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 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 92 enum page_size_enum { 93 __PAGE_SIZE = PAGE_SIZE 94 }; 95 96 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 97 { 98 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 99 struct bpf_prog_aux *aux; 100 struct bpf_prog *fp; 101 102 size = round_up(size, __PAGE_SIZE); 103 fp = __vmalloc(size, gfp_flags); 104 if (fp == NULL) 105 return NULL; 106 107 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 108 if (aux == NULL) { 109 vfree(fp); 110 return NULL; 111 } 112 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 113 if (!fp->active) { 114 vfree(fp); 115 kfree(aux); 116 return NULL; 117 } 118 119 fp->pages = size / PAGE_SIZE; 120 fp->aux = aux; 121 fp->aux->prog = fp; 122 fp->jit_requested = ebpf_jit_enabled(); 123 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 124 #ifdef CONFIG_CGROUP_BPF 125 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 126 #endif 127 128 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 129 #ifdef CONFIG_FINEIBT 130 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 131 #endif 132 mutex_init(&fp->aux->used_maps_mutex); 133 mutex_init(&fp->aux->dst_mutex); 134 135 return fp; 136 } 137 138 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 139 { 140 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 141 struct bpf_prog *prog; 142 int cpu; 143 144 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 145 if (!prog) 146 return NULL; 147 148 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 149 if (!prog->stats) { 150 free_percpu(prog->active); 151 kfree(prog->aux); 152 vfree(prog); 153 return NULL; 154 } 155 156 for_each_possible_cpu(cpu) { 157 struct bpf_prog_stats *pstats; 158 159 pstats = per_cpu_ptr(prog->stats, cpu); 160 u64_stats_init(&pstats->syncp); 161 } 162 return prog; 163 } 164 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 165 166 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 167 { 168 if (!prog->aux->nr_linfo || !prog->jit_requested) 169 return 0; 170 171 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 172 sizeof(*prog->aux->jited_linfo), 173 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 174 if (!prog->aux->jited_linfo) 175 return -ENOMEM; 176 177 return 0; 178 } 179 180 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 181 { 182 if (prog->aux->jited_linfo && 183 (!prog->jited || !prog->aux->jited_linfo[0])) { 184 kvfree(prog->aux->jited_linfo); 185 prog->aux->jited_linfo = NULL; 186 } 187 188 kfree(prog->aux->kfunc_tab); 189 prog->aux->kfunc_tab = NULL; 190 } 191 192 /* The jit engine is responsible to provide an array 193 * for insn_off to the jited_off mapping (insn_to_jit_off). 194 * 195 * The idx to this array is the insn_off. Hence, the insn_off 196 * here is relative to the prog itself instead of the main prog. 197 * This array has one entry for each xlated bpf insn. 198 * 199 * jited_off is the byte off to the end of the jited insn. 200 * 201 * Hence, with 202 * insn_start: 203 * The first bpf insn off of the prog. The insn off 204 * here is relative to the main prog. 205 * e.g. if prog is a subprog, insn_start > 0 206 * linfo_idx: 207 * The prog's idx to prog->aux->linfo and jited_linfo 208 * 209 * jited_linfo[linfo_idx] = prog->bpf_func 210 * 211 * For i > linfo_idx, 212 * 213 * jited_linfo[i] = prog->bpf_func + 214 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 215 */ 216 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 217 const u32 *insn_to_jit_off) 218 { 219 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 220 const struct bpf_line_info *linfo; 221 void **jited_linfo; 222 223 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 224 /* Userspace did not provide linfo */ 225 return; 226 227 linfo_idx = prog->aux->linfo_idx; 228 linfo = &prog->aux->linfo[linfo_idx]; 229 insn_start = linfo[0].insn_off; 230 insn_end = insn_start + prog->len; 231 232 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 233 jited_linfo[0] = prog->bpf_func; 234 235 nr_linfo = prog->aux->nr_linfo - linfo_idx; 236 237 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 238 /* The verifier ensures that linfo[i].insn_off is 239 * strictly increasing 240 */ 241 jited_linfo[i] = prog->bpf_func + 242 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 243 } 244 245 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 246 gfp_t gfp_extra_flags) 247 { 248 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 249 struct bpf_prog *fp; 250 u32 pages; 251 252 size = round_up(size, PAGE_SIZE); 253 pages = size / PAGE_SIZE; 254 if (pages <= fp_old->pages) 255 return fp_old; 256 257 fp = __vmalloc(size, gfp_flags); 258 if (fp) { 259 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 260 fp->pages = pages; 261 fp->aux->prog = fp; 262 263 /* We keep fp->aux from fp_old around in the new 264 * reallocated structure. 265 */ 266 fp_old->aux = NULL; 267 fp_old->stats = NULL; 268 fp_old->active = NULL; 269 __bpf_prog_free(fp_old); 270 } 271 272 return fp; 273 } 274 275 void __bpf_prog_free(struct bpf_prog *fp) 276 { 277 if (fp->aux) { 278 mutex_destroy(&fp->aux->used_maps_mutex); 279 mutex_destroy(&fp->aux->dst_mutex); 280 kfree(fp->aux->poke_tab); 281 kfree(fp->aux); 282 } 283 free_percpu(fp->stats); 284 free_percpu(fp->active); 285 vfree(fp); 286 } 287 288 int bpf_prog_calc_tag(struct bpf_prog *fp) 289 { 290 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 291 u32 raw_size = bpf_prog_tag_scratch_size(fp); 292 u32 digest[SHA1_DIGEST_WORDS]; 293 u32 ws[SHA1_WORKSPACE_WORDS]; 294 u32 i, bsize, psize, blocks; 295 struct bpf_insn *dst; 296 bool was_ld_map; 297 u8 *raw, *todo; 298 __be32 *result; 299 __be64 *bits; 300 301 raw = vmalloc(raw_size); 302 if (!raw) 303 return -ENOMEM; 304 305 sha1_init(digest); 306 memset(ws, 0, sizeof(ws)); 307 308 /* We need to take out the map fd for the digest calculation 309 * since they are unstable from user space side. 310 */ 311 dst = (void *)raw; 312 for (i = 0, was_ld_map = false; i < fp->len; i++) { 313 dst[i] = fp->insnsi[i]; 314 if (!was_ld_map && 315 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 316 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 317 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 318 was_ld_map = true; 319 dst[i].imm = 0; 320 } else if (was_ld_map && 321 dst[i].code == 0 && 322 dst[i].dst_reg == 0 && 323 dst[i].src_reg == 0 && 324 dst[i].off == 0) { 325 was_ld_map = false; 326 dst[i].imm = 0; 327 } else { 328 was_ld_map = false; 329 } 330 } 331 332 psize = bpf_prog_insn_size(fp); 333 memset(&raw[psize], 0, raw_size - psize); 334 raw[psize++] = 0x80; 335 336 bsize = round_up(psize, SHA1_BLOCK_SIZE); 337 blocks = bsize / SHA1_BLOCK_SIZE; 338 todo = raw; 339 if (bsize - psize >= sizeof(__be64)) { 340 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 341 } else { 342 bits = (__be64 *)(todo + bsize + bits_offset); 343 blocks++; 344 } 345 *bits = cpu_to_be64((psize - 1) << 3); 346 347 while (blocks--) { 348 sha1_transform(digest, todo, ws); 349 todo += SHA1_BLOCK_SIZE; 350 } 351 352 result = (__force __be32 *)digest; 353 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 354 result[i] = cpu_to_be32(digest[i]); 355 memcpy(fp->tag, result, sizeof(fp->tag)); 356 357 vfree(raw); 358 return 0; 359 } 360 361 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 362 s32 end_new, s32 curr, const bool probe_pass) 363 { 364 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 365 s32 delta = end_new - end_old; 366 s64 imm = insn->imm; 367 368 if (curr < pos && curr + imm + 1 >= end_old) 369 imm += delta; 370 else if (curr >= end_new && curr + imm + 1 < end_new) 371 imm -= delta; 372 if (imm < imm_min || imm > imm_max) 373 return -ERANGE; 374 if (!probe_pass) 375 insn->imm = imm; 376 return 0; 377 } 378 379 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 380 s32 end_new, s32 curr, const bool probe_pass) 381 { 382 s64 off_min, off_max, off; 383 s32 delta = end_new - end_old; 384 385 if (insn->code == (BPF_JMP32 | BPF_JA)) { 386 off = insn->imm; 387 off_min = S32_MIN; 388 off_max = S32_MAX; 389 } else { 390 off = insn->off; 391 off_min = S16_MIN; 392 off_max = S16_MAX; 393 } 394 395 if (curr < pos && curr + off + 1 >= end_old) 396 off += delta; 397 else if (curr >= end_new && curr + off + 1 < end_new) 398 off -= delta; 399 if (off < off_min || off > off_max) 400 return -ERANGE; 401 if (!probe_pass) { 402 if (insn->code == (BPF_JMP32 | BPF_JA)) 403 insn->imm = off; 404 else 405 insn->off = off; 406 } 407 return 0; 408 } 409 410 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 411 s32 end_new, const bool probe_pass) 412 { 413 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 414 struct bpf_insn *insn = prog->insnsi; 415 int ret = 0; 416 417 for (i = 0; i < insn_cnt; i++, insn++) { 418 u8 code; 419 420 /* In the probing pass we still operate on the original, 421 * unpatched image in order to check overflows before we 422 * do any other adjustments. Therefore skip the patchlet. 