1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Linux Socket Filter - Kernel level socket filtering 4 * 5 * Based on the design of the Berkeley Packet Filter. The new 6 * internal format has been designed by PLUMgrid: 7 * 8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 9 * 10 * Authors: 11 * 12 * Jay Schulist <jschlst@samba.org> 13 * Alexei Starovoitov <ast@plumgrid.com> 14 * Daniel Borkmann <dborkman@redhat.com> 15 * 16 * Andi Kleen - Fix a few bad bugs and races. 17 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 18 */ 19 20 #include <uapi/linux/btf.h> 21 #include <linux/filter.h> 22 #include <linux/skbuff.h> 23 #include <linux/vmalloc.h> 24 #include <linux/random.h> 25 #include <linux/moduleloader.h> 26 #include <linux/bpf.h> 27 #include <linux/btf.h> 28 #include <linux/objtool.h> 29 #include <linux/rbtree_latch.h> 30 #include <linux/kallsyms.h> 31 #include <linux/rcupdate.h> 32 #include <linux/perf_event.h> 33 #include <linux/extable.h> 34 #include <linux/log2.h> 35 #include <linux/bpf_verifier.h> 36 37 #include <asm/barrier.h> 38 #include <asm/unaligned.h> 39 40 /* Registers */ 41 #define BPF_R0 regs[BPF_REG_0] 42 #define BPF_R1 regs[BPF_REG_1] 43 #define BPF_R2 regs[BPF_REG_2] 44 #define BPF_R3 regs[BPF_REG_3] 45 #define BPF_R4 regs[BPF_REG_4] 46 #define BPF_R5 regs[BPF_REG_5] 47 #define BPF_R6 regs[BPF_REG_6] 48 #define BPF_R7 regs[BPF_REG_7] 49 #define BPF_R8 regs[BPF_REG_8] 50 #define BPF_R9 regs[BPF_REG_9] 51 #define BPF_R10 regs[BPF_REG_10] 52 53 /* Named registers */ 54 #define DST regs[insn->dst_reg] 55 #define SRC regs[insn->src_reg] 56 #define FP regs[BPF_REG_FP] 57 #define AX regs[BPF_REG_AX] 58 #define ARG1 regs[BPF_REG_ARG1] 59 #define CTX regs[BPF_REG_CTX] 60 #define IMM insn->imm 61 62 /* No hurry in this branch 63 * 64 * Exported for the bpf jit load helper. 65 */ 66 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 67 { 68 u8 *ptr = NULL; 69 70 if (k >= SKF_NET_OFF) 71 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 72 else if (k >= SKF_LL_OFF) 73 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 74 75 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 76 return ptr; 77 78 return NULL; 79 } 80 81 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 82 { 83 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 84 struct bpf_prog_aux *aux; 85 struct bpf_prog *fp; 86 87 size = round_up(size, PAGE_SIZE); 88 fp = __vmalloc(size, gfp_flags); 89 if (fp == NULL) 90 return NULL; 91 92 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags); 93 if (aux == NULL) { 94 vfree(fp); 95 return NULL; 96 } 97 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags); 98 if (!fp->active) { 99 vfree(fp); 100 kfree(aux); 101 return NULL; 102 } 103 104 fp->pages = size / PAGE_SIZE; 105 fp->aux = aux; 106 fp->aux->prog = fp; 107 fp->jit_requested = ebpf_jit_enabled(); 108 109 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 110 mutex_init(&fp->aux->used_maps_mutex); 111 mutex_init(&fp->aux->dst_mutex); 112 113 return fp; 114 } 115 116 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 117 { 118 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 119 struct bpf_prog *prog; 120 int cpu; 121 122 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 123 if (!prog) 124 return NULL; 125 126 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 127 if (!prog->stats) { 128 free_percpu(prog->active); 129 kfree(prog->aux); 130 vfree(prog); 131 return NULL; 132 } 133 134 for_each_possible_cpu(cpu) { 135 struct bpf_prog_stats *pstats; 136 137 pstats = per_cpu_ptr(prog->stats, cpu); 138 u64_stats_init(&pstats->syncp); 139 } 140 return prog; 141 } 142 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 143 144 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 145 { 146 if (!prog->aux->nr_linfo || !prog->jit_requested) 147 return 0; 148 149 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 150 sizeof(*prog->aux->jited_linfo), 151 GFP_KERNEL_ACCOUNT | __GFP_NOWARN); 152 if (!prog->aux->jited_linfo) 153 return -ENOMEM; 154 155 return 0; 156 } 157 158 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 159 { 160 if (prog->aux->jited_linfo && 161 (!prog->jited || !prog->aux->jited_linfo[0])) { 162 kvfree(prog->aux->jited_linfo); 163 prog->aux->jited_linfo = NULL; 164 } 165 166 kfree(prog->aux->kfunc_tab); 167 prog->aux->kfunc_tab = NULL; 168 } 169 170 /* The jit engine is responsible to provide an array 171 * for insn_off to the jited_off mapping (insn_to_jit_off). 172 * 173 * The idx to this array is the insn_off. Hence, the insn_off 174 * here is relative to the prog itself instead of the main prog. 175 * This array has one entry for each xlated bpf insn. 176 * 177 * jited_off is the byte off to the last byte of the jited insn. 178 * 179 * Hence, with 180 * insn_start: 181 * The first bpf insn off of the prog. The insn off 182 * here is relative to the main prog. 183 * e.g. if prog is a subprog, insn_start > 0 184 * linfo_idx: 185 * The prog's idx to prog->aux->linfo and jited_linfo 186 * 187 * jited_linfo[linfo_idx] = prog->bpf_func 188 * 189 * For i > linfo_idx, 190 * 191 * jited_linfo[i] = prog->bpf_func + 192 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 193 */ 194 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 195 const u32 *insn_to_jit_off) 196 { 197 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 198 const struct bpf_line_info *linfo; 199 void **jited_linfo; 200 201 if (!prog->aux->jited_linfo) 202 /* Userspace did not provide linfo */ 203 return; 204 205 linfo_idx = prog->aux->linfo_idx; 206 linfo = &prog->aux->linfo[linfo_idx]; 207 insn_start = linfo[0].insn_off; 208 insn_end = insn_start + prog->len; 209 210 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 211 jited_linfo[0] = prog->bpf_func; 212 213 nr_linfo = prog->aux->nr_linfo - linfo_idx; 214 215 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 216 /* The verifier ensures that linfo[i].insn_off is 217 * strictly increasing 218 */ 219 jited_linfo[i] = prog->bpf_func + 220 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 221 } 222 223 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 224 gfp_t gfp_extra_flags) 225 { 226 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 227 struct bpf_prog *fp; 228 u32 pages; 229 230 size = round_up(size, PAGE_SIZE); 231 pages = size / PAGE_SIZE; 232 if (pages <= fp_old->pages) 233 return fp_old; 234 235 fp = __vmalloc(size, gfp_flags); 236 if (fp) { 237 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 238 fp->pages = pages; 239 fp->aux->prog = fp; 240 241 /* We keep fp->aux from fp_old around in the new 242 * reallocated structure. 243 */ 244 fp_old->aux = NULL; 245 fp_old->stats = NULL; 246 fp_old->active = NULL; 247 __bpf_prog_free(fp_old); 248 } 249 250 return fp; 251 } 252 253 void __bpf_prog_free(struct bpf_prog *fp) 254 { 255 if (fp->aux) { 256 mutex_destroy(&fp->aux->used_maps_mutex); 257 mutex_destroy(&fp->aux->dst_mutex); 258 kfree(fp->aux->poke_tab); 259 kfree(fp->aux); 260 } 261 free_percpu(fp->stats); 262 free_percpu(fp->active); 263 vfree(fp); 264 } 265 266 int bpf_prog_calc_tag(struct bpf_prog *fp) 267 { 268 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 269 u32 raw_size = bpf_prog_tag_scratch_size(fp); 270 u32 digest[SHA1_DIGEST_WORDS]; 271 u32 ws[SHA1_WORKSPACE_WORDS]; 272 u32 i, bsize, psize, blocks; 273 struct bpf_insn *dst; 274 bool was_ld_map; 275 u8 *raw, *todo; 276 __be32 *result; 277 __be64 *bits; 278 279 raw = vmalloc(raw_size); 280 if (!raw) 281 return -ENOMEM; 282 283 sha1_init(digest); 284 memset(ws, 0, sizeof(ws)); 285 286 /* We need to take out the map fd for the digest calculation 287 * since they are unstable from user space side. 288 */ 289 dst = (void *)raw; 290 for (i = 0, was_ld_map = false; i < fp->len; i++) { 291 dst[i] = fp->insnsi[i]; 292 if (!was_ld_map && 293 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 294 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 295 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 296 was_ld_map = true; 297 dst[i].imm = 0; 298 } else if (was_ld_map && 299 dst[i].code == 0 && 300 dst[i].dst_reg == 0 && 301 dst[i].src_reg == 0 && 302 dst[i].off == 0) { 303 was_ld_map = false; 304 dst[i].imm = 0; 305 } else { 306 was_ld_map = false; 307 } 308 } 309 310 psize = bpf_prog_insn_size(fp); 311 memset(&raw[psize], 0, raw_size - psize); 312 raw[psize++] = 0x80; 313 314 bsize = round_up(psize, SHA1_BLOCK_SIZE); 315 blocks = bsize / SHA1_BLOCK_SIZE; 316 todo = raw; 317 if (bsize - psize >= sizeof(__be64)) { 318 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 319 } else { 320 bits = (__be64 *)(todo + bsize + bits_offset); 321 blocks++; 322 } 323 *bits = cpu_to_be64((psize - 1) << 3); 324 325 while (blocks--) { 326 