1 /* 2 * Linux Socket Filter - Kernel level socket filtering 3 * 4 * Based on the design of the Berkeley Packet Filter. The new 5 * internal format has been designed by PLUMgrid: 6 * 7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 8 * 9 * Authors: 10 * 11 * Jay Schulist <jschlst@samba.org> 12 * Alexei Starovoitov <ast@plumgrid.com> 13 * Daniel Borkmann <dborkman@redhat.com> 14 * 15 * This program is free software; you can redistribute it and/or 16 * modify it under the terms of the GNU General Public License 17 * as published by the Free Software Foundation; either version 18 * 2 of the License, or (at your option) any later version. 19 * 20 * Andi Kleen - Fix a few bad bugs and races. 21 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 22 */ 23 24 #include <uapi/linux/btf.h> 25 #include <linux/filter.h> 26 #include <linux/skbuff.h> 27 #include <linux/vmalloc.h> 28 #include <linux/random.h> 29 #include <linux/moduleloader.h> 30 #include <linux/bpf.h> 31 #include <linux/btf.h> 32 #include <linux/frame.h> 33 #include <linux/rbtree_latch.h> 34 #include <linux/kallsyms.h> 35 #include <linux/rcupdate.h> 36 #include <linux/perf_event.h> 37 38 #include <asm/unaligned.h> 39 40 /* Registers */ 41 #define BPF_R0 regs[BPF_REG_0] 42 #define BPF_R1 regs[BPF_REG_1] 43 #define BPF_R2 regs[BPF_REG_2] 44 #define BPF_R3 regs[BPF_REG_3] 45 #define BPF_R4 regs[BPF_REG_4] 46 #define BPF_R5 regs[BPF_REG_5] 47 #define BPF_R6 regs[BPF_REG_6] 48 #define BPF_R7 regs[BPF_REG_7] 49 #define BPF_R8 regs[BPF_REG_8] 50 #define BPF_R9 regs[BPF_REG_9] 51 #define BPF_R10 regs[BPF_REG_10] 52 53 /* Named registers */ 54 #define DST regs[insn->dst_reg] 55 #define SRC regs[insn->src_reg] 56 #define FP regs[BPF_REG_FP] 57 #define AX regs[BPF_REG_AX] 58 #define ARG1 regs[BPF_REG_ARG1] 59 #define CTX regs[BPF_REG_CTX] 60 #define IMM insn->imm 61 62 /* No hurry in this branch 63 * 64 * Exported for the bpf jit load helper. 65 */ 66 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 67 { 68 u8 *ptr = NULL; 69 70 if (k >= SKF_NET_OFF) 71 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 72 else if (k >= SKF_LL_OFF) 73 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 74 75 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 76 return ptr; 77 78 return NULL; 79 } 80 81 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 82 { 83 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 84 struct bpf_prog_aux *aux; 85 struct bpf_prog *fp; 86 87 size = round_up(size, PAGE_SIZE); 88 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 89 if (fp == NULL) 90 return NULL; 91 92 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags); 93 if (aux == NULL) { 94 vfree(fp); 95 return NULL; 96 } 97 98 fp->pages = size / PAGE_SIZE; 99 fp->aux = aux; 100 fp->aux->prog = fp; 101 fp->jit_requested = ebpf_jit_enabled(); 102 103 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode); 104 105 return fp; 106 } 107 108 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 109 { 110 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 111 struct bpf_prog *prog; 112 int cpu; 113 114 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 115 if (!prog) 116 return NULL; 117 118 prog->aux->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 119 if (!prog->aux->stats) { 120 kfree(prog->aux); 121 vfree(prog); 122 return NULL; 123 } 124 125 for_each_possible_cpu(cpu) { 126 struct bpf_prog_stats *pstats; 127 128 pstats = per_cpu_ptr(prog->aux->stats, cpu); 129 u64_stats_init(&pstats->syncp); 130 } 131 return prog; 132 } 133 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 134 135 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 136 { 137 if (!prog->aux->nr_linfo || !prog->jit_requested) 138 return 0; 139 140 prog->aux->jited_linfo = kcalloc(prog->aux->nr_linfo, 141 sizeof(*prog->aux->jited_linfo), 142 GFP_KERNEL | __GFP_NOWARN); 143 if (!prog->aux->jited_linfo) 144 return -ENOMEM; 145 146 return 0; 147 } 148 149 void bpf_prog_free_jited_linfo(struct bpf_prog *prog) 150 { 151 kfree(prog->aux->jited_linfo); 152 prog->aux->jited_linfo = NULL; 153 } 154 155 void bpf_prog_free_unused_jited_linfo(struct bpf_prog *prog) 156 { 157 if (prog->aux->jited_linfo && !prog->aux->jited_linfo[0]) 158 bpf_prog_free_jited_linfo(prog); 159 } 160 161 /* The jit engine is responsible to provide an array 162 * for insn_off to the jited_off mapping (insn_to_jit_off). 163 * 164 * The idx to this array is the insn_off. Hence, the insn_off 165 * here is relative to the prog itself instead of the main prog. 166 * This array has one entry for each xlated bpf insn. 167 * 168 * jited_off is the byte off to the last byte of the jited insn. 169 * 170 * Hence, with 171 * insn_start: 172 * The first bpf insn off of the prog. The insn off 173 * here is relative to the main prog. 174 * e.g. if prog is a subprog, insn_start > 0 175 * linfo_idx: 176 * The prog's idx to prog->aux->linfo and jited_linfo 177 * 178 * jited_linfo[linfo_idx] = prog->bpf_func 179 * 180 * For i > linfo_idx, 181 * 182 * jited_linfo[i] = prog->bpf_func + 183 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 184 */ 185 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 186 const u32 *insn_to_jit_off) 187 { 188 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 189 const struct bpf_line_info *linfo; 190 void **jited_linfo; 191 192 if (!prog->aux->jited_linfo) 193 /* Userspace did not provide linfo */ 194 return; 195 196 linfo_idx = prog->aux->linfo_idx; 197 linfo = &prog->aux->linfo[linfo_idx]; 198 insn_start = linfo[0].insn_off; 199 insn_end = insn_start + prog->len; 200 201 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 202 jited_linfo[0] = prog->bpf_func; 203 204 nr_linfo = prog->aux->nr_linfo - linfo_idx; 205 206 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 207 /* The verifier ensures that linfo[i].insn_off is 208 * strictly increasing 209 */ 210 jited_linfo[i] = prog->bpf_func + 211 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 212 } 213 214 void bpf_prog_free_linfo(struct bpf_prog *prog) 215 { 216 bpf_prog_free_jited_linfo(prog); 217 kvfree(prog->aux->linfo); 218 } 219 220 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 221 gfp_t gfp_extra_flags) 222 { 223 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 224 struct bpf_prog *fp; 225 u32 pages, delta; 226 int ret; 227 228 BUG_ON(fp_old == NULL); 229 230 size = round_up(size, PAGE_SIZE); 231 pages = size / PAGE_SIZE; 232 if (pages <= fp_old->pages) 233 return fp_old; 234 235 delta = pages - fp_old->pages; 236 ret = __bpf_prog_charge(fp_old->aux->user, delta); 237 if (ret) 238 return NULL; 239 240 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 241 if (fp == NULL) { 242 __bpf_prog_uncharge(fp_old->aux->user, delta); 243 } else { 244 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 245 fp->pages = pages; 246 fp->aux->prog = fp; 247 248 /* We keep fp->aux from fp_old around in the new 249 * reallocated structure. 