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 36 #include <asm/barrier.h> 37 #include <asm/unaligned.h> 38 39 /* Registers */ 40 #define BPF_R0 regs[BPF_REG_0] 41 #define BPF_R1 regs[BPF_REG_1] 42 #define BPF_R2 regs[BPF_REG_2] 43 #define BPF_R3 regs[BPF_REG_3] 44 #define BPF_R4 regs[BPF_REG_4] 45 #define BPF_R5 regs[BPF_REG_5] 46 #define BPF_R6 regs[BPF_REG_6] 47 #define BPF_R7 regs[BPF_REG_7] 48 #define BPF_R8 regs[BPF_REG_8] 49 #define BPF_R9 regs[BPF_REG_9] 50 #define BPF_R10 regs[BPF_REG_10] 51 52 /* Named registers */ 53 #define DST regs[insn->dst_reg] 54 #define SRC regs[insn->src_reg] 55 #define FP regs[BPF_REG_FP] 56 #define AX regs[BPF_REG_AX] 57 #define ARG1 regs[BPF_REG_ARG1] 58 #define CTX regs[BPF_REG_CTX] 59 #define IMM insn->imm 60 61 /* No hurry in this branch 62 * 63 * Exported for the bpf jit load helper. 64 */ 65 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 66 { 67 u8 *ptr = NULL; 68 69 if (k >= SKF_NET_OFF) 70 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 71 else if (k >= SKF_LL_OFF) 72 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 73 74 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 75 return ptr; 76 77 return NULL; 78 } 79 80 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 81 { 82 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 83 struct bpf_prog_aux *aux; 84 struct bpf_prog *fp; 85 86 size = round_up(size, PAGE_SIZE); 87 fp = __vmalloc(size, gfp_flags); 88 if (fp == NULL) 89 return NULL; 90 91 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags); 92 if (aux == NULL) { 93 vfree(fp); 94 return NULL; 95 } 96 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags); 97 if (!fp->active) { 98 vfree(fp); 99 kfree(aux); 100 return NULL; 101 } 102 103 fp->pages = size / PAGE_SIZE; 104 fp->aux = aux; 105 fp->aux->prog = fp; 106 fp->jit_requested = ebpf_jit_enabled(); 107 108 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 109 mutex_init(&fp->aux->used_maps_mutex); 110 mutex_init(&fp->aux->dst_mutex); 111 112 return fp; 113 } 114 115 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 116 { 117 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 118 struct bpf_prog *prog; 119 int cpu; 120 121 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 122 if (!prog) 123 return NULL; 124 125 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 126 if (!prog->stats) { 127 free_percpu(prog->active); 128 kfree(prog->aux); 129 vfree(prog); 130 return NULL; 131 } 132 133 for_each_possible_cpu(cpu) { 134 struct bpf_prog_stats *pstats; 135 136 pstats = per_cpu_ptr(prog->stats, cpu); 137 u64_stats_init(&pstats->syncp); 138 } 139 return prog; 140 } 141 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 142 143 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 144 { 145 if (!prog->aux->nr_linfo || !prog->jit_requested) 146 return 0; 147 148 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 149 sizeof(*prog->aux->jited_linfo), 150 GFP_KERNEL_ACCOUNT | __GFP_NOWARN); 151 if (!prog->aux->jited_linfo) 152 return -ENOMEM; 153 154 return 0; 155 } 156 157 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 158 { 159 if (prog->aux->jited_linfo && 160 (!prog->jited || !prog->aux->jited_linfo[0])) { 161 kvfree(prog->aux->jited_linfo); 162 prog->aux->jited_linfo = NULL; 163 } 164 165 kfree(prog->aux->kfunc_tab); 166 prog->aux->kfunc_tab = NULL; 167 } 168 169 /* The jit engine is responsible to provide an array 170 * for insn_off to the jited_off mapping (insn_to_jit_off). 171 * 172 * The idx to this array is the insn_off. Hence, the insn_off 173 * here is relative to the prog itself instead of the main prog. 174 * This array has one entry for each xlated bpf insn. 175 * 176 * jited_off is the byte off to the last byte of the jited insn. 177 * 178 * Hence, with 179 * insn_start: 180 * The first bpf insn off of the prog. The insn off 181 * here is relative to the main prog. 182 * e.g. if prog is a subprog, insn_start > 0 183 * linfo_idx: 184 * The prog's idx to prog->aux->linfo and jited_linfo 185 * 186 * jited_linfo[linfo_idx] = prog->bpf_func 187 * 188 * For i > linfo_idx, 189 * 190 * jited_linfo[i] = prog->bpf_func + 191 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 192 */ 193 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 194 const u32 *insn_to_jit_off) 195 { 196 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 197 const struct bpf_line_info *linfo; 198 void **jited_linfo; 199 200 if (!prog->aux->jited_linfo) 201 /* Userspace did not provide linfo */ 202 return; 203 204 linfo_idx = prog->aux->linfo_idx; 205 linfo = &prog->aux->linfo[linfo_idx]; 206 insn_start = linfo[0].insn_off; 207 insn_end = insn_start + prog->len; 208 209 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 210 jited_linfo[0] = prog->bpf_func; 211 212 nr_linfo = prog->aux->nr_linfo - linfo_idx; 213 214 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 215 /* The verifier ensures that linfo[i].insn_off is 216 * strictly increasing 217 */ 218 jited_linfo[i] = prog->bpf_func + 219 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 220 } 221 222 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 223 gfp_t gfp_extra_flags) 224 { 225 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 226 struct bpf_prog *fp; 227 u32 pages; 228 229 size = round_up(size, PAGE_SIZE); 230 pages = size / PAGE_SIZE; 231 if (pages <= fp_old->pages) 232 return fp_old; 233 234 fp = __vmalloc(size, gfp_flags); 235 if (fp) { 236 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 237 fp->pages = pages; 238 fp->aux->prog = fp; 239 240 /* We keep fp->aux from fp_old around in the new 241 * reallocated structure. 242 */ 243 fp_old->aux = NULL; 244 fp_old->stats = NULL; 245 fp_old->active = NULL; 246 __bpf_prog_free(fp_old); 247 } 248 249 return fp; 250 } 251 252 void __bpf_prog_free(struct bpf_prog *fp) 253 { 254 if (fp->aux) { 255 mutex_destroy(&fp->aux->used_maps_mutex); 256 mutex_destroy(&fp->aux->dst_mutex); 257 kfree(fp->aux->poke_tab); 258 kfree(fp->aux); 259 } 260 free_percpu(fp->stats); 261 free_percpu(fp->active); 262 vfree(fp); 263 } 264 265 int bpf_prog_calc_tag(struct bpf_prog *fp) 266 { 267 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 268 u32 raw_size = bpf_prog_tag_scratch_size(fp); 269 u32 digest[SHA1_DIGEST_WORDS]; 270 u32 ws[SHA1_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 sha1_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, SHA1_BLOCK_SIZE); 314 blocks = bsize / SHA1_BLOCK_SIZE; 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 sha1_transform(digest, todo, ws); 326 todo += SHA1_BLOCK_SIZE; 327 } 328 329 result = (__force __be32 *)digest; 330 for (i = 0; i < SHA1_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_DEFAULT_ON); 524 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 525 int bpf_jit_harden __read_mostly; 