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 #include <linux/nospec.h> 38 #include <linux/bpf_mem_alloc.h> 39 #include <linux/memcontrol.h> 40 41 #include <asm/barrier.h> 42 #include <asm/unaligned.h> 43 44 /* Registers */ 45 #define BPF_R0 regs[BPF_REG_0] 46 #define BPF_R1 regs[BPF_REG_1] 47 #define BPF_R2 regs[BPF_REG_2] 48 #define BPF_R3 regs[BPF_REG_3] 49 #define BPF_R4 regs[BPF_REG_4] 50 #define BPF_R5 regs[BPF_REG_5] 51 #define BPF_R6 regs[BPF_REG_6] 52 #define BPF_R7 regs[BPF_REG_7] 53 #define BPF_R8 regs[BPF_REG_8] 54 #define BPF_R9 regs[BPF_REG_9] 55 #define BPF_R10 regs[BPF_REG_10] 56 57 /* Named registers */ 58 #define DST regs[insn->dst_reg] 59 #define SRC regs[insn->src_reg] 60 #define FP regs[BPF_REG_FP] 61 #define AX regs[BPF_REG_AX] 62 #define ARG1 regs[BPF_REG_ARG1] 63 #define CTX regs[BPF_REG_CTX] 64 #define OFF insn->off 65 #define IMM insn->imm 66 67 struct bpf_mem_alloc bpf_global_ma; 68 bool bpf_global_ma_set; 69 70 /* No hurry in this branch 71 * 72 * Exported for the bpf jit load helper. 73 */ 74 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 75 { 76 u8 *ptr = NULL; 77 78 if (k >= SKF_NET_OFF) { 79 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 80 } else if (k >= SKF_LL_OFF) { 81 if (unlikely(!skb_mac_header_was_set(skb))) 82 return NULL; 83 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 84 } 85 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 86 return ptr; 87 88 return NULL; 89 } 90 91 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 92 { 93 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 94 struct bpf_prog_aux *aux; 95 struct bpf_prog *fp; 96 97 size = round_up(size, PAGE_SIZE); 98 fp = __vmalloc(size, gfp_flags); 99 if (fp == NULL) 100 return NULL; 101 102 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 103 if (aux == NULL) { 104 vfree(fp); 105 return NULL; 106 } 107 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 108 if (!fp->active) { 109 vfree(fp); 110 kfree(aux); 111 return NULL; 112 } 113 114 fp->pages = size / PAGE_SIZE; 115 fp->aux = aux; 116 fp->aux->prog = fp; 117 fp->jit_requested = ebpf_jit_enabled(); 118 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 119 #ifdef CONFIG_CGROUP_BPF 120 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 121 #endif 122 123 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 124 mutex_init(&fp->aux->used_maps_mutex); 125 mutex_init(&fp->aux->dst_mutex); 126 127 return fp; 128 } 129 130 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 131 { 132 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 133 struct bpf_prog *prog; 134 int cpu; 135 136 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 137 if (!prog) 138 return NULL; 139 140 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 141 if (!prog->stats) { 142 free_percpu(prog->active); 143 kfree(prog->aux); 144 vfree(prog); 145 return NULL; 146 } 147 148 for_each_possible_cpu(cpu) { 149 struct bpf_prog_stats *pstats; 150 151 pstats = per_cpu_ptr(prog->stats, cpu); 152 u64_stats_init(&pstats->syncp); 153 } 154 return prog; 155 } 156 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 157 158 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 159 { 160 if (!prog->aux->nr_linfo || !prog->jit_requested) 161 return 0; 162 163 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 164 sizeof(*prog->aux->jited_linfo), 165 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 166 if (!prog->aux->jited_linfo) 167 return -ENOMEM; 168 169 return 0; 170 } 171 172 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 173 { 174 if (prog->aux->jited_linfo && 175 (!prog->jited || !prog->aux->jited_linfo[0])) { 176 kvfree(prog->aux->jited_linfo); 177 prog->aux->jited_linfo = NULL; 178 } 179 180 kfree(prog->aux->kfunc_tab); 181 prog->aux->kfunc_tab = NULL; 182 } 183 184 /* The jit engine is responsible to provide an array 185 * for insn_off to the jited_off mapping (insn_to_jit_off). 186 * 187 * The idx to this array is the insn_off. Hence, the insn_off 188 * here is relative to the prog itself instead of the main prog. 189 * This array has one entry for each xlated bpf insn. 190 * 191 * jited_off is the byte off to the end of the jited insn. 192 * 193 * Hence, with 194 * insn_start: 195 * The first bpf insn off of the prog. The insn off 196 * here is relative to the main prog. 197 * e.g. if prog is a subprog, insn_start > 0 198 * linfo_idx: 199 * The prog's idx to prog->aux->linfo and jited_linfo 200 * 201 * jited_linfo[linfo_idx] = prog->bpf_func 202 * 203 * For i > linfo_idx, 204 * 205 * jited_linfo[i] = prog->bpf_func + 206 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 207 */ 208 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 209 const u32 *insn_to_jit_off) 210 { 211 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 212 const struct bpf_line_info *linfo; 213 void **jited_linfo; 214 215 if (!prog->aux->jited_linfo) 216 /* Userspace did not provide linfo */ 217 return; 218 219 linfo_idx = prog->aux->linfo_idx; 220 linfo = &prog->aux->linfo[linfo_idx]; 221 insn_start = linfo[0].insn_off; 222 insn_end = insn_start + prog->len; 223 224 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 225 jited_linfo[0] = prog->bpf_func; 226 227 nr_linfo = prog->aux->nr_linfo - linfo_idx; 228 229 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 230 /* The verifier ensures that linfo[i].insn_off is 231 * strictly increasing 232 */ 233 jited_linfo[i] = prog->bpf_func + 234 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 235 } 236 237 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 238 gfp_t gfp_extra_flags) 239 { 240 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 241 struct bpf_prog *fp; 242 u32 pages; 243 244 size = round_up(size, PAGE_SIZE); 245 pages = size / PAGE_SIZE; 246 if (pages <= fp_old->pages) 247 return fp_old; 248 249 fp = __vmalloc(size, gfp_flags); 250 if (fp) { 251 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 252 fp->pages = pages; 253 fp->aux->prog = fp; 254 255 /* We keep fp->aux from fp_old around in the new 256 * reallocated structure. 257 */ 258 fp_old->aux = NULL; 259 fp_old->stats = NULL; 260 fp_old->active = NULL; 261 __bpf_prog_free(fp_old); 262 } 263 264 return fp; 265 } 266 267 void __bpf_prog_free(struct bpf_prog *fp) 268 { 269 if (fp->aux) { 270 mutex_destroy(&fp->aux->used_maps_mutex); 271 mutex_destroy(&fp->aux->dst_mutex); 272 kfree(fp->aux->poke_tab); 273 kfree(fp->aux); 274 } 275 free_percpu(fp->stats); 276 free_percpu(fp->active); 277 vfree(fp); 278 } 279 280 int bpf_prog_calc_tag(struct bpf_prog *fp) 281 { 282 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 283 u32 raw_size = bpf_prog_tag_scratch_size(fp); 284 u32 digest[SHA1_DIGEST_WORDS]; 285 u32 ws[SHA1_WORKSPACE_WORDS]; 286 u32 i, bsize, psize, blocks; 287 struct bpf_insn *dst; 288 bool was_ld_map; 289 u8 *raw, *todo; 290 __be32 *result; 291 __be64 *bits; 292 293 raw = vmalloc(raw_size); 294 if (!raw) 295 return -ENOMEM; 296 297 sha1_init(digest); 298 memset(ws, 0, sizeof(ws)); 299 300 /* We need to take out the map fd for the digest calculation 301 * since they are unstable from user space side. 302 */ 303 dst = (void *)raw; 304 for (i = 0, was_ld_map = false; i < fp->len; i++) { 305 dst[i] = fp->insnsi[i]; 306 if (!was_ld_map && 307 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 308 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 309 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 310 was_ld_map = true; 311 dst[i].imm = 0; 312 } else if (was_ld_map && 313 dst[i].code == 0 && 314 dst[i].dst_reg == 0 && 315 dst[i].src_reg == 0 && 316 dst[i].off == 0) { 317 was_ld_map = false; 318 dst[i].imm = 0; 319 } else { 320 was_ld_map = false; 321 } 322 } 323 324 psize = bpf_prog_insn_size(fp); 325 memset(&raw[psize], 0, raw_size - psize); 326 raw[psize++] = 0x80; 327 328 bsize = round_up(psize, SHA1_BLOCK_SIZE); 329 blocks = bsize / SHA1_BLOCK_SIZE; 330 todo = raw; 331 if (bsize - psize >= sizeof(__be64)) { 332 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 333 } else { 334 bits = (__be64 *)(todo + bsize + bits_offset); 335 blocks++; 336 } 337 *bits = cpu_to_be64((psize - 1) << 3); 338 339 while (blocks--) { 340 sha1_transform(digest, todo, ws); 341 todo += SHA1_BLOCK_SIZE; 342 } 343 344 result = (__force __be32 *)digest; 345 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 346 result[i] = cpu_to_be32(digest[i]); 347 memcpy(fp->tag, result, sizeof(fp->tag)); 348 349 vfree(raw); 350 return 0; 351 } 352 353 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 354 s32 end_new, s32 curr, const bool probe_pass) 355 { 356 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 357 s32 delta = end_new - end_old; 358 s64 imm = insn->imm; 359 360 if (curr < pos && curr + imm + 1 >= end_old) 361 imm += delta; 362 else if (curr >= end_new && curr + imm + 1 < end_new) 363 imm -= delta; 364 if (imm < imm_min || imm > imm_max) 365 return -ERANGE; 366 if (!probe_pass) 367 insn->imm = imm; 368 return 0; 369 } 370 371 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 372 s32 end_new, s32 curr, const bool probe_pass) 373 { 374 s64 off_min, off_max, off; 375 s32 delta = end_new - end_old; 376 377 if (insn->code == (BPF_JMP32 | BPF_JA)) { 378 off = insn->imm; 379 off_min = S32_MIN; 380 off_max = S32_MAX; 381 } else { 382 off = insn->off; 383 off_min = S16_MIN; 384 off_max = S16_MAX; 385 } 386 387 if (curr < pos && curr + off + 1 >= end_old) 388 off += delta; 389 else if (curr >= end_new && curr + off + 1 < end_new) 390 off -= delta; 391 if (off < off_min || off > off_max) 392 return -ERANGE; 393 if (!probe_pass) { 394 if (insn->code == (BPF_JMP32 | BPF_JA)) 395 insn->imm = off; 396 else 397 insn->off = off; 398 } 399 return 0; 400 } 401 402 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 403 s32 end_new, const bool probe_pass) 404 { 405 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 406 struct bpf_insn *insn = prog->insnsi; 407 int ret = 0; 408 409 for (i = 0; i < insn_cnt; i++, insn++) { 410 u8 code; 411 412 /* In the probing pass we still operate on the original, 413 * unpatched image in order to check overflows before we 414 * do any other adjustments. Therefore skip the patchlet. 