_prog_free_jited_linfo(struct bpf_prog *prog) 149 { 150 kfree(prog->aux->jited_linfo); 151 prog->aux->jited_linfo = NULL; 152 } 153 154 void bpf_prog_free_unused_jited_linfo(struct bpf_prog *prog) 155 { 156 if (prog->aux->jited_linfo && !prog->aux->jited_linfo[0]) 157 bpf_prog_free_jited_linfo(prog); 158 } 159 160 /* The jit engine is responsible to provide an array 161 * for insn_off to the jited_off mapping (insn_to_jit_off). 162 * 163 * The idx to this array is the insn_off. Hence, the insn_off 164 * here is relative to the prog itself instead of the main prog. 165 * This array has one entry for each xlated bpf insn. 166 * 167 * jited_off is the byte off to the last byte of the jited insn. 168 * 169 * Hence, with 170 * insn_start: 171 * The first bpf insn off of the prog. The insn off 172 * here is relative to the main prog. 173 * e.g. if prog is a subprog, insn_start > 0 174 * linfo_idx: 175 * The prog's idx to prog->aux->linfo and jited_linfo 176 * 177 * jited_linfo[linfo_idx] = prog->bpf_func 178 * 179 * For i > linfo_idx, 180 * 181 * jited_linfo[i] = prog->bpf_func + 182 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 183 */ 184 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 185 const u32 *insn_to_jit_off) 186 { 187 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 188 const struct bpf_line_info *linfo; 189 void **jited_linfo; 190 191 if (!prog->aux->jited_linfo) 192 /* Userspace did not provide linfo */ 193 return; 194 195 linfo_idx = prog->aux->linfo_idx; 196 linfo = &prog->aux->linfo[linfo_idx]; 197 insn_start = linfo[0].insn_off; 198 insn_end = insn_start + prog->len; 199 200 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 201 jited_linfo[0] = prog->bpf_func; 202 203 nr_linfo = prog->aux->nr_linfo - linfo_idx; 204 205 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 206 /* The verifier ensures that linfo[i].insn_off is 207 * strictly increasing 208 */ 209 jited_linfo[i] = prog->bpf_func + 210 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 211 } 212 213 void bpf_prog_free_linfo(struct bpf_prog *prog) 214 { 215 bpf_prog_free_jited_linfo(prog); 216 kvfree(prog->aux->linfo); 217 } 218 219 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 220 gfp_t gfp_extra_flags) 221 { 222 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 223 struct bpf_prog *fp; 224 u32 pages; 225 226 size = round_up(size, PAGE_SIZE); 227 pages = size / PAGE_SIZE; 228 if (pages <= fp_old->pages) 229 return fp_old; 230 231 fp = __vmalloc(size, gfp_flags); 232 if (fp) { 233 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 234 fp->pages = pages; 235 fp->aux->prog = fp; 236 237 /* We keep fp->aux from fp_old around in the new 238 * reallocated structure. 239 */ 240 fp_old->aux = NULL; 241 __bpf_prog_free(fp_old); 242 } 243 244 return fp; 245 } 246 247 void __bpf_prog_free(struct bpf_prog *fp) 248 { 249 if (fp->aux) { 250 mutex_destroy(&fp->aux->used_maps_mutex); 251 mutex_destroy(&fp->aux->dst_mutex); 252 free_percpu(fp->aux->stats); 253 kfree(fp->aux->poke_tab); 254 kfree(fp->aux); 255 } 256 vfree(fp); 257 } 258 259 int bpf_prog_calc_tag(struct bpf_prog *fp) 260 { 261 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 262 u32 raw_size = bpf_prog_tag_scratch_size(fp); 263 u32 digest[SHA1_DIGEST_WORDS]; 264 u32 ws[SHA1_WORKSPACE_WORDS]; 265 u32 i, bsize, psize, blocks; 266 struct bpf_insn *dst; 267 bool was_ld_map; 268 u8 *raw, *todo; 269 __be32 *result; 270 __be64 *bits; 271 272 raw = vmalloc(raw_size); 273 if (!raw) 274 return -ENOMEM; 275 276 sha1_init(digest); 277 memset(ws, 0, sizeof(ws)); 278 279 /* We need to take out the map fd for the digest calculation 280 * since they are unstable from user space side. 281 */ 282 dst = (void *)raw; 283 for (i = 0, was_ld_map = false; i < fp->len; i++) { 284 dst[i] = fp->insnsi[i]; 285 if (!was_ld_map && 286 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 287 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 288 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 289 was_ld_map = true; 290 dst[i].imm = 0; 291 } else if (was_ld_map && 292 dst[i].code == 0 && 293 dst[i].dst_reg == 0 && 294 dst[i].src_reg == 0 && 295 dst[i].off == 0) { 296 was_ld_map = false; 297 dst[i].imm = 0; 298 } else { 299 was_ld_map = false; 300 } 301 } 302 303 psize = bpf_prog_insn_size(fp); 304 memset(&raw[psize], 0, raw_size - psize); 305 raw[psize++] = 0x80; 306 307 bsize = round_up(psize, SHA1_BLOCK_SIZE); 308 blocks = bsize / SHA1_BLOCK_SIZE; 309 todo = raw; 310 if (bsize - psize >= sizeof(__be64)) { 311 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 312 } else { 313 bits = (__be64 *)(todo + bsize + bits_offset); 314 blocks++; 315 } 316 *bits = cpu_to_be64((psize - 1) << 3); 317 318 while (blocks--) { 319 sha1_transform(digest, todo, ws); 320 todo += SHA1_BLOCK_SIZE; 321 } 322 323 result = (__force __be32 *)digest; 324 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 325 result[i] = cpu_to_be32(digest[i]); 326 memcpy(fp->tag, result, sizeof(fp->tag)); 327 328 vfree(raw); 329 return 0; 330 } 331 332 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 333 s32 end_new, s32 curr, const bool probe_pass) 334 { 335 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 336 s32 delta = end_new - end_old; 337 s64 imm = insn->imm; 338 339 if (curr < pos && curr + imm + 1 >= end_old) 340 imm += delta; 341 else if (curr >= end_new && curr + imm + 1 < end_new) 342 imm -= delta; 343 if (imm < imm_min || imm > imm_max) 344 return -ERANGE; 345 if (!probe_pass) 346 insn->imm = imm; 347 return 0; 348 } 349 350 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 351 s32 end_new, s32 curr, const bool probe_pass) 352 { 353 const s32 off_min = S16_MIN, off_max = S16_MAX; 354 s32 delta = end_new - end_old; 355 s32 off = insn->off; 356 357 if (curr < pos && curr + off + 1 >= end_old) 358 off += delta; 359 else if (curr >= end_new && curr + off + 1 < end_new) 360 off -= delta; 361 if (off < off_min || off > off_max) 362 return -ERANGE; 363 if (!probe_pass) 364 insn->off = off; 365 return 0; 366 } 367 368 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 369 s32 end_new, const bool probe_pass) 370 { 371 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 372 struct bpf_insn *insn = prog->insnsi; 373 int ret = 0; 374 375 for (i = 0; i < insn_cnt; i++, insn++) { 376 u8 code; 377 378 /* In the probing pass we still operate on the original, 379 * unpatched image in order to check overflows before we 380 * do any other adjustments. Therefore skip the patchlet. 381 */ 382 if (probe_pass && i == pos) { 383 i = end_new; 384 insn = prog->insnsi + end_old; 385 } 386 code = insn->code; 387 if ((BPF_CLASS(code) != BPF_JMP && 388 BPF_CLASS(code) != BPF_JMP32) || 389 BPF_OP(code) == BPF_EXIT) 390 continue; 391 /* Adjust offset of jmps if we cross patch boundaries. */ 392 if (BPF_OP(code) == BPF_CALL) { 393 if (insn->src_reg != BPF_PSEUDO_CALL) 394 continue; 395 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 396 end_new, i, probe_pass); 397 } else { 398 ret = bpf_adj_delta_to_off(insn, pos, end_old, 399 end_new, i, probe_pass); 400 } 401 if (ret) 402 break; 403 } 404 405 return ret; 406 } 407 408 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 409 { 410 struct bpf_line_info *linfo; 411 u32 i, nr_linfo; 412 413 nr_linfo = prog->aux->nr_linfo; 414 if (!nr_linfo || !delta) 415 return; 416 417 linfo = prog->aux->linfo; 418 419 for (i = 0; i < nr_linfo; i++) 420 if (off < linfo[i].insn_off) 421 break; 422 423 /* Push all off < linfo[i].insn_off by delta */ 424 for (; i < nr_linfo; i++) 425 linfo[i].insn_off += delta; 426 } 427 428 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 429 const struct bpf_insn *patch, u32 len) 430 { 431 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 432 const u32 cnt_max = S16_MAX; 433 struct bpf_prog *prog_adj; 434 int err; 435 436 /* Since our patchlet doesn't expand the image, we're done. */ 437 if (insn_delta == 0) { 438 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 439 return prog; 440 } 441 442 insn_adj_cnt = prog->len + insn_delta; 443 444 /* Reject anything that would potentially let the insn->off 445 * target overflow when we have excessive program expansions. 