526 long bpf_jit_limit __read_mostly; 527 long bpf_jit_limit_max __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_max = bpf_jit_alloc_exec_limit(); 822 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2, 823 PAGE_SIZE), LONG_MAX); 824 return 0; 825 } 826 pure_initcall(bpf_jit_charge_init); 827 828 int bpf_jit_charge_modmem(u32 pages) 829 { 830 if (atomic_long_add_return(pages, &bpf_jit_current) > 831 (bpf_jit_limit >> PAGE_SHIFT)) { 832 if (!bpf_capable()) { 833 atomic_long_sub(pages, &bpf_jit_current); 834 return -EPERM; 835 } 836 } 837 838 return 0; 839 } 840 841 void bpf_jit_uncharge_modmem(u32 pages) 842 { 843 atomic_long_sub(pages, &bpf_jit_current); 844 } 845 846 void *__weak bpf_jit_alloc_exec(unsigned long size) 847 { 848 return module_alloc(size); 849 } 850 851 void __weak bpf_jit_free_exec(void *addr) 852 { 853 module_memfree(addr); 854 } 855 856 struct bpf_binary_header * 857 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 858 unsigned int alignment, 859 bpf_jit_fill_hole_t bpf_fill_ill_insns) 860 { 861 struct bpf_binary_header *hdr; 862 u32 size, hole, start, pages; 863 864 WARN_ON_ONCE(!is_power_of_2(alignment) || 865 alignment > BPF_IMAGE_ALIGNMENT); 866 867 /* Most of BPF filters are really small, but if some of them 868 * fill a page, allow at least 128 extra bytes to insert a 869 * random section of illegal instructions. 870 */ 871 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 872 pages = size / PAGE_SIZE; 873 874 if (bpf_jit_charge_modmem(pages)) 875 return NULL; 876 hdr = bpf_jit_alloc_exec(size); 877 if (!hdr) { 878 bpf_jit_uncharge_modmem(pages); 879 return NULL; 880 } 881 882 /* Fill space with illegal/arch-dep instructions. */ 883 bpf_fill_ill_insns(hdr, size); 884 885 hdr->pages = pages; 886 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 887 PAGE_SIZE - sizeof(*hdr)); 888 start = (get_random_int() % hole) & ~(alignment - 1); 889 890 /* Leave a random number of instructions before BPF code. */ 891 *image_ptr = &hdr->image[start]; 892 893 return hdr; 894 } 895 896 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 897 { 898 u32 pages = hdr->pages; 899 900 bpf_jit_free_exec(hdr); 901 bpf_jit_uncharge_modmem(pages); 902 } 903 904 /* This symbol is only overridden by archs that have different 905 * requirements than the usual eBPF JITs, f.e. when they only 906 * implement cBPF JIT, do not set images read-only, etc. 907 */ 908 void __weak bpf_jit_free(struct bpf_prog *fp) 909 { 910 if (fp->jited) { 911 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 912 913 bpf_jit_binary_free(hdr); 914 915 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 916 } 917 918 bpf_prog_unlock_free(fp); 919 } 920 921 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 922 const struct bpf_insn *insn, bool extra_pass, 923 u64 *func_addr, bool *func_addr_fixed) 924 { 925 s16 off = insn->off; 926 s32 imm = insn->imm; 927 u8 *addr; 928 929 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 930 if (!*func_addr_fixed) { 931 /* Place-holder address till the last pass has collected 932 * all addresses for JITed subprograms in which case we 933 * can pick them up from prog->aux. 934 */ 935 if (!extra_pass) 936 addr = NULL; 937 else if (prog->aux->func && 938 off >= 0 && off < prog->aux->func_cnt) 939 addr = (u8 *)prog->aux->func[off]->bpf_func; 940 else 941 return -EINVAL; 942 } else { 943 /* Address of a BPF helper call. Since part of the core 944 * kernel, it's always at a fixed location. __bpf_call_base 945 * and the helper with imm relative to it are both in core 946 * kernel. 947 */ 948 addr = (u8 *)__bpf_call_base + imm; 949 } 950 951 *func_addr = (unsigned long)addr; 952 return 0; 953 } 954 955 static int bpf_jit_blind_insn(const struct bpf_insn *from, 956 const struct bpf_insn *aux, 957 struct bpf_insn *to_buff, 958 bool emit_zext) 959 { 960 struct bpf_insn *to = to_buff; 961 u32 imm_rnd = get_random_int(); 962 s16 off; 963 964 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 965 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 966 967 /* Constraints on AX register: 968 * 969 * AX register is inaccessible from user space. It is mapped in 970 * all JITs, and used here for constant blinding rewrites. It is 971 * typically "stateless" meaning its contents are only valid within 972 * the executed instruction, but not across several instructions. 973 * There are a few exceptions however which are further detailed 974 * below. 975 * 976 * Constant blinding is only used by JITs, not in the interpreter. 977 * The interpreter uses AX in some occasions as a local temporary 978 * register e.g. in DIV or MOD instructions. 979 * 980 * In restricted circumstances, the verifier can also use the AX 981 * register for rewrites as long as they do not interfere with 982 * the above cases! 983 */ 984 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 985 goto out; 986 987 if (from->imm == 0 && 988 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 989 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 990 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 991 goto out; 992 } 993 994 switch (from->code) { 995 case BPF_ALU | BPF_ADD | BPF_K: 996 case BPF_ALU | BPF_SUB | BPF_K: 997 case BPF_ALU | BPF_AND | BPF_K: 998 case BPF_ALU | BPF_OR | BPF_K: 999 case BPF_ALU | BPF_XOR | BPF_K: 1000 case BPF_ALU | BPF_MUL | BPF_K: 1001 case BPF_ALU | BPF_MOV | BPF_K: 1002 case BPF_ALU | BPF_DIV | BPF_K: 1003 case BPF_ALU | BPF_MOD | BPF_K: 1004 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1005 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1006 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 1007 break; 1008 1009 case BPF_ALU64 | BPF_ADD | BPF_K: 1010 case BPF_ALU64 | BPF_SUB | BPF_K: 1011 case BPF_ALU64 | BPF_AND | BPF_K: 1012 case BPF_ALU64 | BPF_OR | BPF_K: 1013 case BPF_ALU64 | BPF_XOR | BPF_K: 1014 case BPF_ALU64 | BPF_MUL | BPF_K: 1015 case BPF_ALU64 | BPF_MOV | BPF_K: 1016 case BPF_ALU64 | BPF_DIV | BPF_K: 1017 case BPF_ALU64 | BPF_MOD | BPF_K: 1018 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1019 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1020 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 1021 break; 1022 1023 case BPF_JMP | BPF_JEQ | BPF_K: 1024 case BPF_JMP | BPF_JNE | BPF_K: 1025 case BPF_JMP | BPF_JGT | BPF_K: 1026 case BPF_JMP | BPF_JLT | BPF_K: 1027 case BPF_JMP | BPF_JGE | BPF_K: 1028 case BPF_JMP | BPF_JLE | BPF_K: 1029 case BPF_JMP | BPF_JSGT | BPF_K: 1030 case BPF_JMP | BPF_JSLT | BPF_K: 1031 case BPF_JMP | BPF_JSGE | BPF_K: 1032 case BPF_JMP | BPF_JSLE | BPF_K: 1033 case BPF_JMP | BPF_JSET | BPF_K: 1034 /* Accommodate for extra offset in case of a backjump. */ 1035 off = from->off; 1036 if (off < 0) 1037 off -= 2; 1038 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1039 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1040 