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 static 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 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 672 return n ? 673 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 674 NULL; 675 } 676 677 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 678 unsigned long *off, char *sym) 679 { 680 unsigned long symbol_start, symbol_end; 681 struct bpf_prog *prog; 682 char *ret = NULL; 683 684 rcu_read_lock(); 685 prog = bpf_prog_kallsyms_find(addr); 686 if (prog) { 687 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 688 bpf_get_prog_name(prog, sym); 689 690 ret = sym; 691 if (size) 692 *size = symbol_end - symbol_start; 693 if (off) 694 *off = addr - symbol_start; 695 } 696 rcu_read_unlock(); 697 698 return ret; 699 } 700 701 bool is_bpf_text_address(unsigned long addr) 702 { 703 bool ret; 704 705 rcu_read_lock(); 706 ret = bpf_prog_kallsyms_find(addr) != NULL; 707 rcu_read_unlock(); 708 709 return ret; 710 } 711 712 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 713 { 714 const struct exception_table_entry *e = NULL; 715 struct bpf_prog *prog; 716 717 rcu_read_lock(); 718 prog = bpf_prog_kallsyms_find(addr); 719 if (!prog) 720 goto out; 721 if (!prog->aux->num_exentries) 722 goto out; 723 724 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 725 out: 726 rcu_read_unlock(); 727 return e; 728 } 729 730 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 731 char *sym) 732 { 733 struct bpf_prog_aux *aux; 734 unsigned int it = 0; 735 int ret = -ERANGE; 736 737 if (!bpf_jit_kallsyms_enabled()) 738 return ret; 739 740 rcu_read_lock(); 741 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 742 if (it++ != symnum) 743 continue; 744 745 bpf_get_prog_name(aux->prog, sym); 746 747 *value = (unsigned long)aux->prog->bpf_func; 748 *type = BPF_SYM_ELF_TYPE; 749 750 ret = 0; 751 break; 752 } 753 rcu_read_unlock(); 754 755 return ret; 756 } 757 758 static atomic_long_t bpf_jit_current; 759 760 /* Can be overridden by an arch's JIT compiler if it has a custom, 761 * dedicated BPF backend memory area, or if neither of the two 762 * below apply. 763 */ 764 u64 __weak bpf_jit_alloc_exec_limit(void) 765 { 766 #if defined(MODULES_VADDR) 767 return MODULES_END - MODULES_VADDR; 768 #else 769 return VMALLOC_END - VMALLOC_START; 770 #endif 771 } 772 773 static int __init bpf_jit_charge_init(void) 774 { 775 /* Only used as heuristic here to derive limit. */ 776 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2, 777 PAGE_SIZE), LONG_MAX); 778 return 0; 779 } 780 pure_initcall(bpf_jit_charge_init); 781 782 static int bpf_jit_charge_modmem(u32 pages) 783 { 784 if (atomic_long_add_return(pages, &bpf_jit_current) > 785 (bpf_jit_limit >> PAGE_SHIFT)) { 786 if (!capable(CAP_SYS_ADMIN)) { 787 atomic_long_sub(pages, &bpf_jit_current); 788 return -EPERM; 789 } 790 } 791 792 return 0; 793 } 794 795 static void bpf_jit_uncharge_modmem(u32 pages) 796 { 797 atomic_long_sub(pages, &bpf_jit_current); 798 } 799 800 void *__weak bpf_jit_alloc_exec(unsigned long size) 801 { 802 return module_alloc(size); 803 } 804 805 void __weak bpf_jit_free_exec(void *addr) 806 { 807 module_memfree(addr); 808 } 809 810 struct bpf_binary_header * 811 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 812 unsigned int alignment, 813 bpf_jit_fill_hole_t bpf_fill_ill_insns) 814 { 815 struct bpf_binary_header *hdr; 816 u32 size, hole, start, pages; 817 818 /* Most of BPF filters are really small, but if some of them 819 * fill a page, allow at least 128 extra bytes to insert a 820 * random section of illegal instructions. 821 */ 822 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 823 pages = size / PAGE_SIZE; 824 825 if (bpf_jit_charge_modmem(pages)) 826 return NULL; 827 hdr = bpf_jit_alloc_exec(size); 828 if (!hdr) { 829 bpf_jit_uncharge_modmem(pages); 830 return NULL; 831 } 832 833 /* Fill space with illegal/arch-dep instructions. */ 834 bpf_fill_ill_insns(hdr, size); 835 836 hdr->pages = pages; 837 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 838 PAGE_SIZE - sizeof(*hdr)); 839 start = (get_random_int() % hole) & ~(alignment - 1); 840 841 /* Leave a random number of instructions before BPF code. */ 842 *image_ptr = &hdr->image[start]; 843 844 return hdr; 845 } 846 847 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 848 { 849 u32 pages = hdr->pages; 850 851 bpf_jit_free_exec(hdr); 852 bpf_jit_uncharge_modmem(pages); 853 } 854 855 /* This symbol is only overridden by archs that have different 856 * requirements than the usual eBPF JITs, f.e. when they only 857 * implement cBPF JIT, do not set images read-only, etc. 858 */ 859 void __weak bpf_jit_free(struct bpf_prog *fp) 860 { 861 if (fp->jited) { 862 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 863 864 bpf_jit_binary_free(hdr); 865 866 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 867 } 868 869 bpf_prog_unlock_free(fp); 870 } 871 872 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 873 const struct bpf_insn *insn, bool extra_pass, 874 u64 *func_addr, bool *func_addr_fixed) 875 { 876 s16 off = insn->off; 877 s32 imm = insn->imm; 878 u8 *addr; 879 880 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 881 if (!