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/frame.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 | __GFP_ZERO | gfp_extra_flags; 81 struct bpf_prog_aux *aux; 82 struct bpf_prog *fp; 83 84 size = round_up(size, PAGE_SIZE); 85 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 86 if (fp == NULL) 87 return NULL; 88 89 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags); 90 if (aux == NULL) { 91 vfree(fp); 92 return NULL; 93 } 94 95 fp->pages = size / PAGE_SIZE; 96 fp->aux = aux; 97 fp->aux->prog = fp; 98 fp->jit_requested = ebpf_jit_enabled(); 99 100 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode); 101 102 return fp; 103 } 104 105 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 106 { 107 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 108 struct bpf_prog *prog; 109 int cpu; 110 111 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 112 if (!prog) 113 return NULL; 114 115 prog->aux->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 116 if (!prog->aux->stats) { 117 kfree(prog->aux); 118 vfree(prog); 119 return NULL; 120 } 121 122 for_each_possible_cpu(cpu) { 123 struct bpf_prog_stats *pstats; 124 125 pstats = per_cpu_ptr(prog->aux->stats, cpu); 126 u64_stats_init(&pstats->syncp); 127 } 128 return prog; 129 } 130 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 131 132 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 133 { 134 if (!prog->aux->nr_linfo || !prog->jit_requested) 135 return 0; 136 137 prog->aux->jited_linfo = kcalloc(prog->aux->nr_linfo, 138 sizeof(*prog->aux->jited_linfo), 139 GFP_KERNEL | __GFP_NOWARN); 140 if (!prog->aux->jited_linfo) 141 return -ENOMEM; 142 143 return 0; 144 } 145 146 void bpf_prog_free_jited_linfo(struct bpf_prog *prog) 147 { 148 kfree(prog->aux->jited_linfo); 149 prog->aux->jited_linfo = NULL; 150 } 151 152 void bpf_prog_free_unused_jited_linfo(struct bpf_prog *prog) 153 { 154 if (prog->aux->jited_linfo && !prog->aux->jited_linfo[0]) 155 bpf_prog_free_jited_linfo(prog); 156 } 157 158 /* The jit engine is responsible to provide an array 159 * for insn_off to the jited_off mapping (insn_to_jit_off). 160 * 161 * The idx to this array is the insn_off. Hence, the insn_off 162 * here is relative to the prog itself instead of the main prog. 163 * This array has one entry for each xlated bpf insn. 164 * 165 * jited_off is the byte off to the last byte of the jited insn. 166 * 167 * Hence, with 168 * insn_start: 169 * The first bpf insn off of the prog. The insn off 170 * here is relative to the main prog. 171 * e.g. if prog is a subprog, insn_start > 0 172 * linfo_idx: 173 * The prog's idx to prog->aux->linfo and jited_linfo 174 * 175 * jited_linfo[linfo_idx] = prog->bpf_func 176 * 177 * For i > linfo_idx, 178 * 179 * jited_linfo[i] = prog->bpf_func + 180 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 181 */ 182 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 183 const u32 *insn_to_jit_off) 184 { 185 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 186 const struct bpf_line_info *linfo; 187 void **jited_linfo; 188 189 if (!prog->aux->jited_linfo) 190 /* Userspace did not provide linfo */ 191 return; 192 193 linfo_idx = prog->aux->linfo_idx; 194 linfo = &prog->aux->linfo[linfo_idx]; 195 insn_start = linfo[0].insn_off; 196 insn_end = insn_start + prog->len; 197 198 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 199 jited_linfo[0] = prog->bpf_func; 200 201 nr_linfo = prog->aux->nr_linfo - linfo_idx; 202 203 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 204 /* The verifier ensures that linfo[i].insn_off is 205 * strictly increasing 206 */ 207 jited_linfo[i] = prog->bpf_func + 208 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 209 } 210 211 void bpf_prog_free_linfo(struct bpf_prog *prog) 212 { 213 bpf_prog_free_jited_linfo(prog); 214 kvfree(prog->aux->linfo); 215 } 216 217 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 218 gfp_t gfp_extra_flags) 219 { 220 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 221 struct bpf_prog *fp; 222 u32 pages, delta; 223 int ret; 224 225 BUG_ON(fp_old == NULL); 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 delta = pages - fp_old->pages; 233 ret = __bpf_prog_charge(fp_old->aux->user, delta); 234 if (ret) 235 return NULL; 236 237 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 238 if (fp == NULL) { 239 __bpf_prog_uncharge(fp_old->aux->user, delta); 240 } else { 241 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 242 fp->pages = pages; 243 fp->aux->prog = fp; 244 245 /* We keep fp->aux from fp_old around in the new 246 * reallocated structure. 247 */ 248 fp_old->aux = NULL; 249 __bpf_prog_free(fp_old); 250 } 251 252 return fp; 253 } 254 255 void __bpf_prog_free(struct bpf_prog *fp) 256 { 257 if (fp->aux) { 258 free_percpu(fp->aux->stats); 259 kfree(fp->aux->poke_tab); 260 kfree(fp->aux); 261 } 262 vfree(fp); 263 } 264 265 int bpf_prog_calc_tag(struct bpf_prog *fp) 266 { 267 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64); 268 u32 raw_size = bpf_prog_tag_scratch_size(fp); 269 u32 digest[SHA_DIGEST_WORDS]; 270 u32 ws[SHA_WORKSPACE_WORDS]; 271 u32 i, bsize, psize, blocks; 272 struct bpf_insn *dst; 273 bool was_ld_map; 274 u8 *raw, *todo; 275 __be32 *result; 276 __be64 *bits; 277 278 raw = vmalloc(raw_size); 279 if (!raw) 280 return -ENOMEM; 281 282 sha_init(digest); 283 memset(ws, 0, sizeof(ws)); 284 285 /* We need to take out the map fd for the digest calculation 286 * since they are unstable from user space side. 287 */ 288 dst = (void *)raw; 289 for (i = 0, was_ld_map = false; i < fp->len; i++) { 290 dst[i] = fp->insnsi[i]; 291 if (!was_ld_map && 292 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 293 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 294 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 295 was_ld_map = true; 296 dst[i].imm = 0; 297 } else if (was_ld_map && 298 dst[i].code == 0 && 299 dst[i].dst_reg == 0 && 300 dst[i].src_reg == 0 && 301 dst[i].off == 0) { 302 was_ld_map = false; 303 dst[i].imm = 0; 304 } else { 305 was_ld_map = false; 306 } 307 } 308 309 psize = bpf_prog_insn_size(fp); 310 memset(&raw[psize], 0, raw_size - psize); 311 raw[psize++] = 0x80; 312 313 bsize = round_up(psize, SHA_MESSAGE_BYTES); 314 blocks = bsize / SHA_MESSAGE_BYTES; 315 todo = raw; 316 if (bsize - psize >= sizeof(__be64)) { 317 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 318 } else { 319 bits = (__be64 *)(todo + bsize + bits_offset); 320 blocks++; 321 } 322 *bits = cpu_to_be64((psize - 1) << 3); 323 324 while (blocks--) { 325 sha_transform(digest, todo, ws); 326 todo += SHA_MESSAGE_BYTES; 327 } 328 329 result = (__force __be32 *)digest; 330 for (i = 0; i < SHA_DIGEST_WORDS; i++) 331 result[i] = cpu_to_be32(digest[i]); 332 memcpy(fp->tag, result, sizeof(fp->tag)); 333 334 vfree(raw); 335 return 0; 336 } 337 338 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 339 s32 end_new, s32 curr, const bool probe_pass) 340 { 341 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 342 s32 delta = end_new - end_old; 343 s64 imm = insn->imm; 344 345 if (curr < pos && curr + imm + 1 >= end_old) 346 imm += delta; 347 else if (curr >= end_new && curr + imm + 1 < end_new) 348 imm -= delta; 349 if (imm < imm_min || imm > imm_max) 350 return -ERANGE; 351 if (!probe_pass) 352 insn->imm = imm; 353 return 0; 354 } 355 356 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 357 s32 end_new, s32 curr, const bool probe_pass) 358 { 359 const s32 off_min = S16_MIN, off_max = S16_MAX; 360 s32 delta = end_new - end_old; 361 s32 off = insn->off; 362 363 if (curr < pos && curr + off + 1 >= end_old) 364 off += delta; 365 else if (curr >= end_new && curr + off + 1 < end_new) 366 off -= delta; 367 if (off < off_min || off > off_max) 368 return -ERANGE; 369 if (!probe_pass) 370 insn->off = off; 371 return 0; 372 } 373 374 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 375 s32 end_new, const bool probe_pass) 376 { 377 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 378 struct bpf_insn *insn = prog->insnsi; 379 int ret = 0; 380 381 for (i = 0; i < insn_cnt; i++, insn++) { 382 u8 code; 383 384 /* In the probing pass we still operate on the original, 385 * unpatched image in order to check overflows before we 386 * do any other adjustments. Therefore skip the patchlet. 