250 */ 251 fp_old->aux = NULL; 252 __bpf_prog_free(fp_old); 253 } 254 255 return fp; 256 } 257 258 void __bpf_prog_free(struct bpf_prog *fp) 259 { 260 if (fp->aux) { 261 free_percpu(fp->aux->stats); 262 kfree(fp->aux); 263 } 264 vfree(fp); 265 } 266 267 int bpf_prog_calc_tag(struct bpf_prog *fp) 268 { 269 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64); 270 u32 raw_size = bpf_prog_tag_scratch_size(fp); 271 u32 digest[SHA_DIGEST_WORDS]; 272 u32 ws[SHA_WORKSPACE_WORDS]; 273 u32 i, bsize, psize, blocks; 274 struct bpf_insn *dst; 275 bool was_ld_map; 276 u8 *raw, *todo; 277 __be32 *result; 278 __be64 *bits; 279 280 raw = vmalloc(raw_size); 281 if (!raw) 282 return -ENOMEM; 283 284 sha_init(digest); 285 memset(ws, 0, sizeof(ws)); 286 287 /* We need to take out the map fd for the digest calculation 288 * since they are unstable from user space side. 289 */ 290 dst = (void *)raw; 291 for (i = 0, was_ld_map = false; i < fp->len; i++) { 292 dst[i] = fp->insnsi[i]; 293 if (!was_ld_map && 294 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 295 dst[i].src_reg == BPF_PSEUDO_MAP_FD) { 296 was_ld_map = true; 297 dst[i].imm = 0; 298 } else if (was_ld_map && 299 dst[i].code == 0 && 300 dst[i].dst_reg == 0 && 301 dst[i].src_reg == 0 && 302 dst[i].off == 0) { 303 was_ld_map = false; 304 dst[i].imm = 0; 305 } else { 306 was_ld_map = false; 307 } 308 } 309 310 psize = bpf_prog_insn_size(fp); 311 memset(&raw[psize], 0, raw_size - psize); 312 raw[psize++] = 0x80; 313 314 bsize = round_up(psize, SHA_MESSAGE_BYTES); 315 blocks = bsize / SHA_MESSAGE_BYTES; 316 todo = raw; 317 if (bsize - psize >= sizeof(__be64)) { 318 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 319 } else { 320 bits = (__be64 *)(todo + bsize + bits_offset); 321 blocks++; 322 } 323 *bits = cpu_to_be64((psize - 1) << 3); 324 325 while (blocks--) { 326 sha_transform(digest, todo, ws); 327 todo += SHA_MESSAGE_BYTES; 328 } 329 330 result = (__force __be32 *)digest; 331 for (i = 0; i < SHA_DIGEST_WORDS; i++) 332 result[i] = cpu_to_be32(digest[i]); 333 memcpy(fp->tag, result, sizeof(fp->tag)); 334 335 vfree(raw); 336 return 0; 337 } 338 339 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 340 s32 end_new, u32 curr, const bool probe_pass) 341 { 342 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 343 s32 delta = end_new - end_old; 344 s64 imm = insn->imm; 345 346 if (curr < pos && curr + imm + 1 >= end_old) 347 imm += delta; 348 else if (curr >= end_new && curr + imm + 1 < end_new) 349 imm -= delta; 350 if (imm < imm_min || imm > imm_max) 351 return -ERANGE; 352 if (!probe_pass) 353 insn->imm = imm; 354 return 0; 355 } 356 357 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 358 s32 end_new, u32 curr, const bool probe_pass) 359 { 360 const s32 off_min = S16_MIN, off_max = S16_MAX; 361 s32 delta = end_new - end_old; 362 s32 off = insn->off; 363 364 if (curr < pos && curr + off + 1 >= end_old) 365 off += delta; 366 else if (curr >= end_new && curr + off + 1 < end_new) 367 off -= delta; 368 if (off < off_min || off > off_max) 369 return -ERANGE; 370 if (!probe_pass) 371 insn->off = off; 372 return 0; 373 } 374 375 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 376 s32 end_new, const bool probe_pass) 377 { 378 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 379 struct bpf_insn *insn = prog->insnsi; 380 int ret = 0; 381 382 for (i = 0; i < insn_cnt; i++, insn++) { 383 u8 code; 384 385 /* In the probing pass we still operate on the original, 386 * unpatched image in order to check overflows before we 387 * do any other adjustments. Therefore skip the patchlet. 388 */ 389 if (probe_pass && i == pos) { 390 i = end_new; 391 insn = prog->insnsi + end_old; 392 } 393 code = insn->code; 394 if ((BPF_CLASS(code) != BPF_JMP && 395 BPF_CLASS(code) != BPF_JMP32) || 396 BPF_OP(code) == BPF_EXIT) 397 continue; 398 /* Adjust offset of jmps if we cross patch boundaries. */ 399 if (BPF_OP(code) == BPF_CALL) { 400 if (insn->src_reg != BPF_PSEUDO_CALL) 401 continue; 402 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 403 end_new, i, probe_pass); 404 } else { 405 ret = bpf_adj_delta_to_off(insn, pos, end_old, 406 end_new, i, probe_pass); 407 } 408 if (ret) 409 break; 410 } 411 412 return ret; 413 } 414 415 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 416 { 417 struct bpf_line_info *linfo; 418 u32 i, nr_linfo; 419 420 nr_linfo = prog->aux->nr_linfo; 421 if (!nr_linfo || !delta) 422 return; 423 424 linfo = prog->aux->linfo; 425 426 for (i = 0; i < nr_linfo; i++) 427 if (off < linfo[i].insn_off) 428 break; 429 430 /* Push all off < linfo[i].insn_off by delta */ 431 for (; i < nr_linfo; i++) 432 linfo[i].insn_off += delta; 433 } 434 435 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 436 const struct bpf_insn *patch, u32 len) 437 { 438 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 439 const u32 cnt_max = S16_MAX; 440 struct bpf_prog *prog_adj; 441 442 /* Since our patchlet doesn't expand the image, we're done. */ 443 if (insn_delta == 0) { 444 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 445 return prog; 446 } 447 448 insn_adj_cnt = prog->len + insn_delta; 449 450 /* Reject anything that would potentially let the insn->off 451 * target overflow when we have excessive program expansions. 452 * We need to probe here before we do any reallocation where 453 * we afterwards may not fail anymore. 454 */ 455 if (insn_adj_cnt > cnt_max && 456 bpf_adj_branches(prog, off, off + 1, off + len, true)) 457 return NULL; 458 459 /* Several new instructions need to be inserted. Make room 460 * for them. Likely, there's no need for a new allocation as 461 * last page could have large enough tailroom. 462 */ 463 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 464 GFP_USER); 465 if (!prog_adj) 466 return NULL; 467 468 prog_adj->len = insn_adj_cnt; 469 470 /* Patching happens in 3 steps: 471 * 472 * 1) Move over tail of insnsi from next instruction onwards, 473 * so we can patch the single target insn with one or more 474 * new ones (patching is always from 1 to n insns, n > 0). 475 * 2) Inject new instructions at the target location. 476 * 3) Adjust branch offsets if necessary. 477 */ 478 insn_rest = insn_adj_cnt - off - len; 479 480 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 481 sizeof(*patch) * insn_rest); 482 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 483 484 /* We are guaranteed to not fail at this point, otherwise 485 * the ship has sailed to reverse to the original state. An 486 * overflow cannot happen at this point. 487 */ 488 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 489 490 bpf_adj_linfo(prog_adj, off, insn_delta); 491 492 return prog_adj; 493 } 494 495 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 496 { 497 /* Branch offsets can't overflow when program is shrinking, no need 498 * to call bpf_adj_branches(..., true) here 499 */ 500 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 501 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 502 prog->len -= cnt; 503 504 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 505 } 506 507 void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 508 { 509 int i; 510 511 for (i = 0; i < fp->aux->func_cnt; i++) 512 bpf_prog_kallsyms_del(fp->aux->func[i]); 513 } 514 515 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 516 { 517 bpf_prog_kallsyms_del_subprogs(fp); 518 bpf_prog_kallsyms_del(fp); 519 } 520 521 #ifdef CONFIG_BPF_JIT 522 /* All BPF JIT sysctl knobs here. */ 523 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_ALWAYS_ON); 524 int bpf_jit_harden __read_mostly; 525 int bpf_jit_kallsyms __read_mostly; 526 long bpf_jit_limit __read_mostly; 527 528 static __always_inline void 529 bpf_get_prog_addr_region(const struct bpf_prog *prog, 530 unsigned long *symbol_start, 531 unsigned long *symbol_end) 532 { 533 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 534 unsigned long addr = (unsigned long)hdr; 535 536 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 537 538 *symbol_start = addr; 539 *symbol_end = addr + hdr->pages * PAGE_SIZE; 540 } 541 542 void bpf_get_prog_name(const struct bpf_prog *prog, char *sym) 543 { 544 const char *end = sym + KSYM_NAME_LEN; 545 const struct btf_type *type; 546 const char *func_name; 547 548 BUILD_BUG_ON(sizeof("bpf_prog_") + 549 sizeof(prog->tag) * 2 + 550 /* name has been null terminated. 