sha1_transform(digest, todo, ws); 327 todo += SHA1_BLOCK_SIZE; 328 } 329 330 result = (__force __be32 *)digest; 331 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 332 result[i] = cpu_to_be32(digest[i]); 333 memcpy(fp->tag, result, sizeof(fp->tag)); 334 335 vfree(raw); 336 return 0; 337 } 338 339 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 340 s32 end_new, s32 curr, const bool probe_pass) 341 { 342 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 343 s32 delta = end_new - end_old; 344 s64 imm = insn->imm; 345 346 if (curr < pos && curr + imm + 1 >= end_old) 347 imm += delta; 348 else if (curr >= end_new && curr + imm + 1 < end_new) 349 imm -= delta; 350 if (imm < imm_min || imm > imm_max) 351 return -ERANGE; 352 if (!probe_pass) 353 insn->imm = imm; 354 return 0; 355 } 356 357 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 358 s32 end_new, s32 curr, const bool probe_pass) 359 { 360 const s32 off_min = S16_MIN, off_max = S16_MAX; 361 s32 delta = end_new - end_old; 362 s32 off = insn->off; 363 364 if (curr < pos && curr + off + 1 >= end_old) 365 off += delta; 366 else if (curr >= end_new && curr + off + 1 < end_new) 367 off -= delta; 368 if (off < off_min || off > off_max) 369 return -ERANGE; 370 if (!probe_pass) 371 insn->off = off; 372 return 0; 373 } 374 375 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 376 s32 end_new, const bool probe_pass) 377 { 378 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 379 struct bpf_insn *insn = prog->insnsi; 380 int ret = 0; 381 382 for (i = 0; i < insn_cnt; i++, insn++) { 383 u8 code; 384 385 /* In the probing pass we still operate on the original, 386 * unpatched image in order to check overflows before we 387 * do any other adjustments. Therefore skip the patchlet. 388 */ 389 if (probe_pass && i == pos) { 390 i = end_new; 391 insn = prog->insnsi + end_old; 392 } 393 if (bpf_pseudo_func(insn)) { 394 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 395 end_new, i, probe_pass); 396 if (ret) 397 return ret; 398 continue; 399 } 400 code = insn->code; 401 if ((BPF_CLASS(code) != BPF_JMP && 402 BPF_CLASS(code) != BPF_JMP32) || 403 BPF_OP(code) == BPF_EXIT) 404 continue; 405 /* Adjust offset of jmps if we cross patch boundaries. */ 406 if (BPF_OP(code) == BPF_CALL) { 407 if (insn->src_reg != BPF_PSEUDO_CALL) 408 continue; 409 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 410 end_new, i, probe_pass); 411 } else { 412 ret = bpf_adj_delta_to_off(insn, pos, end_old, 413 end_new, i, probe_pass); 414 } 415 if (ret) 416 break; 417 } 418 419 return ret; 420 } 421 422 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 423 { 424 struct bpf_line_info *linfo; 425 u32 i, nr_linfo; 426 427 nr_linfo = prog->aux->nr_linfo; 428 if (!nr_linfo || !delta) 429 return; 430 431 linfo = prog->aux->linfo; 432 433 for (i = 0; i < nr_linfo; i++) 434 if (off < linfo[i].insn_off) 435 break; 436 437 /* Push all off < linfo[i].insn_off by delta */ 438 for (; i < nr_linfo; i++) 439 linfo[i].insn_off += delta; 440 } 441 442 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 443 const struct bpf_insn *patch, u32 len) 444 { 445 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 446 const u32 cnt_max = S16_MAX; 447 struct bpf_prog *prog_adj; 448 int err; 449 450 /* Since our patchlet doesn't expand the image, we're done. */ 451 if (insn_delta == 0) { 452 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 453 return prog; 454 } 455 456 insn_adj_cnt = prog->len + insn_delta; 457 458 /* Reject anything that would potentially let the insn->off 459 * target overflow when we have excessive program expansions. 460 * We need to probe here before we do any reallocation where 461 * we afterwards may not fail anymore. 462 */ 463 if (insn_adj_cnt > cnt_max && 464 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 465 return ERR_PTR(err); 466 467 /* Several new instructions need to be inserted. Make room 468 * for them. Likely, there's no need for a new allocation as 469 * last page could have large enough tailroom. 470 */ 471 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 472 GFP_USER); 473 if (!prog_adj) 474 return ERR_PTR(-ENOMEM); 475 476 prog_adj->len = insn_adj_cnt; 477 478 /* Patching happens in 3 steps: 479 * 480 * 1) Move over tail of insnsi from next instruction onwards, 481 * so we can patch the single target insn with one or more 482 * new ones (patching is always from 1 to n insns, n > 0). 483 * 2) Inject new instructions at the target location. 484 * 3) Adjust branch offsets if necessary. 485 */ 486 insn_rest = insn_adj_cnt - off - len; 487 488 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 489 sizeof(*patch) * insn_rest); 490 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 491 492 /* We are guaranteed to not fail at this point, otherwise 493 * the ship has sailed to reverse to the original state. An 494 * overflow cannot happen at this point. 495 */ 496 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 497 498 bpf_adj_linfo(prog_adj, off, insn_delta); 499 500 return prog_adj; 501 } 502 503 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 504 { 505 /* Branch offsets can't overflow when program is shrinking, no need 506 * to call bpf_adj_branches(..., true) here 507 */ 508 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 509 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 510 prog->len -= cnt; 511 512 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 513 } 514 515 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 516 { 517 int i; 518 519 for (i = 0; i < fp->aux->func_cnt; i++) 520 bpf_prog_kallsyms_del(fp->aux->func[i]); 521 } 522 523 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 524 { 525 bpf_prog_kallsyms_del_subprogs(fp); 526 bpf_prog_kallsyms_del(fp); 527 } 528 529 #ifdef CONFIG_BPF_JIT 530 /* All BPF JIT sysctl knobs here. */ 531 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 532 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 533 int bpf_jit_harden __read_mostly; 534 long bpf_jit_limit __read_mostly; 535 long bpf_jit_limit_max __read_mostly; 536 537 static void 538 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 539 { 540 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 541 unsigned long addr = (unsigned long)hdr; 542 543 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 544 545 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 546 prog->aux->ksym.end = addr + hdr->pages * PAGE_SIZE; 547 } 548 549 static void 550 bpf_prog_ksym_set_name(struct bpf_prog *prog) 551 { 552 char *sym = prog->aux->ksym.name; 553 const char *end = sym + KSYM_NAME_LEN; 554 const struct btf_type *type; 555 const char *func_name; 556 557 BUILD_BUG_ON(sizeof("bpf_prog_") + 558 sizeof(prog->tag) * 2 + 559 /* name has been null terminated. 560 * We should need +1 for the '_' preceding 561 * the name. However, the null character 562 * is double counted between the name and the 563 * sizeof("bpf_prog_") above, so we omit 564 * the +1 here. 565 */ 566 sizeof(prog->aux->name) > KSYM_NAME_LEN); 567 568 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 569 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 570 571 /* prog->aux->name will be ignored if full btf name is available */ 572 if (prog->aux->func_info_cnt) { 573 type = btf_type_by_id(prog->aux->btf, 574 prog->aux->func_info[prog->aux->func_idx].type_id); 575 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 576 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 577 return; 578 } 579 580 if (prog->aux->name[0]) 581 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 582 else 583 *sym = 0; 584 } 585 586 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 587 { 588 return container_of(n, struct bpf_ksym, tnode)->start; 589 } 590 591 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 592 struct latch_tree_node *b) 593 { 594 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 595 } 596 597 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 598 { 599 unsigned long val = (unsigned long)key; 600 const struct bpf_ksym *ksym; 601 602 ksym = container_of(n, struct bpf_ksym, tnode); 603 604 if (val < ksym->start) 605 return -1; 606 if (val >= ksym->end) 607 return 1; 608 609 return 0; 610 } 611 612 static const struct latch_tree_ops bpf_tree_ops = { 613 .less = bpf_tree_less, 614 .comp = bpf_tree_comp, 615 }; 616 617 static DEFINE_SPINLOCK(bpf_lock); 618 static LIST_HEAD(bpf_kallsyms); 619 static struct latch_tree_root bpf_tree __cacheline_aligned; 620 621 void bpf_ksym_add(struct bpf_ksym *ksym) 622 { 623 spin_lock_bh(&bpf_lock); 624 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 625 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 626 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 627 spin_unlock_bh(&bpf_lock); 