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 # define BPF_JIT_LIMIT_DEFAULT (PAGE_SIZE * 40000) 369 370 /* All BPF JIT sysctl knobs here. */ 371 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_ALWAYS_ON); 372 int bpf_jit_harden __read_mostly; 373 int bpf_jit_kallsyms __read_mostly; 374 int bpf_jit_limit __read_mostly = BPF_JIT_LIMIT_DEFAULT; 375 376 static __always_inline void 377 bpf_get_prog_addr_region(const struct bpf_prog *prog, 378 unsigned long *symbol_start, 379 unsigned long *symbol_end) 380 { 381 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 382 unsigned long addr = (unsigned long)hdr; 383 384 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 385 386 *symbol_start = addr; 387 *symbol_end = addr + hdr->pages * PAGE_SIZE; 388 } 389 390 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym) 391 { 392 const char *end = sym + KSYM_NAME_LEN; 393 394 BUILD_BUG_ON(sizeof("bpf_prog_") + 395 sizeof(prog->tag) * 2 + 396 /* name has been null terminated. 397 * We should need +1 for the '_' preceding 398 * the name. However, the null character 399 * is double counted between the name and the 400 * sizeof("bpf_prog_") above, so we omit 401 * the +1 here. 402 */ 403 sizeof(prog->aux->name) > KSYM_NAME_LEN); 404 405 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 406 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 407 if (prog->aux->name[0]) 408 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 409 else 410 *sym = 0; 411 } 412 413 static __always_inline unsigned long 414 bpf_get_prog_addr_start(struct latch_tree_node *n) 415 { 416 unsigned long symbol_start, symbol_end; 417 const struct bpf_prog_aux *aux; 418 419 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 420 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 421 422 return symbol_start; 423 } 424 425 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 426 struct latch_tree_node *b) 427 { 428 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b); 429 } 430 431 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 432 { 433 unsigned long val = (unsigned long)key; 434 unsigned long symbol_start, symbol_end; 435 const struct bpf_prog_aux *aux; 436 437 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 438 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 439 440 if (val < symbol_start) 441 return -1; 442 if (val >= symbol_end) 443 return 1; 444 445 return 0; 446 } 447 448 static const struct latch_tree_ops bpf_tree_ops = { 449 .less = bpf_tree_less, 450 .comp = bpf_tree_comp, 451 }; 452 453 static DEFINE_SPINLOCK(bpf_lock); 454 static LIST_HEAD(bpf_kallsyms); 455 static struct latch_tree_root bpf_tree __cacheline_aligned; 456 457 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux) 458 { 459 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode)); 460 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms); 461 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 462 } 463 464 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux) 465 { 466 if (list_empty(&aux->ksym_lnode)) 467 return; 468 469 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 470 list_del_rcu(&aux->ksym_lnode); 471 } 472 473 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 474 { 475 return fp->jited && !bpf_prog_was_classic(fp); 476 } 477 478 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 479 { 480 return list_empty(&fp->aux->ksym_lnode) || 481 fp->aux->ksym_lnode.prev == LIST_POISON2; 482 } 483 484 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 485 { 486 if (!bpf_prog_kallsyms_candidate(fp) || 487 !capable(CAP_SYS_ADMIN)) 488 return; 489 490 spin_lock_bh(&bpf_lock); 491 bpf_prog_ksym_node_add(fp->aux); 492 spin_unlock_bh(&bpf_lock); 493 } 494 495 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 496 { 497 if (!bpf_prog_kallsyms_candidate(fp)) 498 return; 499 500 spin_lock_bh(&bpf_lock); 501 bpf_prog_ksym_node_del(fp->aux); 502 spin_unlock_bh(&bpf_lock); 503 } 504 505 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr) 506 { 507 struct latch_tree_node *n; 508 509 if (!bpf_jit_kallsyms_enabled()) 510 return NULL; 511 512 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 513 return n ? 514 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 515 NULL; 516 } 517 518 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 519 unsigned long *off, char *sym) 520 { 521 unsigned long symbol_start, symbol_end; 522 struct bpf_prog *prog; 523 char *ret = NULL; 524 525 rcu_read_lock(); 526 prog = bpf_prog_kallsyms_find(addr); 527 if (prog) { 528 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 529 bpf_get_prog_name(prog, sym); 530 531 ret = sym; 532 if (size) 533 *size = symbol_end - symbol_start; 534 if (off) 535 *off = addr - symbol_start; 536 } 537 rcu_read_unlock(); 538 539 return ret; 540 } 541 542 bool is_bpf_text_address(unsigned long addr) 543 { 544 bool ret; 545 546 rcu_read_lock(); 547 ret = bpf_prog_kallsyms_find(addr) != NULL; 548 rcu_read_unlock(); 549 550 return ret; 551 } 552 553 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 554 char *sym) 555 { 556 unsigned long symbol_start, symbol_end; 557 struct bpf_prog_aux *aux; 558 unsigned int it = 0; 559 int ret = -ERANGE; 560 561 if (!bpf_jit_kallsyms_enabled()) 562 return ret; 563 564 