387 */ 388 if (probe_pass && i == pos) { 389 i = end_new; 390 insn = prog->insnsi + end_old; 391 } 392 code = insn->code; 393 if ((BPF_CLASS(code) != BPF_JMP && 394 BPF_CLASS(code) != BPF_JMP32) || 395 BPF_OP(code) == BPF_EXIT) 396 continue; 397 /* Adjust offset of jmps if we cross patch boundaries. */ 398 if (BPF_OP(code) == BPF_CALL) { 399 if (insn->src_reg != BPF_PSEUDO_CALL) 400 continue; 401 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 402 end_new, i, probe_pass); 403 } else { 404 ret = bpf_adj_delta_to_off(insn, pos, end_old, 405 end_new, i, probe_pass); 406 } 407 if (ret) 408 break; 409 } 410 411 return ret; 412 } 413 414 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 415 { 416 struct bpf_line_info *linfo; 417 u32 i, nr_linfo; 418 419 nr_linfo = prog->aux->nr_linfo; 420 if (!nr_linfo || !delta) 421 return; 422 423 linfo = prog->aux->linfo; 424 425 for (i = 0; i < nr_linfo; i++) 426 if (off < linfo[i].insn_off) 427 break; 428 429 /* Push all off < linfo[i].insn_off by delta */ 430 for (; i < nr_linfo; i++) 431 linfo[i].insn_off += delta; 432 } 433 434 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 435 const struct bpf_insn *patch, u32 len) 436 { 437 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 438 const u32 cnt_max = S16_MAX; 439 struct bpf_prog *prog_adj; 440 int err; 441 442 /* Since our patchlet doesn't expand the image, we're done. */ 443 if (insn_delta == 0) { 444 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 445 return prog; 446 } 447 448 insn_adj_cnt = prog->len + insn_delta; 449 450 /* Reject anything that would potentially let the insn->off 451 * target overflow when we have excessive program expansions. 452 * We need to probe here before we do any reallocation where 453 * we afterwards may not fail anymore. 454 */ 455 if (insn_adj_cnt > cnt_max && 456 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 457 return ERR_PTR(err); 458 459 /* Several new instructions need to be inserted. Make room 460 * for them. Likely, there's no need for a new allocation as 461 * last page could have large enough tailroom. 462 */ 463 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 464 GFP_USER); 465 if (!prog_adj) 466 return ERR_PTR(-ENOMEM); 467 468 prog_adj->len = insn_adj_cnt; 469 470 /* Patching happens in 3 steps: 471 * 472 * 1) Move over tail of insnsi from next instruction onwards, 473 * so we can patch the single target insn with one or more 474 * new ones (patching is always from 1 to n insns, n > 0). 475 * 2) Inject new instructions at the target location. 476 * 3) Adjust branch offsets if necessary. 477 */ 478 insn_rest = insn_adj_cnt - off - len; 479 480 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 481 sizeof(*patch) * insn_rest); 482 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 483 484 /* We are guaranteed to not fail at this point, otherwise 485 * the ship has sailed to reverse to the original state. An 486 * overflow cannot happen at this point. 487 */ 488 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 489 490 bpf_adj_linfo(prog_adj, off, insn_delta); 491 492 return prog_adj; 493 } 494 495 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 496 { 497 /* Branch offsets can't overflow when program is shrinking, no need 498 * to call bpf_adj_branches(..., true) here 499 */ 500 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 501 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 502 prog->len -= cnt; 503 504 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 505 } 506 507 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 508 { 509 int i; 510 511 for (i = 0; i < fp->aux->func_cnt; i++) 512 bpf_prog_kallsyms_del(fp->aux->func[i]); 513 } 514 515 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 516 { 517 bpf_prog_kallsyms_del_subprogs(fp); 518 bpf_prog_kallsyms_del(fp); 519 } 520 521 #ifdef CONFIG_BPF_JIT 522 /* All BPF JIT sysctl knobs here. */ 523 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_ALWAYS_ON); 524 int bpf_jit_harden __read_mostly; 525 int bpf_jit_kallsyms __read_mostly; 526 long bpf_jit_limit __read_mostly; 527 528 static __always_inline void 529 bpf_get_prog_addr_region(const struct bpf_prog *prog, 530 unsigned long *symbol_start, 531 unsigned long *symbol_end) 532 { 533 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 534 unsigned long addr = (unsigned long)hdr; 535 536 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 537 538 *symbol_start = addr; 539 *symbol_end = addr + hdr->pages * PAGE_SIZE; 540 } 541 542 void bpf_get_prog_name(const struct bpf_prog *prog, char *sym) 543 { 544 const char *end = sym + KSYM_NAME_LEN; 545 const struct btf_type *type; 546 const char *func_name; 547 548 BUILD_BUG_ON(sizeof("bpf_prog_") + 549 sizeof(prog->tag) * 2 + 550 /* name has been null terminated. 551 * We should need +1 for the '_' preceding 552 * the name. However, the null character 553 * is double counted between the name and the 554 * sizeof("bpf_prog_") above, so we omit 555 * the +1 here. 556 */ 557 sizeof(prog->aux->name) > KSYM_NAME_LEN); 558 559 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 560 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 561 562 /* prog->aux->name will be ignored if full btf name is available */ 563 if (prog->aux->func_info_cnt) { 564 type = btf_type_by_id(prog->aux->btf, 565 prog->aux->func_info[prog->aux->func_idx].type_id); 566 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 567 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 568 return; 569 } 570 571 if (prog->aux->name[0]) 572 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 573 else 574 *sym = 0; 575 } 576 577 static __always_inline unsigned long 578 bpf_get_prog_addr_start(struct latch_tree_node *n) 579 { 580 unsigned long symbol_start, symbol_end; 581 const struct bpf_prog_aux *aux; 582 583 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 584 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 585 586 return symbol_start; 587 } 588 589 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 590 struct latch_tree_node *b) 591 { 592 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b); 593 } 594 595 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 596 { 597 unsigned long val = (unsigned long)key; 598 unsigned long symbol_start, symbol_end; 599 const struct bpf_prog_aux *aux; 600 601 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 602 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 603 604 if (val < symbol_start) 605 return -1; 606 if (val >= symbol_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 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux) 622 { 623 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode)); 624 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms); 625 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 626 } 627 628 