423 */ 424 if (probe_pass && i == pos) { 425 i = end_new; 426 insn = prog->insnsi + end_old; 427 } 428 if (bpf_pseudo_func(insn)) { 429 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 430 end_new, i, probe_pass); 431 if (ret) 432 return ret; 433 continue; 434 } 435 code = insn->code; 436 if ((BPF_CLASS(code) != BPF_JMP && 437 BPF_CLASS(code) != BPF_JMP32) || 438 BPF_OP(code) == BPF_EXIT) 439 continue; 440 /* Adjust offset of jmps if we cross patch boundaries. */ 441 if (BPF_OP(code) == BPF_CALL) { 442 if (insn->src_reg != BPF_PSEUDO_CALL) 443 continue; 444 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 445 end_new, i, probe_pass); 446 } else { 447 ret = bpf_adj_delta_to_off(insn, pos, end_old, 448 end_new, i, probe_pass); 449 } 450 if (ret) 451 break; 452 } 453 454 return ret; 455 } 456 457 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 458 { 459 struct bpf_line_info *linfo; 460 u32 i, nr_linfo; 461 462 nr_linfo = prog->aux->nr_linfo; 463 if (!nr_linfo || !delta) 464 return; 465 466 linfo = prog->aux->linfo; 467 468 for (i = 0; i < nr_linfo; i++) 469 if (off < linfo[i].insn_off) 470 break; 471 472 /* Push all off < linfo[i].insn_off by delta */ 473 for (; i < nr_linfo; i++) 474 linfo[i].insn_off += delta; 475 } 476 477 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 478 const struct bpf_insn *patch, u32 len) 479 { 480 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 481 const u32 cnt_max = S16_MAX; 482 struct bpf_prog *prog_adj; 483 int err; 484 485 /* Since our patchlet doesn't expand the image, we're done. */ 486 if (insn_delta == 0) { 487 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 488 return prog; 489 } 490 491 insn_adj_cnt = prog->len + insn_delta; 492 493 /* Reject anything that would potentially let the insn->off 494 * target overflow when we have excessive program expansions. 495 * We need to probe here before we do any reallocation where 496 * we afterwards may not fail anymore. 497 */ 498 if (insn_adj_cnt > cnt_max && 499 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 500 return ERR_PTR(err); 501 502 /* Several new instructions need to be inserted. Make room 503 * for them. Likely, there's no need for a new allocation as 504 * last page could have large enough tailroom. 505 */ 506 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 507 GFP_USER); 508 if (!prog_adj) 509 return ERR_PTR(-ENOMEM); 510 511 prog_adj->len = insn_adj_cnt; 512 513 /* Patching happens in 3 steps: 514 * 515 * 1) Move over tail of insnsi from next instruction onwards, 516 * so we can patch the single target insn with one or more 517 * new ones (patching is always from 1 to n insns, n > 0). 518 * 2) Inject new instructions at the target location. 519 * 3) Adjust branch offsets if necessary. 520 */ 521 insn_rest = insn_adj_cnt - off - len; 522 523 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 524 sizeof(*patch) * insn_rest); 525 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 526 527 /* We are guaranteed to not fail at this point, otherwise 528 * the ship has sailed to reverse to the original state. An 529 * overflow cannot happen at this point. 530 */ 531 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 532 533 bpf_adj_linfo(prog_adj, off, insn_delta); 534 535 return prog_adj; 536 } 537 538 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 539 { 540 /* Branch offsets can't overflow when program is shrinking, no need 541 * to call bpf_adj_branches(..., true) here 542 */ 543 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 544 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 545 prog->len -= cnt; 546 547 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 548 } 549 550 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 551 { 552 int i; 553 554 for (i = 0; i < fp->aux->real_func_cnt; i++) 555 bpf_prog_kallsyms_del(fp->aux->func[i]); 556 } 557 558 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 559 { 560 bpf_prog_kallsyms_del_subprogs(fp); 561 bpf_prog_kallsyms_del(fp); 562 } 563 564 #ifdef CONFIG_BPF_JIT 565 /* All BPF JIT sysctl knobs here. */ 566 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 567 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 568 int bpf_jit_harden __read_mostly; 569 long bpf_jit_limit __read_mostly; 570 long bpf_jit_limit_max __read_mostly; 571 572 static void 573 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 574 { 575 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 576 577 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 578 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 579 } 580 581 static void 582 bpf_prog_ksym_set_name(struct bpf_prog *prog) 583 { 584 char *sym = prog->aux->ksym.name; 585 const char *end = sym + KSYM_NAME_LEN; 586 const struct btf_type *type; 587 const char *func_name; 588 589 BUILD_BUG_ON(sizeof("bpf_prog_") + 590 sizeof(prog->tag) * 2 + 591 /* name has been null terminated. 592 * We should need +1 for the '_' preceding 593 * the name. However, the null character 594 * is double counted between the name and the 595 * sizeof("bpf_prog_") above, so we omit 596 * the +1 here. 597 */ 598 sizeof(prog->aux->name) > KSYM_NAME_LEN); 599 600 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 601 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 602 603 /* prog->aux->name will be ignored if full btf name is available */ 604 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 605 type = btf_type_by_id(prog->aux->btf, 606 prog->aux->func_info[prog->aux->func_idx].type_id); 607 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 608 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 609 return; 610 } 611 612 if (prog->aux->name[0]) 613 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 614 else 615 *sym = 0; 616 } 617 618 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 619 { 620 return container_of(n, struct bpf_ksym, tnode)->start; 621 } 622 623 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 624 struct latch_tree_node *b) 625 { 626 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 627 } 628 629 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 630 { 631 unsigned long val = (unsigned long)key; 632 const struct bpf_ksym *ksym; 633 634 ksym = container_of(n, struct bpf_ksym, tnode); 635 636 if (val < ksym->start) 637 return -1; 638 /* Ensure that we detect return addresses as part of the program, when 639 * the final instruction is a call for a program part of the stack 640 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 641 */ 642 if (val > ksym->end) 643 return 1; 644 645 return 0; 646 } 647 648 static const struct latch_tree_ops bpf_tree_ops = { 649 .less = bpf_tree_less, 650 .comp = bpf_tree_comp, 651 }; 652 653 static DEFINE_SPINLOCK(bpf_lock); 654 static LIST_HEAD(bpf_kallsyms); 655 static struct latch_tree_root bpf_tree __cacheline_aligned; 656 657 void bpf_ksym_add(struct bpf_ksym *ksym) 658 { 659 spin_lock_bh(&bpf_lock); 660 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 661 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 662 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 663 spin_unlock_bh(&bpf_lock); 664 } 665 666 static void __bpf_ksym_del(struct bpf_ksym *ksym) 667 { 668 if (list_empty(&ksym->lnode)) 669 return; 670 671 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 672 list_del_rcu(&ksym->lnode); 673 } 674 675 void bpf_ksym_del(struct bpf_ksym *ksym) 676 { 677 spin_lock_bh(&bpf_lock); 678 __bpf_ksym_del(ksym); 679 spin_unlock_bh(&bpf_lock); 680 } 681 682 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 683 { 684 return fp->jited && !bpf_prog_was_classic(fp); 685 } 686 687 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 688 { 689 if (!bpf_prog_kallsyms_candidate(fp) || 690 !bpf_token_capable(fp->aux->token, CAP_BPF)) 691 return; 692 693 bpf_prog_ksym_set_addr(fp); 694 bpf_prog_ksym_set_name(fp); 695 fp->aux->ksym.prog = true; 696 697 bpf_ksym_add(&fp->aux->ksym); 698 699 #ifdef CONFIG_FINEIBT 700 /* 701 * When FineIBT, code in the __cfi_foo() symbols can get executed 702 * and hence unwinder needs help. 703 */ 704 if (cfi_mode != CFI_FINEIBT) 705 return; 706 707 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 708 "__cfi_%s", fp->aux->ksym.name); 709 710 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 711 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 712 713 bpf_ksym_add(&fp->aux->ksym_prefix); 714 #endif 715 } 716 717 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 718 { 719 if (!bpf_prog_kallsyms_candidate(fp)) 720 return; 721 722 bpf_ksym_del(&fp->aux->ksym); 723 #ifdef CONFIG_FINEIBT 724 if (cfi_mode != CFI_FINEIBT) 725 return; 726 bpf_ksym_del(&fp->aux->ksym_prefix); 727 #endif 728 } 729 730 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 731 { 732 struct latch_tree_node *n; 733 734 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 735 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 736 } 737 738 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 739 unsigned long *off, char *sym) 740 { 741 struct bpf_ksym *ksym; 742 char *ret = NULL; 743 744 rcu_read_lock(); 745 ksym = bpf_ksym_find(addr); 746 if (ksym) { 747 unsigned long symbol_start = ksym->start; 748 unsigned long symbol_end = ksym->end; 749 750 strncpy(sym, ksym->name, KSYM_NAME_LEN); 751 752 ret = sym; 753 if (size) 754 *size = symbol_end - symbol_start; 755 if (off) 756 *off = addr - symbol_start; 757 } 758 rcu_read_unlock(); 759 760 return ret; 761 } 762 763 bool is_bpf_text_address(unsigned long addr) 764 { 765 bool ret; 766 767 rcu_read_lock(); 768 ret = bpf_ksym_find(addr) != NULL; 769 rcu_read_unlock(); 770 771 return ret; 772 } 773 774 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 775 { 776 struct bpf_ksym *ksym = bpf_ksym_find(addr); 777 778 return ksym && ksym->prog ? 