628 } 629 630 static void __bpf_ksym_del(struct bpf_ksym *ksym) 631 { 632 if (list_empty(&ksym->lnode)) 633 return; 634 635 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 636 list_del_rcu(&ksym->lnode); 637 } 638 639 void bpf_ksym_del(struct bpf_ksym *ksym) 640 { 641 spin_lock_bh(&bpf_lock); 642 __bpf_ksym_del(ksym); 643 spin_unlock_bh(&bpf_lock); 644 } 645 646 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 647 { 648 return fp->jited && !bpf_prog_was_classic(fp); 649 } 650 651 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 652 { 653 return list_empty(&fp->aux->ksym.lnode) || 654 fp->aux->ksym.lnode.prev == LIST_POISON2; 655 } 656 657 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 658 { 659 if (!bpf_prog_kallsyms_candidate(fp) || 660 !bpf_capable()) 661 return; 662 663 bpf_prog_ksym_set_addr(fp); 664 bpf_prog_ksym_set_name(fp); 665 fp->aux->ksym.prog = true; 666 667 bpf_ksym_add(&fp->aux->ksym); 668 } 669 670 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 671 { 672 if (!bpf_prog_kallsyms_candidate(fp)) 673 return; 674 675 bpf_ksym_del(&fp->aux->ksym); 676 } 677 678 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 679 { 680 struct latch_tree_node *n; 681 682 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 683 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 684 } 685 686 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 687 unsigned long *off, char *sym) 688 { 689 struct bpf_ksym *ksym; 690 char *ret = NULL; 691 692 rcu_read_lock(); 693 ksym = bpf_ksym_find(addr); 694 if (ksym) { 695 unsigned long symbol_start = ksym->start; 696 unsigned long symbol_end = ksym->end; 697 698 strncpy(sym, ksym->name, KSYM_NAME_LEN); 699 700 ret = sym; 701 if (size) 702 *size = symbol_end - symbol_start; 703 if (off) 704 *off = addr - symbol_start; 705 } 706 rcu_read_unlock(); 707 708 return ret; 709 } 710 711 bool is_bpf_text_address(unsigned long addr) 712 { 713 bool ret; 714 715 rcu_read_lock(); 716 ret = bpf_ksym_find(addr) != NULL; 717 rcu_read_unlock(); 718 719 return ret; 720 } 721 722 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 723 { 724 struct bpf_ksym *ksym = bpf_ksym_find(addr); 725 726 return ksym && ksym->prog ? 727 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 728 NULL; 729 } 730 731 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 732 { 733 const struct exception_table_entry *e = NULL; 734 struct bpf_prog *prog; 735 736 rcu_read_lock(); 737 prog = bpf_prog_ksym_find(addr); 738 if (!prog) 739 goto out; 740 if (!prog->aux->num_exentries) 741 goto out; 742 743 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 744 out: 745 rcu_read_unlock(); 746 return e; 747 } 748 749 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 750 char *sym) 751 { 752 struct bpf_ksym *ksym; 753 unsigned int it = 0; 754 int ret = -ERANGE; 755 756 if (!bpf_jit_kallsyms_enabled()) 757 return ret; 758 759 rcu_read_lock(); 760 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 761 if (it++ != symnum) 762 continue; 763 764 strncpy(sym, ksym->name, KSYM_NAME_LEN); 765 766 *value = ksym->start; 767 *type = BPF_SYM_ELF_TYPE; 768 769 ret = 0; 770 break; 771 } 772 rcu_read_unlock(); 773 774 return ret; 775 } 776 777 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 778 struct bpf_jit_poke_descriptor *poke) 779 { 780 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 781 static const u32 poke_tab_max = 1024; 782 u32 slot = prog->aux->size_poke_tab; 783 u32 size = slot + 1; 784 785 if (size > poke_tab_max) 786 return -ENOSPC; 787 if (poke->tailcall_target || poke->tailcall_target_stable || 788 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 789 return -EINVAL; 790 791 switch (poke->reason) { 792 case BPF_POKE_REASON_TAIL_CALL: 793 if (!poke->tail_call.map) 794 return -EINVAL; 795 break; 796 default: 797 return -EINVAL; 798 } 799 800 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL); 801 if (!tab) 802 return -ENOMEM; 803 804 memcpy(&tab[slot], poke, sizeof(*poke)); 805 prog->aux->size_poke_tab = size; 806 prog->aux->poke_tab = tab; 807 808 return slot; 809 } 810 811 static atomic_long_t bpf_jit_current; 812 813 /* Can be overridden by an arch's JIT compiler if it has a custom, 814 * dedicated BPF backend memory area, or if neither of the two 815 * below apply. 816 */ 817 u64 __weak bpf_jit_alloc_exec_limit(void) 818 { 819 #if defined(MODULES_VADDR) 820 return MODULES_END - MODULES_VADDR; 821 #else 822 return VMALLOC_END - VMALLOC_START; 823 #endif 824 } 825 826 static int __init bpf_jit_charge_init(void) 827 { 828 /* Only used as heuristic here to derive limit. */ 829 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 830 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2, 831 PAGE_SIZE), LONG_MAX); 832 return 0; 833 } 834 pure_initcall(bpf_jit_charge_init); 835 836 int bpf_jit_charge_modmem(u32 pages) 837 { 838 if (atomic_long_add_return(pages, &bpf_jit_current) > 839 (bpf_jit_limit >> PAGE_SHIFT)) { 840 if (!bpf_capable()) { 841 atomic_long_sub(pages, &bpf_jit_current); 842 return -EPERM; 843 } 844 } 845 846 return 0; 847 } 848 849 void bpf_jit_uncharge_modmem(u32 pages) 850 { 851 atomic_long_sub(pages, &bpf_jit_current); 852 } 853 854 void *__weak bpf_jit_alloc_exec(unsigned long size) 855 { 856 return module_alloc(size); 857 } 858 859 void __weak bpf_jit_free_exec(void *addr) 860 { 861 module_memfree(addr); 862 } 863 864 struct bpf_binary_header * 865 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 866 unsigned int alignment, 867 bpf_jit_fill_hole_t bpf_fill_ill_insns) 868 { 869 struct bpf_binary_header *hdr; 870 u32 size, hole, start, pages; 871 872 WARN_ON_ONCE(!is_power_of_2(alignment) || 873 alignment > BPF_IMAGE_ALIGNMENT); 874 875 /* Most of BPF filters are really small, but if some of them 876 * fill a page, allow at least 128 extra bytes to insert a 877 * random section of illegal instructions. 878 */ 879 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 880 pages = size / PAGE_SIZE; 881 882 if (bpf_jit_charge_modmem(pages)) 883 return NULL; 884 hdr = bpf_jit_alloc_exec(size); 885 if (!hdr) { 886 bpf_jit_uncharge_modmem(pages); 887 return NULL; 888 } 889 890 /* Fill space with illegal/arch-dep instructions. */ 891 bpf_fill_ill_insns(hdr, size); 892 893 hdr->pages = pages; 894 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 895 PAGE_SIZE - sizeof(*hdr)); 896 start = (get_random_int() % hole) & ~(alignment - 1); 897 898 /* Leave a random number of instructions before BPF code. */ 899 *image_ptr = &hdr->image[start]; 900 901 return hdr; 902 } 903 904 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 905 { 906 u32 pages = hdr->pages; 907 908 bpf_jit_free_exec(hdr); 909 bpf_jit_uncharge_modmem(pages); 910 } 911 912 /* This symbol is only overridden by archs that have different 913 * requirements than the usual eBPF JITs, f.e. when they only 914 * implement cBPF JIT, do not set images read-only, etc. 915 */ 916 void __weak bpf_jit_free(struct bpf_prog *fp) 917 { 918 if (fp->jited) { 919 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 920 921 bpf_jit_binary_free(hdr); 922 923 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 924 } 925 926 bpf_prog_unlock_free(fp); 927 } 928 929 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 930 const struct bpf_insn *insn, bool extra_pass, 931 u64 *func_addr, bool *func_addr_fixed) 932 { 933 s16 off = insn->off; 934 s32 imm = insn->imm; 935 u8 *addr; 936 937 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 938 if (!*func_addr_fixed) { 939 /* Place-holder address till the last pass has collected 940 * all addresses for JITed subprograms in which case we 941 * can pick them up from prog->aux. 942 */ 943 if (!extra_pass) 944 addr = NULL; 945 else if (prog->aux->func && 946 off >= 0 && off < prog->aux->func_cnt) 947 addr = (u8 *)prog->aux->func[off]->bpf_func; 948 else 949 return -EINVAL; 950 } else { 951 /* Address of a BPF helper call. Since part of the core 952 * kernel, it's always at a fixed location. __bpf_call_base 953 * and the helper with imm relative to it are both in core 954 * kernel. 955 */ 956 addr = (u8 *)__bpf_call_base + imm; 957 } 958 959 *func_addr = (unsigned long)addr; 960 return 0; 961 } 962 963 static int bpf_jit_blind_insn(const struct bpf_insn *from, 964 const struct bpf_insn *aux, 965 struct bpf_insn *to_buff, 966 bool emit_zext) 967 { 968 struct bpf_insn *to = to_buff; 969 u32 imm_rnd = get_random_int(); 970 s16 off; 971 972 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 973 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 974 975 /* Constraints on AX register: 976 * 977 * AX register is inaccessible from user space. It is mapped in 978 * all JITs, and used here for constant blinding rewrites. It is 979 * typically "stateless" meaning its contents are only valid within 980 * the executed instruction, but not across several instructions. 