446 * We need to probe here before we do any reallocation where 447 * we afterwards may not fail anymore. 448 */ 449 if (insn_adj_cnt > cnt_max && 450 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 451 return ERR_PTR(err); 452 453 /* Several new instructions need to be inserted. Make room 454 * for them. Likely, there's no need for a new allocation as 455 * last page could have large enough tailroom. 456 */ 457 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 458 GFP_USER); 459 if (!prog_adj) 460 return ERR_PTR(-ENOMEM); 461 462 prog_adj->len = insn_adj_cnt; 463 464 /* Patching happens in 3 steps: 465 * 466 * 1) Move over tail of insnsi from next instruction onwards, 467 * so we can patch the single target insn with one or more 468 * new ones (patching is always from 1 to n insns, n > 0). 469 * 2) Inject new instructions at the target location. 470 * 3) Adjust branch offsets if necessary. 471 */ 472 insn_rest = insn_adj_cnt - off - len; 473 474 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 475 sizeof(*patch) * insn_rest); 476 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 477 478 /* We are guaranteed to not fail at this point, otherwise 479 * the ship has sailed to reverse to the original state. An 480 * overflow cannot happen at this point. 481 */ 482 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 483 484 bpf_adj_linfo(prog_adj, off, insn_delta); 485 486 return prog_adj; 487 } 488 489 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 490 { 491 /* Branch offsets can't overflow when program is shrinking, no need 492 * to call bpf_adj_branches(..., true) here 493 */ 494 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 495 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 496 prog->len -= cnt; 497 498 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 499 } 500 501 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 502 { 503 int i; 504 505 for (i = 0; i < fp->aux->func_cnt; i++) 506 bpf_prog_kallsyms_del(fp->aux->func[i]); 507 } 508 509 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 510 { 511 bpf_prog_kallsyms_del_subprogs(fp); 512 bpf_prog_kallsyms_del(fp); 513 } 514 515 #ifdef CONFIG_BPF_JIT 516 /* All BPF JIT sysctl knobs here. */ 517 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 518 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 519 int bpf_jit_harden __read_mostly; 520 long bpf_jit_limit __read_mostly; 521 522 static void 523 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 524 { 525 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 526 unsigned long addr = (unsigned long)hdr; 527 528 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 529 530 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 531 prog->aux->ksym.end = addr + hdr->pages * PAGE_SIZE; 532 } 533 534 static void 535 bpf_prog_ksym_set_name(struct bpf_prog *prog) 536 { 537 char *sym = prog->aux->ksym.name; 538 const char *end = sym + KSYM_NAME_LEN; 539 const struct btf_type *type; 540 const char *func_name; 541 542 BUILD_BUG_ON(sizeof("bpf_prog_") + 543 sizeof(prog->tag) * 2 + 544 /* name has been null terminated. 545 * We should need +1 for the '_' preceding 546 * the name. However, the null character 547 * is double counted between the name and the 548 * sizeof("bpf_prog_") above, so we omit 549 * the +1 here. 550 */ 551 sizeof(prog->aux->name) > KSYM_NAME_LEN); 552 553 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 554 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 555 556 /* prog->aux->name will be ignored if full btf name is available */ 557 if (prog->aux->func_info_cnt) { 558 type = btf_type_by_id(prog->aux->btf, 559 prog->aux->func_info[prog->aux->func_idx].type_id); 560 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 561 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 562 return; 563 } 564 565 if (prog->aux->name[0]) 566 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 567 else 568 *sym = 0; 569 } 570 571 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 572 { 573 return container_of(n, struct bpf_ksym, tnode)->start; 574 } 575 576 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 577 struct latch_tree_node *b) 578 { 579 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 580 } 581 582 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 583 { 584 unsigned long val = (unsigned long)key; 585 const struct bpf_ksym *ksym; 586 587 ksym = container_of(n, struct bpf_ksym, tnode); 588 589 if (val < ksym->start) 590 return -1; 591 if (val >= ksym->end) 592 return 1; 593 594 return 0; 595 } 596 597 static const struct latch_tree_ops bpf_tree_ops = { 598 .less = bpf_tree_less, 599 .comp = bpf_tree_comp, 600 }; 601 602 static DEFINE_SPINLOCK(bpf_lock); 603 static LIST_HEAD(bpf_kallsyms); 604 static struct latch_tree_root bpf_tree __cacheline_aligned; 605 606 void bpf_ksym_add(struct bpf_ksym *ksym) 607 { 608 spin_lock_bh(&bpf_lock); 609 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 610 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 611 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 612 spin_unlock_bh(&bpf_lock); 613 } 614 615 static void __bpf_ksym_del(struct bpf_ksym *ksym) 616 { 617 if (list_empty(&ksym->lnode)) 618 return; 619 620 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 621 list_del_rcu(&ksym->lnode); 622 } 623 624 void bpf_ksym_del(struct bpf_ksym *ksym) 625 { 626 spin_lock_bh(&bpf_lock); 627 __bpf_ksym_del(ksym); 628 spin_unlock_bh(&bpf_lock); 629 } 630 631 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 632 { 633 return fp->jited && !bpf_prog_was_classic(fp); 634 } 635 636 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 637 { 638 return list_empty(&fp->aux->ksym.lnode) || 639 fp->aux->ksym.lnode.prev == LIST_POISON2; 640 } 641 642 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 643 { 644 if (!bpf_prog_kallsyms_candidate(fp) || 645 !bpf_capable()) 646 return; 647 648 bpf_prog_ksym_set_addr(fp); 649 bpf_prog_ksym_set_name(fp); 650 fp->aux->ksym.prog = true; 651 652 bpf_ksym_add(&fp->aux->ksym); 653 } 654 655 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 656 { 657 if (!bpf_prog_kallsyms_candidate(fp)) 658 return; 659 660 bpf_ksym_del(&fp->aux->ksym); 661 } 662 663 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 664 { 665 struct latch_tree_node *n; 666 667 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 668 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 669 } 670 671 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 672 unsigned long *off, char *sym) 673 { 674 struct bpf_ksym *ksym; 675 char *ret = NULL; 676 677 rcu_read_lock(); 678 ksym = bpf_ksym_find(addr); 679 if (ksym) { 680 unsigned long symbol_start = ksym->start; 681 unsigned long symbol_end = ksym->end; 682 683 strncpy(sym, ksym->name, KSYM_NAME_LEN); 684 685 ret = sym; 686 if (size) 687 *size = symbol_end - symbol_start; 688 if (off) 689 *off = addr - symbol_start; 690 } 691 rcu_read_unlock(); 692 693 return ret; 694 } 695 696 bool is_bpf_text_address(unsigned long addr) 697 { 698 bool ret; 699 700 rcu_read_lock(); 701 ret = bpf_ksym_find(addr) != NULL; 702 rcu_read_unlock(); 703 704 return ret; 705 } 706 707 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 708 { 709 struct bpf_ksym *ksym = bpf_ksym_find(addr); 710 711 return ksym && ksym->prog ? 