rcu_read_lock(); 565 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 566 if (it++ != symnum) 567 continue; 568 569 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 570 bpf_get_prog_name(aux->prog, sym); 571 572 *value = symbol_start; 573 *type = BPF_SYM_ELF_TYPE; 574 575 ret = 0; 576 break; 577 } 578 rcu_read_unlock(); 579 580 return ret; 581 } 582 583 static atomic_long_t bpf_jit_current; 584 585 #if defined(MODULES_VADDR) 586 static int __init bpf_jit_charge_init(void) 587 { 588 /* Only used as heuristic here to derive limit. */ 589 bpf_jit_limit = min_t(u64, round_up((MODULES_END - MODULES_VADDR) >> 2, 590 PAGE_SIZE), INT_MAX); 591 return 0; 592 } 593 pure_initcall(bpf_jit_charge_init); 594 #endif 595 596 static int bpf_jit_charge_modmem(u32 pages) 597 { 598 if (atomic_long_add_return(pages, &bpf_jit_current) > 599 (bpf_jit_limit >> PAGE_SHIFT)) { 600 if (!capable(CAP_SYS_ADMIN)) { 601 atomic_long_sub(pages, &bpf_jit_current); 602 return -EPERM; 603 } 604 } 605 606 return 0; 607 } 608 609 static void bpf_jit_uncharge_modmem(u32 pages) 610 { 611 atomic_long_sub(pages, &bpf_jit_current); 612 } 613 614 struct bpf_binary_header * 615 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 616 unsigned int alignment, 617 bpf_jit_fill_hole_t bpf_fill_ill_insns) 618 { 619 struct bpf_binary_header *hdr; 620 u32 size, hole, start, pages; 621 622 /* Most of BPF filters are really small, but if some of them 623 * fill a page, allow at least 128 extra bytes to insert a 624 * random section of illegal instructions. 625 */ 626 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 627 pages = size / PAGE_SIZE; 628 629 if (bpf_jit_charge_modmem(pages)) 630 return NULL; 631 hdr = module_alloc(size); 632 if (!hdr) { 633 bpf_jit_uncharge_modmem(pages); 634 return NULL; 635 } 636 637 /* Fill space with illegal/arch-dep instructions. */ 638 bpf_fill_ill_insns(hdr, size); 639 640 hdr->pages = pages; 641 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 642 PAGE_SIZE - sizeof(*hdr)); 643 start = (get_random_int() % hole) & ~(alignment - 1); 644 645 /* Leave a random number of instructions before BPF code. */ 646 *image_ptr = &hdr->image[start]; 647 648 return hdr; 649 } 650 651 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 652 { 653 u32 pages = hdr->pages; 654 655 module_memfree(hdr); 656 bpf_jit_uncharge_modmem(pages); 657 } 658 659 /* This symbol is only overridden by archs that have different 660 * requirements than the usual eBPF JITs, f.e. when they only 661 * implement cBPF JIT, do not set images read-only, etc. 662 */ 663 void __weak bpf_jit_free(struct bpf_prog *fp) 664 { 665 if (fp->jited) { 666 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 667 668 bpf_jit_binary_unlock_ro(hdr); 669 bpf_jit_binary_free(hdr); 670 671 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 672 } 673 674 bpf_prog_unlock_free(fp); 675 } 676 677 static int bpf_jit_blind_insn(const struct bpf_insn *from, 678 const struct bpf_insn *aux, 679 struct bpf_insn *to_buff) 680 { 681 struct bpf_insn *to = to_buff; 682 u32 imm_rnd = get_random_int(); 683 s16 off; 684 685 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 686 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 687 688 if (from->imm == 0 && 689 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 690 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 691 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 692 goto out; 693 } 694 695 switch (from->code) { 696 case BPF_ALU | BPF_ADD | BPF_K: 697 case BPF_ALU | BPF_SUB | BPF_K: 698 case BPF_ALU | BPF_AND | BPF_K: 699 case BPF_ALU | BPF_OR | BPF_K: 700 case BPF_ALU | BPF_XOR | BPF_K: 701 case BPF_ALU | BPF_MUL | BPF_K: 702 case BPF_ALU | BPF_MOV | BPF_K: 703 case BPF_ALU | BPF_DIV | BPF_K: 704 case BPF_ALU | BPF_MOD | BPF_K: 705 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 706 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 707 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 708 break; 709 710 case BPF_ALU64 | BPF_ADD | BPF_K: 711 case BPF_ALU64 | BPF_SUB | BPF_K: 712 case BPF_ALU64 | BPF_AND | BPF_K: 713 case BPF_ALU64 | BPF_OR | BPF_K: 714 case BPF_ALU64 | BPF_XOR | BPF_K: 715 case BPF_ALU64 | BPF_MUL | BPF_K: 716 case BPF_ALU64 | BPF_MOV | BPF_K: 717 case BPF_ALU64 | BPF_DIV | BPF_K: 718 case BPF_ALU64 | BPF_MOD | BPF_K: 719 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 720 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 721 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 722 break; 723 724 case BPF_JMP | BPF_JEQ | BPF_K: 725 case BPF_JMP | BPF_JNE | BPF_K: 726 case BPF_JMP | BPF_JGT | BPF_K: 727 case BPF_JMP | BPF_JLT | BPF_K: 728 case BPF_JMP | BPF_JGE | BPF_K: 729 case BPF_JMP | BPF_JLE | BPF_K: 730 case BPF_JMP | BPF_JSGT | BPF_K: 731 case BPF_JMP | BPF_JSLT | BPF_K: 732 case BPF_JMP | BPF_JSGE | BPF_K: 733 case BPF_JMP | BPF_JSLE | BPF_K: 734 case BPF_JMP | BPF_JSET | BPF_K: 735 /* Accommodate for extra offset in case of a backjump. */ 736 off = from->off; 737 if (off < 0) 738 off -= 2; 739 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 740 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 741 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 742 break; 743 744 case BPF_LD | BPF_IMM | BPF_DW: 745 