981 * There are a few exceptions however which are further detailed 982 * below. 983 * 984 * Constant blinding is only used by JITs, not in the interpreter. 985 * The interpreter uses AX in some occasions as a local temporary 986 * register e.g. in DIV or MOD instructions. 987 * 988 * In restricted circumstances, the verifier can also use the AX 989 * register for rewrites as long as they do not interfere with 990 * the above cases! 991 */ 992 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 993 goto out; 994 995 if (from->imm == 0 && 996 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 997 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 998 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 999 goto out; 1000 } 1001 1002 switch (from->code) { 1003 case BPF_ALU | BPF_ADD | BPF_K: 1004 case BPF_ALU | BPF_SUB | BPF_K: 1005 case BPF_ALU | BPF_AND | BPF_K: 1006 case BPF_ALU | BPF_OR | BPF_K: 1007 case BPF_ALU | BPF_XOR | BPF_K: 1008 case BPF_ALU | BPF_MUL | BPF_K: 1009 case BPF_ALU | BPF_MOV | BPF_K: 1010 case BPF_ALU | BPF_DIV | BPF_K: 1011 case BPF_ALU | BPF_MOD | BPF_K: 1012 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1013 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1014 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 1015 break; 1016 1017 case BPF_ALU64 | BPF_ADD | BPF_K: 1018 case BPF_ALU64 | BPF_SUB | BPF_K: 1019 case BPF_ALU64 | BPF_AND | BPF_K: 1020 case BPF_ALU64 | BPF_OR | BPF_K: 1021 case BPF_ALU64 | BPF_XOR | BPF_K: 1022 case BPF_ALU64 | BPF_MUL | BPF_K: 1023 case BPF_ALU64 | BPF_MOV | BPF_K: 1024 case BPF_ALU64 | BPF_DIV | BPF_K: 1025 case BPF_ALU64 | BPF_MOD | BPF_K: 1026 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1027 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1028 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 1029 break; 1030 1031 case BPF_JMP | BPF_JEQ | BPF_K: 1032 case BPF_JMP | BPF_JNE | BPF_K: 1033 case BPF_JMP | BPF_JGT | BPF_K: 1034 case BPF_JMP | BPF_JLT | BPF_K: 1035 case BPF_JMP | BPF_JGE | BPF_K: 1036 case BPF_JMP | BPF_JLE | BPF_K: 1037 case BPF_JMP | BPF_JSGT | BPF_K: 1038 case BPF_JMP | BPF_JSLT | BPF_K: 1039 case BPF_JMP | BPF_JSGE | BPF_K: 1040 case BPF_JMP | BPF_JSLE | BPF_K: 1041 case BPF_JMP | BPF_JSET | BPF_K: 1042 /* Accommodate for extra offset in case of a backjump. */ 1043 off = from->off; 1044 if (off < 0) 1045 off -= 2; 1046 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1047 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1048 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1049 break; 1050 1051 case BPF_JMP32 | BPF_JEQ | BPF_K: 1052 case BPF_JMP32 | BPF_JNE | BPF_K: 1053 case BPF_JMP32 | BPF_JGT | BPF_K: 1054 case BPF_JMP32 | BPF_JLT | BPF_K: 1055 case BPF_JMP32 | BPF_JGE | BPF_K: 1056 case BPF_JMP32 | BPF_JLE | BPF_K: 1057 case BPF_JMP32 | BPF_JSGT | BPF_K: 1058 case BPF_JMP32 | BPF_JSLT | BPF_K: 1059 case BPF_JMP32 | BPF_JSGE | BPF_K: 1060 case BPF_JMP32 | BPF_JSLE | BPF_K: 1061 case BPF_JMP32 | BPF_JSET | BPF_K: 1062 /* Accommodate for extra offset in case of a backjump. */ 1063 off = from->off; 1064 if (off < 0) 1065 off -= 2; 1066 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1067 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1068 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1069 off); 1070 break; 1071 1072 case BPF_LD | BPF_IMM | BPF_DW: 1073 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1074 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1075 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1076 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1077 break; 1078 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1079 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1080 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1081 if (emit_zext) 1082 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1083 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1084 break; 1085 1086 case BPF_ST | BPF_MEM | BPF_DW: 1087 case BPF_ST | BPF_MEM | BPF_W: 1088 case BPF_ST | BPF_MEM | BPF_H: 1089 case BPF_ST | BPF_MEM | BPF_B: 1090 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1091 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1092 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1093 break; 1094 } 1095 out: 1096 return to - to_buff; 1097 } 1098 1099 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1100 gfp_t gfp_extra_flags) 1101 { 1102 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1103 struct bpf_prog *fp; 1104 1105 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1106 if (fp != NULL) { 1107 /* aux->prog still points to the fp_other one, so 1108 * when promoting the clone to the real program, 1109 * this still needs to be adapted. 1110 */ 1111 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1112 } 1113 1114 return fp; 1115 } 1116 1117 static void bpf_prog_clone_free(struct bpf_prog *fp) 1118 { 1119 /* aux was stolen by the other clone, so we cannot free 1120 * it from this path! It will be freed eventually by the 1121 * other program on release. 1122 * 1123 * At this point, we don't need a deferred release since 1124 * clone is guaranteed to not be locked. 1125 */ 1126 fp->aux = NULL; 1127 fp->stats = NULL; 1128 fp->active = NULL; 1129 __bpf_prog_free(fp); 1130 } 1131 1132 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1133 { 1134 /* We have to repoint aux->prog to self, as we don't 1135 * know whether fp here is the clone or the original. 1136 */ 1137 fp->aux->prog = fp; 1138 bpf_prog_clone_free(fp_other); 1139 } 1140 1141 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1142 { 1143 struct bpf_insn insn_buff[16], aux[2]; 1144 struct bpf_prog *clone, *tmp; 1145 int insn_delta, insn_cnt; 1146 struct bpf_insn *insn; 1147 int i, rewritten; 1148 1149 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 1150 return prog; 1151 1152 clone = bpf_prog_clone_create(prog, GFP_USER); 1153 if (!clone) 1154 return ERR_PTR(-ENOMEM); 1155 1156 insn_cnt = clone->len; 1157 insn = clone->insnsi; 1158 1159 for (i = 0; i < insn_cnt; i++, insn++) { 1160 /* We temporarily need to hold the original ld64 insn 1161 * so that we can still access the first part in the 1162 * second blinding run. 1163 */ 1164 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1165 insn[1].code == 0) 1166 memcpy(aux, insn, sizeof(aux)); 1167 1168 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1169 clone->aux->verifier_zext); 1170 if (!rewritten) 1171 continue; 1172 1173 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1174 if (IS_ERR(tmp)) { 1175 /* Patching may have repointed aux->prog during 1176 * realloc from the original one, so we need to 1177 * fix it up here on error. 1178 */ 1179 bpf_jit_prog_release_other(prog, clone); 1180 return tmp; 1181 } 1182 1183 clone = tmp; 1184 insn_delta = rewritten - 1; 1185 1186 /* Walk new program and skip insns we just inserted. */ 1187 insn = clone->insnsi + i + insn_delta; 1188 insn_cnt += insn_delta; 1189 i += insn_delta; 1190 } 1191 1192 clone->blinded = 1; 1193 return clone; 1194 } 1195 #endif /* CONFIG_BPF_JIT */ 1196 1197 /* Base function for offset calculation. Needs to go into .text section, 1198 * therefore keeping it non-static as well; will also be used by JITs 1199 * anyway later on, so do not let the compiler omit it. This also needs 1200 * to go into kallsyms for correlation from e.g. bpftool, so naming 1201 * must not change. 