415 */ 416 if (probe_pass && i == pos) { 417 i = end_new; 418 insn = prog->insnsi + end_old; 419 } 420 if (bpf_pseudo_func(insn)) { 421 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 422 end_new, i, probe_pass); 423 if (ret) 424 return ret; 425 continue; 426 } 427 code = insn->code; 428 if ((BPF_CLASS(code) != BPF_JMP && 429 BPF_CLASS(code) != BPF_JMP32) || 430 BPF_OP(code) == BPF_EXIT) 431 continue; 432 /* Adjust offset of jmps if we cross patch boundaries. */ 433 if (BPF_OP(code) == BPF_CALL) { 434 if (insn->src_reg != BPF_PSEUDO_CALL) 435 continue; 436 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 437 end_new, i, probe_pass); 438 } else { 439 ret = bpf_adj_delta_to_off(insn, pos, end_old, 440 end_new, i, probe_pass); 441 } 442 if (ret) 443 break; 444 } 445 446 return ret; 447 } 448 449 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 450 { 451 struct bpf_line_info *linfo; 452 u32 i, nr_linfo; 453 454 nr_linfo = prog->aux->nr_linfo; 455 if (!nr_linfo || !delta) 456 return; 457 458 linfo = prog->aux->linfo; 459 460 for (i = 0; i < nr_linfo; i++) 461 if (off < linfo[i].insn_off) 462 break; 463 464 /* Push all off < linfo[i].insn_off by delta */ 465 for (; i < nr_linfo; i++) 466 linfo[i].insn_off += delta; 467 } 468 469 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 470 const struct bpf_insn *patch, u32 len) 471 { 472 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 473 const u32 cnt_max = S16_MAX; 474 struct bpf_prog *prog_adj; 475 int err; 476 477 /* Since our patchlet doesn't expand the image, we're done. */ 478 if (insn_delta == 0) { 479 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 480 return prog; 481 } 482 483 insn_adj_cnt = prog->len + insn_delta; 484 485 /* Reject anything that would potentially let the insn->off 486 * target overflow when we have excessive program expansions. 487 * We need to probe here before we do any reallocation where 488 * we afterwards may not fail anymore. 489 */ 490 if (insn_adj_cnt > cnt_max && 491 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 492 return ERR_PTR(err); 493 494 /* Several new instructions need to be inserted. Make room 495 * for them. Likely, there's no need for a new allocation as 496 * last page could have large enough tailroom. 497 */ 498 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 499 GFP_USER); 500 if (!prog_adj) 501 return ERR_PTR(-ENOMEM); 502 503 prog_adj->len = insn_adj_cnt; 504 505 /* Patching happens in 3 steps: 506 * 507 * 1) Move over tail of insnsi from next instruction onwards, 508 * so we can patch the single target insn with one or more 509 * new ones (patching is always from 1 to n insns, n > 0). 510 * 2) Inject new instructions at the target location. 511 * 3) Adjust branch offsets if necessary. 512 */ 513 insn_rest = insn_adj_cnt - off - len; 514 515 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 516 sizeof(*patch) * insn_rest); 517 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 518 519 /* We are guaranteed to not fail at this point, otherwise 520 * the ship has sailed to reverse to the original state. An 521 * overflow cannot happen at this point. 522 */ 523 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 524 525 bpf_adj_linfo(prog_adj, off, insn_delta); 526 527 return prog_adj; 528 } 529 530 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 531 { 532 /* Branch offsets can't overflow when program is shrinking, no need 533 * to call bpf_adj_branches(..., true) here 534 */ 535 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 536 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 537 prog->len -= cnt; 538 539 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 540 } 541 542 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 543 { 544 int i; 545 546 for (i = 0; i < fp->aux->func_cnt; i++) 547 bpf_prog_kallsyms_del(fp->aux->func[i]); 548 } 549 550 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 551 { 552 bpf_prog_kallsyms_del_subprogs(fp); 553 bpf_prog_kallsyms_del(fp); 554 } 555 556 #ifdef CONFIG_BPF_JIT 557 /* All BPF JIT sysctl knobs here. */ 558 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 559 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 560 int bpf_jit_harden __read_mostly; 561 long bpf_jit_limit __read_mostly; 562 long bpf_jit_limit_max __read_mostly; 563 564 static void 565 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 566 { 567 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 568 569 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 570 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 571 } 572 573 static void 574 bpf_prog_ksym_set_name(struct bpf_prog *prog) 575 { 576 char *sym = prog->aux->ksym.name; 577 const char *end = sym + KSYM_NAME_LEN; 578 const struct btf_type *type; 579 const char *func_name; 580 581 BUILD_BUG_ON(sizeof("bpf_prog_") + 582 sizeof(prog->tag) * 2 + 583 /* name has been null terminated. 584 * We should need +1 for the '_' preceding 585 * the name. However, the null character 586 * is double counted between the name and the 587 * sizeof("bpf_prog_") above, so we omit 588 * the +1 here. 589 */ 590 sizeof(prog->aux->name) > KSYM_NAME_LEN); 591 592 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 593 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 594 595 /* prog->aux->name will be ignored if full btf name is available */ 596 if (prog->aux->func_info_cnt) { 597 type = btf_type_by_id(prog->aux->btf, 598 prog->aux->func_info[prog->aux->func_idx].type_id); 599 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 600 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 601 return; 602 } 603 604 if (prog->aux->name[0]) 605 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 606 else 607 *sym = 0; 608 } 609 610 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 611 { 612 return container_of(n, struct bpf_ksym, tnode)->start; 613 } 614 615 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 616 struct latch_tree_node *b) 617 { 618 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 619 } 620 621 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 622 { 623 unsigned long val = (unsigned long)key; 624 const struct bpf_ksym *ksym; 625 626 ksym = container_of(n, struct bpf_ksym, tnode); 627 628 if (val < ksym->start) 629 return -1; 630 /* Ensure that we detect return addresses as part of the program, when 631 * the final instruction is a call for a program part of the stack 632 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 633 */ 634 if (val > ksym->end) 635 return 1; 636 637 return 0; 638 } 639 640 static const struct latch_tree_ops bpf_tree_ops = { 641 .less = bpf_tree_less, 642 .comp = bpf_tree_comp, 643 }; 644 645 static DEFINE_SPINLOCK(bpf_lock); 646 static LIST_HEAD(bpf_kallsyms); 647 static struct latch_tree_root bpf_tree __cacheline_aligned; 648 649 void bpf_ksym_add(struct bpf_ksym *ksym) 650 { 651 spin_lock_bh(&bpf_lock); 652 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 653 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 654 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 655 spin_unlock_bh(&bpf_lock); 656 } 657 658 static void __bpf_ksym_del(struct bpf_ksym *ksym) 659 { 660 if (list_empty(&ksym->lnode)) 661 return; 662 663 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 664 list_del_rcu(&ksym->lnode); 665 } 666 667 void bpf_ksym_del(struct bpf_ksym *ksym) 668 { 669 spin_lock_bh(&bpf_lock); 670 __bpf_ksym_del(ksym); 671 spin_unlock_bh(&bpf_lock); 672 } 673 674 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 675 { 676 return fp->jited && !bpf_prog_was_classic(fp); 677 } 678 679 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 680 { 681 if (!bpf_prog_kallsyms_candidate(fp) || 682 !bpf_capable()) 683 return; 684 685 bpf_prog_ksym_set_addr(fp); 686 bpf_prog_ksym_set_name(fp); 687 fp->aux->ksym.prog = true; 688 689 bpf_ksym_add(&fp->aux->ksym); 690 } 691 692 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 693 { 694 if (!bpf_prog_kallsyms_candidate(fp)) 695 return; 696 697 bpf_ksym_del(&fp->aux->ksym); 698 } 699 700 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 701 { 702 struct latch_tree_node *n; 703 704 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 705 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 706 } 707 708 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 709 unsigned long *off, char *sym) 710 { 711 struct bpf_ksym *ksym; 712 char *ret = NULL; 713 714 rcu_read_lock(); 715 ksym = bpf_ksym_find(addr); 716 if (ksym) { 717 unsigned long symbol_start = ksym->start; 718 unsigned long symbol_end = ksym->end; 719 720 strncpy(sym, ksym->name, KSYM_NAME_LEN); 721 722 ret = sym; 723 if (size) 724 *size = symbol_end - symbol_start; 725 if (off) 726 *off = addr - symbol_start; 727 } 728 rcu_read_unlock(); 729 730 return ret; 731 } 732 733 bool is_bpf_text_address(unsigned long addr) 734 { 735 bool ret; 736 737 rcu_read_lock(); 738 ret = bpf_ksym_find(addr) != NULL; 739 rcu_read_unlock(); 740 741 return ret; 742 } 743 744 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 745 { 746 struct bpf_ksym *ksym = bpf_ksym_find(addr); 747 748 return ksym && ksym->prog ? 749 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 750 NULL; 751 } 752 753 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 754 { 755 const struct exception_table_entry *e = NULL; 756 struct bpf_prog *prog; 757 758 rcu_read_lock(); 759 prog = bpf_prog_ksym_find(addr); 760 if (!prog) 761 goto out; 762 if (!prog->aux->num_exentries) 763 goto out; 764 765 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 766 out: 767 rcu_read_unlock(); 768 return e; 769 } 770 771 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 772 char *sym) 773 { 774 struct bpf_ksym *ksym; 775 unsigned int it = 0; 776 int ret = -ERANGE; 777 778 if (!bpf_jit_kallsyms_enabled()) 779 return ret; 780 781 rcu_read_lock(); 782 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 783 if (it++ != symnum) 784 continue; 785 786 strncpy(sym, ksym->name, KSYM_NAME_LEN); 787 788 *value = ksym->start; 789 *type = BPF_SYM_ELF_TYPE; 790 791 ret = 0; 792 break; 793 } 794 rcu_read_unlock(); 795 796 return ret; 797 } 798 799 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 800 struct bpf_jit_poke_descriptor *poke) 801 { 802 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 803 static const u32 poke_tab_max = 1024; 804 u32 slot = prog->aux->size_poke_tab; 805 u32 size = slot + 1; 806 807 if (size > poke_tab_max) 808 return -ENOSPC; 809 if (poke->tailcall_target || poke->tailcall_target_stable || 810 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 811 return -EINVAL; 812 813 switch (poke->reason) { 814 case BPF_POKE_REASON_TAIL_CALL: 815 if (!poke->tail_call.map) 816 return -EINVAL; 817 break; 818 default: 819 return -EINVAL; 820 } 821 822 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL); 823 if (!tab) 824 return -ENOMEM; 825 826 memcpy(&tab[slot], poke, sizeof(*poke)); 827 prog->aux->size_poke_tab = size; 828 prog->aux->poke_tab = tab; 829 830 return slot; 831 } 832 833 /* 834 * BPF program pack allocator. 