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 #include <linux/nodemask.h> 37 38 #include <asm/barrier.h> 39 #include <asm/unaligned.h> 40 41 /* Registers */ 42 #define BPF_R0 regs[BPF_REG_0] 43 #define BPF_R1 regs[BPF_REG_1] 44 #define BPF_R2 regs[BPF_REG_2] 45 #define BPF_R3 regs[BPF_REG_3] 46 #define BPF_R4 regs[BPF_REG_4] 47 #define BPF_R5 regs[BPF_REG_5] 48 #define BPF_R6 regs[BPF_REG_6] 49 #define BPF_R7 regs[BPF_REG_7] 50 #define BPF_R8 regs[BPF_REG_8] 51 #define BPF_R9 regs[BPF_REG_9] 52 #define BPF_R10 regs[BPF_REG_10] 53 54 /* Named registers */ 55 #define DST regs[insn->dst_reg] 56 #define SRC regs[insn->src_reg] 57 #define FP regs[BPF_REG_FP] 58 #define AX regs[BPF_REG_AX] 59 #define ARG1 regs[BPF_REG_ARG1] 60 #define CTX regs[BPF_REG_CTX] 61 #define IMM insn->imm 62 63 /* No hurry in this branch 64 * 65 * Exported for the bpf jit load helper. 66 */ 67 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 68 { 69 u8 *ptr = NULL; 70 71 if (k >= SKF_NET_OFF) 72 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 73 else if (k >= SKF_LL_OFF) 74 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 75 76 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 77 return ptr; 78 79 return NULL; 80 } 81 82 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 83 { 84 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 85 struct bpf_prog_aux *aux; 86 struct bpf_prog *fp; 87 88 size = round_up(size, PAGE_SIZE); 89 fp = __vmalloc(size, gfp_flags); 90 if (fp == NULL) 91 return NULL; 92 93 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags); 94 if (aux == NULL) { 95 vfree(fp); 96 return NULL; 97 } 98 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags); 99 if (!fp->active) { 100 vfree(fp); 101 kfree(aux); 102 return NULL; 103 } 104 105 fp->pages = size / PAGE_SIZE; 106 fp->aux = aux; 107 fp->aux->prog = fp; 108 fp->jit_requested = ebpf_jit_enabled(); 109 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 110 111 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 112 mutex_init(&fp->aux->used_maps_mutex); 113 mutex_init(&fp->aux->dst_mutex); 114 115 return fp; 116 } 117 118 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 119 { 120 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 121 struct bpf_prog *prog; 122 int cpu; 123 124 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 125 if (!prog) 126 return NULL; 127 128 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 129 if (!prog->stats) { 130 free_percpu(prog->active); 131 kfree(prog->aux); 132 vfree(prog); 133 return NULL; 134 } 135 136 for_each_possible_cpu(cpu) { 137 struct bpf_prog_stats *pstats; 138 139 pstats = per_cpu_ptr(prog->stats, cpu); 140 u64_stats_init(&pstats->syncp); 141 } 142 return prog; 143 } 144 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 145 146 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 147 { 148 if (!prog->aux->nr_linfo || !prog->jit_requested) 149 return 0; 150 151 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 152 sizeof(*prog->aux->jited_linfo), 153 GFP_KERNEL_ACCOUNT | __GFP_NOWARN); 154 if (!prog->aux->jited_linfo) 155 return -ENOMEM; 156 157 return 0; 158 } 159 160 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 161 { 162 if (prog->aux->jited_linfo && 163 (!prog->jited || !prog->aux->jited_linfo[0])) { 164 kvfree(prog->aux->jited_linfo); 165 prog->aux->jited_linfo = NULL; 166 } 167 168 kfree(prog->aux->kfunc_tab); 169 prog->aux->kfunc_tab = NULL; 170 } 171 172 /* The jit engine is responsible to provide an array 173 * for insn_off to the jited_off mapping (insn_to_jit_off). 174 * 175 * The idx to this array is the insn_off. Hence, the insn_off 176 * here is relative to the prog itself instead of the main prog. 177 * This array has one entry for each xlated bpf insn. 178 * 179 * jited_off is the byte off to the last byte of the jited insn. 180 * 181 * Hence, with 182 * insn_start: 183 * The first bpf insn off of the prog. The insn off 184 * here is relative to the main prog. 185 * e.g. if prog is a subprog, insn_start > 0 186 * linfo_idx: 187 * The prog's idx to prog->aux->linfo and jited_linfo 188 * 189 * jited_linfo[linfo_idx] = prog->bpf_func 190 * 191 * For i > linfo_idx, 192 * 193 * jited_linfo[i] = prog->bpf_func + 194 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 195 */ 196 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 197 const u32 *insn_to_jit_off) 198 { 199 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 200 const struct bpf_line_info *linfo; 201 void **jited_linfo; 202 203 if (!prog->aux->jited_linfo) 204 /* Userspace did not provide linfo */ 205 return; 206 207 linfo_idx = prog->aux->linfo_idx; 208 linfo = &prog->aux->linfo[linfo_idx]; 209 insn_start = linfo[0].insn_off; 210 insn_end = insn_start + prog->len; 211 212 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 213 jited_linfo[0] = prog->bpf_func; 214 215 nr_linfo = prog->aux->nr_linfo - linfo_idx; 216 217 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 218 /* The verifier ensures that linfo[i].insn_off is 219 * strictly increasing 220 */ 221 jited_linfo[i] = prog->bpf_func + 222 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 223 } 224 225 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 226 gfp_t gfp_extra_flags) 227 { 228 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags; 229 struct bpf_prog *fp; 230 u32 pages; 231 232 size = round_up(size, PAGE_SIZE); 233 pages = size / PAGE_SIZE; 234 if (pages <= fp_old->pages) 235 return fp_old; 236 237 fp = __vmalloc(size, gfp_flags); 238 if (fp) { 239 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 240 fp->pages = pages; 241 fp->aux->prog = fp; 242 243 /* We keep fp->aux from fp_old around in the new 244 * reallocated structure. 245 */ 246 fp_old->aux = NULL; 247 fp_old->stats = NULL; 248 fp_old->active = NULL; 249 __bpf_prog_free(fp_old); 250 } 251 252 return fp; 253 } 254 255 void __bpf_prog_free(struct bpf_prog *fp) 256 { 257 if (fp->aux) { 258 mutex_destroy(&fp->aux->used_maps_mutex); 259 mutex_destroy(&fp->aux->dst_mutex); 260 kfree(fp->aux->poke_tab); 261 kfree(fp->aux); 262 } 263 free_percpu(fp->stats); 264 free_percpu(fp->active); 265 vfree(fp); 266 } 267 268 int bpf_prog_calc_tag(struct bpf_prog *fp) 269 { 270 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 271 u32 raw_size = bpf_prog_tag_scratch_size(fp); 272 u32 digest[SHA1_DIGEST_WORDS]; 273 u32 ws[SHA1_WORKSPACE_WORDS]; 274 u32 i, bsize, psize, blocks; 275 struct bpf_insn *dst; 276 bool was_ld_map; 277 u8 *raw, *todo; 278 __be32 *result; 279 __be64 *bits; 280 281 raw = vmalloc(raw_size); 282 if (!raw) 283 return -ENOMEM; 284 285 sha1_init(digest); 286 memset(ws, 0, sizeof(ws)); 287 288 /* We need to take out the map fd for the digest calculation 289 * since they are unstable from user space side. 290 */ 291 dst = (void *)raw; 292 for (i = 0, was_ld_map = false; i < fp->len; i++) { 293 dst[i] = fp->insnsi[i]; 294 if (!was_ld_map && 295 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 296 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 297 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 298 was_ld_map = true; 299 dst[i].imm = 0; 300 } else if (was_ld_map && 301 dst[i].code == 0 && 302 dst[i].dst_reg == 0 && 303 dst[i].src_reg == 0 && 304 dst[i].off == 0) { 305 was_ld_map = false; 306 dst[i].imm = 0; 307 } else { 308 was_ld_map = false; 309 } 310 } 311 312 psize = bpf_prog_insn_size(fp); 313 memset(&raw[psize], 0, raw_size - psize); 314 raw[psize++] = 0x80; 315 316 bsize = round_up(psize, SHA1_BLOCK_SIZE); 317 blocks = bsize / SHA1_BLOCK_SIZE; 318 todo = raw; 319 if (bsize - psize >= sizeof(__be64)) { 320 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 321 } else { 322 bits = (__be64 *)(todo + bsize + bits_offset); 323 blocks++; 324 } 325 *bits = cpu_to_be64((psize - 1) << 3); 326 327 while (blocks--) { 328 sha1_transform(digest, todo, ws); 329 todo += SHA1_BLOCK_SIZE; 330 } 331 332 result = (__force __be32 *)digest; 333 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 334 result[i] = cpu_to_be32(digest[i]); 335 memcpy(fp->tag, result, sizeof(fp->tag)); 336 337 vfree(raw); 338 return 0; 339 } 340 341 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 342 s32 end_new, s32 curr, const bool probe_pass) 343 { 344 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 345 s32 delta = end_new - end_old; 346 s64 imm = insn->imm; 347 348 if (curr < pos && curr + imm + 1 >= end_old) 349 imm += delta; 350 else if (curr >= end_new && curr + imm + 1 < end_new) 351 imm -= delta; 352 if (imm < imm_min || imm > imm_max) 353 return -ERANGE; 354 if (!probe_pass) 355 insn->imm = imm; 356 return 0; 357 } 358 359 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 360 s32 end_new, s32 curr, const bool probe_pass) 361 { 362 const s32 off_min = S16_MIN, off_max = S16_MAX; 363 s32 delta = end_new - end_old; 364 s32 off = insn->off; 365 366 if (curr < pos && curr + off + 1 >= end_old) 367 off += delta; 368 else if (curr >= end_new && curr + off + 1 < end_new) 369 off -= delta; 370 if (off < off_min || off > off_max) 371 return -ERANGE; 372 if (!probe_pass) 373 insn->off = off; 374 return 0; 375 } 376 377 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 378 s32 end_new, const bool probe_pass) 379 { 380 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 381 struct bpf_insn *insn = prog->insnsi; 382 int ret = 0; 383 384 for (i = 0; i < insn_cnt; i++, insn++) { 385 u8 code; 386 387 /* In the probing pass we still operate on the original, 388 * unpatched image in order to check overflows before we 389 * do any other adjustments. Therefore skip the patchlet. 