551 * We should need +1 for the '_' preceding 552 * the name. However, the null character 553 * is double counted between the name and the 554 * sizeof("bpf_prog_") above, so we omit 555 * the +1 here. 556 */ 557 sizeof(prog->aux->name) > KSYM_NAME_LEN); 558 559 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 560 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 561 562 /* prog->aux->name will be ignored if full btf name is available */ 563 if (prog->aux->func_info_cnt) { 564 type = btf_type_by_id(prog->aux->btf, 565 prog->aux->func_info[prog->aux->func_idx].type_id); 566 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 567 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 568 return; 569 } 570 571 if (prog->aux->name[0]) 572 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 573 else 574 *sym = 0; 575 } 576 577 static __always_inline unsigned long 578 bpf_get_prog_addr_start(struct latch_tree_node *n) 579 { 580 unsigned long symbol_start, symbol_end; 581 const struct bpf_prog_aux *aux; 582 583 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 584 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 585 586 return symbol_start; 587 } 588 589 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 590 struct latch_tree_node *b) 591 { 592 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b); 593 } 594 595 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 596 { 597 unsigned long val = (unsigned long)key; 598 unsigned long symbol_start, symbol_end; 599 const struct bpf_prog_aux *aux; 600 601 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 602 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 603 604 if (val < symbol_start) 605 return -1; 606 if (val >= symbol_end) 607 return 1; 608 609 return 0; 610 } 611 612 static const struct latch_tree_ops bpf_tree_ops = { 613 .less = bpf_tree_less, 614 .comp = bpf_tree_comp, 615 }; 616 617 static DEFINE_SPINLOCK(bpf_lock); 618 static LIST_HEAD(bpf_kallsyms); 619 static struct latch_tree_root bpf_tree __cacheline_aligned; 620 621 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux) 622 { 623 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode)); 624 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms); 625 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 626 } 627 628 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux) 629 { 630 if (list_empty(&aux->ksym_lnode)) 631 return; 632 633 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 634 list_del_rcu(&aux->ksym_lnode); 635 } 636 637 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 638 { 639 return fp->jited && !bpf_prog_was_classic(fp); 640 } 641 642 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 643 { 644 return list_empty(&fp->aux->ksym_lnode) || 645 fp->aux->ksym_lnode.prev == LIST_POISON2; 646 } 647 648 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 649 { 650 if (!bpf_prog_kallsyms_candidate(fp) || 651 !capable(CAP_SYS_ADMIN)) 652 return; 653 654 spin_lock_bh(&bpf_lock); 655 bpf_prog_ksym_node_add(fp->aux); 656 spin_unlock_bh(&bpf_lock); 657 } 658 659 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 660 { 661 if (!bpf_prog_kallsyms_candidate(fp)) 662 return; 663 664 spin_lock_bh(&bpf_lock); 665 bpf_prog_ksym_node_del(fp->aux); 666 spin_unlock_bh(&bpf_lock); 667 } 668 669 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr) 670 { 671 struct latch_tree_node *n; 672 673 if (!bpf_jit_kallsyms_enabled()) 674 return NULL; 675 676 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 677 return n ? 678 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 679 NULL; 680 } 681 682 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 683 unsigned long *off, char *sym) 684 { 685 unsigned long symbol_start, symbol_end; 686 struct bpf_prog *prog; 687 char *ret = NULL; 688 689 rcu_read_lock(); 690 prog = bpf_prog_kallsyms_find(addr); 691 if (prog) { 692 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 693 bpf_get_prog_name(prog, sym); 694 695 ret = sym; 696 if (size) 697 *size = symbol_end - symbol_start; 698 if (off) 699 *off = addr - symbol_start; 700 } 701 rcu_read_unlock(); 702 703 return ret; 704 } 705 706 bool is_bpf_text_address(unsigned long addr) 707 { 708 bool ret; 709 710 rcu_read_lock(); 711 ret = bpf_prog_kallsyms_find(addr) != NULL; 712 rcu_read_unlock(); 713 714 return ret; 715 } 716 717 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 718 char *sym) 719 { 720 struct bpf_prog_aux *aux; 721 unsigned int it = 0; 722 int ret = -ERANGE; 723 724 if (!bpf_jit_kallsyms_enabled()) 725 return ret; 726 727 rcu_read_lock(); 728 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 729 if (it++ != symnum) 730 continue; 731 732 bpf_get_prog_name(aux->prog, sym); 733 734 *value = (unsigned long)aux->prog->bpf_func; 735 *type = BPF_SYM_ELF_TYPE; 736 737 ret = 0; 738 break; 739 } 740 rcu_read_unlock(); 741 742 return ret; 743 } 744 745 static atomic_long_t bpf_jit_current; 746 747 /* Can be overridden by an arch's JIT compiler if it has a custom, 748 * dedicated BPF backend memory area, or if neither of the two 749 * below apply. 750 */ 751 u64 __weak bpf_jit_alloc_exec_limit(void) 752 { 753 #if defined(MODULES_VADDR) 754 return MODULES_END - MODULES_VADDR; 755 #else 756 return VMALLOC_END - VMALLOC_START; 757 #endif 758 } 759 760 static int __init bpf_jit_charge_init(void) 761 { 762 /* Only used as heuristic here to derive limit. */ 763 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2, 764 PAGE_SIZE), LONG_MAX); 765 return 0; 766 } 767 pure_initcall(bpf_jit_charge_init); 768 769 static int bpf_jit_charge_modmem(u32 pages) 770 { 771 if (atomic_long_add_return(pages, &bpf_jit_current) > 772 (bpf_jit_limit >> PAGE_SHIFT)) { 773 if (!capable(CAP_SYS_ADMIN)) { 774 atomic_long_sub(pages, &bpf_jit_current); 775 return -EPERM; 776 } 777 } 778 779 return 0; 780 } 781 782 static void bpf_jit_uncharge_modmem(u32 pages) 783 { 784 atomic_long_sub(pages, &bpf_jit_current); 785 } 786 787 void *__weak bpf_jit_alloc_exec(unsigned long size) 788 { 789 return module_alloc(size); 790 } 791 792 void __weak bpf_jit_free_exec(void *addr) 793 { 794 module_memfree(addr); 795 } 796 797 struct bpf_binary_header * 798 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 799 unsigned int alignment, 800 bpf_jit_fill_hole_t bpf_fill_ill_insns) 801 { 802 struct bpf_binary_header *hdr; 803 u32 size, hole, start, pages; 804 805 /* Most of BPF filters are really small, but if some of them 806 * fill a page, allow at least 128 extra bytes to insert a 807 * random section of illegal instructions. 