*func_addr_fixed) { 882 /* Place-holder address till the last pass has collected 883 * all addresses for JITed subprograms in which case we 884 * can pick them up from prog->aux. 885 */ 886 if (!extra_pass) 887 addr = NULL; 888 else if (prog->aux->func && 889 off >= 0 && off < prog->aux->func_cnt) 890 addr = (u8 *)prog->aux->func[off]->bpf_func; 891 else 892 return -EINVAL; 893 } else { 894 /* Address of a BPF helper call. Since part of the core 895 * kernel, it's always at a fixed location. __bpf_call_base 896 * and the helper with imm relative to it are both in core 897 * kernel. 898 */ 899 addr = (u8 *)__bpf_call_base + imm; 900 } 901 902 *func_addr = (unsigned long)addr; 903 return 0; 904 } 905 906 static int bpf_jit_blind_insn(const struct bpf_insn *from, 907 const struct bpf_insn *aux, 908 struct bpf_insn *to_buff, 909 bool emit_zext) 910 { 911 struct bpf_insn *to = to_buff; 912 u32 imm_rnd = get_random_int(); 913 s16 off; 914 915 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 916 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 917 918 /* Constraints on AX register: 919 * 920 * AX register is inaccessible from user space. It is mapped in 921 * all JITs, and used here for constant blinding rewrites. It is 922 * typically "stateless" meaning its contents are only valid within 923 * the executed instruction, but not across several instructions. 924 * There are a few exceptions however which are further detailed 925 * below. 926 * 927 * Constant blinding is only used by JITs, not in the interpreter. 928 * The interpreter uses AX in some occasions as a local temporary 929 * register e.g. in DIV or MOD instructions. 930 * 931 * In restricted circumstances, the verifier can also use the AX 932 * register for rewrites as long as they do not interfere with 933 * the above cases! 934 */ 935 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 936 goto out; 937 938 if (from->imm == 0 && 939 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 940 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 941 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 942 goto out; 943 } 944 945 switch (from->code) { 946 case BPF_ALU | BPF_ADD | BPF_K: 947 case BPF_ALU | BPF_SUB | BPF_K: 948 case BPF_ALU | BPF_AND | BPF_K: 949 case BPF_ALU | BPF_OR | BPF_K: 950 case BPF_ALU | BPF_XOR | BPF_K: 951 case BPF_ALU | BPF_MUL | BPF_K: 952 case BPF_ALU | BPF_MOV | BPF_K: 953 case BPF_ALU | BPF_DIV | BPF_K: 954 case BPF_ALU | BPF_MOD | BPF_K: 955 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 956 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 957 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 958 break; 959 960 case BPF_ALU64 | BPF_ADD | BPF_K: 961 case BPF_ALU64 | BPF_SUB | BPF_K: 962 case BPF_ALU64 | BPF_AND | BPF_K: 963 case BPF_ALU64 | BPF_OR | BPF_K: 964 case BPF_ALU64 | BPF_XOR | BPF_K: 965 case BPF_ALU64 | BPF_MUL | BPF_K: 966 case BPF_ALU64 | BPF_MOV | BPF_K: 967 case BPF_ALU64 | BPF_DIV | BPF_K: 968 case BPF_ALU64 | BPF_MOD | BPF_K: 969 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 970 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 971 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 972 break; 973 974 case BPF_JMP | BPF_JEQ | BPF_K: 975 case BPF_JMP | BPF_JNE | BPF_K: 976 case BPF_JMP | BPF_JGT | BPF_K: 977 case BPF_JMP | BPF_JLT | BPF_K: 978 case BPF_JMP | BPF_JGE | BPF_K: 979 case BPF_JMP | BPF_JLE | BPF_K: 980 case BPF_JMP | BPF_JSGT | BPF_K: 981 case BPF_JMP | BPF_JSLT | BPF_K: 982 case BPF_JMP | BPF_JSGE | BPF_K: 983 case BPF_JMP | BPF_JSLE | BPF_K: 984 case BPF_JMP | BPF_JSET | BPF_K: 985 /* Accommodate for extra offset in case of a backjump. */ 986 off = from->off; 987 if (off < 0) 988 off -= 2; 989 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 990 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 991 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 992 break; 993 994 case BPF_JMP32 | BPF_JEQ | BPF_K: 995 case BPF_JMP32 | BPF_JNE | BPF_K: 996 case BPF_JMP32 | BPF_JGT | BPF_K: 997 case BPF_JMP32 | BPF_JLT | BPF_K: 998 case BPF_JMP32 | BPF_JGE | BPF_K: 999 case BPF_JMP32 | BPF_JLE | BPF_K: 1000 case BPF_JMP32 | BPF_JSGT | BPF_K: 1001 case BPF_JMP32 | BPF_JSLT | BPF_K: 1002 case BPF_JMP32 | BPF_JSGE | BPF_K: 1003 case BPF_JMP32 | BPF_JSLE | BPF_K: 1004 case BPF_JMP32 | BPF_JSET | BPF_K: 1005 /* Accommodate for extra offset in case of a backjump. */ 1006 off = from->off; 1007 if (off < 0) 1008 off -= 2; 1009 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1010 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1011 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1012 off); 1013 break; 1014 1015 case BPF_LD | BPF_IMM | BPF_DW: 1016 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1017 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1018 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1019 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1020 break; 1021 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1022 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1023 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1024 if (emit_zext) 1025 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1026 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1027 break; 1028 1029 case BPF_ST | BPF_MEM | BPF_DW: 1030 case BPF_ST | BPF_MEM | BPF_W: 1031 case BPF_ST | BPF_MEM | BPF_H: 1032 case BPF_ST | BPF_MEM | BPF_B: 1033 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1034 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1035 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1036 break; 1037 } 1038 out: 1039 return to - to_buff; 1040 } 1041 1042 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1043 gfp_t gfp_extra_flags) 1044 { 1045 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1046 struct bpf_prog *fp; 1047 1048 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 1049 if (fp != NULL) { 1050 /* aux->prog still points to the fp_other one, so 1051 * when promoting the clone to the real program, 1052 * this still needs to be adapted. 