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 static size_t bpf_prog_pack_mask = -1; 833 834 static int bpf_prog_chunk_count(void) 835 { 836 WARN_ON_ONCE(bpf_prog_pack_size == -1); 837 return bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE; 838 } 839 840 static DEFINE_MUTEX(pack_mutex); 841 static LIST_HEAD(pack_list); 842 843 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 844 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 845 */ 846 #ifdef PMD_SIZE 847 #define BPF_HPAGE_SIZE PMD_SIZE 848 #define BPF_HPAGE_MASK PMD_MASK 849 #else 850 #define BPF_HPAGE_SIZE PAGE_SIZE 851 #define BPF_HPAGE_MASK PAGE_MASK 852 #endif 853 854 static size_t select_bpf_prog_pack_size(void) 855 { 856 size_t size; 857 void *ptr; 858 859 size = BPF_HPAGE_SIZE * num_online_nodes(); 860 ptr = module_alloc(size); 861 862 /* Test whether we can get huge pages. If not just use PAGE_SIZE 863 * packs. 864 */ 865 if (!ptr || !is_vm_area_hugepages(ptr)) { 866 size = PAGE_SIZE; 867 bpf_prog_pack_mask = PAGE_MASK; 868 } else { 869 bpf_prog_pack_mask = BPF_HPAGE_MASK; 870 } 871 872 vfree(ptr); 873 return size; 874 } 875 876 static struct bpf_prog_pack *alloc_new_pack(void) 877 { 878 struct bpf_prog_pack *pack; 879 880 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(bpf_prog_chunk_count())), 881 GFP_KERNEL); 882 if (!pack) 883 return NULL; 884 pack->ptr = module_alloc(bpf_prog_pack_size); 885 if (!pack->ptr) { 886 kfree(pack); 887 return NULL; 888 } 889 bitmap_zero(pack->bitmap, bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE); 890 list_add_tail(&pack->list, &pack_list); 891 892 set_vm_flush_reset_perms(pack->ptr); 893 set_memory_ro((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE); 894 set_memory_x((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE); 895 return pack; 896 } 897 898 static void *bpf_prog_pack_alloc(u32 size) 899 { 900 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 901 struct bpf_prog_pack *pack; 902 unsigned long pos; 903 void *ptr = NULL; 904 905 mutex_lock(&pack_mutex); 906 if (bpf_prog_pack_size == -1) 907 bpf_prog_pack_size = select_bpf_prog_pack_size(); 908 909 if (size > bpf_prog_pack_size) { 910 size = round_up(size, PAGE_SIZE); 911 ptr = module_alloc(size); 912 if (ptr) { 913 set_vm_flush_reset_perms(ptr); 914 set_memory_ro((unsigned long)ptr, size / PAGE_SIZE); 915 set_memory_x((unsigned long)ptr, size / PAGE_SIZE); 916 } 917 goto out; 918 } 919 list_for_each_entry(pack, &pack_list, list) { 920 pos = bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0, 921 nbits, 0); 922 if (pos < bpf_prog_chunk_count()) 923 goto found_free_area; 924 } 925 926 pack = alloc_new_pack(); 927 if (!pack) 928 goto out; 929 930 pos = 0; 931 932 found_free_area: 933 bitmap_set(pack->bitmap, pos, nbits); 934 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 935 936 out: 937 mutex_unlock(&pack_mutex); 938 return ptr; 939 } 940 941 static void bpf_prog_pack_free(struct bpf_binary_header *hdr) 942 { 943 struct bpf_prog_pack *pack = NULL, *tmp; 944 unsigned int nbits; 945 unsigned long pos; 946 void *pack_ptr; 947 948 mutex_lock(&pack_mutex); 949 if (hdr->size > bpf_prog_pack_size) { 950 module_memfree(hdr); 951 goto out; 952 } 953 954 pack_ptr = (void *)((unsigned long)hdr & bpf_prog_pack_mask); 955 956 list_for_each_entry(tmp, &pack_list, list) { 957 if (tmp->ptr == pack_ptr) { 958 pack = tmp; 959 break; 960 } 961 } 962 963 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 964 goto out; 965 966 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size); 967 pos = ((unsigned long)hdr - (unsigned long)pack_ptr) >> BPF_PROG_CHUNK_SHIFT; 968 969 bitmap_clear(pack->bitmap, pos, nbits); 970 if (bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0, 971 bpf_prog_chunk_count(), 0) == 0) { 972 list_del(&pack->list); 973 module_memfree(pack->ptr); 974 kfree(pack); 975 } 976 out: 977 mutex_unlock(&pack_mutex); 978 } 979 980 static atomic_long_t bpf_jit_current; 981 982 /* Can be overridden by an arch's JIT compiler if it has a custom, 983 * dedicated BPF backend memory area, or if neither of the two 984 * below apply. 985 */ 986 u64 __weak bpf_jit_alloc_exec_limit(void) 987 { 988 #if defined(MODULES_VADDR) 989 return MODULES_END - MODULES_VADDR; 990 #else 991 return VMALLOC_END - VMALLOC_START; 992 #endif 993 } 994 995 static int __init bpf_jit_charge_init(void) 996 { 997 /* Only used as heuristic here to derive limit. */ 998 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 999 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2, 1000 PAGE_SIZE), LONG_MAX); 1001 return 0; 1002 } 1003 pure_initcall(bpf_jit_charge_init); 1004 1005 int bpf_jit_charge_modmem(u32 size) 1006 { 1007 if (atomic_long_add_return(size, &bpf_jit_current) > bpf_jit_limit) { 1008 if (!bpf_capable()) { 1009 atomic_long_sub(size, &bpf_jit_current); 1010 return -EPERM; 1011 } 1012 } 1013 1014 return 0; 1015 } 1016 1017 void bpf_jit_uncharge_modmem(u32 size) 1018 { 1019 atomic_long_sub(size, &bpf_jit_current); 1020 } 1021 1022 void *__weak bpf_jit_alloc_exec(unsigned long size) 1023 { 1024 return module_alloc(size); 1025 } 1026 1027 void __weak bpf_jit_free_exec(void *addr) 1028 { 1029 module_memfree(addr); 1030 } 1031 1032 struct bpf_binary_header * 1033 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1034 unsigned int alignment, 1035 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1036 { 1037 struct bpf_binary_header *hdr; 1038 u32 size, hole, start; 1039 1040 WARN_ON_ONCE(!is_power_of_2(alignment) || 1041 alignment > BPF_IMAGE_ALIGNMENT); 1042 1043 /* Most of BPF filters are really small, but if some of them 1044 * fill a page, allow at least 128 extra bytes to insert a 1045 * random section of illegal instructions. 1046 */ 1047 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1048 1049 if (bpf_jit_charge_modmem(size)) 1050 return NULL; 1051 hdr = bpf_jit_alloc_exec(size); 1052 if (!hdr) { 1053 bpf_jit_uncharge_modmem(size); 1054 return NULL; 1055 } 1056 1057 /* Fill space with illegal/arch-dep instructions. */ 1058 bpf_fill_ill_insns(hdr, size); 1059 1060 hdr->size = size; 1061 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1062 PAGE_SIZE - sizeof(*hdr)); 1063 start = (get_random_int() % hole) & ~(alignment - 1); 1064 1065 /* Leave a random number of instructions before BPF code. */ 1066 *image_ptr = &hdr->image[start]; 1067 1068 return hdr; 1069 } 1070 1071 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1072 { 1073 u32 size = hdr->size; 1074 1075 bpf_jit_free_exec(hdr); 1076 bpf_jit_uncharge_modmem(size); 1077 } 1078 1079 /* Allocate jit binary from bpf_prog_pack allocator. 1080 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1081 * to the memory. To solve this problem, a RW buffer is also allocated at 1082 * as the same time. The JIT engine should calculate offsets based on the 1083 * RO memory address, but write JITed program to the RW buffer. Once the 1084 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1085 * the JITed program to the RO memory. 