390 */ 391 if (probe_pass && i == pos) { 392 i = end_new; 393 insn = prog->insnsi + end_old; 394 } 395 if (bpf_pseudo_func(insn)) { 396 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 397 end_new, i, probe_pass); 398 if (ret) 399 return ret; 400 continue; 401 } 402 code = insn->code; 403 if ((BPF_CLASS(code) != BPF_JMP && 404 BPF_CLASS(code) != BPF_JMP32) || 405 BPF_OP(code) == BPF_EXIT) 406 continue; 407 /* Adjust offset of jmps if we cross patch boundaries. */ 408 if (BPF_OP(code) == BPF_CALL) { 409 if (insn->src_reg != BPF_PSEUDO_CALL) 410 continue; 411 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 412 end_new, i, probe_pass); 413 } else { 414 ret = bpf_adj_delta_to_off(insn, pos, end_old, 415 end_new, i, probe_pass); 416 } 417 if (ret) 418 break; 419 } 420 421 return ret; 422 } 423 424 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 425 { 426 struct bpf_line_info *linfo; 427 u32 i, nr_linfo; 428 429 nr_linfo = prog->aux->nr_linfo; 430 if (!nr_linfo || !delta) 431 return; 432 433 linfo = prog->aux->linfo; 434 435 for (i = 0; i < nr_linfo; i++) 436 if (off < linfo[i].insn_off) 437 break; 438 439 /* Push all off < linfo[i].insn_off by delta */ 440 for (; i < nr_linfo; i++) 441 linfo[i].insn_off += delta; 442 } 443 444 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 445 const struct bpf_insn *patch, u32 len) 446 { 447 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 448 const u32 cnt_max = S16_MAX; 449 struct bpf_prog *prog_adj; 450 int err; 451 452 /* Since our patchlet doesn't expand the image, we're done. */ 453 if (insn_delta == 0) { 454 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 455 return prog; 456 } 457 458 insn_adj_cnt = prog->len + insn_delta; 459 460 /* Reject anything that would potentially let the insn->off 461 * target overflow when we have excessive program expansions. 462 * We need to probe here before we do any reallocation where 463 * we afterwards may not fail anymore. 464 */ 465 if (insn_adj_cnt > cnt_max && 466 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 467 return ERR_PTR(err); 468 469 /* Several new instructions need to be inserted. Make room 470 * for them. Likely, there's no need for a new allocation as 471 * last page could have large enough tailroom. 472 */ 473 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 474 GFP_USER); 475 if (!prog_adj) 476 return ERR_PTR(-ENOMEM); 477 478 prog_adj->len = insn_adj_cnt; 479 480 /* Patching happens in 3 steps: 481 * 482 * 1) Move over tail of insnsi from next instruction onwards, 483 * so we can patch the single target insn with one or more 484 * new ones (patching is always from 1 to n insns, n > 0). 485 * 2) Inject new instructions at the target location. 486 * 3) Adjust branch offsets if necessary. 487 */ 488 insn_rest = insn_adj_cnt - off - len; 489 490 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 491 sizeof(*patch) * insn_rest); 492 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 493 494 /* We are guaranteed to not fail at this point, otherwise 495 * the ship has sailed to reverse to the original state. An 496 * overflow cannot happen at this point. 497 */ 498 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 499 500 bpf_adj_linfo(prog_adj, off, insn_delta); 501 502 return prog_adj; 503 } 504 505 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 506 { 507 /* Branch offsets can't overflow when program is shrinking, no need 508 * to call bpf_adj_branches(..., true) here 509 */ 510 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 511 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 512 prog->len -= cnt; 513 514 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 515 } 516 517 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 518 { 519 int i; 520 521 for (i = 0; i < fp->aux->func_cnt; i++) 522 bpf_prog_kallsyms_del(fp->aux->func[i]); 523 } 524 525 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 526 { 527 bpf_prog_kallsyms_del_subprogs(fp); 528 bpf_prog_kallsyms_del(fp); 529 } 530 531 #ifdef CONFIG_BPF_JIT 532 /* All BPF JIT sysctl knobs here. */ 533 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 534 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 535 int bpf_jit_harden __read_mostly; 536 long bpf_jit_limit __read_mostly; 537 long bpf_jit_limit_max __read_mostly; 538 539 static void 540 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 541 { 542 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 543 544 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 545 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 546 } 547 548 static void 549 bpf_prog_ksym_set_name(struct bpf_prog *prog) 550 { 551 char *sym = prog->aux->ksym.name; 552 const char *end = sym + KSYM_NAME_LEN; 553 const struct btf_type *type; 554 const char *func_name; 555 556 BUILD_BUG_ON(sizeof("bpf_prog_") + 557 sizeof(prog->tag) * 2 + 558 /* name has been null terminated. 559 * We should need +1 for the '_' preceding 560 * the name. However, the null character 561 * is double counted between the name and the 562 * sizeof("bpf_prog_") above, so we omit 563 * the +1 here. 564 */ 565 sizeof(prog->aux->name) > KSYM_NAME_LEN); 566 567 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 568 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 569 570 /* prog->aux->name will be ignored if full btf name is available */ 571 if (prog->aux->func_info_cnt) { 572 type = btf_type_by_id(prog->aux->btf, 573 prog->aux->func_info[prog->aux->func_idx].type_id); 574 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 575 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 576 return; 577 } 578 579 if (prog->aux->name[0]) 580 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 581 else 582 *sym = 0; 583 } 584 585 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 586 { 587 return container_of(n, struct bpf_ksym, tnode)->start; 588 } 589 590 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 591 struct latch_tree_node *b) 592 { 593 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 594 } 595 596 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 597 { 598 unsigned long val = (unsigned long)key; 599 const struct bpf_ksym *ksym; 600 601 ksym = container_of(n, struct bpf_ksym, tnode); 602 603 if (val < ksym->start) 604 return -1; 605 if (val >= ksym->end) 606 return 1; 607 608 return 0; 609 } 610 611 static const struct latch_tree_ops bpf_tree_ops = { 612 .less = bpf_tree_less, 613 .comp = bpf_tree_comp, 614 }; 615 616 static DEFINE_SPINLOCK(bpf_lock); 617 static LIST_HEAD(bpf_kallsyms); 618 static struct latch_tree_root bpf_tree __cacheline_aligned; 619 620 void bpf_ksym_add(struct bpf_ksym *ksym) 621 { 622 spin_lock_bh(&bpf_lock); 623 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 624 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 625 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 626 spin_unlock_bh(&bpf_lock); 627 } 628 629 static void __bpf_ksym_del(struct bpf_ksym *ksym) 630 { 631 if (list_empty(&ksym->lnode)) 632 return; 633 634 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 635 list_del_rcu(&ksym->lnode); 636 } 637 638 void bpf_ksym_del(struct bpf_ksym *ksym) 639 { 640 spin_lock_bh(&bpf_lock); 641 __bpf_ksym_del(ksym); 642 spin_unlock_bh(&bpf_lock); 643 } 644 645 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 646 { 647 return fp->jited && !bpf_prog_was_classic(fp); 648 } 649 650 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 651 { 652 return list_empty(&fp->aux->ksym.lnode) || 653 fp->aux->ksym.lnode.prev == LIST_POISON2; 654 } 655 656 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 657 { 658 if (!bpf_prog_kallsyms_candidate(fp) || 659 !bpf_capable()) 660 return; 661 662 bpf_prog_ksym_set_addr(fp); 663 bpf_prog_ksym_set_name(fp); 664 fp->aux->ksym.prog = true; 665 666 bpf_ksym_add(&fp->aux->ksym); 667 } 668 669 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 670 { 671 if (!bpf_prog_kallsyms_candidate(fp)) 672 return; 673 674 bpf_ksym_del(&fp->aux->ksym); 675 } 676 677 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 678 { 679 struct latch_tree_node *n; 680 681 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 682 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 683 } 684 685 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 686 unsigned long *off, char *sym) 687 { 688 struct bpf_ksym *ksym; 689 char *ret = NULL; 690 691 rcu_read_lock(); 692 ksym = bpf_ksym_find(addr); 693 if (ksym) { 694 unsigned long symbol_start = ksym->start; 695 unsigned long symbol_end = ksym->end; 696 697 strncpy(sym, ksym->name, KSYM_NAME_LEN); 698 699 ret = sym; 700 if (size) 701 *size = symbol_end - symbol_start; 702 if (off) 703 *off = addr - symbol_start; 704 } 705 rcu_read_unlock(); 706 707 return ret; 708 } 709 710 bool is_bpf_text_address(unsigned long addr) 711 { 712 bool ret; 713 714 rcu_read_lock(); 715 ret = bpf_ksym_find(addr) != NULL; 716 rcu_read_unlock(); 717 718 return ret; 719 } 720 721 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 722 { 723 struct bpf_ksym *ksym = bpf_ksym_find(addr); 724 725 return ksym && ksym->prog ? 