835 * 836 * Most BPF programs are pretty small. Allocating a hole page for each 837 * program is sometime a waste. Many small bpf program also adds pressure 838 * to instruction TLB. To solve this issue, we introduce a BPF program pack 839 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 840 * to host BPF programs. 841 */ 842 #define BPF_PROG_CHUNK_SHIFT 6 843 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 844 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 845 846 struct bpf_prog_pack { 847 struct list_head list; 848 void *ptr; 849 unsigned long bitmap[]; 850 }; 851 852 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 853 { 854 memset(area, 0, size); 855 } 856 857 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 858 859 static DEFINE_MUTEX(pack_mutex); 860 static LIST_HEAD(pack_list); 861 862 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 863 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 864 */ 865 #ifdef PMD_SIZE 866 /* PMD_SIZE is really big for some archs. It doesn't make sense to 867 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 868 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 869 * greater than or equal to 2MB. 870 */ 871 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 872 #else 873 #define BPF_PROG_PACK_SIZE PAGE_SIZE 874 #endif 875 876 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 877 878 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 879 { 880 struct bpf_prog_pack *pack; 881 882 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 883 GFP_KERNEL); 884 if (!pack) 885 return NULL; 886 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 887 if (!pack->ptr) { 888 kfree(pack); 889 return NULL; 890 } 891 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 892 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 893 list_add_tail(&pack->list, &pack_list); 894 895 set_vm_flush_reset_perms(pack->ptr); 896 set_memory_rox((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE); 897 return pack; 898 } 899 900 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 901 { 902 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 903 struct bpf_prog_pack *pack; 904 unsigned long pos; 905 void *ptr = NULL; 906 907 mutex_lock(&pack_mutex); 908 if (size > BPF_PROG_PACK_SIZE) { 909 size = round_up(size, PAGE_SIZE); 910 ptr = bpf_jit_alloc_exec(size); 911 if (ptr) { 912 bpf_fill_ill_insns(ptr, size); 913 set_vm_flush_reset_perms(ptr); 914 set_memory_rox((unsigned long)ptr, size / PAGE_SIZE); 915 } 916 goto out; 917 } 918 list_for_each_entry(pack, &pack_list, list) { 919 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 920 nbits, 0); 921 if (pos < BPF_PROG_CHUNK_COUNT) 922 goto found_free_area; 923 } 924 925 pack = alloc_new_pack(bpf_fill_ill_insns); 926 if (!pack) 927 goto out; 928 929 pos = 0; 930 931 found_free_area: 932 bitmap_set(pack->bitmap, pos, nbits); 933 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 934 935 out: 936 mutex_unlock(&pack_mutex); 937 return ptr; 938 } 939 940 void bpf_prog_pack_free(struct bpf_binary_header *hdr) 941 { 942 struct bpf_prog_pack *pack = NULL, *tmp; 943 unsigned int nbits; 944 unsigned long pos; 945 946 mutex_lock(&pack_mutex); 947 if (hdr->size > BPF_PROG_PACK_SIZE) { 948 bpf_jit_free_exec(hdr); 949 goto out; 950 } 951 952 list_for_each_entry(tmp, &pack_list, list) { 953 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) { 954 pack = tmp; 955 break; 956 } 957 } 958 959 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 960 goto out; 961 962 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size); 963 pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 964 965 WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size), 966 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 967 968 bitmap_clear(pack->bitmap, pos, nbits); 969 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 970 BPF_PROG_CHUNK_COUNT, 0) == 0) { 971 list_del(&pack->list); 972 bpf_jit_free_exec(pack->ptr); 973 kfree(pack); 974 } 975 out: 976 mutex_unlock(&pack_mutex); 977 } 978 979 static atomic_long_t bpf_jit_current; 980 981 /* Can be overridden by an arch's JIT compiler if it has a custom, 982 * dedicated BPF backend memory area, or if neither of the two 983 * below apply. 984 */ 985 u64 __weak bpf_jit_alloc_exec_limit(void) 986 { 987 #if defined(MODULES_VADDR) 988 return MODULES_END - MODULES_VADDR; 989 #else 990 return VMALLOC_END - VMALLOC_START; 991 #endif 992 } 993 994 static int __init bpf_jit_charge_init(void) 995 { 996 /* Only used as heuristic here to derive limit. */ 997 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 998 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 999 PAGE_SIZE), LONG_MAX); 1000 return 0; 1001 } 1002 pure_initcall(bpf_jit_charge_init); 1003 1004 int bpf_jit_charge_modmem(u32 size) 1005 { 1006 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1007 if (!bpf_capable()) { 1008 atomic_long_sub(size, &bpf_jit_current); 1009 return -EPERM; 1010 } 1011 } 1012 1013 return 0; 1014 } 1015 1016 void bpf_jit_uncharge_modmem(u32 size) 1017 { 1018 atomic_long_sub(size, &bpf_jit_current); 1019 } 1020 1021 void *__weak bpf_jit_alloc_exec(unsigned long size) 1022 { 1023 return module_alloc(size); 1024 } 1025 1026 void __weak bpf_jit_free_exec(void *addr) 1027 { 1028 module_memfree(addr); 1029 } 1030 1031 struct bpf_binary_header * 1032 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1033 unsigned int alignment, 1034 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1035 { 1036 struct bpf_binary_header *hdr; 1037 u32 size, hole, start; 1038 1039 WARN_ON_ONCE(!is_power_of_2(alignment) || 1040 alignment > BPF_IMAGE_ALIGNMENT); 1041 1042 /* Most of BPF filters are really small, but if some of them 1043 * fill a page, allow at least 128 extra bytes to insert a 1044 * random section of illegal instructions. 1045 */ 1046 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1047 1048 if (bpf_jit_charge_modmem(size)) 1049 return NULL; 1050 hdr = bpf_jit_alloc_exec(size); 1051 if (!hdr) { 1052 bpf_jit_uncharge_modmem(size); 1053 return NULL; 1054 } 1055 1056 /* Fill space with illegal/arch-dep instructions. */ 1057 bpf_fill_ill_insns(hdr, size); 1058 1059 hdr->size = size; 1060 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1061 PAGE_SIZE - sizeof(*hdr)); 1062 start = get_random_u32_below(hole) & ~(alignment - 1); 1063 1064 /* Leave a random number of instructions before BPF code. */ 1065 *image_ptr = &hdr->image[start]; 1066 1067 return hdr; 1068 } 1069 1070 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1071 { 1072 u32 size = hdr->size; 1073 1074 bpf_jit_free_exec(hdr); 1075 bpf_jit_uncharge_modmem(size); 1076 } 1077 1078 /* Allocate jit binary from bpf_prog_pack allocator. 1079 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1080 * to the memory. To solve this problem, a RW buffer is also allocated at 1081 * as the same time. The JIT engine should calculate offsets based on the 1082 * RO memory address, but write JITed program to the RW buffer. Once the 1083 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1084 * the JITed program to the RO memory. 1085 */ 1086 struct bpf_binary_header * 1087 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1088 unsigned int alignment, 1089 struct bpf_binary_header **rw_header, 1090 u8 **rw_image, 1091 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1092 { 1093 struct bpf_binary_header *ro_header; 1094 u32 size, hole, start; 1095 1096 WARN_ON_ONCE(!is_power_of_2(alignment) || 1097 alignment > BPF_IMAGE_ALIGNMENT); 1098 1099 /* add 16 bytes for a random section of illegal instructions */ 1100 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1101 1102 if (bpf_jit_charge_modmem(size)) 1103 return NULL; 1104 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1105 if (!ro_header) { 1106 bpf_jit_uncharge_modmem(size); 1107 return NULL; 1108 } 1109 1110 *rw_header = kvmalloc(size, GFP_KERNEL); 1111 if (!*rw_header) { 1112 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size)); 1113 bpf_prog_pack_free(ro_header); 1114 bpf_jit_uncharge_modmem(size); 1115 return NULL; 1116 } 1117 1118 /* Fill space with illegal/arch-dep instructions. */ 1119 bpf_fill_ill_insns(*rw_header, size); 1120 (*rw_header)->size = size; 1121 1122 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1123 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1124 start = get_random_u32_below(hole) & ~(alignment - 1); 1125 1126 *image_ptr = &ro_header->image[start]; 1127 *rw_image = &(*rw_header)->image[start]; 1128 1129 return ro_header; 1130 } 1131 1132 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1133 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog, 1134 struct bpf_binary_header *ro_header, 1135 struct bpf_binary_header *rw_header) 1136 { 1137 void *ptr; 1138 1139 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1140 1141 kvfree(rw_header); 1142 1143 if (IS_ERR(ptr)) { 1144 bpf_prog_pack_free(ro_header); 1145 return PTR_ERR(ptr); 1146 } 1147 return 0; 1148 } 1149 1150 /* bpf_jit_binary_pack_free is called in two different scenarios: 1151 * 1) when the program is freed after; 1152 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1153 * For case 2), we need to free both the RO memory and the RW buffer. 1154 * 1155 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1156 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1157 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1158 * bpf_arch_text_copy (when jit fails). 