1053 */ 1054 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1055 } 1056 1057 return fp; 1058 } 1059 1060 static void bpf_prog_clone_free(struct bpf_prog *fp) 1061 { 1062 /* aux was stolen by the other clone, so we cannot free 1063 * it from this path! It will be freed eventually by the 1064 * other program on release. 1065 * 1066 * At this point, we don't need a deferred release since 1067 * clone is guaranteed to not be locked. 1068 */ 1069 fp->aux = NULL; 1070 __bpf_prog_free(fp); 1071 } 1072 1073 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1074 { 1075 /* We have to repoint aux->prog to self, as we don't 1076 * know whether fp here is the clone or the original. 1077 */ 1078 fp->aux->prog = fp; 1079 bpf_prog_clone_free(fp_other); 1080 } 1081 1082 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1083 { 1084 struct bpf_insn insn_buff[16], aux[2]; 1085 struct bpf_prog *clone, *tmp; 1086 int insn_delta, insn_cnt; 1087 struct bpf_insn *insn; 1088 int i, rewritten; 1089 1090 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 1091 return prog; 1092 1093 clone = bpf_prog_clone_create(prog, GFP_USER); 1094 if (!clone) 1095 return ERR_PTR(-ENOMEM); 1096 1097 insn_cnt = clone->len; 1098 insn = clone->insnsi; 1099 1100 for (i = 0; i < insn_cnt; i++, insn++) { 1101 /* We temporarily need to hold the original ld64 insn 1102 * so that we can still access the first part in the 1103 * second blinding run. 1104 */ 1105 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1106 insn[1].code == 0) 1107 memcpy(aux, insn, sizeof(aux)); 1108 1109 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1110 clone->aux->verifier_zext); 1111 if (!rewritten) 1112 continue; 1113 1114 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1115 if (IS_ERR(tmp)) { 1116 /* Patching may have repointed aux->prog during 1117 * realloc from the original one, so we need to 1118 * fix it up here on error. 1119 */ 1120 bpf_jit_prog_release_other(prog, clone); 1121 return tmp; 1122 } 1123 1124 clone = tmp; 1125 insn_delta = rewritten - 1; 1126 1127 /* Walk new program and skip insns we just inserted. */ 1128 insn = clone->insnsi + i + insn_delta; 1129 insn_cnt += insn_delta; 1130 i += insn_delta; 1131 } 1132 1133 clone->blinded = 1; 1134 return clone; 1135 } 1136 #endif /* CONFIG_BPF_JIT */ 1137 1138 /* Base function for offset calculation. Needs to go into .text section, 1139 * therefore keeping it non-static as well; will also be used by JITs 1140 * anyway later on, so do not let the compiler omit it. This also needs 1141 * to go into kallsyms for correlation from e.g. bpftool, so naming 1142 * must not change. 1143 */ 1144 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1145 { 1146 return 0; 1147 } 1148 EXPORT_SYMBOL_GPL(__bpf_call_base); 1149 1150 /* All UAPI available opcodes. */ 1151 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1152 /* 32 bit ALU operations. */ \ 1153 /* Register based. */ \ 1154 INSN_3(ALU, ADD, X), \ 1155 INSN_3(ALU, SUB, X), \ 1156 INSN_3(ALU, AND, X), \ 1157 INSN_3(ALU, OR, X), \ 1158 INSN_3(ALU, LSH, X), \ 1159 INSN_3(ALU, RSH, X), \ 1160 INSN_3(ALU, XOR, X), \ 1161 INSN_3(ALU, MUL, X), \ 1162 INSN_3(ALU, MOV, X), \ 1163 INSN_3(ALU, ARSH, X), \ 1164 INSN_3(ALU, DIV, X), \ 1165 INSN_3(ALU, MOD, X), \ 1166 INSN_2(ALU, NEG), \ 1167 INSN_3(ALU, END, TO_BE), \ 1168 INSN_3(ALU, END, TO_LE), \ 1169 /* Immediate based. */ \ 1170 INSN_3(ALU, ADD, K), \ 1171 INSN_3(ALU, SUB, K), \ 1172 INSN_3(ALU, AND, K), \ 1173 INSN_3(ALU, OR, K), \ 1174 INSN_3(ALU, LSH, K), \ 1175 INSN_3(ALU, RSH, K), \ 1176 INSN_3(ALU, XOR, K), \ 1177 INSN_3(ALU, MUL, K), \ 1178 INSN_3(ALU, MOV, K), \ 1179 INSN_3(ALU, ARSH, K), \ 1180 INSN_3(ALU, DIV, K), \ 1181 INSN_3(ALU, MOD, K), \ 1182 /* 64 bit ALU operations. */ \ 1183 /* Register based. */ \ 1184 INSN_3(ALU64, ADD, X), \ 1185 INSN_3(ALU64, SUB, X), \ 1186 INSN_3(ALU64, AND, X), \ 1187 INSN_3(ALU64, OR, X), \ 1188 INSN_3(ALU64, LSH, X), \ 1189 INSN_3(ALU64, RSH, X), \ 1190 INSN_3(ALU64, XOR, X), \ 1191 INSN_3(ALU64, MUL, X), \ 1192 INSN_3(ALU64, MOV, X), \ 1193 INSN_3(ALU64, ARSH, X), \ 1194 INSN_3(ALU64, DIV, X), \ 1195 INSN_3(ALU64, MOD, X), \ 1196 INSN_2(ALU64, NEG), \ 1197 /* Immediate based. */ \ 1198 INSN_3(ALU64, ADD, K), \ 1199 INSN_3(ALU64, SUB, K), \ 1200 INSN_3(ALU64, AND, K), \ 1201 INSN_3(ALU64, OR, K), \ 1202 INSN_3(ALU64, LSH, K), \ 1203 INSN_3(ALU64, RSH, K), \ 1204 INSN_3(ALU64, XOR, K), \ 1205 INSN_3(ALU64, MUL, K), \ 1206 INSN_3(ALU64, MOV, K), \ 1207 INSN_3(ALU64, ARSH, K), \ 1208 INSN_3(ALU64, DIV, K), \ 1209 INSN_3(ALU64, MOD, K), \ 1210 /* Call instruction. */ \ 1211 INSN_2(JMP, CALL), \ 1212 /* Exit instruction. */ \ 1213 INSN_2(JMP, EXIT), \ 