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux) 629 { 630 if (list_empty(&aux->ksym_lnode)) 631 return; 632 633 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 634 list_del_rcu(&aux->ksym_lnode); 635 } 636 637 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 638 { 639 return fp->jited && !bpf_prog_was_classic(fp); 640 } 641 642 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 643 { 644 return list_empty(&fp->aux->ksym_lnode) || 645 fp->aux->ksym_lnode.prev == LIST_POISON2; 646 } 647 648 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 649 { 650 if (!bpf_prog_kallsyms_candidate(fp) || 651 !capable(CAP_SYS_ADMIN)) 652 return; 653 654 spin_lock_bh(&bpf_lock); 655 bpf_prog_ksym_node_add(fp->aux); 656 spin_unlock_bh(&bpf_lock); 657 } 658 659 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 660 { 661 if (!bpf_prog_kallsyms_candidate(fp)) 662 return; 663 664 spin_lock_bh(&bpf_lock); 665 bpf_prog_ksym_node_del(fp->aux); 666 spin_unlock_bh(&bpf_lock); 667 } 668 669 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr) 670 { 671 struct latch_tree_node *n; 672 673 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 674 return n ? 675 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 676 NULL; 677 } 678 679 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 680 unsigned long *off, char *sym) 681 { 682 unsigned long symbol_start, symbol_end; 683 struct bpf_prog *prog; 684 char *ret = NULL; 685 686 rcu_read_lock(); 687 prog = bpf_prog_kallsyms_find(addr); 688 if (prog) { 689 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 690 bpf_get_prog_name(prog, sym); 691 692 ret = sym; 693 if (size) 694 *size = symbol_end - symbol_start; 695 if (off) 696 *off = addr - symbol_start; 697 } 698 rcu_read_unlock(); 699 700 return ret; 701 } 702 703 bool is_bpf_text_address(unsigned long addr) 704 { 705 bool ret; 706 707 rcu_read_lock(); 708 ret = bpf_prog_kallsyms_find(addr) != NULL; 709 rcu_read_unlock(); 710 711 return ret; 712 } 713 714 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 715 { 716 const struct exception_table_entry *e = NULL; 717 struct bpf_prog *prog; 718 719 rcu_read_lock(); 720 prog = bpf_prog_kallsyms_find(addr); 721 if (!prog) 722 goto out; 723 if (!prog->aux->num_exentries) 724 goto out; 725 726 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 727 out: 728 rcu_read_unlock(); 729 return e; 730 } 731 732 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 733 char *sym) 734 { 735 struct bpf_prog_aux *aux; 736 unsigned int it = 0; 737 int ret = -ERANGE; 738 739 if (!bpf_jit_kallsyms_enabled()) 740 return ret; 741 742 rcu_read_lock(); 743 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 744 if (it++ != symnum) 745 continue; 746 747 bpf_get_prog_name(aux->prog, sym); 748 749 *value = (unsigned long)aux->prog->bpf_func; 750 *type = BPF_SYM_ELF_TYPE; 751 752 ret = 0; 753 break; 754 } 755 rcu_read_unlock(); 756 757 return ret; 758 } 759 760 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 761 struct bpf_jit_poke_descriptor *poke) 762 { 763 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 764 static const u32 poke_tab_max = 1024; 765 u32 slot = prog->aux->size_poke_tab; 766 u32 size = slot + 1; 767 768 if (size > poke_tab_max) 769 return -ENOSPC; 770 if (poke->ip || poke->ip_stable || poke->adj_off) 771 return -EINVAL; 772 773 switch (poke->reason) { 774 case BPF_POKE_REASON_TAIL_CALL: 775 if (!poke->tail_call.map) 776 return -EINVAL; 777 break; 778 default: 779 return -EINVAL; 780 } 781 782 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL); 783 if (!tab) 784 return -ENOMEM; 785 786 memcpy(&tab[slot], poke, sizeof(*poke)); 787 prog->aux->size_poke_tab = size; 788 prog->aux->poke_tab = tab; 789 790 return slot; 791 } 792 793 static atomic_long_t bpf_jit_current; 794 795 /* Can be overridden by an arch's JIT compiler if it has a custom, 796 * dedicated BPF backend memory area, or if neither of the two 797 * below apply. 798 */ 799 u64 __weak bpf_jit_alloc_exec_limit(void) 800 { 801 #if defined(MODULES_VADDR) 802 return MODULES_END - MODULES_VADDR; 803 #else 804 return VMALLOC_END - VMALLOC_START; 805 #endif 806 } 807 808 static int __init bpf_jit_charge_init(void) 809 { 810 /* Only used as heuristic here to derive limit. */ 811 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2, 812 PAGE_SIZE), LONG_MAX); 813 return 0; 814 } 815 pure_initcall(bpf_jit_charge_init); 816 817 static int bpf_jit_charge_modmem(u32 pages) 818 { 819 if (atomic_long_add_return(pages, &bpf_jit_current) > 820 (bpf_jit_limit >> PAGE_SHIFT)) { 821 if (!capable(CAP_SYS_ADMIN)) { 822 atomic_long_sub(pages, &bpf_jit_current); 823 return -EPERM; 824 } 825 } 826 827 return 0; 828 } 829 830 static void bpf_jit_uncharge_modmem(u32 pages) 831 { 832 atomic_long_sub(pages, &bpf_jit_current); 833 } 834 835 void *__weak bpf_jit_alloc_exec(unsigned long size) 836 { 837 return module_alloc(size); 838 } 839 840 void __weak bpf_jit_free_exec(void *addr) 841 { 842 module_memfree(addr); 843 } 844 845 struct bpf_binary_header * 846 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 847 unsigned int alignment, 848 bpf_jit_fill_hole_t bpf_fill_ill_insns) 849 { 850 struct bpf_binary_header *hdr; 851 u32 size, hole, start, pages; 852 853 WARN_ON_ONCE(!is_power_of_2(alignment) || 854 alignment > BPF_IMAGE_ALIGNMENT); 855 856 /* Most of BPF filters are really small, but if some of them 857 * fill a page, allow at least 128 extra bytes to insert a 858 * random section of illegal instructions. 859 */ 860 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 861 pages = size / PAGE_SIZE; 862 863 if (bpf_jit_charge_modmem(pages)) 864 return NULL; 865 hdr = bpf_jit_alloc_exec(size); 866 if (!hdr) { 867 bpf_jit_uncharge_modmem(pages); 868 return NULL; 869 } 870 871 /* Fill space with illegal/arch-dep instructions. */ 872 bpf_fill_ill_insns(hdr, size); 873 874 hdr->pages = pages; 875 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 876 PAGE_SIZE - sizeof(*hdr)); 877 start = (get_random_int() % hole) & ~(alignment - 1); 878 879 /* Leave a random number of instructions before BPF code. */ 880 *image_ptr = &hdr->image[start]; 881 882 return hdr; 883 } 884 885 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 886 { 887 u32 pages = hdr->pages; 888 889 bpf_jit_free_exec(hdr); 890 bpf_jit_uncharge_modmem(pages); 891 } 892 893 /* This symbol is only overridden by archs that have different 894 * requirements than the usual eBPF JITs, f.e. when they only 895 * implement cBPF JIT, do not set images read-only, etc. 896 */ 897 void __weak bpf_jit_free(struct bpf_prog *fp) 898 { 899 if (fp->jited) { 900 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 901 902 bpf_jit_binary_free(hdr); 903 904 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 905 } 906 907 bpf_prog_unlock_free(fp); 908 } 909 910 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 911 const struct bpf_insn *insn, bool extra_pass, 912 u64 *func_addr, bool *func_addr_fixed) 913 { 914 s16 off = insn->off; 915 s32 imm = insn->imm; 916 u8 *addr; 917 918 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 919 if (!*func_addr_fixed) { 920 /* Place-holder address till the last pass has collected 921 * all addresses for JITed subprograms in which case we 922 * can pick them up from prog->aux. 923 */ 924 if (!extra_pass) 925 addr = NULL; 926 else if (prog->aux->func && 927 off >= 0 && off < prog->aux->func_cnt) 928 addr = (u8 *)prog->aux->func[off]->bpf_func; 929 else 930 return -EINVAL; 931 } else { 932 /* Address of a BPF helper call. Since part of the core 933 * kernel, it's always at a fixed location. __bpf_call_base 934 * and the helper with imm relative to it are both in core 935 * kernel. 