808 */ 809 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 810 pages = size / PAGE_SIZE; 811 812 if (bpf_jit_charge_modmem(pages)) 813 return NULL; 814 hdr = bpf_jit_alloc_exec(size); 815 if (!hdr) { 816 bpf_jit_uncharge_modmem(pages); 817 return NULL; 818 } 819 820 /* Fill space with illegal/arch-dep instructions. */ 821 bpf_fill_ill_insns(hdr, size); 822 823 hdr->pages = pages; 824 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 825 PAGE_SIZE - sizeof(*hdr)); 826 start = (get_random_int() % hole) & ~(alignment - 1); 827 828 /* Leave a random number of instructions before BPF code. */ 829 *image_ptr = &hdr->image[start]; 830 831 return hdr; 832 } 833 834 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 835 { 836 u32 pages = hdr->pages; 837 838 bpf_jit_free_exec(hdr); 839 bpf_jit_uncharge_modmem(pages); 840 } 841 842 /* This symbol is only overridden by archs that have different 843 * requirements than the usual eBPF JITs, f.e. when they only 844 * implement cBPF JIT, do not set images read-only, etc. 845 */ 846 void __weak bpf_jit_free(struct bpf_prog *fp) 847 { 848 if (fp->jited) { 849 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 850 851 bpf_jit_binary_free(hdr); 852 853 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 854 } 855 856 bpf_prog_unlock_free(fp); 857 } 858 859 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 860 const struct bpf_insn *insn, bool extra_pass, 861 u64 *func_addr, bool *func_addr_fixed) 862 { 863 s16 off = insn->off; 864 s32 imm = insn->imm; 865 u8 *addr; 866 867 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 868 if (!*func_addr_fixed) { 869 /* Place-holder address till the last pass has collected 870 * all addresses for JITed subprograms in which case we 871 * can pick them up from prog->aux. 872 */ 873 if (!extra_pass) 874 addr = NULL; 875 else if (prog->aux->func && 876 off >= 0 && off < prog->aux->func_cnt) 877 addr = (u8 *)prog->aux->func[off]->bpf_func; 878 else 879 return -EINVAL; 880 } else { 881 /* Address of a BPF helper call. Since part of the core 882 * kernel, it's always at a fixed location. __bpf_call_base 883 * and the helper with imm relative to it are both in core 884 * kernel. 885 */ 886 addr = (u8 *)__bpf_call_base + imm; 887 } 888 889 *func_addr = (unsigned long)addr; 890 return 0; 891 } 892 893 static int bpf_jit_blind_insn(const struct bpf_insn *from, 894 const struct bpf_insn *aux, 895 struct bpf_insn *to_buff) 896 { 897 struct bpf_insn *to = to_buff; 898 u32 imm_rnd = get_random_int(); 899 s16 off; 900 901 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 902 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 903 904 /* Constraints on AX register: 905 * 906 * AX register is inaccessible from user space. It is mapped in 907 * all JITs, and used here for constant blinding rewrites. It is 908 * typically "stateless" meaning its contents are only valid within 909 * the executed instruction, but not across several instructions. 910 * There are a few exceptions however which are further detailed 911 * below. 912 * 913 * Constant blinding is only used by JITs, not in the interpreter. 914 * The interpreter uses AX in some occasions as a local temporary 915 * register e.g. in DIV or MOD instructions. 916 * 917 * In restricted circumstances, the verifier can also use the AX 918 * register for rewrites as long as they do not interfere with 919 * the above cases! 920 */ 921 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 922 goto out; 923 924 if (from->imm == 0 && 925 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 926 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 927 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 928 goto out; 929 } 930 931 switch (from->code) { 932 case BPF_ALU | BPF_ADD | BPF_K: 933 case BPF_ALU | BPF_SUB | BPF_K: 934 case BPF_ALU | BPF_AND | BPF_K: 935 case BPF_ALU | BPF_OR | BPF_K: 936 case BPF_ALU | BPF_XOR | BPF_K: 937 case BPF_ALU | BPF_MUL | BPF_K: 938 case BPF_ALU | BPF_MOV | BPF_K: 939 case BPF_ALU | BPF_DIV | BPF_K: 940 case BPF_ALU | BPF_MOD | BPF_K: 941 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 942 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 943 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 944 break; 945 946 case BPF_ALU64 | BPF_ADD | BPF_K: 947 case BPF_ALU64 | BPF_SUB | BPF_K: 948 case BPF_ALU64 | BPF_AND | BPF_K: 949 case BPF_ALU64 | BPF_OR | BPF_K: 950 case BPF_ALU64 | BPF_XOR | BPF_K: 951 case BPF_ALU64 | BPF_MUL | BPF_K: 952 case BPF_ALU64 | BPF_MOV | BPF_K: 953 case BPF_ALU64 | BPF_DIV | BPF_K: 954 case BPF_ALU64 | BPF_MOD | BPF_K: 955 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 956 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 957 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 958 break; 959 960 case BPF_JMP | BPF_JEQ | BPF_K: 961 case BPF_JMP | BPF_JNE | BPF_K: 962 case BPF_JMP | BPF_JGT | BPF_K: 963 case BPF_JMP | BPF_JLT | BPF_K: 964 case BPF_JMP | BPF_JGE | BPF_K: 965 case BPF_JMP | BPF_JLE | BPF_K: 966 case BPF_JMP | BPF_JSGT | BPF_K: 967 case BPF_JMP | BPF_JSLT | BPF_K: 968 case BPF_JMP | BPF_JSGE | BPF_K: 969 case BPF_JMP | BPF_JSLE | BPF_K: 970 case BPF_JMP | BPF_JSET | BPF_K: 971 /* Accommodate for extra offset in case of a backjump. */ 972 off = from->off; 973 if (off < 0) 974 off -= 2; 975 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 976 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 977 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 978 break; 979 980 case BPF_JMP32 | BPF_JEQ | BPF_K: 981 case BPF_JMP32 | BPF_JNE | BPF_K: 982 case BPF_JMP32 | BPF_JGT | BPF_K: 983 case BPF_JMP32 | BPF_JLT | BPF_K: 984 case BPF_JMP32 | BPF_JGE | BPF_K: 985 case BPF_JMP32 | BPF_JLE | BPF_K: 986 case BPF_JMP32 | BPF_JSGT | BPF_K: 987 case BPF_JMP32 | BPF_JSLT | BPF_K: 988 case BPF_JMP32 | BPF_JSGE | BPF_K: 989 case BPF_JMP32 | BPF_JSLE | BPF_K: 990 case BPF_JMP32 | BPF_JSET | BPF_K: 991 /* Accommodate for extra offset in case of a backjump. */ 992 off = from->off; 993 if (off < 0) 994 off -= 2; 995 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 996 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 997 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 998 off); 999 break; 1000 1001 case BPF_LD | BPF_IMM | BPF_DW: 1002 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1003 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1004 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1005 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1006 break; 1007 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1008 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1009 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1010 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1011 break; 1012 1013 case BPF_ST | BPF_MEM | BPF_DW: 1014 case BPF_ST | BPF_MEM | BPF_W: 1015 case BPF_ST | BPF_MEM | BPF_H: 1016 case BPF_ST | BPF_MEM | BPF_B: 1017 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1018 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1019 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1020 break; 1021 } 1022 out: 1023 return to - to_buff; 1024 } 1025 1026 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1027 gfp_t gfp_extra_flags) 1028 { 1029 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1030 struct bpf_prog *fp; 1031 1032 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 1033 if (fp != NULL) { 1034 /* aux->prog still points to the fp_other one, so 1035 * when promoting the clone to the real program, 1036 * this still needs to be adapted. 1037 */ 1038 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1039 } 1040 1041 return fp; 1042 } 1043 1044 static void bpf_prog_clone_free(struct bpf_prog *fp) 1045 { 1046 /* aux was stolen by the other clone, so we cannot free 1047 * it from this path! It will be freed eventually by the 1048 * other program on release. 1049 * 1050 * At this point, we don't need a deferred release since 1051 * clone is guaranteed to not be locked. 1052 */ 1053 fp->aux = NULL; 1054 __bpf_prog_free(fp); 1055 } 1056 1057 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1058 { 1059 /* We have to repoint aux->prog to self, as we don't 1060 * know whether fp here is the clone or the original. 