1214 /* 32-bit Jump instructions. */ \ 1215 /* Register based. */ \ 1216 INSN_3(JMP32, JEQ, X), \ 1217 INSN_3(JMP32, JNE, X), \ 1218 INSN_3(JMP32, JGT, X), \ 1219 INSN_3(JMP32, JLT, X), \ 1220 INSN_3(JMP32, JGE, X), \ 1221 INSN_3(JMP32, JLE, X), \ 1222 INSN_3(JMP32, JSGT, X), \ 1223 INSN_3(JMP32, JSLT, X), \ 1224 INSN_3(JMP32, JSGE, X), \ 1225 INSN_3(JMP32, JSLE, X), \ 1226 INSN_3(JMP32, JSET, X), \ 1227 /* Immediate based. */ \ 1228 INSN_3(JMP32, JEQ, K), \ 1229 INSN_3(JMP32, JNE, K), \ 1230 INSN_3(JMP32, JGT, K), \ 1231 INSN_3(JMP32, JLT, K), \ 1232 INSN_3(JMP32, JGE, K), \ 1233 INSN_3(JMP32, JLE, K), \ 1234 INSN_3(JMP32, JSGT, K), \ 1235 INSN_3(JMP32, JSLT, K), \ 1236 INSN_3(JMP32, JSGE, K), \ 1237 INSN_3(JMP32, JSLE, K), \ 1238 INSN_3(JMP32, JSET, K), \ 1239 /* Jump instructions. */ \ 1240 /* Register based. */ \ 1241 INSN_3(JMP, JEQ, X), \ 1242 INSN_3(JMP, JNE, X), \ 1243 INSN_3(JMP, JGT, X), \ 1244 INSN_3(JMP, JLT, X), \ 1245 INSN_3(JMP, JGE, X), \ 1246 INSN_3(JMP, JLE, X), \ 1247 INSN_3(JMP, JSGT, X), \ 1248 INSN_3(JMP, JSLT, X), \ 1249 INSN_3(JMP, JSGE, X), \ 1250 INSN_3(JMP, JSLE, X), \ 1251 INSN_3(JMP, JSET, X), \ 1252 /* Immediate based. */ \ 1253 INSN_3(JMP, JEQ, K), \ 1254 INSN_3(JMP, JNE, K), \ 1255 INSN_3(JMP, JGT, K), \ 1256 INSN_3(JMP, JLT, K), \ 1257 INSN_3(JMP, JGE, K), \ 1258 INSN_3(JMP, JLE, K), \ 1259 INSN_3(JMP, JSGT, K), \ 1260 INSN_3(JMP, JSLT, K), \ 1261 INSN_3(JMP, JSGE, K), \ 1262 INSN_3(JMP, JSLE, K), \ 1263 INSN_3(JMP, JSET, K), \ 1264 INSN_2(JMP, JA), \ 1265 /* Store instructions. */ \ 1266 /* Register based. */ \ 1267 INSN_3(STX, MEM, B), \ 1268 INSN_3(STX, MEM, H), \ 1269 INSN_3(STX, MEM, W), \ 1270 INSN_3(STX, MEM, DW), \ 1271 INSN_3(STX, XADD, W), \ 1272 INSN_3(STX, XADD, DW), \ 1273 /* Immediate based. */ \ 1274 INSN_3(ST, MEM, B), \ 1275 INSN_3(ST, MEM, H), \ 1276 INSN_3(ST, MEM, W), \ 1277 INSN_3(ST, MEM, DW), \ 1278 /* Load instructions. */ \ 1279 /* Register based. */ \ 1280 INSN_3(LDX, MEM, B), \ 1281 INSN_3(LDX, MEM, H), \ 1282 INSN_3(LDX, MEM, W), \ 1283 INSN_3(LDX, MEM, DW), \ 1284 /* Immediate based. */ \ 1285 INSN_3(LD, IMM, DW) 1286 1287 bool bpf_opcode_in_insntable(u8 code) 1288 { 1289 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1290 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1291 static const bool public_insntable[256] = { 1292 [0 ... 255] = false, 1293 /* Now overwrite non-defaults ... */ 1294 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1295 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1296 [BPF_LD | BPF_ABS | BPF_B] = true, 1297 [BPF_LD | BPF_ABS | BPF_H] = true, 1298 [BPF_LD | BPF_ABS | BPF_W] = true, 1299 [BPF_LD | BPF_IND | BPF_B] = true, 1300 [BPF_LD | BPF_IND | BPF_H] = true, 1301 [BPF_LD | BPF_IND | BPF_W] = true, 1302 }; 1303 #undef BPF_INSN_3_TBL 1304 #undef BPF_INSN_2_TBL 1305 return public_insntable[code]; 1306 } 1307 1308 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1309 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1310 { 1311 memset(dst, 0, size); 1312 return -EFAULT; 1313 } 1314 1315 /** 1316 * __bpf_prog_run - run eBPF program on a given context 1317 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1318 * @insn: is the array of eBPF instructions 1319 * @stack: is the eBPF storage stack 1320 * 1321 * Decode and execute eBPF instructions. 1322 */ 1323 static u64 __no_fgcse ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 1324 { 1325 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1326 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1327 static const void * const jumptable[256] __annotate_jump_table = { 1328 [0 ... 255] = &&default_label, 1329 /* Now overwrite non-defaults ... */ 1330 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1331 /* Non-UAPI available opcodes. */ 1332 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1333 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1334 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1335 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1336 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1337 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1338 }; 1339 #undef BPF_INSN_3_LBL 1340 #undef BPF_INSN_2_LBL 1341 u32 tail_call_cnt = 0; 1342 1343 #define CONT ({ insn++; goto select_insn; }) 1344 #define CONT_JMP ({ insn++; goto select_insn; }) 1345 1346 select_insn: 1347 goto *jumptable[insn->code]; 1348 1349 /* ALU */ 1350 #define ALU(OPCODE, OP) \ 1351 ALU64_##OPCODE##_X: \ 1352 DST = DST OP SRC; \ 1353 CONT; \ 1354 ALU_##OPCODE##_X: \ 1355 DST = (u32) DST OP (u32) SRC; \ 1356 CONT; \ 1357 ALU64_##OPCODE##_K: \ 1358 DST = DST OP IMM; \ 1359 CONT; \ 1360 ALU_##OPCODE##_K: \ 1361 DST = (u32) DST OP (u32) IMM; \ 1362 CONT; 1363 1364 ALU(ADD, +) 1365 ALU(SUB, -) 1366 ALU(AND, &) 1367 ALU(OR, |) 1368 ALU(LSH, <<) 1369 ALU(RSH, >>) 1370 ALU(XOR, ^) 1371 ALU(MUL, *) 1372 #undef ALU 1373 ALU_NEG: 1374 DST = (u32) -DST; 1375 CONT; 1376 ALU64_NEG: 1377 DST = -DST; 1378 CONT; 1379 ALU_MOV_X: 1380 DST = (u32) SRC; 1381 CONT; 1382 ALU_MOV_K: 1383 DST = (u32) IMM; 1384 CONT; 1385 ALU64_MOV_X: 1386 DST = SRC; 1387 CONT; 1388 ALU64_MOV_K: 1389 DST = IMM; 1390 CONT; 1391 LD_IMM_DW: 1392 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1393 insn++; 1394 CONT; 1395 ALU_ARSH_X: 1396 DST = (u64) (u32) (((s32) DST) >> SRC); 1397 CONT; 1398 