779 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 780 NULL; 781 } 782 783 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 784 { 785 const struct exception_table_entry *e = NULL; 786 struct bpf_prog *prog; 787 788 rcu_read_lock(); 789 prog = bpf_prog_ksym_find(addr); 790 if (!prog) 791 goto out; 792 if (!prog->aux->num_exentries) 793 goto out; 794 795 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 796 out: 797 rcu_read_unlock(); 798 return e; 799 } 800 801 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 802 char *sym) 803 { 804 struct bpf_ksym *ksym; 805 unsigned int it = 0; 806 int ret = -ERANGE; 807 808 if (!bpf_jit_kallsyms_enabled()) 809 return ret; 810 811 rcu_read_lock(); 812 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 813 if (it++ != symnum) 814 continue; 815 816 strncpy(sym, ksym->name, KSYM_NAME_LEN); 817 818 *value = ksym->start; 819 *type = BPF_SYM_ELF_TYPE; 820 821 ret = 0; 822 break; 823 } 824 rcu_read_unlock(); 825 826 return ret; 827 } 828 829 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 830 struct bpf_jit_poke_descriptor *poke) 831 { 832 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 833 static const u32 poke_tab_max = 1024; 834 u32 slot = prog->aux->size_poke_tab; 835 u32 size = slot + 1; 836 837 if (size > poke_tab_max) 838 return -ENOSPC; 839 if (poke->tailcall_target || poke->tailcall_target_stable || 840 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 841 return -EINVAL; 842 843 switch (poke->reason) { 844 case BPF_POKE_REASON_TAIL_CALL: 845 if (!poke->tail_call.map) 846 return -EINVAL; 847 break; 848 default: 849 return -EINVAL; 850 } 851 852 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL); 853 if (!tab) 854 return -ENOMEM; 855 856 memcpy(&tab[slot], poke, sizeof(*poke)); 857 prog->aux->size_poke_tab = size; 858 prog->aux->poke_tab = tab; 859 860 return slot; 861 } 862 863 /* 864 * BPF program pack allocator. 865 * 866 * Most BPF programs are pretty small. Allocating a hole page for each 867 * program is sometime a waste. Many small bpf program also adds pressure 868 * to instruction TLB. To solve this issue, we introduce a BPF program pack 869 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 870 * to host BPF programs. 871 */ 872 #define BPF_PROG_CHUNK_SHIFT 6 873 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 874 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 875 876 struct bpf_prog_pack { 877 struct list_head list; 878 void *ptr; 879 unsigned long bitmap[]; 880 }; 881 882 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 883 { 884 memset(area, 0, size); 885 } 886 887 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 888 889 static DEFINE_MUTEX(pack_mutex); 890 static LIST_HEAD(pack_list); 891 892 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 893 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 894 */ 895 #ifdef PMD_SIZE 896 /* PMD_SIZE is really big for some archs. It doesn't make sense to 897 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 898 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 899 * greater than or equal to 2MB. 900 */ 901 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 902 #else 903 #define BPF_PROG_PACK_SIZE PAGE_SIZE 904 #endif 905 906 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 907 908 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 909 { 910 struct bpf_prog_pack *pack; 911 int err; 912 913 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 914 GFP_KERNEL); 915 if (!pack) 916 return NULL; 917 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 918 if (!pack->ptr) 919 goto out; 920 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 921 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 922 923 set_vm_flush_reset_perms(pack->ptr); 924 err = set_memory_rox((unsigned long)pack->ptr, 925 BPF_PROG_PACK_SIZE / PAGE_SIZE); 926 if (err) 927 goto out; 928 list_add_tail(&pack->list, &pack_list); 929 return pack; 930 931 out: 932 bpf_jit_free_exec(pack->ptr); 933 kfree(pack); 934 return NULL; 935 } 936 937 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 938 { 939 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 940 struct bpf_prog_pack *pack; 941 unsigned long pos; 942 void *ptr = NULL; 943 944 mutex_lock(&pack_mutex); 945 if (size > BPF_PROG_PACK_SIZE) { 946 size = round_up(size, PAGE_SIZE); 947 ptr = bpf_jit_alloc_exec(size); 948 if (ptr) { 949 int err; 950 951 bpf_fill_ill_insns(ptr, size); 952 set_vm_flush_reset_perms(ptr); 953 err = set_memory_rox((unsigned long)ptr, 954 size / PAGE_SIZE); 955 if (err) { 956 bpf_jit_free_exec(ptr); 957 ptr = NULL; 958 } 959 } 960 goto out; 961 } 962 list_for_each_entry(pack, &pack_list, list) { 963 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 964 nbits, 0); 965 if (pos < BPF_PROG_CHUNK_COUNT) 966 goto found_free_area; 967 } 968 969 pack = alloc_new_pack(bpf_fill_ill_insns); 970 if (!pack) 971 goto out; 972 973 pos = 0; 974 975 found_free_area: 976 bitmap_set(pack->bitmap, pos, nbits); 977 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 978 979 out: 980 mutex_unlock(&pack_mutex); 981 return ptr; 982 } 983 984 void bpf_prog_pack_free(void *ptr, u32 size) 985 { 986 struct bpf_prog_pack *pack = NULL, *tmp; 987 unsigned int nbits; 988 unsigned long pos; 989 990 mutex_lock(&pack_mutex); 991 if (size > BPF_PROG_PACK_SIZE) { 992 bpf_jit_free_exec(ptr); 993 goto out; 994 } 995 996 list_for_each_entry(tmp, &pack_list, list) { 997 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 998 pack = tmp; 999 break; 1000 } 1001 } 1002 1003 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1004 goto out; 1005 1006 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1007 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1008 1009 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1010 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1011 1012 bitmap_clear(pack->bitmap, pos, nbits); 1013 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1014 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1015 list_del(&pack->list); 1016 bpf_jit_free_exec(pack->ptr); 1017 kfree(pack); 1018 } 1019 out: 1020 mutex_unlock(&pack_mutex); 1021 } 1022 1023 static atomic_long_t bpf_jit_current; 1024 1025 /* Can be overridden by an arch's JIT compiler if it has a custom, 1026 * dedicated BPF backend memory area, or if neither of the two 1027 * below apply. 1028 */ 1029 u64 __weak bpf_jit_alloc_exec_limit(void) 1030 { 1031 #if defined(MODULES_VADDR) 1032 return MODULES_END - MODULES_VADDR; 1033 #else 1034 return VMALLOC_END - VMALLOC_START; 1035 #endif 1036 } 1037 1038 static int __init bpf_jit_charge_init(void) 1039 { 1040 /* Only used as heuristic here to derive limit. */ 1041 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1042 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1043 PAGE_SIZE), LONG_MAX); 1044 return 0; 1045 } 1046 pure_initcall(bpf_jit_charge_init); 1047 1048 int bpf_jit_charge_modmem(u32 size) 1049 { 1050 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1051 if (!bpf_capable()) { 1052 atomic_long_sub(size, &bpf_jit_current); 1053 return -EPERM; 1054 } 1055 } 1056 1057 return 0; 1058 } 1059 1060 void bpf_jit_uncharge_modmem(u32 size) 1061 { 1062 atomic_long_sub(size, &bpf_jit_current); 1063 } 1064 1065 void *__weak bpf_jit_alloc_exec(unsigned long size) 1066 { 1067 return module_alloc(size); 1068 } 1069 1070 void __weak bpf_jit_free_exec(void *addr) 1071 { 1072 module_memfree(addr); 1073 } 1074 1075 struct bpf_binary_header * 1076 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1077 unsigned int alignment, 1078 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1079 { 1080 struct bpf_binary_header *hdr; 1081 u32 size, hole, start; 1082 1083 WARN_ON_ONCE(!is_power_of_2(alignment) || 1084 alignment > BPF_IMAGE_ALIGNMENT); 1085 1086 /* Most of BPF filters are really small, but if some of them 1087 * fill a page, allow at least 128 extra bytes to insert a 1088 * random section of illegal instructions. 1089 */ 1090 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1091 1092 if (bpf_jit_charge_modmem(size)) 1093 return NULL; 1094 hdr = bpf_jit_alloc_exec(size); 1095 if (!hdr) { 1096 bpf_jit_uncharge_modmem(size); 1097 return NULL; 1098 } 1099 1100 /* Fill space with illegal/arch-dep instructions. */ 1101 bpf_fill_ill_insns(hdr, size); 1102 1103 hdr->size = size; 1104 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1105 PAGE_SIZE - sizeof(*hdr)); 1106 start = get_random_u32_below(hole) & ~(alignment - 1); 1107 1108 /* Leave a random number of instructions before BPF code. */ 1109 *image_ptr = &hdr->image[start]; 1110 1111 return hdr; 1112 } 1113 1114 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1115 { 1116 u32 size = hdr->size; 1117 1118 bpf_jit_free_exec(hdr); 1119 bpf_jit_uncharge_modmem(size); 1120 } 1121 1122 /* Allocate jit binary from bpf_prog_pack allocator. 