1086 */ 1087 struct bpf_binary_header * 1088 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1089 unsigned int alignment, 1090 struct bpf_binary_header **rw_header, 1091 u8 **rw_image, 1092 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1093 { 1094 struct bpf_binary_header *ro_header; 1095 u32 size, hole, start; 1096 1097 WARN_ON_ONCE(!is_power_of_2(alignment) || 1098 alignment > BPF_IMAGE_ALIGNMENT); 1099 1100 /* add 16 bytes for a random section of illegal instructions */ 1101 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1102 1103 if (bpf_jit_charge_modmem(size)) 1104 return NULL; 1105 ro_header = bpf_prog_pack_alloc(size); 1106 if (!ro_header) { 1107 bpf_jit_uncharge_modmem(size); 1108 return NULL; 1109 } 1110 1111 *rw_header = kvmalloc(size, GFP_KERNEL); 1112 if (!*rw_header) { 1113 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size)); 1114 bpf_prog_pack_free(ro_header); 1115 bpf_jit_uncharge_modmem(size); 1116 return NULL; 1117 } 1118 1119 /* Fill space with illegal/arch-dep instructions. */ 1120 bpf_fill_ill_insns(*rw_header, size); 1121 (*rw_header)->size = size; 1122 1123 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1124 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1125 start = (get_random_int() % hole) & ~(alignment - 1); 1126 1127 *image_ptr = &ro_header->image[start]; 1128 *rw_image = &(*rw_header)->image[start]; 1129 1130 return ro_header; 1131 } 1132 1133 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1134 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog, 1135 struct bpf_binary_header *ro_header, 1136 struct bpf_binary_header *rw_header) 1137 { 1138 void *ptr; 1139 1140 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1141 1142 kvfree(rw_header); 1143 1144 if (IS_ERR(ptr)) { 1145 bpf_prog_pack_free(ro_header); 1146 return PTR_ERR(ptr); 1147 } 1148 prog->aux->use_bpf_prog_pack = true; 1149 return 0; 1150 } 1151 1152 /* bpf_jit_binary_pack_free is called in two different scenarios: 1153 * 1) when the program is freed after; 1154 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1155 * For case 2), we need to free both the RO memory and the RW buffer. 1156 * 1157 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1158 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1159 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1160 * bpf_arch_text_copy (when jit fails). 1161 */ 1162 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1163 struct bpf_binary_header *rw_header) 1164 { 1165 u32 size = ro_header->size; 1166 1167 bpf_prog_pack_free(ro_header); 1168 kvfree(rw_header); 1169 bpf_jit_uncharge_modmem(size); 1170 } 1171 1172 static inline struct bpf_binary_header * 1173 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1174 { 1175 unsigned long real_start = (unsigned long)fp->bpf_func; 1176 unsigned long addr; 1177 1178 if (fp->aux->use_bpf_prog_pack) 1179 addr = real_start & BPF_PROG_CHUNK_MASK; 1180 else 1181 addr = real_start & PAGE_MASK; 1182 1183 return (void *)addr; 1184 } 1185 1186 /* This symbol is only overridden by archs that have different 1187 * requirements than the usual eBPF JITs, f.e. when they only 1188 * implement cBPF JIT, do not set images read-only, etc. 1189 */ 1190 void __weak bpf_jit_free(struct bpf_prog *fp) 1191 { 1192 if (fp->jited) { 1193 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1194 1195 if (fp->aux->use_bpf_prog_pack) 1196 bpf_jit_binary_pack_free(hdr, NULL /* rw_buffer */); 1197 else 1198 bpf_jit_binary_free(hdr); 1199 1200 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1201 } 1202 1203 bpf_prog_unlock_free(fp); 1204 } 1205 1206 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1207 const struct bpf_insn *insn, bool extra_pass, 1208 u64 *func_addr, bool *func_addr_fixed) 1209 { 1210 s16 off = insn->off; 1211 s32 imm = insn->imm; 1212 u8 *addr; 1213 1214 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1215 if (!*func_addr_fixed) { 1216 /* Place-holder address till the last pass has collected 1217 * all addresses for JITed subprograms in which case we 1218 * can pick them up from prog->aux. 1219 */ 1220 if (!extra_pass) 1221 addr = NULL; 1222 else if (prog->aux->func && 1223 off >= 0 && off < prog->aux->func_cnt) 1224 addr = (u8 *)prog->aux->func[off]->bpf_func; 1225 else 1226 return -EINVAL; 1227 } else { 1228 /* Address of a BPF helper call. Since part of the core 1229 * kernel, it's always at a fixed location. __bpf_call_base 1230 * and the helper with imm relative to it are both in core 1231 * kernel. 1232 */ 1233 addr = (u8 *)__bpf_call_base + imm; 1234 } 1235 1236 *func_addr = (unsigned long)addr; 1237 return 0; 1238 } 1239 1240 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1241 const struct bpf_insn *aux, 1242 struct bpf_insn *to_buff, 1243 bool emit_zext) 1244 { 1245 struct bpf_insn *to = to_buff; 1246 u32 imm_rnd = get_random_int(); 1247 s16 off; 1248 1249 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1250 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1251 1252 /* Constraints on AX register: 1253 * 1254 * AX register is inaccessible from user space. It is mapped in 1255 * all JITs, and used here for constant blinding rewrites. It is 1256 * typically "stateless" meaning its contents are only valid within 1257 * the executed instruction, but not across several instructions. 1258 * There are a few exceptions however which are further detailed 1259 * below. 1260 * 1261 * Constant blinding is only used by JITs, not in the interpreter. 1262 * The interpreter uses AX in some occasions as a local temporary 1263 * register e.g. in DIV or MOD instructions. 1264 * 1265 * In restricted circumstances, the verifier can also use the AX 1266 * register for rewrites as long as they do not interfere with 1267 * the above cases! 1268 */ 1269 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1270 goto out; 1271 1272 if (from->imm == 0 && 1273 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1274 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1275 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1276 goto out; 1277 } 1278 1279 switch (from->code) { 1280 case BPF_ALU | BPF_ADD | BPF_K: 1281 case BPF_ALU | BPF_SUB | BPF_K: 1282 case BPF_ALU | BPF_AND | BPF_K: 1283 case BPF_ALU | BPF_OR | BPF_K: 1284 case BPF_ALU | BPF_XOR | BPF_K: 1285 case BPF_ALU | BPF_MUL | BPF_K: 1286 case BPF_ALU | BPF_MOV | BPF_K: 1287 case BPF_ALU | BPF_DIV | BPF_K: 1288 case BPF_ALU | BPF_MOD | BPF_K: 1289 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1290 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1291 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 1292 break; 1293 1294 case BPF_ALU64 | BPF_ADD | BPF_K: 1295 case BPF_ALU64 | BPF_SUB | BPF_K: 1296 case BPF_ALU64 | BPF_AND | BPF_K: 1297 case BPF_ALU64 | BPF_OR | BPF_K: 1298 case BPF_ALU64 | BPF_XOR | BPF_K: 1299 case BPF_ALU64 | BPF_MUL | BPF_K: 1300 case BPF_ALU64 | BPF_MOV | BPF_K: 