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