ALU_ARSH_K: 1399 DST = (u64) (u32) (((s32) DST) >> IMM); 1400 CONT; 1401 ALU64_ARSH_X: 1402 (*(s64 *) &DST) >>= SRC; 1403 CONT; 1404 ALU64_ARSH_K: 1405 (*(s64 *) &DST) >>= IMM; 1406 CONT; 1407 ALU64_MOD_X: 1408 div64_u64_rem(DST, SRC, &AX); 1409 DST = AX; 1410 CONT; 1411 ALU_MOD_X: 1412 AX = (u32) DST; 1413 DST = do_div(AX, (u32) SRC); 1414 CONT; 1415 ALU64_MOD_K: 1416 div64_u64_rem(DST, IMM, &AX); 1417 DST = AX; 1418 CONT; 1419 ALU_MOD_K: 1420 AX = (u32) DST; 1421 DST = do_div(AX, (u32) IMM); 1422 CONT; 1423 ALU64_DIV_X: 1424 DST = div64_u64(DST, SRC); 1425 CONT; 1426 ALU_DIV_X: 1427 AX = (u32) DST; 1428 do_div(AX, (u32) SRC); 1429 DST = (u32) AX; 1430 CONT; 1431 ALU64_DIV_K: 1432 DST = div64_u64(DST, IMM); 1433 CONT; 1434 ALU_DIV_K: 1435 AX = (u32) DST; 1436 do_div(AX, (u32) IMM); 1437 DST = (u32) AX; 1438 CONT; 1439 ALU_END_TO_BE: 1440 switch (IMM) { 1441 case 16: 1442 DST = (__force u16) cpu_to_be16(DST); 1443 break; 1444 case 32: 1445 DST = (__force u32) cpu_to_be32(DST); 1446 break; 1447 case 64: 1448 DST = (__force u64) cpu_to_be64(DST); 1449 break; 1450 } 1451 CONT; 1452 ALU_END_TO_LE: 1453 switch (IMM) { 1454 case 16: 1455 DST = (__force u16) cpu_to_le16(DST); 1456 break; 1457 case 32: 1458 DST = (__force u32) cpu_to_le32(DST); 1459 break; 1460 case 64: 1461 DST = (__force u64) cpu_to_le64(DST); 1462 break; 1463 } 1464 CONT; 1465 1466 /* CALL */ 1467 JMP_CALL: 1468 /* Function call scratches BPF_R1-BPF_R5 registers, 1469 * preserves BPF_R6-BPF_R9, and stores return value 1470 * into BPF_R0. 1471 */ 1472 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1473 BPF_R4, BPF_R5); 1474 CONT; 1475 1476 JMP_CALL_ARGS: 1477 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1478 BPF_R3, BPF_R4, 1479 BPF_R5, 1480 insn + insn->off + 1); 1481 CONT; 1482 1483 JMP_TAIL_CALL: { 1484 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1485 struct bpf_array *array = container_of(map, struct bpf_array, map); 1486 struct bpf_prog *prog; 1487 u32 index = BPF_R3; 1488 1489 if (unlikely(index >= array->map.max_entries)) 1490 goto out; 1491 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1492 goto out; 1493 1494 tail_call_cnt++; 1495 1496 prog = READ_ONCE(array->ptrs[index]); 1497 if (!prog) 1498 goto out; 1499 1500 /* ARG1 at this point is guaranteed to point to CTX from 1501 * the verifier side due to the fact that the tail call is 1502 * handeled like a helper, that is, bpf_tail_call_proto, 1503 * where arg1_type is ARG_PTR_TO_CTX. 1504 */ 1505 insn = prog->insnsi; 1506 goto select_insn; 1507 out: 1508 CONT; 1509 } 1510 JMP_JA: 1511 insn += insn->off; 1512 CONT; 1513 JMP_EXIT: 1514 return BPF_R0; 1515 /* JMP */ 1516 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1517 JMP_##OPCODE##_X: \ 1518 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1519 insn += insn->off; \ 1520 CONT_JMP; \ 1521 } \ 1522 CONT; \ 1523 JMP32_##OPCODE##_X: \ 1524 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1525 insn += insn->off; \ 1526 CONT_JMP; \ 1527 } \ 1528 CONT; \ 1529 JMP_##OPCODE##_K: \ 1530 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1531 insn += insn->off; \ 1532 CONT_JMP; \ 1533 } \ 1534 CONT; \ 1535 JMP32_##OPCODE##_K: \ 1536 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1537 insn += insn->off; \ 1538 CONT_JMP; \ 1539 } \ 1540 CONT; 1541 COND_JMP(u, JEQ, ==) 1542 COND_JMP(u, JNE, !=) 1543 COND_JMP(u, JGT, >) 1544 COND_JMP(u, JLT, <) 1545 COND_JMP(u, JGE, >=) 1546 COND_JMP(u, JLE, <=) 1547 COND_JMP(u, JSET, &) 1548 COND_JMP(s, JSGT, >) 1549 COND_JMP(s, JSLT, <) 1550 COND_JMP(s, JSGE, >=) 1551 COND_JMP(s, JSLE, <=) 1552 #undef COND_JMP 1553 /* STX and ST and LDX*/ 1554 #define LDST(SIZEOP, SIZE) \ 1555 STX_MEM_##SIZEOP: \ 1556 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1557 CONT; \ 1558 ST_MEM_##SIZEOP: \ 1559 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1560 CONT; \ 1561 LDX_MEM_##SIZEOP: \ 1562 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1563 CONT; 1564 1565 LDST(B, u8) 1566 LDST(H, u16) 1567 LDST(W, u32) 1568 LDST(DW, u64) 1569 #undef LDST 1570 #define LDX_PROBE(SIZEOP, SIZE) \ 1571 LDX_PROBE_MEM_##SIZEOP: \ 1572 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) SRC); \ 1573 CONT; 1574 LDX_PROBE(B, 1) 1575 LDX_PROBE(H, 2) 1576 LDX_PROBE(W, 4) 1577 LDX_PROBE(DW, 8) 1578 #undef LDX_PROBE 1579 1580 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1581 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1582 (DST + insn->off)); 1583 CONT; 1584 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1585 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1586 (DST + insn->off)); 1587 CONT; 1588 1589 default_label: 1590 /* If we ever reach this, we have a bug somewhere. Die hard here 1591 * instead of just returning 0; we could be somewhere in a subprog, 1592 * so execution could continue otherwise which we do /not/ want. 1593 * 1594 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1595 */ 1596 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code); 1597 BUG_ON(1); 1598 return 0; 1599 } 1600 1601 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1602 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1603 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1604 { \ 1605 u64 stack[stack_size / sizeof(u64)]; \ 1606 u64 regs[MAX_BPF_EXT_REG]; \ 1607 \ 1608 