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 746 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 747 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 748 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 749 break; 750 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 751 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 752 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 753 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 754 break; 755 756 case BPF_ST | BPF_MEM | BPF_DW: 757 case BPF_ST | BPF_MEM | BPF_W: 758 case BPF_ST | BPF_MEM | BPF_H: 759 case BPF_ST | BPF_MEM | BPF_B: 760 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 761 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 762 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 763 break; 764 } 765 out: 766 return to - to_buff; 767 } 768 769 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 770 gfp_t gfp_extra_flags) 771 { 772 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 773 struct bpf_prog *fp; 774 775 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 776 if (fp != NULL) { 777 /* aux->prog still points to the fp_other one, so 778 * when promoting the clone to the real program, 779 * this still needs to be adapted. 780 */ 781 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 782 } 783 784 return fp; 785 } 786 787 static void bpf_prog_clone_free(struct bpf_prog *fp) 788 { 789 /* aux was stolen by the other clone, so we cannot free 790 * it from this path! It will be freed eventually by the 791 * other program on release. 792 * 793 * At this point, we don't need a deferred release since 794 * clone is guaranteed to not be locked. 795 */ 796 fp->aux = NULL; 797 __bpf_prog_free(fp); 798 } 799 800 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 801 { 802 /* We have to repoint aux->prog to self, as we don't 803 * know whether fp here is the clone or the original. 804 */ 805 fp->aux->prog = fp; 806 bpf_prog_clone_free(fp_other); 807 } 808 809 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 810 { 811 struct bpf_insn insn_buff[16], aux[2]; 812 struct bpf_prog *clone, *tmp; 813 int insn_delta, insn_cnt; 814 struct bpf_insn *insn; 815 int i, rewritten; 816 817 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 818 return prog; 819 820 clone = bpf_prog_clone_create(prog, GFP_USER); 821 if (!clone) 822 return ERR_PTR(-ENOMEM); 823 824 insn_cnt = clone->len; 825 insn = clone->insnsi; 826 827 for (i = 0; i < insn_cnt; i++, insn++) { 828 /* We temporarily need to hold the original ld64 insn 829 * so that we can still access the first part in the 830 * second blinding run. 831 */ 832 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 833 insn[1].code == 0) 834 memcpy(aux, insn, sizeof(aux)); 835 836 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff); 837 if (!rewritten) 838 continue; 839 840 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 841 if (!tmp) { 842 /* Patching may have repointed aux->prog during 843 * realloc from the original one, so we need to 844 * fix it up here on error. 845 */ 846 bpf_jit_prog_release_other(prog, clone); 847 return ERR_PTR(-ENOMEM); 848 } 849 850 clone = tmp; 851 insn_delta = rewritten - 1; 852 853 /* Walk new program and skip insns we just inserted. */ 854 insn = clone->insnsi + i + insn_delta; 855 insn_cnt += insn_delta; 856 i += insn_delta; 857 } 858 859 clone->blinded = 1; 860 return clone; 861 } 862 #endif /* CONFIG_BPF_JIT */ 863 864 /* Base function for offset calculation. Needs to go into .text section, 865 * therefore keeping it non-static as well; will also be used by JITs 866 * anyway later on, so do not let the compiler omit it. This also needs 867 * to go into kallsyms for correlation from e.g. bpftool, so naming 868 * must not change. 869 */ 870 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 871 { 872 return 0; 873 } 874 EXPORT_SYMBOL_GPL(__bpf_call_base); 875 876 /* All UAPI available opcodes. */ 877 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 878 /* 32 bit ALU operations. */ \ 879 /* Register based. */ \ 880 INSN_3(ALU, ADD, X), \ 881 INSN_3(ALU, SUB, X), \ 882 INSN_3(ALU, AND, X), \ 883 INSN_3(ALU, OR, X), \ 884 INSN_3(ALU, LSH, X), \ 885 INSN_3(ALU, RSH, X), \ 886 INSN_3(ALU, XOR, X), \ 887 INSN_3(ALU, MUL, X), \ 888 INSN_3(ALU, MOV, X), \ 889 INSN_3(ALU, DIV, X), \ 890 INSN_3(ALU, MOD, X), \ 891 INSN_2(ALU, NEG), \ 892 INSN_3(ALU, END, TO_BE), \ 893 INSN_3(ALU, END, TO_LE), \ 894 /* Immediate based. */ \ 895 INSN_3(ALU, ADD, K), \ 896 INSN_3(ALU, SUB, K), \ 897 INSN_3(ALU, AND, K), \ 898 INSN_3(ALU, OR, K), \ 899 INSN_3(ALU, LSH, K), \ 900 INSN_3(ALU, RSH, K), \ 901 INSN_3(ALU, XOR, K), \ 902 INSN_3(ALU, MUL, K), \ 903 INSN_3(ALU, MOV, K), \ 904 INSN_3(ALU, DIV, K), \ 905 INSN_3(ALU, MOD, K), \ 906 /* 64 bit ALU operations. */ \ 907 /* Register based. */ \ 908 INSN_3(ALU64, ADD, X), \ 909 INSN_3(ALU64, SUB, X), \ 910 INSN_3(ALU64, AND, X), \ 911 INSN_3(ALU64, OR, X), \ 912 INSN_3(ALU64, LSH, X), \ 913 INSN_3(ALU64, RSH, X), \ 914 INSN_3(ALU64, XOR, X), \ 915 INSN_3(ALU64, MUL, X), \ 916 INSN_3(ALU64, MOV, X), \ 917 INSN_3(ALU64, ARSH, X), \ 918 INSN_3(ALU64, DIV, X), \ 919 INSN_3(ALU64, MOD, X), \ 920 INSN_2(ALU64, NEG), \ 921 /* Immediate based. */ \ 922 INSN_3(ALU64, ADD, K), \ 923 INSN_3(ALU64, SUB, K), \ 924 INSN_3(ALU64, AND, K), \ 925 INSN_3(ALU64, OR, K), \ 926 INSN_3(ALU64, LSH, K), \ 927 INSN_3(ALU64, RSH, K), \ 928 INSN_3(ALU64, XOR, K), \ 929 INSN_3(ALU64, MUL, K), \ 930 INSN_3(ALU64, MOV, K), \ 931 INSN_3(ALU64, ARSH, K), \ 932 INSN_3(ALU64, DIV, K), \ 933 INSN_3(ALU64, MOD, K), \ 934 /* Call instruction. */ \ 935 INSN_2(JMP, CALL), \ 936 /* Exit instruction. */ \ 937 INSN_2(JMP, EXIT), \ 938 /* Jump instructions. */ \ 939 /* Register based. */ \ 940 INSN_3(JMP, JEQ, X), \ 941 INSN_3(JMP, JNE, X), \ 942 INSN_3(JMP, JGT, X), \ 943 INSN_3(JMP, JLT, X), \ 944 INSN_3(JMP, JGE, X), \ 945 INSN_3(JMP, JLE, X), \ 946 INSN_3(JMP, JSGT, X), \ 947 INSN_3(JMP, JSLT, X), \ 948 INSN_3(JMP, JSGE, X), \ 949 INSN_3(JMP, JSLE, X), \ 950 INSN_3(JMP, JSET, X), \ 951 /* Immediate based. */ \ 952 INSN_3(JMP, JEQ, K), \ 953 INSN_3(JMP, JNE, K), \ 954 INSN_3(JMP, JGT, K), \ 955 INSN_3(JMP, JLT, K), \ 956 INSN_3(JMP, JGE, K), \ 957 INSN_3(JMP, JLE, K), \ 958 INSN_3(JMP, JSGT, K), \ 959 INSN_3(JMP, JSLT, K), \ 960 INSN_3(JMP, JSGE, K), \ 961 INSN_3(JMP, JSLE, K), \ 962 INSN_3(JMP, JSET, K), \ 963 INSN_2(JMP, JA), \ 964 /* Store instructions. */ \ 965 /* Register based. */ \ 966 INSN_3(STX, MEM, B), \ 967 INSN_3(STX, MEM, H), \ 968 INSN_3(STX, MEM, W), \ 969 INSN_3(STX, MEM, DW), \ 970 INSN_3(STX, XADD, W), \ 971 INSN_3(STX, XADD, DW), \ 972 /* Immediate based. */ \ 973 INSN_3(ST, MEM, B), \ 974 INSN_3(ST, MEM, H), \ 975 INSN_3(ST, MEM, W), \ 976 INSN_3(ST, MEM, DW), \ 977 /* Load instructions. */ \ 978 /* Register based. */ \ 979 INSN_3(LDX, MEM, B), \ 980 INSN_3(LDX, MEM, H), \ 981 INSN_3(LDX, MEM, W), \ 982 INSN_3(LDX, MEM, DW), \ 983 /* Immediate based. */ \ 984 INSN_3(LD, IMM, DW) 985 986 bool bpf_opcode_in_insntable(u8 code) 987 { 988 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 989 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 990 static const bool public_insntable[256] = { 991 [0 ... 255] = false, 992 /* Now overwrite non-defaults ... */ 993 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 994 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 995 [BPF_LD | BPF_ABS | BPF_B] = true, 996 [BPF_LD | BPF_ABS | BPF_H] = true, 997 [BPF_LD | BPF_ABS | BPF_W] = true, 998 [BPF_LD | BPF_IND | BPF_B] = true, 999 [BPF_LD | BPF_IND | BPF_H] = true, 1000 [BPF_LD | BPF_IND | BPF_W] = true, 1001 }; 1002 #undef BPF_INSN_3_TBL 1003 #undef BPF_INSN_2_TBL 1004 return public_insntable[code]; 1005 } 1006 1007 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1008 /** 1009 * __bpf_prog_run - run eBPF program on a given context 1010 * @ctx: is the data we are operating on 1011 * @insn: is the array of eBPF instructions 1012 * 1013 * Decode and execute eBPF instructions. 1014 */ 1015 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 1016 { 1017 u64 tmp; 1018 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1019 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1020 static const void *jumptable[256] = { 1021 [0 ... 255] = &&default_label, 1022 /* Now overwrite non-defaults ... */ 1023 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1024 /* Non-UAPI available opcodes. */ 1025 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1026 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1027 }; 1028 #undef BPF_INSN_3_LBL 1029 #undef BPF_INSN_2_LBL 1030 u32 tail_call_cnt = 0; 1031 1032 #define CONT ({ insn++; goto select_insn; }) 1033 #define CONT_JMP ({ insn++; goto select_insn; }) 1034 1035 select_insn: 1036 goto *jumptable[insn->code]; 1037 1038 /* ALU */ 1039 #define ALU(OPCODE, OP) \ 1040 ALU64_##OPCODE##_X: \ 1041 DST = DST OP SRC; \ 1042 CONT; \ 1043 ALU_##OPCODE##_X: \ 1044 DST = (u32) DST OP (u32) SRC; \ 1045 CONT; \ 1046 ALU64_##OPCODE##_K: \ 1047 DST = DST OP IMM; \ 1048 CONT; \ 1049 ALU_##OPCODE##_K: \ 1050 DST = (u32) DST OP (u32) IMM; \ 1051 CONT; 1052 1053 ALU(ADD, +) 1054 ALU(SUB, -) 1055 ALU(AND, &) 1056 ALU(OR, |) 1057 ALU(LSH, <<) 1058 ALU(RSH, >>) 1059 ALU(XOR, ^) 1060 ALU(MUL, *) 1061 #undef ALU 1062 ALU_NEG: 1063 DST = (u32) -DST; 1064 CONT; 1065 ALU64_NEG: 1066 DST = -DST; 1067 CONT; 1068 ALU_MOV_X: 1069 DST = (u32) SRC; 1070 CONT; 1071 ALU_MOV_K: 1072 DST = (u32) IMM; 1073 CONT; 1074 ALU64_MOV_X: 1075 DST = SRC; 1076 CONT; 1077 ALU64_MOV_K: 1078 DST = IMM; 1079 CONT; 1080 LD_IMM_DW: 1081 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1082 insn++; 1083 CONT; 1084 ALU64_ARSH_X: 1085 (*(s64 *) &DST) >>= SRC; 1086 CONT; 1087 ALU64_ARSH_K: 1088 (*(s64 *) &DST) >>= IMM; 1089 CONT; 1090 ALU64_MOD_X: 1091 div64_u64_rem(DST, SRC, &tmp); 1092 DST = tmp; 1093 CONT; 1094 ALU_MOD_X: 1095 tmp = (u32) DST; 1096 DST = do_div(tmp, (u32) SRC); 1097 CONT; 1098 ALU64_MOD_K: 1099 div64_u64_rem(DST, IMM, &tmp); 1100 DST = tmp; 1101 CONT; 1102 ALU_MOD_K: 1103 tmp = (u32) DST; 1104 DST = do_div(tmp, (u32) IMM); 1105 CONT; 1106 ALU64_DIV_X: 1107 DST = div64_u64(DST, SRC); 1108 CONT; 1109 ALU_DIV_X: 1110 tmp = (u32) DST; 1111 do_div(tmp, (u32) SRC); 1112 DST = (u32) tmp; 1113 CONT; 1114 ALU64_DIV_K: 1115 DST = div64_u64(DST, IMM); 1116 CONT; 1117 ALU_DIV_K: 1118 tmp = (u32) DST; 1119 do_div(tmp, (u32) IMM); 1120 DST = (u32) tmp; 1121 CONT; 1122 ALU_END_TO_BE: 1123 switch (IMM) { 1124 case 16: 1125 DST = (__force u16) cpu_to_be16(DST); 1126 break; 1127 case 32: 1128 DST = (__force u32) cpu_to_be32(DST); 1129 break; 1130 case 64: 1131 DST = (__force u64) cpu_to_be64(DST); 1132 break; 1133 } 1134 CONT; 1135 ALU_END_TO_LE: 1136 switch (IMM) { 1137 case 16: 1138 DST = (__force u16) cpu_to_le16(DST); 1139 break; 1140 case 32: 1141 DST = (__force u32) cpu_to_le32(DST); 1142 break; 1143 case 64: 1144 DST = (__force u64) cpu_to_le64(DST); 1145 break; 1146 } 1147 CONT; 1148 1149 /* CALL */ 1150 JMP_CALL: 1151 /* Function call scratches BPF_R1-BPF_R5 registers, 1152 * preserves BPF_R6-BPF_R9, and stores return value 1153 * into BPF_R0. 