1123 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1124 * to the memory. To solve this problem, a RW buffer is also allocated at 1125 * as the same time. The JIT engine should calculate offsets based on the 1126 * RO memory address, but write JITed program to the RW buffer. Once the 1127 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1128 * the JITed program to the RO memory. 1129 */ 1130 struct bpf_binary_header * 1131 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1132 unsigned int alignment, 1133 struct bpf_binary_header **rw_header, 1134 u8 **rw_image, 1135 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1136 { 1137 struct bpf_binary_header *ro_header; 1138 u32 size, hole, start; 1139 1140 WARN_ON_ONCE(!is_power_of_2(alignment) || 1141 alignment > BPF_IMAGE_ALIGNMENT); 1142 1143 /* add 16 bytes for a random section of illegal instructions */ 1144 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1145 1146 if (bpf_jit_charge_modmem(size)) 1147 return NULL; 1148 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1149 if (!ro_header) { 1150 bpf_jit_uncharge_modmem(size); 1151 return NULL; 1152 } 1153 1154 *rw_header = kvmalloc(size, GFP_KERNEL); 1155 if (!*rw_header) { 1156 bpf_prog_pack_free(ro_header, size); 1157 bpf_jit_uncharge_modmem(size); 1158 return NULL; 1159 } 1160 1161 /* Fill space with illegal/arch-dep instructions. */ 1162 bpf_fill_ill_insns(*rw_header, size); 1163 (*rw_header)->size = size; 1164 1165 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1166 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1167 start = get_random_u32_below(hole) & ~(alignment - 1); 1168 1169 *image_ptr = &ro_header->image[start]; 1170 *rw_image = &(*rw_header)->image[start]; 1171 1172 return ro_header; 1173 } 1174 1175 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1176 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog, 1177 struct bpf_binary_header *ro_header, 1178 struct bpf_binary_header *rw_header) 1179 { 1180 void *ptr; 1181 1182 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1183 1184 kvfree(rw_header); 1185 1186 if (IS_ERR(ptr)) { 1187 bpf_prog_pack_free(ro_header, ro_header->size); 1188 return PTR_ERR(ptr); 1189 } 1190 return 0; 1191 } 1192 1193 /* bpf_jit_binary_pack_free is called in two different scenarios: 1194 * 1) when the program is freed after; 1195 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1196 * For case 2), we need to free both the RO memory and the RW buffer. 1197 * 1198 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1199 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1200 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1201 * bpf_arch_text_copy (when jit fails). 1202 */ 1203 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1204 struct bpf_binary_header *rw_header) 1205 { 1206 u32 size = ro_header->size; 1207 1208 bpf_prog_pack_free(ro_header, size); 1209 kvfree(rw_header); 1210 bpf_jit_uncharge_modmem(size); 1211 } 1212 1213 struct bpf_binary_header * 1214 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1215 { 1216 unsigned long real_start = (unsigned long)fp->bpf_func; 1217 unsigned long addr; 1218 1219 addr = real_start & BPF_PROG_CHUNK_MASK; 1220 return (void *)addr; 1221 } 1222 1223 static inline struct bpf_binary_header * 1224 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1225 { 1226 unsigned long real_start = (unsigned long)fp->bpf_func; 1227 unsigned long addr; 1228 1229 addr = real_start & PAGE_MASK; 1230 return (void *)addr; 1231 } 1232 1233 /* This symbol is only overridden by archs that have different 1234 * requirements than the usual eBPF JITs, f.e. when they only 1235 * implement cBPF JIT, do not set images read-only, etc. 1236 */ 1237 void __weak bpf_jit_free(struct bpf_prog *fp) 1238 { 1239 if (fp->jited) { 1240 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1241 1242 bpf_jit_binary_free(hdr); 1243 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1244 } 1245 1246 bpf_prog_unlock_free(fp); 1247 } 1248 1249 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1250 const struct bpf_insn *insn, bool extra_pass, 1251 u64 *func_addr, bool *func_addr_fixed) 1252 { 1253 s16 off = insn->off; 1254 s32 imm = insn->imm; 1255 u8 *addr; 1256 int err; 1257 1258 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1259 if (!*func_addr_fixed) { 1260 /* Place-holder address till the last pass has collected 1261 * all addresses for JITed subprograms in which case we 1262 * can pick them up from prog->aux. 1263 */ 1264 if (!extra_pass) 1265 addr = NULL; 1266 else if (prog->aux->func && 1267 off >= 0 && off < prog->aux->real_func_cnt) 1268 addr = (u8 *)prog->aux->func[off]->bpf_func; 1269 else 1270 return -EINVAL; 1271 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1272 bpf_jit_supports_far_kfunc_call()) { 1273 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1274 if (err) 1275 return err; 1276 } else { 1277 /* Address of a BPF helper call. Since part of the core 1278 * kernel, it's always at a fixed location. __bpf_call_base 1279 * and the helper with imm relative to it are both in core 1280 * kernel. 1281 */ 1282 addr = (u8 *)__bpf_call_base + imm; 1283 } 1284 1285 *func_addr = (unsigned long)addr; 1286 return 0; 1287 } 1288 1289 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1290 const struct bpf_insn *aux, 1291 struct bpf_insn *to_buff, 1292 bool emit_zext) 1293 { 1294 struct bpf_insn *to = to_buff; 1295 u32 imm_rnd = get_random_u32(); 1296 s16 off; 1297 1298 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1299 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1300 1301 /* Constraints on AX register: 1302 * 1303 * AX register is inaccessible from user space. It is mapped in 1304 * all JITs, and used here for constant blinding rewrites. It is 1305 * typically "stateless" meaning its contents are only valid within 1306 * the executed instruction, but not across several instructions. 1307 * There are a few exceptions however which are further detailed 1308 * below. 1309 * 1310 * Constant blinding is only used by JITs, not in the interpreter. 1311 * The interpreter uses AX in some occasions as a local temporary 1312 * register e.g. in DIV or MOD instructions. 1313 * 1314 * In restricted circumstances, the verifier can also use the AX 1315 * register for rewrites as long as they do not interfere with 1316 * the above cases! 1317 */ 1318 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1319 goto out; 1320 1321 if (from->imm == 0 && 1322 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1323 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1324 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1325 goto out; 1326 } 1327 1328 switch (from->code) { 1329 case BPF_ALU | BPF_ADD | BPF_K: 1330 case BPF_ALU | BPF_SUB | BPF_K: 1331 case BPF_ALU | BPF_AND | BPF_K: 1332 case BPF_ALU | BPF_OR | BPF_K: 1333 case BPF_ALU | BPF_XOR | BPF_K: 1334 case BPF_ALU | BPF_MUL | BPF_K: 1335 case BPF_ALU | BPF_MOV | BPF_K: 1336 case BPF_ALU | BPF_DIV | BPF_K: 1337 case BPF_ALU | BPF_MOD | BPF_K: 1338 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1339 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1340 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1341 break; 1342 1343 case BPF_ALU64 | BPF_ADD | BPF_K: 1344 case BPF_ALU64 | BPF_SUB | BPF_K: 1345 case BPF_ALU64 | BPF_AND | BPF_K: 1346 case BPF_ALU64 | BPF_OR | BPF_K: 1347 case BPF_ALU64 | BPF_XOR | BPF_K: 1348 case BPF_ALU64 | BPF_MUL | BPF_K: 1349 case BPF_ALU64 | BPF_MOV | BPF_K: 1350 case BPF_ALU64 | BPF_DIV | BPF_K: 1351 case BPF_ALU64 | BPF_MOD | BPF_K: 1352 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1353 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1354 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1355 break; 1356 1357 case BPF_JMP | BPF_JEQ | BPF_K: 1358 case BPF_JMP | BPF_JNE | BPF_K: 1359 case BPF_JMP | BPF_JGT | BPF_K: 1360 case BPF_JMP | BPF_JLT | BPF_K: 1361 case BPF_JMP | BPF_JGE | BPF_K: 1362 case BPF_JMP | BPF_JLE | BPF_K: 1363 case BPF_JMP | BPF_JSGT | BPF_K: 1364 case BPF_JMP | BPF_JSLT | BPF_K: 1365 case BPF_JMP | BPF_JSGE | BPF_K: 1366 case BPF_JMP | BPF_JSLE | BPF_K: 1367 case BPF_JMP | BPF_JSET | BPF_K: 1368 /* Accommodate for extra offset in case of a backjump. */ 1369 off = from->off; 1370 if (off < 0) 1371 off -= 2; 1372 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1373 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1374 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1375 break; 1376 1377 case BPF_JMP32 | BPF_JEQ | BPF_K: 1378 case BPF_JMP32 | BPF_JNE | BPF_K: 1379 case BPF_JMP32 | BPF_JGT | BPF_K: 1380 case BPF_JMP32 | BPF_JLT | BPF_K: 1381 case BPF_JMP32 | BPF_JGE | BPF_K: 1382 case BPF_JMP32 | BPF_JLE | BPF_K: 1383 case BPF_JMP32 | BPF_JSGT | BPF_K: 1384 case BPF_JMP32 | BPF_JSLT | BPF_K: 1385 case BPF_JMP32 | BPF_JSGE | BPF_K: 1386 case BPF_JMP32 | BPF_JSLE | BPF_K: 1387 case BPF_JMP32 | BPF_JSET | BPF_K: 1388 /* Accommodate for extra offset in case of a backjump. */ 1389 off = from->off; 1390 if (off < 0) 1391 off -= 2; 1392 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1393 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1394 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1395 off); 1396 break; 