712 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 713 NULL; 714 } 715 716 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 717 { 718 const struct exception_table_entry *e = NULL; 719 struct bpf_prog *prog; 720 721 rcu_read_lock(); 722 prog = bpf_prog_ksym_find(addr); 723 if (!prog) 724 goto out; 725 if (!prog->aux->num_exentries) 726 goto out; 727 728 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 729 out: 730 rcu_read_unlock(); 731 return e; 732 } 733 734 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 735 char *sym) 736 { 737 struct bpf_ksym *ksym; 738 unsigned int it = 0; 739 int ret = -ERANGE; 740 741 if (!bpf_jit_kallsyms_enabled()) 742 return ret; 743 744 rcu_read_lock(); 745 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 746 if (it++ != symnum) 747 continue; 748 749 strncpy(sym, ksym->name, KSYM_NAME_LEN); 750 751 *value = ksym->start; 752 *type = BPF_SYM_ELF_TYPE; 753 754 ret = 0; 755 break; 756 } 757 rcu_read_unlock(); 758 759 return ret; 760 } 761 762 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 763 struct bpf_jit_poke_descriptor *poke) 764 { 765 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 766 static const u32 poke_tab_max = 1024; 767 u32 slot = prog->aux->size_poke_tab; 768 u32 size = slot + 1; 769 770 if (size > poke_tab_max) 771 return -ENOSPC; 772 if (poke->tailcall_target || poke->tailcall_target_stable || 773 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 774 return -EINVAL; 775 776 switch (poke->reason) { 777 case BPF_POKE_REASON_TAIL_CALL: 778 if (!poke->tail_call.map) 779 return -EINVAL; 780 break; 781 default: 782 return -EINVAL; 783 } 784 785 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL); 786 if (!tab) 787 return -ENOMEM; 788 789 memcpy(&tab[slot], poke, sizeof(*poke)); 790 prog->aux->size_poke_tab = size; 791 prog->aux->poke_tab = tab; 792 793 return slot; 794 } 795 796 static atomic_long_t bpf_jit_current; 797 798 /* Can be overridden by an arch's JIT compiler if it has a custom, 799 * dedicated BPF backend memory area, or if neither of the two 800 * below apply. 801 */ 802 u64 __weak bpf_jit_alloc_exec_limit(void) 803 { 804 #if defined(MODULES_VADDR) 805 return MODULES_END - MODULES_VADDR; 806 #else 807 return VMALLOC_END - VMALLOC_START; 808 #endif 809 } 810 811 static int __init bpf_jit_charge_init(void) 812 { 813 /* Only used as heuristic here to derive limit. */ 814 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2, 815 PAGE_SIZE), LONG_MAX); 816 return 0; 817 } 818 pure_initcall(bpf_jit_charge_init); 819 820 static int bpf_jit_charge_modmem(u32 pages) 821 { 822 if (atomic_long_add_return(pages, &bpf_jit_current) > 823 (bpf_jit_limit >> PAGE_SHIFT)) { 824 if (!capable(CAP_SYS_ADMIN)) { 825 atomic_long_sub(pages, &bpf_jit_current); 826 return -EPERM; 827 } 828 } 829 830 return 0; 831 } 832 833 static void bpf_jit_uncharge_modmem(u32 pages) 834 { 835 atomic_long_sub(pages, &bpf_jit_current); 836 } 837 838 void *__weak bpf_jit_alloc_exec(unsigned long size) 839 { 840 return module_alloc(size); 841 } 842 843 void __weak bpf_jit_free_exec(void *addr) 844 { 845 module_memfree(addr); 846 } 847 848 struct bpf_binary_header * 849 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 850 unsigned int alignment, 851 bpf_jit_fill_hole_t bpf_fill_ill_insns) 852 { 853 struct bpf_binary_header *hdr; 854 u32 size, hole, start, pages; 855 856 WARN_ON_ONCE(!is_power_of_2(alignment) || 857 alignment > BPF_IMAGE_ALIGNMENT); 858 859 /* Most of BPF filters are really small, but if some of them 860 * fill a page, allow at least 128 extra bytes to insert a 861 * random section of illegal instructions. 862 */ 863 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 864 pages = size / PAGE_SIZE; 865 866 if (bpf_jit_charge_modmem(pages)) 867 return NULL; 868 hdr = bpf_jit_alloc_exec(size); 869 if (!hdr) { 870 bpf_jit_uncharge_modmem(pages); 871 return NULL; 872 } 873 874 /* Fill space with illegal/arch-dep instructions. */ 875 bpf_fill_ill_insns(hdr, size); 876 877 hdr->pages = pages; 878 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 879 PAGE_SIZE - sizeof(*hdr)); 880 start = (get_random_int() % hole) & ~(alignment - 1); 881 882 /* Leave a random number of instructions before BPF code. */ 883 *image_ptr = &hdr->image[start]; 884 885 return hdr; 886 } 887 888 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 889 { 890 u32 pages = hdr->pages; 891 892 bpf_jit_free_exec(hdr); 893 bpf_jit_uncharge_modmem(pages); 894 } 895 896 /* This symbol is only overridden by archs that have different 897 * requirements than the usual eBPF JITs, f.e. when they only 898 * implement cBPF JIT, do not set images read-only, etc. 899 */ 900 void __weak bpf_jit_free(struct bpf_prog *fp) 901 { 902 if (fp->jited) { 903 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 904 905 bpf_jit_binary_free(hdr); 906 907 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 908 } 909 910 bpf_prog_unlock_free(fp); 911 } 912 913 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 914 const struct bpf_insn *insn, bool extra_pass, 915 u64 *func_addr, bool *func_addr_fixed) 916 { 917 s16 off = insn->off; 918 s32 imm = insn->imm; 919 u8 *addr; 920 921 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 922 if (!*func_addr_fixed) { 923 /* Place-holder address till the last pass has collected 924 * all addresses for JITed subprograms in which case we 925 * can pick them up from prog->aux. 926 */ 927 if (!extra_pass) 928 addr = NULL; 929 else if (prog->aux->func && 930 off >= 0 && off < prog->aux->func_cnt) 931 addr = (u8 *)prog->aux->func[off]->bpf_func; 932 else 933 return -EINVAL; 934 } else { 935 /* Address of a BPF helper call. Since part of the core 936 * kernel, it's always at a fixed location. __bpf_call_base 937 * and the helper with imm relative to it are both in core 938 * kernel. 939 */ 940 addr = (u8 *)__bpf_call_base + imm; 941 } 942 943 *func_addr = (unsigned long)addr; 944 return 0; 945 } 946 947 static int bpf_jit_blind_insn(const struct bpf_insn *from, 948 const struct bpf_insn *aux, 949 struct bpf_insn *to_buff, 950 bool emit_zext) 951 { 952 struct bpf_insn *to = to_buff; 953 u32 imm_rnd = get_random_int(); 954 s16 off; 955 956 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 957 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 958 959 /* Constraints on AX register: 960 * 961 * AX register is inaccessible from user space. It is mapped in 962 * all JITs, and used here for constant blinding rewrites. It is 963 * typically "stateless" meaning its contents are only valid within 964 * the executed instruction, but not across several instructions. 965 * There are a few exceptions however which are further detailed 966 * below. 967 * 968 * Constant blinding is only used by JITs, not in the interpreter. 969 * The interpreter uses AX in some occasions as a local temporary 970 * register e.g. in DIV or MOD instructions. 971 * 972 * In restricted circumstances, the verifier can also use the AX 973 * register for rewrites as long as they do not interfere with 974 * the above cases! 