1154 */ 1155 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1156 BPF_R4, BPF_R5); 1157 CONT; 1158 1159 JMP_CALL_ARGS: 1160 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1161 BPF_R3, BPF_R4, 1162 BPF_R5, 1163 insn + insn->off + 1); 1164 CONT; 1165 1166 JMP_TAIL_CALL: { 1167 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1168 struct bpf_array *array = container_of(map, struct bpf_array, map); 1169 struct bpf_prog *prog; 1170 u32 index = BPF_R3; 1171 1172 if (unlikely(index >= array->map.max_entries)) 1173 goto out; 1174 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1175 goto out; 1176 1177 tail_call_cnt++; 1178 1179 prog = READ_ONCE(array->ptrs[index]); 1180 if (!prog) 1181 goto out; 1182 1183 /* ARG1 at this point is guaranteed to point to CTX from 1184 * the verifier side due to the fact that the tail call is 1185 * handeled like a helper, that is, bpf_tail_call_proto, 1186 * where arg1_type is ARG_PTR_TO_CTX. 1187 */ 1188 insn = prog->insnsi; 1189 goto select_insn; 1190 out: 1191 CONT; 1192 } 1193 /* JMP */ 1194 JMP_JA: 1195 insn += insn->off; 1196 CONT; 1197 JMP_JEQ_X: 1198 if (DST == SRC) { 1199 insn += insn->off; 1200 CONT_JMP; 1201 } 1202 CONT; 1203 JMP_JEQ_K: 1204 if (DST == IMM) { 1205 insn += insn->off; 1206 CONT_JMP; 1207 } 1208 CONT; 1209 JMP_JNE_X: 1210 if (DST != SRC) { 1211 insn += insn->off; 1212 CONT_JMP; 1213 } 1214 CONT; 1215 JMP_JNE_K: 1216 if (DST != IMM) { 1217 insn += insn->off; 1218 CONT_JMP; 1219 } 1220 CONT; 1221 JMP_JGT_X: 1222 if (DST > SRC) { 1223 insn += insn->off; 1224 CONT_JMP; 1225 } 1226 CONT; 1227 JMP_JGT_K: 1228 if (DST > IMM) { 1229 insn += insn->off; 1230 CONT_JMP; 1231 } 1232 CONT; 1233 JMP_JLT_X: 1234 if (DST < SRC) { 1235 insn += insn->off; 1236 CONT_JMP; 1237 } 1238 CONT; 1239 JMP_JLT_K: 1240 if (DST < IMM) { 1241 insn += insn->off; 1242 CONT_JMP; 1243 } 1244 CONT; 1245 JMP_JGE_X: 1246 if (DST >= SRC) { 1247 insn += insn->off; 1248 CONT_JMP; 1249 } 1250 CONT; 1251 JMP_JGE_K: 1252 if (DST >= IMM) { 1253 insn += insn->off; 1254 CONT_JMP; 1255 } 1256 CONT; 1257 JMP_JLE_X: 1258 if (DST <= SRC) { 1259 insn += insn->off; 1260 CONT_JMP; 1261 } 1262 CONT; 1263 JMP_JLE_K: 1264 if (DST <= IMM) { 1265 insn += insn->off; 1266 CONT_JMP; 1267 } 1268 CONT; 1269 JMP_JSGT_X: 1270 if (((s64) DST) > ((s64) SRC)) { 1271 insn += insn->off; 1272 CONT_JMP; 1273 } 1274 CONT; 1275 JMP_JSGT_K: 1276 if (((s64) DST) > ((s64) IMM)) { 1277 insn += insn->off; 1278 CONT_JMP; 1279 } 1280 CONT; 1281 JMP_JSLT_X: 1282 if (((s64) DST) < ((s64) SRC)) { 1283 insn += insn->off; 1284 CONT_JMP; 1285 } 1286 CONT; 1287 JMP_JSLT_K: 1288 if (((s64) DST) < ((s64) IMM)) { 1289 insn += insn->off; 1290 CONT_JMP; 1291 } 1292 CONT; 1293 JMP_JSGE_X: 1294 if (((s64) DST) >= ((s64) SRC)) { 1295 insn += insn->off; 1296 CONT_JMP; 1297 } 1298 CONT; 1299 JMP_JSGE_K: 1300 if (((s64) DST) >= ((s64) IMM)) { 1301 insn += insn->off; 1302 CONT_JMP; 1303 } 1304 CONT; 1305 JMP_JSLE_X: 1306 if (((s64) DST) <= ((s64) SRC)) { 1307 insn += insn->off; 1308 CONT_JMP; 1309 } 1310 CONT; 1311 JMP_JSLE_K: 1312 if (((s64) DST) <= ((s64) IMM)) { 1313 insn += insn->off; 1314 CONT_JMP; 1315 } 1316 CONT; 1317 JMP_JSET_X: 1318 if (DST & SRC) { 1319 insn += insn->off; 1320 CONT_JMP; 1321 } 1322 CONT; 1323 JMP_JSET_K: 1324 if (DST & IMM) { 1325 insn += insn->off; 1326 CONT_JMP; 1327 } 1328 CONT; 1329 JMP_EXIT: 1330 return BPF_R0; 1331 1332 /* STX and ST and LDX*/ 1333 #define LDST(SIZEOP, SIZE) \ 1334 STX_MEM_##SIZEOP: \ 1335 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1336 CONT; \ 1337 ST_MEM_##SIZEOP: \ 1338 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1339 CONT; \ 1340 LDX_MEM_##SIZEOP: \ 1341 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1342 CONT; 1343 1344 LDST(B, u8) 1345 LDST(H, u16) 1346 LDST(W, u32) 1347 LDST(DW, u64) 1348 #undef LDST 1349 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1350 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1351 (DST + insn->off)); 1352 CONT; 1353 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1354 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1355 (DST + insn->off)); 1356 CONT; 1357 1358 default_label: 1359 /* If we ever reach this, we have a bug somewhere. Die hard here 1360 * instead of just returning 0; we could be somewhere in a subprog, 1361 * so execution could continue otherwise which we do /not/ want. 1362 * 1363 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1364 */ 1365 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code); 1366 BUG_ON(1); 1367 return 0; 1368 } 1369 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */ 1370 1371 