1397 1398 case BPF_LD | BPF_IMM | BPF_DW: 1399 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1400 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1401 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1402 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1403 break; 1404 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1405 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1406 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1407 if (emit_zext) 1408 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1409 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1410 break; 1411 1412 case BPF_ST | BPF_MEM | BPF_DW: 1413 case BPF_ST | BPF_MEM | BPF_W: 1414 case BPF_ST | BPF_MEM | BPF_H: 1415 case BPF_ST | BPF_MEM | BPF_B: 1416 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1417 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1418 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1419 break; 1420 } 1421 out: 1422 return to - to_buff; 1423 } 1424 1425 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1426 gfp_t gfp_extra_flags) 1427 { 1428 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1429 struct bpf_prog *fp; 1430 1431 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1432 if (fp != NULL) { 1433 /* aux->prog still points to the fp_other one, so 1434 * when promoting the clone to the real program, 1435 * this still needs to be adapted. 1436 */ 1437 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1438 } 1439 1440 return fp; 1441 } 1442 1443 static void bpf_prog_clone_free(struct bpf_prog *fp) 1444 { 1445 /* aux was stolen by the other clone, so we cannot free 1446 * it from this path! It will be freed eventually by the 1447 * other program on release. 1448 * 1449 * At this point, we don't need a deferred release since 1450 * clone is guaranteed to not be locked. 1451 */ 1452 fp->aux = NULL; 1453 fp->stats = NULL; 1454 fp->active = NULL; 1455 __bpf_prog_free(fp); 1456 } 1457 1458 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1459 { 1460 /* We have to repoint aux->prog to self, as we don't 1461 * know whether fp here is the clone or the original. 1462 */ 1463 fp->aux->prog = fp; 1464 bpf_prog_clone_free(fp_other); 1465 } 1466 1467 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1468 { 1469 struct bpf_insn insn_buff[16], aux[2]; 1470 struct bpf_prog *clone, *tmp; 1471 int insn_delta, insn_cnt; 1472 struct bpf_insn *insn; 1473 int i, rewritten; 1474 1475 if (!prog->blinding_requested || prog->blinded) 1476 return prog; 1477 1478 clone = bpf_prog_clone_create(prog, GFP_USER); 1479 if (!clone) 1480 return ERR_PTR(-ENOMEM); 1481 1482 insn_cnt = clone->len; 1483 insn = clone->insnsi; 1484 1485 for (i = 0; i < insn_cnt; i++, insn++) { 1486 if (bpf_pseudo_func(insn)) { 1487 /* ld_imm64 with an address of bpf subprog is not 1488 * a user controlled constant. Don't randomize it, 1489 * since it will conflict with jit_subprogs() logic. 1490 */ 1491 insn++; 1492 i++; 1493 continue; 1494 } 1495 1496 /* We temporarily need to hold the original ld64 insn 1497 * so that we can still access the first part in the 1498 * second blinding run. 1499 */ 1500 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1501 insn[1].code == 0) 1502 memcpy(aux, insn, sizeof(aux)); 1503 1504 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1505 clone->aux->verifier_zext); 1506 if (!rewritten) 1507 continue; 1508 1509 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1510 if (IS_ERR(tmp)) { 1511 /* Patching may have repointed aux->prog during 1512 * realloc from the original one, so we need to 1513 * fix it up here on error. 1514 */ 1515 bpf_jit_prog_release_other(prog, clone); 1516 return tmp; 1517 } 1518 1519 clone = tmp; 1520 insn_delta = rewritten - 1; 1521 1522 /* Walk new program and skip insns we just inserted. */ 1523 insn = clone->insnsi + i + insn_delta; 1524 insn_cnt += insn_delta; 1525 i += insn_delta; 1526 } 1527 1528 clone->blinded = 1; 1529 return clone; 1530 } 1531 #endif /* CONFIG_BPF_JIT */ 1532 1533 /* Base function for offset calculation. Needs to go into .text section, 1534 * therefore keeping it non-static as well; will also be used by JITs 1535 * anyway later on, so do not let the compiler omit it. This also needs 1536 * to go into kallsyms for correlation from e.g. bpftool, so naming 1537 * must not change. 1538 */ 1539 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1540 { 1541 return 0; 1542 } 1543 EXPORT_SYMBOL_GPL(__bpf_call_base); 1544 1545 /* All UAPI available opcodes. */ 1546 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1547 /* 32 bit ALU operations. */ \ 1548 /* Register based. */ \ 1549 INSN_3(ALU, ADD, X), \ 1550 INSN_3(ALU, SUB, X), \ 1551 INSN_3(ALU, AND, X), \ 1552 INSN_3(ALU, OR, X), \ 1553 INSN_3(ALU, LSH, X), \ 1554 INSN_3(ALU, RSH, X), \ 1555 INSN_3(ALU, XOR, X), \ 1556 INSN_3(ALU, MUL, X), \ 1557 INSN_3(ALU, MOV, X), \ 1558 INSN_3(ALU, ARSH, X), \ 1559 INSN_3(ALU, DIV, X), \ 1560 INSN_3(ALU, MOD, X), \ 1561 INSN_2(ALU, NEG), \ 1562 INSN_3(ALU, END, TO_BE), \ 1563 INSN_3(ALU, END, TO_LE), \ 1564 /* Immediate based. */ \ 1565 INSN_3(ALU, ADD, K), \ 1566 INSN_3(ALU, SUB, K), \ 1567 INSN_3(ALU, AND, K), \ 1568 INSN_3(ALU, OR, K), \ 1569 INSN_3(ALU, LSH, K), \ 1570 INSN_3(ALU, RSH, K), \ 1571 INSN_3(ALU, XOR, K), \ 1572 INSN_3(ALU, MUL, K), \ 1573 INSN_3(ALU, MOV, K), \ 1574 INSN_3(ALU, ARSH, K), \ 1575 INSN_3(ALU, DIV, K), \ 1576 INSN_3(ALU, MOD, K), \ 1577 /* 64 bit ALU operations. */ \ 1578 /* Register based. */ \ 1579 INSN_3(ALU64, ADD, X), \ 1580 INSN_3(ALU64, SUB, X), \ 1581 INSN_3(ALU64, AND, X), \ 1582 INSN_3(ALU64, OR, X), \ 1583 INSN_3(ALU64, LSH, X), \ 1584 INSN_3(ALU64, RSH, X), \ 1585 INSN_3(ALU64, XOR, X), \ 1586 INSN_3(ALU64, MUL, X), \ 1587 INSN_3(ALU64, MOV, X), \ 1588 INSN_3(ALU64, ARSH, X), \ 1589 INSN_3(ALU64, DIV, X), \ 1590 INSN_3(ALU64, MOD, X), \ 1591 INSN_2(ALU64, NEG), \ 1592 INSN_3(ALU64, END, TO_LE), \ 1593 /* Immediate based. */ \ 1594 INSN_3(ALU64, ADD, K), \ 1595 INSN_3(ALU64, SUB, K), \ 1596 INSN_3(ALU64, AND, K), \ 1597 INSN_3(ALU64, OR, K), \ 1598 INSN_3(ALU64, LSH, K), \ 1599 INSN_3(ALU64, RSH, K), \ 1600 INSN_3(ALU64, XOR, K), \ 1601 INSN_3(ALU64, MUL, K), \ 1602 INSN_3(ALU64, MOV, K), \ 1603 INSN_3(ALU64, ARSH, K), \ 1604 INSN_3(ALU64, DIV, K), \ 1605 INSN_3(ALU64, MOD, K), \ 1606 /* Call instruction. */ \ 1607 INSN_2(JMP, CALL), \ 1608 /* Exit instruction. */ \ 1609 INSN_2(JMP, EXIT), \ 1610 /* 32-bit Jump instructions. */ \ 1611 /* Register based. */ \ 1612 INSN_3(JMP32, JEQ, X), \ 1613 INSN_3(JMP32, JNE, X), \ 1614 INSN_3(JMP32, JGT, X), \ 1615 INSN_3(JMP32, JLT, X), \ 1616 INSN_3(JMP32, JGE, X), \ 1617 INSN_3(JMP32, JLE, X), \ 1618 INSN_3(JMP32, JSGT, X), \ 1619 INSN_3(JMP32, JSLT, X), \ 1620 INSN_3(JMP32, JSGE, X), \ 1621 INSN_3(JMP32, JSLE, X), \ 1622 INSN_3(JMP32, JSET, X), \ 1623 /* Immediate based. */ \ 1624 INSN_3(JMP32, JEQ, K), \ 1625 INSN_3(JMP32, JNE, K), \ 1626 INSN_3(JMP32, JGT, K), \ 1627 INSN_3(JMP32, JLT, K), \ 1628 INSN_3(JMP32, JGE, K), \ 1629 INSN_3(JMP32, JLE, K), \ 1630 INSN_3(JMP32, JSGT, K), \ 1631 INSN_3(JMP32, JSLT, K), \ 1632 INSN_3(JMP32, JSGE, K), \ 1633 INSN_3(JMP32, JSLE, K), \ 1634 INSN_3(JMP32, JSET, K), \ 1635 /* Jump instructions. */ \ 1636 /* Register based. */ \ 1637 INSN_3(JMP, JEQ, X), \ 1638 INSN_3(JMP, JNE, X), \ 1639 INSN_3(JMP, JGT, X), \ 1640 INSN_3(JMP, JLT, X), \ 1641 INSN_3(JMP, JGE, X), \ 1642 INSN_3(JMP, JLE, X), \ 1643 INSN_3(JMP, JSGT, X), \ 1644 INSN_3(JMP, JSLT, X), \ 1645 INSN_3(JMP, JSGE, X), \ 1646 INSN_3(JMP, JSLE, X), \ 1647 INSN_3(JMP, JSET, X), \ 1648 /* Immediate based. */ \ 1649 INSN_3(JMP, JEQ, K), \ 1650 INSN_3(JMP, JNE, K), \ 1651 INSN_3(JMP, JGT, K), \ 1652 INSN_3(JMP, JLT, K), \ 1653 INSN_3(JMP, JGE, K), \ 1654 INSN_3(JMP, JLE, K), \ 1655 INSN_3(JMP, JSGT, K), \ 1656 INSN_3(JMP, JSLT, K), \ 1657 INSN_3(JMP, JSGE, K), \ 1658 INSN_3(JMP, JSLE, K), \ 1659 INSN_3(JMP, JSET, K), \ 1660 INSN_2(JMP, JA), \ 1661 INSN_2(JMP32, JA), \ 1662 /* Store instructions. */ \ 1663 /* Register based. */ \ 1664 INSN_3(STX, MEM, B), \ 1665 INSN_3(STX, MEM, H), \ 1666 INSN_3(STX, MEM, W), \ 1667 INSN_3(STX, MEM, DW), \ 1668 INSN_3(STX, ATOMIC, W), \ 1669 INSN_3(STX, ATOMIC, DW), \ 1670 /* Immediate based. */ \ 1671 INSN_3(ST, MEM, B), \ 1672 INSN_3(ST, MEM, H), \ 1673 INSN_3(ST, MEM, W), \ 1674 INSN_3(ST, MEM, DW), \ 1675 /* Load instructions. */ \ 1676 /* Register based. */ \ 1677 INSN_3(LDX, MEM, B), \ 1678 INSN_3(LDX, MEM, H), \ 1679 INSN_3(LDX, MEM, W), \ 1680 INSN_3(LDX, MEM, DW), \ 1681 INSN_3(LDX, MEMSX, B), \ 1682 INSN_3(LDX, MEMSX, H), \ 1683 INSN_3(LDX, MEMSX, W), \ 1684 /* Immediate based. */ \ 1685 INSN_3(LD, IMM, DW) 1686 1687 bool bpf_opcode_in_insntable(u8 code) 1688 { 1689 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1690 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1691 static const bool public_insntable[256] = { 1692 [0 ... 