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