726 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 727 NULL; 728 } 729 730 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 731 { 732 const struct exception_table_entry *e = NULL; 733 struct bpf_prog *prog; 734 735 rcu_read_lock(); 736 prog = bpf_prog_ksym_find(addr); 737 if (!prog) 738 goto out; 739 if (!prog->aux->num_exentries) 740 goto out; 741 742 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 743 out: 744 rcu_read_unlock(); 745 return e; 746 } 747 748 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 749 char *sym) 750 { 751 struct bpf_ksym *ksym; 752 unsigned int it = 0; 753 int ret = -ERANGE; 754 755 if (!bpf_jit_kallsyms_enabled()) 756 return ret; 757 758 rcu_read_lock(); 759 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 760 if (it++ != symnum) 761 continue; 762 763 strncpy(sym, ksym->name, KSYM_NAME_LEN); 764 765 *value = ksym->start; 766 *type = BPF_SYM_ELF_TYPE; 767 768 ret = 0; 769 break; 770 } 771 rcu_read_unlock(); 772 773 return ret; 774 } 775 776 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 777 struct bpf_jit_poke_descriptor *poke) 778 { 779 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 780 static const u32 poke_tab_max = 1024; 781 u32 slot = prog->aux->size_poke_tab; 782 u32 size = slot + 1; 783 784 if (size > poke_tab_max) 785 return -ENOSPC; 786 if (poke->tailcall_target || poke->tailcall_target_stable || 787 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 788 return -EINVAL; 789 790 switch (poke->reason) { 791 case BPF_POKE_REASON_TAIL_CALL: 792 if (!poke->tail_call.map) 793 return -EINVAL; 794 break; 795 default: 796 return -EINVAL; 797 } 798 799 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL); 800 if (!tab) 801 return -ENOMEM; 802 803 memcpy(&tab[slot], poke, sizeof(*poke)); 804 prog->aux->size_poke_tab = size; 805 prog->aux->poke_tab = tab; 806 807 return slot; 808 } 809 810 /* 811 * BPF program pack allocator. 812 * 813 * Most BPF programs are pretty small. Allocating a hole page for each 814 * program is sometime a waste. Many small bpf program also adds pressure 815 * to instruction TLB. To solve this issue, we introduce a BPF program pack 816 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 817 * to host BPF programs. 818 */ 819 #define BPF_PROG_CHUNK_SHIFT 6 820 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 821 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 822 823 struct bpf_prog_pack { 824 struct list_head list; 825 void *ptr; 826 unsigned long bitmap[]; 827 }; 828 829 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 830 831 static size_t bpf_prog_pack_size = -1; 832 833 static int bpf_prog_chunk_count(void) 834 { 835 WARN_ON_ONCE(bpf_prog_pack_size == -1); 836 return bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE; 837 } 838 839 static DEFINE_MUTEX(pack_mutex); 840 static LIST_HEAD(pack_list); 841 842 static size_t select_bpf_prog_pack_size(void) 843 { 844 size_t size; 845 void *ptr; 846 847 size = PMD_SIZE * num_online_nodes(); 848 ptr = module_alloc(size); 849 850 /* Test whether we can get huge pages. If not just use PAGE_SIZE 851 * packs. 852 */ 853 if (!ptr || !is_vm_area_hugepages(ptr)) 854 size = PAGE_SIZE; 855 856 vfree(ptr); 857 return size; 858 } 859 860 static struct bpf_prog_pack *alloc_new_pack(void) 861 { 862 struct bpf_prog_pack *pack; 863 864 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(bpf_prog_chunk_count())), 865 GFP_KERNEL); 866 if (!pack) 867 return NULL; 868 pack->ptr = module_alloc(bpf_prog_pack_size); 869 if (!pack->ptr) { 870 kfree(pack); 871 return NULL; 872 } 873 bitmap_zero(pack->bitmap, bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE); 874 list_add_tail(&pack->list, &pack_list); 875 876 set_vm_flush_reset_perms(pack->ptr); 877 set_memory_ro((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE); 878 set_memory_x((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE); 879 return pack; 880 } 881 882 static void *bpf_prog_pack_alloc(u32 size) 883 { 884 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 885 struct bpf_prog_pack *pack; 886 unsigned long pos; 887 void *ptr = NULL; 888 889 mutex_lock(&pack_mutex); 890 if (bpf_prog_pack_size == -1) 891 bpf_prog_pack_size = select_bpf_prog_pack_size(); 892 893 if (size > bpf_prog_pack_size) { 894 size = round_up(size, PAGE_SIZE); 895 ptr = module_alloc(size); 896 if (ptr) { 897 set_vm_flush_reset_perms(ptr); 898 set_memory_ro((unsigned long)ptr, size / PAGE_SIZE); 899 set_memory_x((unsigned long)ptr, size / PAGE_SIZE); 900 } 901 goto out; 902 } 903 list_for_each_entry(pack, &pack_list, list) { 904 pos = bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0, 905 nbits, 0); 906 if (pos < bpf_prog_chunk_count()) 907 goto found_free_area; 908 } 909 910 pack = alloc_new_pack(); 911 if (!pack) 912 goto out; 913 914 pos = 0; 915 916 found_free_area: 917 bitmap_set(pack->bitmap, pos, nbits); 918 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 919 920 out: 921 mutex_unlock(&pack_mutex); 922 return ptr; 923 } 924 925 static void bpf_prog_pack_free(struct bpf_binary_header *hdr) 926 { 927 struct bpf_prog_pack *pack = NULL, *tmp; 928 unsigned int nbits; 929 unsigned long pos; 930 void *pack_ptr; 931 932 mutex_lock(&pack_mutex); 933 if (hdr->size > bpf_prog_pack_size) { 934 module_memfree(hdr); 935 goto out; 936 } 937 938 pack_ptr = (void *)((unsigned long)hdr & ~(bpf_prog_pack_size - 1)); 939 940 list_for_each_entry(tmp, &pack_list, list) { 941 if (tmp->ptr == pack_ptr) { 942 pack = tmp; 943 break; 944 } 945 } 946 947 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 948 goto out; 949 950 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size); 951 pos = ((unsigned long)hdr - (unsigned long)pack_ptr) >> BPF_PROG_CHUNK_SHIFT; 952 953 bitmap_clear(pack->bitmap, pos, nbits); 954 if (bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0, 955 bpf_prog_chunk_count(), 0) == 0) { 956 list_del(&pack->list); 957 module_memfree(pack->ptr); 958 kfree(pack); 959 } 960 out: 961 mutex_unlock(&pack_mutex); 962 } 963 964 static atomic_long_t bpf_jit_current; 965 966 /* Can be overridden by an arch's JIT compiler if it has a custom, 967 * dedicated BPF backend memory area, or if neither of the two 968 * below apply. 969 */ 970 u64 __weak bpf_jit_alloc_exec_limit(void) 971 { 972 #if defined(MODULES_VADDR) 973 return MODULES_END - MODULES_VADDR; 974 #else 975 return VMALLOC_END - VMALLOC_START; 976 #endif 977 } 978 979 static int __init bpf_jit_charge_init(void) 980 { 981 /* Only used as heuristic here to derive limit. */ 982 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 983 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2, 984 PAGE_SIZE), LONG_MAX); 985 return 0; 986 } 987 pure_initcall(bpf_jit_charge_init); 988 989 int bpf_jit_charge_modmem(u32 size) 990 { 991 if (atomic_long_add_return(size, &bpf_jit_current) > bpf_jit_limit) { 992 if (!bpf_capable()) { 993 atomic_long_sub(size, &bpf_jit_current); 994 return -EPERM; 995 } 996 } 997 998 return 0; 999 } 1000 1001 void bpf_jit_uncharge_modmem(u32 size) 1002 { 1003 atomic_long_sub(size, &bpf_jit_current); 1004 } 1005 1006 void *__weak bpf_jit_alloc_exec(unsigned long size) 1007 { 1008 return module_alloc(size); 1009 } 1010 1011 void __weak bpf_jit_free_exec(void *addr) 1012 { 1013 module_memfree(addr); 1014 } 1015 1016 struct bpf_binary_header * 1017 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1018 unsigned int alignment, 1019 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1020 { 1021 struct bpf_binary_header *hdr; 1022 u32 size, hole, start; 1023 1024 WARN_ON_ONCE(!is_power_of_2(alignment) || 1025 alignment > BPF_IMAGE_ALIGNMENT); 1026 1027 /* Most of BPF filters are really small, but if some of them 1028 * fill a page, allow at least 128 extra bytes to insert a 1029 * random section of illegal instructions. 1030 */ 1031 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1032 1033 if (bpf_jit_charge_modmem(size)) 1034 return NULL; 1035 hdr = bpf_jit_alloc_exec(size); 1036 if (!hdr) { 1037 bpf_jit_uncharge_modmem(size); 1038 return NULL; 1039 } 1040 1041 /* Fill space with illegal/arch-dep instructions. */ 1042 bpf_fill_ill_insns(hdr, size); 1043 1044 hdr->size = size; 1045 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1046 PAGE_SIZE - sizeof(*hdr)); 1047 start = (get_random_int() % hole) & ~(alignment - 1); 1048 1049 /* Leave a random number of instructions before BPF code. */ 1050 *image_ptr = &hdr->image[start]; 1051 1052 return hdr; 1053 } 1054 1055 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1056 { 1057 u32 size = hdr->size; 1058 1059 bpf_jit_free_exec(hdr); 1060 bpf_jit_uncharge_modmem(size); 1061 } 1062 1063 /* Allocate jit binary from bpf_prog_pack allocator. 1064 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1065 * to the memory. To solve this problem, a RW buffer is also allocated at 1066 * as the same time. The JIT engine should calculate offsets based on the 1067 * RO memory address, but write JITed program to the RW buffer. Once the 1068 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1069 * the JITed program to the RO memory. 1070 */ 1071 struct bpf_binary_header * 1072 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1073 unsigned int alignment, 1074 struct bpf_binary_header **rw_header, 1075 u8 **rw_image, 1076 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1077 { 1078 struct bpf_binary_header *ro_header; 1079 u32 size, hole, start; 1080 1081 WARN_ON_ONCE(!is_power_of_2(alignment) || 1082 alignment > BPF_IMAGE_ALIGNMENT); 1083 1084 /* add 16 bytes for a random section of illegal instructions */ 1085 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1086 1087 if (bpf_jit_charge_modmem(size)) 1088 return NULL; 1089 ro_header = bpf_prog_pack_alloc(size); 1090 if (!ro_header) { 1091 bpf_jit_uncharge_modmem(size); 1092 return NULL; 1093 } 1094 1095 *rw_header = kvmalloc(size, GFP_KERNEL); 1096 if (!