1061 */ 1062 fp->aux->prog = fp; 1063 bpf_prog_clone_free(fp_other); 1064 } 1065 1066 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1067 { 1068 struct bpf_insn insn_buff[16], aux[2]; 1069 struct bpf_prog *clone, *tmp; 1070 int insn_delta, insn_cnt; 1071 struct bpf_insn *insn; 1072 int i, rewritten; 1073 1074 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 1075 return prog; 1076 1077 clone = bpf_prog_clone_create(prog, GFP_USER); 1078 if (!clone) 1079 return ERR_PTR(-ENOMEM); 1080 1081 insn_cnt = clone->len; 1082 insn = clone->insnsi; 1083 1084 for (i = 0; i < insn_cnt; i++, insn++) { 1085 /* We temporarily need to hold the original ld64 insn 1086 * so that we can still access the first part in the 1087 * second blinding run. 1088 */ 1089 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1090 insn[1].code == 0) 1091 memcpy(aux, insn, sizeof(aux)); 1092 1093 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff); 1094 if (!rewritten) 1095 continue; 1096 1097 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1098 if (!tmp) { 1099 /* Patching may have repointed aux->prog during 1100 * realloc from the original one, so we need to 1101 * fix it up here on error. 1102 */ 1103 bpf_jit_prog_release_other(prog, clone); 1104 return ERR_PTR(-ENOMEM); 1105 } 1106 1107 clone = tmp; 1108 insn_delta = rewritten - 1; 1109 1110 /* Walk new program and skip insns we just inserted. */ 1111 insn = clone->insnsi + i + insn_delta; 1112 insn_cnt += insn_delta; 1113 i += insn_delta; 1114 } 1115 1116 clone->blinded = 1; 1117 return clone; 1118 } 1119 #endif /* CONFIG_BPF_JIT */ 1120 1121 /* Base function for offset calculation. Needs to go into .text section, 1122 * therefore keeping it non-static as well; will also be used by JITs 1123 * anyway later on, so do not let the compiler omit it. This also needs 1124 * to go into kallsyms for correlation from e.g. bpftool, so naming 1125 * must not change. 1126 */ 1127 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1128 { 1129 return 0; 1130 } 1131 EXPORT_SYMBOL_GPL(__bpf_call_base); 1132 1133 /* All UAPI available opcodes. */ 1134 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1135 /* 32 bit ALU operations. */ \ 1136 /* Register based. */ \ 1137 INSN_3(ALU, ADD, X), \ 1138 INSN_3(ALU, SUB, X), \ 1139 INSN_3(ALU, AND, X), \ 1140 INSN_3(ALU, OR, X), \ 1141 INSN_3(ALU, LSH, X), \ 1142 INSN_3(ALU, RSH, X), \ 1143 INSN_3(ALU, XOR, X), \ 1144 INSN_3(ALU, MUL, X), \ 1145 INSN_3(ALU, MOV, X), \ 1146 INSN_3(ALU, ARSH, X), \ 1147 INSN_3(ALU, DIV, X), \ 1148 INSN_3(ALU, MOD, X), \ 1149 INSN_2(ALU, NEG), \ 1150 INSN_3(ALU, END, TO_BE), \ 1151 INSN_3(ALU, END, TO_LE), \ 1152 /* Immediate based. */ \ 1153 INSN_3(ALU, ADD, K), \ 1154 INSN_3(ALU, SUB, K), \ 1155 INSN_3(ALU, AND, K), \ 1156 INSN_3(ALU, OR, K), \ 1157 INSN_3(ALU, LSH, K), \ 1158 INSN_3(ALU, RSH, K), \ 1159 INSN_3(ALU, XOR, K), \ 1160 INSN_3(ALU, MUL, K), \ 1161 INSN_3(ALU, MOV, K), \ 1162 INSN_3(ALU, ARSH, K), \ 1163 INSN_3(ALU, DIV, K), \ 1164 INSN_3(ALU, MOD, K), \ 1165 /* 64 bit ALU operations. */ \ 1166 /* Register based. */ \ 1167 INSN_3(ALU64, ADD, X), \ 1168 INSN_3(ALU64, SUB, X), \ 1169 INSN_3(ALU64, AND, X), \ 1170 INSN_3(ALU64, OR, X), \ 1171 INSN_3(ALU64, LSH, X), \ 1172 INSN_3(ALU64, RSH, X), \ 1173 INSN_3(ALU64, XOR, X), \ 1174 INSN_3(ALU64, MUL, X), \ 1175 INSN_3(ALU64, MOV, X), \ 1176 INSN_3(ALU64, ARSH, X), \ 1177 INSN_3(ALU64, DIV, X), \ 1178 INSN_3(ALU64, MOD, X), \ 1179 INSN_2(ALU64, NEG), \ 1180 /* Immediate based. */ \ 1181 INSN_3(ALU64, ADD, K), \ 1182 INSN_3(ALU64, SUB, K), \ 1183 INSN_3(ALU64, AND, K), \ 1184 INSN_3(ALU64, OR, K), \ 1185 INSN_3(ALU64, LSH, K), \ 1186 INSN_3(ALU64, RSH, K), \ 1187 INSN_3(ALU64, XOR, K), \ 1188 INSN_3(ALU64, MUL, K), \ 1189 INSN_3(ALU64, MOV, K), \ 1190 INSN_3(ALU64, ARSH, K), \ 1191 INSN_3(ALU64, DIV, K), \ 1192 INSN_3(ALU64, MOD, K), \ 1193 /* Call instruction. */ \ 1194 INSN_2(JMP, CALL), \ 1195 /* Exit instruction. */ \ 1196 INSN_2(JMP, EXIT), \ 1197 /* 32-bit Jump instructions. */ \ 1198 /* Register based. */ \ 1199 INSN_3(JMP32, JEQ, X), \ 1200 INSN_3(JMP32, JNE, X), \ 1201 INSN_3(JMP32, JGT, X), \ 1202 INSN_3(JMP32, JLT, X), \ 1203 INSN_3(JMP32, JGE, X), \ 1204 INSN_3(JMP32, JLE, X), \ 1205 INSN_3(JMP32, JSGT, X), \ 1206 INSN_3(JMP32, JSLT, X), \ 1207 INSN_3(JMP32, JSGE, X), \ 1208 INSN_3(JMP32, JSLE, X), \ 1209 INSN_3(JMP32, JSET, X), \ 1210 /* Immediate based. */ \ 1211 INSN_3(JMP32, JEQ, K), \ 1212 INSN_3(JMP32, JNE, K), \ 1213 INSN_3(JMP32, JGT, K), \ 1214 INSN_3(JMP32, JLT, K), \ 1215 INSN_3(JMP32, JGE, K), \ 1216 INSN_3(JMP32, JLE, K), \ 1217 INSN_3(JMP32, JSGT, K), \ 1218 INSN_3(JMP32, JSLT, K), \ 1219 INSN_3(JMP32, JSGE, K), \ 1220 INSN_3(JMP32, JSLE, K), \ 1221 INSN_3(JMP32, JSET, K), \ 1222 /* Jump instructions. */ \ 1223 /* Register based. */ \ 1224 INSN_3(JMP, JEQ, X), \ 1225 INSN_3(JMP, JNE, X), \ 1226 INSN_3(JMP, JGT, X), \ 1227 INSN_3(JMP, JLT, X), \ 1228 INSN_3(JMP, JGE, X), \ 1229 INSN_3(JMP, JLE, X), \ 1230 INSN_3(JMP, JSGT, X), \ 1231 INSN_3(JMP, JSLT, X), \ 1232 INSN_3(JMP, JSGE, X), \ 1233 INSN_3(JMP, JSLE, X), \ 1234 INSN_3(JMP, JSET, X), \ 1235 /* Immediate based. */ \ 1236 INSN_3(JMP, JEQ, K), \ 1237 INSN_3(JMP, JNE, K), \ 1238 INSN_3(JMP, JGT, K), \ 1239 INSN_3(JMP, JLT, K), \ 1240 INSN_3(JMP, JGE, K), \ 1241 INSN_3(JMP, JLE, K), \ 1242 INSN_3(JMP, JSGT, K), \ 1243 INSN_3(JMP, JSLT, K), \ 1244 INSN_3(JMP, JSGE, K), \ 1245 INSN_3(JMP, JSLE, K), \ 1246 INSN_3(JMP, JSET, K), \ 1247 INSN_2(JMP, JA), \ 1248 /* Store instructions. */ \ 1249 /* Register based. */ \ 1250 INSN_3(STX, MEM, B), \ 1251 INSN_3(STX, MEM, H), \ 1252 INSN_3(STX, MEM, W), \ 1253 INSN_3(STX, MEM, DW), \ 1254 INSN_3(STX, XADD, W), \ 1255 INSN_3(STX, XADD, DW), \ 1256 /* Immediate based. */ \ 1257 INSN_3(ST, MEM, B), \ 1258 INSN_3(ST, MEM, H), \ 1259 INSN_3(ST, MEM, W), \ 1260 INSN_3(ST, MEM, DW), \ 1261 /* Load instructions. */ \ 1262 /* Register based. */ \ 1263 INSN_3(LDX, MEM, B), \ 1264 INSN_3(LDX, MEM, H), \ 1265 INSN_3(LDX, MEM, W), \ 1266 INSN_3(LDX, MEM, DW), \ 1267 /* Immediate based. */ \ 1268 INSN_3(LD, IMM, DW) 1269 1270 bool bpf_opcode_in_insntable(u8 code) 1271 { 1272 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1273 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1274 static const bool public_insntable[256] = { 1275 [0 ... 255] = false, 1276 /* Now overwrite non-defaults ... */ 1277 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1278 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1279 [BPF_LD | BPF_ABS | BPF_B] = true, 1280 [BPF_LD | BPF_ABS | BPF_H] = true, 1281 [BPF_LD | BPF_ABS | BPF_W] = true, 1282 [BPF_LD | BPF_IND | BPF_B] = true, 1283 [BPF_LD | BPF_IND | BPF_H] = true, 1284 [BPF_LD | BPF_IND | BPF_W] = true, 1285 }; 1286 #undef BPF_INSN_3_TBL 1287 #undef BPF_INSN_2_TBL 1288 return public_insntable[code]; 1289 } 1290 1291 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1292 /** 1293 * __bpf_prog_run - run eBPF program on a given context 1294 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1295 * @insn: is the array of eBPF instructions 1296 * @stack: is the eBPF storage stack 1297 * 1298 * Decode and execute eBPF instructions. 1299 */ 1300 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 1301 { 1302 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1303 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1304 static const void *jumptable[256] = { 1305 [0 ... 