936 */ 937 addr = (u8 *)__bpf_call_base + imm; 938 } 939 940 *func_addr = (unsigned long)addr; 941 return 0; 942 } 943 944 static int bpf_jit_blind_insn(const struct bpf_insn *from, 945 const struct bpf_insn *aux, 946 struct bpf_insn *to_buff, 947 bool emit_zext) 948 { 949 struct bpf_insn *to = to_buff; 950 u32 imm_rnd = get_random_int(); 951 s16 off; 952 953 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 954 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 955 956 /* Constraints on AX register: 957 * 958 * AX register is inaccessible from user space. It is mapped in 959 * all JITs, and used here for constant blinding rewrites. It is 960 * typically "stateless" meaning its contents are only valid within 961 * the executed instruction, but not across several instructions. 962 * There are a few exceptions however which are further detailed 963 * below. 964 * 965 * Constant blinding is only used by JITs, not in the interpreter. 966 * The interpreter uses AX in some occasions as a local temporary 967 * register e.g. in DIV or MOD instructions. 968 * 969 * In restricted circumstances, the verifier can also use the AX 970 * register for rewrites as long as they do not interfere with 971 * the above cases! 972 */ 973 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 974 goto out; 975 976 if (from->imm == 0 && 977 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 978 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 979 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 980 goto out; 981 } 982 983 switch (from->code) { 984 case BPF_ALU | BPF_ADD | BPF_K: 985 case BPF_ALU | BPF_SUB | BPF_K: 986 case BPF_ALU | BPF_AND | BPF_K: 987 case BPF_ALU | BPF_OR | BPF_K: 988 case BPF_ALU | BPF_XOR | BPF_K: 989 case BPF_ALU | BPF_MUL | BPF_K: 990 case BPF_ALU | BPF_MOV | BPF_K: 991 case BPF_ALU | BPF_DIV | BPF_K: 992 case BPF_ALU | BPF_MOD | BPF_K: 993 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 994 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 995 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 996 break; 997 998 case BPF_ALU64 | BPF_ADD | BPF_K: 999 case BPF_ALU64 | BPF_SUB | BPF_K: 1000 case BPF_ALU64 | BPF_AND | BPF_K: 1001 case BPF_ALU64 | BPF_OR | BPF_K: 1002 case BPF_ALU64 | BPF_XOR | BPF_K: 1003 case BPF_ALU64 | BPF_MUL | BPF_K: 1004 case BPF_ALU64 | BPF_MOV | BPF_K: 1005 case BPF_ALU64 | BPF_DIV | BPF_K: 1006 case BPF_ALU64 | BPF_MOD | BPF_K: 1007 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1008 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1009 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 1010 break; 1011 1012 case BPF_JMP | BPF_JEQ | BPF_K: 1013 case BPF_JMP | BPF_JNE | BPF_K: 1014 case BPF_JMP | BPF_JGT | BPF_K: 1015 case BPF_JMP | BPF_JLT | BPF_K: 1016 case BPF_JMP | BPF_JGE | BPF_K: 1017 case BPF_JMP | BPF_JLE | BPF_K: 1018 case BPF_JMP | BPF_JSGT | BPF_K: 1019 case BPF_JMP | BPF_JSLT | BPF_K: 1020 case BPF_JMP | BPF_JSGE | BPF_K: 1021 case BPF_JMP | BPF_JSLE | BPF_K: 1022 case BPF_JMP | BPF_JSET | BPF_K: 1023 /* Accommodate for extra offset in case of a backjump. */ 1024 off = from->off; 1025 if (off < 0) 1026 off -= 2; 1027 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1028 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1029 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1030 break; 1031 1032 case BPF_JMP32 | BPF_JEQ | BPF_K: 1033 case BPF_JMP32 | BPF_JNE | BPF_K: 1034 case BPF_JMP32 | BPF_JGT | BPF_K: 1035 case BPF_JMP32 | BPF_JLT | BPF_K: 1036 case BPF_JMP32 | BPF_JGE | BPF_K: 1037 case BPF_JMP32 | BPF_JLE | BPF_K: 1038 case BPF_JMP32 | BPF_JSGT | BPF_K: 1039 case BPF_JMP32 | BPF_JSLT | BPF_K: 1040 case BPF_JMP32 | BPF_JSGE | BPF_K: 1041 case BPF_JMP32 | BPF_JSLE | BPF_K: 1042 case BPF_JMP32 | BPF_JSET | BPF_K: 1043 /* Accommodate for extra offset in case of a backjump. */ 1044 off = from->off; 1045 if (off < 0) 1046 off -= 2; 1047 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1048 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1049 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1050 off); 1051 break; 1052 1053 case BPF_LD | BPF_IMM | BPF_DW: 1054 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1055 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1056 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1057 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1058 break; 1059 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1060 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1061 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1062 if (emit_zext) 1063 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1064 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1065 break; 1066 1067 case BPF_ST | BPF_MEM | BPF_DW: 1068 case BPF_ST | BPF_MEM | BPF_W: 1069 case BPF_ST | BPF_MEM | BPF_H: 1070 case BPF_ST | BPF_MEM | BPF_B: 1071 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1072 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1073 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1074 break; 1075 } 1076 out: 1077 return to - to_buff; 1078 } 1079 1080 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1081 gfp_t gfp_extra_flags) 1082 { 1083 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1084 struct bpf_prog *fp; 1085 1086 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 1087 if (fp != NULL) { 1088 /* aux->prog still points to the fp_other one, so 1089 * when promoting the clone to the real program, 1090 * this still needs to be adapted. 1091 */ 1092 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1093 } 1094 1095 return fp; 1096 } 1097 1098 static void bpf_prog_clone_free(struct bpf_prog *fp) 1099 { 1100 /* aux was stolen by the other clone, so we cannot free 1101 * it from this path! It will be freed eventually by the 1102 * other program on release. 1103 * 1104 * At this point, we don't need a deferred release since 1105 * clone is guaranteed to not be locked. 1106 */ 1107 fp->aux = NULL; 1108 __bpf_prog_free(fp); 1109 } 1110 1111 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1112 { 1113 /* We have to repoint aux->prog to self, as we don't 1114 * know whether fp here is the clone or the original. 1115 */ 1116 fp->aux->prog = fp; 1117 bpf_prog_clone_free(fp_other); 1118 } 1119 1120 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1121 { 1122 struct bpf_insn insn_buff[16], aux[2]; 1123 struct bpf_prog *clone, *tmp; 1124 int insn_delta, insn_cnt; 1125 struct bpf_insn *insn; 1126 int i, rewritten; 1127 1128 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 1129 return prog; 1130 1131 clone = bpf_prog_clone_create(prog, GFP_USER); 1132 if (!clone) 1133 return ERR_PTR(-ENOMEM); 1134 1135 insn_cnt = clone->len; 1136 insn = clone->insnsi; 1137 1138 for (i = 0; i < insn_cnt; i++, insn++) { 1139 /* We temporarily need to hold the original ld64 insn 1140 * so that we can still access the first part in the 1141 * second blinding run. 1142 */ 1143 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1144 insn[1].code == 0) 1145 memcpy(aux, insn, sizeof(aux)); 1146 1147 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1148 clone->aux->verifier_zext); 1149 if (!rewritten) 1150 continue; 1151 1152 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1153 if (IS_ERR(tmp)) { 1154 /* Patching may have repointed aux->prog during 1155 * realloc from the original one, so we need to 1156 * fix it up here on error. 