1202 */ 1203 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1204 { 1205 return 0; 1206 } 1207 EXPORT_SYMBOL_GPL(__bpf_call_base); 1208 1209 /* All UAPI available opcodes. */ 1210 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1211 /* 32 bit ALU operations. */ \ 1212 /* Register based. */ \ 1213 INSN_3(ALU, ADD, X), \ 1214 INSN_3(ALU, SUB, X), \ 1215 INSN_3(ALU, AND, X), \ 1216 INSN_3(ALU, OR, X), \ 1217 INSN_3(ALU, LSH, X), \ 1218 INSN_3(ALU, RSH, X), \ 1219 INSN_3(ALU, XOR, X), \ 1220 INSN_3(ALU, MUL, X), \ 1221 INSN_3(ALU, MOV, X), \ 1222 INSN_3(ALU, ARSH, X), \ 1223 INSN_3(ALU, DIV, X), \ 1224 INSN_3(ALU, MOD, X), \ 1225 INSN_2(ALU, NEG), \ 1226 INSN_3(ALU, END, TO_BE), \ 1227 INSN_3(ALU, END, TO_LE), \ 1228 /* Immediate based. */ \ 1229 INSN_3(ALU, ADD, K), \ 1230 INSN_3(ALU, SUB, K), \ 1231 INSN_3(ALU, AND, K), \ 1232 INSN_3(ALU, OR, K), \ 1233 INSN_3(ALU, LSH, K), \ 1234 INSN_3(ALU, RSH, K), \ 1235 INSN_3(ALU, XOR, K), \ 1236 INSN_3(ALU, MUL, K), \ 1237 INSN_3(ALU, MOV, K), \ 1238 INSN_3(ALU, ARSH, K), \ 1239 INSN_3(ALU, DIV, K), \ 1240 INSN_3(ALU, MOD, K), \ 1241 /* 64 bit ALU operations. */ \ 1242 /* Register based. */ \ 1243 INSN_3(ALU64, ADD, X), \ 1244 INSN_3(ALU64, SUB, X), \ 1245 INSN_3(ALU64, AND, X), \ 1246 INSN_3(ALU64, OR, X), \ 1247 INSN_3(ALU64, LSH, X), \ 1248 INSN_3(ALU64, RSH, X), \ 1249 INSN_3(ALU64, XOR, X), \ 1250 INSN_3(ALU64, MUL, X), \ 1251 INSN_3(ALU64, MOV, X), \ 1252 INSN_3(ALU64, ARSH, X), \ 1253 INSN_3(ALU64, DIV, X), \ 1254 INSN_3(ALU64, MOD, X), \ 1255 INSN_2(ALU64, NEG), \ 1256 /* Immediate based. */ \ 1257 INSN_3(ALU64, ADD, K), \ 1258 INSN_3(ALU64, SUB, K), \ 1259 INSN_3(ALU64, AND, K), \ 1260 INSN_3(ALU64, OR, K), \ 1261 INSN_3(ALU64, LSH, K), \ 1262 INSN_3(ALU64, RSH, K), \ 1263 INSN_3(ALU64, XOR, K), \ 1264 INSN_3(ALU64, MUL, K), \ 1265 INSN_3(ALU64, MOV, K), \ 1266 INSN_3(ALU64, ARSH, K), \ 1267 INSN_3(ALU64, DIV, K), \ 1268 INSN_3(ALU64, MOD, K), \ 1269 /* Call instruction. */ \ 1270 INSN_2(JMP, CALL), \ 1271 /* Exit instruction. */ \ 1272 INSN_2(JMP, EXIT), \ 1273 /* 32-bit Jump instructions. */ \ 1274 /* Register based. */ \ 1275 INSN_3(JMP32, JEQ, X), \ 1276 INSN_3(JMP32, JNE, X), \ 1277 INSN_3(JMP32, JGT, X), \ 1278 INSN_3(JMP32, JLT, X), \ 1279 INSN_3(JMP32, JGE, X), \ 1280 INSN_3(JMP32, JLE, X), \ 1281 INSN_3(JMP32, JSGT, X), \ 1282 INSN_3(JMP32, JSLT, X), \ 1283 INSN_3(JMP32, JSGE, X), \ 1284 INSN_3(JMP32, JSLE, X), \ 1285 INSN_3(JMP32, JSET, X), \ 1286 /* Immediate based. */ \ 1287 INSN_3(JMP32, JEQ, K), \ 1288 INSN_3(JMP32, JNE, K), \ 1289 INSN_3(JMP32, JGT, K), \ 1290 INSN_3(JMP32, JLT, K), \ 1291 INSN_3(JMP32, JGE, K), \ 1292 INSN_3(JMP32, JLE, K), \ 1293 INSN_3(JMP32, JSGT, K), \ 1294 INSN_3(JMP32, JSLT, K), \ 1295 INSN_3(JMP32, JSGE, K), \ 1296 INSN_3(JMP32, JSLE, K), \ 1297 INSN_3(JMP32, JSET, K), \ 1298 /* Jump instructions. */ \ 1299 /* Register based. */ \ 1300 INSN_3(JMP, JEQ, X), \ 1301 INSN_3(JMP, JNE, X), \ 1302 INSN_3(JMP, JGT, X), \ 1303 INSN_3(JMP, JLT, X), \ 1304 INSN_3(JMP, JGE, X), \ 1305 INSN_3(JMP, JLE, X), \ 1306 INSN_3(JMP, JSGT, X), \ 1307 INSN_3(JMP, JSLT, X), \ 1308 INSN_3(JMP, JSGE, X), \ 1309 INSN_3(JMP, JSLE, X), \ 1310 INSN_3(JMP, JSET, X), \ 1311 /* Immediate based. */ \ 1312 INSN_3(JMP, JEQ, K), \ 1313 INSN_3(JMP, JNE, K), \ 1314 INSN_3(JMP, JGT, K), \ 1315 INSN_3(JMP, JLT, K), \ 1316 INSN_3(JMP, JGE, K), \ 1317 INSN_3(JMP, JLE, K), \ 1318 INSN_3(JMP, JSGT, K), \ 1319 INSN_3(JMP, JSLT, K), \ 1320 INSN_3(JMP, JSGE, K), \ 1321 INSN_3(JMP, JSLE, K), \ 1322 INSN_3(JMP, JSET, K), \ 1323 INSN_2(JMP, JA), \ 1324 /* Store instructions. */ \ 1325 /* Register based. */ \ 1326 INSN_3(STX, MEM, B), \ 1327 INSN_3(STX, MEM, H), \ 1328 INSN_3(STX, MEM, W), \ 1329 INSN_3(STX, MEM, DW), \ 1330 INSN_3(STX, ATOMIC, W), \ 1331 INSN_3(STX, ATOMIC, DW), \ 1332 /* Immediate based. */ \ 1333 INSN_3(ST, MEM, B), \ 1334 INSN_3(ST, MEM, H), \ 1335 INSN_3(ST, MEM, W), \ 1336 INSN_3(ST, MEM, DW), \ 1337 /* Load instructions. */ \ 1338 /* Register based. */ \ 1339 INSN_3(LDX, MEM, B), \ 1340 INSN_3(LDX, MEM, H), \ 1341 INSN_3(LDX, MEM, W), \ 1342 INSN_3(LDX, MEM, DW), \ 1343 /* Immediate based. */ \ 1344 INSN_3(LD, IMM, DW) 1345 1346 bool bpf_opcode_in_insntable(u8 code) 1347 { 1348 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1349 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1350 static const bool public_insntable[256] = { 1351 [0 ... 255] = false, 1352 /* Now overwrite non-defaults ... */ 1353 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1354 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1355 [BPF_LD | BPF_ABS | BPF_B] = true, 1356 [BPF_LD | BPF_ABS | BPF_H] = true, 1357 [BPF_LD | BPF_ABS | BPF_W] = true, 1358 [BPF_LD | BPF_IND | BPF_B] = true, 1359 [BPF_LD | BPF_IND | BPF_H] = true, 1360 [BPF_LD | BPF_IND | BPF_W] = true, 1361 }; 1362 #undef BPF_INSN_3_TBL 1363 #undef BPF_INSN_2_TBL 1364 return public_insntable[code]; 1365 } 1366 1367 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1368 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1369 { 1370 memset(dst, 0, size); 1371 return -EFAULT; 1372 } 1373 1374 /** 1375 * ___bpf_prog_run - run eBPF program on a given context 1376 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1377 * @insn: is the array of eBPF instructions 1378 * 1379 * Decode and execute eBPF instructions. 1380 * 1381 * Return: whatever value is in %BPF_R0 at program exit 1382 */ 1383 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1384 { 1385 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1386 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1387 static const void * const jumptable[256] __annotate_jump_table = { 1388 [0 ... 255] = &&default_label, 1389 /* Now overwrite non-defaults ... */ 1390 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1391 /* Non-UAPI available opcodes. */ 1392 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1393 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1394 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1395 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1396 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1397 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1398 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1399 }; 1400 #undef BPF_INSN_3_LBL 1401 #undef BPF_INSN_2_LBL 1402 u32 tail_call_cnt = 0; 1403 1404 #define CONT ({ insn++; goto select_insn; }) 1405 #define CONT_JMP ({ insn++; goto select_insn; }) 1406 1407 select_insn: 1408 goto *jumptable[insn->code]; 1409 1410 /* Explicitly mask the register-based shift amounts with 63 or 31 1411 * to avoid undefined behavior. Normally this won't affect the 1412 * generated code, for example, in case of native 64 bit archs such 1413 * as x86-64 or arm64, the compiler is optimizing the AND away for 1414 * the interpreter. In case of JITs, each of the JIT backends compiles 1415 * the BPF shift operations to machine instructions which produce 1416 * implementation-defined results in such a case; the resulting 1417 * contents of the register may be arbitrary, but program behaviour 1418 * as a whole remains defined. In other words, in case of JIT backends, 1419 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1420 */ 1421 /* ALU (shifts) */ 1422 #define SHT(OPCODE, OP) \ 1423 ALU64_##OPCODE##_X: \ 1424 DST = DST OP (SRC & 63); \ 1425 CONT; \ 1426 ALU_##OPCODE##_X: \ 1427 DST = (u32) DST OP ((u32) SRC & 31); \ 1428 CONT; \ 1429 ALU64_##OPCODE##_K: \ 1430 DST = DST OP IMM; \ 1431 CONT; \ 1432 ALU_##OPCODE##_K: \ 1433 DST = (u32) DST OP (u32) IMM; \ 1434 CONT; 1435 /* ALU (rest) */ 1436 #define ALU(OPCODE, OP) \ 1437 ALU64_##OPCODE##_X: \ 1438 DST = DST OP SRC; \ 1439 CONT; \ 1440 ALU_##OPCODE##_X: \ 1441 DST = (u32) DST OP (u32) SRC; \ 1442 CONT; \ 1443 ALU64_##OPCODE##_K: \ 1444 DST = DST OP IMM; \ 1445 CONT; \ 1446 ALU_##OPCODE##_K: \ 1447 DST = (u32) DST OP (u32) IMM; \ 1448 CONT; 1449 ALU(ADD, +) 1450 ALU(SUB, -) 1451 ALU(AND, &) 1452 ALU(OR, |) 1453 ALU(XOR, ^) 1454 ALU(MUL, *) 1455 SHT(LSH, <<) 1456 SHT(RSH, >>) 1457 #undef SHT 1458 #undef ALU 1459 ALU_NEG: 1460 DST = (u32) -DST; 1461 CONT; 1462 ALU64_NEG: 1463 DST = -DST; 1464 CONT; 1465 ALU_MOV_X: 1466 DST = (u32) SRC; 1467 CONT; 1468 ALU_MOV_K: 1469 DST = (u32) IMM; 1470 CONT; 1471 ALU64_MOV_X: 1472 DST = SRC; 1473 CONT; 1474 ALU64_MOV_K: 1475 DST = IMM; 1476 CONT; 1477 LD_IMM_DW: 1478 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1479 insn++; 1480 CONT; 1481 ALU_ARSH_X: 1482 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1483 CONT; 1484 ALU_ARSH_K: 1485 DST = (u64) (u32) (((s32) DST) >> IMM); 1486 CONT; 1487 ALU64_ARSH_X: 1488 (*(s64 *) &DST) >>= (SRC & 63); 1489 CONT; 1490 ALU64_ARSH_K: 1491 (*(s64 *) &DST) >>= IMM; 1492 CONT; 1493 ALU64_MOD_X: 1494 div64_u64_rem(DST, SRC, &AX); 1495 DST = AX; 1496 CONT; 1497 ALU_MOD_X: 1498 AX = (u32) DST; 1499 DST = do_div(AX, (u32) SRC); 1500 CONT; 1501 ALU64_MOD_K: 1502 div64_u64_rem(DST, IMM, &AX); 1503 DST = AX; 1504 CONT; 1505 ALU_MOD_K: 1506 AX = (u32) DST; 1507 DST = do_div(AX, (u32) IMM); 1508 CONT; 1509 ALU64_DIV_X: 1510 DST = div64_u64(DST, SRC); 1511 CONT; 1512 ALU_DIV_X: 1513 AX = (u32) DST; 1514 do_div(AX, (u32) SRC); 1515 DST = (u32) AX; 1516 CONT; 1517 ALU64_DIV_K: 1518 DST = div64_u64(DST, IMM); 1519 CONT; 1520 ALU_DIV_K: 1521 AX = (u32) DST; 1522 do_div(AX, (u32) IMM); 1523 DST = (u32) AX; 1524 CONT; 1525 ALU_END_TO_BE: 1526 switch (IMM) { 1527 case 16: 1528 DST = (__force u16) cpu_to_be16(DST); 1529 break; 1530 case 32: 1531 DST = (__force u32) cpu_to_be32(DST); 1532 break; 1533 case 64: 1534 DST = (__force u64) cpu_to_be64(DST); 1535 break; 1536 } 1537 CONT; 1538 ALU_END_TO_LE: 1539 switch (IMM) { 1540 case 16: 1541 DST = (__force u16) cpu_to_le16(DST); 1542 break; 1543 case 32: 1544 DST = (__force u32) cpu_to_le32(DST); 1545 break; 1546 case 64: 1547 DST = (__force u64) cpu_to_le64(DST); 1548 break; 1549 } 1550 CONT; 1551 1552 /* CALL */ 1553 JMP_CALL: 1554 /* Function call scratches BPF_R1-BPF_R5 registers, 1555 * preserves BPF_R6-BPF_R9, and stores return value 1556 * into BPF_R0. 