975 */ 976 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 977 goto out; 978 979 if (from->imm == 0 && 980 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 981 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 982 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 983 goto out; 984 } 985 986 switch (from->code) { 987 case BPF_ALU | BPF_ADD | BPF_K: 988 case BPF_ALU | BPF_SUB | BPF_K: 989 case BPF_ALU | BPF_AND | BPF_K: 990 case BPF_ALU | BPF_OR | BPF_K: 991 case BPF_ALU | BPF_XOR | BPF_K: 992 case BPF_ALU | BPF_MUL | BPF_K: 993 case BPF_ALU | BPF_MOV | BPF_K: 994 case BPF_ALU | BPF_DIV | BPF_K: 995 case BPF_ALU | BPF_MOD | BPF_K: 996 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 997 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 998 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 999 break; 1000 1001 case BPF_ALU64 | BPF_ADD | BPF_K: 1002 case BPF_ALU64 | BPF_SUB | BPF_K: 1003 case BPF_ALU64 | BPF_AND | BPF_K: 1004 case BPF_ALU64 | BPF_OR | BPF_K: 1005 case BPF_ALU64 | BPF_XOR | BPF_K: 1006 case BPF_ALU64 | BPF_MUL | BPF_K: 1007 case BPF_ALU64 | BPF_MOV | BPF_K: 1008 case BPF_ALU64 | BPF_DIV | BPF_K: 1009 case BPF_ALU64 | BPF_MOD | BPF_K: 1010 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1011 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1012 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 1013 break; 1014 1015 case BPF_JMP | BPF_JEQ | BPF_K: 1016 case BPF_JMP | BPF_JNE | BPF_K: 1017 case BPF_JMP | BPF_JGT | BPF_K: 1018 case BPF_JMP | BPF_JLT | BPF_K: 1019 case BPF_JMP | BPF_JGE | BPF_K: 1020 case BPF_JMP | BPF_JLE | BPF_K: 1021 case BPF_JMP | BPF_JSGT | BPF_K: 1022 case BPF_JMP | BPF_JSLT | BPF_K: 1023 case BPF_JMP | BPF_JSGE | BPF_K: 1024 case BPF_JMP | BPF_JSLE | BPF_K: 1025 case BPF_JMP | BPF_JSET | BPF_K: 1026 /* Accommodate for extra offset in case of a backjump. */ 1027 off = from->off; 1028 if (off < 0) 1029 off -= 2; 1030 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1031 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1032 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1033 break; 1034 1035 case BPF_JMP32 | BPF_JEQ | BPF_K: 1036 case BPF_JMP32 | BPF_JNE | BPF_K: 1037 case BPF_JMP32 | BPF_JGT | BPF_K: 1038 case BPF_JMP32 | BPF_JLT | BPF_K: 1039 case BPF_JMP32 | BPF_JGE | BPF_K: 1040 case BPF_JMP32 | BPF_JLE | BPF_K: 1041 case BPF_JMP32 | BPF_JSGT | BPF_K: 1042 case BPF_JMP32 | BPF_JSLT | BPF_K: 1043 case BPF_JMP32 | BPF_JSGE | BPF_K: 1044 case BPF_JMP32 | BPF_JSLE | BPF_K: 1045 case BPF_JMP32 | BPF_JSET | BPF_K: 1046 /* Accommodate for extra offset in case of a backjump. */ 1047 off = from->off; 1048 if (off < 0) 1049 off -= 2; 1050 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1051 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1052 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1053 off); 1054 break; 1055 1056 case BPF_LD | BPF_IMM | BPF_DW: 1057 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1058 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1059 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1060 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1061 break; 1062 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1063 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1064 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1065 if (emit_zext) 1066 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1067 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1068 break; 1069 1070 case BPF_ST | BPF_MEM | BPF_DW: 1071 case BPF_ST | BPF_MEM | BPF_W: 1072 case BPF_ST | BPF_MEM | BPF_H: 1073 case BPF_ST | BPF_MEM | BPF_B: 1074 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1075 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1076 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1077 break; 1078 } 1079 out: 1080 return to - to_buff; 1081 } 1082 1083 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1084 gfp_t gfp_extra_flags) 1085 { 1086 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1087 struct bpf_prog *fp; 1088 1089 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1090 if (fp != NULL) { 1091 /* aux->prog still points to the fp_other one, so 1092 * when promoting the clone to the real program, 1093 * this still needs to be adapted. 1094 */ 1095 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1096 } 1097 1098 return fp; 1099 } 1100 1101 static void bpf_prog_clone_free(struct bpf_prog *fp) 1102 { 1103 /* aux was stolen by the other clone, so we cannot free 1104 * it from this path! It will be freed eventually by the 1105 * other program on release. 1106 * 1107 * At this point, we don't need a deferred release since 1108 * clone is guaranteed to not be locked. 1109 */ 1110 fp->aux = NULL; 1111 __bpf_prog_free(fp); 1112 } 1113 1114 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1115 { 1116 /* We have to repoint aux->prog to self, as we don't 1117 * know whether fp here is the clone or the original. 1118 */ 1119 fp->aux->prog = fp; 1120 bpf_prog_clone_free(fp_other); 1121 } 1122 1123 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1124 { 1125 struct bpf_insn insn_buff[16], aux[2]; 1126 struct bpf_prog *clone, *tmp; 1127 int insn_delta, insn_cnt; 1128 struct bpf_insn *insn; 1129 int i, rewritten; 1130 1131 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 1132 return prog; 1133 1134 clone = bpf_prog_clone_create(prog, GFP_USER); 1135 if (!clone) 1136 return ERR_PTR(-ENOMEM); 1137 1138 insn_cnt = clone->len; 1139 insn = clone->insnsi; 1140 1141 for (i = 0; i < insn_cnt; i++, insn++) { 1142 /* We temporarily need to hold the original ld64 insn 1143 * so that we can still access the first part in the 1144 * second blinding run. 1145 */ 1146 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1147 insn[1].code == 0) 1148 memcpy(aux, insn, sizeof(aux)); 1149 1150 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1151 clone->aux->verifier_zext); 1152 if (!rewritten) 1153 continue; 1154 1155 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1156 if (IS_ERR(tmp)) { 1157 /* Patching may have repointed aux->prog during 1158 * realloc from the original one, so we need to 1159 * fix it up here on error. 1160 */ 1161 bpf_jit_prog_release_other(prog, clone); 1162 return tmp; 1163 } 1164 1165 clone = tmp; 1166 insn_delta = rewritten - 1; 1167 1168 /* Walk new program and skip insns we just inserted. */ 1169 insn = clone->insnsi + i + insn_delta; 1170 insn_cnt += insn_delta; 1171 i += insn_delta; 1172 } 1173 1174 clone->blinded = 1; 1175 return clone; 1176 } 1177 #endif /* CONFIG_BPF_JIT */ 1178 1179 /* Base function for offset calculation. Needs to go into .text section, 1180 * therefore keeping it non-static as well; will also be used by JITs 1181 * anyway later on, so do not let the compiler omit it. This also needs 1182 * to go into kallsyms for correlation from e.g. bpftool, so naming 1183 * must not change. 