*rw_header) { 1097 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size)); 1098 bpf_prog_pack_free(ro_header); 1099 bpf_jit_uncharge_modmem(size); 1100 return NULL; 1101 } 1102 1103 /* Fill space with illegal/arch-dep instructions. */ 1104 bpf_fill_ill_insns(*rw_header, size); 1105 (*rw_header)->size = size; 1106 1107 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1108 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1109 start = (get_random_int() % hole) & ~(alignment - 1); 1110 1111 *image_ptr = &ro_header->image[start]; 1112 *rw_image = &(*rw_header)->image[start]; 1113 1114 return ro_header; 1115 } 1116 1117 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1118 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog, 1119 struct bpf_binary_header *ro_header, 1120 struct bpf_binary_header *rw_header) 1121 { 1122 void *ptr; 1123 1124 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1125 1126 kvfree(rw_header); 1127 1128 if (IS_ERR(ptr)) { 1129 bpf_prog_pack_free(ro_header); 1130 return PTR_ERR(ptr); 1131 } 1132 prog->aux->use_bpf_prog_pack = true; 1133 return 0; 1134 } 1135 1136 /* bpf_jit_binary_pack_free is called in two different scenarios: 1137 * 1) when the program is freed after; 1138 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1139 * For case 2), we need to free both the RO memory and the RW buffer. 1140 * 1141 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1142 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1143 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1144 * bpf_arch_text_copy (when jit fails). 1145 */ 1146 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1147 struct bpf_binary_header *rw_header) 1148 { 1149 u32 size = ro_header->size; 1150 1151 bpf_prog_pack_free(ro_header); 1152 kvfree(rw_header); 1153 bpf_jit_uncharge_modmem(size); 1154 } 1155 1156 static inline struct bpf_binary_header * 1157 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1158 { 1159 unsigned long real_start = (unsigned long)fp->bpf_func; 1160 unsigned long addr; 1161 1162 if (fp->aux->use_bpf_prog_pack) 1163 addr = real_start & BPF_PROG_CHUNK_MASK; 1164 else 1165 addr = real_start & PAGE_MASK; 1166 1167 return (void *)addr; 1168 } 1169 1170 /* This symbol is only overridden by archs that have different 1171 * requirements than the usual eBPF JITs, f.e. when they only 1172 * implement cBPF JIT, do not set images read-only, etc. 1173 */ 1174 void __weak bpf_jit_free(struct bpf_prog *fp) 1175 { 1176 if (fp->jited) { 1177 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1178 1179 if (fp->aux->use_bpf_prog_pack) 1180 bpf_jit_binary_pack_free(hdr, NULL /* rw_buffer */); 1181 else 1182 bpf_jit_binary_free(hdr); 1183 1184 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1185 } 1186 1187 bpf_prog_unlock_free(fp); 1188 } 1189 1190 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1191 const struct bpf_insn *insn, bool extra_pass, 1192 u64 *func_addr, bool *func_addr_fixed) 1193 { 1194 s16 off = insn->off; 1195 s32 imm = insn->imm; 1196 u8 *addr; 1197 1198 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1199 if (!*func_addr_fixed) { 1200 /* Place-holder address till the last pass has collected 1201 * all addresses for JITed subprograms in which case we 1202 * can pick them up from prog->aux. 1203 */ 1204 if (!extra_pass) 1205 addr = NULL; 1206 else if (prog->aux->func && 1207 off >= 0 && off < prog->aux->func_cnt) 1208 addr = (u8 *)prog->aux->func[off]->bpf_func; 1209 else 1210 return -EINVAL; 1211 } else { 1212 /* Address of a BPF helper call. Since part of the core 1213 * kernel, it's always at a fixed location. __bpf_call_base 1214 * and the helper with imm relative to it are both in core 1215 * kernel. 1216 */ 1217 addr = (u8 *)__bpf_call_base + imm; 1218 } 1219 1220 *func_addr = (unsigned long)addr; 1221 return 0; 1222 } 1223 1224 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1225 const struct bpf_insn *aux, 1226 struct bpf_insn *to_buff, 1227 bool emit_zext) 1228 { 1229 struct bpf_insn *to = to_buff; 1230 u32 imm_rnd = get_random_int(); 1231 s16 off; 1232 1233 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1234 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1235 1236 /* Constraints on AX register: 1237 * 1238 * AX register is inaccessible from user space. It is mapped in 1239 * all JITs, and used here for constant blinding rewrites. It is 1240 * typically "stateless" meaning its contents are only valid within 1241 * the executed instruction, but not across several instructions. 1242 * There are a few exceptions however which are further detailed 1243 * below. 1244 * 1245 * Constant blinding is only used by JITs, not in the interpreter. 1246 * The interpreter uses AX in some occasions as a local temporary 1247 * register e.g. in DIV or MOD instructions. 1248 * 1249 * In restricted circumstances, the verifier can also use the AX 1250 * register for rewrites as long as they do not interfere with 1251 * the above cases! 1252 */ 1253 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1254 goto out; 1255 1256 if (from->imm == 0 && 1257 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1258 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1259 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1260 goto out; 1261 } 1262 1263 switch (from->code) { 1264 case BPF_ALU | BPF_ADD | BPF_K: 1265 case BPF_ALU | BPF_SUB | BPF_K: 1266 case BPF_ALU | BPF_AND | BPF_K: 1267 case BPF_ALU | BPF_OR | BPF_K: 1268 case BPF_ALU | BPF_XOR | BPF_K: 1269 case BPF_ALU | BPF_MUL | BPF_K: 1270 case BPF_ALU | BPF_MOV | BPF_K: 1271 case BPF_ALU | BPF_DIV | BPF_K: 1272 case BPF_ALU | BPF_MOD | BPF_K: 1273 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1274 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1275 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 1276 break; 1277 1278 case BPF_ALU64 | BPF_ADD | BPF_K: 1279 case BPF_ALU64 | BPF_SUB | BPF_K: 1280 case BPF_ALU64 | BPF_AND | BPF_K: 1281 case BPF_ALU64 | BPF_OR | BPF_K: 1282 case BPF_ALU64 | BPF_XOR | BPF_K: 1283 case BPF_ALU64 | BPF_MUL | BPF_K: 1284 case BPF_ALU64 | BPF_MOV | BPF_K: 1285 case BPF_ALU64 | BPF_DIV | BPF_K: 1286 case BPF_ALU64 | BPF_MOD | BPF_K: 1287 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1288 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1289 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 1290 break; 1291 1292 case BPF_JMP | BPF_JEQ | BPF_K: 1293 case BPF_JMP | BPF_JNE | BPF_K: 1294 case BPF_JMP | BPF_JGT | BPF_K: 1295 case BPF_JMP | BPF_JLT | BPF_K: 1296 case BPF_JMP | BPF_JGE | BPF_K: 1297 case BPF_JMP | BPF_JLE | BPF_K: 1298 case BPF_JMP | BPF_JSGT | BPF_K: 1299 case BPF_JMP | BPF_JSLT | BPF_K: 1300 case BPF_JMP | BPF_JSGE | BPF_K: 1301 case BPF_JMP | BPF_JSLE | BPF_K: 1302 case BPF_JMP | BPF_JSET | BPF_K: 1303 /* Accommodate for extra offset in case of a backjump. */ 1304 off = from->off; 1305 if (off < 0) 1306 off -= 2; 1307 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1308 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1309 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1310 break; 1311 1312 case BPF_JMP32 | BPF_JEQ | BPF_K: 1313 case BPF_JMP32 | BPF_JNE | BPF_K: 1314 case BPF_JMP32 | BPF_JGT | BPF_K: 1315 case BPF_JMP32 | BPF_JLT | BPF_K: 1316 case BPF_JMP32 | BPF_JGE | BPF_K: 1317 case BPF_JMP32 | BPF_JLE | BPF_K: 1318 case BPF_JMP32 | BPF_JSGT | BPF_K: 1319 case BPF_JMP32 | BPF_JSLT | BPF_K: 1320 case BPF_JMP32 | BPF_JSGE | BPF_K: 1321 case BPF_JMP32 | BPF_JSLE | BPF_K: 1322 case BPF_JMP32 | BPF_JSET | BPF_K: 1323 /* Accommodate for extra offset in case of a backjump. */ 1324 off = from->off; 1325 if (off < 0) 1326 off -= 2; 1327 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1328 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1329 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1330 off); 1331 break; 1332 1333 case BPF_LD | BPF_IMM | BPF_DW: 1334 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1335 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1336 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1337 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1338 break; 1339 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1340 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1341 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1342 if (emit_zext) 1343 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1344 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1345 break; 1346 1347 case BPF_ST | BPF_MEM | BPF_DW: 1348 case BPF_ST | BPF_MEM | BPF_W: 1349 case BPF_ST | BPF_MEM | BPF_H: 1350 case BPF_ST | BPF_MEM | BPF_B: 1351 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1352 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1353 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1354 break; 1355 } 1356 out: 1357 return to - to_buff; 1358 } 1359 1360 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1361 gfp_t gfp_extra_flags) 1362 { 1363 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1364 struct bpf_prog *fp; 1365 1366 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1367 if (fp != NULL) { 1368 /* aux->prog still points to the fp_other one, so 1369 * when promoting the clone to the real program, 1370 * this still needs to be adapted. 1371 */ 1372 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1373 } 1374 1375 return fp; 1376 } 1377 1378 static void bpf_prog_clone_free(struct bpf_prog *fp) 1379 { 1380 /* aux was stolen by the other clone, so we cannot free 1381 * it from this path! It will be freed eventually by the 1382 * other program on release. 