1159 */ 1160 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1161 struct bpf_binary_header *rw_header) 1162 { 1163 u32 size = ro_header->size; 1164 1165 bpf_prog_pack_free(ro_header); 1166 kvfree(rw_header); 1167 bpf_jit_uncharge_modmem(size); 1168 } 1169 1170 struct bpf_binary_header * 1171 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1172 { 1173 unsigned long real_start = (unsigned long)fp->bpf_func; 1174 unsigned long addr; 1175 1176 addr = real_start & BPF_PROG_CHUNK_MASK; 1177 return (void *)addr; 1178 } 1179 1180 static inline struct bpf_binary_header * 1181 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1182 { 1183 unsigned long real_start = (unsigned long)fp->bpf_func; 1184 unsigned long addr; 1185 1186 addr = real_start & PAGE_MASK; 1187 return (void *)addr; 1188 } 1189 1190 /* This symbol is only overridden by archs that have different 1191 * requirements than the usual eBPF JITs, f.e. when they only 1192 * implement cBPF JIT, do not set images read-only, etc. 1193 */ 1194 void __weak bpf_jit_free(struct bpf_prog *fp) 1195 { 1196 if (fp->jited) { 1197 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1198 1199 bpf_jit_binary_free(hdr); 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 int err; 1214 1215 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1216 if (!*func_addr_fixed) { 1217 /* Place-holder address till the last pass has collected 1218 * all addresses for JITed subprograms in which case we 1219 * can pick them up from prog->aux. 1220 */ 1221 if (!extra_pass) 1222 addr = NULL; 1223 else if (prog->aux->func && 1224 off >= 0 && off < prog->aux->func_cnt) 1225 addr = (u8 *)prog->aux->func[off]->bpf_func; 1226 else 1227 return -EINVAL; 1228 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1229 bpf_jit_supports_far_kfunc_call()) { 1230 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1231 if (err) 1232 return err; 1233 } else { 1234 /* Address of a BPF helper call. Since part of the core 1235 * kernel, it's always at a fixed location. __bpf_call_base 1236 * and the helper with imm relative to it are both in core 1237 * kernel. 1238 */ 1239 addr = (u8 *)__bpf_call_base + imm; 1240 } 1241 1242 *func_addr = (unsigned long)addr; 1243 return 0; 1244 } 1245 1246 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1247 const struct bpf_insn *aux, 1248 struct bpf_insn *to_buff, 1249 bool emit_zext) 1250 { 1251 struct bpf_insn *to = to_buff; 1252 u32 imm_rnd = get_random_u32(); 1253 s16 off; 1254 1255 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1256 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1257 1258 /* Constraints on AX register: 1259 * 1260 * AX register is inaccessible from user space. It is mapped in 1261 * all JITs, and used here for constant blinding rewrites. It is 1262 * typically "stateless" meaning its contents are only valid within 1263 * the executed instruction, but not across several instructions. 1264 * There are a few exceptions however which are further detailed 1265 * below. 1266 * 1267 * Constant blinding is only used by JITs, not in the interpreter. 1268 * The interpreter uses AX in some occasions as a local temporary 1269 * register e.g. in DIV or MOD instructions. 1270 * 1271 * In restricted circumstances, the verifier can also use the AX 1272 * register for rewrites as long as they do not interfere with 1273 * the above cases! 1274 */ 1275 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1276 goto out; 1277 1278 if (from->imm == 0 && 1279 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1280 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1281 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1282 goto out; 1283 } 1284 1285 switch (from->code) { 1286 case BPF_ALU | BPF_ADD | BPF_K: 1287 case BPF_ALU | BPF_SUB | BPF_K: 1288 case BPF_ALU | BPF_AND | BPF_K: 1289 case BPF_ALU | BPF_OR | BPF_K: 1290 case BPF_ALU | BPF_XOR | BPF_K: 1291 case BPF_ALU | BPF_MUL | BPF_K: 1292 case BPF_ALU | BPF_MOV | BPF_K: 1293 case BPF_ALU | BPF_DIV | BPF_K: 1294 case BPF_ALU | BPF_MOD | BPF_K: 1295 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1296 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1297 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1298 break; 1299 1300 case BPF_ALU64 | BPF_ADD | BPF_K: 1301 case BPF_ALU64 | BPF_SUB | BPF_K: 1302 case BPF_ALU64 | BPF_AND | BPF_K: 1303 case BPF_ALU64 | BPF_OR | BPF_K: 1304 case BPF_ALU64 | BPF_XOR | BPF_K: 1305 case BPF_ALU64 | BPF_MUL | BPF_K: 1306 case BPF_ALU64 | BPF_MOV | BPF_K: 1307 case BPF_ALU64 | BPF_DIV | BPF_K: 1308 case BPF_ALU64 | BPF_MOD | BPF_K: 1309 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1310 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1311 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1312 break; 1313 1314 case BPF_JMP | BPF_JEQ | BPF_K: 1315 case BPF_JMP | BPF_JNE | BPF_K: 1316 case BPF_JMP | BPF_JGT | BPF_K: 1317 case BPF_JMP | BPF_JLT | BPF_K: 1318 case BPF_JMP | BPF_JGE | BPF_K: 1319 case BPF_JMP | BPF_JLE | BPF_K: 1320 case BPF_JMP | BPF_JSGT | BPF_K: 1321 case BPF_JMP | BPF_JSLT | BPF_K: 1322 case BPF_JMP | BPF_JSGE | BPF_K: 1323 case BPF_JMP | BPF_JSLE | BPF_K: 1324 case BPF_JMP | BPF_JSET | BPF_K: 1325 /* Accommodate for extra offset in case of a backjump. */ 1326 off = from->off; 1327 if (off < 0) 1328 off -= 2; 1329 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1330 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1331 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1332 break; 1333 1334 case BPF_JMP32 | BPF_JEQ | BPF_K: 1335 case BPF_JMP32 | BPF_JNE | BPF_K: 1336 case BPF_JMP32 | BPF_JGT | BPF_K: 1337 case BPF_JMP32 | BPF_JLT | BPF_K: 1338 case BPF_JMP32 | BPF_JGE | BPF_K: 1339 case BPF_JMP32 | BPF_JLE | BPF_K: 1340 case BPF_JMP32 | BPF_JSGT | BPF_K: 1341 case BPF_JMP32 | BPF_JSLT | BPF_K: 1342 case BPF_JMP32 | BPF_JSGE | BPF_K: 1343 case BPF_JMP32 | BPF_JSLE | BPF_K: 1344 case BPF_JMP32 | BPF_JSET | BPF_K: 1345 /* Accommodate for extra offset in case of a backjump. */ 1346 off = from->off; 1347 if (off < 0) 1348 off -= 2; 1349 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1350 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1351 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1352 off); 1353 break; 1354 1355 case BPF_LD | BPF_IMM | BPF_DW: 1356 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1357 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1358 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1359 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1360 break; 1361 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1362 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1363 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1364 if (emit_zext) 1365 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1366 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1367 break; 1368 1369 case BPF_ST | BPF_MEM | BPF_DW: 1370 case BPF_ST | BPF_MEM | BPF_W: 1371 case BPF_ST | BPF_MEM | BPF_H: 1372 case BPF_ST | BPF_MEM | BPF_B: 1373 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1374 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1375 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1376 break; 1377 } 1378 out: 1379 return to - to_buff; 1380 } 1381 1382 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1383 gfp_t gfp_extra_flags) 1384 { 1385 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1386 struct bpf_prog *fp; 1387 1388 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1389 if (fp != NULL) { 1390 /* aux->prog still points to the fp_other one, so 1391 * when promoting the clone to the real program, 1392 * this still needs to be adapted. 1393 */ 1394 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1395 } 1396 1397 return fp; 1398 } 1399 1400 static void bpf_prog_clone_free(struct bpf_prog *fp) 1401 { 1402 /* aux was stolen by the other clone, so we cannot free 1403 * it from this path! It will be freed eventually by the 1404 * other program on release. 1405 * 1406 * At this point, we don't need a deferred release since 1407 * clone is guaranteed to not be locked. 1408 */ 1409 fp->aux = NULL; 1410 fp->stats = NULL; 1411 fp->active = NULL; 1412 __bpf_prog_free(fp); 1413 } 1414 1415 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1416 { 1417 /* We have to repoint aux->prog to self, as we don't 1418 * know whether fp here is the clone or the original. 1419 */ 1420 fp->aux->prog = fp; 1421 bpf_prog_clone_free(fp_other); 1422 } 1423 1424 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1425 { 1426 struct bpf_insn insn_buff[16], aux[2]; 1427 struct bpf_prog *clone, *tmp; 1428 int insn_delta, insn_cnt; 1429 struct bpf_insn *insn; 1430 int i, rewritten; 1431 1432 if (!prog->blinding_requested || prog->blinded) 1433 return prog; 1434 1435 clone = bpf_prog_clone_create(prog, GFP_USER); 1436 if (!clone) 1437 return ERR_PTR(-ENOMEM); 1438 1439 insn_cnt = clone->len; 1440 insn = clone->insnsi; 1441 1442 for (i = 0; i < insn_cnt; i++, insn++) { 1443 if (bpf_pseudo_func(insn)) { 1444 /* ld_imm64 with an address of bpf subprog is not 1445 * a user controlled constant. Don't randomize it, 1446 * since it will conflict with jit_subprogs() logic. 1447 */ 1448 insn++; 1449 i++; 1450 continue; 1451 } 1452 1453 /* We temporarily need to hold the original ld64 insn 1454 * so that we can still access the first part in the 1455 * second blinding run. 1456 */ 1457 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1458 insn[1].code == 0) 1459 memcpy(aux, insn, sizeof(aux)); 1460 1461 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1462 clone->aux->verifier_zext); 1463 if (!rewritten) 1464 continue; 1465 1466 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1467 if (IS_ERR(tmp)) { 1468 /* Patching may have repointed aux->prog during 1469 * realloc from the original one, so we need to 1470 * fix it up here on error. 1471 */ 1472 bpf_jit_prog_release_other(prog, clone); 1473 return tmp; 1474 } 1475 1476 clone = tmp; 1477 insn_delta = rewritten - 1; 1478 1479 /* Walk new program and skip insns we just inserted. */ 1480 insn = clone->insnsi + i + insn_delta; 1481 insn_cnt += insn_delta; 1482 i += insn_delta; 1483 } 1484 1485 clone->blinded = 1; 1486 return clone; 1487 } 1488 #endif /* CONFIG_BPF_JIT */ 1489 1490 /* Base function for offset calculation. Needs to go into .text section, 1491 * therefore keeping it non-static as well; will also be used by JITs 1492 * anyway later on, so do not let the compiler omit it. This also needs 1493 * to go into kallsyms for correlation from e.g. bpftool, so naming 1494 * must not change. 