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1041 break; 1042 1043 case BPF_JMP32 | BPF_JEQ | BPF_K: 1044 case BPF_JMP32 | BPF_JNE | BPF_K: 1045 case BPF_JMP32 | BPF_JGT | BPF_K: 1046 case BPF_JMP32 | BPF_JLT | BPF_K: 1047 case BPF_JMP32 | BPF_JGE | BPF_K: 1048 case BPF_JMP32 | BPF_JLE | BPF_K: 1049 case BPF_JMP32 | BPF_JSGT | BPF_K: 1050 case BPF_JMP32 | BPF_JSLT | BPF_K: 1051 case BPF_JMP32 | BPF_JSGE | BPF_K: 1052 case BPF_JMP32 | BPF_JSLE | BPF_K: 1053 case BPF_JMP32 | BPF_JSET | BPF_K: 1054 /* Accommodate for extra offset in case of a backjump. */ 1055 off = from->off; 1056 if (off < 0) 1057 off -= 2; 1058 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1059 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1060 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1061 off); 1062 break; 1063 1064 case BPF_LD | BPF_IMM | BPF_DW: 1065 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1066 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1067 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1068 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1069 break; 1070 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1071 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1072 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1073 if (emit_zext) 1074 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1075 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1076 break; 1077 1078 case BPF_ST | BPF_MEM | BPF_DW: 1079 case BPF_ST | BPF_MEM | BPF_W: 1080 case BPF_ST | BPF_MEM | BPF_H: 1081 case BPF_ST | BPF_MEM | BPF_B: 1082 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1083 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1084 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1085 break; 1086 } 1087 out: 1088 return to - to_buff; 1089 } 1090 1091 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1092 gfp_t gfp_extra_flags) 1093 { 1094 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1095 struct bpf_prog *fp; 1096 1097 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1098 if (fp != NULL) { 1099 /* aux->prog still points to the fp_other one, so 1100 * when promoting the clone to the real program, 1101 * this still needs to be adapted. 1102 */ 1103 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1104 } 1105 1106 return fp; 1107 } 1108 1109 static void bpf_prog_clone_free(struct bpf_prog *fp) 1110 { 1111 /* aux was stolen by the other clone, so we cannot free 1112 * it from this path! It will be freed eventually by the 1113 * other program on release. 1114 * 1115 * At this point, we don't need a deferred release since 1116 * clone is guaranteed to not be locked. 1117 */ 1118 fp->aux = NULL; 1119 fp->stats = NULL; 1120 fp->active = NULL; 1121 __bpf_prog_free(fp); 1122 } 1123 1124 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1125 { 1126 /* We have to repoint aux->prog to self, as we don't 1127 * know whether fp here is the clone or the original. 1128 */ 1129 fp->aux->prog = fp; 1130 bpf_prog_clone_free(fp_other); 1131 } 1132 1133 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1134 { 1135 struct bpf_insn insn_buff[16], aux[2]; 1136 struct bpf_prog *clone, *tmp; 1137 int insn_delta, insn_cnt; 1138 struct bpf_insn *insn; 1139 int i, rewritten; 1140 1141 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 1142 return prog; 1143 1144 clone = bpf_prog_clone_create(prog, GFP_USER); 1145 if (!clone) 1146 return ERR_PTR(-ENOMEM); 1147 1148 insn_cnt = clone->len; 1149 insn = clone->insnsi; 1150 1151 for (i = 0; i < insn_cnt; i++, insn++) { 1152 /* We temporarily need to hold the original ld64 insn 1153 * so that we can still access the first part in the 1154 * second blinding run. 1155 */ 1156 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1157 insn[1].code == 0) 1158 memcpy(aux, insn, sizeof(aux)); 1159 1160 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1161 clone->aux->verifier_zext); 1162 if (!rewritten) 1163 continue; 1164 1165 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1166 if (IS_ERR(tmp)) { 1167 /* Patching may have repointed aux->prog during 1168 * realloc from the original one, so we need to 1169 * fix it up here on error. 1170 */ 1171 bpf_jit_prog_release_other(prog, clone); 1172 return tmp; 1173 } 1174 1175 clone = tmp; 1176 insn_delta = rewritten - 1; 1177 1178 /* Walk new program and skip insns we just inserted. */ 1179 insn = clone->insnsi + i + insn_delta; 1180 insn_cnt += insn_delta; 1181 i += insn_delta; 1182 } 1183 1184 clone->blinded = 1; 1185 return clone; 1186 } 1187 #endif /* CONFIG_BPF_JIT */ 1188 1189 /* Base function for offset calculation. Needs to go into .text section, 1190 * therefore keeping it non-static as well; will also be used by JITs 1191 * anyway later on, so do not let the compiler omit it. This also needs 1192 * to go into kallsyms for correlation from e.g. bpftool, so naming 1193 * must not change. 1194 */ 1195 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1196 { 1197 return 0; 1198 } 1199 EXPORT_SYMBOL_GPL(__bpf_call_base); 1200 1201 /* All UAPI available opcodes. */ 1202 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1203 /* 32 bit ALU operations. */ \ 1204 /* Register based. */ \ 1205 INSN_3(ALU, ADD, X), \ 1206 INSN_3(ALU, SUB, X), \ 1207 INSN_3(ALU, AND, X), \ 1208 INSN_3(ALU, OR, X), \ 1209 INSN_3(ALU, LSH, X), \ 1210 INSN_3(ALU, RSH, X), \ 1211 INSN_3(ALU, XOR, X), \ 1212 INSN_3(ALU, MUL, X), \ 1213 INSN_3(ALU, MOV, X), \ 1214 INSN_3(ALU, ARSH, X), \ 1215 INSN_3(ALU, DIV, X), \ 1216 INSN_3(ALU, MOD, X), \ 1217 INSN_2(ALU, NEG), \ 1218 INSN_3(ALU, END, TO_BE), \ 1219 INSN_3(ALU, END, TO_LE), \ 1220 /* Immediate based. */ \ 1221 INSN_3(ALU, ADD, K), \ 1222 INSN_3(ALU, SUB, K), \ 1223 INSN_3(ALU, AND, K), \ 1224 INSN_3(ALU, OR, K), \ 1225 INSN_3(ALU, LSH, K), \ 1226 INSN_3(ALU, RSH, K), \ 1227 INSN_3(ALU, XOR, K), \ 1228 INSN_3(ALU, MUL, K), \ 1229 INSN_3(ALU, MOV, K), \ 1230 INSN_3(ALU, ARSH, K), \ 1231 INSN_3(ALU, DIV, K), \ 1232 INSN_3(ALU, MOD, K), \ 1233 /* 64 bit ALU operations. */ \ 1234 /* Register based. */ \ 1235 INSN_3(ALU64, ADD, X), \ 1236 INSN_3(ALU64, SUB, X), \ 1237 INSN_3(ALU64, AND, X), \ 1238 INSN_3(ALU64, OR, X), \ 1239 INSN_3(ALU64, LSH, X), \ 1240 INSN_3(ALU64, RSH, X), \ 1241 INSN_3(ALU64, XOR, X), \ 1242 INSN_3(ALU64, MUL, X), \ 1243 INSN_3(ALU64, MOV, X), \ 1244 INSN_3(ALU64, ARSH, X), \ 1245 INSN_3(ALU64, DIV, X), \ 1246 INSN_3(ALU64, MOD, X), \ 1247 INSN_2(ALU64, NEG), \ 1248 /* Immediate based. */ \ 1249 INSN_3(ALU64, ADD, K), \ 1250 INSN_3(ALU64, SUB, K), \ 1251 INSN_3(ALU64, AND, K), \ 1252 INSN_3(ALU64, OR, K), \ 1253 INSN_3(ALU64, LSH, K), \ 1254 INSN_3(ALU64, RSH, K), \ 1255 INSN_3(ALU64, XOR, K), \ 1256 INSN_3(ALU64, MUL, K), \ 1257 INSN_3(ALU64, MOV, K), \ 1258 INSN_3(ALU64, ARSH, K), \ 1259 INSN_3(ALU64, DIV, K), \ 1260 