1383 * 1384 * At this point, we don't need a deferred release since 1385 * clone is guaranteed to not be locked. 1386 */ 1387 fp->aux = NULL; 1388 fp->stats = NULL; 1389 fp->active = NULL; 1390 __bpf_prog_free(fp); 1391 } 1392 1393 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1394 { 1395 /* We have to repoint aux->prog to self, as we don't 1396 * know whether fp here is the clone or the original. 1397 */ 1398 fp->aux->prog = fp; 1399 bpf_prog_clone_free(fp_other); 1400 } 1401 1402 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1403 { 1404 struct bpf_insn insn_buff[16], aux[2]; 1405 struct bpf_prog *clone, *tmp; 1406 int insn_delta, insn_cnt; 1407 struct bpf_insn *insn; 1408 int i, rewritten; 1409 1410 if (!prog->blinding_requested || prog->blinded) 1411 return prog; 1412 1413 clone = bpf_prog_clone_create(prog, GFP_USER); 1414 if (!clone) 1415 return ERR_PTR(-ENOMEM); 1416 1417 insn_cnt = clone->len; 1418 insn = clone->insnsi; 1419 1420 for (i = 0; i < insn_cnt; i++, insn++) { 1421 /* We temporarily need to hold the original ld64 insn 1422 * so that we can still access the first part in the 1423 * second blinding run. 1424 */ 1425 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1426 insn[1].code == 0) 1427 memcpy(aux, insn, sizeof(aux)); 1428 1429 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1430 clone->aux->verifier_zext); 1431 if (!rewritten) 1432 continue; 1433 1434 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1435 if (IS_ERR(tmp)) { 1436 /* Patching may have repointed aux->prog during 1437 * realloc from the original one, so we need to 1438 * fix it up here on error. 1439 */ 1440 bpf_jit_prog_release_other(prog, clone); 1441 return tmp; 1442 } 1443 1444 clone = tmp; 1445 insn_delta = rewritten - 1; 1446 1447 /* Walk new program and skip insns we just inserted. */ 1448 insn = clone->insnsi + i + insn_delta; 1449 insn_cnt += insn_delta; 1450 i += insn_delta; 1451 } 1452 1453 clone->blinded = 1; 1454 return clone; 1455 } 1456 #endif /* CONFIG_BPF_JIT */ 1457 1458 /* Base function for offset calculation. Needs to go into .text section, 1459 * therefore keeping it non-static as well; will also be used by JITs 1460 * anyway later on, so do not let the compiler omit it. This also needs 1461 * to go into kallsyms for correlation from e.g. bpftool, so naming 1462 * must not change. 1463 */ 1464 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1465 { 1466 return 0; 1467 } 1468 EXPORT_SYMBOL_GPL(__bpf_call_base); 1469 1470 /* All UAPI available opcodes. */ 1471 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1472 /* 32 bit ALU operations. */ \ 1473 /* Register based. */ \ 1474 INSN_3(ALU, ADD, X), \ 1475 INSN_3(ALU, SUB, X), \ 1476 INSN_3(ALU, AND, X), \ 1477 INSN_3(ALU, OR, X), \ 1478 INSN_3(ALU, LSH, X), \ 1479 INSN_3(ALU, RSH, X), \ 1480 INSN_3(ALU, XOR, X), \ 1481 INSN_3(ALU, MUL, X), \ 1482 INSN_3(ALU, MOV, X), \ 1483 INSN_3(ALU, ARSH, X), \ 1484 INSN_3(ALU, DIV, X), \ 1485 INSN_3(ALU, MOD, X), \ 1486 INSN_2(ALU, NEG), \ 1487 INSN_3(ALU, END, TO_BE), \ 1488 INSN_3(ALU, END, TO_LE), \ 1489 /* Immediate based. */ \ 1490 INSN_3(ALU, ADD, K), \ 1491 INSN_3(ALU, SUB, K), \ 1492 INSN_3(ALU, AND, K), \ 1493 INSN_3(ALU, OR, K), \ 1494 INSN_3(ALU, LSH, K), \ 1495 INSN_3(ALU, RSH, K), \ 1496 INSN_3(ALU, XOR, K), \ 1497 INSN_3(ALU, MUL, K), \ 1498 INSN_3(ALU, MOV, K), \ 1499 INSN_3(ALU, ARSH, K), \ 1500 INSN_3(ALU, DIV, K), \ 1501 INSN_3(ALU, MOD, K), \ 1502 /* 64 bit ALU operations. */ \ 1503 /* Register based. */ \ 1504 INSN_3(ALU64, ADD, X), \ 1505 INSN_3(ALU64, SUB, X), \ 1506 INSN_3(ALU64, AND, X), \ 1507 INSN_3(ALU64, OR, X), \ 1508 INSN_3(ALU64, LSH, X), \ 1509 INSN_3(ALU64, RSH, X), \ 1510 INSN_3(ALU64, XOR, X), \ 1511 INSN_3(ALU64, MUL, X), \ 1512 INSN_3(ALU64, MOV, X), \ 1513 INSN_3(ALU64, ARSH, X), \ 1514 INSN_3(ALU64, DIV, X), \ 1515 INSN_3(ALU64, MOD, X), \ 1516 INSN_2(ALU64, NEG), \ 1517 /* Immediate based. */ \ 1518 INSN_3(ALU64, ADD, K), \ 1519 INSN_3(ALU64, SUB, K), \ 1520 INSN_3(ALU64, AND, K), \ 1521 INSN_3(ALU64, OR, K), \ 1522 INSN_3(ALU64, LSH, K), \ 1523 INSN_3(ALU64, RSH, K), \ 1524 INSN_3(ALU64, XOR, K), \ 1525 INSN_3(ALU64, MUL, K), \ 1526 INSN_3(ALU64, MOV, K), \ 1527 INSN_3(ALU64, ARSH, K), \ 1528 INSN_3(ALU64, DIV, K), \ 1529 INSN_3(ALU64, MOD, K), \ 1530 /* Call instruction. */ \ 1531 INSN_2(JMP, CALL), \ 1532 /* Exit instruction. */ \ 1533 INSN_2(JMP, EXIT), \ 1534 /* 32-bit Jump instructions. */ \ 1535 /* Register based. */ \ 1536 INSN_3(JMP32, JEQ, X), \ 1537 INSN_3(JMP32, JNE, X), \ 1538 INSN_3(JMP32, JGT, X), \ 1539 INSN_3(JMP32, JLT, X), \ 1540 INSN_3(JMP32, JGE, X), \ 1541 INSN_3(JMP32, JLE, X), \ 1542 INSN_3(JMP32, JSGT, X), \ 1543 INSN_3(JMP32, JSLT, X), \ 1544 INSN_3(JMP32, JSGE, X), \ 1545 INSN_3(JMP32, JSLE, X), \ 1546 INSN_3(JMP32, JSET, X), \ 1547 /* Immediate based. */ \ 1548 INSN_3(JMP32, JEQ, K), \ 1549 INSN_3(JMP32, JNE, K), \ 1550 INSN_3(JMP32, JGT, K), \ 1551 INSN_3(JMP32, JLT, K), \ 1552 INSN_3(JMP32, JGE, K), \ 1553 INSN_3(JMP32, JLE, K), \ 1554 INSN_3(JMP32, JSGT, K), \ 1555 INSN_3(JMP32, JSLT, K), \ 1556 INSN_3(JMP32, JSGE, K), \ 1557 INSN_3(JMP32, JSLE, K), \ 1558 INSN_3(JMP32, JSET, K), \ 1559 /* Jump instructions. */ \ 1560 /* Register based. */ \ 1561 INSN_3(JMP, JEQ, X), \ 1562 INSN_3(JMP, JNE, X), \ 1563 INSN_3(JMP, JGT, X), \ 1564 INSN_3(JMP, JLT, X), \ 1565 INSN_3(JMP, JGE, X), \ 1566 INSN_3(JMP, JLE, X), \ 1567 INSN_3(JMP, JSGT, X), \ 1568 INSN_3(JMP, JSLT, X), \ 1569 INSN_3(JMP, JSGE, X), \ 1570 INSN_3(JMP, JSLE, X), \ 1571 INSN_3(JMP, JSET, X), \ 1572 /* Immediate based. */ \ 1573 INSN_3(JMP, JEQ, K), \ 1574 INSN_3(JMP, JNE, K), \ 1575 INSN_3(JMP, JGT, K), \ 1576 INSN_3(JMP, JLT, K), \ 1577 INSN_3(JMP, JGE, K), \ 1578 INSN_3(JMP, JLE, K), \ 1579 INSN_3(JMP, JSGT, K), \ 1580 INSN_3(JMP, JSLT, K), \ 1581 INSN_3(JMP, JSGE, K), \ 1582 INSN_3(JMP, JSLE, K), \ 1583 INSN_3(JMP, JSET, K), \ 1584 INSN_2(JMP, JA), \ 1585 /* Store instructions. */ \ 1586 /* Register based. */ \ 1587 INSN_3(STX, MEM, B), \ 1588 INSN_3(STX, MEM, H), \ 1589 INSN_3(STX, MEM, W), \ 1590 INSN_3(STX, MEM, DW), \ 1591 INSN_3(STX, ATOMIC, W), \ 1592 INSN_3(STX, ATOMIC, DW), \ 1593 /* Immediate based. */ \ 1594 INSN_3(ST, MEM, B), \ 1595 INSN_3(ST, MEM, H), \ 1596 INSN_3(ST, MEM, W), \ 1597 INSN_3(ST, MEM, DW), \ 1598 /* Load instructions. */ \ 1599 /* Register based. */ \ 1600 INSN_3(LDX, MEM, B), \ 1601 INSN_3(LDX, MEM, H), \ 1602 INSN_3(LDX, MEM, W), \ 1603 INSN_3(LDX, MEM, DW), \ 1604 /* Immediate based. */ \ 1605 INSN_3(LD, IMM, DW) 1606 1607 bool bpf_opcode_in_insntable(u8 code) 1608 { 1609 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1610 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1611 static const bool public_insntable[256] = { 1612 [0 ... 255] = false, 1613 /* Now overwrite non-defaults ... */ 1614 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1615 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1616 [BPF_LD | BPF_ABS | BPF_B] = true, 1617 [BPF_LD | BPF_ABS | BPF_H] = true, 1618 [BPF_LD | BPF_ABS | BPF_W] = true, 1619 [BPF_LD | BPF_IND | BPF_B] = true, 1620 [BPF_LD | BPF_IND | BPF_H] = true, 1621 [BPF_LD | BPF_IND | BPF_W] = true, 1622 }; 1623 #undef BPF_INSN_3_TBL 1624 #undef BPF_INSN_2_TBL 1625 return public_insntable[code]; 1626 } 1627 1628 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1629 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1630 { 1631 memset(dst, 0, size); 1632 return -EFAULT; 1633 } 1634 1635 /** 1636 * ___bpf_prog_run - run eBPF program on a given context 1637 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1638 * @insn: is the array of eBPF instructions 1639 * 1640 * Decode and execute eBPF instructions. 1641 * 1642 * Return: whatever value is in %BPF_R0 at program exit 1643 */ 1644 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1645 { 1646 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1647 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1648 static const void * const jumptable[256] __annotate_jump_table = { 1649 [0 ... 255] = &&default_label, 1650 /* Now overwrite non-defaults ... */ 1651 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1652 /* Non-UAPI available opcodes. */ 1653 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1654 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1655 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1656 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1657 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1658 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1659 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1660 }; 1661 #undef BPF_INSN_3_LBL 1662 #undef BPF_INSN_2_LBL 1663 u32 tail_call_cnt = 0; 1664 1665 #define CONT ({ insn++; goto select_insn; }) 1666 #define CONT_JMP ({ insn++; goto select_insn; }) 1667 1668 select_insn: 1669 goto *jumptable[insn->code]; 1670 1671 /* Explicitly mask the register-based shift amounts with 63 or 31 1672 * to avoid undefined behavior. Normally this won't affect the 1673 * generated code, for example, in case of native 64 bit archs such 1674 * as x86-64 or arm64, the compiler is optimizing the AND away for 1675 * the interpreter. In case of JITs, each of the JIT backends compiles 1676 * the BPF shift operations to machine instructions which produce 1677 * implementation-defined results in such a case; the resulting 1678 * contents of the register may be arbitrary, but program behaviour 1679 * as a whole remains defined. In other words, in case of JIT backends, 1680 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1681 */ 1682 /* ALU (shifts) */ 1683 #define SHT(OPCODE, OP) \ 