255] = false, 1693 /* Now overwrite non-defaults ... */ 1694 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1695 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1696 [BPF_LD | BPF_ABS | BPF_B] = true, 1697 [BPF_LD | BPF_ABS | BPF_H] = true, 1698 [BPF_LD | BPF_ABS | BPF_W] = true, 1699 [BPF_LD | BPF_IND | BPF_B] = true, 1700 [BPF_LD | BPF_IND | BPF_H] = true, 1701 [BPF_LD | BPF_IND | BPF_W] = true, 1702 [BPF_JMP | BPF_JCOND] = true, 1703 }; 1704 #undef BPF_INSN_3_TBL 1705 #undef BPF_INSN_2_TBL 1706 return public_insntable[code]; 1707 } 1708 1709 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1710 /** 1711 * ___bpf_prog_run - run eBPF program on a given context 1712 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1713 * @insn: is the array of eBPF instructions 1714 * 1715 * Decode and execute eBPF instructions. 1716 * 1717 * Return: whatever value is in %BPF_R0 at program exit 1718 */ 1719 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1720 { 1721 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1722 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1723 static const void * const jumptable[256] __annotate_jump_table = { 1724 [0 ... 255] = &&default_label, 1725 /* Now overwrite non-defaults ... */ 1726 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1727 /* Non-UAPI available opcodes. */ 1728 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1729 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1730 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1731 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1732 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1733 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1734 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1735 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1736 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1737 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1738 }; 1739 #undef BPF_INSN_3_LBL 1740 #undef BPF_INSN_2_LBL 1741 u32 tail_call_cnt = 0; 1742 1743 #define CONT ({ insn++; goto select_insn; }) 1744 #define CONT_JMP ({ insn++; goto select_insn; }) 1745 1746 select_insn: 1747 goto *jumptable[insn->code]; 1748 1749 /* Explicitly mask the register-based shift amounts with 63 or 31 1750 * to avoid undefined behavior. Normally this won't affect the 1751 * generated code, for example, in case of native 64 bit archs such 1752 * as x86-64 or arm64, the compiler is optimizing the AND away for 1753 * the interpreter. In case of JITs, each of the JIT backends compiles 1754 * the BPF shift operations to machine instructions which produce 1755 * implementation-defined results in such a case; the resulting 1756 * contents of the register may be arbitrary, but program behaviour 1757 * as a whole remains defined. In other words, in case of JIT backends, 1758 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1759 */ 1760 /* ALU (shifts) */ 1761 #define SHT(OPCODE, OP) \ 1762 ALU64_##OPCODE##_X: \ 1763 DST = DST OP (SRC & 63); \ 1764 CONT; \ 1765 ALU_##OPCODE##_X: \ 1766 DST = (u32) DST OP ((u32) SRC & 31); \ 1767 CONT; \ 1768 ALU64_##OPCODE##_K: \ 1769 DST = DST OP IMM; \ 1770 CONT; \ 1771 ALU_##OPCODE##_K: \ 1772 DST = (u32) DST OP (u32) IMM; \ 1773 CONT; 1774 /* ALU (rest) */ 1775 #define ALU(OPCODE, OP) \ 1776 ALU64_##OPCODE##_X: \ 1777 DST = DST OP SRC; \ 1778 CONT; \ 1779 ALU_##OPCODE##_X: \ 1780 DST = (u32) DST OP (u32) SRC; \ 1781 CONT; \ 1782 ALU64_##OPCODE##_K: \ 1783 DST = DST OP IMM; \ 1784 CONT; \ 1785 ALU_##OPCODE##_K: \ 1786 DST = (u32) DST OP (u32) IMM; \ 1787 CONT; 1788 ALU(ADD, +) 1789 ALU(SUB, -) 1790 ALU(AND, &) 1791 ALU(OR, |) 1792 ALU(XOR, ^) 1793 ALU(MUL, *) 1794 SHT(LSH, <<) 1795 SHT(RSH, >>) 1796 #undef SHT 1797 #undef ALU 1798 ALU_NEG: 1799 DST = (u32) -DST; 1800 CONT; 1801 ALU64_NEG: 1802 DST = -DST; 1803 CONT; 1804 ALU_MOV_X: 1805 switch (OFF) { 1806 case 0: 1807 DST = (u32) SRC; 1808 break; 1809 case 8: 1810 DST = (u32)(s8) SRC; 1811 break; 1812 case 16: 1813 DST = (u32)(s16) SRC; 1814 break; 1815 } 1816 CONT; 1817 ALU_MOV_K: 1818 DST = (u32) IMM; 1819 CONT; 1820 ALU64_MOV_X: 1821 switch (OFF) { 1822 case 0: 1823 DST = SRC; 1824 break; 1825 case 8: 1826 DST = (s8) SRC; 1827 break; 1828 case 16: 1829 DST = (s16) SRC; 1830 break; 1831 case 32: 1832 DST = (s32) SRC; 1833 break; 1834 } 1835 CONT; 1836 ALU64_MOV_K: 1837 DST = IMM; 1838 CONT; 1839 LD_IMM_DW: 1840 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1841 insn++; 1842 CONT; 1843 ALU_ARSH_X: 1844 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1845 CONT; 1846 ALU_ARSH_K: 1847 DST = (u64) (u32) (((s32) DST) >> IMM); 1848 CONT; 1849 ALU64_ARSH_X: 1850 (*(s64 *) &DST) >>= (SRC & 63); 1851 CONT; 1852 ALU64_ARSH_K: 1853 (*(s64 *) &DST) >>= IMM; 1854 CONT; 1855 ALU64_MOD_X: 1856 switch (OFF) { 1857 case 0: 1858 div64_u64_rem(DST, SRC, &AX); 1859 DST = AX; 1860 break; 1861 case 1: 1862 AX = div64_s64(DST, SRC); 1863 DST = DST - AX * SRC; 1864 break; 1865 } 1866 CONT; 1867 ALU_MOD_X: 1868 switch (OFF) { 1869 case 0: 1870 AX = (u32) DST; 1871 DST = do_div(AX, (u32) SRC); 1872 break; 1873 case 1: 1874 AX = abs((s32)DST); 1875 AX = do_div(AX, abs((s32)SRC)); 1876 if ((s32)DST < 0) 1877 DST = (u32)-AX; 1878 else 1879 DST = (u32)AX; 1880 break; 1881 } 1882 CONT; 1883 ALU64_MOD_K: 1884 switch (OFF) { 1885 case 0: 1886 div64_u64_rem(DST, IMM, &AX); 1887 DST = AX; 1888 break; 1889 case 1: 1890 AX = div64_s64(DST, IMM); 1891 DST = DST - AX * IMM; 1892 break; 1893 } 1894 CONT; 1895 ALU_MOD_K: 1896 switch (OFF) { 1897 case 0: 1898 AX = (u32) DST; 1899 DST = do_div(AX, (u32) IMM); 1900 break; 1901 case 1: 1902 AX = abs((s32)DST); 1903 AX = do_div(AX, abs((s32)IMM)); 1904 if ((s32)DST < 0) 1905 DST = (u32)-AX; 1906 else 1907 DST = (u32)AX; 1908 break; 1909 } 1910 CONT; 1911 ALU64_DIV_X: 1912 switch (OFF) { 1913 case 0: 1914 DST = div64_u64(DST, SRC); 1915 break; 1916 case 1: 1917 DST = div64_s64(DST, SRC); 1918 break; 1919 } 1920 CONT; 1921 ALU_DIV_X: 1922 switch (OFF) { 1923 case 0: 1924 AX = (u32) DST; 1925 do_div(AX, (u32) SRC); 1926 DST = (u32) AX; 1927 break; 1928 case 1: 1929 AX = abs((s32)DST); 1930 do_div(AX, abs((s32)SRC)); 1931 if (((s32)DST < 0) == ((s32)SRC < 0)) 1932 DST = (u32)AX; 1933 else 1934 DST = (u32)-AX; 1935 break; 1936 } 1937 CONT; 1938 ALU64_DIV_K: 1939 switch (OFF) { 1940 case 0: 1941 DST = div64_u64(DST, IMM); 1942 break; 1943 case 1: 1944 DST = div64_s64(DST, IMM); 1945 break; 1946 } 1947 CONT; 1948 ALU_DIV_K: 1949 switch (OFF) { 1950 case 0: 1951 AX = (u32) DST; 1952 do_div(AX, (u32) IMM); 1953 DST = (u32) AX; 1954 break; 1955 case 1: 1956 AX = abs((s32)DST); 1957 do_div(AX, abs((s32)IMM)); 1958 if (((s32)DST < 0) == ((s32)IMM < 0)) 1959 DST = (u32)AX; 1960 else 1961 DST = (u32)-AX; 1962 break; 1963 } 1964 CONT; 1965 ALU_END_TO_BE: 1966 switch (IMM) { 1967 case 16: 1968 DST = (__force u16) cpu_to_be16(DST); 1969 break; 1970 case 32: 1971 DST = (__force u32) cpu_to_be32(DST); 1972 break; 1973 case 64: 1974 DST = (__force u64) cpu_to_be64(DST); 1975 break; 1976 } 1977 CONT; 1978 ALU_END_TO_LE: 1979 switch (IMM) { 1980 case 16: 1981 DST = (__force u16) cpu_to_le16(DST); 1982 break; 1983 case 32: 1984 DST = (__force u32) cpu_to_le32(DST); 1985 break; 1986 case 64: 1987 DST = (__force u64) cpu_to_le64(DST); 1988 break; 1989 } 1990 CONT; 1991 ALU64_END_TO_LE: 1992 switch (IMM) { 1993 case 16: 1994 DST = (__force u16) __swab16(DST); 1995 break; 1996 case 32: 1997 DST = (__force u32) __swab32(DST); 1998 break; 1999 case 64: 2000 DST = (__force u64) __swab64(DST); 2001 break; 2002 } 2003 CONT; 2004 2005 /* CALL */ 2006 JMP_CALL: 2007 /* Function call scratches BPF_R1-BPF_R5 registers, 2008 * preserves BPF_R6-BPF_R9, and stores return value 2009 * into BPF_R0. 2010 */ 2011 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2012 BPF_R4, BPF_R5); 2013 CONT; 2014 2015 JMP_CALL_ARGS: 2016 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2017 BPF_R3, BPF_R4, 2018 BPF_R5, 2019 insn + insn->off + 1); 2020 CONT; 2021 2022 JMP_TAIL_CALL: { 2023 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2024 struct bpf_array *array = container_of(map, struct bpf_array, map); 2025 struct bpf_prog *prog; 2026 u32 index = BPF_R3; 2027 2028 if (unlikely(index >= array->map.max_entries)) 2029 goto out; 2030 2031 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2032 goto out; 2033 2034 tail_call_cnt++; 2035 2036 prog = READ_ONCE(array->ptrs[index]); 2037 if (!prog) 2038 goto out; 2039 2040 /* ARG1 at this point is guaranteed to point to CTX from 2041 * the verifier side due to the fact that the tail call is 2042 * handled like a helper, that is, bpf_tail_call_proto, 2043 * where arg1_type is ARG_PTR_TO_CTX. 2044 */ 2045 insn = prog->insnsi; 2046 goto select_insn; 2047 out: 2048 CONT; 2049 } 2050 JMP_JA: 2051 insn += insn->off; 2052 CONT; 2053 JMP32_JA: 2054 insn += insn->imm; 2055 CONT; 2056 JMP_EXIT: 2057 return BPF_R0; 2058 /* JMP */ 2059 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2060 JMP_##OPCODE##_X: \ 2061 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2062 insn += insn->off; \ 2063 CONT_JMP; \ 2064 } \ 2065 CONT; \ 2066 JMP32_##OPCODE##_X: \ 2067 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2068 insn += insn->off; \ 2069 CONT_JMP; \ 2070 } \ 2071 CONT; \ 2072 JMP_##OPCODE##_K: \ 2073 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2074 insn += insn->off; \ 2075 CONT_JMP; \ 2076 } \ 2077 CONT; \ 2078 JMP32_##OPCODE##_K: \ 2079 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2080 insn += insn->off; \ 2081 CONT_JMP; \ 2082 } \ 2083 CONT; 2084 COND_JMP(u, JEQ, ==) 2085 COND_JMP(u, JNE, !