INSN_3(ALU64, MOD, K), \ 1261 /* Call instruction. */ \ 1262 INSN_2(JMP, CALL), \ 1263 /* Exit instruction. */ \ 1264 INSN_2(JMP, EXIT), \ 1265 /* 32-bit Jump instructions. */ \ 1266 /* Register based. */ \ 1267 INSN_3(JMP32, JEQ, X), \ 1268 INSN_3(JMP32, JNE, X), \ 1269 INSN_3(JMP32, JGT, X), \ 1270 INSN_3(JMP32, JLT, X), \ 1271 INSN_3(JMP32, JGE, X), \ 1272 INSN_3(JMP32, JLE, X), \ 1273 INSN_3(JMP32, JSGT, X), \ 1274 INSN_3(JMP32, JSLT, X), \ 1275 INSN_3(JMP32, JSGE, X), \ 1276 INSN_3(JMP32, JSLE, X), \ 1277 INSN_3(JMP32, JSET, X), \ 1278 /* Immediate based. */ \ 1279 INSN_3(JMP32, JEQ, K), \ 1280 INSN_3(JMP32, JNE, K), \ 1281 INSN_3(JMP32, JGT, K), \ 1282 INSN_3(JMP32, JLT, K), \ 1283 INSN_3(JMP32, JGE, K), \ 1284 INSN_3(JMP32, JLE, K), \ 1285 INSN_3(JMP32, JSGT, K), \ 1286 INSN_3(JMP32, JSLT, K), \ 1287 INSN_3(JMP32, JSGE, K), \ 1288 INSN_3(JMP32, JSLE, K), \ 1289 INSN_3(JMP32, JSET, K), \ 1290 /* Jump instructions. */ \ 1291 /* Register based. */ \ 1292 INSN_3(JMP, JEQ, X), \ 1293 INSN_3(JMP, JNE, X), \ 1294 INSN_3(JMP, JGT, X), \ 1295 INSN_3(JMP, JLT, X), \ 1296 INSN_3(JMP, JGE, X), \ 1297 INSN_3(JMP, JLE, X), \ 1298 INSN_3(JMP, JSGT, X), \ 1299 INSN_3(JMP, JSLT, X), \ 1300 INSN_3(JMP, JSGE, X), \ 1301 INSN_3(JMP, JSLE, X), \ 1302 INSN_3(JMP, JSET, X), \ 1303 /* Immediate based. */ \ 1304 INSN_3(JMP, JEQ, K), \ 1305 INSN_3(JMP, JNE, K), \ 1306 INSN_3(JMP, JGT, K), \ 1307 INSN_3(JMP, JLT, K), \ 1308 INSN_3(JMP, JGE, K), \ 1309 INSN_3(JMP, JLE, K), \ 1310 INSN_3(JMP, JSGT, K), \ 1311 INSN_3(JMP, JSLT, K), \ 1312 INSN_3(JMP, JSGE, K), \ 1313 INSN_3(JMP, JSLE, K), \ 1314 INSN_3(JMP, JSET, K), \ 1315 INSN_2(JMP, JA), \ 1316 /* Store instructions. */ \ 1317 /* Register based. */ \ 1318 INSN_3(STX, MEM, B), \ 1319 INSN_3(STX, MEM, H), \ 1320 INSN_3(STX, MEM, W), \ 1321 INSN_3(STX, MEM, DW), \ 1322 INSN_3(STX, ATOMIC, W), \ 1323 INSN_3(STX, ATOMIC, DW), \ 1324 /* Immediate based. */ \ 1325 INSN_3(ST, MEM, B), \ 1326 INSN_3(ST, MEM, H), \ 1327 INSN_3(ST, MEM, W), \ 1328 INSN_3(ST, MEM, DW), \ 1329 /* Load instructions. */ \ 1330 /* Register based. */ \ 1331 INSN_3(LDX, MEM, B), \ 1332 INSN_3(LDX, MEM, H), \ 1333 INSN_3(LDX, MEM, W), \ 1334 INSN_3(LDX, MEM, DW), \ 1335 /* Immediate based. */ \ 1336 INSN_3(LD, IMM, DW) 1337 1338 bool bpf_opcode_in_insntable(u8 code) 1339 { 1340 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1341 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1342 static const bool public_insntable[256] = { 1343 [0 ... 255] = false, 1344 /* Now overwrite non-defaults ... */ 1345 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1346 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1347 [BPF_LD | BPF_ABS | BPF_B] = true, 1348 [BPF_LD | BPF_ABS | BPF_H] = true, 1349 [BPF_LD | BPF_ABS | BPF_W] = true, 1350 [BPF_LD | BPF_IND | BPF_B] = true, 1351 [BPF_LD | BPF_IND | BPF_H] = true, 1352 [BPF_LD | BPF_IND | BPF_W] = true, 1353 }; 1354 #undef BPF_INSN_3_TBL 1355 #undef BPF_INSN_2_TBL 1356 return public_insntable[code]; 1357 } 1358 1359 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1360 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1361 { 1362 memset(dst, 0, size); 1363 return -EFAULT; 1364 } 1365 1366 /** 1367 * ___bpf_prog_run - run eBPF program on a given context 1368 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1369 * @insn: is the array of eBPF instructions 1370 * 1371 * Decode and execute eBPF instructions. 1372 * 1373 * Return: whatever value is in %BPF_R0 at program exit 1374 */ 1375 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1376 { 1377 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1378 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1379 static const void * const jumptable[256] __annotate_jump_table = { 1380 [0 ... 255] = &&default_label, 1381 /* Now overwrite non-defaults ... */ 1382 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1383 /* Non-UAPI available opcodes. */ 1384 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1385 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1386 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1387 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1388 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1389 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1390 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1391 }; 1392 #undef BPF_INSN_3_LBL 1393 #undef BPF_INSN_2_LBL 1394 u32 tail_call_cnt = 0; 1395 1396 #define CONT ({ insn++; goto select_insn; }) 1397 #define CONT_JMP ({ insn++; goto select_insn; }) 1398 1399 select_insn: 1400 goto *jumptable[insn->code]; 1401 1402 /* Explicitly mask the register-based shift amounts with 63 or 31 1403 * to avoid undefined behavior. Normally this won't affect the 1404 * generated code, for example, in case of native 64 bit archs such 1405 * as x86-64 or arm64, the compiler is optimizing the AND away for 1406 * the interpreter. In case of JITs, each of the JIT backends compiles 1407 * the BPF shift operations to machine instructions which produce 1408 * implementation-defined results in such a case; the resulting 1409 * contents of the register may be arbitrary, but program behaviour 1410 * as a whole remains defined. In other words, in case of JIT backends, 1411 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1412 */ 1413 /* ALU (shifts) */ 1414 #define SHT(OPCODE, OP) \ 1415 ALU64_##OPCODE##_X: \ 1416 DST = DST OP (SRC & 63); \ 1417 CONT; \ 1418 ALU_##OPCODE##_X: \ 1419 DST = (u32) DST OP ((u32) SRC & 31); \ 1420 CONT; \ 1421 ALU64_##OPCODE##_K: \ 1422 DST = DST OP IMM; \ 1423 CONT; \ 1424 ALU_##OPCODE##_K: \ 1425 DST = (u32) DST OP (u32) IMM; \ 1426 CONT; 1427 /* ALU (rest) */ 1428 #define ALU(OPCODE, OP) \ 1429 ALU64_##OPCODE##_X: \ 1430 DST = DST OP SRC; \ 1431 CONT; \ 1432 ALU_##OPCODE##_X: \ 1433 DST = (u32) DST OP (u32) SRC; \ 1434 CONT; \ 1435 ALU64_##OPCODE##_K: \ 1436 DST = DST OP IMM; \ 1437 CONT; \ 1438 ALU_##OPCODE##_K: \ 1439 DST = (u32) DST OP (u32) IMM; \ 1440 CONT; 1441 ALU(ADD, +) 1442 ALU(SUB, -) 1443 ALU(AND, &) 1444 ALU(OR, |) 1445 ALU(XOR, ^) 1446 ALU(MUL, *) 1447 SHT(LSH, <<) 1448 SHT(RSH, >>) 1449 #undef SHT 1450 #undef ALU 1451 ALU_NEG: 1452 DST = (u32) -DST; 1453 CONT; 1454 ALU64_NEG: 1455 DST = -DST; 1456 CONT; 1457 ALU_MOV_X: 1458 DST = (u32) SRC; 1459 CONT; 1460 ALU_MOV_K: 1461 DST = (u32) IMM; 1462 CONT; 1463 ALU64_MOV_X: 1464 DST = SRC; 1465 CONT; 1466 ALU64_MOV_K: 1467 DST = IMM; 1468 CONT; 1469 LD_IMM_DW: 1470 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1471 insn++; 1472 CONT; 1473 ALU_ARSH_X: 1474 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1475 CONT; 1476 ALU_ARSH_K: 1477 DST = (u64) (u32) (((s32) DST) >> IMM); 1478 CONT; 1479 ALU64_ARSH_X: 1480 (*(s64 *) &DST) >>= (SRC & 63); 1481 CONT; 1482 ALU64_ARSH_K: 1483 (*(s64 *) &DST) >>= IMM; 1484 CONT; 1485 ALU64_MOD_X: 1486 