1557 */ 1558 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1559 BPF_R4, BPF_R5); 1560 CONT; 1561 1562 JMP_CALL_ARGS: 1563 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1564 BPF_R3, BPF_R4, 1565 BPF_R5, 1566 insn + insn->off + 1); 1567 CONT; 1568 1569 JMP_TAIL_CALL: { 1570 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1571 struct bpf_array *array = container_of(map, struct bpf_array, map); 1572 struct bpf_prog *prog; 1573 u32 index = BPF_R3; 1574 1575 if (unlikely(index >= array->map.max_entries)) 1576 goto out; 1577 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1578 goto out; 1579 1580 tail_call_cnt++; 1581 1582 prog = READ_ONCE(array->ptrs[index]); 1583 if (!prog) 1584 goto out; 1585 1586 /* ARG1 at this point is guaranteed to point to CTX from 1587 * the verifier side due to the fact that the tail call is 1588 * handled like a helper, that is, bpf_tail_call_proto, 1589 * where arg1_type is ARG_PTR_TO_CTX. 1590 */ 1591 insn = prog->insnsi; 1592 goto select_insn; 1593 out: 1594 CONT; 1595 } 1596 JMP_JA: 1597 insn += insn->off; 1598 CONT; 1599 JMP_EXIT: 1600 return BPF_R0; 1601 /* JMP */ 1602 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1603 JMP_##OPCODE##_X: \ 1604 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1605 insn += insn->off; \ 1606 CONT_JMP; \ 1607 } \ 1608 CONT; \ 1609 JMP32_##OPCODE##_X: \ 1610 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1611 insn += insn->off; \ 1612 CONT_JMP; \ 1613 } \ 1614 CONT; \ 1615 JMP_##OPCODE##_K: \ 1616 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1617 insn += insn->off; \ 1618 CONT_JMP; \ 1619 } \ 1620 CONT; \ 1621 JMP32_##OPCODE##_K: \ 1622 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1623 insn += insn->off; \ 1624 CONT_JMP; \ 1625 } \ 1626 CONT; 1627 COND_JMP(u, JEQ, ==) 1628 COND_JMP(u, JNE, !=) 1629 COND_JMP(u, JGT, >) 1630 COND_JMP(u, JLT, <) 1631 COND_JMP(u, JGE, >=) 1632 COND_JMP(u, JLE, <=) 1633 COND_JMP(u, JSET, &) 1634 COND_JMP(s, JSGT, >) 1635 COND_JMP(s, JSLT, <) 1636 COND_JMP(s, JSGE, >=) 1637 COND_JMP(s, JSLE, <=) 1638 #undef COND_JMP 1639 /* ST, STX and LDX*/ 1640 ST_NOSPEC: 1641 /* Speculation barrier for mitigating Speculative Store Bypass. 1642 * In case of arm64, we rely on the firmware mitigation as 1643 * controlled via the ssbd kernel parameter. Whenever the 1644 * mitigation is enabled, it works for all of the kernel code 1645 * with no need to provide any additional instructions here. 1646 * In case of x86, we use 'lfence' insn for mitigation. We 1647 * reuse preexisting logic from Spectre v1 mitigation that 1648 * happens to produce the required code on x86 for v4 as well. 1649 */ 1650 #ifdef CONFIG_X86 1651 barrier_nospec(); 1652 #endif 1653 CONT; 1654 #define LDST(SIZEOP, SIZE) \ 1655 STX_MEM_##SIZEOP: \ 1656 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1657 CONT; \ 1658 ST_MEM_##SIZEOP: \ 1659 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1660 CONT; \ 1661 LDX_MEM_##SIZEOP: \ 1662 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1663 CONT; 1664 1665 LDST(B, u8) 1666 LDST(H, u16) 1667 LDST(W, u32) 1668 LDST(DW, u64) 1669 #undef LDST 1670 #define LDX_PROBE(SIZEOP, SIZE) \ 1671 LDX_PROBE_MEM_##SIZEOP: \ 1672 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \ 1673 CONT; 1674 LDX_PROBE(B, 1) 1675 LDX_PROBE(H, 2) 1676 LDX_PROBE(W, 4) 1677 LDX_PROBE(DW, 8) 1678 #undef LDX_PROBE 1679 1680 #define ATOMIC_ALU_OP(BOP, KOP) \ 1681 case BOP: \ 1682 if (BPF_SIZE(insn->code) == BPF_W) \ 1683 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 1684 (DST + insn->off)); \ 1685 else \ 1686 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 1687 (DST + insn->off)); \ 1688 break; \ 1689 case BOP | BPF_FETCH: \ 1690 if (BPF_SIZE(insn->code) == BPF_W) \ 1691 SRC = (u32) atomic_fetch_##KOP( \ 1692 (u32) SRC, \ 1693 (atomic_t *)(unsigned long) (DST + insn->off)); \ 1694 else \ 1695 SRC = (u64) atomic64_fetch_##KOP( \ 1696 (u64) SRC, \ 1697 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 1698 break; 1699 1700 STX_ATOMIC_DW: 1701 STX_ATOMIC_W: 1702 switch (IMM) { 1703 ATOMIC_ALU_OP(BPF_ADD, add) 1704 ATOMIC_ALU_OP(BPF_AND, and) 1705 ATOMIC_ALU_OP(BPF_OR, or) 1706 ATOMIC_ALU_OP(BPF_XOR, xor) 1707 #undef ATOMIC_ALU_OP 1708 1709 case BPF_XCHG: 1710 if (BPF_SIZE(insn->code) == BPF_W) 1711 SRC = (u32) atomic_xchg( 1712 (atomic_t *)(unsigned long) (DST + insn->off), 1713 (u32) SRC); 1714 else 1715 SRC = (u64) atomic64_xchg( 1716 (atomic64_t *)(unsigned long) (DST + insn->off), 1717 (u64) SRC); 1718 break; 1719 case BPF_CMPXCHG: 1720 if (BPF_SIZE(insn->code) == BPF_W) 1721 BPF_R0 = (u32) atomic_cmpxchg( 1722 (atomic_t *)(unsigned long) (DST + insn->off), 1723 (u32) BPF_R0, (u32) SRC); 1724 else 1725 BPF_R0 = (u64) atomic64_cmpxchg( 1726 (atomic64_t *)(unsigned long) (DST + insn->off), 1727 (u64) BPF_R0, (u64) SRC); 1728 break; 1729 1730 default: 1731 goto default_label; 1732 } 1733 CONT; 1734 1735 default_label: 1736 /* If we ever reach this, we have a bug somewhere. Die hard here 1737 * instead of just returning 0; we could be somewhere in a subprog, 1738 * so execution could continue otherwise which we do /not/ want. 1739 * 1740 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1741 */ 1742 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 1743 insn->code, insn->imm); 1744 BUG_ON(1); 1745 return 0; 1746 } 1747 1748 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1749 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1750 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1751 { \ 1752 u64 stack[stack_size / sizeof(u64)]; \ 1753 u64 regs[MAX_BPF_EXT_REG]; \ 1754 \ 1755 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1756 ARG1 = (u64) (unsigned long) ctx; \ 1757 return ___bpf_prog_run(regs, insn); \ 1758 } 1759 1760 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1761 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1762 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1763 const struct bpf_insn *insn) \ 1764 { \ 1765 u64 stack[stack_size / sizeof(u64)]; \ 1766 u64 regs[MAX_BPF_EXT_REG]; \ 1767 \ 1768 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1769 BPF_R1 = r1; \ 1770 BPF_R2 = r2; \ 1771 BPF_R3 = r3; \ 1772 BPF_R4 = r4; \ 1773 BPF_R5 = r5; \ 1774 return ___bpf_prog_run(regs, insn); \ 1775 } 1776 1777 #define EVAL1(FN, X) FN(X) 1778 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1779 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1780 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1781 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1782 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1783 1784 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1785 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1786 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1787 1788 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1789 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1790 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1791 1792 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1793 1794 static unsigned int (*interpreters[])(const void *ctx, 1795 const struct bpf_insn *insn) = { 1796 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1797 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1798 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1799 }; 1800 #undef PROG_NAME_LIST 1801 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1802 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1803 const struct bpf_insn *insn) = { 1804 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1805 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1806 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1807 }; 1808 #undef PROG_NAME_LIST 1809 1810 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1811 { 1812 stack_depth = max_t(u32, stack_depth, 1); 1813 insn->off = (s16) insn->imm; 1814 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1815 __bpf_call_base_args; 1816 insn->code = BPF_JMP | BPF_CALL_ARGS; 1817 } 1818 1819 #else 1820 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1821 const struct bpf_insn *insn) 1822 { 1823 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1824 * is not working properly, so warn about it! 1825 */ 1826 WARN_ON_ONCE(1); 1827 return 0; 1828 } 1829 #endif 1830 1831 bool bpf_prog_array_compatible(struct bpf_array *array, 1832 const struct bpf_prog *fp) 1833 { 1834 bool ret; 1835 1836 if (fp->kprobe_override) 1837 return false; 1838 1839 spin_lock(&array->aux->owner.lock); 1840 1841 if (!array->aux->owner.type) { 1842 /* There's no owner yet where we could check for 1843 * compatibility. 