1684 ALU64_##OPCODE##_X: \ 1685 DST = DST OP (SRC & 63); \ 1686 CONT; \ 1687 ALU_##OPCODE##_X: \ 1688 DST = (u32) DST OP ((u32) SRC & 31); \ 1689 CONT; \ 1690 ALU64_##OPCODE##_K: \ 1691 DST = DST OP IMM; \ 1692 CONT; \ 1693 ALU_##OPCODE##_K: \ 1694 DST = (u32) DST OP (u32) IMM; \ 1695 CONT; 1696 /* ALU (rest) */ 1697 #define ALU(OPCODE, OP) \ 1698 ALU64_##OPCODE##_X: \ 1699 DST = DST OP SRC; \ 1700 CONT; \ 1701 ALU_##OPCODE##_X: \ 1702 DST = (u32) DST OP (u32) SRC; \ 1703 CONT; \ 1704 ALU64_##OPCODE##_K: \ 1705 DST = DST OP IMM; \ 1706 CONT; \ 1707 ALU_##OPCODE##_K: \ 1708 DST = (u32) DST OP (u32) IMM; \ 1709 CONT; 1710 ALU(ADD, +) 1711 ALU(SUB, -) 1712 ALU(AND, &) 1713 ALU(OR, |) 1714 ALU(XOR, ^) 1715 ALU(MUL, *) 1716 SHT(LSH, <<) 1717 SHT(RSH, >>) 1718 #undef SHT 1719 #undef ALU 1720 ALU_NEG: 1721 DST = (u32) -DST; 1722 CONT; 1723 ALU64_NEG: 1724 DST = -DST; 1725 CONT; 1726 ALU_MOV_X: 1727 DST = (u32) SRC; 1728 CONT; 1729 ALU_MOV_K: 1730 DST = (u32) IMM; 1731 CONT; 1732 ALU64_MOV_X: 1733 DST = SRC; 1734 CONT; 1735 ALU64_MOV_K: 1736 DST = IMM; 1737 CONT; 1738 LD_IMM_DW: 1739 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1740 insn++; 1741 CONT; 1742 ALU_ARSH_X: 1743 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1744 CONT; 1745 ALU_ARSH_K: 1746 DST = (u64) (u32) (((s32) DST) >> IMM); 1747 CONT; 1748 ALU64_ARSH_X: 1749 (*(s64 *) &DST) >>= (SRC & 63); 1750 CONT; 1751 ALU64_ARSH_K: 1752 (*(s64 *) &DST) >>= IMM; 1753 CONT; 1754 ALU64_MOD_X: 1755 div64_u64_rem(DST, SRC, &AX); 1756 DST = AX; 1757 CONT; 1758 ALU_MOD_X: 1759 AX = (u32) DST; 1760 DST = do_div(AX, (u32) SRC); 1761 CONT; 1762 ALU64_MOD_K: 1763 div64_u64_rem(DST, IMM, &AX); 1764 DST = AX; 1765 CONT; 1766 ALU_MOD_K: 1767 AX = (u32) DST; 1768 DST = do_div(AX, (u32) IMM); 1769 CONT; 1770 ALU64_DIV_X: 1771 DST = div64_u64(DST, SRC); 1772 CONT; 1773 ALU_DIV_X: 1774 AX = (u32) DST; 1775 do_div(AX, (u32) SRC); 1776 DST = (u32) AX; 1777 CONT; 1778 ALU64_DIV_K: 1779 DST = div64_u64(DST, IMM); 1780 CONT; 1781 ALU_DIV_K: 1782 AX = (u32) DST; 1783 do_div(AX, (u32) IMM); 1784 DST = (u32) AX; 1785 CONT; 1786 ALU_END_TO_BE: 1787 switch (IMM) { 1788 case 16: 1789 DST = (__force u16) cpu_to_be16(DST); 1790 break; 1791 case 32: 1792 DST = (__force u32) cpu_to_be32(DST); 1793 break; 1794 case 64: 1795 DST = (__force u64) cpu_to_be64(DST); 1796 break; 1797 } 1798 CONT; 1799 ALU_END_TO_LE: 1800 switch (IMM) { 1801 case 16: 1802 DST = (__force u16) cpu_to_le16(DST); 1803 break; 1804 case 32: 1805 DST = (__force u32) cpu_to_le32(DST); 1806 break; 1807 case 64: 1808 DST = (__force u64) cpu_to_le64(DST); 1809 break; 1810 } 1811 CONT; 1812 1813 /* CALL */ 1814 JMP_CALL: 1815 /* Function call scratches BPF_R1-BPF_R5 registers, 1816 * preserves BPF_R6-BPF_R9, and stores return value 1817 * into BPF_R0. 1818 */ 1819 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1820 BPF_R4, BPF_R5); 1821 CONT; 1822 1823 JMP_CALL_ARGS: 1824 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1825 BPF_R3, BPF_R4, 1826 BPF_R5, 1827 insn + insn->off + 1); 1828 CONT; 1829 1830 JMP_TAIL_CALL: { 1831 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1832 struct bpf_array *array = container_of(map, struct bpf_array, map); 1833 struct bpf_prog *prog; 1834 u32 index = BPF_R3; 1835 1836 if (unlikely(index >= array->map.max_entries)) 1837 goto out; 1838 1839 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 1840 goto out; 1841 1842 tail_call_cnt++; 1843 1844 prog = READ_ONCE(array->ptrs[index]); 1845 if (!prog) 1846 goto out; 1847 1848 /* ARG1 at this point is guaranteed to point to CTX from 1849 * the verifier side due to the fact that the tail call is 1850 * handled like a helper, that is, bpf_tail_call_proto, 1851 * where arg1_type is ARG_PTR_TO_CTX. 1852 */ 1853 insn = prog->insnsi; 1854 goto select_insn; 1855 out: 1856 CONT; 1857 } 1858 JMP_JA: 1859 insn += insn->off; 1860 CONT; 1861 JMP_EXIT: 1862 return BPF_R0; 1863 /* JMP */ 1864 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1865 JMP_##OPCODE##_X: \ 1866 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1867 insn += insn->off; \ 1868 CONT_JMP; \ 1869 } \ 1870 CONT; \ 1871 JMP32_##OPCODE##_X: \ 1872 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1873 insn += insn->off; \ 1874 CONT_JMP; \ 1875 } \ 1876 CONT; \ 1877 JMP_##OPCODE##_K: \ 1878 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1879 insn += insn->off; \ 1880 CONT_JMP; \ 1881 } \ 1882 CONT; \ 1883 JMP32_##OPCODE##_K: \ 1884 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1885 insn += insn->off; \ 1886 CONT_JMP; \ 1887 } \ 1888 CONT; 1889 COND_JMP(u, JEQ, ==) 1890 COND_JMP(u, JNE, !=) 1891 COND_JMP(u, JGT, >) 1892 COND_JMP(u, JLT, <) 1893 COND_JMP(u, JGE, >=) 1894 COND_JMP(u, JLE, <=) 1895 COND_JMP(u, JSET, &) 1896 COND_JMP(s, JSGT, >) 1897 COND_JMP(s, JSLT, <) 1898 COND_JMP(s, JSGE, >=) 1899 COND_JMP(s, JSLE, <=) 1900 #undef COND_JMP 1901 /* ST, STX and LDX*/ 1902 ST_NOSPEC: 1903 /* Speculation barrier for mitigating Speculative Store Bypass. 1904 * In case of arm64, we rely on the firmware mitigation as 1905 * controlled via the ssbd kernel parameter. Whenever the 1906 * mitigation is enabled, it works for all of the kernel code 1907 * with no need to provide any additional instructions here. 1908 * In case of x86, we use 'lfence' insn for mitigation. We 1909 * reuse preexisting logic from Spectre v1 mitigation that 1910 * happens to produce the required code on x86 for v4 as well. 1911 */ 1912 #ifdef CONFIG_X86 1913 barrier_nospec(); 1914 #endif 1915 CONT; 1916 #define LDST(SIZEOP, SIZE) \ 1917 STX_MEM_##SIZEOP: \ 1918 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1919 CONT; \ 1920 ST_MEM_##SIZEOP: \ 1921 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1922 CONT; \ 1923 LDX_MEM_##SIZEOP: \ 1924 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1925 CONT; 1926 1927 LDST(B, u8) 1928 LDST(H, u16) 1929 LDST(W, u32) 1930 LDST(DW, u64) 1931 #undef LDST 1932 #define LDX_PROBE(SIZEOP, SIZE) \ 1933 LDX_PROBE_MEM_##SIZEOP: \ 1934 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \ 1935 CONT; 1936 LDX_PROBE(B, 1) 1937 LDX_PROBE(H, 2) 1938 LDX_PROBE(W, 4) 1939 LDX_PROBE(DW, 8) 1940 #undef LDX_PROBE 1941 1942 #define ATOMIC_ALU_OP(BOP, KOP) \ 1943 case BOP: \ 1944 if (BPF_SIZE(insn->code) == BPF_W) \ 1945 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 1946 (DST + insn->off)); \ 1947 else \ 1948 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 1949 (DST + insn->off)); \ 1950 break; \ 1951 case BOP | BPF_FETCH: \ 1952 if (BPF_SIZE(insn->code) == BPF_W) \ 1953 SRC = (u32) atomic_fetch_##KOP( \ 1954 (u32) SRC, \ 1955 (atomic_t *)(unsigned long) (DST + insn->off)); \ 1956 else \ 1957 SRC = (u64) atomic64_fetch_##KOP( \ 1958 (u64) SRC, \ 1959 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 1960 break; 1961 1962 STX_ATOMIC_DW: 1963 STX_ATOMIC_W: 1964 switch (IMM) { 1965 ATOMIC_ALU_OP(BPF_ADD, add) 1966 ATOMIC_ALU_OP(BPF_AND, and) 1967 ATOMIC_ALU_OP(BPF_OR, or) 1968 ATOMIC_ALU_OP(BPF_XOR, xor) 1969 #undef ATOMIC_ALU_OP 1970 1971 case BPF_XCHG: 1972 if (BPF_SIZE(insn->code) == BPF_W) 1973 SRC = (u32) atomic_xchg( 1974 (atomic_t *)(unsigned long) (DST + insn->off), 1975 (u32) SRC); 1976 else 1977 SRC = (u64) atomic64_xchg( 1978 (atomic64_t *)(unsigned long) (DST + insn->off), 1979 (u64) SRC); 1980 break; 1981 case BPF_CMPXCHG: 1982 if (BPF_SIZE(insn->code) == BPF_W) 1983 BPF_R0 = (u32) atomic_cmpxchg( 1984 (atomic_t *)(unsigned long) (DST + insn->off), 1985 (u32) BPF_R0, (u32) SRC); 1986 else 1987 BPF_R0 = (u64) atomic64_cmpxchg( 1988 (atomic64_t *)(unsigned long) (DST + insn->off), 1989 (u64) BPF_R0, (u64) SRC); 1990 break; 1991 1992 default: 1993 goto default_label; 1994 } 1995 CONT; 1996 1997 default_label: 1998 /* If we ever reach this, we have a bug somewhere. Die hard here 1999 * instead of just returning 0; we could be somewhere in a subprog, 2000 * so execution could continue otherwise which we do /not/ want. 2001 * 2002 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2003 */ 2004 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2005 insn->code, insn->imm); 2006 BUG_ON(1); 2007 return 0; 2008 } 2009 2010 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2011 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2012 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2013 { \ 2014 u64 stack[stack_size / sizeof(u64)]; \ 2015 u64 regs[MAX_BPF_EXT_REG]; \ 2016 \ 2017 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2018 ARG1 = (u64) (unsigned long) ctx; \ 2019 return ___bpf_prog_run(regs, insn); \ 2020 } 2021 2022 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2023 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2024 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2025 const struct bpf_insn *insn) \ 2026 { \ 2027 u64 stack[stack_size / sizeof(u64)]; \ 2028 u64 regs[MAX_BPF_EXT_REG]; \ 2029 \ 2030 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2031 BPF_R1 = r1; \ 2032 BPF_R2 = r2; \ 2033 BPF_R3 = r3; \ 2034 BPF_R4 = r4; \ 2035 BPF_R5 = r5; \ 2036 return ___bpf_prog_run(regs, insn); \ 2037 } 2038 2039 #define EVAL1(FN, X) FN(X) 2040 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2041 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2042 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2043 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2044 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2045 2046 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2047 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2048 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2049 