div64_u64_rem(DST, SRC, &AX); 1487 DST = AX; 1488 CONT; 1489 ALU_MOD_X: 1490 AX = (u32) DST; 1491 DST = do_div(AX, (u32) SRC); 1492 CONT; 1493 ALU64_MOD_K: 1494 div64_u64_rem(DST, IMM, &AX); 1495 DST = AX; 1496 CONT; 1497 ALU_MOD_K: 1498 AX = (u32) DST; 1499 DST = do_div(AX, (u32) IMM); 1500 CONT; 1501 ALU64_DIV_X: 1502 DST = div64_u64(DST, SRC); 1503 CONT; 1504 ALU_DIV_X: 1505 AX = (u32) DST; 1506 do_div(AX, (u32) SRC); 1507 DST = (u32) AX; 1508 CONT; 1509 ALU64_DIV_K: 1510 DST = div64_u64(DST, IMM); 1511 CONT; 1512 ALU_DIV_K: 1513 AX = (u32) DST; 1514 do_div(AX, (u32) IMM); 1515 DST = (u32) AX; 1516 CONT; 1517 ALU_END_TO_BE: 1518 switch (IMM) { 1519 case 16: 1520 DST = (__force u16) cpu_to_be16(DST); 1521 break; 1522 case 32: 1523 DST = (__force u32) cpu_to_be32(DST); 1524 break; 1525 case 64: 1526 DST = (__force u64) cpu_to_be64(DST); 1527 break; 1528 } 1529 CONT; 1530 ALU_END_TO_LE: 1531 switch (IMM) { 1532 case 16: 1533 DST = (__force u16) cpu_to_le16(DST); 1534 break; 1535 case 32: 1536 DST = (__force u32) cpu_to_le32(DST); 1537 break; 1538 case 64: 1539 DST = (__force u64) cpu_to_le64(DST); 1540 break; 1541 } 1542 CONT; 1543 1544 /* CALL */ 1545 JMP_CALL: 1546 /* Function call scratches BPF_R1-BPF_R5 registers, 1547 * preserves BPF_R6-BPF_R9, and stores return value 1548 * into BPF_R0. 1549 */ 1550 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1551 BPF_R4, BPF_R5); 1552 CONT; 1553 1554 JMP_CALL_ARGS: 1555 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1556 BPF_R3, BPF_R4, 1557 BPF_R5, 1558 insn + insn->off + 1); 1559 CONT; 1560 1561 JMP_TAIL_CALL: { 1562 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1563 struct bpf_array *array = container_of(map, struct bpf_array, map); 1564 struct bpf_prog *prog; 1565 u32 index = BPF_R3; 1566 1567 if (unlikely(index >= array->map.max_entries)) 1568 goto out; 1569 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1570 goto out; 1571 1572 tail_call_cnt++; 1573 1574 prog = READ_ONCE(array->ptrs[index]); 1575 if (!prog) 1576 goto out; 1577 1578 /* ARG1 at this point is guaranteed to point to CTX from 1579 * the verifier side due to the fact that the tail call is 1580 * handled like a helper, that is, bpf_tail_call_proto, 1581 * where arg1_type is ARG_PTR_TO_CTX. 1582 */ 1583 insn = prog->insnsi; 1584 goto select_insn; 1585 out: 1586 CONT; 1587 } 1588 JMP_JA: 1589 insn += insn->off; 1590 CONT; 1591 JMP_EXIT: 1592 return BPF_R0; 1593 /* JMP */ 1594 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1595 JMP_##OPCODE##_X: \ 1596 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1597 insn += insn->off; \ 1598 CONT_JMP; \ 1599 } \ 1600 CONT; \ 1601 JMP32_##OPCODE##_X: \ 1602 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1603 insn += insn->off; \ 1604 CONT_JMP; \ 1605 } \ 1606 CONT; \ 1607 JMP_##OPCODE##_K: \ 1608 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1609 insn += insn->off; \ 1610 CONT_JMP; \ 1611 } \ 1612 CONT; \ 1613 JMP32_##OPCODE##_K: \ 1614 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1615 insn += insn->off; \ 1616 CONT_JMP; \ 1617 } \ 1618 CONT; 1619 COND_JMP(u, JEQ, ==) 1620 COND_JMP(u, JNE, !=) 1621 COND_JMP(u, JGT, >) 1622 COND_JMP(u, JLT, <) 1623 COND_JMP(u, JGE, >=) 1624 COND_JMP(u, JLE, <=) 1625 COND_JMP(u, JSET, &) 1626 COND_JMP(s, JSGT, >) 1627 COND_JMP(s, JSLT, <) 1628 COND_JMP(s, JSGE, >=) 1629 COND_JMP(s, JSLE, <=) 1630 #undef COND_JMP 1631 /* ST, STX and LDX*/ 1632 ST_NOSPEC: 1633 /* Speculation barrier for mitigating Speculative Store Bypass. 1634 * In case of arm64, we rely on the firmware mitigation as 1635 * controlled via the ssbd kernel parameter. Whenever the 1636 * mitigation is enabled, it works for all of the kernel code 1637 * with no need to provide any additional instructions here. 1638 * In case of x86, we use 'lfence' insn for mitigation. We 1639 * reuse preexisting logic from Spectre v1 mitigation that 1640 * happens to produce the required code on x86 for v4 as well. 1641 */ 1642 #ifdef CONFIG_X86 1643 barrier_nospec(); 1644 #endif 1645 CONT; 1646 #define LDST(SIZEOP, SIZE) \ 1647 STX_MEM_##SIZEOP: \ 1648 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1649 CONT; \ 1650 ST_MEM_##SIZEOP: \ 1651 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1652 CONT; \ 1653 LDX_MEM_##SIZEOP: \ 1654 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1655 CONT; 1656 1657 LDST(B, u8) 1658 LDST(H, u16) 1659 LDST(W, u32) 1660 LDST(DW, u64) 1661 #undef LDST 1662 #define LDX_PROBE(SIZEOP, SIZE) \ 1663 LDX_PROBE_MEM_##SIZEOP: \ 1664 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \ 1665 CONT; 1666 LDX_PROBE(B, 1) 1667 LDX_PROBE(H, 2) 1668 LDX_PROBE(W, 4) 1669 LDX_PROBE(DW, 8) 1670 #undef LDX_PROBE 1671 1672 #define ATOMIC_ALU_OP(BOP, KOP) \ 1673 case BOP: \ 1674 if (BPF_SIZE(insn->code) == BPF_W) \ 1675 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 1676 (DST + insn->off)); \ 1677 else \ 1678 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 1679 (DST + insn->off)); \ 1680 break; \ 1681 case BOP | BPF_FETCH: \ 1682 if (BPF_SIZE(insn->code) == BPF_W) \ 1683 SRC = (u32) atomic_fetch_##KOP( \ 1684 (u32) SRC, \ 1685 (atomic_t *)(unsigned long) (DST + insn->off)); \ 1686 else \ 1687 SRC = (u64) atomic64_fetch_##KOP( \ 1688 (u64) SRC, \ 1689 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 1690 break; 1691 1692 STX_ATOMIC_DW: 1693 STX_ATOMIC_W: 1694 switch (IMM) { 1695 ATOMIC_ALU_OP(BPF_ADD, add) 1696 ATOMIC_ALU_OP(BPF_AND, and) 1697 ATOMIC_ALU_OP(BPF_OR, or) 1698 ATOMIC_ALU_OP(BPF_XOR, xor) 1699 #undef ATOMIC_ALU_OP 1700 1701 case BPF_XCHG: 1702 if (BPF_SIZE(insn->code) == BPF_W) 1703 SRC = (u32) atomic_xchg( 1704 (atomic_t *)(unsigned long) (DST + insn->off), 1705 (u32) SRC); 1706 else 1707 SRC = (u64) atomic64_xchg( 1708 (atomic64_t *)(unsigned long) (DST + insn->off), 1709 (u64) SRC); 1710 break; 1711 case BPF_CMPXCHG: 1712 if (BPF_SIZE(insn->code) == BPF_W) 1713 BPF_R0 = (u32) atomic_cmpxchg( 1714 (atomic_t *)(unsigned long) (DST + insn->off), 1715 (u32) BPF_R0, (u32) SRC); 1716 else 1717 BPF_R0 = (u64) atomic64_cmpxchg( 1718 (atomic64_t *)(unsigned long) (DST + insn->off), 1719 (u64) BPF_R0, (u64) SRC); 1720 break; 1721 1722 default: 1723 goto default_label; 1724 } 1725 CONT; 1726 1727 default_label: 1728 /* If we ever reach this, we have a bug somewhere. Die hard here 1729 * instead of just returning 0; we could be somewhere in a subprog, 1730 * so execution could continue otherwise which we do /not/ want. 1731 * 1732 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1733 */ 1734 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 1735 insn->code, insn->imm); 1736 BUG_ON(1); 1737 return 0; 1738 } 1739 1740 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1741 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1742 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1743 { \ 1744 u64 stack[stack_size / sizeof(u64)]; \ 1745 u64 regs[MAX_BPF_EXT_REG]; \ 1746 \ 1747 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1748 ARG1 = (u64) (unsigned long) ctx; \ 1749 return ___bpf_prog_run(regs, insn); \ 1750 } 1751 1752 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1753 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1754 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1755 const struct bpf_insn *insn) \ 1756 { \ 1757 u64 stack[stack_size / sizeof(u64)]; \ 1758 u64 regs[MAX_BPF_EXT_REG]; \ 1759 \ 1760 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1761 BPF_R1 = r1; \ 1762 BPF_R2 = r2; \ 1763 BPF_R3 = r3; \ 1764 BPF_R4 = r4; \ 1765 BPF_R5 = r5; \ 1766 return ___bpf_prog_run(regs, insn); \ 1767 } 1768 1769 #define EVAL1(FN, X) FN(X) 1770 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1771 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1772 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1773 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1774 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1775 1776 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1777 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1778 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1779 1780 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1781 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1782 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1783 1784 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1785 1786 static unsigned int (*interpreters[])(const void *ctx, 1787 const struct bpf_insn *insn) = { 1788 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1789 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1790 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1791 }; 1792 #undef PROG_NAME_LIST 1793 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1794 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1795 const struct bpf_insn *insn) = { 1796 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1797 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1798 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1799 }; 1800 #undef PROG_NAME_LIST 1801 1802 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1803 { 1804 stack_depth = max_t(u32, stack_depth, 1); 1805 insn->off = (s16) insn->imm; 1806 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1807 __bpf_call_base_args; 1808 insn->code = BPF_JMP | BPF_CALL_ARGS; 1809 } 1810 1811 #else 1812 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1813 const struct bpf_insn *insn) 1814 { 1815 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1816 * is not working properly, so warn about it! 1817 */ 1818 WARN_ON_ONCE(1); 1819 return 0; 1820 } 1821 #endif 1822 1823 bool bpf_prog_array_compatible(struct bpf_array *array, 1824 const struct bpf_prog *fp) 1825 { 1826 bool ret; 1827 1828 if (fp->kprobe_override) 1829 return false; 1830 1831 spin_lock(&array->aux->owner.lock); 1832 1833 if (!array->aux->owner.type) { 1834 /* There's no owner yet where we could check for 1835 * compatibility. 1836 */ 1837 array->aux->owner.type = fp->type; 1838 array->aux->owner.jited = fp->jited; 1839 ret = true; 1840 } else { 1841 ret = array->aux->owner.type == fp->type && 1842 array->aux->owner.jited == fp->jited; 1843 } 1844 spin_unlock(&array->aux->owner.lock); 1845 return ret; 1846 } 1847 1848 static int bpf_check_tail_call(const struct bpf_prog *fp) 1849 { 1850 struct bpf_prog_aux *aux = fp->aux; 1851 int i, ret = 0; 1852 1853 mutex_lock(&aux->used_maps_mutex); 1854 for (i = 0; i < aux->used_map_cnt; i++) { 1855 struct bpf_map *map = aux->used_maps[i]; 1856 struct bpf_array *array; 1857 1858 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1859 continue; 1860 1861 array = container_of(map, struct bpf_array, map); 1862 if (!bpf_prog_array_compatible(array, fp)) { 1863 ret = -EINVAL; 1864 goto out; 1865 } 1866 } 1867 1868 out: 1869 mutex_unlock(&aux->used_maps_mutex); 1870 return ret; 1871 } 1872 1873 static void bpf_prog_select_func(struct bpf_prog *fp) 1874 { 1875 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1876 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1877 1878 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1879 #else 1880 fp->bpf_func = __bpf_prog_ret0_warn; 1881 #endif 1882 } 1883 1884 /** 1885 * bpf_prog_select_runtime - select exec runtime for BPF program 1886 * @fp: bpf_prog populated with internal BPF program 1887 * @err: pointer to error variable 1888 * 1889 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1890 * The BPF program will be executed via bpf_prog_run() function. 1891 * 1892 * Return: the &fp argument along with &err set to 0 for success or 1893 * a negative errno code on failure 1894 */ 1895 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1896 { 1897 /* In case of BPF to BPF calls, verifier did all the prep 1898 * work with regards to JITing, etc. 1899 */ 1900 bool jit_needed = false; 1901 1902 if (fp->bpf_func) 1903 goto finalize; 1904 1905 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 1906 bpf_prog_has_kfunc_call(fp)) 1907 jit_needed = true; 1908 1909 bpf_prog_select_func(fp); 1910 1911 /* eBPF JITs can rewrite the program in case constant 1912 * blinding is active. However, in case of error during 1913 * blinding, bpf_int_jit_compile() must always return a 1914 * valid program, which in this case would simply not 1915 * be JITed, but falls back to the interpreter. 1916 */ 1917 if (!bpf_prog_is_dev_bound(fp->aux)) { 1918 *err = bpf_prog_alloc_jited_linfo(fp); 1919 if (*err) 1920 return fp; 1921 1922 fp = bpf_int_jit_compile(fp); 1923 bpf_prog_jit_attempt_done(fp); 1924 if (!fp->jited && jit_needed) { 1925 *err = -ENOTSUPP; 1926 return fp; 1927 } 1928 } else { 1929 *err = bpf_prog_offload_compile(fp); 1930 if (*err) 1931 return fp; 1932 } 1933 1934 finalize: 1935 bpf_prog_lock_ro(fp); 1936 1937 /* The tail call compatibility check can only be done at 1938 * this late stage as we need to determine, if we deal 1939 * with JITed or non JITed program concatenations and not 1940 * all eBPF JITs might immediately support all features. 