1301 case BPF_ALU64 | BPF_DIV | BPF_K: 1302 case BPF_ALU64 | BPF_MOD | BPF_K: 1303 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1304 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1305 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 1306 break; 1307 1308 case BPF_JMP | BPF_JEQ | BPF_K: 1309 case BPF_JMP | BPF_JNE | BPF_K: 1310 case BPF_JMP | BPF_JGT | BPF_K: 1311 case BPF_JMP | BPF_JLT | BPF_K: 1312 case BPF_JMP | BPF_JGE | BPF_K: 1313 case BPF_JMP | BPF_JLE | BPF_K: 1314 case BPF_JMP | BPF_JSGT | BPF_K: 1315 case BPF_JMP | BPF_JSLT | BPF_K: 1316 case BPF_JMP | BPF_JSGE | BPF_K: 1317 case BPF_JMP | BPF_JSLE | BPF_K: 1318 case BPF_JMP | BPF_JSET | BPF_K: 1319 /* Accommodate for extra offset in case of a backjump. */ 1320 off = from->off; 1321 if (off < 0) 1322 off -= 2; 1323 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1324 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1325 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1326 break; 1327 1328 case BPF_JMP32 | BPF_JEQ | BPF_K: 1329 case BPF_JMP32 | BPF_JNE | BPF_K: 1330 case BPF_JMP32 | BPF_JGT | BPF_K: 1331 case BPF_JMP32 | BPF_JLT | BPF_K: 1332 case BPF_JMP32 | BPF_JGE | BPF_K: 1333 case BPF_JMP32 | BPF_JLE | BPF_K: 1334 case BPF_JMP32 | BPF_JSGT | BPF_K: 1335 case BPF_JMP32 | BPF_JSLT | BPF_K: 1336 case BPF_JMP32 | BPF_JSGE | BPF_K: 1337 case BPF_JMP32 | BPF_JSLE | BPF_K: 1338 case BPF_JMP32 | BPF_JSET | BPF_K: 1339 /* Accommodate for extra offset in case of a backjump. */ 1340 off = from->off; 1341 if (off < 0) 1342 off -= 2; 1343 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1344 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1345 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1346 off); 1347 break; 1348 1349 case BPF_LD | BPF_IMM | BPF_DW: 1350 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1351 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1352 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1353 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1354 break; 1355 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1356 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1357 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1358 if (emit_zext) 1359 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1360 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1361 break; 1362 1363 case BPF_ST | BPF_MEM | BPF_DW: 1364 case BPF_ST | BPF_MEM | BPF_W: 1365 case BPF_ST | BPF_MEM | BPF_H: 1366 case BPF_ST | BPF_MEM | BPF_B: 1367 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1368 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1369 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1370 break; 1371 } 1372 out: 1373 return to - to_buff; 1374 } 1375 1376 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1377 gfp_t gfp_extra_flags) 1378 { 1379 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1380 struct bpf_prog *fp; 1381 1382 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1383 if (fp != NULL) { 1384 /* aux->prog still points to the fp_other one, so 1385 * when promoting the clone to the real program, 1386 * this still needs to be adapted. 1387 */ 1388 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1389 } 1390 1391 return fp; 1392 } 1393 1394 static void bpf_prog_clone_free(struct bpf_prog *fp) 1395 { 1396 /* aux was stolen by the other clone, so we cannot free 1397 * it from this path! It will be freed eventually by the 1398 * other program on release. 1399 * 1400 * At this point, we don't need a deferred release since 1401 * clone is guaranteed to not be locked. 1402 */ 1403 fp->aux = NULL; 1404 fp->stats = NULL; 1405 fp->active = NULL; 1406 __bpf_prog_free(fp); 1407 } 1408 1409 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1410 { 1411 /* We have to repoint aux->prog to self, as we don't 1412 * know whether fp here is the clone or the original. 1413 */ 1414 fp->aux->prog = fp; 1415 bpf_prog_clone_free(fp_other); 1416 } 1417 1418 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1419 { 1420 struct bpf_insn insn_buff[16], aux[2]; 1421 struct bpf_prog *clone, *tmp; 1422 int insn_delta, insn_cnt; 1423 struct bpf_insn *insn; 1424 int i, rewritten; 1425 1426 if (!prog->blinding_requested || prog->blinded) 1427 return prog; 1428 1429 clone = bpf_prog_clone_create(prog, GFP_USER); 1430 if (!clone) 1431 return ERR_PTR(-ENOMEM); 1432 1433 insn_cnt = clone->len; 1434 insn = clone->insnsi; 1435 1436 for (i = 0; i < insn_cnt; i++, insn++) { 1437 /* We temporarily need to hold the original ld64 insn 1438 * so that we can still access the first part in the 1439 * second blinding run. 1440 */ 1441 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1442 insn[1].code == 0) 1443 memcpy(aux, insn, sizeof(aux)); 1444 1445 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1446 clone->aux->verifier_zext); 1447 if (!rewritten) 1448 continue; 1449 1450 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1451 if (IS_ERR(tmp)) { 1452 /* Patching may have repointed aux->prog during 1453 * realloc from the original one, so we need to 1454 * fix it up here on error. 1455 */ 1456 bpf_jit_prog_release_other(prog, clone); 1457 return tmp; 1458 } 1459 1460 clone = tmp; 1461 insn_delta = rewritten - 1; 1462 1463 /* Walk new program and skip insns we just inserted. */ 1464 insn = clone->insnsi + i + insn_delta; 1465 insn_cnt += insn_delta; 1466 i += insn_delta; 1467 } 1468 1469 clone->blinded = 1; 1470 return clone; 1471 } 1472 #endif /* CONFIG_BPF_JIT */ 1473 1474 /* Base function for offset calculation. Needs to go into .text section, 1475 * therefore keeping it non-static as well; will also be used by JITs 1476 * anyway later on, so do not let the compiler omit it. This also needs 1477 * to go into kallsyms for correlation from e.g. bpftool, so naming 1478 * must not change. 1479 */ 1480 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1481 { 1482 return 0; 1483 } 1484 EXPORT_SYMBOL_GPL(__bpf_call_base); 1485 1486 /* All UAPI available opcodes. */ 1487 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1488 /* 32 bit ALU operations. */ \ 1489 /* Register based. */ \ 1490 INSN_3(ALU, ADD, X), \ 1491 INSN_3(ALU, SUB, X), \ 1492 INSN_3(ALU, AND, X), \ 1493 INSN_3(ALU, OR, X), \ 1494 INSN_3(ALU, LSH, X), \ 1495 INSN_3(ALU, RSH, X), \ 1496 INSN_3(ALU, XOR, X), \ 1497 INSN_3(ALU, MUL, X), \ 1498 INSN_3(ALU, MOV, X), \ 1499 INSN_3(ALU, ARSH, X), \ 1500 INSN_3(ALU, DIV, X), \ 1501 INSN_3(ALU, MOD, X), \ 1502 INSN_2(ALU, NEG), \ 1503 INSN_3(ALU, END, TO_BE), \ 1504 INSN_3(ALU, END, TO_LE), \ 1505 /* Immediate based. */ \ 1506 INSN_3(ALU, ADD, K), \ 1507 INSN_3(ALU, SUB, K), \ 1508 INSN_3(ALU, AND, K), \ 1509 INSN_3(ALU, OR, K), \ 1510 INSN_3(ALU, LSH, K), \ 1511 INSN_3(ALU, RSH, K), \ 1512 INSN_3(ALU, XOR, K), \ 1513 INSN_3(ALU, MUL, K), \ 1514 INSN_3(ALU, MOV, K), \ 1515 INSN_3(ALU, ARSH, K), \ 1516 INSN_3(ALU, DIV, K), \ 1517 INSN_3(ALU, MOD, K), \ 1518 /* 64 bit ALU operations. */ \ 1519 /* Register