1157 */ 1158 bpf_jit_prog_release_other(prog, clone); 1159 return tmp; 1160 } 1161 1162 clone = tmp; 1163 insn_delta = rewritten - 1; 1164 1165 /* Walk new program and skip insns we just inserted. */ 1166 insn = clone->insnsi + i + insn_delta; 1167 insn_cnt += insn_delta; 1168 i += insn_delta; 1169 } 1170 1171 clone->blinded = 1; 1172 return clone; 1173 } 1174 #endif /* CONFIG_BPF_JIT */ 1175 1176 /* Base function for offset calculation. Needs to go into .text section, 1177 * therefore keeping it non-static as well; will also be used by JITs 1178 * anyway later on, so do not let the compiler omit it. This also needs 1179 * to go into kallsyms for correlation from e.g. bpftool, so naming 1180 * must not change. 1181 */ 1182 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1183 { 1184 return 0; 1185 } 1186 EXPORT_SYMBOL_GPL(__bpf_call_base); 1187 1188 /* All UAPI available opcodes. */ 1189 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1190 /* 32 bit ALU operations. */ \ 1191 /* Register based. */ \ 1192 INSN_3(ALU, ADD, X), \ 1193 INSN_3(ALU, SUB, X), \ 1194 INSN_3(ALU, AND, X), \ 1195 INSN_3(ALU, OR, X), \ 1196 INSN_3(ALU, LSH, X), \ 1197 INSN_3(ALU, RSH, X), \ 1198 INSN_3(ALU, XOR, X), \ 1199 INSN_3(ALU, MUL, X), \ 1200 INSN_3(ALU, MOV, X), \ 1201 INSN_3(ALU, ARSH, X), \ 1202 INSN_3(ALU, DIV, X), \ 1203 INSN_3(ALU, MOD, X), \ 1204 INSN_2(ALU, NEG), \ 1205 INSN_3(ALU, END, TO_BE), \ 1206 INSN_3(ALU, END, TO_LE), \ 1207 /* Immediate based. */ \ 1208 INSN_3(ALU, ADD, K), \ 1209 INSN_3(ALU, SUB, K), \ 1210 INSN_3(ALU, AND, K), \ 1211 INSN_3(ALU, OR, K), \ 1212 INSN_3(ALU, LSH, K), \ 1213 INSN_3(ALU, RSH, K), \ 1214 INSN_3(ALU, XOR, K), \ 1215 INSN_3(ALU, MUL, K), \ 1216 INSN_3(ALU, MOV, K), \ 1217 INSN_3(ALU, ARSH, K), \ 1218 INSN_3(ALU, DIV, K), \ 1219 INSN_3(ALU, MOD, K), \ 1220 /* 64 bit ALU operations. */ \ 1221 /* Register based. */ \ 1222 INSN_3(ALU64, ADD, X), \ 1223 INSN_3(ALU64, SUB, X), \ 1224 INSN_3(ALU64, AND, X), \ 1225 INSN_3(ALU64, OR, X), \ 1226 INSN_3(ALU64, LSH, X), \ 1227 INSN_3(ALU64, RSH, X), \ 1228 INSN_3(ALU64, XOR, X), \ 1229 INSN_3(ALU64, MUL, X), \ 1230 INSN_3(ALU64, MOV, X), \ 1231 INSN_3(ALU64, ARSH, X), \ 1232 INSN_3(ALU64, DIV, X), \ 1233 INSN_3(ALU64, MOD, X), \ 1234 INSN_2(ALU64, NEG), \ 1235 /* Immediate based. */ \ 1236 INSN_3(ALU64, ADD, K), \ 1237 INSN_3(ALU64, SUB, K), \ 1238 INSN_3(ALU64, AND, K), \ 1239 INSN_3(ALU64, OR, K), \ 1240 INSN_3(ALU64, LSH, K), \ 1241 INSN_3(ALU64, RSH, K), \ 1242 INSN_3(ALU64, XOR, K), \ 1243 INSN_3(ALU64, MUL, K), \ 1244 INSN_3(ALU64, MOV, K), \ 1245 INSN_3(ALU64, ARSH, K), \ 1246 INSN_3(ALU64, DIV, K), \ 1247 INSN_3(ALU64, MOD, K), \ 1248 /* Call instruction. */ \ 1249 INSN_2(JMP, CALL), \ 1250 /* Exit instruction. */ \ 1251 INSN_2(JMP, EXIT), \ 1252 /* 32-bit Jump instructions. */ \ 1253 /* Register based. */ \ 1254 INSN_3(JMP32, JEQ, X), \ 1255 INSN_3(JMP32, JNE, X), \ 1256 INSN_3(JMP32, JGT, X), \ 1257 INSN_3(JMP32, JLT, X), \ 1258 INSN_3(JMP32, JGE, X), \ 1259 INSN_3(JMP32, JLE, X), \ 1260 INSN_3(JMP32, JSGT, X), \ 1261 INSN_3(JMP32, JSLT, X), \ 1262 INSN_3(JMP32, JSGE, X), \ 1263 INSN_3(JMP32, JSLE, X), \ 1264 INSN_3(JMP32, JSET, X), \ 1265 /* Immediate based. */ \ 1266 INSN_3(JMP32, JEQ, K), \ 1267 INSN_3(JMP32, JNE, K), \ 1268 INSN_3(JMP32, JGT, K), \ 1269 INSN_3(JMP32, JLT, K), \ 1270 INSN_3(JMP32, JGE, K), \ 1271 INSN_3(JMP32, JLE, K), \ 1272 INSN_3(JMP32, JSGT, K), \ 1273 INSN_3(JMP32, JSLT, K), \ 1274 INSN_3(JMP32, JSGE, K), \ 1275 INSN_3(JMP32, JSLE, K), \ 1276 INSN_3(JMP32, JSET, K), \ 1277 /* Jump instructions. */ \ 1278 /* Register based. */ \ 1279 INSN_3(JMP, JEQ, X), \ 1280 INSN_3(JMP, JNE, X), \ 1281 INSN_3(JMP, JGT, X), \ 1282 INSN_3(JMP, JLT, X), \ 1283 INSN_3(JMP, JGE, X), \ 1284 INSN_3(JMP, JLE, X), \ 1285 INSN_3(JMP, JSGT, X), \ 1286 INSN_3(JMP, JSLT, X), \ 1287 INSN_3(JMP, JSGE, X), \ 1288 INSN_3(JMP, JSLE, X), \ 1289 INSN_3(JMP, JSET, X), \ 1290 /* Immediate based. */ \ 1291 INSN_3(JMP, JEQ, K), \ 1292 INSN_3(JMP, JNE, K), \ 1293 INSN_3(JMP, JGT, K), \ 1294 INSN_3(JMP, JLT, K), \ 1295 INSN_3(JMP, JGE, K), \ 1296 INSN_3(JMP, JLE, K), \ 1297 INSN_3(JMP, JSGT, K), \ 1298 INSN_3(JMP, JSLT, K), \ 1299 INSN_3(JMP, JSGE, K), \ 1300 INSN_3(JMP, JSLE, K), \ 1301 INSN_3(JMP, JSET, K), \ 1302 INSN_2(JMP, JA), \ 1303 /* Store instructions. */ \ 1304 /* Register based. */ \ 1305 INSN_3(STX, MEM, B), \ 1306 INSN_3(STX, MEM, H), \ 1307 INSN_3(STX, MEM, W), \ 1308 INSN_3(STX, MEM, DW), \ 1309 INSN_3(STX, XADD, W), \ 1310 INSN_3(STX, XADD, DW), \ 1311 /* Immediate based. */ \ 1312 INSN_3(ST, MEM, B), \ 1313 INSN_3(ST, MEM, H), \ 1314 INSN_3(ST, MEM, W), \ 1315 INSN_3(ST, MEM, DW), \ 1316 /* Load instructions. */ \ 1317 /* Register based. */ \ 1318 INSN_3(LDX, MEM, B), \ 1319 INSN_3(LDX, MEM, H), \ 1320 INSN_3(LDX, MEM, W), \ 1321 INSN_3(LDX, MEM, DW), \ 1322 /* Immediate based. */ \ 1323 INSN_3(LD, IMM, DW) 1324 1325 bool bpf_opcode_in_insntable(u8 code) 1326 { 1327 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1328 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1329 static const bool public_insntable[256] = { 1330 [0 ... 255] = false, 1331 /* Now overwrite non-defaults ... */ 1332 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1333 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1334 [BPF_LD | BPF_ABS | BPF_B] = true, 1335 [BPF_LD | BPF_ABS | BPF_H] = true, 1336 [BPF_LD | BPF_ABS | BPF_W] = true, 1337 [BPF_LD | BPF_IND | BPF_B] = true, 1338 [BPF_LD | BPF_IND | BPF_H] = true, 1339 [BPF_LD | BPF_IND | BPF_W] = true, 1340 }; 1341 #undef BPF_INSN_3_TBL 1342 #undef BPF_INSN_2_TBL 1343 return public_insntable[code]; 1344 } 1345 1346 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1347 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1348 { 1349 memset(dst, 0, size); 1350 return -EFAULT; 1351 } 1352 1353 /** 1354 * __bpf_prog_run - run eBPF program on a given context 1355 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1356 * @insn: is the array of eBPF instructions 1357 * @stack: is the eBPF storage stack 1358 * 1359 * Decode and execute eBPF instructions. 1360 */ 1361 static u64 __no_fgcse ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 1362 { 1363 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1364 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1365 static const void * const jumptable[256] __annotate_jump_table = { 1366 [0 ... 255] = &&default_label, 1367 /* Now overwrite non-defaults ... */ 1368 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1369 /* Non-UAPI available opcodes. */ 1370 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1371 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1372 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1373 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1374 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1375 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1376 }; 1377 #undef BPF_INSN_3_LBL 1378 #undef BPF_INSN_2_LBL 1379 u32 tail_call_cnt = 0; 1380 1381 #define CONT ({ insn++; goto select_insn; }) 1382 #define CONT_JMP ({ insn++; goto select_insn; }) 1383 1384 select_insn: 1385 goto *jumptable[insn->code]; 1386 1387 /* ALU */ 1388 #define ALU(OPCODE, OP) \ 1389 ALU64_##OPCODE##_X: \ 1390 DST = DST OP SRC; \ 1391 CONT; \ 1392 ALU_##OPCODE##_X: \ 1393 DST = (u32) DST OP (u32) SRC; \ 1394 CONT; \ 1395 ALU64_##OPCODE##_K: \ 1396 DST = DST OP IMM; \ 1397 CONT; \ 1398 ALU_##OPCODE##_K: \ 1399 DST = (u32) DST OP (u32) IMM; \ 1400 CONT; 1401 1402 ALU(ADD, +) 1403 ALU(SUB, -) 1404 ALU(AND, &) 1405 ALU(OR, |) 1406 ALU(LSH, <<) 1407 ALU(RSH, >>) 1408 ALU(XOR, ^) 1409 ALU(MUL, *) 1410 #undef ALU 1411 ALU_NEG: 1412 DST = (u32) -DST; 