255] = &&default_label, 1306 /* Now overwrite non-defaults ... */ 1307 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1308 /* Non-UAPI available opcodes. */ 1309 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1310 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1311 }; 1312 #undef BPF_INSN_3_LBL 1313 #undef BPF_INSN_2_LBL 1314 u32 tail_call_cnt = 0; 1315 1316 #define CONT ({ insn++; goto select_insn; }) 1317 #define CONT_JMP ({ insn++; goto select_insn; }) 1318 1319 select_insn: 1320 goto *jumptable[insn->code]; 1321 1322 /* ALU */ 1323 #define ALU(OPCODE, OP) \ 1324 ALU64_##OPCODE##_X: \ 1325 DST = DST OP SRC; \ 1326 CONT; \ 1327 ALU_##OPCODE##_X: \ 1328 DST = (u32) DST OP (u32) SRC; \ 1329 CONT; \ 1330 ALU64_##OPCODE##_K: \ 1331 DST = DST OP IMM; \ 1332 CONT; \ 1333 ALU_##OPCODE##_K: \ 1334 DST = (u32) DST OP (u32) IMM; \ 1335 CONT; 1336 1337 ALU(ADD, +) 1338 ALU(SUB, -) 1339 ALU(AND, &) 1340 ALU(OR, |) 1341 ALU(LSH, <<) 1342 ALU(RSH, >>) 1343 ALU(XOR, ^) 1344 ALU(MUL, *) 1345 #undef ALU 1346 ALU_NEG: 1347 DST = (u32) -DST; 1348 CONT; 1349 ALU64_NEG: 1350 DST = -DST; 1351 CONT; 1352 ALU_MOV_X: 1353 DST = (u32) SRC; 1354 CONT; 1355 ALU_MOV_K: 1356 DST = (u32) IMM; 1357 CONT; 1358 ALU64_MOV_X: 1359 DST = SRC; 1360 CONT; 1361 ALU64_MOV_K: 1362 DST = IMM; 1363 CONT; 1364 LD_IMM_DW: 1365 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1366 insn++; 1367 CONT; 1368 ALU_ARSH_X: 1369 DST = (u64) (u32) ((*(s32 *) &DST) >> SRC); 1370 CONT; 1371 ALU_ARSH_K: 1372 DST = (u64) (u32) ((*(s32 *) &DST) >> IMM); 1373 CONT; 1374 ALU64_ARSH_X: 1375 (*(s64 *) &DST) >>= SRC; 1376 CONT; 1377 ALU64_ARSH_K: 1378 (*(s64 *) &DST) >>= IMM; 1379 CONT; 1380 ALU64_MOD_X: 1381 div64_u64_rem(DST, SRC, &AX); 1382 DST = AX; 1383 CONT; 1384 ALU_MOD_X: 1385 AX = (u32) DST; 1386 DST = do_div(AX, (u32) SRC); 1387 CONT; 1388 ALU64_MOD_K: 1389 div64_u64_rem(DST, IMM, &AX); 1390 DST = AX; 1391 CONT; 1392 ALU_MOD_K: 1393 AX = (u32) DST; 1394 DST = do_div(AX, (u32) IMM); 1395 CONT; 1396 ALU64_DIV_X: 1397 DST = div64_u64(DST, SRC); 1398 CONT; 1399 ALU_DIV_X: 1400 AX = (u32) DST; 1401 do_div(AX, (u32) SRC); 1402 DST = (u32) AX; 1403 CONT; 1404 ALU64_DIV_K: 1405 DST = div64_u64(DST, IMM); 1406 CONT; 1407 ALU_DIV_K: 1408 AX = (u32) DST; 1409 do_div(AX, (u32) IMM); 1410 DST = (u32) AX; 1411 CONT; 1412 ALU_END_TO_BE: 1413 switch (IMM) { 1414 case 16: 1415 DST = (__force u16) cpu_to_be16(DST); 1416 break; 1417 case 32: 1418 DST = (__force u32) cpu_to_be32(DST); 1419 break; 1420 case 64: 1421 DST = (__force u64) cpu_to_be64(DST); 1422 break; 1423 } 1424 CONT; 1425 ALU_END_TO_LE: 1426 switch (IMM) { 1427 case 16: 1428 DST = (__force u16) cpu_to_le16(DST); 1429 break; 1430 case 32: 1431 DST = (__force u32) cpu_to_le32(DST); 1432 break; 1433 case 64: 1434 DST = (__force u64) cpu_to_le64(DST); 1435 break; 1436 } 1437 CONT; 1438 1439 /* CALL */ 1440 JMP_CALL: 1441 /* Function call scratches BPF_R1-BPF_R5 registers, 1442 * preserves BPF_R6-BPF_R9, and stores return value 1443 * into BPF_R0. 1444 */ 1445 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1446 BPF_R4, BPF_R5); 1447 CONT; 1448 1449 JMP_CALL_ARGS: 1450 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1451 BPF_R3, BPF_R4, 1452 BPF_R5, 1453 insn + insn->off + 1); 1454 CONT; 1455 1456 JMP_TAIL_CALL: { 1457 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1458 struct bpf_array *array = container_of(map, struct bpf_array, map); 1459 struct bpf_prog *prog; 1460 u32 index = BPF_R3; 1461 1462 if (unlikely(index >= array->map.max_entries)) 1463 goto out; 1464 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1465 goto out; 1466 1467 tail_call_cnt++; 1468 1469 prog = READ_ONCE(array->ptrs[index]); 1470 if (!prog) 1471 goto out; 1472 1473 /* ARG1 at this point is guaranteed to point to CTX from 1474 * the verifier side due to the fact that the tail call is 1475 * handeled like a helper, that is, bpf_tail_call_proto, 1476 * where arg1_type is ARG_PTR_TO_CTX. 1477 */ 1478 insn = prog->insnsi; 1479 goto select_insn; 1480 out: 1481 CONT; 1482 } 1483 JMP_JA: 1484 insn += insn->off; 1485 CONT; 1486 JMP_EXIT: 1487 return BPF_R0; 1488 /* JMP */ 1489 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1490 JMP_##OPCODE##_X: \ 1491 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1492 insn += insn->off; \ 1493 CONT_JMP; \ 1494 } \ 1495 CONT; \ 1496 JMP32_##OPCODE##_X: \ 1497 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1498 insn += insn->off; \ 1499 CONT_JMP; \ 1500 } \ 1501 CONT; \ 1502 JMP_##OPCODE##_K: \ 1503 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1504 insn += insn->off; \ 1505 CONT_JMP; \ 1506 } \ 1507 CONT; \ 1508 JMP32_##OPCODE##_K: \ 1509 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1510 insn += insn->off; \ 1511 CONT_JMP; \ 1512 } \ 1513 CONT; 1514 COND_JMP(u, JEQ, ==) 1515 COND_JMP(u, JNE, !=) 1516 COND_JMP(u, JGT, >) 1517 COND_JMP(u, JLT, <) 1518 COND_JMP(u, JGE, >=) 1519 COND_JMP(u, JLE, <=) 1520 COND_JMP(u, JSET, &) 1521 COND_JMP(s, JSGT, >) 1522 COND_JMP(s, JSLT, <) 1523 COND_JMP(s, JSGE, >=) 1524 COND_JMP(s, JSLE, <=) 1525 #undef COND_JMP 1526 /* STX and ST and LDX*/ 1527 #define LDST(SIZEOP, SIZE) \ 1528 STX_MEM_##SIZEOP: \ 1529 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1530 CONT; \ 1531 ST_MEM_##SIZEOP: \ 1532 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1533 CONT; \ 1534 LDX_MEM_##SIZEOP: \ 1535 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1536 CONT; 1537 1538 LDST(B, u8) 1539 LDST(H, u16) 1540 LDST(W, u32) 1541 LDST(DW, u64) 1542 #undef LDST 1543 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1544 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1545 (DST + insn->off)); 1546 CONT; 1547 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1548 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1549 (DST + insn->off)); 1550 CONT; 1551 1552 default_label: 1553 /* If we ever reach this, we have a bug somewhere. Die hard here 1554 * instead of just returning 0; we could be somewhere in a subprog, 1555 * so execution could continue otherwise which we do /not/ want. 1556 * 1557 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1558 */ 1559 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code); 1560 BUG_ON(1); 1561 return 0; 1562 } 1563 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */ 1564 1565 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1566 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1567 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1568 { \ 1569 u64 stack[stack_size / sizeof(u64)]; \ 1570 u64 regs[MAX_BPF_EXT_REG]; \ 1571 \ 1572 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1573 ARG1 = (u64) (unsigned long) ctx; \ 1574 return ___bpf_prog_run(regs, insn, stack); \ 1575 } 1576 1577 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1578 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1579 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1580 const struct bpf_insn *insn) \ 1581 { \ 1582 u64 stack[stack_size / sizeof(u64)]; \ 1583 u64 regs[MAX_BPF_EXT_REG]; \ 1584 \ 1585 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1586 BPF_R1 = r1; \ 1587 BPF_R2 = r2; \ 1588 BPF_R3 = r3; \ 1589 BPF_R4 = r4; \ 1590 BPF_R5 = r5; \ 1591 return ___bpf_prog_run(regs, insn, stack); \ 1592 } 1593 1594 #define EVAL1(FN, X) FN(X) 1595 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1596 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1597 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1598 