1184 */ 1185 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1186 { 1187 return 0; 1188 } 1189 EXPORT_SYMBOL_GPL(__bpf_call_base); 1190 1191 /* All UAPI available opcodes. */ 1192 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1193 /* 32 bit ALU operations. */ \ 1194 /* Register based. */ \ 1195 INSN_3(ALU, ADD, X), \ 1196 INSN_3(ALU, SUB, X), \ 1197 INSN_3(ALU, AND, X), \ 1198 INSN_3(ALU, OR, X), \ 1199 INSN_3(ALU, LSH, X), \ 1200 INSN_3(ALU, RSH, X), \ 1201 INSN_3(ALU, XOR, X), \ 1202 INSN_3(ALU, MUL, X), \ 1203 INSN_3(ALU, MOV, X), \ 1204 INSN_3(ALU, ARSH, X), \ 1205 INSN_3(ALU, DIV, X), \ 1206 INSN_3(ALU, MOD, X), \ 1207 INSN_2(ALU, NEG), \ 1208 INSN_3(ALU, END, TO_BE), \ 1209 INSN_3(ALU, END, TO_LE), \ 1210 /* Immediate based. */ \ 1211 INSN_3(ALU, ADD, K), \ 1212 INSN_3(ALU, SUB, K), \ 1213 INSN_3(ALU, AND, K), \ 1214 INSN_3(ALU, OR, K), \ 1215 INSN_3(ALU, LSH, K), \ 1216 INSN_3(ALU, RSH, K), \ 1217 INSN_3(ALU, XOR, K), \ 1218 INSN_3(ALU, MUL, K), \ 1219 INSN_3(ALU, MOV, K), \ 1220 INSN_3(ALU, ARSH, K), \ 1221 INSN_3(ALU, DIV, K), \ 1222 INSN_3(ALU, MOD, K), \ 1223 /* 64 bit ALU operations. */ \ 1224 /* Register based. */ \ 1225 INSN_3(ALU64, ADD, X), \ 1226 INSN_3(ALU64, SUB, X), \ 1227 INSN_3(ALU64, AND, X), \ 1228 INSN_3(ALU64, OR, X), \ 1229 INSN_3(ALU64, LSH, X), \ 1230 INSN_3(ALU64, RSH, X), \ 1231 INSN_3(ALU64, XOR, X), \ 1232 INSN_3(ALU64, MUL, X), \ 1233 INSN_3(ALU64, MOV, X), \ 1234 INSN_3(ALU64, ARSH, X), \ 1235 INSN_3(ALU64, DIV, X), \ 1236 INSN_3(ALU64, MOD, X), \ 1237 INSN_2(ALU64, NEG), \ 1238 /* Immediate based. */ \ 1239 INSN_3(ALU64, ADD, K), \ 1240 INSN_3(ALU64, SUB, K), \ 1241 INSN_3(ALU64, AND, K), \ 1242 INSN_3(ALU64, OR, K), \ 1243 INSN_3(ALU64, LSH, K), \ 1244 INSN_3(ALU64, RSH, K), \ 1245 INSN_3(ALU64, XOR, K), \ 1246 INSN_3(ALU64, MUL, K), \ 1247 INSN_3(ALU64, MOV, K), \ 1248 INSN_3(ALU64, ARSH, K), \ 1249 INSN_3(ALU64, DIV, K), \ 1250 INSN_3(ALU64, MOD, K), \ 1251 /* Call instruction. */ \ 1252 INSN_2(JMP, CALL), \ 1253 /* Exit instruction. */ \ 1254 INSN_2(JMP, EXIT), \ 1255 /* 32-bit Jump instructions. */ \ 1256 /* Register based. */ \ 1257 INSN_3(JMP32, JEQ, X), \ 1258 INSN_3(JMP32, JNE, X), \ 1259 INSN_3(JMP32, JGT, X), \ 1260 INSN_3(JMP32, JLT, X), \ 1261 INSN_3(JMP32, JGE, X), \ 1262 INSN_3(JMP32, JLE, X), \ 1263 INSN_3(JMP32, JSGT, X), \ 1264 INSN_3(JMP32, JSLT, X), \ 1265 INSN_3(JMP32, JSGE, X), \ 1266 INSN_3(JMP32, JSLE, X), \ 1267 INSN_3(JMP32, JSET, X), \ 1268 /* Immediate based. */ \ 1269 INSN_3(JMP32, JEQ, K), \ 1270 INSN_3(JMP32, JNE, K), \ 1271 INSN_3(JMP32, JGT, K), \ 1272 INSN_3(JMP32, JLT, K), \ 1273 INSN_3(JMP32, JGE, K), \ 1274 INSN_3(JMP32, JLE, K), \ 1275 INSN_3(JMP32, JSGT, K), \ 1276 INSN_3(JMP32, JSLT, K), \ 1277 INSN_3(JMP32, JSGE, K), \ 1278 INSN_3(JMP32, JSLE, K), \ 1279 INSN_3(JMP32, JSET, K), \ 1280 /* Jump instructions. */ \ 1281 /* Register based. */ \ 1282 INSN_3(JMP, JEQ, X), \ 1283 INSN_3(JMP, JNE, X), \ 1284 INSN_3(JMP, JGT, X), \ 1285 INSN_3(JMP, JLT, X), \ 1286 INSN_3(JMP, JGE, X), \ 1287 INSN_3(JMP, JLE, X), \ 1288 INSN_3(JMP, JSGT, X), \ 1289 INSN_3(JMP, JSLT, X), \ 1290 INSN_3(JMP, JSGE, X), \ 1291 INSN_3(JMP, JSLE, X), \ 1292 INSN_3(JMP, JSET, X), \ 1293 /* Immediate based. */ \ 1294 INSN_3(JMP, JEQ, K), \ 1295 INSN_3(JMP, JNE, K), \ 1296 INSN_3(JMP, JGT, K), \ 1297 INSN_3(JMP, JLT, K), \ 1298 INSN_3(JMP, JGE, K), \ 1299 INSN_3(JMP, JLE, K), \ 1300 INSN_3(JMP, JSGT, K), \ 1301 INSN_3(JMP, JSLT, K), \ 1302 INSN_3(JMP, JSGE, K), \ 1303 INSN_3(JMP, JSLE, K), \ 1304 INSN_3(JMP, JSET, K), \ 1305 INSN_2(JMP, JA), \ 1306 /* Store instructions. */ \ 1307 /* Register based. */ \ 1308 INSN_3(STX, MEM, B), \ 1309 INSN_3(STX, MEM, H), \ 1310 INSN_3(STX, MEM, W), \ 1311 INSN_3(STX, MEM, DW), \ 1312 INSN_3(STX, XADD, W), \ 1313 INSN_3(STX, XADD, DW), \ 1314 /* Immediate based. */ \ 1315 INSN_3(ST, MEM, B), \ 1316 INSN_3(ST, MEM, H), \ 1317 INSN_3(ST, MEM, W), \ 1318 INSN_3(ST, MEM, DW), \ 1319 /* Load instructions. */ \ 1320 /* Register based. */ \ 1321 INSN_3(LDX, MEM, B), \ 1322 INSN_3(LDX, MEM, H), \ 1323 INSN_3(LDX, MEM, W), \ 1324 INSN_3(LDX, MEM, DW), \ 1325 /* Immediate based. */ \ 1326 INSN_3(LD, IMM, DW) 1327 1328 bool bpf_opcode_in_insntable(u8 code) 1329 { 1330 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1331 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1332 static const bool public_insntable[256] = { 1333 [0 ... 255] = false, 1334 /* Now overwrite non-defaults ... */ 1335 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1336 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1337 [BPF_LD | BPF_ABS | BPF_B] = true, 1338 [BPF_LD | BPF_ABS | BPF_H] = true, 1339 [BPF_LD | BPF_ABS | BPF_W] = true, 1340 [BPF_LD | BPF_IND | BPF_B] = true, 1341 [BPF_LD | BPF_IND | BPF_H] = true, 1342 [BPF_LD | BPF_IND | BPF_W] = true, 1343 }; 1344 #undef BPF_INSN_3_TBL 1345 #undef BPF_INSN_2_TBL 1346 return public_insntable[code]; 1347 } 1348 1349 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1350 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1351 { 1352 memset(dst, 0, size); 1353 return -EFAULT; 1354 } 1355 1356 /** 1357 * __bpf_prog_run - run eBPF program on a given context 1358 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1359 * @insn: is the array of eBPF instructions 1360 * @stack: is the eBPF storage stack 1361 * 1362 * Decode and execute eBPF instructions. 1363 */ 1364 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 1365 { 1366 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1367 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1368 static const void * const jumptable[256] __annotate_jump_table = { 1369 [0 ... 255] = &&default_label, 1370 /* Now overwrite non-defaults ... */ 1371 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1372 /* Non-UAPI available opcodes. */ 1373 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1374 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1375 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1376 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1377 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1378 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1379 }; 1380 #undef BPF_INSN_3_LBL 1381 #undef BPF_INSN_2_LBL 1382 u32 tail_call_cnt = 0; 1383 1384 #define CONT ({ insn++; goto select_insn; }) 1385 #define CONT_JMP ({ insn++; goto select_insn; }) 1386 1387 select_insn: 1388 goto *jumptable[insn->code]; 1389 1390 /* ALU */ 1391 #define ALU(OPCODE, OP) \ 1392 ALU64_##OPCODE##_X: \ 1393 DST = DST OP SRC; \ 1394 CONT; \ 1395 ALU_##OPCODE##_X: \ 1396 DST = (u32) DST OP (u32) SRC; \ 1397 CONT; \ 1398 ALU64_##OPCODE##_K: \ 1399 DST = DST OP IMM; \ 1400 CONT; \ 1401 ALU_##OPCODE##_K: \ 1402 DST = (u32) DST OP (u32) IMM; \ 1403 CONT; 1404 1405 ALU(ADD, +) 1406 ALU(SUB, -) 1407 ALU(AND, &) 1408 ALU(OR, |) 1409 ALU(LSH, <<) 1410 ALU(RSH, >>) 1411 ALU(XOR, ^) 1412 ALU(MUL, *) 1413 #undef ALU 1414 ALU_NEG: 1415 DST = (u32) -DST; 1416 CONT; 1417 ALU64_NEG: 1418 DST = -DST; 1419 CONT; 1420 ALU_MOV_X: 1421 DST = (u32) SRC; 1422 CONT; 1423 ALU_MOV_K: 1424 DST = (u32) IMM; 1425 CONT; 1426 ALU64_MOV_X: 1427 DST = SRC; 1428 CONT; 1429 ALU64_MOV_K: 1430 DST = IMM; 1431 CONT; 1432 LD_IMM_DW: 1433 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1434 insn++; 1435 CONT; 1436 ALU_ARSH_X: 1437 DST = (u64) (u32) (((s32) DST) >> SRC); 1438 CONT; 1439 ALU_ARSH_K: 1440 DST = (u64) (u32) (((s32) DST) >> IMM); 1441 CONT; 1442 ALU64_ARSH_X: 1443 (*(s64 *) &DST) >>= SRC; 1444 CONT; 1445 ALU64_ARSH_K: 1446 (*(s64 *) &DST) >>= IMM; 1447 CONT; 1448 ALU64_MOD_X: 1449 div64_u64_rem(DST, SRC, &AX); 1450 DST = AX; 1451 CONT; 1452 ALU_MOD_X: 1453 AX = (u32) DST; 1454 DST = do_div(AX, (u32) SRC); 1455 CONT; 1456 ALU64_MOD_K: 1457 div64_u64_rem(DST, IMM, &AX); 1458 DST = AX; 1459 CONT; 1460 ALU_MOD_K: 1461 AX = (u32) DST; 1462 DST = do_div(AX, (u32) IMM); 1463 CONT; 1464 ALU64_DIV_X: 1465 DST = div64_u64(DST, SRC); 1466 CONT; 1467 ALU_DIV_X: 1468 AX = (u32) DST; 1469 do_div(AX, (u32) SRC); 1470 DST = (u32) AX; 1471 CONT; 1472 ALU64_DIV_K: 1473 DST = div64_u64(DST, IMM); 1474 CONT; 1475 ALU_DIV_K: 1476 AX = (u32) DST; 1477 do_div(AX, (u32) IMM); 1478 DST = (u32) AX; 1479 CONT; 1480 ALU_END_TO_BE: 1481 switch (IMM) { 1482 case 16: 1483 DST = (__force u16) cpu_to_be16(DST); 1484 break; 1485 case 32: 1486 DST = (__force u32) cpu_to_be32(DST); 1487 break; 1488 case 64: 1489 DST = (__force u64) cpu_to_be64(DST); 1490 break; 1491 } 1492 CONT; 1493 ALU_END_TO_LE: 1494 switch (IMM) { 1495 case 16: 1496 DST = (__force u16) cpu_to_le16(DST); 1497 break; 1498 case 32: 1499 DST = (__force u32) cpu_to_le32(DST); 1500 break; 1501 case 64: 1502 DST = (__force u64) cpu_to_le64(DST); 1503 break; 1504 } 1505 CONT; 1506 1507 /* CALL */ 1508 JMP_CALL: 1509 /* Function call scratches BPF_R1-BPF_R5 registers, 1510 * preserves BPF_R6-BPF_R9, and stores return value 1511 * into BPF_R0. 