based. */ \ 1520 INSN_3(ALU64, ADD, X), \ 1521 INSN_3(ALU64, SUB, X), \ 1522 INSN_3(ALU64, AND, X), \ 1523 INSN_3(ALU64, OR, X), \ 1524 INSN_3(ALU64, LSH, X), \ 1525 INSN_3(ALU64, RSH, X), \ 1526 INSN_3(ALU64, XOR, X), \ 1527 INSN_3(ALU64, MUL, X), \ 1528 INSN_3(ALU64, MOV, X), \ 1529 INSN_3(ALU64, ARSH, X), \ 1530 INSN_3(ALU64, DIV, X), \ 1531 INSN_3(ALU64, MOD, X), \ 1532 INSN_2(ALU64, NEG), \ 1533 /* Immediate based. */ \ 1534 INSN_3(ALU64, ADD, K), \ 1535 INSN_3(ALU64, SUB, K), \ 1536 INSN_3(ALU64, AND, K), \ 1537 INSN_3(ALU64, OR, K), \ 1538 INSN_3(ALU64, LSH, K), \ 1539 INSN_3(ALU64, RSH, K), \ 1540 INSN_3(ALU64, XOR, K), \ 1541 INSN_3(ALU64, MUL, K), \ 1542 INSN_3(ALU64, MOV, K), \ 1543 INSN_3(ALU64, ARSH, K), \ 1544 INSN_3(ALU64, DIV, K), \ 1545 INSN_3(ALU64, MOD, K), \ 1546 /* Call instruction. */ \ 1547 INSN_2(JMP, CALL), \ 1548 /* Exit instruction. */ \ 1549 INSN_2(JMP, EXIT), \ 1550 /* 32-bit Jump instructions. */ \ 1551 /* Register based. */ \ 1552 INSN_3(JMP32, JEQ, X), \ 1553 INSN_3(JMP32, JNE, X), \ 1554 INSN_3(JMP32, JGT, X), \ 1555 INSN_3(JMP32, JLT, X), \ 1556 INSN_3(JMP32, JGE, X), \ 1557 INSN_3(JMP32, JLE, X), \ 1558 INSN_3(JMP32, JSGT, X), \ 1559 INSN_3(JMP32, JSLT, X), \ 1560 INSN_3(JMP32, JSGE, X), \ 1561 INSN_3(JMP32, JSLE, X), \ 1562 INSN_3(JMP32, JSET, X), \ 1563 /* Immediate based. */ \ 1564 INSN_3(JMP32, JEQ, K), \ 1565 INSN_3(JMP32, JNE, K), \ 1566 INSN_3(JMP32, JGT, K), \ 1567 INSN_3(JMP32, JLT, K), \ 1568 INSN_3(JMP32, JGE, K), \ 1569 INSN_3(JMP32, JLE, K), \ 1570 INSN_3(JMP32, JSGT, K), \ 1571 INSN_3(JMP32, JSLT, K), \ 1572 INSN_3(JMP32, JSGE, K), \ 1573 INSN_3(JMP32, JSLE, K), \ 1574 INSN_3(JMP32, JSET, K), \ 1575 /* Jump instructions. */ \ 1576 /* Register based. */ \ 1577 INSN_3(JMP, JEQ, X), \ 1578 INSN_3(JMP, JNE, X), \ 1579 INSN_3(JMP, JGT, X), \ 1580 INSN_3(JMP, JLT, X), \ 1581 INSN_3(JMP, JGE, X), \ 1582 INSN_3(JMP, JLE, X), \ 1583 INSN_3(JMP, JSGT, X), \ 1584 INSN_3(JMP, JSLT, X), \ 1585 INSN_3(JMP, JSGE, X), \ 1586 INSN_3(JMP, JSLE, X), \ 1587 INSN_3(JMP, JSET, X), \ 1588 /* Immediate based. */ \ 1589 INSN_3(JMP, JEQ, K), \ 1590 INSN_3(JMP, JNE, K), \ 1591 INSN_3(JMP, JGT, K), \ 1592 INSN_3(JMP, JLT, K), \ 1593 INSN_3(JMP, JGE, K), \ 1594 INSN_3(JMP, JLE, K), \ 1595 INSN_3(JMP, JSGT, K), \ 1596 INSN_3(JMP, JSLT, K), \ 1597 INSN_3(JMP, JSGE, K), \ 1598 INSN_3(JMP, JSLE, K), \ 1599 INSN_3(JMP, JSET, K), \ 1600 INSN_2(JMP, JA), \ 1601 /* Store instructions. */ \ 1602 /* Register based. */ \ 1603 INSN_3(STX, MEM, B), \ 1604 INSN_3(STX, MEM, H), \ 1605 INSN_3(STX, MEM, W), \ 1606 INSN_3(STX, MEM, DW), \ 1607 INSN_3(STX, ATOMIC, W), \ 1608 INSN_3(STX, ATOMIC, DW), \ 1609 /* Immediate based. */ \ 1610 INSN_3(ST, MEM, B), \ 1611 INSN_3(ST, MEM, H), \ 1612 INSN_3(ST, MEM, W), \ 1613 INSN_3(ST, MEM, DW), \ 1614 /* Load instructions. */ \ 1615 /* Register based. */ \ 1616 INSN_3(LDX, MEM, B), \ 1617 INSN_3(LDX, MEM, H), \ 1618 INSN_3(LDX, MEM, W), \ 1619 INSN_3(LDX, MEM, DW), \ 1620 /* Immediate based. */ \ 1621 INSN_3(LD, IMM, DW) 1622 1623 bool bpf_opcode_in_insntable(u8 code) 1624 { 1625 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1626 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1627 static const bool public_insntable[256] = { 1628 [0 ... 255] = false, 1629 /* Now overwrite non-defaults ... */ 1630 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1631 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1632 [BPF_LD | BPF_ABS | BPF_B] = true, 1633 [BPF_LD | BPF_ABS | BPF_H] = true, 1634 [BPF_LD | BPF_ABS | BPF_W] = true, 1635 [BPF_LD | BPF_IND | BPF_B] = true, 1636 [BPF_LD | BPF_IND | BPF_H] = true, 1637 [BPF_LD | BPF_IND | BPF_W] = true, 1638 }; 1639 #undef BPF_INSN_3_TBL 1640 #undef BPF_INSN_2_TBL 1641 return public_insntable[code]; 1642 } 1643 1644 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1645 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr) 1646 { 1647 memset(dst, 0, size); 1648 return -EFAULT; 1649 } 1650 1651 /** 1652 * ___bpf_prog_run - run eBPF program on a given context 1653 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1654 * @insn: is the array of eBPF instructions 1655 * 1656 * Decode and execute eBPF instructions. 1657 * 1658 * Return: whatever value is in %BPF_R0 at program exit 1659 */ 1660 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1661 { 1662 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1663 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1664 static const void * const jumptable[256] __annotate_jump_table = { 1665 [0 ... 255] = &&default_label, 1666 /* Now overwrite non-defaults ... */ 1667 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1668 /* Non-UAPI available opcodes. */ 1669 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1670 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1671 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1672 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1673 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1674 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1675 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1676 }; 1677 #undef BPF_INSN_3_LBL 1678 #undef BPF_INSN_2_LBL 1679 u32 tail_call_cnt = 0; 1680 1681 #define CONT ({ insn++; goto select_insn; }) 1682 #define CONT_JMP ({ insn++; goto select_insn; }) 1683 1684 select_insn: 1685 goto *jumptable[insn->code]; 1686 1687 /* Explicitly mask the register-based shift amounts with 63 or 31 1688 * to avoid undefined behavior. Normally this won't affect the 1689 * generated code, for example, in case of native 64 bit archs such 1690 * as x86-64 or arm64, the compiler is optimizing the AND away for 1691 * the interpreter. In case of JITs, each of the JIT backends compiles 1692 * the BPF shift operations to machine instructions which produce 1693 * implementation-defined results in such a case; the resulting 1694 * contents of the register may be arbitrary, but program behaviour 1695 * as a whole remains defined. In other words, in case of JIT backends, 1696 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1697 */ 1698 /* ALU (shifts) */ 1699 #define SHT(OPCODE, OP) \ 1700 ALU64_##OPCODE##_X: \ 1701 DST = DST OP (SRC & 63); \ 1702 CONT; \ 1703 ALU_##OPCODE##_X: \ 1704 DST = (u32) DST OP ((u32) SRC & 31); \ 1705 CONT; \ 1706 ALU64_##OPCODE##_K: \ 1707 DST = DST OP IMM; \ 1708 CONT; \ 1709 ALU_##OPCODE##_K: \ 1710 DST = (u32) DST OP (u32) IMM; \ 1711 CONT; 1712 /* ALU (rest) */ 1713 #define ALU(OPCODE, OP) \ 1714 ALU64_##OPCODE##_X: \ 1715 DST = DST OP SRC; \ 1716 CONT; \ 1717 ALU_##OPCODE##_X: \ 1718 DST = (u32) DST OP (u32) SRC; \ 1719 CONT; \ 1720 ALU64_##OPCODE##_K: \ 1721 DST = DST OP IMM; \ 1722 CONT; \ 1723 ALU_##OPCODE##_K: \ 1724 DST = (u32) DST OP (u32) IMM; \ 1725 CONT; 1726 ALU(ADD, +) 1727 ALU(SUB, -) 1728 ALU(AND, &) 1729 ALU(OR, |) 1730 ALU(XOR, ^) 1731 ALU(MUL, *) 1732 SHT(LSH, <<) 1733 SHT(RSH, >>) 1734 #undef SHT 1735 #undef ALU 1736 ALU_NEG: 1737 DST = (u32) -DST; 1738 CONT; 1739 