2050 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2051 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2052 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2053 2054 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2055 2056 static unsigned int (*interpreters[])(const void *ctx, 2057 const struct bpf_insn *insn) = { 2058 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2059 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2060 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2061 }; 2062 #undef PROG_NAME_LIST 2063 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2064 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2065 const struct bpf_insn *insn) = { 2066 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2067 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2068 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2069 }; 2070 #undef PROG_NAME_LIST 2071 2072 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2073 { 2074 stack_depth = max_t(u32, stack_depth, 1); 2075 insn->off = (s16) insn->imm; 2076 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2077 __bpf_call_base_args; 2078 insn->code = BPF_JMP | BPF_CALL_ARGS; 2079 } 2080 2081 #else 2082 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2083 const struct bpf_insn *insn) 2084 { 2085 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2086 * is not working properly, so warn about it! 2087 */ 2088 WARN_ON_ONCE(1); 2089 return 0; 2090 } 2091 #endif 2092 2093 bool bpf_prog_map_compatible(struct bpf_map *map, 2094 const struct bpf_prog *fp) 2095 { 2096 bool ret; 2097 2098 if (fp->kprobe_override) 2099 return false; 2100 2101 spin_lock(&map->owner.lock); 2102 if (!map->owner.type) { 2103 /* There's no owner yet where we could check for 2104 * compatibility. 2105 */ 2106 map->owner.type = fp->type; 2107 map->owner.jited = fp->jited; 2108 map->owner.xdp_has_frags = fp->aux->xdp_has_frags; 2109 ret = true; 2110 } else { 2111 ret = map->owner.type == fp->type && 2112 map->owner.jited == fp->jited && 2113 map->owner.xdp_has_frags == fp->aux->xdp_has_frags; 2114 } 2115 spin_unlock(&map->owner.lock); 2116 2117 return ret; 2118 } 2119 2120 static int bpf_check_tail_call(const struct bpf_prog *fp) 2121 { 2122 struct bpf_prog_aux *aux = fp->aux; 2123 int i, ret = 0; 2124 2125 mutex_lock(&aux->used_maps_mutex); 2126 for (i = 0; i < aux->used_map_cnt; i++) { 2127 struct bpf_map *map = aux->used_maps[i]; 2128 2129 if (!map_type_contains_progs(map)) 2130 continue; 2131 2132 if (!bpf_prog_map_compatible(map, fp)) { 2133 ret = -EINVAL; 2134 goto out; 2135 } 2136 } 2137 2138 out: 2139 mutex_unlock(&aux->used_maps_mutex); 2140 return ret; 2141 } 2142 2143 static void bpf_prog_select_func(struct bpf_prog *fp) 2144 { 2145 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2146 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2147 2148 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2149 #else 2150 fp->bpf_func = __bpf_prog_ret0_warn; 2151 #endif 2152 } 2153 2154 /** 2155 * bpf_prog_select_runtime - select exec runtime for BPF program 2156 * @fp: bpf_prog populated with BPF program 2157 * @err: pointer to error variable 2158 * 2159 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2160 * The BPF program will be executed via bpf_prog_run() function. 2161 * 2162 * Return: the &fp argument along with &err set to 0 for success or 2163 * a negative errno code on failure 2164 */ 2165 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2166 { 2167 /* In case of BPF to BPF calls, verifier did all the prep 2168 * work with regards to JITing, etc. 2169 */ 2170 bool jit_needed = false; 2171 2172 if (fp->bpf_func) 2173 goto finalize; 2174 2175 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2176 bpf_prog_has_kfunc_call(fp)) 2177 jit_needed = true; 2178 2179 bpf_prog_select_func(fp); 2180 2181 /* eBPF JITs can rewrite the program in case constant 2182 * blinding is active. However, in case of error during 2183 * blinding, bpf_int_jit_compile() must always return a 2184 * valid program, which in this case would simply not 2185 * be JITed, but falls back to the interpreter. 2186 */ 2187 if (!bpf_prog_is_dev_bound(fp->aux)) { 2188 *err = bpf_prog_alloc_jited_linfo(fp); 2189 if (*err) 2190 return fp; 2191 2192 fp = bpf_int_jit_compile(fp); 2193 bpf_prog_jit_attempt_done(fp); 2194 if (!fp->jited && jit_needed) { 2195 *err = -ENOTSUPP; 2196 return fp; 2197 } 2198 } else { 2199 *err = bpf_prog_offload_compile(fp); 2200 if (*err) 2201 return fp; 2202 } 2203 2204 finalize: 2205 bpf_prog_lock_ro(fp); 2206 2207 /* The tail call compatibility check can only be done at 2208 * this late stage as we need to determine, if we deal 2209 * with JITed or non JITed program concatenations and not 2210 * all eBPF JITs might immediately support all features. 2211 */ 2212 *err = bpf_check_tail_call(fp); 2213 2214 return fp; 2215 } 2216 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2217 2218 static unsigned int __bpf_prog_ret1(const void *ctx, 2219 const struct bpf_insn *insn) 2220 { 2221 return 1; 2222 } 2223 2224 static struct bpf_prog_dummy { 2225 struct bpf_prog prog; 2226 } dummy_bpf_prog = { 2227 .prog = { 2228 .bpf_func = __bpf_prog_ret1, 2229 }, 2230 }; 2231 2232 struct bpf_empty_prog_array bpf_empty_prog_array = { 2233 .null_prog = NULL, 2234 }; 2235 EXPORT_SYMBOL(bpf_empty_prog_array); 2236 2237 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2238 { 2239 if (prog_cnt) 2240 return kzalloc(sizeof(struct bpf_prog_array) + 2241 sizeof(struct bpf_prog_array_item) * 2242 (prog_cnt + 1), 2243 flags); 2244 2245 return &bpf_empty_prog_array.hdr; 2246 } 2247 2248 void bpf_prog_array_free(struct bpf_prog_array *progs) 2249 { 2250 if (!progs || progs == &bpf_empty_prog_array.hdr) 2251 return; 2252 kfree_rcu(progs, rcu); 2253 } 2254 2255 int bpf_prog_array_length(struct bpf_prog_array *array) 2256 { 2257 struct bpf_prog_array_item *item; 2258 u32 cnt = 0; 2259 2260 for (item = array->items; item->prog; item++) 2261 if (item->prog != &dummy_bpf_prog.prog) 2262 cnt++; 2263 return cnt; 2264 } 2265 2266 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2267 { 2268 struct bpf_prog_array_item *item; 2269 2270 for (item = array->items; item->prog; item++) 2271 if (item->prog != &dummy_bpf_prog.prog) 2272 return false; 2273 return true; 2274 } 2275 2276 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2277 u32 *prog_ids, 2278 u32 request_cnt) 2279 { 2280 struct bpf_prog_array_item *item; 2281 int i = 0; 2282 2283 for (item = array->items; item->prog; item++) { 2284 if (item->prog == &dummy_bpf_prog.prog) 2285 continue; 2286 prog_ids[i] = item->prog->aux->id; 2287 if (++i == request_cnt) { 2288 item++; 2289 break; 2290 } 2291 } 2292 2293 return !!(item->prog); 2294 } 2295 2296 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2297 __u32 __user *prog_ids, u32 cnt) 2298 { 2299 unsigned long err = 0; 2300 bool nospc; 2301 u32 *ids; 2302 2303 /* users of this function are doing: 2304 * cnt = bpf_prog_array_length(); 2305 * if (cnt > 0) 2306 * bpf_prog_array_copy_to_user(..., cnt); 2307 * so below kcalloc doesn't need extra cnt > 0 check. 2308 */ 2309 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2310 if (!ids) 2311 return -ENOMEM; 2312 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2313 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2314 kfree(ids); 2315 if (err) 2316 return -EFAULT; 2317 if (nospc) 2318 return -ENOSPC; 2319 return 0; 2320 } 2321 2322 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2323 struct bpf_prog *old_prog) 2324 { 2325 struct bpf_prog_array_item *item; 2326 2327 for (item = array->items; item->prog; item++) 2328 if (item->prog == old_prog) { 2329 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2330 break; 2331 } 2332 } 2333 2334 /** 2335 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2336 * index into the program array with 2337 * a dummy no-op program. 2338 * @array: a bpf_prog_array 2339 * @index: the index of the program to replace 2340 * 2341 * Skips over dummy programs, by not counting them, when calculating 2342 * the position of the program to replace. 2343 * 2344 * Return: 2345 * * 0 - Success 2346 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2347 * * -ENOENT - Index out of range 2348 */ 2349 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2350 { 2351 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2352 } 2353 2354 /** 2355 * bpf_prog_array_update_at() - Updates the program at the given index 2356 * into the program array. 2357 * @array: a bpf_prog_array 2358 * @index: the index of the program to update 2359 * @prog: the program to insert into the array 2360 * 2361 * Skips over dummy programs, by not counting them, when calculating 2362 * the position of the program to update. 2363 * 2364 * Return: 2365 * * 0 - Success 2366 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2367 * * -ENOENT - Index out of range 2368 */ 2369 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2370 struct bpf_prog *prog) 2371 { 2372 struct bpf_prog_array_item *item; 2373 2374 if (unlikely(index < 0)) 2375 return -EINVAL; 2376 2377 for (item = array->items; item->prog; item++) { 2378 if (item->prog == &dummy_bpf_prog.prog) 2379 continue; 2380 if (!index) { 2381 WRITE_ONCE(item->prog, prog); 2382 return 0; 2383 } 2384 index--; 2385 } 2386 return -ENOENT; 2387 } 2388 2389 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2390 struct bpf_prog *exclude_prog, 2391 struct bpf_prog *include_prog, 2392 u64 bpf_cookie, 2393 struct bpf_prog_array **new_array) 2394 { 2395 int new_prog_cnt, carry_prog_cnt = 0; 2396 struct bpf_prog_array_item *existing, *new; 2397 struct bpf_prog_array *array; 2398 bool found_exclude = false; 2399 2400 /* Figure out how many existing progs we need to carry over to 2401 * the new array. 