1495 */ 1496 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1497 { 1498 return 0; 1499 } 1500 EXPORT_SYMBOL_GPL(__bpf_call_base); 1501 1502 /* All UAPI available opcodes. */ 1503 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1504 /* 32 bit ALU operations. */ \ 1505 /* Register based. */ \ 1506 INSN_3(ALU, ADD, X), \ 1507 INSN_3(ALU, SUB, X), \ 1508 INSN_3(ALU, AND, X), \ 1509 INSN_3(ALU, OR, X), \ 1510 INSN_3(ALU, LSH, X), \ 1511 INSN_3(ALU, RSH, X), \ 1512 INSN_3(ALU, XOR, X), \ 1513 INSN_3(ALU, MUL, X), \ 1514 INSN_3(ALU, MOV, X), \ 1515 INSN_3(ALU, ARSH, X), \ 1516 INSN_3(ALU, DIV, X), \ 1517 INSN_3(ALU, MOD, X), \ 1518 INSN_2(ALU, NEG), \ 1519 INSN_3(ALU, END, TO_BE), \ 1520 INSN_3(ALU, END, TO_LE), \ 1521 /* Immediate based. */ \ 1522 INSN_3(ALU, ADD, K), \ 1523 INSN_3(ALU, SUB, K), \ 1524 INSN_3(ALU, AND, K), \ 1525 INSN_3(ALU, OR, K), \ 1526 INSN_3(ALU, LSH, K), \ 1527 INSN_3(ALU, RSH, K), \ 1528 INSN_3(ALU, XOR, K), \ 1529 INSN_3(ALU, MUL, K), \ 1530 INSN_3(ALU, MOV, K), \ 1531 INSN_3(ALU, ARSH, K), \ 1532 INSN_3(ALU, DIV, K), \ 1533 INSN_3(ALU, MOD, K), \ 1534 /* 64 bit ALU operations. */ \ 1535 /* Register based. */ \ 1536 INSN_3(ALU64, ADD, X), \ 1537 INSN_3(ALU64, SUB, X), \ 1538 INSN_3(ALU64, AND, X), \ 1539 INSN_3(ALU64, OR, X), \ 1540 INSN_3(ALU64, LSH, X), \ 1541 INSN_3(ALU64, RSH, X), \ 1542 INSN_3(ALU64, XOR, X), \ 1543 INSN_3(ALU64, MUL, X), \ 1544 INSN_3(ALU64, MOV, X), \ 1545 INSN_3(ALU64, ARSH, X), \ 1546 INSN_3(ALU64, DIV, X), \ 1547 INSN_3(ALU64, MOD, X), \ 1548 INSN_2(ALU64, NEG), \ 1549 INSN_3(ALU64, END, TO_LE), \ 1550 /* Immediate based. */ \ 1551 INSN_3(ALU64, ADD, K), \ 1552 INSN_3(ALU64, SUB, K), \ 1553 INSN_3(ALU64, AND, K), \ 1554 INSN_3(ALU64, OR, K), \ 1555 INSN_3(ALU64, LSH, K), \ 1556 INSN_3(ALU64, RSH, K), \ 1557 INSN_3(ALU64, XOR, K), \ 1558 INSN_3(ALU64, MUL, K), \ 1559 INSN_3(ALU64, MOV, K), \ 1560 INSN_3(ALU64, ARSH, K), \ 1561 INSN_3(ALU64, DIV, K), \ 1562 INSN_3(ALU64, MOD, K), \ 1563 /* Call instruction. */ \ 1564 INSN_2(JMP, CALL), \ 1565 /* Exit instruction. */ \ 1566 INSN_2(JMP, EXIT), \ 1567 /* 32-bit Jump instructions. */ \ 1568 /* Register based. */ \ 1569 INSN_3(JMP32, JEQ, X), \ 1570 INSN_3(JMP32, JNE, X), \ 1571 INSN_3(JMP32, JGT, X), \ 1572 INSN_3(JMP32, JLT, X), \ 1573 INSN_3(JMP32, JGE, X), \ 1574 INSN_3(JMP32, JLE, X), \ 1575 INSN_3(JMP32, JSGT, X), \ 1576 INSN_3(JMP32, JSLT, X), \ 1577 INSN_3(JMP32, JSGE, X), \ 1578 INSN_3(JMP32, JSLE, X), \ 1579 INSN_3(JMP32, JSET, X), \ 1580 /* Immediate based. */ \ 1581 INSN_3(JMP32, JEQ, K), \ 1582 INSN_3(JMP32, JNE, K), \ 1583 INSN_3(JMP32, JGT, K), \ 1584 INSN_3(JMP32, JLT, K), \ 1585 INSN_3(JMP32, JGE, K), \ 1586 INSN_3(JMP32, JLE, K), \ 1587 INSN_3(JMP32, JSGT, K), \ 1588 INSN_3(JMP32, JSLT, K), \ 1589 INSN_3(JMP32, JSGE, K), \ 1590 INSN_3(JMP32, JSLE, K), \ 1591 INSN_3(JMP32, JSET, K), \ 1592 /* Jump instructions. */ \ 1593 /* Register based. */ \ 1594 INSN_3(JMP, JEQ, X), \ 1595 INSN_3(JMP, JNE, X), \ 1596 INSN_3(JMP, JGT, X), \ 1597 INSN_3(JMP, JLT, X), \ 1598 INSN_3(JMP, JGE, X), \ 1599 INSN_3(JMP, JLE, X), \ 1600 INSN_3(JMP, JSGT, X), \ 1601 INSN_3(JMP, JSLT, X), \ 1602 INSN_3(JMP, JSGE, X), \ 1603 INSN_3(JMP, JSLE, X), \ 1604 INSN_3(JMP, JSET, X), \ 1605 /* Immediate based. */ \ 1606 INSN_3(JMP, JEQ, K), \ 1607 INSN_3(JMP, JNE, K), \ 1608 INSN_3(JMP, JGT, K), \ 1609 INSN_3(JMP, JLT, K), \ 1610 INSN_3(JMP, JGE, K), \ 1611 INSN_3(JMP, JLE, K), \ 1612 INSN_3(JMP, JSGT, K), \ 1613 INSN_3(JMP, JSLT, K), \ 1614 INSN_3(JMP, JSGE, K), \ 1615 INSN_3(JMP, JSLE, K), \ 1616 INSN_3(JMP, JSET, K), \ 1617 INSN_2(JMP, JA), \ 1618 INSN_2(JMP32, JA), \ 1619 /* Store instructions. */ \ 1620 /* Register based. */ \ 1621 INSN_3(STX, MEM, B), \ 1622 INSN_3(STX, MEM, H), \ 1623 INSN_3(STX, MEM, W), \ 1624 INSN_3(STX, MEM, DW), \ 1625 INSN_3(STX, ATOMIC, W), \ 1626 INSN_3(STX, ATOMIC, DW), \ 1627 /* Immediate based. */ \ 1628 INSN_3(ST, MEM, B), \ 1629 INSN_3(ST, MEM, H), \ 1630 INSN_3(ST, MEM, W), \ 1631 INSN_3(ST, MEM, DW), \ 1632 /* Load instructions. */ \ 1633 /* Register based. */ \ 1634 INSN_3(LDX, MEM, B), \ 1635 INSN_3(LDX, MEM, H), \ 1636 INSN_3(LDX, MEM, W), \ 1637 INSN_3(LDX, MEM, DW), \ 1638 INSN_3(LDX, MEMSX, B), \ 1639 INSN_3(LDX, MEMSX, H), \ 1640 INSN_3(LDX, MEMSX, W), \ 1641 /* Immediate based. */ \ 1642 INSN_3(LD, IMM, DW) 1643 1644 bool bpf_opcode_in_insntable(u8 code) 1645 { 1646 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1647 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1648 static const bool public_insntable[256] = { 1649 [0 ... 255] = false, 1650 /* Now overwrite non-defaults ... */ 1651 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1652 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1653 [BPF_LD | BPF_ABS | BPF_B] = true, 1654 [BPF_LD | BPF_ABS | BPF_H] = true, 1655 [BPF_LD | BPF_ABS | BPF_W] = true, 1656 [BPF_LD | BPF_IND | BPF_B] = true, 1657 [BPF_LD | BPF_IND | BPF_H] = true, 1658 [BPF_LD | BPF_IND | BPF_W] = true, 1659 }; 1660 #undef BPF_INSN_3_TBL 1661 #undef BPF_INSN_2_TBL 1662 return public_insntable[code]; 1663 } 1664 1665 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1666 /** 1667 * ___bpf_prog_run - run eBPF program on a given context 1668 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1669 * @insn: is the array of eBPF instructions 1670 * 1671 * Decode and execute eBPF instructions. 1672 * 1673 * Return: whatever value is in %BPF_R0 at program exit 1674 */ 1675 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1676 { 1677 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1678 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1679 static const void * const jumptable[256] __annotate_jump_table = { 1680 [0 ... 255] = &&default_label, 1681 /* Now overwrite non-defaults ... */ 1682 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1683 /* Non-UAPI available opcodes. */ 1684 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1685 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1686 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1687 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1688 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1689 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1690 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1691 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1692 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1693 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1694 }; 1695 #undef BPF_INSN_3_LBL 1696 #undef BPF_INSN_2_LBL 1697 u32 tail_call_cnt = 0; 1698 1699 #define CONT ({ insn++; goto select_insn; }) 1700 #define CONT_JMP ({ insn++; goto select_insn; }) 1701 1702 select_insn: 1703 goto *jumptable[insn->code]; 1704 1705 /* Explicitly mask the register-based shift amounts with 63 or 31 1706 * to avoid undefined behavior. Normally this won't affect the 1707 * generated code, for example, in case of native 64 bit archs such 1708 * as x86-64 or arm64, the compiler is optimizing the AND away for 1709 * the interpreter. In case of JITs, each of the JIT backends compiles 1710 * the BPF shift operations to machine instructions which produce 1711 * implementation-defined results in such a case; the resulting 1712 * contents of the register may be arbitrary, but program behaviour 1713 * as a whole remains defined. In other words, in case of JIT backends, 1714 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1715 */ 1716 /* ALU (shifts) */ 1717 #define SHT(OPCODE, OP) \ 1718 ALU64_##OPCODE##_X: \ 1719 DST = DST OP (SRC & 63); \ 1720 CONT; \ 1721 ALU_##OPCODE##_X: \ 1722 DST = (u32) DST OP ((u32) SRC & 31); \ 1723 CONT; \ 1724 ALU64_##OPCODE##_K: \ 1725 DST = DST OP IMM; \ 1726 CONT; \ 1727 ALU_##OPCODE##_K: \ 1728 DST = (u32) DST OP (u32) IMM; \ 1729 CONT; 1730 /* ALU (rest) */ 1731 #define ALU(OPCODE, OP) \ 1732 ALU64_##OPCODE##_X: \ 1733 DST = DST OP SRC; \ 1734 CONT; \ 1735 ALU_##OPCODE##_X: \ 1736 DST = (u32) DST OP (u32) SRC; \ 1737 CONT; \ 1738 ALU64_##OPCODE##_K: \ 1739 DST = DST OP IMM; \ 1740 CONT; \ 1741 ALU_##OPCODE##_K: \ 1742 DST = (u32) DST OP (u32) IMM; \ 1743 CONT; 1744 ALU(ADD, +) 1745 ALU(SUB, -) 1746 ALU(AND, &) 1747 ALU(OR, |) 1748 ALU(XOR, ^) 1749 ALU(MUL, *) 1750 SHT(LSH, <<) 1751 SHT(RSH, >>) 1752 #undef SHT 1753 #undef ALU 1754 ALU_NEG: 1755 DST = (u32) -DST; 1756 CONT; 1757 ALU64_NEG: 1758 DST = -DST; 1759 CONT; 1760 ALU_MOV_X: 1761 switch (OFF) { 1762 case 0: 1763 DST = (u32) SRC; 1764 break; 1765 case 8: 1766 DST = (u32)(s8) SRC; 1767 break; 1768 case 16: 1769 DST = (u32)(s16) SRC; 1770 break; 1771 } 1772 CONT; 1773 ALU_MOV_K: 1774 DST = (u32) IMM; 1775 CONT; 1776 ALU64_MOV_X: 1777 switch (OFF) { 1778 case 0: 1779 DST = SRC; 1780 break; 1781 case 8: 1782 DST = (s8) SRC; 1783 break; 1784 case 16: 1785 DST = (s16) SRC; 1786 break; 1787 case 32: 1788 DST = (s32) SRC; 1789 break; 1790 } 1791 CONT; 1792 ALU64_MOV_K: 1793 DST = IMM; 1794 CONT; 1795 LD_IMM_DW: 1796 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1797 insn++; 1798 CONT; 1799 ALU_ARSH_X: 1800 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1801 CONT; 1802 ALU_ARSH_K: 1803 DST = (u64) (u32) (((s32) DST) >> IMM); 1804 CONT; 1805 ALU64_ARSH_X: 1806 (*(s64 *) &DST) >>= (SRC & 63); 1807 CONT; 1808 ALU64_ARSH_K: 1809 (*(s64 *) &DST) >>= IMM; 1810 CONT; 1811 ALU64_MOD_X: 1812 switch (OFF) { 1813 case 0: 1814 div64_u64_rem(DST, SRC, &AX); 1815 DST = AX; 1816 break; 1817 case 1: 1818 AX = div64_s64(DST, SRC); 1819 DST = DST - AX * SRC; 1820 break; 1821 } 1822 CONT; 1823 ALU_MOD_X: 1824 switch (OFF) { 1825 case 0: 1826 AX = (u32) DST; 1827 DST = do_div(AX, (u32) SRC); 1828 break; 1829 case 1: 1830 AX = abs((s32)DST); 1831 AX = do_div(AX, abs((s32)SRC)); 1832 if ((s32)DST < 0) 1833 DST = (u32)-AX; 1834 else 1835 DST = (u32)AX; 1836 break; 1837 } 1838 CONT; 1839 ALU64_MOD_K: 1840 switch (OFF) { 1841 case 0: 1842 div64_u64_rem(DST, IMM, &AX); 1843 DST = AX; 1844 break; 1845 case 1: 1846 AX = div64_s64(DST, IMM); 1847 DST = DST - AX * IMM; 1848 break; 1849 } 1850 CONT; 1851 ALU_MOD_K: 1852 switch (OFF) { 1853 case 0: 1854 AX = (u32) DST; 1855 DST = do_div(AX, (u32) IMM); 1856 break; 1857 case 1: 1858 AX = abs((s32)DST); 1859 AX = do_div(AX, abs((s32)IMM)); 1860 if ((s32)DST < 0) 1861 DST = (u32)-AX; 1862 else 1863 DST = (u32)AX; 