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1372 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1373 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1374 { \ 1375 u64 stack[stack_size / sizeof(u64)]; \ 1376 u64 regs[MAX_BPF_REG]; \ 1377 \ 1378 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1379 ARG1 = (u64) (unsigned long) ctx; \ 1380 return ___bpf_prog_run(regs, insn, stack); \ 1381 } 1382 1383 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1384 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1385 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1386 const struct bpf_insn *insn) \ 1387 { \ 1388 u64 stack[stack_size / sizeof(u64)]; \ 1389 u64 regs[MAX_BPF_REG]; \ 1390 \ 1391 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1392 BPF_R1 = r1; \ 1393 BPF_R2 = r2; \ 1394 BPF_R3 = r3; \ 1395 BPF_R4 = r4; \ 1396 BPF_R5 = r5; \ 1397 return ___bpf_prog_run(regs, insn, stack); \ 1398 } 1399 1400 #define EVAL1(FN, X) FN(X) 1401 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1402 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1403 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1404 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1405 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1406 1407 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1408 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1409 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1410 1411 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1412 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1413 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1414 1415 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1416 1417 static unsigned int (*interpreters[])(const void *ctx, 1418 const struct bpf_insn *insn) = { 1419 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1420 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1421 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1422 }; 1423 #undef PROG_NAME_LIST 1424 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1425 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1426 const struct bpf_insn *insn) = { 1427 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1428 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1429 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1430 }; 1431 #undef PROG_NAME_LIST 1432 1433 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1434 { 1435 stack_depth = max_t(u32, stack_depth, 1); 1436 insn->off = (s16) insn->imm; 1437 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1438 __bpf_call_base_args; 1439 insn->code = BPF_JMP | BPF_CALL_ARGS; 1440 } 1441 1442 #else 1443 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1444 const struct bpf_insn *insn) 1445 { 1446 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1447 * is not working properly, so warn about it! 1448 */ 1449 WARN_ON_ONCE(1); 1450 return 0; 1451 } 1452 #endif 1453 1454 bool bpf_prog_array_compatible(struct bpf_array *array, 1455 const struct bpf_prog *fp) 1456 { 1457 if (fp->kprobe_override) 1458 return false; 1459 1460 if (!array->owner_prog_type) { 1461 /* There's no owner yet where we could check for 1462 * compatibility. 1463 */ 1464 array->owner_prog_type = fp->type; 1465 array->owner_jited = fp->jited; 1466 1467 return true; 1468 } 1469 1470 return array->owner_prog_type == fp->type && 1471 array->owner_jited == fp->jited; 1472 } 1473 1474 static int bpf_check_tail_call(const struct bpf_prog *fp) 1475 { 1476 struct bpf_prog_aux *aux = fp->aux; 1477 int i; 1478 1479 for (i = 0; i < aux->used_map_cnt; i++) { 1480 struct bpf_map *map = aux->used_maps[i]; 1481 struct bpf_array *array; 1482 1483 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1484 continue; 1485 1486 array = container_of(map, struct bpf_array, map); 1487 if (!bpf_prog_array_compatible(array, fp)) 1488 return -EINVAL; 1489 } 1490 1491 return 0; 1492 } 1493 1494 static void bpf_prog_select_func(struct bpf_prog *fp) 1495 { 1496 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1497 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1498 1499 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1500 #else 1501 fp->bpf_func = __bpf_prog_ret0_warn; 1502 #endif 1503 } 1504 1505 /** 1506 * bpf_prog_select_runtime - select exec runtime for BPF program 1507 * @fp: bpf_prog populated with internal BPF program 1508 * @err: pointer to error variable 1509 * 1510 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1511 * The BPF program will be executed via BPF_PROG_RUN() macro. 1512 */ 1513 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1514 { 1515 /* In case of BPF to BPF calls, verifier did all the prep 1516 * work with regards to JITing, etc. 1517 */ 1518 if (fp->bpf_func) 1519 goto finalize; 1520 1521 bpf_prog_select_func(fp); 1522 1523 /* eBPF JITs can rewrite the program in case constant 1524 * blinding is active. However, in case of error during 1525 * blinding, bpf_int_jit_compile() must always return a 1526 * valid program, which in this case would simply not 1527 * be JITed, but falls back to the interpreter. 