2402 */ 2403 if (old_array) { 2404 existing = old_array->items; 2405 for (; existing->prog; existing++) { 2406 if (existing->prog == exclude_prog) { 2407 found_exclude = true; 2408 continue; 2409 } 2410 if (existing->prog != &dummy_bpf_prog.prog) 2411 carry_prog_cnt++; 2412 if (existing->prog == include_prog) 2413 return -EEXIST; 2414 } 2415 } 2416 2417 if (exclude_prog && !found_exclude) 2418 return -ENOENT; 2419 2420 /* How many progs (not NULL) will be in the new array? */ 2421 new_prog_cnt = carry_prog_cnt; 2422 if (include_prog) 2423 new_prog_cnt += 1; 2424 2425 /* Do we have any prog (not NULL) in the new array? */ 2426 if (!new_prog_cnt) { 2427 *new_array = NULL; 2428 return 0; 2429 } 2430 2431 /* +1 as the end of prog_array is marked with NULL */ 2432 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2433 if (!array) 2434 return -ENOMEM; 2435 new = array->items; 2436 2437 /* Fill in the new prog array */ 2438 if (carry_prog_cnt) { 2439 existing = old_array->items; 2440 for (; existing->prog; existing++) { 2441 if (existing->prog == exclude_prog || 2442 existing->prog == &dummy_bpf_prog.prog) 2443 continue; 2444 2445 new->prog = existing->prog; 2446 new->bpf_cookie = existing->bpf_cookie; 2447 new++; 2448 } 2449 } 2450 if (include_prog) { 2451 new->prog = include_prog; 2452 new->bpf_cookie = bpf_cookie; 2453 new++; 2454 } 2455 new->prog = NULL; 2456 *new_array = array; 2457 return 0; 2458 } 2459 2460 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2461 u32 *prog_ids, u32 request_cnt, 2462 u32 *prog_cnt) 2463 { 2464 u32 cnt = 0; 2465 2466 if (array) 2467 cnt = bpf_prog_array_length(array); 2468 2469 *prog_cnt = cnt; 2470 2471 /* return early if user requested only program count or nothing to copy */ 2472 if (!request_cnt || !cnt) 2473 return 0; 2474 2475 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2476 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2477 : 0; 2478 } 2479 2480 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2481 struct bpf_map **used_maps, u32 len) 2482 { 2483 struct bpf_map *map; 2484 u32 i; 2485 2486 for (i = 0; i < len; i++) { 2487 map = used_maps[i]; 2488 if (map->ops->map_poke_untrack) 2489 map->ops->map_poke_untrack(map, aux); 2490 bpf_map_put(map); 2491 } 2492 } 2493 2494 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2495 { 2496 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2497 kfree(aux->used_maps); 2498 } 2499 2500 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2501 struct btf_mod_pair *used_btfs, u32 len) 2502 { 2503 #ifdef CONFIG_BPF_SYSCALL 2504 struct btf_mod_pair *btf_mod; 2505 u32 i; 2506 2507 for (i = 0; i < len; i++) { 2508 btf_mod = &used_btfs[i]; 2509 if (btf_mod->module) 2510 module_put(btf_mod->module); 2511 btf_put(btf_mod->btf); 2512 } 2513 #endif 2514 } 2515 2516 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2517 { 2518 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2519 kfree(aux->used_btfs); 2520 } 2521 2522 static void bpf_prog_free_deferred(struct work_struct *work) 2523 { 2524 struct bpf_prog_aux *aux; 2525 int i; 2526 2527 aux = container_of(work, struct bpf_prog_aux, work); 2528 #ifdef CONFIG_BPF_SYSCALL 2529 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2530 #endif 2531 bpf_free_used_maps(aux); 2532 bpf_free_used_btfs(aux); 2533 if (bpf_prog_is_dev_bound(aux)) 2534 bpf_prog_offload_destroy(aux->prog); 2535 #ifdef CONFIG_PERF_EVENTS 2536 if (aux->prog->has_callchain_buf) 2537 put_callchain_buffers(); 2538 #endif 2539 if (aux->dst_trampoline) 2540 bpf_trampoline_put(aux->dst_trampoline); 2541 for (i = 0; i < aux->func_cnt; i++) { 2542 /* We can just unlink the subprog poke descriptor table as 2543 * it was originally linked to the main program and is also 2544 * released along with it. 2545 */ 2546 aux->func[i]->aux->poke_tab = NULL; 2547 bpf_jit_free(aux->func[i]); 2548 } 2549 if (aux->func_cnt) { 2550 kfree(aux->func); 2551 bpf_prog_unlock_free(aux->prog); 2552 } else { 2553 bpf_jit_free(aux->prog); 2554 } 2555 } 2556 2557 void bpf_prog_free(struct bpf_prog *fp) 2558 { 2559 struct bpf_prog_aux *aux = fp->aux; 2560 2561 if (aux->dst_prog) 2562 bpf_prog_put(aux->dst_prog); 2563 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2564 schedule_work(&aux->work); 2565 } 2566 EXPORT_SYMBOL_GPL(bpf_prog_free); 2567 2568 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2569 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2570 2571 void bpf_user_rnd_init_once(void) 2572 { 2573 prandom_init_once(&bpf_user_rnd_state); 2574 } 2575 2576 BPF_CALL_0(bpf_user_rnd_u32) 2577 { 2578 /* Should someone ever have the rather unwise idea to use some 2579 * of the registers passed into this function, then note that 2580 * this function is called from native eBPF and classic-to-eBPF 2581 * transformations. Register assignments from both sides are 2582 * different, f.e. classic always sets fn(ctx, A, X) here. 2583 */ 2584 struct rnd_state *state; 2585 u32 res; 2586 2587 state = &get_cpu_var(bpf_user_rnd_state); 2588 res = prandom_u32_state(state); 2589 put_cpu_var(bpf_user_rnd_state); 2590 2591 return res; 2592 } 2593 2594 BPF_CALL_0(bpf_get_raw_cpu_id) 2595 { 2596 return raw_smp_processor_id(); 2597 } 2598 2599 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2600 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2601 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2602 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2603 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2604 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2605 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2606 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2607 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2608 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2609 2610 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2611 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2612 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2613 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2614 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2615 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2616 2617 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2618 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2619 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2620 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2621 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2622 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2623 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2624 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2625 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2626 2627 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2628 { 2629 return NULL; 2630 } 2631 2632 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2633 { 2634 return NULL; 2635 } 2636 2637 u64 __weak 2638 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2639 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2640 { 2641 return -ENOTSUPP; 2642 } 2643 EXPORT_SYMBOL_GPL(bpf_event_output); 2644 2645 /* Always built-in helper functions. */ 2646 const struct bpf_func_proto bpf_tail_call_proto = { 2647 .func = NULL, 2648 .gpl_only = false, 2649 .ret_type = RET_VOID, 2650 .arg1_type = ARG_PTR_TO_CTX, 2651 .arg2_type = ARG_CONST_MAP_PTR, 2652 .arg3_type = ARG_ANYTHING, 2653 }; 2654 2655 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2656 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2657 * eBPF and implicitly also cBPF can get JITed! 2658 */ 2659 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2660 { 2661 return prog; 2662 } 2663 2664 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2665 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2666 */ 2667 void __weak bpf_jit_compile(struct bpf_prog *prog) 2668 { 2669 } 2670 2671 bool __weak bpf_helper_changes_pkt_data(void *func) 2672 { 2673 return false; 2674 } 2675 2676 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2677 * analysis code and wants explicit zero extension inserted by verifier. 2678 * Otherwise, return FALSE. 2679 * 2680 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2681 * you don't override this. JITs that don't want these extra insns can detect 2682 * them using insn_is_zext. 2683 */ 2684 bool __weak bpf_jit_needs_zext(void) 2685 { 2686 return false; 2687 } 2688 2689 bool __weak bpf_jit_supports_kfunc_call(void) 2690 { 2691 return false; 2692 } 2693 2694 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2695 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2696 */ 2697 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2698 int len) 2699 { 2700 return -EFAULT; 2701 } 2702 2703 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2704 void *addr1, void *addr2) 2705 { 2706 return -ENOTSUPP; 2707 } 2708 2709 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 2710 { 2711 return ERR_PTR(-ENOTSUPP); 2712 } 2713 2714 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2715 EXPORT_SYMBOL(bpf_stats_enabled_key); 2716 2717 /* All definitions of tracepoints related to BPF. */ 2718 #define CREATE_TRACE_POINTS 2719 #include <linux/bpf_trace.h> 2720 2721 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2722 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2723