1413 CONT; 1414 ALU64_NEG: 1415 DST = -DST; 1416 CONT; 1417 ALU_MOV_X: 1418 DST = (u32) SRC; 1419 CONT; 1420 ALU_MOV_K: 1421 DST = (u32) IMM; 1422 CONT; 1423 ALU64_MOV_X: 1424 DST = SRC; 1425 CONT; 1426 ALU64_MOV_K: 1427 DST = IMM; 1428 CONT; 1429 LD_IMM_DW: 1430 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1431 insn++; 1432 CONT; 1433 ALU_ARSH_X: 1434 DST = (u64) (u32) (((s32) DST) >> SRC); 1435 CONT; 1436 ALU_ARSH_K: 1437 DST = (u64) (u32) (((s32) DST) >> IMM); 1438 CONT; 1439 ALU64_ARSH_X: 1440 (*(s64 *) &DST) >>= SRC; 1441 CONT; 1442 ALU64_ARSH_K: 1443 (*(s64 *) &DST) >>= IMM; 1444 CONT; 1445 ALU64_MOD_X: 1446 div64_u64_rem(DST, SRC, &AX); 1447 DST = AX; 1448 CONT; 1449 ALU_MOD_X: 1450 AX = (u32) DST; 1451 DST = do_div(AX, (u32) SRC); 1452 CONT; 1453 ALU64_MOD_K: 1454 div64_u64_rem(DST, IMM, &AX); 1455 DST = AX; 1456 CONT; 1457 ALU_MOD_K: 1458 AX = (u32) DST; 1459 DST = do_div(AX, (u32) IMM); 1460 CONT; 1461 ALU64_DIV_X: 1462 DST = div64_u64(DST, SRC); 1463 CONT; 1464 ALU_DIV_X: 1465 AX = (u32) DST; 1466 do_div(AX, (u32) SRC); 1467 DST = (u32) AX; 1468 CONT; 1469 ALU64_DIV_K: 1470 DST = div64_u64(DST, IMM); 1471 CONT; 1472 ALU_DIV_K: 1473 AX = (u32) DST; 1474 do_div(AX, (u32) IMM); 1475 DST = (u32) AX; 1476 CONT; 1477 ALU_END_TO_BE: 1478 switch (IMM) { 1479 case 16: 1480 DST = (__force u16) cpu_to_be16(DST); 1481 break; 1482 case 32: 1483 DST = (__force u32) cpu_to_be32(DST); 1484 break; 1485 case 64: 1486 DST = (__force u64) cpu_to_be64(DST); 1487 break; 1488 } 1489 CONT; 1490 ALU_END_TO_LE: 1491 switch (IMM) { 1492 case 16: 1493 DST = (__force u16) cpu_to_le16(DST); 1494 break; 1495 case 32: 1496 DST = (__force u32) cpu_to_le32(DST); 1497 break; 1498 case 64: 1499 DST = (__force u64) cpu_to_le64(DST); 1500 break; 1501 } 1502 CONT; 1503 1504 /* CALL */ 1505 JMP_CALL: 1506 /* Function call scratches BPF_R1-BPF_R5 registers, 1507 * preserves BPF_R6-BPF_R9, and stores return value 1508 * into BPF_R0. 1509 */ 1510 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1511 BPF_R4, BPF_R5); 1512 CONT; 1513 1514 JMP_CALL_ARGS: 1515 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1516 BPF_R3, BPF_R4, 1517 BPF_R5, 1518 insn + insn->off + 1); 1519 CONT; 1520 1521 JMP_TAIL_CALL: { 1522 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1523 struct bpf_array *array = container_of(map, struct bpf_array, map); 1524 struct bpf_prog *prog; 1525 u32 index = BPF_R3; 1526 1527 if (unlikely(index >= array->map.max_entries)) 1528 goto out; 1529 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1530 goto out; 1531 1532 tail_call_cnt++; 1533 1534 prog = READ_ONCE(array->ptrs[index]); 1535 if (!prog) 1536 goto out; 1537 1538 /* ARG1 at this point is guaranteed to point to CTX from 1539 * the verifier side due to the fact that the tail call is 1540 * handeled like a helper, that is, bpf_tail_call_proto, 1541 * where arg1_type is ARG_PTR_TO_CTX. 1542 */ 1543 insn = prog->insnsi; 1544 goto select_insn; 1545 out: 1546 CONT; 1547 } 1548 JMP_JA: 1549 insn += insn->off; 1550 CONT; 1551 JMP_EXIT: 1552 return BPF_R0; 1553 /* JMP */ 1554 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1555 JMP_##OPCODE##_X: \ 1556 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1557 insn += insn->off; \ 1558 CONT_JMP; \ 1559 } \ 1560 CONT; \ 1561 JMP32_##OPCODE##_X: \ 1562 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1563 insn += insn->off; \ 1564 CONT_JMP; \ 1565 } \ 1566 CONT; \ 1567 JMP_##OPCODE##_K: \ 1568 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1569 insn += insn->off; \ 1570 CONT_JMP; \ 1571 } \ 1572 CONT; \ 1573 JMP32_##OPCODE##_K: \ 1574 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1575 insn += insn->off; \ 1576 CONT_JMP; \ 1577 } \ 1578 CONT; 1579 COND_JMP(u, JEQ, ==) 1580 COND_JMP(u, JNE, !=) 1581 COND_JMP(u, JGT, >) 1582 COND_JMP(u, JLT, <) 1583 COND_JMP(u, JGE, >=) 1584 COND_JMP(u, JLE, <=) 1585 COND_JMP(u, JSET, &) 1586 COND_JMP(s, JSGT, >) 1587 COND_JMP(s, JSLT, <) 1588 COND_JMP(s, JSGE, >=) 1589 COND_JMP(s, JSLE, <=) 1590 #undef COND_JMP 1591 /* STX and ST and LDX*/ 1592 #define LDST(SIZEOP, SIZE) \ 1593 STX_MEM_##SIZEOP: \ 1594 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1595 CONT; \ 1596 ST_MEM_##SIZEOP: \ 1597 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1598 CONT; \ 1599 LDX_MEM_##SIZEOP: \ 1600 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1601 CONT; 1602 1603 LDST(B, u8) 1604 LDST(H, u16) 1605 LDST(W, u32) 1606 LDST(DW, u64) 1607 #undef LDST 1608 #define LDX_PROBE(SIZEOP, SIZE) \ 1609 LDX_PROBE_MEM_##SIZEOP: \ 1610 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \ 1611 CONT; 1612 LDX_PROBE(B, 1) 1613 LDX_PROBE(H, 2) 1614 LDX_PROBE(W, 4) 1615 LDX_PROBE(DW, 8) 1616 #undef LDX_PROBE 1617 1618 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1619 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1620 (DST + insn->off)); 1621 CONT; 1622 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1623 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1624 (DST + insn->off)); 1625 CONT; 1626 1627 default_label: 1628 /* If we ever reach this, we have a bug somewhere. Die hard here 1629 * instead of just returning 0; we could be somewhere in a subprog, 1630 * so execution could continue otherwise which we do /not/ want. 1631 * 1632 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1633 */ 1634 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code); 1635 BUG_ON(1); 1636 return 0; 1637 } 1638 1639 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1640 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1641 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1642 { \ 1643 u64 stack[stack_size / sizeof(u64)]; \ 1644 u64 regs[MAX_BPF_EXT_REG]; \ 1645 \ 1646 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1647 ARG1 = (u64) (unsigned long) ctx; \ 1648 return ___bpf_prog_run(regs, insn, stack); \ 1649 } 1650 1651 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1652 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1653 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1654 const struct bpf_insn *insn) \ 1655 { \ 1656 u64 stack[stack_size / sizeof(u64)]; \ 1657 u64 regs[MAX_BPF_EXT_REG]; \ 1658 \ 1659 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1660 BPF_R1 = r1; \ 1661 BPF_R2 = r2; \ 1662 BPF_R3 = r3; \ 1663 BPF_R4 = r4; \ 1664 BPF_R5 = r5; \ 1665 return ___bpf_prog_run(regs, insn, stack); \ 1666 } 1667 1668 #define EVAL1(FN, X) FN(X) 1669 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1670 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1671 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1672 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1673 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1674 1675 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1676 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1677 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1678 1679 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1680 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1681 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1682 1683 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1684 1685 static unsigned int (*interpreters[])(const void *ctx, 1686 const struct bpf_insn *insn) = { 1687 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1688 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1689 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1690 }; 1691 #undef PROG_NAME_LIST 1692 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1693 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1694 const struct bpf_insn *insn) = { 1695 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1696 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1697 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1698 }; 1699 #undef PROG_NAME_LIST 1700 1701 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1702 { 1703 stack_depth = max_t(u32, stack_depth, 1); 1704 insn->off = (s16) insn->imm; 1705 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1706 __bpf_call_base_args; 1707 insn->code = BPF_JMP | BPF_CALL_ARGS; 1708 } 1709 1710 #else 1711 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1712 const struct bpf_insn *insn) 1713 { 1714 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1715 * is not working properly, so warn about it! 1716 */ 1717 WARN_ON_ONCE(1); 1718 return 0; 1719 } 1720 #endif 1721 1722 bool bpf_prog_array_compatible(struct bpf_array *array, 1723 const struct bpf_prog *fp) 1724 { 1725 if (fp->kprobe_override) 1726 return false; 1727 1728 if (!array->aux->type) { 1729 /* There's no owner yet where we could check for 1730 * compatibility. 