1864 break; 1865 } 1866 CONT; 1867 ALU64_DIV_X: 1868 switch (OFF) { 1869 case 0: 1870 DST = div64_u64(DST, SRC); 1871 break; 1872 case 1: 1873 DST = div64_s64(DST, SRC); 1874 break; 1875 } 1876 CONT; 1877 ALU_DIV_X: 1878 switch (OFF) { 1879 case 0: 1880 AX = (u32) DST; 1881 do_div(AX, (u32) SRC); 1882 DST = (u32) AX; 1883 break; 1884 case 1: 1885 AX = abs((s32)DST); 1886 do_div(AX, abs((s32)SRC)); 1887 if (((s32)DST < 0) == ((s32)SRC < 0)) 1888 DST = (u32)AX; 1889 else 1890 DST = (u32)-AX; 1891 break; 1892 } 1893 CONT; 1894 ALU64_DIV_K: 1895 switch (OFF) { 1896 case 0: 1897 DST = div64_u64(DST, IMM); 1898 break; 1899 case 1: 1900 DST = div64_s64(DST, IMM); 1901 break; 1902 } 1903 CONT; 1904 ALU_DIV_K: 1905 switch (OFF) { 1906 case 0: 1907 AX = (u32) DST; 1908 do_div(AX, (u32) IMM); 1909 DST = (u32) AX; 1910 break; 1911 case 1: 1912 AX = abs((s32)DST); 1913 do_div(AX, abs((s32)IMM)); 1914 if (((s32)DST < 0) == ((s32)IMM < 0)) 1915 DST = (u32)AX; 1916 else 1917 DST = (u32)-AX; 1918 break; 1919 } 1920 CONT; 1921 ALU_END_TO_BE: 1922 switch (IMM) { 1923 case 16: 1924 DST = (__force u16) cpu_to_be16(DST); 1925 break; 1926 case 32: 1927 DST = (__force u32) cpu_to_be32(DST); 1928 break; 1929 case 64: 1930 DST = (__force u64) cpu_to_be64(DST); 1931 break; 1932 } 1933 CONT; 1934 ALU_END_TO_LE: 1935 switch (IMM) { 1936 case 16: 1937 DST = (__force u16) cpu_to_le16(DST); 1938 break; 1939 case 32: 1940 DST = (__force u32) cpu_to_le32(DST); 1941 break; 1942 case 64: 1943 DST = (__force u64) cpu_to_le64(DST); 1944 break; 1945 } 1946 CONT; 1947 ALU64_END_TO_LE: 1948 switch (IMM) { 1949 case 16: 1950 DST = (__force u16) __swab16(DST); 1951 break; 1952 case 32: 1953 DST = (__force u32) __swab32(DST); 1954 break; 1955 case 64: 1956 DST = (__force u64) __swab64(DST); 1957 break; 1958 } 1959 CONT; 1960 1961 /* CALL */ 1962 JMP_CALL: 1963 /* Function call scratches BPF_R1-BPF_R5 registers, 1964 * preserves BPF_R6-BPF_R9, and stores return value 1965 * into BPF_R0. 1966 */ 1967 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1968 BPF_R4, BPF_R5); 1969 CONT; 1970 1971 JMP_CALL_ARGS: 1972 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1973 BPF_R3, BPF_R4, 1974 BPF_R5, 1975 insn + insn->off + 1); 1976 CONT; 1977 1978 JMP_TAIL_CALL: { 1979 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1980 struct bpf_array *array = container_of(map, struct bpf_array, map); 1981 struct bpf_prog *prog; 1982 u32 index = BPF_R3; 1983 1984 if (unlikely(index >= array->map.max_entries)) 1985 goto out; 1986 1987 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 1988 goto out; 1989 1990 tail_call_cnt++; 1991 1992 prog = READ_ONCE(array->ptrs[index]); 1993 if (!prog) 1994 goto out; 1995 1996 /* ARG1 at this point is guaranteed to point to CTX from 1997 * the verifier side due to the fact that the tail call is 1998 * handled like a helper, that is, bpf_tail_call_proto, 1999 * where arg1_type is ARG_PTR_TO_CTX. 2000 */ 2001 insn = prog->insnsi; 2002 goto select_insn; 2003 out: 2004 CONT; 2005 } 2006 JMP_JA: 2007 insn += insn->off; 2008 CONT; 2009 JMP32_JA: 2010 insn += insn->imm; 2011 CONT; 2012 JMP_EXIT: 2013 return BPF_R0; 2014 /* JMP */ 2015 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2016 JMP_##OPCODE##_X: \ 2017 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2018 insn += insn->off; \ 2019 CONT_JMP; \ 2020 } \ 2021 CONT; \ 2022 JMP32_##OPCODE##_X: \ 2023 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2024 insn += insn->off; \ 2025 CONT_JMP; \ 2026 } \ 2027 CONT; \ 2028 JMP_##OPCODE##_K: \ 2029 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2030 insn += insn->off; \ 2031 CONT_JMP; \ 2032 } \ 2033 CONT; \ 2034 JMP32_##OPCODE##_K: \ 2035 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2036 insn += insn->off; \ 2037 CONT_JMP; \ 2038 } \ 2039 CONT; 2040 COND_JMP(u, JEQ, ==) 2041 COND_JMP(u, JNE, !=) 2042 COND_JMP(u, JGT, >) 2043 COND_JMP(u, JLT, <) 2044 COND_JMP(u, JGE, >=) 2045 COND_JMP(u, JLE, <=) 2046 COND_JMP(u, JSET, &) 2047 COND_JMP(s, JSGT, >) 2048 COND_JMP(s, JSLT, <) 2049 COND_JMP(s, JSGE, >=) 2050 COND_JMP(s, JSLE, <=) 2051 #undef COND_JMP 2052 /* ST, STX and LDX*/ 2053 ST_NOSPEC: 2054 /* Speculation barrier for mitigating Speculative Store Bypass. 2055 * In case of arm64, we rely on the firmware mitigation as 2056 * controlled via the ssbd kernel parameter. Whenever the 2057 * mitigation is enabled, it works for all of the kernel code 2058 * with no need to provide any additional instructions here. 2059 * In case of x86, we use 'lfence' insn for mitigation. We 2060 * reuse preexisting logic from Spectre v1 mitigation that 2061 * happens to produce the required code on x86 for v4 as well. 2062 */ 2063 barrier_nospec(); 2064 CONT; 2065 #define LDST(SIZEOP, SIZE) \ 2066 STX_MEM_##SIZEOP: \ 2067 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2068 CONT; \ 2069 ST_MEM_##SIZEOP: \ 2070 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2071 CONT; \ 2072 LDX_MEM_##SIZEOP: \ 2073 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2074 CONT; \ 2075 LDX_PROBE_MEM_##SIZEOP: \ 2076 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2077 (const void *)(long) (SRC + insn->off)); \ 2078 DST = *((SIZE *)&DST); \ 2079 CONT; 2080 2081 LDST(B, u8) 2082 LDST(H, u16) 2083 LDST(W, u32) 2084 LDST(DW, u64) 2085 #undef LDST 2086 2087 #define LDSX(SIZEOP, SIZE) \ 2088 LDX_MEMSX_##SIZEOP: \ 2089 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2090 CONT; \ 2091 LDX_PROBE_MEMSX_##SIZEOP: \ 2092 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2093 (const void *)(long) (SRC + insn->off)); \ 2094 DST = *((SIZE *)&DST); \ 2095 CONT; 2096 2097 LDSX(B, s8) 2098 LDSX(H, s16) 2099 LDSX(W, s32) 2100 #undef LDSX 2101 2102 #define ATOMIC_ALU_OP(BOP, KOP) \ 2103 case BOP: \ 2104 if (BPF_SIZE(insn->code) == BPF_W) \ 2105 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2106 (DST + insn->off)); \ 2107 else \ 2108 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2109 (DST + insn->off)); \ 2110 break; \ 2111 case BOP | BPF_FETCH: \ 2112 if (BPF_SIZE(insn->code) == BPF_W) \ 2113 SRC = (u32) atomic_fetch_##KOP( \ 2114 (u32) SRC, \ 2115 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2116 else \ 2117 SRC = (u64) atomic64_fetch_##KOP( \ 2118 (u64) SRC, \ 2119 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2120 break; 2121 2122 STX_ATOMIC_DW: 2123 STX_ATOMIC_W: 2124 switch (IMM) { 2125 ATOMIC_ALU_OP(BPF_ADD, add) 2126 ATOMIC_ALU_OP(BPF_AND, and) 2127 ATOMIC_ALU_OP(BPF_OR, or) 2128 ATOMIC_ALU_OP(BPF_XOR, xor) 2129 #undef ATOMIC_ALU_OP 2130 2131 case BPF_XCHG: 2132 if (BPF_SIZE(insn->code) == BPF_W) 2133 SRC = (u32) atomic_xchg( 2134 (atomic_t *)(unsigned long) (DST + insn->off), 2135 (u32) SRC); 2136 else 2137 SRC = (u64) atomic64_xchg( 2138 (atomic64_t *)(unsigned long) (DST + insn->off), 2139 (u64) SRC); 2140 break; 2141 case BPF_CMPXCHG: 2142 if (BPF_SIZE(insn->code) == BPF_W) 2143 BPF_R0 = (u32) atomic_cmpxchg( 2144 (atomic_t *)(unsigned long) (DST + insn->off), 2145 (u32) BPF_R0, (u32) SRC); 2146 else 2147 BPF_R0 = (u64) atomic64_cmpxchg( 2148 (atomic64_t *)(unsigned long) (DST + insn->off), 2149 (u64) BPF_R0, (u64) SRC); 2150 break; 2151 2152 default: 2153 goto default_label; 2154 } 2155 CONT; 2156 2157 default_label: 2158 /* If we ever reach this, we have a bug somewhere. Die hard here 2159 * instead of just returning 0; we could be somewhere in a subprog, 2160 * so execution could continue otherwise which we do /not/ want. 2161 * 2162 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2163 */ 2164 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2165 insn->code, insn->imm); 2166 BUG_ON(1); 2167 return 0; 2168 } 2169 2170 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2171 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2172 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2173 { \ 2174 u64 stack[stack_size / sizeof(u64)]; \ 2175 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2176 \ 2177 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2178 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2179 ARG1 = (u64) (unsigned long) ctx; \ 2180 return ___bpf_prog_run(regs, insn); \ 2181 } 2182 2183 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2184 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2185 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2186 const struct bpf_insn *insn) \ 2187 { \ 2188 u64 stack[stack_size / sizeof(u64)]; \ 2189 u64 regs[MAX_BPF_EXT_REG]; \ 2190 \ 2191 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2192 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2193 BPF_R1 = r1; \ 2194 BPF_R2 = r2; \ 2195 BPF_R3 = r3; \ 2196 BPF_R4 = r4; \ 2197 BPF_R5 = r5; \ 2198 return ___bpf_prog_run(regs, insn); \ 2199 } 2200 2201 #define EVAL1(FN, X) FN(X) 2202 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2203 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2204 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2205 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2206 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2207 2208 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2209 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2210 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2211 2212 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2213 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2214 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2215 2216 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2217 2218 static unsigned int (*interpreters[])(const void *ctx, 2219 const struct bpf_insn *insn) = { 2220 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2221 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2222 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2223 }; 2224 #undef PROG_NAME_LIST 2225 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2226 static __maybe_unused 2227 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2228 const struct bpf_insn *insn) = { 2229 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2230 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2231 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2232 }; 2233 #undef PROG_NAME_LIST 2234 2235 #ifdef CONFIG_BPF_SYSCALL 2236 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2237 { 2238 stack_depth = max_t(u32, stack_depth, 1); 2239 insn->off = (s16) insn->imm; 2240 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2241 __bpf_call_base_args; 2242 insn->code = BPF_JMP | BPF_CALL_ARGS; 2243 } 2244 #endif 2245 #else 2246 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2247 const struct bpf_insn *insn) 2248 { 2249 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2250 * is not working properly, so warn about it! 2251 */ 2252 WARN_ON_ONCE(1); 2253 return 0; 2254 } 2255 #endif 2256 2257 bool bpf_prog_map_compatible(struct bpf_map *map, 2258 const struct bpf_prog *fp) 2259 { 2260 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2261 bool ret; 2262 struct bpf_prog_aux *aux = fp->aux; 2263 2264 if (fp->kprobe_override) 2265 return false; 2266 2267 /* XDP programs inserted into maps are not guaranteed to run on 2268 * a particular netdev (and can run outside driver context entirely 2269 * in the case of devmap and cpumap). Until device checks 2270 * are implemented, prohibit adding dev-bound programs to program maps. 2271 */ 2272 if (bpf_prog_is_dev_bound(aux)) 2273 return false; 2274 2275 spin_lock(&map->owner.lock); 2276 if (!map->owner.type) { 2277 /* There's no owner yet where we could check for 2278 * compatibility. 