1844 */ 1845 array->aux->owner.type = fp->type; 1846 array->aux->owner.jited = fp->jited; 1847 ret = true; 1848 } else { 1849 ret = array->aux->owner.type == fp->type && 1850 array->aux->owner.jited == fp->jited; 1851 } 1852 spin_unlock(&array->aux->owner.lock); 1853 return ret; 1854 } 1855 1856 static int bpf_check_tail_call(const struct bpf_prog *fp) 1857 { 1858 struct bpf_prog_aux *aux = fp->aux; 1859 int i, ret = 0; 1860 1861 mutex_lock(&aux->used_maps_mutex); 1862 for (i = 0; i < aux->used_map_cnt; i++) { 1863 struct bpf_map *map = aux->used_maps[i]; 1864 struct bpf_array *array; 1865 1866 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1867 continue; 1868 1869 array = container_of(map, struct bpf_array, map); 1870 if (!bpf_prog_array_compatible(array, fp)) { 1871 ret = -EINVAL; 1872 goto out; 1873 } 1874 } 1875 1876 out: 1877 mutex_unlock(&aux->used_maps_mutex); 1878 return ret; 1879 } 1880 1881 static void bpf_prog_select_func(struct bpf_prog *fp) 1882 { 1883 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1884 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1885 1886 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1887 #else 1888 fp->bpf_func = __bpf_prog_ret0_warn; 1889 #endif 1890 } 1891 1892 /** 1893 * bpf_prog_select_runtime - select exec runtime for BPF program 1894 * @fp: bpf_prog populated with internal BPF program 1895 * @err: pointer to error variable 1896 * 1897 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1898 * The BPF program will be executed via bpf_prog_run() function. 1899 * 1900 * Return: the &fp argument along with &err set to 0 for success or 1901 * a negative errno code on failure 1902 */ 1903 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1904 { 1905 /* In case of BPF to BPF calls, verifier did all the prep 1906 * work with regards to JITing, etc. 1907 */ 1908 bool jit_needed = false; 1909 1910 if (fp->bpf_func) 1911 goto finalize; 1912 1913 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 1914 bpf_prog_has_kfunc_call(fp)) 1915 jit_needed = true; 1916 1917 bpf_prog_select_func(fp); 1918 1919 /* eBPF JITs can rewrite the program in case constant 1920 * blinding is active. However, in case of error during 1921 * blinding, bpf_int_jit_compile() must always return a 1922 * valid program, which in this case would simply not 1923 * be JITed, but falls back to the interpreter. 1924 */ 1925 if (!bpf_prog_is_dev_bound(fp->aux)) { 1926 *err = bpf_prog_alloc_jited_linfo(fp); 1927 if (*err) 1928 return fp; 1929 1930 fp = bpf_int_jit_compile(fp); 1931 bpf_prog_jit_attempt_done(fp); 1932 if (!fp->jited && jit_needed) { 1933 *err = -ENOTSUPP; 1934 return fp; 1935 } 1936 } else { 1937 *err = bpf_prog_offload_compile(fp); 1938 if (*err) 1939 return fp; 1940 } 1941 1942 finalize: 1943 bpf_prog_lock_ro(fp); 1944 1945 /* The tail call compatibility check can only be done at 1946 * this late stage as we need to determine, if we deal 1947 * with JITed or non JITed program concatenations and not 1948 * all eBPF JITs might immediately support all features. 1949 */ 1950 *err = bpf_check_tail_call(fp); 1951 1952 return fp; 1953 } 1954 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1955 1956 static unsigned int __bpf_prog_ret1(const void *ctx, 1957 const struct bpf_insn *insn) 1958 { 1959 return 1; 1960 } 1961 1962 static struct bpf_prog_dummy { 1963 struct bpf_prog prog; 1964 } dummy_bpf_prog = { 1965 .prog = { 1966 .bpf_func = __bpf_prog_ret1, 1967 }, 1968 }; 1969 1970 /* to avoid allocating empty bpf_prog_array for cgroups that 1971 * don't have bpf program attached use one global 'empty_prog_array' 1972 * It will not be modified the caller of bpf_prog_array_alloc() 1973 * (since caller requested prog_cnt == 0) 1974 * that pointer should be 'freed' by bpf_prog_array_free() 1975 */ 1976 static struct { 1977 struct bpf_prog_array hdr; 1978 struct bpf_prog *null_prog; 1979 } empty_prog_array = { 1980 .null_prog = NULL, 1981 }; 1982 1983 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1984 { 1985 if (prog_cnt) 1986 return kzalloc(sizeof(struct bpf_prog_array) + 1987 sizeof(struct bpf_prog_array_item) * 1988 (prog_cnt + 1), 1989 flags); 1990 1991 return &empty_prog_array.hdr; 1992 } 1993 1994 void bpf_prog_array_free(struct bpf_prog_array *progs) 1995 { 1996 if (!progs || progs == &empty_prog_array.hdr) 1997 return; 1998 kfree_rcu(progs, rcu); 1999 } 2000 2001 int bpf_prog_array_length(struct bpf_prog_array *array) 2002 { 2003 struct bpf_prog_array_item *item; 2004 u32 cnt = 0; 2005 2006 for (item = array->items; item->prog; item++) 2007 if (item->prog != &dummy_bpf_prog.prog) 2008 cnt++; 2009 return cnt; 2010 } 2011 2012 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2013 { 2014 struct bpf_prog_array_item *item; 2015 2016 for (item = array->items; item->prog; item++) 2017 if (item->prog != &dummy_bpf_prog.prog) 2018 return false; 2019 return true; 2020 } 2021 2022 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2023 u32 *prog_ids, 2024 u32 request_cnt) 2025 { 2026 struct bpf_prog_array_item *item; 2027 int i = 0; 2028 2029 for (item = array->items; item->prog; item++) { 2030 if (item->prog == &dummy_bpf_prog.prog) 2031 continue; 2032 prog_ids[i] = item->prog->aux->id; 2033 if (++i == request_cnt) { 2034 item++; 2035 break; 2036 } 2037 } 2038 2039 return !!(item->prog); 2040 } 2041 2042 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2043 __u32 __user *prog_ids, u32 cnt) 2044 { 2045 unsigned long err = 0; 2046 bool nospc; 2047 u32 *ids; 2048 2049 /* users of this function are doing: 2050 * cnt = bpf_prog_array_length(); 2051 * if (cnt > 0) 2052 * bpf_prog_array_copy_to_user(..., cnt); 2053 * so below kcalloc doesn't need extra cnt > 0 check. 2054 */ 2055 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2056 if (!ids) 2057 return -ENOMEM; 2058 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2059 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2060 kfree(ids); 2061 if (err) 2062 return -EFAULT; 2063 if (nospc) 2064 return -ENOSPC; 2065 return 0; 2066 } 2067 2068 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2069 struct bpf_prog *old_prog) 2070 { 2071 struct bpf_prog_array_item *item; 2072 2073 for (item = array->items; item->prog; item++) 2074 if (item->prog == old_prog) { 2075 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2076 break; 2077 } 2078 } 2079 2080 /** 2081 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2082 * index into the program array with 2083 * a dummy no-op program. 2084 * @array: a bpf_prog_array 2085 * @index: the index of the program to replace 2086 * 2087 * Skips over dummy programs, by not counting them, when calculating 2088 * the position of the program to replace. 2089 * 2090 * Return: 2091 * * 0 - Success 2092 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2093 * * -ENOENT - Index out of range 2094 */ 2095 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2096 { 2097 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2098 } 2099 2100 /** 2101 * bpf_prog_array_update_at() - Updates the program at the given index 2102 * into the program array. 2103 * @array: a bpf_prog_array 2104 * @index: the index of the program to update 2105 * @prog: the program to insert into the array 2106 * 2107 * Skips over dummy programs, by not counting them, when calculating 2108 * the position of the program to update. 2109 * 2110 * Return: 2111 * * 0 - Success 2112 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2113 * * -ENOENT - Index out of range 2114 */ 2115 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2116 struct bpf_prog *prog) 2117 { 2118 struct bpf_prog_array_item *item; 2119 2120 if (unlikely(index < 0)) 2121 return -EINVAL; 2122 2123 for (item = array->items; item->prog; item++) { 2124 if (item->prog == &dummy_bpf_prog.prog) 2125 continue; 2126 if (!index) { 2127 WRITE_ONCE(item->prog, prog); 2128 return 0; 2129 } 2130 index--; 2131 } 2132 return -ENOENT; 2133 } 2134 2135 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2136 struct bpf_prog *exclude_prog, 2137 struct bpf_prog *include_prog, 2138 u64 bpf_cookie, 2139 struct bpf_prog_array **new_array) 2140 { 2141 int new_prog_cnt, carry_prog_cnt = 0; 2142 struct bpf_prog_array_item *existing, *new; 2143 struct bpf_prog_array *array; 2144 bool found_exclude = false; 2145 2146 /* Figure out how many existing progs we need to carry over to 2147 * the new array. 