=) 2086 COND_JMP(u, JGT, >) 2087 COND_JMP(u, JLT, <) 2088 COND_JMP(u, JGE, >=) 2089 COND_JMP(u, JLE, <=) 2090 COND_JMP(u, JSET, &) 2091 COND_JMP(s, JSGT, >) 2092 COND_JMP(s, JSLT, <) 2093 COND_JMP(s, JSGE, >=) 2094 COND_JMP(s, JSLE, <=) 2095 #undef COND_JMP 2096 /* ST, STX and LDX*/ 2097 ST_NOSPEC: 2098 /* Speculation barrier for mitigating Speculative Store Bypass. 2099 * In case of arm64, we rely on the firmware mitigation as 2100 * controlled via the ssbd kernel parameter. Whenever the 2101 * mitigation is enabled, it works for all of the kernel code 2102 * with no need to provide any additional instructions here. 2103 * In case of x86, we use 'lfence' insn for mitigation. We 2104 * reuse preexisting logic from Spectre v1 mitigation that 2105 * happens to produce the required code on x86 for v4 as well. 2106 */ 2107 barrier_nospec(); 2108 CONT; 2109 #define LDST(SIZEOP, SIZE) \ 2110 STX_MEM_##SIZEOP: \ 2111 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2112 CONT; \ 2113 ST_MEM_##SIZEOP: \ 2114 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2115 CONT; \ 2116 LDX_MEM_##SIZEOP: \ 2117 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2118 CONT; \ 2119 LDX_PROBE_MEM_##SIZEOP: \ 2120 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2121 (const void *)(long) (SRC + insn->off)); \ 2122 DST = *((SIZE *)&DST); \ 2123 CONT; 2124 2125 LDST(B, u8) 2126 LDST(H, u16) 2127 LDST(W, u32) 2128 LDST(DW, u64) 2129 #undef LDST 2130 2131 #define LDSX(SIZEOP, SIZE) \ 2132 LDX_MEMSX_##SIZEOP: \ 2133 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2134 CONT; \ 2135 LDX_PROBE_MEMSX_##SIZEOP: \ 2136 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2137 (const void *)(long) (SRC + insn->off)); \ 2138 DST = *((SIZE *)&DST); \ 2139 CONT; 2140 2141 LDSX(B, s8) 2142 LDSX(H, s16) 2143 LDSX(W, s32) 2144 #undef LDSX 2145 2146 #define ATOMIC_ALU_OP(BOP, KOP) \ 2147 case BOP: \ 2148 if (BPF_SIZE(insn->code) == BPF_W) \ 2149 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2150 (DST + insn->off)); \ 2151 else \ 2152 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2153 (DST + insn->off)); \ 2154 break; \ 2155 case BOP | BPF_FETCH: \ 2156 if (BPF_SIZE(insn->code) == BPF_W) \ 2157 SRC = (u32) atomic_fetch_##KOP( \ 2158 (u32) SRC, \ 2159 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2160 else \ 2161 SRC = (u64) atomic64_fetch_##KOP( \ 2162 (u64) SRC, \ 2163 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2164 break; 2165 2166 STX_ATOMIC_DW: 2167 STX_ATOMIC_W: 2168 switch (IMM) { 2169 ATOMIC_ALU_OP(BPF_ADD, add) 2170 ATOMIC_ALU_OP(BPF_AND, and) 2171 ATOMIC_ALU_OP(BPF_OR, or) 2172 ATOMIC_ALU_OP(BPF_XOR, xor) 2173 #undef ATOMIC_ALU_OP 2174 2175 case BPF_XCHG: 2176 if (BPF_SIZE(insn->code) == BPF_W) 2177 SRC = (u32) atomic_xchg( 2178 (atomic_t *)(unsigned long) (DST + insn->off), 2179 (u32) SRC); 2180 else 2181 SRC = (u64) atomic64_xchg( 2182 (atomic64_t *)(unsigned long) (DST + insn->off), 2183 (u64) SRC); 2184 break; 2185 case BPF_CMPXCHG: 2186 if (BPF_SIZE(insn->code) == BPF_W) 2187 BPF_R0 = (u32) atomic_cmpxchg( 2188 (atomic_t *)(unsigned long) (DST + insn->off), 2189 (u32) BPF_R0, (u32) SRC); 2190 else 2191 BPF_R0 = (u64) atomic64_cmpxchg( 2192 (atomic64_t *)(unsigned long) (DST + insn->off), 2193 (u64) BPF_R0, (u64) SRC); 2194 break; 2195 2196 default: 2197 goto default_label; 2198 } 2199 CONT; 2200 2201 default_label: 2202 /* If we ever reach this, we have a bug somewhere. Die hard here 2203 * instead of just returning 0; we could be somewhere in a subprog, 2204 * so execution could continue otherwise which we do /not/ want. 2205 * 2206 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2207 */ 2208 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2209 insn->code, insn->imm); 2210 BUG_ON(1); 2211 return 0; 2212 } 2213 2214 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2215 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2216 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2217 { \ 2218 u64 stack[stack_size / sizeof(u64)]; \ 2219 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2220 \ 2221 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2222 ARG1 = (u64) (unsigned long) ctx; \ 2223 return ___bpf_prog_run(regs, insn); \ 2224 } 2225 2226 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2227 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2228 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2229 const struct bpf_insn *insn) \ 2230 { \ 2231 u64 stack[stack_size / sizeof(u64)]; \ 2232 u64 regs[MAX_BPF_EXT_REG]; \ 2233 \ 2234 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2235 BPF_R1 = r1; \ 2236 BPF_R2 = r2; \ 2237 BPF_R3 = r3; \ 2238 BPF_R4 = r4; \ 2239 BPF_R5 = r5; \ 2240 return ___bpf_prog_run(regs, insn); \ 2241 } 2242 2243 #define EVAL1(FN, X) FN(X) 2244 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2245 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2246 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2247 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2248 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2249 2250 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2251 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2252 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2253 2254 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2255 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2256 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2257 2258 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2259 2260 static unsigned int (*interpreters[])(const void *ctx, 2261 const struct bpf_insn *insn) = { 2262 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2263 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2264 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2265 }; 2266 #undef PROG_NAME_LIST 2267 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2268 static __maybe_unused 2269 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2270 const struct bpf_insn *insn) = { 2271 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2272 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2273 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2274 }; 2275 #undef PROG_NAME_LIST 2276 2277 #ifdef CONFIG_BPF_SYSCALL 2278 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2279 { 2280 stack_depth = max_t(u32, stack_depth, 1); 2281 insn->off = (s16) insn->imm; 2282 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2283 __bpf_call_base_args; 2284 insn->code = BPF_JMP | BPF_CALL_ARGS; 2285 } 2286 #endif 2287 #else 2288 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2289 const struct bpf_insn *insn) 2290 { 2291 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2292 * is not working properly, so warn about it! 2293 */ 2294 WARN_ON_ONCE(1); 2295 return 0; 2296 } 2297 #endif 2298 2299 bool bpf_prog_map_compatible(struct bpf_map *map, 2300 const struct bpf_prog *fp) 2301 { 2302 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2303 bool ret; 2304 2305 if (fp->kprobe_override) 2306 return false; 2307 2308 /* XDP programs inserted into maps are not guaranteed to run on 2309 * a particular netdev (and can run outside driver context entirely 2310 * in the case of devmap and cpumap). Until device checks 2311 * are implemented, prohibit adding dev-bound programs to program maps. 2312 */ 2313 if (bpf_prog_is_dev_bound(fp->aux)) 2314 return false; 2315 2316 spin_lock(&map->owner.lock); 2317 if (!map->owner.type) { 2318 /* There's no owner yet where we could check for 2319 * compatibility. 2320 */ 2321 map->owner.type = prog_type; 2322 map->owner.jited = fp->jited; 2323 map->owner.xdp_has_frags = fp->aux->xdp_has_frags; 2324 ret = true; 2325 } else { 2326 ret = map->owner.type == prog_type && 2327 map->owner.jited == fp->jited && 2328 map->owner.xdp_has_frags == fp->aux->xdp_has_frags; 2329 } 2330 spin_unlock(&map->owner.lock); 2331 2332 return ret; 2333 } 2334 2335 static int bpf_check_tail_call(const struct bpf_prog *fp) 2336 { 2337 struct bpf_prog_aux *aux = fp->aux; 2338 int i, ret = 0; 2339 2340 mutex_lock(&aux->used_maps_mutex); 2341 for (i = 0; i < aux->used_map_cnt; i++) { 2342 struct bpf_map *map = aux->used_maps[i]; 2343 2344 if (!map_type_contains_progs(map)) 2345 continue; 2346 2347 if (!bpf_prog_map_compatible(map, fp)) { 2348 ret = -EINVAL; 2349 goto out; 2350 } 2351 } 2352 2353 out: 2354 mutex_unlock(&aux->used_maps_mutex); 2355 return ret; 2356 } 2357 2358 static void bpf_prog_select_func(struct bpf_prog *fp) 2359 { 2360 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2361 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2362 2363 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2364 #else 2365 fp->bpf_func = __bpf_prog_ret0_warn; 2366 #endif 2367 } 2368 2369 /** 2370 * bpf_prog_select_runtime - select exec runtime for BPF program 2371 * @fp: bpf_prog populated with BPF program 2372 * @err: pointer to error variable 2373 * 2374 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2375 * The BPF program will be executed via bpf_prog_run() function. 