2279 */ 2280 map->owner.type = prog_type; 2281 map->owner.jited = fp->jited; 2282 map->owner.xdp_has_frags = aux->xdp_has_frags; 2283 map->owner.attach_func_proto = aux->attach_func_proto; 2284 ret = true; 2285 } else { 2286 ret = map->owner.type == prog_type && 2287 map->owner.jited == fp->jited && 2288 map->owner.xdp_has_frags == aux->xdp_has_frags; 2289 if (ret && 2290 map->owner.attach_func_proto != aux->attach_func_proto) { 2291 switch (prog_type) { 2292 case BPF_PROG_TYPE_TRACING: 2293 case BPF_PROG_TYPE_LSM: 2294 case BPF_PROG_TYPE_EXT: 2295 case BPF_PROG_TYPE_STRUCT_OPS: 2296 ret = false; 2297 break; 2298 default: 2299 break; 2300 } 2301 } 2302 } 2303 spin_unlock(&map->owner.lock); 2304 2305 return ret; 2306 } 2307 2308 static int bpf_check_tail_call(const struct bpf_prog *fp) 2309 { 2310 struct bpf_prog_aux *aux = fp->aux; 2311 int i, ret = 0; 2312 2313 mutex_lock(&aux->used_maps_mutex); 2314 for (i = 0; i < aux->used_map_cnt; i++) { 2315 struct bpf_map *map = aux->used_maps[i]; 2316 2317 if (!map_type_contains_progs(map)) 2318 continue; 2319 2320 if (!bpf_prog_map_compatible(map, fp)) { 2321 ret = -EINVAL; 2322 goto out; 2323 } 2324 } 2325 2326 out: 2327 mutex_unlock(&aux->used_maps_mutex); 2328 return ret; 2329 } 2330 2331 static void bpf_prog_select_func(struct bpf_prog *fp) 2332 { 2333 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2334 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2335 2336 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2337 #else 2338 fp->bpf_func = __bpf_prog_ret0_warn; 2339 #endif 2340 } 2341 2342 /** 2343 * bpf_prog_select_runtime - select exec runtime for BPF program 2344 * @fp: bpf_prog populated with BPF program 2345 * @err: pointer to error variable 2346 * 2347 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2348 * The BPF program will be executed via bpf_prog_run() function. 2349 * 2350 * Return: the &fp argument along with &err set to 0 for success or 2351 * a negative errno code on failure 2352 */ 2353 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2354 { 2355 /* In case of BPF to BPF calls, verifier did all the prep 2356 * work with regards to JITing, etc. 2357 */ 2358 bool jit_needed = false; 2359 2360 if (fp->bpf_func) 2361 goto finalize; 2362 2363 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2364 bpf_prog_has_kfunc_call(fp)) 2365 jit_needed = true; 2366 2367 bpf_prog_select_func(fp); 2368 2369 /* eBPF JITs can rewrite the program in case constant 2370 * blinding is active. However, in case of error during 2371 * blinding, bpf_int_jit_compile() must always return a 2372 * valid program, which in this case would simply not 2373 * be JITed, but falls back to the interpreter. 2374 */ 2375 if (!bpf_prog_is_offloaded(fp->aux)) { 2376 *err = bpf_prog_alloc_jited_linfo(fp); 2377 if (*err) 2378 return fp; 2379 2380 fp = bpf_int_jit_compile(fp); 2381 bpf_prog_jit_attempt_done(fp); 2382 if (!fp->jited && jit_needed) { 2383 *err = -ENOTSUPP; 2384 return fp; 2385 } 2386 } else { 2387 *err = bpf_prog_offload_compile(fp); 2388 if (*err) 2389 return fp; 2390 } 2391 2392 finalize: 2393 bpf_prog_lock_ro(fp); 2394 2395 /* The tail call compatibility check can only be done at 2396 * this late stage as we need to determine, if we deal 2397 * with JITed or non JITed program concatenations and not 2398 * all eBPF JITs might immediately support all features. 2399 */ 2400 *err = bpf_check_tail_call(fp); 2401 2402 return fp; 2403 } 2404 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2405 2406 static unsigned int __bpf_prog_ret1(const void *ctx, 2407 const struct bpf_insn *insn) 2408 { 2409 return 1; 2410 } 2411 2412 static struct bpf_prog_dummy { 2413 struct bpf_prog prog; 2414 } dummy_bpf_prog = { 2415 .prog = { 2416 .bpf_func = __bpf_prog_ret1, 2417 }, 2418 }; 2419 2420 struct bpf_empty_prog_array bpf_empty_prog_array = { 2421 .null_prog = NULL, 2422 }; 2423 EXPORT_SYMBOL(bpf_empty_prog_array); 2424 2425 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2426 { 2427 if (prog_cnt) 2428 return kzalloc(sizeof(struct bpf_prog_array) + 2429 sizeof(struct bpf_prog_array_item) * 2430 (prog_cnt + 1), 2431 flags); 2432 2433 return &bpf_empty_prog_array.hdr; 2434 } 2435 2436 void bpf_prog_array_free(struct bpf_prog_array *progs) 2437 { 2438 if (!progs || progs == &bpf_empty_prog_array.hdr) 2439 return; 2440 kfree_rcu(progs, rcu); 2441 } 2442 2443 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2444 { 2445 struct bpf_prog_array *progs; 2446 2447 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2448 * no need to call kfree_rcu(), just call kfree() directly. 2449 */ 2450 progs = container_of(rcu, struct bpf_prog_array, rcu); 2451 if (rcu_trace_implies_rcu_gp()) 2452 kfree(progs); 2453 else 2454 kfree_rcu(progs, rcu); 2455 } 2456 2457 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2458 { 2459 if (!progs || progs == &bpf_empty_prog_array.hdr) 2460 return; 2461 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2462 } 2463 2464 int bpf_prog_array_length(struct bpf_prog_array *array) 2465 { 2466 struct bpf_prog_array_item *item; 2467 u32 cnt = 0; 2468 2469 for (item = array->items; item->prog; item++) 2470 if (item->prog != &dummy_bpf_prog.prog) 2471 cnt++; 2472 return cnt; 2473 } 2474 2475 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2476 { 2477 struct bpf_prog_array_item *item; 2478 2479 for (item = array->items; item->prog; item++) 2480 if (item->prog != &dummy_bpf_prog.prog) 2481 return false; 2482 return true; 2483 } 2484 2485 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2486 u32 *prog_ids, 2487 u32 request_cnt) 2488 { 2489 struct bpf_prog_array_item *item; 2490 int i = 0; 2491 2492 for (item = array->items; item->prog; item++) { 2493 if (item->prog == &dummy_bpf_prog.prog) 2494 continue; 2495 prog_ids[i] = item->prog->aux->id; 2496 if (++i == request_cnt) { 2497 item++; 2498 break; 2499 } 2500 } 2501 2502 return !!(item->prog); 2503 } 2504 2505 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2506 __u32 __user *prog_ids, u32 cnt) 2507 { 2508 unsigned long err = 0; 2509 bool nospc; 2510 u32 *ids; 2511 2512 /* users of this function are doing: 2513 * cnt = bpf_prog_array_length(); 2514 * if (cnt > 0) 2515 * bpf_prog_array_copy_to_user(..., cnt); 2516 * so below kcalloc doesn't need extra cnt > 0 check. 2517 */ 2518 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2519 if (!ids) 2520 return -ENOMEM; 2521 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2522 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2523 kfree(ids); 2524 if (err) 2525 return -EFAULT; 2526 if (nospc) 2527 return -ENOSPC; 2528 return 0; 2529 } 2530 2531 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2532 struct bpf_prog *old_prog) 2533 { 2534 struct bpf_prog_array_item *item; 2535 2536 for (item = array->items; item->prog; item++) 2537 if (item->prog == old_prog) { 2538 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2539 break; 2540 } 2541 } 2542 2543 /** 2544 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2545 * index into the program array with 2546 * a dummy no-op program. 2547 * @array: a bpf_prog_array 2548 * @index: the index of the program to replace 2549 * 2550 * Skips over dummy programs, by not counting them, when calculating 2551 * the position of the program to replace. 2552 * 2553 * Return: 2554 * * 0 - Success 2555 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2556 * * -ENOENT - Index out of range 2557 */ 2558 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2559 { 2560 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2561 } 2562 2563 /** 2564 * bpf_prog_array_update_at() - Updates the program at the given index 2565 * into the program array. 2566 * @array: a bpf_prog_array 2567 * @index: the index of the program to update 2568 * @prog: the program to insert into the array 2569 * 2570 * Skips over dummy programs, by not counting them, when calculating 2571 * the position of the program to update. 2572 * 2573 * Return: 2574 * * 0 - Success 2575 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2576 * * -ENOENT - Index out of range 2577 */ 2578 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2579 struct bpf_prog *prog) 2580 { 2581 struct bpf_prog_array_item *item; 2582 2583 if (unlikely(index < 0)) 2584 return -EINVAL; 2585 2586 for (item = array->items; item->prog; item++) { 2587 if (item->prog == &dummy_bpf_prog.prog) 2588 continue; 2589 if (!index) { 2590 WRITE_ONCE(item->prog, prog); 2591 return 0; 2592 } 2593 index--; 2594 } 2595 return -ENOENT; 2596 } 2597 2598 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2599 struct bpf_prog *exclude_prog, 2600 struct bpf_prog *include_prog, 2601 u64 bpf_cookie, 2602 struct bpf_prog_array **new_array) 2603 { 2604 int new_prog_cnt, carry_prog_cnt = 0; 2605 struct bpf_prog_array_item *existing, *new; 2606 struct bpf_prog_array *array; 2607 bool found_exclude = false; 2608 2609 /* Figure out how many existing progs we need to carry over to 2610 * the new array. 