ALU64_NEG: 1740 DST = -DST; 1741 CONT; 1742 ALU_MOV_X: 1743 DST = (u32) SRC; 1744 CONT; 1745 ALU_MOV_K: 1746 DST = (u32) IMM; 1747 CONT; 1748 ALU64_MOV_X: 1749 DST = SRC; 1750 CONT; 1751 ALU64_MOV_K: 1752 DST = IMM; 1753 CONT; 1754 LD_IMM_DW: 1755 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1756 insn++; 1757 CONT; 1758 ALU_ARSH_X: 1759 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1760 CONT; 1761 ALU_ARSH_K: 1762 DST = (u64) (u32) (((s32) DST) >> IMM); 1763 CONT; 1764 ALU64_ARSH_X: 1765 (*(s64 *) &DST) >>= (SRC & 63); 1766 CONT; 1767 ALU64_ARSH_K: 1768 (*(s64 *) &DST) >>= IMM; 1769 CONT; 1770 ALU64_MOD_X: 1771 div64_u64_rem(DST, SRC, &AX); 1772 DST = AX; 1773 CONT; 1774 ALU_MOD_X: 1775 AX = (u32) DST; 1776 DST = do_div(AX, (u32) SRC); 1777 CONT; 1778 ALU64_MOD_K: 1779 div64_u64_rem(DST, IMM, &AX); 1780 DST = AX; 1781 CONT; 1782 ALU_MOD_K: 1783 AX = (u32) DST; 1784 DST = do_div(AX, (u32) IMM); 1785 CONT; 1786 ALU64_DIV_X: 1787 DST = div64_u64(DST, SRC); 1788 CONT; 1789 ALU_DIV_X: 1790 AX = (u32) DST; 1791 do_div(AX, (u32) SRC); 1792 DST = (u32) AX; 1793 CONT; 1794 ALU64_DIV_K: 1795 DST = div64_u64(DST, IMM); 1796 CONT; 1797 ALU_DIV_K: 1798 AX = (u32) DST; 1799 do_div(AX, (u32) IMM); 1800 DST = (u32) AX; 1801 CONT; 1802 ALU_END_TO_BE: 1803 switch (IMM) { 1804 case 16: 1805 DST = (__force u16) cpu_to_be16(DST); 1806 break; 1807 case 32: 1808 DST = (__force u32) cpu_to_be32(DST); 1809 break; 1810 case 64: 1811 DST = (__force u64) cpu_to_be64(DST); 1812 break; 1813 } 1814 CONT; 1815 ALU_END_TO_LE: 1816 switch (IMM) { 1817 case 16: 1818 DST = (__force u16) cpu_to_le16(DST); 1819 break; 1820 case 32: 1821 DST = (__force u32) cpu_to_le32(DST); 1822 break; 1823 case 64: 1824 DST = (__force u64) cpu_to_le64(DST); 1825 break; 1826 } 1827 CONT; 1828 1829 /* CALL */ 1830 JMP_CALL: 1831 /* Function call scratches BPF_R1-BPF_R5 registers, 1832 * preserves BPF_R6-BPF_R9, and stores return value 1833 * into BPF_R0. 1834 */ 1835 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1836 BPF_R4, BPF_R5); 1837 CONT; 1838 1839 JMP_CALL_ARGS: 1840 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1841 BPF_R3, BPF_R4, 1842 BPF_R5, 1843 insn + insn->off + 1); 1844 CONT; 1845 1846 JMP_TAIL_CALL: { 1847 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1848 struct bpf_array *array = container_of(map, struct bpf_array, map); 1849 struct bpf_prog *prog; 1850 u32 index = BPF_R3; 1851 1852 if (unlikely(index >= array->map.max_entries)) 1853 goto out; 1854 1855 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 1856 goto out; 1857 1858 tail_call_cnt++; 1859 1860 prog = READ_ONCE(array->ptrs[index]); 1861 if (!prog) 1862 goto out; 1863 1864 /* ARG1 at this point is guaranteed to point to CTX from 1865 * the verifier side due to the fact that the tail call is 1866 * handled like a helper, that is, bpf_tail_call_proto, 1867 * where arg1_type is ARG_PTR_TO_CTX. 1868 */ 1869 insn = prog->insnsi; 1870 goto select_insn; 1871 out: 1872 CONT; 1873 } 1874 JMP_JA: 1875 insn += insn->off; 1876 CONT; 1877 JMP_EXIT: 1878 return BPF_R0; 1879 /* JMP */ 1880 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 1881 JMP_##OPCODE##_X: \ 1882 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 1883 insn += insn->off; \ 1884 CONT_JMP; \ 1885 } \ 1886 CONT; \ 1887 JMP32_##OPCODE##_X: \ 1888 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 1889 insn += insn->off; \ 1890 CONT_JMP; \ 1891 } \ 1892 CONT; \ 1893 JMP_##OPCODE##_K: \ 1894 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 1895 insn += insn->off; \ 1896 CONT_JMP; \ 1897 } \ 1898 CONT; \ 1899 JMP32_##OPCODE##_K: \ 1900 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 1901 insn += insn->off; \ 1902 CONT_JMP; \ 1903 } \ 1904 CONT; 1905 COND_JMP(u, JEQ, ==) 1906 COND_JMP(u, JNE, !=) 1907 COND_JMP(u, JGT, >) 1908 COND_JMP(u, JLT, <) 1909 COND_JMP(u, JGE, >=) 1910 COND_JMP(u, JLE, <=) 1911 COND_JMP(u, JSET, &) 1912 COND_JMP(s, JSGT, >) 1913 COND_JMP(s, JSLT, <) 1914 COND_JMP(s, JSGE, >=) 1915 COND_JMP(s, JSLE, <=) 1916 #undef COND_JMP 1917 /* ST, STX and LDX*/ 1918 ST_NOSPEC: 1919 /* Speculation barrier for mitigating Speculative Store Bypass. 1920 * In case of arm64, we rely on the firmware mitigation as 1921 * controlled via the ssbd kernel parameter. Whenever the 1922 * mitigation is enabled, it works for all of the kernel code 1923 * with no need to provide any additional instructions here. 1924 * In case of x86, we use 'lfence' insn for mitigation. We 1925 * reuse preexisting logic from Spectre v1 mitigation that 1926 * happens to produce the required code on x86 for v4 as well. 1927 */ 1928 #ifdef CONFIG_X86 1929 barrier_nospec(); 1930 #endif 1931 CONT; 1932 #define LDST(SIZEOP, SIZE) \ 1933 STX_MEM_##SIZEOP: \ 1934 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1935 CONT; \ 1936 ST_MEM_##SIZEOP: \ 1937 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1938 CONT; \ 1939 LDX_MEM_##SIZEOP: \ 1940 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1941 CONT; 1942 1943 LDST(B, u8) 1944 LDST(H, u16) 1945 LDST(W, u32) 1946 LDST(DW, u64) 1947 #undef LDST 1948 #define LDX_PROBE(SIZEOP, SIZE) \ 1949 LDX_PROBE_MEM_##SIZEOP: \ 1950 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \ 1951 CONT; 1952 LDX_PROBE(B, 1) 1953 LDX_PROBE(H, 2) 1954 LDX_PROBE(W, 4) 1955 LDX_PROBE(DW, 8) 1956 #undef LDX_PROBE 1957 1958 #define ATOMIC_ALU_OP(BOP, KOP) \ 1959 case BOP: \ 1960 if (BPF_SIZE(insn->code) == BPF_W) \ 1961 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 1962 (DST + insn->off)); \ 1963 else \ 1964 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 1965 (DST + insn->off)); \ 1966 break; \ 1967 case BOP | BPF_FETCH: \ 1968 if (BPF_SIZE(insn->code) == BPF_W) \ 1969 SRC = (u32) atomic_fetch_##KOP( \ 1970 (u32) SRC, \ 1971 (atomic_t *)(unsigned long) (DST + insn->off)); \ 1972 else \ 1973 SRC = (u64) atomic64_fetch_##KOP( \ 1974 (u64) SRC, \ 1975 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 1976 break; 1977 1978 STX_ATOMIC_DW: 1979 STX_ATOMIC_W: 1980 switch (IMM) { 1981 ATOMIC_ALU_OP(BPF_ADD, add) 1982 ATOMIC_ALU_OP(BPF_AND, and) 1983 ATOMIC_ALU_OP(BPF_OR, or) 1984 ATOMIC_ALU_OP(BPF_XOR, xor) 1985 #undef ATOMIC_ALU_OP 1986 1987 case BPF_XCHG: 1988 if (BPF_SIZE(insn->code) == BPF_W) 1989 SRC = (u32) atomic_xchg( 1990 (atomic_t *)(unsigned long) (DST + insn->off), 1991 (u32) SRC); 1992 else 1993 SRC = (u64) atomic64_xchg( 1994 (atomic64_t *)(unsigned long) (DST + insn->off), 1995 (u64) SRC); 1996 break; 1997 case BPF_CMPXCHG: 1998 if (BPF_SIZE(insn->code) == BPF_W) 1999 BPF_R0 = (u32) atomic_cmpxchg( 2000 (atomic_t *)(unsigned long) (DST + insn->off), 2001 (u32) BPF_R0, (u32) SRC); 2002 else 2003 BPF_R0 = (u64) atomic64_cmpxchg( 2004 (atomic64_t *)(unsigned long) (DST + insn->off), 2005 (u64) BPF_R0, (u64) SRC); 2006 break; 2007 2008 default: 2009 goto default_label; 2010 } 2011 CONT; 2012 2013 default_label: 2014 /* If we ever reach this, we have a bug somewhere. Die hard here 2015 * instead of just returning 0; we could be somewhere in a subprog, 2016 * so execution could continue otherwise which we do /not/ want. 2017 * 2018 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2019 */ 2020 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2021 insn->code, insn->imm); 2022 BUG_ON(1); 2023 return 0; 2024 } 2025 2026 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2027 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2028 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2029 { \ 2030 u64 stack[stack_size / sizeof(u64)]; \ 2031 u64 regs[MAX_BPF_EXT_REG]; \ 2032 \ 2033 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2034 ARG1 = (u64) (unsigned long) ctx; \ 2035 return ___bpf_prog_run(regs, insn); \ 2036 } 2037 2038 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2039 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2040 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2041 const struct bpf_insn *insn) \ 2042 { \ 2043 u64 stack[stack_size / sizeof(u64)]; \ 2044 u64 regs[MAX_BPF_EXT_REG]; \ 2045 \ 2046 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2047 BPF_R1 = r1; \ 2048 BPF_R2 = r2; \ 2049 BPF_R3 = r3; \ 2050 BPF_R4 = r4; \ 2051 BPF_R5 = r5; \ 2052 return ___bpf_prog_run(regs, insn); \ 2053 } 2054 2055 #define EVAL1(FN, X) FN(X) 2056 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2057 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2058 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2059 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2060 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2061 2062 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2063 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2064 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2065 2066 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2067 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2068 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2069 2070 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2071 2072 static unsigned int (*interpreters[])(const void *ctx, 2073 const struct bpf_insn *insn) = { 2074 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2075 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2076 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2077 }; 2078 #undef PROG_NAME_LIST 2079 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2080 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2081 const struct bpf_insn *insn) = { 2082 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2083 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2084 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2085 }; 2086 #undef PROG_NAME_LIST 2087 2088 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2089 { 2090 stack_depth = max_t(u32, stack_depth, 1); 2091 insn->off = (s16) insn->imm; 2092 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2093 __bpf_call_base_args; 2094 insn->code = BPF_JMP | BPF_CALL_ARGS; 2095 } 2096 2097 #else 2098 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2099 const struct bpf_insn *insn) 2100 { 2101 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2102 * is not working properly, so warn about it! 2103 */ 2104 WARN_ON_ONCE(1); 2105 return 0; 2106 } 2107 #endif 2108 2109 bool bpf_prog_map_compatible(struct bpf_map *map, 2110 const struct bpf_prog *fp) 2111 { 2112 bool ret; 2113 2114 if (fp->kprobe_override) 2115 return false; 2116 2117 spin_lock(&map->owner.lock); 2118 if (!map->owner.type) { 2119 /* There's no owner yet where we could check for 2120 * compatibility. 2121 */ 2122 map->owner.type = fp->type; 2123 map->owner.jited = fp->jited; 2124 map->owner.xdp_has_frags = fp->aux->xdp_has_frags; 2125 ret = true; 2126 } else { 2127 ret = map->owner.type == fp->type && 2128 map->owner.jited == fp->jited && 2129 map->owner.xdp_has_frags == fp->aux->xdp_has_frags; 2130 } 2131 spin_unlock(&map->owner.lock); 2132 2133 return ret; 2134 } 2135 2136 static int bpf_check_tail_call(const struct bpf_prog *fp) 2137 { 2138 struct bpf_prog_aux *aux = fp->aux; 2139 int i, ret = 0; 2140 2141 mutex_lock(&aux->used_maps_mutex); 2142 for (i = 0; i < aux->used_map_cnt; i++) { 2143 struct bpf_map *map = aux->used_maps[i]; 2144 2145 if (!map_type_contains_progs(map)) 2146 continue; 2147 2148 if (!bpf_prog_map_compatible(map, fp)) { 2149 ret = -EINVAL; 2150 goto out; 2151 } 2152 } 2153 2154 out: 2155 mutex_unlock(&aux->used_maps_mutex); 2156 return ret; 2157 } 2158 2159 static void bpf_prog_select_func(struct bpf_prog *fp) 2160 { 2161 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2162 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2163 2164 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2165 #else 2166 fp->bpf_func = __bpf_prog_ret0_warn; 2167 #endif 2168 } 2169 2170 /** 2171 * bpf_prog_select_runtime - select exec runtime for BPF program 2172 * @fp: bpf_prog populated with BPF program 2173 * @err: pointer to error variable 2174 * 2175 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2176 * The BPF program will be executed via bpf_prog_run() function. 2177 * 2178 * Return: the &fp argument along with &err set to 0 for success or 2179 * a negative errno code on failure 2180 */ 2181 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2182 { 2183 /* In case of BPF to BPF calls, verifier did all the prep 2184 * work with regards to JITing, etc. 2185 */ 2186 bool jit_needed = false; 2187 2188 if (fp->bpf_func) 2189 goto finalize; 2190 2191 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2192 bpf_prog_has_kfunc_call(fp)) 2193 jit_needed = true; 2194 2195 bpf_prog_select_func(fp); 2196 2197 /* eBPF JITs can rewrite the program in case constant 2198 * blinding is active. However, in case of error during 2199 * blinding, bpf_int_jit_compile() must always return a 2200 * valid program, which in this case would simply not 2201 * be JITed, but falls back to the interpreter. 2202 */ 2203 if (!bpf_prog_is_dev_bound(fp->aux)) { 2204 *err = bpf_prog_alloc_jited_linfo(fp); 2205 if (*err) 2206 return fp; 2207 2208 fp = bpf_int_jit_compile(fp); 2209 bpf_prog_jit_attempt_done(fp); 2210 if (!fp->jited && jit_needed) { 2211 *err = -ENOTSUPP; 2212 return fp; 2213 } 2214 } else { 2215 *err = bpf_prog_offload_compile(fp); 2216 if (*err) 2217 return fp; 2218 } 2219 2220 finalize: 2221 bpf_prog_lock_ro(fp); 2222 2223 /* The tail call compatibility check can only be done at 2224 * this late stage as we need to determine, if we deal 2225 * with JITed or non JITed program concatenations and not 2226 * all eBPF JITs might immediately support all features. 