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1599 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1600 1601 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1602 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1603 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1604 1605 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1606 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1607 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1608 1609 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1610 1611 static unsigned int (*interpreters[])(const void *ctx, 1612 const struct bpf_insn *insn) = { 1613 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1614 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1615 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1616 }; 1617 #undef PROG_NAME_LIST 1618 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1619 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1620 const struct bpf_insn *insn) = { 1621 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1622 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1623 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1624 }; 1625 #undef PROG_NAME_LIST 1626 1627 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1628 { 1629 stack_depth = max_t(u32, stack_depth, 1); 1630 insn->off = (s16) insn->imm; 1631 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1632 __bpf_call_base_args; 1633 insn->code = BPF_JMP | BPF_CALL_ARGS; 1634 } 1635 1636 #else 1637 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1638 const struct bpf_insn *insn) 1639 { 1640 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1641 * is not working properly, so warn about it! 1642 */ 1643 WARN_ON_ONCE(1); 1644 return 0; 1645 } 1646 #endif 1647 1648 bool bpf_prog_array_compatible(struct bpf_array *array, 1649 const struct bpf_prog *fp) 1650 { 1651 if (fp->kprobe_override) 1652 return false; 1653 1654 if (!array->owner_prog_type) { 1655 /* There's no owner yet where we could check for 1656 * compatibility. 1657 */ 1658 array->owner_prog_type = fp->type; 1659 array->owner_jited = fp->jited; 1660 1661 return true; 1662 } 1663 1664 return array->owner_prog_type == fp->type && 1665 array->owner_jited == fp->jited; 1666 } 1667 1668 static int bpf_check_tail_call(const struct bpf_prog *fp) 1669 { 1670 struct bpf_prog_aux *aux = fp->aux; 1671 int i; 1672 1673 for (i = 0; i < aux->used_map_cnt; i++) { 1674 struct bpf_map *map = aux->used_maps[i]; 1675 struct bpf_array *array; 1676 1677 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1678 continue; 1679 1680 array = container_of(map, struct bpf_array, map); 1681 if (!bpf_prog_array_compatible(array, fp)) 1682 return -EINVAL; 1683 } 1684 1685 return 0; 1686 } 1687 1688 static void bpf_prog_select_func(struct bpf_prog *fp) 1689 { 1690 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1691 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1692 1693 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1694 #else 1695 fp->bpf_func = __bpf_prog_ret0_warn; 1696 #endif 1697 } 1698 1699 /** 1700 * bpf_prog_select_runtime - select exec runtime for BPF program 1701 * @fp: bpf_prog populated with internal BPF program 1702 * @err: pointer to error variable 1703 * 1704 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1705 * The BPF program will be executed via BPF_PROG_RUN() macro. 1706 */ 1707 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1708 { 1709 /* In case of BPF to BPF calls, verifier did all the prep 1710 * work with regards to JITing, etc. 1711 */ 1712 if (fp->bpf_func) 1713 goto finalize; 1714 1715 bpf_prog_select_func(fp); 1716 1717 /* eBPF JITs can rewrite the program in case constant 1718 * blinding is active. However, in case of error during 1719 * blinding, bpf_int_jit_compile() must always return a 1720 * valid program, which in this case would simply not 1721 * be JITed, but falls back to the interpreter. 1722 */ 1723 if (!bpf_prog_is_dev_bound(fp->aux)) { 1724 *err = bpf_prog_alloc_jited_linfo(fp); 1725 if (*err) 1726 return fp; 1727 1728 fp = bpf_int_jit_compile(fp); 1729 if (!fp->jited) { 1730 bpf_prog_free_jited_linfo(fp); 1731 #ifdef CONFIG_BPF_JIT_ALWAYS_ON 1732 *err = -ENOTSUPP; 1733 return fp; 1734 #endif 1735 } else { 1736 bpf_prog_free_unused_jited_linfo(fp); 1737 } 1738 } else { 1739 *err = bpf_prog_offload_compile(fp); 1740 if (*err) 1741 return fp; 1742 } 1743 1744 finalize: 1745 bpf_prog_lock_ro(fp); 1746 1747 /* The tail call compatibility check can only be done at 1748 * this late stage as we need to determine, if we deal 1749 * with JITed or non JITed program concatenations and not 1750 * all eBPF JITs might immediately support all features. 1751 */ 1752 *err = bpf_check_tail_call(fp); 1753 1754 return fp; 1755 } 1756 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1757 1758 static unsigned int __bpf_prog_ret1(const void *ctx, 1759 const struct bpf_insn *insn) 1760 { 1761 return 1; 1762 } 1763 1764 static struct bpf_prog_dummy { 1765 struct bpf_prog prog; 1766 } dummy_bpf_prog = { 1767 .prog = { 1768 .bpf_func = __bpf_prog_ret1, 1769 }, 1770 }; 1771 1772 /* to avoid allocating empty bpf_prog_array for cgroups that 1773 * don't have bpf program attached use one global 'empty_prog_array' 1774 * It will not be modified the caller of bpf_prog_array_alloc() 1775 * (since caller requested prog_cnt == 0) 1776 * that pointer should be 'freed' by bpf_prog_array_free() 1777 */ 1778 static struct { 1779 struct bpf_prog_array hdr; 1780 struct bpf_prog *null_prog; 1781 } empty_prog_array = { 1782 .null_prog = NULL, 1783 }; 1784 1785 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1786 { 1787 if (prog_cnt) 1788 return kzalloc(sizeof(struct bpf_prog_array) + 1789 sizeof(struct bpf_prog_array_item) * 1790 (prog_cnt + 1), 1791 flags); 1792 1793 return &empty_prog_array.hdr; 1794 } 1795 1796 void bpf_prog_array_free(struct bpf_prog_array __rcu *progs) 1797 { 1798 if (!progs || 1799 progs == (struct bpf_prog_array __rcu *)&empty_prog_array.hdr) 1800 return; 1801 kfree_rcu(progs, rcu); 1802 } 1803 1804 int bpf_prog_array_length(struct bpf_prog_array __rcu *array) 1805 { 1806 struct bpf_prog_array_item *item; 1807 u32 cnt = 0; 1808 1809 rcu_read_lock(); 1810 item = rcu_dereference(array)->items; 1811 for (; item->prog; item++) 1812 if (item->prog != &dummy_bpf_prog.prog) 1813 cnt++; 1814 rcu_read_unlock(); 1815 return cnt; 1816 } 1817 1818 1819 static bool bpf_prog_array_copy_core(struct bpf_prog_array __rcu *array, 1820 u32 *prog_ids, 1821 u32 request_cnt) 1822 { 1823 struct bpf_prog_array_item *item; 1824 int i = 0; 1825 1826 item = rcu_dereference_check(array, 1)->items; 1827 for (; item->prog; item++) { 1828 if (item->prog == &dummy_bpf_prog.prog) 1829 continue; 1830 prog_ids[i] = item->prog->aux->id; 1831 if (++i == request_cnt) { 1832 item++; 1833 break; 1834 } 1835 } 1836 1837 return !!