1528 */ 1529 if (!bpf_prog_is_dev_bound(fp->aux)) { 1530 fp = bpf_int_jit_compile(fp); 1531 #ifdef CONFIG_BPF_JIT_ALWAYS_ON 1532 if (!fp->jited) { 1533 *err = -ENOTSUPP; 1534 return fp; 1535 } 1536 #endif 1537 } else { 1538 *err = bpf_prog_offload_compile(fp); 1539 if (*err) 1540 return fp; 1541 } 1542 1543 finalize: 1544 bpf_prog_lock_ro(fp); 1545 1546 /* The tail call compatibility check can only be done at 1547 * this late stage as we need to determine, if we deal 1548 * with JITed or non JITed program concatenations and not 1549 * all eBPF JITs might immediately support all features. 1550 */ 1551 *err = bpf_check_tail_call(fp); 1552 1553 return fp; 1554 } 1555 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1556 1557 static unsigned int __bpf_prog_ret1(const void *ctx, 1558 const struct bpf_insn *insn) 1559 { 1560 return 1; 1561 } 1562 1563 static struct bpf_prog_dummy { 1564 struct bpf_prog prog; 1565 } dummy_bpf_prog = { 1566 .prog = { 1567 .bpf_func = __bpf_prog_ret1, 1568 }, 1569 }; 1570 1571 /* to avoid allocating empty bpf_prog_array for cgroups that 1572 * don't have bpf program attached use one global 'empty_prog_array' 1573 * It will not be modified the caller of bpf_prog_array_alloc() 1574 * (since caller requested prog_cnt == 0) 1575 * that pointer should be 'freed' by bpf_prog_array_free() 1576 */ 1577 static struct { 1578 struct bpf_prog_array hdr; 1579 struct bpf_prog *null_prog; 1580 } empty_prog_array = { 1581 .null_prog = NULL, 1582 }; 1583 1584 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1585 { 1586 if (prog_cnt) 1587 return kzalloc(sizeof(struct bpf_prog_array) + 1588 sizeof(struct bpf_prog_array_item) * 1589 (prog_cnt + 1), 1590 flags); 1591 1592 return &empty_prog_array.hdr; 1593 } 1594 1595 void bpf_prog_array_free(struct bpf_prog_array __rcu *progs) 1596 { 1597 if (!progs || 1598 progs == (struct bpf_prog_array __rcu *)&empty_prog_array.hdr) 1599 return; 1600 kfree_rcu(progs, rcu); 1601 } 1602 1603 int bpf_prog_array_length(struct bpf_prog_array __rcu *array) 1604 { 1605 struct bpf_prog_array_item *item; 1606 u32 cnt = 0; 1607 1608 rcu_read_lock(); 1609 item = rcu_dereference(array)->items; 1610 for (; item->prog; item++) 1611 if (item->prog != &dummy_bpf_prog.prog) 1612 cnt++; 1613 rcu_read_unlock(); 1614 return cnt; 1615 } 1616 1617 1618 static bool bpf_prog_array_copy_core(struct bpf_prog_array __rcu *array, 1619 u32 *prog_ids, 1620 u32 request_cnt) 1621 { 1622 struct bpf_prog_array_item *item; 1623 int i = 0; 1624 1625 item = rcu_dereference_check(array, 1)->items; 1626 for (; item->prog; item++) { 1627 if (item->prog == &dummy_bpf_prog.prog) 1628 continue; 1629 prog_ids[i] = item->prog->aux->id; 1630 if (++i == request_cnt) { 1631 item++; 1632 break; 1633 } 1634 } 1635 1636 return !!(item->prog); 1637 } 1638 1639 int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *array, 1640 __u32 __user *prog_ids, u32 cnt) 1641 { 1642 unsigned long err = 0; 1643 bool nospc; 1644 u32 *ids; 1645 1646 /* users of this function are doing: 1647 * cnt = bpf_prog_array_length(); 1648 * if (cnt > 0) 1649 * bpf_prog_array_copy_to_user(..., cnt); 1650 * so below kcalloc doesn't need extra cnt > 0 check, but 1651 * bpf_prog_array_length() releases rcu lock and 1652 * prog array could have been swapped with empty or larger array, 1653 * so always copy 'cnt' prog_ids to the user. 1654 * In a rare race the user will see zero prog_ids 1655 */ 1656 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 1657 if (!ids) 1658 return -ENOMEM; 1659 rcu_read_lock(); 1660 nospc = bpf_prog_array_copy_core(array, ids, cnt); 1661 rcu_read_unlock(); 1662 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 1663 kfree(ids); 1664 if (err) 1665 return -EFAULT; 1666 if (nospc) 1667 return -ENOSPC; 1668 return 0; 1669 } 1670 1671 void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *array, 1672 struct bpf_prog *old_prog) 1673 { 1674 struct bpf_prog_array_item *item = array->items; 1675 1676 for (; item->prog; item++) 1677 if (item->prog == old_prog) { 1678 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 1679 break; 1680 } 1681 } 1682 1683 int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array, 1684 struct bpf_prog *exclude_prog, 1685 struct bpf_prog *include_prog, 1686 struct bpf_prog_array **new_array) 1687 { 1688 int new_prog_cnt, carry_prog_cnt = 0; 1689 struct bpf_prog_array_item *existing; 1690 struct bpf_prog_array *array; 1691 bool found_exclude = false; 1692 int new_prog_idx = 0; 1693 1694 /* Figure out how many existing progs we need to carry over to 1695 * the new array. 1696 */ 1697 if (old_array) { 1698 existing = old_array->items; 1699 for (; existing->prog; existing++) { 1700 if (existing->prog == exclude_prog) { 1701 found_exclude = true; 1702 continue; 1703 } 1704 if (existing->prog != &dummy_bpf_prog.prog) 1705 carry_prog_cnt++; 1706 if (existing->prog == include_prog) 1707 return -EEXIST; 1708 } 1709 } 1710 1711 if (exclude_prog && !found_exclude) 1712 return -ENOENT; 1713 1714 /* How many progs (not NULL) will be in the new array? */ 1715 new_prog_cnt = carry_prog_cnt; 1716 if (include_prog) 1717 new_prog_cnt += 1; 1718 1719 /* Do we have any prog (not NULL) in the new array? */ 1720 if (!new_prog_cnt) { 1721 *new_array = NULL; 1722 return 0; 1723 } 1724 1725 /* +1 as the end of prog_array is marked with NULL */ 1726 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1727 if (!array) 1728 return -ENOMEM; 1729 1730 /* Fill in the new prog array */ 1731 if (carry_prog_cnt) { 1732 existing = old_array->items; 1733 for (; existing->prog; existing++) 1734 if (existing->prog != exclude_prog && 1735 existing->prog != &dummy_bpf_prog.prog) { 1736 array->items[new_prog_idx++].prog = 1737 existing->prog; 1738 } 1739 } 1740 if (include_prog) 1741 array->items[new_prog_idx++].prog = include_prog; 1742 array->items[new_prog_idx].prog = NULL; 1743 *new_array = array; 1744 return 0; 1745 } 1746 1747 int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array, 1748 u32 *prog_ids, u32 request_cnt, 1749 u32 *prog_cnt) 1750 { 1751 u32 cnt = 0; 1752 1753 if (array) 1754 cnt = bpf_prog_array_length(array); 1755 1756 *prog_cnt = cnt; 1757 1758 /* return early if user requested only program count or nothing to copy */ 1759 if (!request_cnt || !cnt) 1760 return 0; 1761 1762 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 1763 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 1764 : 0; 1765 } 1766 1767 static void bpf_prog_free_deferred(struct work_struct *work) 1768 { 1769 struct bpf_prog_aux *aux; 1770 int i; 1771 1772 aux = container_of(work, struct bpf_prog_aux, work); 1773 if (bpf_prog_is_dev_bound(aux)) 1774 bpf_prog_offload_destroy(aux->prog); 1775 #ifdef CONFIG_PERF_EVENTS 1776 if (aux->prog->has_callchain_buf) 1777 put_callchain_buffers(); 1778 #endif 1779 for (i = 0; i < aux->func_cnt; i++) 1780 bpf_jit_free(aux->func[i]); 1781 if (aux->func_cnt) { 1782 kfree(aux->func); 1783 bpf_prog_unlock_free(aux->prog); 1784 } else { 1785 bpf_jit_free(aux->prog); 1786 } 1787 } 1788 1789 /* Free internal BPF program */ 1790 void bpf_prog_free(struct bpf_prog *fp) 1791 { 1792 struct bpf_prog_aux *aux = fp->aux; 1793 1794 INIT_WORK(&aux->work, bpf_prog_free_deferred); 1795 schedule_work(&aux->work); 1796 } 1797 EXPORT_SYMBOL_GPL(bpf_prog_free); 1798 1799 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 1800 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 1801 1802 void bpf_user_rnd_init_once(void) 1803 { 1804 prandom_init_once(&bpf_user_rnd_state); 1805 } 1806 1807 BPF_CALL_0(bpf_user_rnd_u32) 1808 { 1809 /* Should someone ever have the rather unwise idea to use some 1810 * of the registers passed into this function, then note that 1811 * this function is called from native eBPF and classic-to-eBPF 1812 * transformations. Register assignments from both sides are 1813 * different, f.e. classic always sets fn(ctx, A, X) here. 1814 */ 1815 struct rnd_state *state; 1816 u32 res; 1817 1818 state = &get_cpu_var(bpf_user_rnd_state); 1819 res = prandom_u32_state(state); 1820 put_cpu_var(bpf_user_rnd_state); 1821 1822 return res; 1823 } 1824 1825 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 1826 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 1827 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 1828 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 1829 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 1830 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 1831 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 1832 1833 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 1834 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 1835 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 1836 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 1837 1838 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 1839 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 1840 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 1841 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 1842 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 1843 1844 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 1845 { 1846 return NULL; 1847 } 1848 1849 u64 __weak 1850 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 1851 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 1852 { 1853 return -ENOTSUPP; 1854 } 1855 EXPORT_SYMBOL_GPL(bpf_event_output); 1856 1857 /* Always built-in helper functions. */ 1858 const struct bpf_func_proto bpf_tail_call_proto = { 1859 .func = NULL, 1860 .gpl_only = false, 1861 .ret_type = RET_VOID, 1862 .arg1_type = ARG_PTR_TO_CTX, 1863 .arg2_type = ARG_CONST_MAP_PTR, 1864 .arg3_type = ARG_ANYTHING, 1865 }; 1866 1867 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 1868 * It is encouraged to implement bpf_int_jit_compile() instead, so that 1869 * eBPF and implicitly also cBPF can get JITed! 1870 */ 1871 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 1872 { 1873 return prog; 1874 } 1875 1876 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 1877 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 1878 */ 1879 void __weak bpf_jit_compile(struct bpf_prog *prog) 1880 { 1881 } 1882 1883 bool __weak bpf_helper_changes_pkt_data(void *func) 1884 { 1885 return false; 1886 } 1887 1888 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 1889 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 1890 */ 1891 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 1892 int len) 1893 { 1894 return -EFAULT; 1895 } 1896 1897 /* All definitions of tracepoints related to BPF. */ 1898 #define CREATE_TRACE_POINTS 1899 #include <linux/bpf_trace.h> 1900 1901 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 1902