2611 */ 2612 if (old_array) { 2613 existing = old_array->items; 2614 for (; existing->prog; existing++) { 2615 if (existing->prog == exclude_prog) { 2616 found_exclude = true; 2617 continue; 2618 } 2619 if (existing->prog != &dummy_bpf_prog.prog) 2620 carry_prog_cnt++; 2621 if (existing->prog == include_prog) 2622 return -EEXIST; 2623 } 2624 } 2625 2626 if (exclude_prog && !found_exclude) 2627 return -ENOENT; 2628 2629 /* How many progs (not NULL) will be in the new array? */ 2630 new_prog_cnt = carry_prog_cnt; 2631 if (include_prog) 2632 new_prog_cnt += 1; 2633 2634 /* Do we have any prog (not NULL) in the new array? */ 2635 if (!new_prog_cnt) { 2636 *new_array = NULL; 2637 return 0; 2638 } 2639 2640 /* +1 as the end of prog_array is marked with NULL */ 2641 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2642 if (!array) 2643 return -ENOMEM; 2644 new = array->items; 2645 2646 /* Fill in the new prog array */ 2647 if (carry_prog_cnt) { 2648 existing = old_array->items; 2649 for (; existing->prog; existing++) { 2650 if (existing->prog == exclude_prog || 2651 existing->prog == &dummy_bpf_prog.prog) 2652 continue; 2653 2654 new->prog = existing->prog; 2655 new->bpf_cookie = existing->bpf_cookie; 2656 new++; 2657 } 2658 } 2659 if (include_prog) { 2660 new->prog = include_prog; 2661 new->bpf_cookie = bpf_cookie; 2662 new++; 2663 } 2664 new->prog = NULL; 2665 *new_array = array; 2666 return 0; 2667 } 2668 2669 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2670 u32 *prog_ids, u32 request_cnt, 2671 u32 *prog_cnt) 2672 { 2673 u32 cnt = 0; 2674 2675 if (array) 2676 cnt = bpf_prog_array_length(array); 2677 2678 *prog_cnt = cnt; 2679 2680 /* return early if user requested only program count or nothing to copy */ 2681 if (!request_cnt || !cnt) 2682 return 0; 2683 2684 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2685 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2686 : 0; 2687 } 2688 2689 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2690 struct bpf_map **used_maps, u32 len) 2691 { 2692 struct bpf_map *map; 2693 bool sleepable; 2694 u32 i; 2695 2696 sleepable = aux->sleepable; 2697 for (i = 0; i < len; i++) { 2698 map = used_maps[i]; 2699 if (map->ops->map_poke_untrack) 2700 map->ops->map_poke_untrack(map, aux); 2701 if (sleepable) 2702 atomic64_dec(&map->sleepable_refcnt); 2703 bpf_map_put(map); 2704 } 2705 } 2706 2707 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2708 { 2709 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2710 kfree(aux->used_maps); 2711 } 2712 2713 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2714 struct btf_mod_pair *used_btfs, u32 len) 2715 { 2716 #ifdef CONFIG_BPF_SYSCALL 2717 struct btf_mod_pair *btf_mod; 2718 u32 i; 2719 2720 for (i = 0; i < len; i++) { 2721 btf_mod = &used_btfs[i]; 2722 if (btf_mod->module) 2723 module_put(btf_mod->module); 2724 btf_put(btf_mod->btf); 2725 } 2726 #endif 2727 } 2728 2729 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2730 { 2731 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2732 kfree(aux->used_btfs); 2733 } 2734 2735 static void bpf_prog_free_deferred(struct work_struct *work) 2736 { 2737 struct bpf_prog_aux *aux; 2738 int i; 2739 2740 aux = container_of(work, struct bpf_prog_aux, work); 2741 #ifdef CONFIG_BPF_SYSCALL 2742 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2743 #endif 2744 #ifdef CONFIG_CGROUP_BPF 2745 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2746 bpf_cgroup_atype_put(aux->cgroup_atype); 2747 #endif 2748 bpf_free_used_maps(aux); 2749 bpf_free_used_btfs(aux); 2750 if (bpf_prog_is_dev_bound(aux)) 2751 bpf_prog_dev_bound_destroy(aux->prog); 2752 #ifdef CONFIG_PERF_EVENTS 2753 if (aux->prog->has_callchain_buf) 2754 put_callchain_buffers(); 2755 #endif 2756 if (aux->dst_trampoline) 2757 bpf_trampoline_put(aux->dst_trampoline); 2758 for (i = 0; i < aux->func_cnt; i++) { 2759 /* We can just unlink the subprog poke descriptor table as 2760 * it was originally linked to the main program and is also 2761 * released along with it. 2762 */ 2763 aux->func[i]->aux->poke_tab = NULL; 2764 bpf_jit_free(aux->func[i]); 2765 } 2766 if (aux->func_cnt) { 2767 kfree(aux->func); 2768 bpf_prog_unlock_free(aux->prog); 2769 } else { 2770 bpf_jit_free(aux->prog); 2771 } 2772 } 2773 2774 void bpf_prog_free(struct bpf_prog *fp) 2775 { 2776 struct bpf_prog_aux *aux = fp->aux; 2777 2778 if (aux->dst_prog) 2779 bpf_prog_put(aux->dst_prog); 2780 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2781 schedule_work(&aux->work); 2782 } 2783 EXPORT_SYMBOL_GPL(bpf_prog_free); 2784 2785 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2786 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2787 2788 void bpf_user_rnd_init_once(void) 2789 { 2790 prandom_init_once(&bpf_user_rnd_state); 2791 } 2792 2793 BPF_CALL_0(bpf_user_rnd_u32) 2794 { 2795 /* Should someone ever have the rather unwise idea to use some 2796 * of the registers passed into this function, then note that 2797 * this function is called from native eBPF and classic-to-eBPF 2798 * transformations. Register assignments from both sides are 2799 * different, f.e. classic always sets fn(ctx, A, X) here. 2800 */ 2801 struct rnd_state *state; 2802 u32 res; 2803 2804 state = &get_cpu_var(bpf_user_rnd_state); 2805 res = prandom_u32_state(state); 2806 put_cpu_var(bpf_user_rnd_state); 2807 2808 return res; 2809 } 2810 2811 BPF_CALL_0(bpf_get_raw_cpu_id) 2812 { 2813 return raw_smp_processor_id(); 2814 } 2815 2816 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2817 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2818 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2819 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2820 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2821 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2822 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2823 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2824 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2825 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2826 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2827 2828 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2829 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2830 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2831 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2832 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2833 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2834 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 2835 2836 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2837 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2838 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2839 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2840 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2841 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2842 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2843 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2844 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2845 const struct bpf_func_proto bpf_set_retval_proto __weak; 2846 const struct bpf_func_proto bpf_get_retval_proto __weak; 2847 2848 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2849 { 2850 return NULL; 2851 } 2852 2853 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2854 { 2855 return NULL; 2856 } 2857 2858 u64 __weak 2859 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2860 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2861 { 2862 return -ENOTSUPP; 2863 } 2864 EXPORT_SYMBOL_GPL(bpf_event_output); 2865 2866 /* Always built-in helper functions. */ 2867 const struct bpf_func_proto bpf_tail_call_proto = { 2868 .func = NULL, 2869 .gpl_only = false, 2870 .ret_type = RET_VOID, 2871 .arg1_type = ARG_PTR_TO_CTX, 2872 .arg2_type = ARG_CONST_MAP_PTR, 2873 .arg3_type = ARG_ANYTHING, 2874 }; 2875 2876 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2877 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2878 * eBPF and implicitly also cBPF can get JITed! 2879 */ 2880 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2881 { 2882 return prog; 2883 } 2884 2885 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2886 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2887 */ 2888 void __weak bpf_jit_compile(struct bpf_prog *prog) 2889 { 2890 } 2891 2892 bool __weak bpf_helper_changes_pkt_data(void *func) 2893 { 2894 return false; 2895 } 2896 2897 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2898 * analysis code and wants explicit zero extension inserted by verifier. 2899 * Otherwise, return FALSE. 2900 * 2901 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2902 * you don't override this. JITs that don't want these extra insns can detect 2903 * them using insn_is_zext. 2904 */ 2905 bool __weak bpf_jit_needs_zext(void) 2906 { 2907 return false; 2908 } 2909 2910 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 2911 bool __weak bpf_jit_supports_subprog_tailcalls(void) 2912 { 2913 return false; 2914 } 2915 2916 bool __weak bpf_jit_supports_kfunc_call(void) 2917 { 2918 return false; 2919 } 2920 2921 bool __weak bpf_jit_supports_far_kfunc_call(void) 2922 { 2923 return false; 2924 } 2925 2926 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2927 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2928 */ 2929 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2930 int len) 2931 { 2932 return -EFAULT; 2933 } 2934 2935 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2936 void *addr1, void *addr2) 2937 { 2938 return -ENOTSUPP; 2939 } 2940 2941 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 2942 { 2943 return ERR_PTR(-ENOTSUPP); 2944 } 2945 2946 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 2947 { 2948 return -ENOTSUPP; 2949 } 2950 2951 #ifdef CONFIG_BPF_SYSCALL 2952 static int __init bpf_global_ma_init(void) 2953 { 2954 int ret; 2955 2956 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 2957 bpf_global_ma_set = !ret; 2958 return ret; 2959 } 2960 late_initcall(bpf_global_ma_init); 2961 #endif 2962 2963 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 2964 EXPORT_SYMBOL(bpf_stats_enabled_key); 2965 2966 /* All definitions of tracepoints related to BPF. */ 2967 #define CREATE_TRACE_POINTS 2968 #include <linux/bpf_trace.h> 2969 2970 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 2971 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 2972