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1609 ARG1 = (u64) (unsigned long) ctx; \ 1610 return ___bpf_prog_run(regs, insn, stack); \ 1611 } 1612 1613 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1614 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1615 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1616 const struct bpf_insn *insn) \ 1617 { \ 1618 u64 stack[stack_size / sizeof(u64)]; \ 1619 u64 regs[MAX_BPF_EXT_REG]; \ 1620 \ 1621 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1622 BPF_R1 = r1; \ 1623 BPF_R2 = r2; \ 1624 BPF_R3 = r3; \ 1625 BPF_R4 = r4; \ 1626 BPF_R5 = r5; \ 1627 return ___bpf_prog_run(regs, insn, stack); \ 1628 } 1629 1630 #define EVAL1(FN, X) FN(X) 1631 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1632 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1633 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1634 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1635 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1636 1637 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1638 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1639 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1640 1641 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1642 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1643 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1644 1645 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1646 1647 static unsigned int (*interpreters[])(const void *ctx, 1648 const struct bpf_insn *insn) = { 1649 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1650 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1651 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1652 }; 1653 #undef PROG_NAME_LIST 1654 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1655 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1656 const struct bpf_insn *insn) = { 1657 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1658 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1659 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1660 }; 1661 #undef PROG_NAME_LIST 1662 1663 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1664 { 1665 stack_depth = max_t(u32, stack_depth, 1); 1666 insn->off = (s16) insn->imm; 1667 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1668 __bpf_call_base_args; 1669 insn->code = BPF_JMP | BPF_CALL_ARGS; 1670 } 1671 1672 #else 1673 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1674 const struct bpf_insn *insn) 1675 { 1676 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1677 * is not working properly, so warn about it! 1678 */ 1679 WARN_ON_ONCE(1); 1680 return 0; 1681 } 1682 #endif 1683 1684 bool bpf_prog_array_compatible(struct bpf_array *array, 1685 const struct bpf_prog *fp) 1686 { 1687 if (fp->kprobe_override) 1688 return false; 1689 1690 if (!array->owner_prog_type) { 1691 /* There's no owner yet where we could check for 1692 * compatibility. 1693 */ 1694 array->owner_prog_type = fp->type; 1695 array->owner_jited = fp->jited; 1696 1697 return true; 1698 } 1699 1700 return array->owner_prog_type == fp->type && 1701 array->owner_jited == fp->jited; 1702 } 1703 1704 static int bpf_check_tail_call(const struct bpf_prog *fp) 1705 { 1706 struct bpf_prog_aux *aux = fp->aux; 1707 int i; 1708 1709 for (i = 0; i < aux->used_map_cnt; i++) { 1710 struct bpf_map *map = aux->used_maps[i]; 1711 struct bpf_array *array; 1712 1713 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1714 continue; 1715 1716 array = container_of(map, struct bpf_array, map); 1717 if (!bpf_prog_array_compatible(array, fp)) 1718 return -EINVAL; 1719 } 1720 1721 return 0; 1722 } 1723 1724 static void bpf_prog_select_func(struct bpf_prog *fp) 1725 { 1726 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1727 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1728 1729 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1730 #else 1731 fp->bpf_func = __bpf_prog_ret0_warn; 1732 #endif 1733 } 1734 1735 /** 1736 * bpf_prog_select_runtime - select exec runtime for BPF program 1737 * @fp: bpf_prog populated with internal BPF program 1738 * @err: pointer to error variable 1739 * 1740 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1741 * The BPF program will be executed via BPF_PROG_RUN() macro. 1742 */ 1743 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1744 { 1745 /* In case of BPF to BPF calls, verifier did all the prep 1746 * work with regards to JITing, etc. 1747 */ 1748 if (fp->bpf_func) 1749 goto finalize; 1750 1751 bpf_prog_select_func(fp); 1752 1753 /* eBPF JITs can rewrite the program in case constant 1754 * blinding is active. However, in case of error during 1755 * blinding, bpf_int_jit_compile() must always return a 1756 * valid program, which in this case would simply not 1757 * be JITed, but falls back to the interpreter. 1758 */ 1759 if (!bpf_prog_is_dev_bound(fp->aux)) { 1760 *err = bpf_prog_alloc_jited_linfo(fp); 1761 if (*err) 1762 return fp; 1763 1764 fp = bpf_int_jit_compile(fp); 1765 if (!fp->jited) { 1766 bpf_prog_free_jited_linfo(fp); 1767 #ifdef CONFIG_BPF_JIT_ALWAYS_ON 1768 *err = -ENOTSUPP; 1769 return fp; 1770 #endif 1771 } else { 1772 bpf_prog_free_unused_jited_linfo(fp); 1773 } 1774 } else { 1775 *err = bpf_prog_offload_compile(fp); 1776 if (*err) 1777 return fp; 1778 } 1779 1780 finalize: 1781 bpf_prog_lock_ro(fp); 1782 1783 /* The tail call compatibility check can only be done at 1784 * this late stage as we need to determine, if we deal 1785 * with JITed or non JITed program concatenations and not 1786 * all eBPF JITs might immediately support all features. 