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