(item->prog); 1838 } 1839 1840 int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *array, 1841 __u32 __user *prog_ids, u32 cnt) 1842 { 1843 unsigned long err = 0; 1844 bool nospc; 1845 u32 *ids; 1846 1847 /* users of this function are doing: 1848 * cnt = bpf_prog_array_length(); 1849 * if (cnt > 0) 1850 * bpf_prog_array_copy_to_user(..., cnt); 1851 * so below kcalloc doesn't need extra cnt > 0 check, but 1852 * bpf_prog_array_length() releases rcu lock and 1853 * prog array could have been swapped with empty or larger array, 1854 * so always copy 'cnt' prog_ids to the user. 1855 * In a rare race the user will see zero prog_ids 1856 */ 1857 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 1858 if (!ids) 1859 return -ENOMEM; 1860 rcu_read_lock(); 1861 nospc = bpf_prog_array_copy_core(array, ids, cnt); 1862 rcu_read_unlock(); 1863 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 1864 kfree(ids); 1865 if (err) 1866 return -EFAULT; 1867 if (nospc) 1868 return -ENOSPC; 1869 return 0; 1870 } 1871 1872 void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *array, 1873 struct bpf_prog *old_prog) 1874 { 1875 struct bpf_prog_array_item *item = array->items; 1876 1877 for (; item->prog; item++) 1878 if (item->prog == old_prog) { 1879 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 1880 break; 1881 } 1882 } 1883 1884 int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array, 1885 struct bpf_prog *exclude_prog, 1886 struct bpf_prog *include_prog, 1887 struct bpf_prog_array **new_array) 1888 { 1889 int new_prog_cnt, carry_prog_cnt = 0; 1890 struct bpf_prog_array_item *existing; 1891 struct bpf_prog_array *array; 1892 bool found_exclude = false; 1893 int new_prog_idx = 0; 1894 1895 /* Figure out how many existing progs we need to carry over to 1896 * the new array. 1897 */ 1898 if (old_array) { 1899 existing = old_array->items; 1900 for (; existing->prog; existing++) { 1901 if (existing->prog == exclude_prog) { 1902 found_exclude = true; 1903 continue; 1904 } 1905 if (existing->prog != &dummy_bpf_prog.prog) 1906 carry_prog_cnt++; 1907 if (existing->prog == include_prog) 1908 return -EEXIST; 1909 } 1910 } 1911 1912 if (exclude_prog && !found_exclude) 1913 return -ENOENT; 1914 1915 /* How many progs (not NULL) will be in the new array? */ 1916 new_prog_cnt = carry_prog_cnt; 1917 if (include_prog) 1918 new_prog_cnt += 1; 1919 1920 /* Do we have any prog (not NULL) in the new array? */ 1921 if (!new_prog_cnt) { 1922 *new_array = NULL; 1923 return 0; 1924 } 1925 1926 /* +1 as the end of prog_array is marked with NULL */ 1927 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1928 if (!array) 1929 return -ENOMEM; 1930 1931 /* Fill in the new prog array */ 1932 if (carry_prog_cnt) { 1933 existing = old_array->items; 1934 for (; existing->prog; existing++) 1935 if (existing->prog != exclude_prog && 1936 existing->prog != &dummy_bpf_prog.prog) { 1937 array->items[new_prog_idx++].prog = 1938 existing->prog; 1939 } 1940 } 1941 if (include_prog) 1942 array->items[new_prog_idx++].prog = include_prog; 1943 array->items[new_prog_idx].prog = NULL; 1944 *new_array = array; 1945 return 0; 1946 } 1947 1948 int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array, 1949 u32 *prog_ids, u32 request_cnt, 1950 u32 *prog_cnt) 1951 { 1952 u32 cnt = 0; 1953 1954 if (array) 1955 cnt = bpf_prog_array_length(array); 1956 1957 *prog_cnt = cnt; 1958 1959 /* return early if user requested only program count or nothing to copy */ 1960 if (!request_cnt || !cnt) 1961 return 0; 1962 1963 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 1964 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 1965 : 0; 1966 } 1967 1968 static void bpf_prog_free_deferred(struct work_struct *work) 1969 { 1970 struct bpf_prog_aux *aux; 1971 int i; 1972 1973 aux = container_of(work, struct bpf_prog_aux, work); 1974 if (bpf_prog_is_dev_bound(aux)) 1975 bpf_prog_offload_destroy(aux->prog); 1976 #ifdef CONFIG_PERF_EVENTS 1977 if (aux->prog->has_callchain_buf) 1978 put_callchain_buffers(); 1979 #endif 1980 for (i = 0; i < aux->func_cnt; i++) 1981 bpf_jit_free(aux->func[i]); 1982 if (aux->func_cnt) { 1983 kfree(aux->func); 1984 bpf_prog_unlock_free(aux->prog); 1985 } else { 1986 bpf_jit_free(aux->prog); 1987 } 1988 } 1989 1990 /* Free internal BPF program */ 1991 void bpf_prog_free(struct bpf_prog *fp) 1992 { 1993 struct bpf_prog_aux *aux = fp->aux; 1994 1995 INIT_WORK(&aux->work, bpf_prog_free_deferred); 1996 schedule_work(&aux->work); 1997 } 1998 EXPORT_SYMBOL_GPL(bpf_prog_free); 1999 2000 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2001 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2002 2003 void bpf_user_rnd_init_once(void) 2004 { 2005 prandom_init_once(&bpf_user_rnd_state); 2006 } 2007 2008 BPF_CALL_0(bpf_user_rnd_u32) 2009 { 2010 /* Should someone ever have the rather unwise idea to use some 2011 * of the registers passed into this function, then note that 2012 * this function is called from native eBPF and classic-to-eBPF 2013 * transformations. Register assignments from both sides are 2014 * different, f.e. classic always sets fn(ctx, A, X) here. 2015 */ 2016 struct rnd_state *state; 2017 u32 res; 2018 2019 state = &get_cpu_var(bpf_user_rnd_state); 2020 res = prandom_u32_state(state); 2021 put_cpu_var(bpf_user_rnd_state); 2022 2023 return res; 2024 } 2025 2026 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2027 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2028 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2029 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2030 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2031 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2032 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2033 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2034 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2035 2036 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2037 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2038 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2039 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2040 2041 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2042 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2043 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2044 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2045 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2046 2047 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2048 { 2049 return NULL; 2050 } 2051 2052 u64 __weak 2053 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2054 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2055 { 2056 return -ENOTSUPP; 2057 } 2058 EXPORT_SYMBOL_GPL(bpf_event_output); 2059 2060 /* Always built-in helper functions. */ 2061 const struct bpf_func_proto bpf_tail_call_proto = { 2062 .func = NULL, 2063 .gpl_only = false, 2064 .ret_type = RET_VOID, 2065 .arg1_type = ARG_PTR_TO_CTX, 2066 .arg2_type = ARG_CONST_MAP_PTR, 2067 .arg3_type = ARG_ANYTHING, 2068 }; 2069 2070 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2071 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2072 * eBPF and implicitly also cBPF can get JITed! 2073 */ 2074 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2075 { 2076 return prog; 2077 } 2078 2079 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2080 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2081 */ 2082 void __weak bpf_jit_compile(struct bpf_prog *prog) 2083 { 2084 } 2085 2086 bool __weak bpf_helper_changes_pkt_data(void *func) 2087 { 2088 return false; 2089 } 2090 2091 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2092 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2093 */ 2094 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2095 int len) 2096 { 2097 return -EFAULT; 2098 } 2099 2100 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2101 EXPORT_SYMBOL(bpf_stats_enabled_key); 2102 int sysctl_bpf_stats_enabled __read_mostly; 2103 2104 /* All definitions of tracepoints related to BPF. */ 2105 #define CREATE_TRACE_POINTS 2106 #include <linux/bpf_trace.h> 2107 2108 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2109