1512 */ 1513 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1514 BPF_R4, BPF_R5); 1515 CONT; 1516 1517 JMP_CALL_ARGS: 1518 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1519 BPF_R3, BPF_R4, 1520 BPF_R5, 1521 insn + insn->off + 1); 1522 CONT; 1523 1524 JMP_TAIL_CALL: { 1525 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1526 struct bpf_array *array = container_of(map, struct bpf_array, map); 1527 struct bpf_prog *prog; 1528 u32 index = BPF_R3; 1529 1530 if (unlikely(index >= array->map.max_entries)) 1531 goto out; 1532 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1533 goto out; 1534 1535 tail_call_cnt++; 1536 1537 prog = READ_ONCE(array->ptrs[index]); 1538 if (!prog) 1539 goto out; 1540 1541 /* ARG1 at this point is guaranteed to point to CTX from 1542 * the verifier side due to the fact that the tail call is 1543 * handled like a helper, that is, bpf_tail_call_proto, 1544 * where arg1_type is ARG_PTR_TO_CTX. 1545 */ 1546 insn = prog->insnsi; 1547 goto select_insn; 1548 out: 1549 CONT; 1550 } 1551 JMP_JA: 1552 insn += insn->off; 1553 CONT; 1554 JMP_EXIT: 1555 return BPF_R0; 1556 /* JMP */ 1557 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1558 JMP_##OPCODE##_X: \ 1559 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1560 insn += insn->off; \ 1561 CONT_JMP; \ 1562 } \ 1563 CONT; \ 1564 JMP32_##OPCODE##_X: \ 1565 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1566 insn += insn->off; \ 1567 CONT_JMP; \ 1568 } \ 1569 CONT; \ 1570 JMP_##OPCODE##_K: \ 1571 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1572 insn += insn->off; \ 1573 CONT_JMP; \ 1574 } \ 1575 CONT; \ 1576 JMP32_##OPCODE##_K: \ 1577 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1578 insn += insn->off; \ 1579 CONT_JMP; \ 1580 } \ 1581 CONT; 1582 COND_JMP(u, JEQ, ==) 1583 COND_JMP(u, JNE, !=) 1584 COND_JMP(u, JGT, >) 1585 COND_JMP(u, JLT, <) 1586 COND_JMP(u, JGE, >=) 1587 COND_JMP(u, JLE, <=) 1588 COND_JMP(u, JSET, &) 1589 COND_JMP(s, JSGT, >) 1590 COND_JMP(s, JSLT, <) 1591 COND_JMP(s, JSGE, >=) 1592 COND_JMP(s, JSLE, <=) 1593 #undef COND_JMP 1594 /* STX and ST and LDX*/ 1595 #define LDST(SIZEOP, SIZE) \ 1596 STX_MEM_##SIZEOP: \ 1597 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1598 CONT; \ 1599 ST_MEM_##SIZEOP: \ 1600 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1601 CONT; \ 1602 LDX_MEM_##SIZEOP: \ 1603 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1604 CONT; 1605 1606 LDST(B, u8) 1607 LDST(H, u16) 1608 LDST(W, u32) 1609 LDST(DW, u64) 1610 #undef LDST 1611 #define LDX_PROBE(SIZEOP, SIZE) \ 1612 LDX_PROBE_MEM_##SIZEOP: \ 1613 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \ 1614 CONT; 1615 LDX_PROBE(B, 1) 1616 LDX_PROBE(H, 2) 1617 LDX_PROBE(W, 4) 1618 LDX_PROBE(DW, 8) 1619 #undef LDX_PROBE 1620 1621 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1622 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1623 (DST + insn->off)); 1624 CONT; 1625 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1626 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1627 (DST + insn->off)); 1628 CONT; 1629 1630 default_label: 1631 /* If we ever reach this, we have a bug somewhere. Die hard here 1632 * instead of just returning 0; we could be somewhere in a subprog, 1633 * so execution could continue otherwise which we do /not/ want. 1634 * 1635 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1636 */ 1637 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code); 1638 BUG_ON(1); 1639 return 0; 1640 } 1641 1642 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1643 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1644 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1645 { \ 1646 u64 stack[stack_size / sizeof(u64)]; \ 1647 u64 regs[MAX_BPF_EXT_REG]; \ 1648 \ 1649 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1650 ARG1 = (u64) (unsigned long) ctx; \ 1651 return ___bpf_prog_run(regs, insn, stack); \ 1652 } 1653 1654 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1655 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1656 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1657 const struct bpf_insn *insn) \ 1658 { \ 1659 u64 stack[stack_size / sizeof(u64)]; \ 1660 u64 regs[MAX_BPF_EXT_REG]; \ 1661 \ 1662 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1663 BPF_R1 = r1; \ 1664 BPF_R2 = r2; \ 1665 BPF_R3 = r3; \ 1666 BPF_R4 = r4; \ 1667 BPF_R5 = r5; \ 1668 return ___bpf_prog_run(regs, insn, stack); \ 1669 } 1670 1671 #define EVAL1(FN, X) FN(X) 1672 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1673 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1674 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1675 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1676 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1677 1678 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1679 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1680 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1681 1682 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1683 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1684 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1685 1686 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1687 1688 static unsigned int (*interpreters[])(const void *ctx, 1689 const struct bpf_insn *insn) = { 1690 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1691 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1692 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1693 }; 1694 #undef PROG_NAME_LIST 1695 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1696 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1697 const struct bpf_insn *insn) = { 1698 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1699 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1700 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1701 }; 1702 #undef PROG_NAME_LIST 1703 1704 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1705 { 1706 stack_depth = max_t(u32, stack_depth, 1); 1707 insn->off = (s16) insn->imm; 1708 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1709 __bpf_call_base_args; 1710 insn->code = BPF_JMP | BPF_CALL_ARGS; 1711 } 1712 1713 #else 1714 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1715 const struct bpf_insn *insn) 1716 { 1717 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1718 * is not working properly, so warn about it! 1719 */ 1720 WARN_ON_ONCE(1); 1721 return 0; 1722 } 1723 #endif 1724 1725 bool bpf_prog_array_compatible(struct bpf_array *array, 1726 const struct bpf_prog *fp) 1727 { 1728 if (fp->kprobe_override) 1729 return false; 1730 1731 if (!array->aux->type) { 1732 /* There's no owner yet where we could check for 1733 * compatibility. 1734 */ 1735 array->aux->type = fp->type; 1736 array->aux->jited = fp->jited; 1737 return true; 1738 } 1739 1740 return array->aux->type == fp->type && 1741 array->aux->jited == fp->jited; 1742 } 1743 1744 static int bpf_check_tail_call(const struct bpf_prog *fp) 1745 { 1746 struct bpf_prog_aux *aux = fp->aux; 1747 int i, ret = 0; 1748 1749 mutex_lock(&aux->used_maps_mutex); 1750 for (i = 0; i < aux->used_map_cnt; i++) { 1751 struct bpf_map *map = aux->used_maps[i]; 1752 struct bpf_array *array; 1753 1754 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1755 continue; 1756 1757 array = container_of(map, struct bpf_array, map); 1758 if (!bpf_prog_array_compatible(array, fp)) { 1759 ret = -EINVAL; 1760 goto out; 1761 } 1762 } 1763 1764 out: 1765 mutex_unlock(&aux->used_maps_mutex); 1766 return ret; 1767 } 1768 1769 static void bpf_prog_select_func(struct bpf_prog *fp) 1770 { 1771 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1772 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1773 1774 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1775 #else 1776 fp->bpf_func = __bpf_prog_ret0_warn; 1777 #endif 1778 } 1779 1780 /** 1781 * bpf_prog_select_runtime - select exec runtime for BPF program 1782 * @fp: bpf_prog populated with internal BPF program 1783 * @err: pointer to error variable 1784 * 1785 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1786 * The BPF program will be executed via BPF_PROG_RUN() macro. 