1787 */ 1788 *err = bpf_check_tail_call(fp); 1789 1790 return fp; 1791 } 1792 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1793 1794 static unsigned int __bpf_prog_ret1(const void *ctx, 1795 const struct bpf_insn *insn) 1796 { 1797 return 1; 1798 } 1799 1800 static struct bpf_prog_dummy { 1801 struct bpf_prog prog; 1802 } dummy_bpf_prog = { 1803 .prog = { 1804 .bpf_func = __bpf_prog_ret1, 1805 }, 1806 }; 1807 1808 /* to avoid allocating empty bpf_prog_array for cgroups that 1809 * don't have bpf program attached use one global 'empty_prog_array' 1810 * It will not be modified the caller of bpf_prog_array_alloc() 1811 * (since caller requested prog_cnt == 0) 1812 * that pointer should be 'freed' by bpf_prog_array_free() 1813 */ 1814 static struct { 1815 struct bpf_prog_array hdr; 1816 struct bpf_prog *null_prog; 1817 } empty_prog_array = { 1818 .null_prog = NULL, 1819 }; 1820 1821 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1822 { 1823 if (prog_cnt) 1824 return kzalloc(sizeof(struct bpf_prog_array) + 1825 sizeof(struct bpf_prog_array_item) * 1826 (prog_cnt + 1), 1827 flags); 1828 1829 return &empty_prog_array.hdr; 1830 } 1831 1832 void bpf_prog_array_free(struct bpf_prog_array *progs) 1833 { 1834 if (!progs || progs == &empty_prog_array.hdr) 1835 return; 1836 kfree_rcu(progs, rcu); 1837 } 1838 1839 int bpf_prog_array_length(struct bpf_prog_array *array) 1840 { 1841 struct bpf_prog_array_item *item; 1842 u32 cnt = 0; 1843 1844 for (item = array->items; item->prog; item++) 1845 if (item->prog != &dummy_bpf_prog.prog) 1846 cnt++; 1847 return cnt; 1848 } 1849 1850 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 1851 { 1852 struct bpf_prog_array_item *item; 1853 1854 for (item = array->items; item->prog; item++) 1855 if (item->prog != &dummy_bpf_prog.prog) 1856 return false; 1857 return true; 1858 } 1859 1860 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 1861 u32 *prog_ids, 1862 u32 request_cnt) 1863 { 1864 struct bpf_prog_array_item *item; 1865 int i = 0; 1866 1867 for (item = array->items; item->prog; item++) { 1868 if (item->prog == &dummy_bpf_prog.prog) 1869 continue; 1870 prog_ids[i] = item->prog->aux->id; 1871 if (++i == request_cnt) { 1872 item++; 1873 break; 1874 } 1875 } 1876 1877 return !!(item->prog); 1878 } 1879 1880 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 1881 __u32 __user *prog_ids, u32 cnt) 1882 { 1883 unsigned long err = 0; 1884 bool nospc; 1885 u32 *ids; 1886 1887 /* users of this function are doing: 1888 * cnt = bpf_prog_array_length(); 1889 * if (cnt > 0) 1890 * bpf_prog_array_copy_to_user(..., cnt); 1891 * so below kcalloc doesn't need extra cnt > 0 check. 1892 */ 1893 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 1894 if (!ids) 1895 return -ENOMEM; 1896 nospc = bpf_prog_array_copy_core(array, ids, cnt); 1897 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 1898 kfree(ids); 1899 if (err) 1900 return -EFAULT; 1901 if (nospc) 1902 return -ENOSPC; 1903 return 0; 1904 } 1905 1906 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 1907 struct bpf_prog *old_prog) 1908 { 1909 struct bpf_prog_array_item *item; 1910 1911 for (item = array->items; item->prog; item++) 1912 if (item->prog == old_prog) { 1913 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 1914 break; 1915 } 1916 } 1917 1918 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 1919 struct bpf_prog *exclude_prog, 1920 struct bpf_prog *include_prog, 1921 struct bpf_prog_array **new_array) 1922 { 1923 int new_prog_cnt, carry_prog_cnt = 0; 1924 struct bpf_prog_array_item *existing; 1925 struct bpf_prog_array *array; 1926 bool found_exclude = false; 1927 int new_prog_idx = 0; 1928 1929 /* Figure out how many existing progs we need to carry over to 1930 * the new array. 1931 */ 1932 if (old_array) { 1933 existing = old_array->items; 1934 for (; existing->prog; existing++) { 1935 if (existing->prog == exclude_prog) { 1936 found_exclude = true; 1937 continue; 1938 } 1939 if (existing->prog != &dummy_bpf_prog.prog) 1940 carry_prog_cnt++; 1941 if (existing->prog == include_prog) 1942 return -EEXIST; 1943 } 1944 } 1945 1946 if (exclude_prog && !found_exclude) 1947 return -ENOENT; 1948 1949 /* How many progs (not NULL) will be in the new array? */ 1950 new_prog_cnt = carry_prog_cnt; 1951 if (include_prog) 1952 new_prog_cnt += 1; 1953 1954 /* Do we have any prog (not NULL) in the new array? */ 1955 if (!new_prog_cnt) { 1956 *new_array = NULL; 1957 return 0; 1958 } 1959 1960 /* +1 as the end of prog_array is marked with NULL */ 1961 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1962 if (!array) 1963 return -ENOMEM; 1964 1965 /* Fill in the new prog array */ 1966 if (carry_prog_cnt) { 1967 existing = old_array->items; 1968 for (; existing->prog; existing++) 1969 if (existing->prog != exclude_prog && 1970 existing->prog != &dummy_bpf_prog.prog) { 1971 array->items[new_prog_idx++].prog = 1972 existing->prog; 1973 } 1974 } 1975 if (include_prog) 1976 array->items[new_prog_idx++].prog = include_prog; 1977 array->items[new_prog_idx].prog = NULL; 1978 *new_array = array; 1979 return 0; 1980 } 1981 1982 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 1983 u32 *prog_ids, u32 request_cnt, 1984 u32 *prog_cnt) 1985 { 1986 u32 cnt = 0; 1987 1988 if (array) 1989 cnt = bpf_prog_array_length(array); 1990 1991 *prog_cnt = cnt; 1992 1993 /* return early if user requested only program count or nothing to copy */ 1994 if (!request_cnt || !cnt) 1995 return 0; 1996 1997 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 1998 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 1999 : 0; 2000 } 2001 2002 static void bpf_prog_free_deferred(struct work_struct *work) 2003 { 2004 struct bpf_prog_aux *aux; 2005 int i; 2006 2007 aux = container_of(work, struct bpf_prog_aux, work); 2008 if (bpf_prog_is_dev_bound(aux)) 2009 bpf_prog_offload_destroy(aux->prog); 2010 #ifdef CONFIG_PERF_EVENTS 2011 if (aux->prog->has_callchain_buf) 2012 put_callchain_buffers(); 2013 #endif 2014 for (i = 0; i < aux->func_cnt; i++) 2015 bpf_jit_free(aux->func[i]); 2016 if (aux->func_cnt) { 2017 kfree(aux->func); 2018 bpf_prog_unlock_free(aux->prog); 2019 } else { 2020 bpf_jit_free(aux->prog); 2021 } 2022 } 2023 2024 /* Free internal BPF program */ 2025 void bpf_prog_free(struct bpf_prog *fp) 2026 { 2027 struct bpf_prog_aux *aux = fp->aux; 2028 2029 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2030 schedule_work(&aux->work); 2031 } 2032 EXPORT_SYMBOL_GPL(bpf_prog_free); 2033 2034 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2035 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2036 2037 void bpf_user_rnd_init_once(void) 2038 { 2039 prandom_init_once(&bpf_user_rnd_state); 2040 } 2041 2042 BPF_CALL_0(bpf_user_rnd_u32) 2043 { 2044 /* Should someone ever have the rather unwise idea to use some 2045 * of the registers passed into this function, then note that 2046 * this function is called from native eBPF and classic-to-eBPF 2047 * transformations. Register assignments from both sides are 2048 * different, f.e. classic always sets fn(ctx, A, X) here. 2049 */ 2050 struct rnd_state *state; 2051 u32 res; 2052 2053 state = &get_cpu_var(bpf_user_rnd_state); 2054 res = prandom_u32_state(state); 2055 put_cpu_var(bpf_user_rnd_state); 2056 2057 return res; 2058 } 2059 2060 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2061 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2062 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2063 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2064 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2065 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2066 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2067 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2068 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2069 2070 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2071 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2072 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2073 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2074 2075 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2076 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2077 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2078 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2079 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2080 2081 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2082 { 2083 return NULL; 2084 } 2085 2086 u64 __weak 2087 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2088 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2089 { 2090 return -ENOTSUPP; 2091 } 2092 EXPORT_SYMBOL_GPL(bpf_event_output); 2093 2094 /* Always built-in helper functions. */ 2095 const struct bpf_func_proto bpf_tail_call_proto = { 2096 .func = NULL, 2097 .gpl_only = false, 2098 .ret_type = RET_VOID, 2099 .arg1_type = ARG_PTR_TO_CTX, 2100 .arg2_type = ARG_CONST_MAP_PTR, 2101 .arg3_type = ARG_ANYTHING, 2102 }; 2103 2104 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2105 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2106 * eBPF and implicitly also cBPF can get JITed! 2107 */ 2108 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2109 { 2110 return prog; 2111 } 2112 2113 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2114 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2115 */ 2116 void __weak bpf_jit_compile(struct bpf_prog *prog) 2117 { 2118 } 2119 2120 bool __weak bpf_helper_changes_pkt_data(void *func) 2121 { 2122 return false; 2123 } 2124 2125 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2126 * analysis code and wants explicit zero extension inserted by verifier. 2127 * Otherwise, return FALSE. 2128 */ 2129 bool __weak bpf_jit_needs_zext(void) 2130 { 2131 return false; 2132 } 2133 2134 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2135 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2136 */ 2137 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2138 int len) 2139 { 2140 return -EFAULT; 2141 } 2142 2143 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2144 EXPORT_SYMBOL(bpf_stats_enabled_key); 2145 2146 /* All definitions of tracepoints related to BPF. */ 2147 #define CREATE_TRACE_POINTS 2148 #include <linux/bpf_trace.h> 2149 2150 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2151 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2152