1941 */ 1942 *err = bpf_check_tail_call(fp); 1943 1944 return fp; 1945 } 1946 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1947 1948 static unsigned int __bpf_prog_ret1(const void *ctx, 1949 const struct bpf_insn *insn) 1950 { 1951 return 1; 1952 } 1953 1954 static struct bpf_prog_dummy { 1955 struct bpf_prog prog; 1956 } dummy_bpf_prog = { 1957 .prog = { 1958 .bpf_func = __bpf_prog_ret1, 1959 }, 1960 }; 1961 1962 /* to avoid allocating empty bpf_prog_array for cgroups that 1963 * don't have bpf program attached use one global 'empty_prog_array' 1964 * It will not be modified the caller of bpf_prog_array_alloc() 1965 * (since caller requested prog_cnt == 0) 1966 * that pointer should be 'freed' by bpf_prog_array_free() 1967 */ 1968 static struct { 1969 struct bpf_prog_array hdr; 1970 struct bpf_prog *null_prog; 1971 } empty_prog_array = { 1972 .null_prog = NULL, 1973 }; 1974 1975 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1976 { 1977 if (prog_cnt) 1978 return kzalloc(sizeof(struct bpf_prog_array) + 1979 sizeof(struct bpf_prog_array_item) * 1980 (prog_cnt + 1), 1981 flags); 1982 1983 return &empty_prog_array.hdr; 1984 } 1985 1986 void bpf_prog_array_free(struct bpf_prog_array *progs) 1987 { 1988 if (!progs || progs == &empty_prog_array.hdr) 1989 return; 1990 kfree_rcu(progs, rcu); 1991 } 1992 1993 int bpf_prog_array_length(struct bpf_prog_array *array) 1994 { 1995 struct bpf_prog_array_item *item; 1996 u32 cnt = 0; 1997 1998 for (item = array->items; item->prog; item++) 1999 if (item->prog != &dummy_bpf_prog.prog) 2000 cnt++; 2001 return cnt; 2002 } 2003 2004 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2005 { 2006 struct bpf_prog_array_item *item; 2007 2008 for (item = array->items; item->prog; item++) 2009 if (item->prog != &dummy_bpf_prog.prog) 2010 return false; 2011 return true; 2012 } 2013 2014 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2015 u32 *prog_ids, 2016 u32 request_cnt) 2017 { 2018 struct bpf_prog_array_item *item; 2019 int i = 0; 2020 2021 for (item = array->items; item->prog; item++) { 2022 if (item->prog == &dummy_bpf_prog.prog) 2023 continue; 2024 prog_ids[i] = item->prog->aux->id; 2025 if (++i == request_cnt) { 2026 item++; 2027 break; 2028 } 2029 } 2030 2031 return !!(item->prog); 2032 } 2033 2034 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2035 __u32 __user *prog_ids, u32 cnt) 2036 { 2037 unsigned long err = 0; 2038 bool nospc; 2039 u32 *ids; 2040 2041 /* users of this function are doing: 2042 * cnt = bpf_prog_array_length(); 2043 * if (cnt > 0) 2044 * bpf_prog_array_copy_to_user(..., cnt); 2045 * so below kcalloc doesn't need extra cnt > 0 check. 2046 */ 2047 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2048 if (!ids) 2049 return -ENOMEM; 2050 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2051 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2052 kfree(ids); 2053 if (err) 2054 return -EFAULT; 2055 if (nospc) 2056 return -ENOSPC; 2057 return 0; 2058 } 2059 2060 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2061 struct bpf_prog *old_prog) 2062 { 2063 struct bpf_prog_array_item *item; 2064 2065 for (item = array->items; item->prog; item++) 2066 if (item->prog == old_prog) { 2067 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2068 break; 2069 } 2070 } 2071 2072 /** 2073 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2074 * index into the program array with 2075 * a dummy no-op program. 2076 * @array: a bpf_prog_array 2077 * @index: the index of the program to replace 2078 * 2079 * Skips over dummy programs, by not counting them, when calculating 2080 * the position of the program to replace. 2081 * 2082 * Return: 2083 * * 0 - Success 2084 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2085 * * -ENOENT - Index out of range 2086 */ 2087 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2088 { 2089 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2090 } 2091 2092 /** 2093 * bpf_prog_array_update_at() - Updates the program at the given index 2094 * into the program array. 2095 * @array: a bpf_prog_array 2096 * @index: the index of the program to update 2097 * @prog: the program to insert into the array 2098 * 2099 * Skips over dummy programs, by not counting them, when calculating 2100 * the position of the program to update. 2101 * 2102 * Return: 2103 * * 0 - Success 2104 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2105 * * -ENOENT - Index out of range 2106 */ 2107 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2108 struct bpf_prog *prog) 2109 { 2110 struct bpf_prog_array_item *item; 2111 2112 if (unlikely(index < 0)) 2113 return -EINVAL; 2114 2115 for (item = array->items; item->prog; item++) { 2116 if (item->prog == &dummy_bpf_prog.prog) 2117 continue; 2118 if (!index) { 2119 WRITE_ONCE(item->prog, prog); 2120 return 0; 2121 } 2122 index--; 2123 } 2124 return -ENOENT; 2125 } 2126 2127 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2128 struct bpf_prog *exclude_prog, 2129 struct bpf_prog *include_prog, 2130 u64 bpf_cookie, 2131 struct bpf_prog_array **new_array) 2132 { 2133 int new_prog_cnt, carry_prog_cnt = 0; 2134 struct bpf_prog_array_item *existing, *new; 2135 struct bpf_prog_array *array; 2136 bool found_exclude = false; 2137 2138 /* Figure out how many existing progs we need to carry over to 2139 * the new array. 2140 */ 2141 if (old_array) { 2142 existing = old_array->items; 2143 for (; existing->prog; existing++) { 2144 if (existing->prog == exclude_prog) { 2145 found_exclude = true; 2146 continue; 2147 } 2148 if (existing->prog != &dummy_bpf_prog.prog) 2149 carry_prog_cnt++; 2150 if (existing->prog == include_prog) 2151 return -EEXIST; 2152 } 2153 } 2154 2155 if (exclude_prog && !found_exclude) 2156 return -ENOENT; 2157 2158 /* How many progs (not NULL) will be in the new array? */ 2159 new_prog_cnt = carry_prog_cnt; 2160 if (include_prog) 2161 new_prog_cnt += 1; 2162 2163 /* Do we have any prog (not NULL) in the new array? */ 2164 if (!new_prog_cnt) { 2165 *new_array = NULL; 2166 return 0; 2167 } 2168 2169 /* +1 as the end of prog_array is marked with NULL */ 2170 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2171 if (!array) 2172 return -ENOMEM; 2173 new = array->items; 2174 2175 /* Fill in the new prog array */ 2176 if (carry_prog_cnt) { 2177 existing = old_array->items; 2178 for (; existing->prog; existing++) { 2179 if (existing->prog == exclude_prog || 2180 existing->prog == &dummy_bpf_prog.prog) 2181 continue; 2182 2183 new->prog = existing->prog; 2184 new->bpf_cookie = existing->bpf_cookie; 2185 new++; 2186 } 2187 } 2188 if (include_prog) { 2189 new->prog = include_prog; 2190 new->bpf_cookie = bpf_cookie; 2191 new++; 2192 } 2193 new->prog = NULL; 2194 *new_array = array; 2195 return 0; 2196 } 2197 2198 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2199 u32 *prog_ids, u32 request_cnt, 2200 u32 *prog_cnt) 2201 { 2202 u32 cnt = 0; 2203 2204 if (array) 2205 cnt = bpf_prog_array_length(array); 2206 2207 *prog_cnt = cnt; 2208 2209 /* return early if user requested only program count or nothing to copy */ 2210 if (!request_cnt || !cnt) 2211 return 0; 2212 2213 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2214 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2215 : 0; 2216 } 2217 2218 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2219 struct bpf_map **used_maps, u32 len) 2220 { 2221 struct bpf_map *map; 2222 u32 i; 2223 2224 for (i = 0; i < len; i++) { 2225 map = used_maps[i]; 2226 if (map->ops->map_poke_untrack) 2227 