1787 */ 1788 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1789 { 1790 /* In case of BPF to BPF calls, verifier did all the prep 1791 * work with regards to JITing, etc. 1792 */ 1793 if (fp->bpf_func) 1794 goto finalize; 1795 1796 bpf_prog_select_func(fp); 1797 1798 /* eBPF JITs can rewrite the program in case constant 1799 * blinding is active. However, in case of error during 1800 * blinding, bpf_int_jit_compile() must always return a 1801 * valid program, which in this case would simply not 1802 * be JITed, but falls back to the interpreter. 1803 */ 1804 if (!bpf_prog_is_dev_bound(fp->aux)) { 1805 *err = bpf_prog_alloc_jited_linfo(fp); 1806 if (*err) 1807 return fp; 1808 1809 fp = bpf_int_jit_compile(fp); 1810 if (!fp->jited) { 1811 bpf_prog_free_jited_linfo(fp); 1812 #ifdef CONFIG_BPF_JIT_ALWAYS_ON 1813 *err = -ENOTSUPP; 1814 return fp; 1815 #endif 1816 } else { 1817 bpf_prog_free_unused_jited_linfo(fp); 1818 } 1819 } else { 1820 *err = bpf_prog_offload_compile(fp); 1821 if (*err) 1822 return fp; 1823 } 1824 1825 finalize: 1826 bpf_prog_lock_ro(fp); 1827 1828 /* The tail call compatibility check can only be done at 1829 * this late stage as we need to determine, if we deal 1830 * with JITed or non JITed program concatenations and not 1831 * all eBPF JITs might immediately support all features. 1832 */ 1833 *err = bpf_check_tail_call(fp); 1834 1835 return fp; 1836 } 1837 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1838 1839 static unsigned int __bpf_prog_ret1(const void *ctx, 1840 const struct bpf_insn *insn) 1841 { 1842 return 1; 1843 } 1844 1845 static struct bpf_prog_dummy { 1846 struct bpf_prog prog; 1847 } dummy_bpf_prog = { 1848 .prog = { 1849 .bpf_func = __bpf_prog_ret1, 1850 }, 1851 }; 1852 1853 /* to avoid allocating empty bpf_prog_array for cgroups that 1854 * don't have bpf program attached use one global 'empty_prog_array' 1855 * It will not be modified the caller of bpf_prog_array_alloc() 1856 * (since caller requested prog_cnt == 0) 1857 * that pointer should be 'freed' by bpf_prog_array_free() 1858 */ 1859 static struct { 1860 struct bpf_prog_array hdr; 1861 struct bpf_prog *null_prog; 1862 } empty_prog_array = { 1863 .null_prog = NULL, 1864 }; 1865 1866 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1867 { 1868 if (prog_cnt) 1869 return kzalloc(sizeof(struct bpf_prog_array) + 1870 sizeof(struct bpf_prog_array_item) * 1871 (prog_cnt + 1), 1872 flags); 1873 1874 return &empty_prog_array.hdr; 1875 } 1876 1877 void bpf_prog_array_free(struct bpf_prog_array *progs) 1878 { 1879 if (!progs || progs == &empty_prog_array.hdr) 1880 return; 1881 kfree_rcu(progs, rcu); 1882 } 1883 1884 int bpf_prog_array_length(struct bpf_prog_array *array) 1885 { 1886 struct bpf_prog_array_item *item; 1887 u32 cnt = 0; 1888 1889 for (item = array->items; item->prog; item++) 1890 if (item->prog != &dummy_bpf_prog.prog) 1891 cnt++; 1892 return cnt; 1893 } 1894 1895 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 1896 { 1897 struct bpf_prog_array_item *item; 1898 1899 for (item = array->items; item->prog; item++) 1900 if (item->prog != &dummy_bpf_prog.prog) 1901 return false; 1902 return true; 1903 } 1904 1905 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 1906 u32 *prog_ids, 1907 u32 request_cnt) 1908 { 1909 struct bpf_prog_array_item *item; 1910 int i = 0; 1911 1912 for (item = array->items; item->prog; item++) { 1913 if (item->prog == &dummy_bpf_prog.prog) 1914 continue; 1915 prog_ids[i] = item->prog->aux->id; 1916 if (++i == request_cnt) { 1917 item++; 1918 break; 1919 } 1920 } 1921 1922 return !!(item->prog); 1923 } 1924 1925 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 1926 __u32 __user *prog_ids, u32 cnt) 1927 { 1928 unsigned long err = 0; 1929 bool nospc; 1930 u32 *ids; 1931 1932 /* users of this function are doing: 1933 * cnt = bpf_prog_array_length(); 1934 * if (cnt > 0) 1935 * bpf_prog_array_copy_to_user(..., cnt); 1936 * so below kcalloc doesn't need extra cnt > 0 check. 1937 */ 1938 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 1939 if (!ids) 1940 return -ENOMEM; 1941 nospc = bpf_prog_array_copy_core(array, ids, cnt); 1942 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 1943 kfree(ids); 1944 if (err) 1945 return -EFAULT; 1946 if (nospc) 1947 return -ENOSPC; 1948 return 0; 1949 } 1950 1951 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 1952 struct bpf_prog *old_prog) 1953 { 1954 struct bpf_prog_array_item *item; 1955 1956 for (item = array->items; item->prog; item++) 1957 if (item->prog == old_prog) { 1958 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 1959 break; 1960 } 1961 } 1962 1963 /** 1964 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 1965 * index into the program array with 1966 * a dummy no-op program. 1967 * @array: a bpf_prog_array 1968 * @index: the index of the program to replace 1969 * 1970 * Skips over dummy programs, by not counting them, when calculating 1971 * the position of the program to replace. 1972 * 1973 * Return: 1974 * * 0 - Success 1975 * * -EINVAL - Invalid index value. Must be a non-negative integer. 1976 * * -ENOENT - Index out of range 1977 */ 1978 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 1979 { 1980 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 1981 } 1982 1983 /** 1984 * bpf_prog_array_update_at() - Updates the program at the given index 1985 * into the program array. 1986 * @array: a bpf_prog_array 1987 * @index: the index of the program to update 1988 * @prog: the program to insert into the array 1989 * 1990 * Skips over dummy programs, by not counting them, when calculating 1991 * the position of the program to update. 1992 * 1993 * Return: 1994 * * 0 - Success 1995 * * -EINVAL - Invalid index value. Must be a non-negative integer. 1996 * * -ENOENT - Index out of range 1997 */ 1998 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 1999 struct bpf_prog *prog) 2000 { 2001 struct bpf_prog_array_item *item; 2002 2003 if (unlikely(index < 0)) 2004 return -EINVAL; 2005 2006 for (item = array->items; item->prog; item++) { 2007 if (item->prog == &dummy_bpf_prog.prog) 2008 continue; 2009 if (!index) { 2010 WRITE_ONCE(item->prog, prog); 2011 return 0; 2012 } 2013 index--; 2014 } 2015 return -ENOENT; 2016 } 2017 2018 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2019 struct bpf_prog *exclude_prog, 2020 struct bpf_prog *include_prog, 2021 struct bpf_prog_array **new_array) 2022 { 2023 int new_prog_cnt, carry_prog_cnt = 0; 2024 struct bpf_prog_array_item *existing; 2025 struct bpf_prog_array *array; 2026 bool found_exclude = false; 2027 int new_prog_idx = 0; 2028 2029 /* Figure out how many existing progs we need to carry over to 2030 * the new array. 