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