map->ops->map_poke_untrack(map, aux); 2228 bpf_map_put(map); 2229 } 2230 } 2231 2232 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2233 { 2234 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2235 kfree(aux->used_maps); 2236 } 2237 2238 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2239 struct btf_mod_pair *used_btfs, u32 len) 2240 { 2241 #ifdef CONFIG_BPF_SYSCALL 2242 struct btf_mod_pair *btf_mod; 2243 u32 i; 2244 2245 for (i = 0; i < len; i++) { 2246 btf_mod = &used_btfs[i]; 2247 if (btf_mod->module) 2248 module_put(btf_mod->module); 2249 btf_put(btf_mod->btf); 2250 } 2251 #endif 2252 } 2253 2254 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2255 { 2256 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2257 kfree(aux->used_btfs); 2258 } 2259 2260 static void bpf_prog_free_deferred(struct work_struct *work) 2261 { 2262 struct bpf_prog_aux *aux; 2263 int i; 2264 2265 aux = container_of(work, struct bpf_prog_aux, work); 2266 bpf_free_used_maps(aux); 2267 bpf_free_used_btfs(aux); 2268 if (bpf_prog_is_dev_bound(aux)) 2269 bpf_prog_offload_destroy(aux->prog); 2270 #ifdef CONFIG_PERF_EVENTS 2271 if (aux->prog->has_callchain_buf) 2272 put_callchain_buffers(); 2273 #endif 2274 if (aux->dst_trampoline) 2275 bpf_trampoline_put(aux->dst_trampoline); 2276 for (i = 0; i < aux->func_cnt; i++) { 2277 /* We can just unlink the subprog poke descriptor table as 2278 * it was originally linked to the main program and is also 2279 * released along with it. 2280 */ 2281 aux->func[i]->aux->poke_tab = NULL; 2282 bpf_jit_free(aux->func[i]); 2283 } 2284 if (aux->func_cnt) { 2285 kfree(aux->func); 2286 bpf_prog_unlock_free(aux->prog); 2287 } else { 2288 bpf_jit_free(aux->prog); 2289 } 2290 } 2291 2292 /* Free internal BPF program */ 2293 void bpf_prog_free(struct bpf_prog *fp) 2294 { 2295 struct bpf_prog_aux *aux = fp->aux; 2296 2297 if (aux->dst_prog) 2298 bpf_prog_put(aux->dst_prog); 2299 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2300 schedule_work(&aux->work); 2301 } 2302 EXPORT_SYMBOL_GPL(bpf_prog_free); 2303 2304 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2305 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2306 2307 void bpf_user_rnd_init_once(void) 2308 { 2309 prandom_init_once(&bpf_user_rnd_state); 2310 } 2311 2312 BPF_CALL_0(bpf_user_rnd_u32) 2313 { 2314 /* Should someone ever have the rather unwise idea to use some 2315 * of the registers passed into this function, then note that 2316 * this function is called from native eBPF and classic-to-eBPF 2317 * transformations. Register assignments from both sides are 2318 * different, f.e. classic always sets fn(ctx, A, X) here. 2319 */ 2320 struct rnd_state *state; 2321 u32 res; 2322 2323 state = &get_cpu_var(bpf_user_rnd_state); 2324 res = prandom_u32_state(state); 2325 put_cpu_var(bpf_user_rnd_state); 2326 2327 return res; 2328 } 2329 2330 BPF_CALL_0(bpf_get_raw_cpu_id) 2331 { 2332 return raw_smp_processor_id(); 2333 } 2334 2335 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2336 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2337 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2338 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2339 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2340 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2341 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2342 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2343 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2344 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2345 2346 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2347 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2348 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2349 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2350 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2351 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2352 2353 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2354 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2355 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2356 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2357 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2358 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2359 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2360 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2361 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2362 2363 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2364 { 2365 return NULL; 2366 } 2367 2368 u64 __weak 2369 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2370 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2371 { 2372 return -ENOTSUPP; 2373 } 2374 EXPORT_SYMBOL_GPL(bpf_event_output); 2375 2376 /* Always built-in helper functions. */ 2377 const struct bpf_func_proto bpf_tail_call_proto = { 2378 .func = NULL, 2379 .gpl_only = false, 2380 .ret_type = RET_VOID, 2381 .arg1_type = ARG_PTR_TO_CTX, 2382 .arg2_type = ARG_CONST_MAP_PTR, 2383 .arg3_type = ARG_ANYTHING, 2384 }; 2385 2386 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2387 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2388 * eBPF and implicitly also cBPF can get JITed! 2389 */ 2390 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2391 { 2392 return prog; 2393 } 2394 2395 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2396 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2397 */ 2398 void __weak bpf_jit_compile(struct bpf_prog *prog) 2399 { 2400 } 2401 2402 bool __weak bpf_helper_changes_pkt_data(void *func) 2403 { 2404 return false; 2405 } 2406 2407 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2408 * analysis code and wants explicit zero extension inserted by verifier. 2409 * Otherwise, return FALSE. 2410 * 2411 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2412 * you don't override this. JITs that don't want these extra insns can detect 2413 * them using insn_is_zext. 2414 */ 2415 bool __weak bpf_jit_needs_zext(void) 2416 { 2417 return false; 2418 } 2419 2420 bool __weak bpf_jit_supports_kfunc_call(void) 2421 { 2422 return false; 2423 } 2424 2425 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2426 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2427 */ 2428 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2429 int len) 2430 { 2431 return -EFAULT; 2432 } 2433 2434 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2435 void *addr1, void *addr2) 2436 { 2437 return -ENOTSUPP; 2438 } 2439 2440 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2441 EXPORT_SYMBOL(bpf_stats_enabled_key); 2442 2443 /* All definitions of tracepoints related to BPF. */ 2444 #define CREATE_TRACE_POINTS 2445 #include <linux/bpf_trace.h> 2446 2447 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2448 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2449