1731 */ 1732 array->aux->type = fp->type; 1733 array->aux->jited = fp->jited; 1734 return true; 1735 } 1736 1737 return array->aux->type == fp->type && 1738 array->aux->jited == fp->jited; 1739 } 1740 1741 static int bpf_check_tail_call(const struct bpf_prog *fp) 1742 { 1743 struct bpf_prog_aux *aux = fp->aux; 1744 int i; 1745 1746 for (i = 0; i < aux->used_map_cnt; i++) { 1747 struct bpf_map *map = aux->used_maps[i]; 1748 struct bpf_array *array; 1749 1750 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1751 continue; 1752 1753 array = container_of(map, struct bpf_array, map); 1754 if (!bpf_prog_array_compatible(array, fp)) 1755 return -EINVAL; 1756 } 1757 1758 return 0; 1759 } 1760 1761 static void bpf_prog_select_func(struct bpf_prog *fp) 1762 { 1763 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1764 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1765 1766 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1767 #else 1768 fp->bpf_func = __bpf_prog_ret0_warn; 1769 #endif 1770 } 1771 1772 /** 1773 * bpf_prog_select_runtime - select exec runtime for BPF program 1774 * @fp: bpf_prog populated with internal BPF program 1775 * @err: pointer to error variable 1776 * 1777 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1778 * The BPF program will be executed via BPF_PROG_RUN() macro. 1779 */ 1780 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1781 { 1782 /* In case of BPF to BPF calls, verifier did all the prep 1783 * work with regards to JITing, etc. 1784 */ 1785 if (fp->bpf_func) 1786 goto finalize; 1787 1788 bpf_prog_select_func(fp); 1789 1790 /* eBPF JITs can rewrite the program in case constant 1791 * blinding is active. However, in case of error during 1792 * blinding, bpf_int_jit_compile() must always return a 1793 * valid program, which in this case would simply not 1794 * be JITed, but falls back to the interpreter. 1795 */ 1796 if (!bpf_prog_is_dev_bound(fp->aux)) { 1797 *err = bpf_prog_alloc_jited_linfo(fp); 1798 if (*err) 1799 return fp; 1800 1801 fp = bpf_int_jit_compile(fp); 1802 if (!fp->jited) { 1803 bpf_prog_free_jited_linfo(fp); 1804 #ifdef CONFIG_BPF_JIT_ALWAYS_ON 1805 *err = -ENOTSUPP; 1806 return fp; 1807 #endif 1808 } else { 1809 bpf_prog_free_unused_jited_linfo(fp); 1810 } 1811 } else { 1812 *err = bpf_prog_offload_compile(fp); 1813 if (*err) 1814 return fp; 1815 } 1816 1817 finalize: 1818 bpf_prog_lock_ro(fp); 1819 1820 /* The tail call compatibility check can only be done at 1821 * this late stage as we need to determine, if we deal 1822 * with JITed or non JITed program concatenations and not 1823 * all eBPF JITs might immediately support all features. 1824 */ 1825 *err = bpf_check_tail_call(fp); 1826 1827 return fp; 1828 } 1829 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1830 1831 static unsigned int __bpf_prog_ret1(const void *ctx, 1832 const struct bpf_insn *insn) 1833 { 1834 return 1; 1835 } 1836 1837 static struct bpf_prog_dummy { 1838 struct bpf_prog prog; 1839 } dummy_bpf_prog = { 1840 .prog = { 1841 .bpf_func = __bpf_prog_ret1, 1842 }, 1843 }; 1844 1845 /* to avoid allocating empty bpf_prog_array for cgroups that 1846 * don't have bpf program attached use one global 'empty_prog_array' 1847 * It will not be modified the caller of bpf_prog_array_alloc() 1848 * (since caller requested prog_cnt == 0) 1849 * that pointer should be 'freed' by bpf_prog_array_free() 1850 */ 1851 static struct { 1852 struct bpf_prog_array hdr; 1853 struct bpf_prog *null_prog; 1854 } empty_prog_array = { 1855 .null_prog = NULL, 1856 }; 1857 1858 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1859 { 1860 if (prog_cnt) 1861 return kzalloc(sizeof(struct bpf_prog_array) + 1862 sizeof(struct bpf_prog_array_item) * 1863 (prog_cnt + 1), 1864 flags); 1865 1866 return &empty_prog_array.hdr; 1867 } 1868 1869 void bpf_prog_array_free(struct bpf_prog_array *progs) 1870 { 1871 if (!progs || progs == &empty_prog_array.hdr) 1872 return; 1873 kfree_rcu(progs, rcu); 1874 } 1875 1876 int bpf_prog_array_length(struct bpf_prog_array *array) 1877 { 1878 struct bpf_prog_array_item *item; 1879 u32 cnt = 0; 1880 1881 for (item = array->items; item->prog; item++) 1882 if (item->prog != &dummy_bpf_prog.prog) 1883 cnt++; 1884 return cnt; 1885 } 1886 1887 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 1888 { 1889 struct bpf_prog_array_item *item; 1890 1891 for (item = array->items; item->prog; item++) 1892 if (item->prog != &dummy_bpf_prog.prog) 1893 return false; 1894 return true; 1895 } 1896 1897 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 1898 u32 *prog_ids, 1899 u32 request_cnt) 1900 { 1901 struct bpf_prog_array_item *item; 1902 int i = 0; 1903 1904 for (item = array->items; item->prog; item++) { 1905 if (item->prog == &dummy_bpf_prog.prog) 1906 continue; 1907 prog_ids[i] = item->prog->aux->id; 1908 if (++i == request_cnt) { 1909 item++; 1910 break; 1911 } 1912 } 1913 1914 return !!(item->prog); 1915 } 1916 1917 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 1918 __u32 __user *prog_ids, u32 cnt) 1919 { 1920 unsigned long err = 0; 1921 bool nospc; 1922 u32 *ids; 1923 1924 /* users of this function are doing: 1925 * cnt = bpf_prog_array_length(); 1926 * if (cnt > 0) 1927 * bpf_prog_array_copy_to_user(..., cnt); 1928 * so below kcalloc doesn't need extra cnt > 0 check. 1929 */ 1930 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 1931 if (!ids) 1932 return -ENOMEM; 1933 nospc = bpf_prog_array_copy_core(array, ids, cnt); 1934 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 1935 kfree(ids); 1936 if (err) 1937 return -EFAULT; 1938 if (nospc) 1939 return -ENOSPC; 1940 return 0; 1941 } 1942 1943 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 1944 struct bpf_prog *old_prog) 1945 { 1946 struct bpf_prog_array_item *item; 1947 1948 for (item = array->items; item->prog; item++) 1949 if (item->prog == old_prog) { 1950 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 1951 break; 1952 } 1953 } 1954 1955 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 1956 struct bpf_prog *exclude_prog, 1957 struct bpf_prog *include_prog, 1958 struct bpf_prog_array **new_array) 1959 { 1960 int new_prog_cnt, carry_prog_cnt = 0; 1961 struct bpf_prog_array_item *existing; 1962 struct bpf_prog_array *array; 1963 bool found_exclude = false; 1964 int new_prog_idx = 0; 1965 1966 /* Figure out how many existing progs we need to carry over to 1967 * the new array. 