2148 */ 2149 if (old_array) { 2150 existing = old_array->items; 2151 for (; existing->prog; existing++) { 2152 if (existing->prog == exclude_prog) { 2153 found_exclude = true; 2154 continue; 2155 } 2156 if (existing->prog != &dummy_bpf_prog.prog) 2157 carry_prog_cnt++; 2158 if (existing->prog == include_prog) 2159 return -EEXIST; 2160 } 2161 } 2162 2163 if (exclude_prog && !found_exclude) 2164 return -ENOENT; 2165 2166 /* How many progs (not NULL) will be in the new array? */ 2167 new_prog_cnt = carry_prog_cnt; 2168 if (include_prog) 2169 new_prog_cnt += 1; 2170 2171 /* Do we have any prog (not NULL) in the new array? */ 2172 if (!new_prog_cnt) { 2173 *new_array = NULL; 2174 return 0; 2175 } 2176 2177 /* +1 as the end of prog_array is marked with NULL */ 2178 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2179 if (!array) 2180 return -ENOMEM; 2181 new = array->items; 2182 2183 /* Fill in the new prog array */ 2184 if (carry_prog_cnt) { 2185 existing = old_array->items; 2186 for (; existing->prog; existing++) { 2187 if (existing->prog == exclude_prog || 2188 existing->prog == &dummy_bpf_prog.prog) 2189 continue; 2190 2191 new->prog = existing->prog; 2192 new->bpf_cookie = existing->bpf_cookie; 2193 new++; 2194 } 2195 } 2196 if (include_prog) { 2197 new->prog = include_prog; 2198 new->bpf_cookie = bpf_cookie; 2199 new++; 2200 } 2201 new->prog = NULL; 2202 *new_array = array; 2203 return 0; 2204 } 2205 2206 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2207 u32 *prog_ids, u32 request_cnt, 2208 u32 *prog_cnt) 2209 { 2210 u32 cnt = 0; 2211 2212 if (array) 2213 cnt = bpf_prog_array_length(array); 2214 2215 *prog_cnt = cnt; 2216 2217 /* return early if user requested only program count or nothing to copy */ 2218 if (!request_cnt || !cnt) 2219 return 0; 2220 2221 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2222 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2223 : 0; 2224 } 2225 2226 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2227 struct bpf_map **used_maps, u32 len) 2228 { 2229 struct bpf_map *map; 2230 u32 i; 2231 2232 for (i = 0; i < len; i++) { 2233 map = used_maps[i]; 2234 if (map->ops->map_poke_untrack) 2235 map->ops->map_poke_untrack(map, aux); 2236 bpf_map_put(map); 2237 } 2238 } 2239 2240 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2241 { 2242 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2243 kfree(aux->used_maps); 2244 } 2245 2246 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2247 struct btf_mod_pair *used_btfs, u32 len) 2248 { 2249 #ifdef CONFIG_BPF_SYSCALL 2250 struct btf_mod_pair *btf_mod; 2251 u32 i; 2252 2253 for (i = 0; i < len; i++) { 2254 btf_mod = &used_btfs[i]; 2255 if (btf_mod->module) 2256 module_put(btf_mod->module); 2257 btf_put(btf_mod->btf); 2258 } 2259 #endif 2260 } 2261 2262 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2263 { 2264 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2265 kfree(aux->used_btfs); 2266 } 2267 2268 static void bpf_prog_free_deferred(struct work_struct *work) 2269 { 2270 struct bpf_prog_aux *aux; 2271 int i; 2272 2273 aux = container_of(work, struct bpf_prog_aux, work); 2274 #ifdef CONFIG_BPF_SYSCALL 2275 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2276 #endif 2277 bpf_free_used_maps(aux); 2278 bpf_free_used_btfs(aux); 2279 if (bpf_prog_is_dev_bound(aux)) 2280 bpf_prog_offload_destroy(aux->prog); 2281 #ifdef CONFIG_PERF_EVENTS 2282 if (aux->prog->has_callchain_buf) 2283 put_callchain_buffers(); 2284 #endif 2285 if (aux->dst_trampoline) 2286 bpf_trampoline_put(aux->dst_trampoline); 2287 for (i = 0; i < aux->func_cnt; i++) { 2288 /* We can just unlink the subprog poke descriptor table as 2289 * it was originally linked to the main program and is also 2290 * released along with it. 2291 */ 2292 aux->func[i]->aux->poke_tab = NULL; 2293 bpf_jit_free(aux->func[i]); 2294 } 2295 if (aux->func_cnt) { 2296 kfree(aux->func); 2297 bpf_prog_unlock_free(aux->prog); 2298 } else { 2299 bpf_jit_free(aux->prog); 2300 } 2301 } 2302 2303 /* Free internal BPF program */ 2304 void bpf_prog_free(struct bpf_prog *fp) 2305 { 2306 struct bpf_prog_aux *aux = fp->aux; 2307 2308 if (aux->dst_prog) 2309 bpf_prog_put(aux->dst_prog); 2310 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2311 schedule_work(&aux->work); 2312 } 2313 EXPORT_SYMBOL_GPL(bpf_prog_free); 2314 2315 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2316 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2317 2318 void bpf_user_rnd_init_once(void) 2319 { 2320 prandom_init_once(&bpf_user_rnd_state); 2321 } 2322 2323 BPF_CALL_0(bpf_user_rnd_u32) 2324 { 2325 /* Should someone ever have the rather unwise idea to use some 2326 * of the registers passed into this function, then note that 2327 * this function is called from native eBPF and classic-to-eBPF 2328 * transformations. Register assignments from both sides are 2329 * different, f.e. classic always sets fn(ctx, A, X) here. 2330 */ 2331 struct rnd_state *state; 2332 u32 res; 2333 2334 state = &get_cpu_var(bpf_user_rnd_state); 2335 res = prandom_u32_state(state); 2336 put_cpu_var(bpf_user_rnd_state); 2337 2338 return res; 2339 } 2340 2341 BPF_CALL_0(bpf_get_raw_cpu_id) 2342 { 2343 return raw_smp_processor_id(); 2344 } 2345 2346 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2347 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2348 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2349 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2350 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2351 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2352 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2353 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2354 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2355 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2356 2357 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2358 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2359 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2360 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2361 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2362 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2363 2364 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2365 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2366 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2367 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2368 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2369 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2370 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2371 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2372 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2373 2374 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2375 { 2376 return NULL; 2377 } 2378 2379 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2380 { 2381 return NULL; 2382 } 2383 2384 u64 __weak 2385 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2386 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2387 { 2388 return -ENOTSUPP; 2389 } 2390 EXPORT_SYMBOL_GPL(bpf_event_output); 2391 2392 /* Always built-in helper functions. */ 2393 const struct bpf_func_proto bpf_tail_call_proto = { 2394 .func = NULL, 2395 .gpl_only = false, 2396 .ret_type = RET_VOID, 2397 .arg1_type = ARG_PTR_TO_CTX, 2398 .arg2_type = ARG_CONST_MAP_PTR, 2399 .arg3_type = ARG_ANYTHING, 2400 }; 2401 2402 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2403 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2404 * eBPF and implicitly also cBPF can get JITed! 2405 */ 2406 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2407 { 2408 return prog; 2409 } 2410 2411 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2412 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2413 */ 2414 void __weak bpf_jit_compile(struct bpf_prog *prog) 2415 { 2416 } 2417 2418 bool __weak bpf_helper_changes_pkt_data(void *func) 2419 { 2420 return false; 2421 } 2422 2423 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2424 * analysis code and wants explicit zero extension inserted by verifier. 2425 * Otherwise, return FALSE. 2426 * 2427 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2428 * you don't override this. JITs that don't want these extra insns can detect 2429 * them using insn_is_zext. 2430 */ 2431 bool __weak bpf_jit_needs_zext(void) 2432 { 2433 return false; 2434 } 2435 2436 bool __weak bpf_jit_supports_kfunc_call(void) 2437 { 2438 return false; 2439 } 2440 2441 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2442 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2443 */ 2444 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2445 int len) 2446 { 2447 return -EFAULT; 2448 } 2449 2450 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2451 void *addr1, void *addr2) 2452 { 2453 return -ENOTSUPP; 2454 } 2455 2456 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2457 EXPORT_SYMBOL(bpf_stats_enabled_key); 2458 2459 /* All definitions of tracepoints related to BPF. */ 2460 #define CREATE_TRACE_POINTS 2461 #include <linux/bpf_trace.h> 2462 2463 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2464 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2465