2227 */ 2228 *err = bpf_check_tail_call(fp); 2229 2230 return fp; 2231 } 2232 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2233 2234 static unsigned int __bpf_prog_ret1(const void *ctx, 2235 const struct bpf_insn *insn) 2236 { 2237 return 1; 2238 } 2239 2240 static struct bpf_prog_dummy { 2241 struct bpf_prog prog; 2242 } dummy_bpf_prog = { 2243 .prog = { 2244 .bpf_func = __bpf_prog_ret1, 2245 }, 2246 }; 2247 2248 struct bpf_empty_prog_array bpf_empty_prog_array = { 2249 .null_prog = NULL, 2250 }; 2251 EXPORT_SYMBOL(bpf_empty_prog_array); 2252 2253 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2254 { 2255 if (prog_cnt) 2256 return kzalloc(sizeof(struct bpf_prog_array) + 2257 sizeof(struct bpf_prog_array_item) * 2258 (prog_cnt + 1), 2259 flags); 2260 2261 return &bpf_empty_prog_array.hdr; 2262 } 2263 2264 void bpf_prog_array_free(struct bpf_prog_array *progs) 2265 { 2266 if (!progs || progs == &bpf_empty_prog_array.hdr) 2267 return; 2268 kfree_rcu(progs, rcu); 2269 } 2270 2271 int bpf_prog_array_length(struct bpf_prog_array *array) 2272 { 2273 struct bpf_prog_array_item *item; 2274 u32 cnt = 0; 2275 2276 for (item = array->items; item->prog; item++) 2277 if (item->prog != &dummy_bpf_prog.prog) 2278 cnt++; 2279 return cnt; 2280 } 2281 2282 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2283 { 2284 struct bpf_prog_array_item *item; 2285 2286 for (item = array->items; item->prog; item++) 2287 if (item->prog != &dummy_bpf_prog.prog) 2288 return false; 2289 return true; 2290 } 2291 2292 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2293 u32 *prog_ids, 2294 u32 request_cnt) 2295 { 2296 struct bpf_prog_array_item *item; 2297 int i = 0; 2298 2299 for (item = array->items; item->prog; item++) { 2300 if (item->prog == &dummy_bpf_prog.prog) 2301 continue; 2302 prog_ids[i] = item->prog->aux->id; 2303 if (++i == request_cnt) { 2304 item++; 2305 break; 2306 } 2307 } 2308 2309 return !!(item->prog); 2310 } 2311 2312 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2313 __u32 __user *prog_ids, u32 cnt) 2314 { 2315 unsigned long err = 0; 2316 bool nospc; 2317 u32 *ids; 2318 2319 /* users of this function are doing: 2320 * cnt = bpf_prog_array_length(); 2321 * if (cnt > 0) 2322 * bpf_prog_array_copy_to_user(..., cnt); 2323 * so below kcalloc doesn't need extra cnt > 0 check. 2324 */ 2325 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2326 if (!ids) 2327 return -ENOMEM; 2328 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2329 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2330 kfree(ids); 2331 if (err) 2332 return -EFAULT; 2333 if (nospc) 2334 return -ENOSPC; 2335 return 0; 2336 } 2337 2338 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2339 struct bpf_prog *old_prog) 2340 { 2341 struct bpf_prog_array_item *item; 2342 2343 for (item = array->items; item->prog; item++) 2344 if (item->prog == old_prog) { 2345 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2346 break; 2347 } 2348 } 2349 2350 /** 2351 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2352 * index into the program array with 2353 * a dummy no-op program. 2354 * @array: a bpf_prog_array 2355 * @index: the index of the program to replace 2356 * 2357 * Skips over dummy programs, by not counting them, when calculating 2358 * the position of the program to replace. 2359 * 2360 * Return: 2361 * * 0 - Success 2362 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2363 * * -ENOENT - Index out of range 2364 */ 2365 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2366 { 2367 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2368 } 2369 2370 /** 2371 * bpf_prog_array_update_at() - Updates the program at the given index 2372 * into the program array. 2373 * @array: a bpf_prog_array 2374 * @index: the index of the program to update 2375 * @prog: the program to insert into the array 2376 * 2377 * Skips over dummy programs, by not counting them, when calculating 2378 * the position of the program to update. 2379 * 2380 * Return: 2381 * * 0 - Success 2382 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2383 * * -ENOENT - Index out of range 2384 */ 2385 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2386 struct bpf_prog *prog) 2387 { 2388 struct bpf_prog_array_item *item; 2389 2390 if (unlikely(index < 0)) 2391 return -EINVAL; 2392 2393 for (item = array->items; item->prog; item++) { 2394 if (item->prog == &dummy_bpf_prog.prog) 2395 continue; 2396 if (!index) { 2397 WRITE_ONCE(item->prog, prog); 2398 return 0; 2399 } 2400 index--; 2401 } 2402 return -ENOENT; 2403 } 2404 2405 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2406 struct bpf_prog *exclude_prog, 2407 struct bpf_prog *include_prog, 2408 u64 bpf_cookie, 2409 struct bpf_prog_array **new_array) 2410 { 2411 int new_prog_cnt, carry_prog_cnt = 0; 2412 struct bpf_prog_array_item *existing, *new; 2413 struct bpf_prog_array *array; 2414 bool found_exclude = false; 2415 2416 /* Figure out how many existing progs we need to carry over to 2417 * the new array. 2418 */ 2419 if (old_array) { 2420 existing = old_array->items; 2421 for (; existing->prog; existing++) { 2422 if (existing->prog == exclude_prog) { 2423 found_exclude = true; 2424 continue; 2425 } 2426 if (existing->prog != &dummy_bpf_prog.prog) 2427 carry_prog_cnt++; 2428 if (existing->prog == include_prog) 2429 return -EEXIST; 2430 } 2431 } 2432 2433 if (exclude_prog && !found_exclude) 2434 return -ENOENT; 2435 2436 /* How many progs (not NULL) will be in the new array? */ 2437 new_prog_cnt = carry_prog_cnt; 2438 if (include_prog) 2439 new_prog_cnt += 1; 2440 2441 /* Do we have any prog (not NULL) in the new array? */ 2442 if (!new_prog_cnt) { 2443 *new_array = NULL; 2444 return 0; 2445 } 2446 2447 /* +1 as the end of prog_array is marked with NULL */ 2448 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2449 if (!array) 2450 return -ENOMEM; 2451 new = array->items; 2452 2453 /* Fill in the new prog array */ 2454 if (carry_prog_cnt) { 2455 existing = old_array->items; 2456 for (; existing->prog; existing++) { 2457 if (existing->prog == exclude_prog || 2458 existing->prog == &dummy_bpf_prog.prog) 2459 continue; 2460 2461 new->prog = existing->prog; 2462 new->bpf_cookie = existing->bpf_cookie; 2463 new++; 2464 } 2465 } 2466 if (include_prog) { 2467 new->prog = include_prog; 2468 new->bpf_cookie = bpf_cookie; 2469 new++; 2470 } 2471 new->prog = NULL; 2472 *new_array = array; 2473 return 0; 2474 } 2475 2476 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2477 u32 *prog_ids, u32 request_cnt, 2478 u32 *prog_cnt) 2479 { 2480 u32 cnt = 0; 2481 2482 if (array) 2483 cnt = bpf_prog_array_length(array); 2484 2485 *prog_cnt = cnt; 2486 2487 /* return early if user requested only program count or nothing to copy */ 2488 if (!request_cnt || !cnt) 2489 return 0; 2490 2491 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2492 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2493 : 0; 2494 } 2495 2496 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2497 struct bpf_map **used_maps, u32 len) 2498 { 2499 struct bpf_map *map; 2500 u32 i; 2501 2502 for (i = 0; i < len; i++) { 2503 map = used_maps[i]; 2504 if (map->ops->map_poke_untrack) 2505 map->ops->map_poke_untrack(map, aux); 2506 bpf_map_put(map); 2507 } 2508 } 2509 2510 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2511 { 2512 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2513 kfree(aux->used_maps); 2514 } 2515 2516 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2517 struct btf_mod_pair *used_btfs, u32 len) 2518 { 2519 #ifdef CONFIG_BPF_SYSCALL 2520 struct btf_mod_pair *btf_mod; 2521 u32 i; 2522 2523 for (i = 0; i < len; i++) { 2524 btf_mod = &used_btfs[i]; 2525 if (btf_mod->module) 2526 module_put(btf_mod->module); 2527 btf_put(btf_mod->btf); 2528 } 2529 #endif 2530 } 2531 2532 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2533 { 2534 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2535 kfree(aux->used_btfs); 2536 } 2537 2538 static void bpf_prog_free_deferred(struct work_struct *work) 2539 { 2540 struct bpf_prog_aux *aux; 2541 int i; 2542 2543 aux = container_of(work, struct bpf_prog_aux, work); 2544 #ifdef CONFIG_BPF_SYSCALL 2545 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2546 #endif 2547 bpf_free_used_maps(aux); 2548 bpf_free_used_btfs(aux); 2549 if (bpf_prog_is_dev_bound(aux)) 2550 bpf_prog_offload_destroy(aux->prog); 2551 #ifdef CONFIG_PERF_EVENTS 2552 if (aux->prog->has_callchain_buf) 2553 put_callchain_buffers(); 2554 #endif 2555 if (aux->dst_trampoline) 2556 bpf_trampoline_put(aux->dst_trampoline); 2557 for (i = 0; i < aux->func_cnt; i++) { 2558 /* We can just unlink the subprog poke descriptor table as 2559 * it was originally linked to the main program and is also 2560 * released along with it. 2561 */ 2562 aux->func[i]->aux->poke_tab = NULL; 2563 bpf_jit_free(aux->func[i]); 2564 } 2565 if (aux->func_cnt) { 2566 kfree(aux->func); 2567 bpf_prog_unlock_free(aux->prog); 2568 } else { 2569 bpf_jit_free(aux->prog); 2570 } 2571 } 2572 2573 void bpf_prog_free(struct bpf_prog *fp) 2574 { 2575 struct bpf_prog_aux *aux = fp->aux; 2576 2577 if (aux->dst_prog) 2578 bpf_prog_put(aux->dst_prog); 2579 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2580 schedule_work(&aux->work); 2581 } 2582 EXPORT_SYMBOL_GPL(bpf_prog_free); 2583 2584 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2585 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2586 2587 void bpf_user_rnd_init_once(void) 2588 { 2589 prandom_init_once(&bpf_user_rnd_state); 2590 } 2591 2592 BPF_CALL_0(bpf_user_rnd_u32) 2593 { 2594 /* Should someone ever have the rather unwise idea to use some 2595 * of the registers passed into this function, then note that 2596 * this function is called from native eBPF and classic-to-eBPF 2597 * transformations. Register assignments from both sides are 2598 * different, f.e. classic always sets fn(ctx, A, X) here. 2599 */ 2600 struct rnd_state *state; 2601 u32 res; 2602 2603 state = &get_cpu_var(bpf_user_rnd_state); 2604 res = prandom_u32_state(state); 2605 put_cpu_var(bpf_user_rnd_state); 2606 2607 return res; 2608 } 2609 2610 BPF_CALL_0(bpf_get_raw_cpu_id) 2611 { 2612 return raw_smp_processor_id(); 2613 } 2614 2615 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2616 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2617 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2618 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2619 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2620 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2621 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2622 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2623 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2624 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2625 2626 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2627 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2628 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2629 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2630 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2631 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2632 2633 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2634 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2635 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2636 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2637 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2638 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2639 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2640 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2641 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2642 2643 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2644 { 2645 return NULL; 2646 } 2647 2648 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2649 { 2650 return NULL; 2651 } 2652 2653 u64 __weak 2654 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2655 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2656 { 2657 return -ENOTSUPP; 2658 } 2659 EXPORT_SYMBOL_GPL(bpf_event_output); 2660 2661 /* Always built-in helper functions. */ 2662 const struct bpf_func_proto bpf_tail_call_proto = { 2663 .func = NULL, 2664 .gpl_only = false, 2665 .ret_type = RET_VOID, 2666 .arg1_type = ARG_PTR_TO_CTX, 2667 .arg2_type = ARG_CONST_MAP_PTR, 2668 .arg3_type = ARG_ANYTHING, 2669 }; 2670 2671 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2672 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2673 * eBPF and implicitly also cBPF can get JITed! 2674 */ 2675 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2676 { 2677 return prog; 2678 } 2679 2680 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2681 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2682 */ 2683 void __weak bpf_jit_compile(struct bpf_prog *prog) 2684 { 2685 } 2686 2687 bool __weak bpf_helper_changes_pkt_data(void *func) 2688 { 2689 return false; 2690 } 2691 2692 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2693 * analysis code and wants explicit zero extension inserted by verifier. 2694 * Otherwise, return FALSE. 2695 * 2696 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2697 * you don't override this. JITs that don't want these extra insns can detect 2698 * them using insn_is_zext. 2699 */ 2700 bool __weak bpf_jit_needs_zext(void) 2701 { 2702 return false; 2703 } 2704 2705 bool __weak bpf_jit_supports_kfunc_call(void) 2706 { 2707 return false; 2708 } 2709 2710 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2711 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2712 */ 2713 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2714 int len) 2715 { 2716 return -EFAULT; 2717 } 2718 2719 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2720 void *addr1, void *addr2) 2721 { 2722 return -ENOTSUPP; 2723 } 2724 2725 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 2726 { 2727 return ERR_PTR(-ENOTSUPP); 2728 } 2729 2730 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2731 EXPORT_SYMBOL(bpf_stats_enabled_key); 2732 2733 /* All definitions of tracepoints related to BPF. */ 2734 #define CREATE_TRACE_POINTS 2735 #include <linux/bpf_trace.h> 2736 2737 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2738 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2739