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