1968 */ 1969 if (old_array) { 1970 existing = old_array->items; 1971 for (; existing->prog; existing++) { 1972 if (existing->prog == exclude_prog) { 1973 found_exclude = true; 1974 continue; 1975 } 1976 if (existing->prog != &dummy_bpf_prog.prog) 1977 carry_prog_cnt++; 1978 if (existing->prog == include_prog) 1979 return -EEXIST; 1980 } 1981 } 1982 1983 if (exclude_prog && !found_exclude) 1984 return -ENOENT; 1985 1986 /* How many progs (not NULL) will be in the new array? */ 1987 new_prog_cnt = carry_prog_cnt; 1988 if (include_prog) 1989 new_prog_cnt += 1; 1990 1991 /* Do we have any prog (not NULL) in the new array? */ 1992 if (!new_prog_cnt) { 1993 *new_array = NULL; 1994 return 0; 1995 } 1996 1997 /* +1 as the end of prog_array is marked with NULL */ 1998 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1999 if (!array) 2000 return -ENOMEM; 2001 2002 /* Fill in the new prog array */ 2003 if (carry_prog_cnt) { 2004 existing = old_array->items; 2005 for (; existing->prog; existing++) 2006 if (existing->prog != exclude_prog && 2007 existing->prog != &dummy_bpf_prog.prog) { 2008 array->items[new_prog_idx++].prog = 2009 existing->prog; 2010 } 2011 } 2012 if (include_prog) 2013 array->items[new_prog_idx++].prog = include_prog; 2014 array->items[new_prog_idx].prog = NULL; 2015 *new_array = array; 2016 return 0; 2017 } 2018 2019 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2020 u32 *prog_ids, u32 request_cnt, 2021 u32 *prog_cnt) 2022 { 2023 u32 cnt = 0; 2024 2025 if (array) 2026 cnt = bpf_prog_array_length(array); 2027 2028 *prog_cnt = cnt; 2029 2030 /* return early if user requested only program count or nothing to copy */ 2031 if (!request_cnt || !cnt) 2032 return 0; 2033 2034 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2035 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2036 : 0; 2037 } 2038 2039 static void bpf_free_cgroup_storage(struct bpf_prog_aux *aux) 2040 { 2041 enum bpf_cgroup_storage_type stype; 2042 2043 for_each_cgroup_storage_type(stype) { 2044 if (!aux->cgroup_storage[stype]) 2045 continue; 2046 bpf_cgroup_storage_release(aux, aux->cgroup_storage[stype]); 2047 } 2048 } 2049 2050 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2051 struct bpf_map **used_maps, u32 len) 2052 { 2053 struct bpf_map *map; 2054 u32 i; 2055 2056 bpf_free_cgroup_storage(aux); 2057 for (i = 0; i < len; i++) { 2058 map = used_maps[i]; 2059 if (map->ops->map_poke_untrack) 2060 map->ops->map_poke_untrack(map, aux); 2061 bpf_map_put(map); 2062 } 2063 } 2064 2065 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2066 { 2067 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2068 kfree(aux->used_maps); 2069 } 2070 2071 static void bpf_prog_free_deferred(struct work_struct *work) 2072 { 2073 struct bpf_prog_aux *aux; 2074 int i; 2075 2076 aux = container_of(work, struct bpf_prog_aux, work); 2077 bpf_free_used_maps(aux); 2078 if (bpf_prog_is_dev_bound(aux)) 2079 bpf_prog_offload_destroy(aux->prog); 2080 #ifdef CONFIG_PERF_EVENTS 2081 if (aux->prog->has_callchain_buf) 2082 put_callchain_buffers(); 2083 #endif 2084 bpf_trampoline_put(aux->trampoline); 2085 for (i = 0; i < aux->func_cnt; i++) 2086 bpf_jit_free(aux->func[i]); 2087 if (aux->func_cnt) { 2088 kfree(aux->func); 2089 bpf_prog_unlock_free(aux->prog); 2090 } else { 2091 bpf_jit_free(aux->prog); 2092 } 2093 } 2094 2095 /* Free internal BPF program */ 2096 void bpf_prog_free(struct bpf_prog *fp) 2097 { 2098 struct bpf_prog_aux *aux = fp->aux; 2099 2100 if (aux->linked_prog) 2101 bpf_prog_put(aux->linked_prog); 2102 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2103 schedule_work(&aux->work); 2104 } 2105 EXPORT_SYMBOL_GPL(bpf_prog_free); 2106 2107 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2108 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2109 2110 void bpf_user_rnd_init_once(void) 2111 { 2112 prandom_init_once(&bpf_user_rnd_state); 2113 } 2114 2115 BPF_CALL_0(bpf_user_rnd_u32) 2116 { 2117 /* Should someone ever have the rather unwise idea to use some 2118 * of the registers passed into this function, then note that 2119 * this function is called from native eBPF and classic-to-eBPF 2120 * transformations. Register assignments from both sides are 2121 * different, f.e. classic always sets fn(ctx, A, X) here. 2122 */ 2123 struct rnd_state *state; 2124 u32 res; 2125 2126 state = &get_cpu_var(bpf_user_rnd_state); 2127 res = prandom_u32_state(state); 2128 put_cpu_var(bpf_user_rnd_state); 2129 2130 return res; 2131 } 2132 2133 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2134 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2135 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2136 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2137 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2138 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2139 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2140 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2141 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2142 2143 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2144 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2145 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2146 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2147 2148 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2149 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2150 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2151 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2152 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2153 2154 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2155 { 2156 return NULL; 2157 } 2158 2159 u64 __weak 2160 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2161 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2162 { 2163 return -ENOTSUPP; 2164 } 2165 EXPORT_SYMBOL_GPL(bpf_event_output); 2166 2167 /* Always built-in helper functions. */ 2168 const struct bpf_func_proto bpf_tail_call_proto = { 2169 .func = NULL, 2170 .gpl_only = false, 2171 .ret_type = RET_VOID, 2172 .arg1_type = ARG_PTR_TO_CTX, 2173 .arg2_type = ARG_CONST_MAP_PTR, 2174 .arg3_type = ARG_ANYTHING, 2175 }; 2176 2177 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2178 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2179 * eBPF and implicitly also cBPF can get JITed! 2180 */ 2181 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2182 { 2183 return prog; 2184 } 2185 2186 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2187 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2188 */ 2189 void __weak bpf_jit_compile(struct bpf_prog *prog) 2190 { 2191 } 2192 2193 bool __weak bpf_helper_changes_pkt_data(void *func) 2194 { 2195 return false; 2196 } 2197 2198 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2199 * analysis code and wants explicit zero extension inserted by verifier. 2200 * Otherwise, return FALSE. 2201 */ 2202 bool __weak bpf_jit_needs_zext(void) 2203 { 2204 return false; 2205 } 2206 2207 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2208 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2209 */ 2210 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2211 int len) 2212 { 2213 return -EFAULT; 2214 } 2215 2216 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2217 void *addr1, void *addr2) 2218 { 2219 return -ENOTSUPP; 2220 } 2221 2222 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2223 EXPORT_SYMBOL(bpf_stats_enabled_key); 2224 2225 /* All definitions of tracepoints related to BPF. */ 2226 #define CREATE_TRACE_POINTS 2227 #include <linux/bpf_trace.h> 2228 2229 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2230 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2231