2031 */ 2032 if (old_array) { 2033 existing = old_array->items; 2034 for (; existing->prog; existing++) { 2035 if (existing->prog == exclude_prog) { 2036 found_exclude = true; 2037 continue; 2038 } 2039 if (existing->prog != &dummy_bpf_prog.prog) 2040 carry_prog_cnt++; 2041 if (existing->prog == include_prog) 2042 return -EEXIST; 2043 } 2044 } 2045 2046 if (exclude_prog && !found_exclude) 2047 return -ENOENT; 2048 2049 /* How many progs (not NULL) will be in the new array? */ 2050 new_prog_cnt = carry_prog_cnt; 2051 if (include_prog) 2052 new_prog_cnt += 1; 2053 2054 /* Do we have any prog (not NULL) in the new array? */ 2055 if (!new_prog_cnt) { 2056 *new_array = NULL; 2057 return 0; 2058 } 2059 2060 /* +1 as the end of prog_array is marked with NULL */ 2061 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2062 if (!array) 2063 return -ENOMEM; 2064 2065 /* Fill in the new prog array */ 2066 if (carry_prog_cnt) { 2067 existing = old_array->items; 2068 for (; existing->prog; existing++) 2069 if (existing->prog != exclude_prog && 2070 existing->prog != &dummy_bpf_prog.prog) { 2071 array->items[new_prog_idx++].prog = 2072 existing->prog; 2073 } 2074 } 2075 if (include_prog) 2076 array->items[new_prog_idx++].prog = include_prog; 2077 array->items[new_prog_idx].prog = NULL; 2078 *new_array = array; 2079 return 0; 2080 } 2081 2082 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2083 u32 *prog_ids, u32 request_cnt, 2084 u32 *prog_cnt) 2085 { 2086 u32 cnt = 0; 2087 2088 if (array) 2089 cnt = bpf_prog_array_length(array); 2090 2091 *prog_cnt = cnt; 2092 2093 /* return early if user requested only program count or nothing to copy */ 2094 if (!request_cnt || !cnt) 2095 return 0; 2096 2097 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2098 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2099 : 0; 2100 } 2101 2102 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2103 struct bpf_map **used_maps, u32 len) 2104 { 2105 struct bpf_map *map; 2106 u32 i; 2107 2108 for (i = 0; i < len; i++) { 2109 map = used_maps[i]; 2110 if (map->ops->map_poke_untrack) 2111 map->ops->map_poke_untrack(map, aux); 2112 bpf_map_put(map); 2113 } 2114 } 2115 2116 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2117 { 2118 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2119 kfree(aux->used_maps); 2120 } 2121 2122 static void bpf_prog_free_deferred(struct work_struct *work) 2123 { 2124 struct bpf_prog_aux *aux; 2125 int i; 2126 2127 aux = container_of(work, struct bpf_prog_aux, work); 2128 bpf_free_used_maps(aux); 2129 if (bpf_prog_is_dev_bound(aux)) 2130 bpf_prog_offload_destroy(aux->prog); 2131 #ifdef CONFIG_PERF_EVENTS 2132 if (aux->prog->has_callchain_buf) 2133 put_callchain_buffers(); 2134 #endif 2135 if (aux->dst_trampoline) 2136 bpf_trampoline_put(aux->dst_trampoline); 2137 for (i = 0; i < aux->func_cnt; i++) 2138 bpf_jit_free(aux->func[i]); 2139 if (aux->func_cnt) { 2140 kfree(aux->func); 2141 bpf_prog_unlock_free(aux->prog); 2142 } else { 2143 bpf_jit_free(aux->prog); 2144 } 2145 } 2146 2147 /* Free internal BPF program */ 2148 void bpf_prog_free(struct bpf_prog *fp) 2149 { 2150 struct bpf_prog_aux *aux = fp->aux; 2151 2152 if (aux->dst_prog) 2153 bpf_prog_put(aux->dst_prog); 2154 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2155 schedule_work(&aux->work); 2156 } 2157 EXPORT_SYMBOL_GPL(bpf_prog_free); 2158 2159 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2160 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2161 2162 void bpf_user_rnd_init_once(void) 2163 { 2164 prandom_init_once(&bpf_user_rnd_state); 2165 } 2166 2167 BPF_CALL_0(bpf_user_rnd_u32) 2168 { 2169 /* Should someone ever have the rather unwise idea to use some 2170 * of the registers passed into this function, then note that 2171 * this function is called from native eBPF and classic-to-eBPF 2172 * transformations. Register assignments from both sides are 2173 * different, f.e. classic always sets fn(ctx, A, X) here. 2174 */ 2175 struct rnd_state *state; 2176 u32 res; 2177 2178 state = &get_cpu_var(bpf_user_rnd_state); 2179 res = prandom_u32_state(state); 2180 put_cpu_var(bpf_user_rnd_state); 2181 2182 return res; 2183 } 2184 2185 BPF_CALL_0(bpf_get_raw_cpu_id) 2186 { 2187 return raw_smp_processor_id(); 2188 } 2189 2190 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2191 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2192 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2193 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2194 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2195 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2196 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2197 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2198 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2199 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2200 2201 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2202 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2203 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2204 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2205 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2206 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2207 2208 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2209 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2210 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2211 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2212 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2213 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2214 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2215 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2216 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2217 2218 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2219 { 2220 return NULL; 2221 } 2222 2223 u64 __weak 2224 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2225 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2226 { 2227 return -ENOTSUPP; 2228 } 2229 EXPORT_SYMBOL_GPL(bpf_event_output); 2230 2231 /* Always built-in helper functions. */ 2232 const struct bpf_func_proto bpf_tail_call_proto = { 2233 .func = NULL, 2234 .gpl_only = false, 2235 .ret_type = RET_VOID, 2236 .arg1_type = ARG_PTR_TO_CTX, 2237 .arg2_type = ARG_CONST_MAP_PTR, 2238 .arg3_type = ARG_ANYTHING, 2239 }; 2240 2241 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2242 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2243 * eBPF and implicitly also cBPF can get JITed! 2244 */ 2245 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2246 { 2247 return prog; 2248 } 2249 2250 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2251 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2252 */ 2253 void __weak bpf_jit_compile(struct bpf_prog *prog) 2254 { 2255 } 2256 2257 bool __weak bpf_helper_changes_pkt_data(void *func) 2258 { 2259 return false; 2260 } 2261 2262 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2263 * analysis code and wants explicit zero extension inserted by verifier. 2264 * Otherwise, return FALSE. 2265 */ 2266 bool __weak bpf_jit_needs_zext(void) 2267 { 2268 return false; 2269 } 2270 2271 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2272 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2273 */ 2274 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2275 int len) 2276 { 2277 return -EFAULT; 2278 } 2279 2280 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2281 void *addr1, void *addr2) 2282 { 2283 return -ENOTSUPP; 2284 } 2285 2286 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2287 EXPORT_SYMBOL(bpf_stats_enabled_key); 2288 2289 /* All definitions of tracepoints related to BPF. */ 2290 #define CREATE_TRACE_POINTS 2291 #include <linux/bpf_trace.h> 2292 2293 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2294 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2295