2376 * 2377 * Return: the &fp argument along with &err set to 0 for success or 2378 * a negative errno code on failure 2379 */ 2380 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2381 { 2382 /* In case of BPF to BPF calls, verifier did all the prep 2383 * work with regards to JITing, etc. 2384 */ 2385 bool jit_needed = false; 2386 2387 if (fp->bpf_func) 2388 goto finalize; 2389 2390 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2391 bpf_prog_has_kfunc_call(fp)) 2392 jit_needed = true; 2393 2394 bpf_prog_select_func(fp); 2395 2396 /* eBPF JITs can rewrite the program in case constant 2397 * blinding is active. However, in case of error during 2398 * blinding, bpf_int_jit_compile() must always return a 2399 * valid program, which in this case would simply not 2400 * be JITed, but falls back to the interpreter. 2401 */ 2402 if (!bpf_prog_is_offloaded(fp->aux)) { 2403 *err = bpf_prog_alloc_jited_linfo(fp); 2404 if (*err) 2405 return fp; 2406 2407 fp = bpf_int_jit_compile(fp); 2408 bpf_prog_jit_attempt_done(fp); 2409 if (!fp->jited && jit_needed) { 2410 *err = -ENOTSUPP; 2411 return fp; 2412 } 2413 } else { 2414 *err = bpf_prog_offload_compile(fp); 2415 if (*err) 2416 return fp; 2417 } 2418 2419 finalize: 2420 *err = bpf_prog_lock_ro(fp); 2421 if (*err) 2422 return fp; 2423 2424 /* The tail call compatibility check can only be done at 2425 * this late stage as we need to determine, if we deal 2426 * with JITed or non JITed program concatenations and not 2427 * all eBPF JITs might immediately support all features. 2428 */ 2429 *err = bpf_check_tail_call(fp); 2430 2431 return fp; 2432 } 2433 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2434 2435 static unsigned int __bpf_prog_ret1(const void *ctx, 2436 const struct bpf_insn *insn) 2437 { 2438 return 1; 2439 } 2440 2441 static struct bpf_prog_dummy { 2442 struct bpf_prog prog; 2443 } dummy_bpf_prog = { 2444 .prog = { 2445 .bpf_func = __bpf_prog_ret1, 2446 }, 2447 }; 2448 2449 struct bpf_empty_prog_array bpf_empty_prog_array = { 2450 .null_prog = NULL, 2451 }; 2452 EXPORT_SYMBOL(bpf_empty_prog_array); 2453 2454 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2455 { 2456 if (prog_cnt) 2457 return kzalloc(sizeof(struct bpf_prog_array) + 2458 sizeof(struct bpf_prog_array_item) * 2459 (prog_cnt + 1), 2460 flags); 2461 2462 return &bpf_empty_prog_array.hdr; 2463 } 2464 2465 void bpf_prog_array_free(struct bpf_prog_array *progs) 2466 { 2467 if (!progs || progs == &bpf_empty_prog_array.hdr) 2468 return; 2469 kfree_rcu(progs, rcu); 2470 } 2471 2472 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2473 { 2474 struct bpf_prog_array *progs; 2475 2476 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2477 * no need to call kfree_rcu(), just call kfree() directly. 2478 */ 2479 progs = container_of(rcu, struct bpf_prog_array, rcu); 2480 if (rcu_trace_implies_rcu_gp()) 2481 kfree(progs); 2482 else 2483 kfree_rcu(progs, rcu); 2484 } 2485 2486 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2487 { 2488 if (!progs || progs == &bpf_empty_prog_array.hdr) 2489 return; 2490 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2491 } 2492 2493 int bpf_prog_array_length(struct bpf_prog_array *array) 2494 { 2495 struct bpf_prog_array_item *item; 2496 u32 cnt = 0; 2497 2498 for (item = array->items; item->prog; item++) 2499 if (item->prog != &dummy_bpf_prog.prog) 2500 cnt++; 2501 return cnt; 2502 } 2503 2504 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2505 { 2506 struct bpf_prog_array_item *item; 2507 2508 for (item = array->items; item->prog; item++) 2509 if (item->prog != &dummy_bpf_prog.prog) 2510 return false; 2511 return true; 2512 } 2513 2514 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2515 u32 *prog_ids, 2516 u32 request_cnt) 2517 { 2518 struct bpf_prog_array_item *item; 2519 int i = 0; 2520 2521 for (item = array->items; item->prog; item++) { 2522 if (item->prog == &dummy_bpf_prog.prog) 2523 continue; 2524 prog_ids[i] = item->prog->aux->id; 2525 if (++i == request_cnt) { 2526 item++; 2527 break; 2528 } 2529 } 2530 2531 return !!(item->prog); 2532 } 2533 2534 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2535 __u32 __user *prog_ids, u32 cnt) 2536 { 2537 unsigned long err = 0; 2538 bool nospc; 2539 u32 *ids; 2540 2541 /* users of this function are doing: 2542 * cnt = bpf_prog_array_length(); 2543 * if (cnt > 0) 2544 * bpf_prog_array_copy_to_user(..., cnt); 2545 * so below kcalloc doesn't need extra cnt > 0 check. 2546 */ 2547 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2548 if (!ids) 2549 return -ENOMEM; 2550 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2551 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2552 kfree(ids); 2553 if (err) 2554 return -EFAULT; 2555 if (nospc) 2556 return -ENOSPC; 2557 return 0; 2558 } 2559 2560 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2561 struct bpf_prog *old_prog) 2562 { 2563 struct bpf_prog_array_item *item; 2564 2565 for (item = array->items; item->prog; item++) 2566 if (item->prog == old_prog) { 2567 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2568 break; 2569 } 2570 } 2571 2572 /** 2573 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2574 * index into the program array with 2575 * a dummy no-op program. 2576 * @array: a bpf_prog_array 2577 * @index: the index of the program to replace 2578 * 2579 * Skips over dummy programs, by not counting them, when calculating 2580 * the position of the program to replace. 2581 * 2582 * Return: 2583 * * 0 - Success 2584 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2585 * * -ENOENT - Index out of range 2586 */ 2587 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2588 { 2589 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2590 } 2591 2592 /** 2593 * bpf_prog_array_update_at() - Updates the program at the given index 2594 * into the program array. 2595 * @array: a bpf_prog_array 2596 * @index: the index of the program to update 2597 * @prog: the program to insert into the array 2598 * 2599 * Skips over dummy programs, by not counting them, when calculating 2600 * the position of the program to update. 2601 * 2602 * Return: 2603 * * 0 - Success 2604 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2605 * * -ENOENT - Index out of range 2606 */ 2607 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2608 struct bpf_prog *prog) 2609 { 2610 struct bpf_prog_array_item *item; 2611 2612 if (unlikely(index < 0)) 2613 return -EINVAL; 2614 2615 for (item = array->items; item->prog; item++) { 2616 if (item->prog == &dummy_bpf_prog.prog) 2617 continue; 2618 if (!index) { 2619 WRITE_ONCE(item->prog, prog); 2620 return 0; 2621 } 2622 index--; 2623 } 2624 return -ENOENT; 2625 } 2626 2627 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2628 struct bpf_prog *exclude_prog, 2629 struct bpf_prog *include_prog, 2630 u64 bpf_cookie, 2631 struct bpf_prog_array **new_array) 2632 { 2633 int new_prog_cnt, carry_prog_cnt = 0; 2634 struct bpf_prog_array_item *existing, *new; 2635 struct bpf_prog_array *array; 2636 bool found_exclude = false; 2637 2638 /* Figure out how many existing progs we need to carry over to 2639 * the new array. 2640 */ 2641 if (old_array) { 2642 existing = old_array->items; 2643 for (; existing->prog; existing++) { 2644 if (existing->prog == exclude_prog) { 2645 found_exclude = true; 2646 continue; 2647 } 2648 if (existing->prog != &dummy_bpf_prog.prog) 2649 carry_prog_cnt++; 2650 if (existing->prog == include_prog) 2651 return -EEXIST; 2652 } 2653 } 2654 2655 if (exclude_prog && !found_exclude) 2656 return -ENOENT; 2657 2658 /* How many progs (not NULL) will be in the new array? */ 2659 new_prog_cnt = carry_prog_cnt; 2660 if (include_prog) 2661 new_prog_cnt += 1; 2662 2663 /* Do we have any prog (not NULL) in the new array? */ 2664 if (!new_prog_cnt) { 2665 *new_array = NULL; 2666 return 0; 2667 } 2668 2669 /* +1 as the end of prog_array is marked with NULL */ 2670 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2671 if (!array) 2672 return -ENOMEM; 2673 new = array->items; 2674 2675 /* Fill in the new prog array */ 2676 if (carry_prog_cnt) { 2677 existing = old_array->items; 2678 for (; existing->prog; existing++) { 2679 if (existing->prog == exclude_prog || 2680 existing->prog == &dummy_bpf_prog.prog) 2681 continue; 2682 2683 new->prog = existing->prog; 2684 new->bpf_cookie = existing->bpf_cookie; 2685 new++; 2686 } 2687 } 2688 if (include_prog) { 2689 new->prog = include_prog; 2690 new->bpf_cookie = bpf_cookie; 2691 new++; 2692 } 2693 new->prog = NULL; 2694 *new_array = array; 2695 return 0; 2696 } 2697 2698 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2699 u32 *prog_ids, u32 request_cnt, 2700 u32 *prog_cnt) 2701 { 2702 u32 cnt = 0; 2703 2704 if (array) 2705 cnt = bpf_prog_array_length(array); 2706 2707 *prog_cnt = cnt; 2708 2709 /* return early if user requested only program count or nothing to copy */ 2710 if (!request_cnt || !cnt) 2711 return 0; 2712 2713 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2714 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2715 : 0; 2716 } 2717 2718 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2719 struct bpf_map **used_maps, u32 len) 2720 { 2721 struct bpf_map *map; 2722 bool sleepable; 2723 u32 i; 2724 2725 sleepable = aux->prog->sleepable; 2726 for (i = 0; i < len; i++) { 2727 map = used_maps[i]; 2728 if (map->ops->map_poke_untrack) 2729 map->ops->map_poke_untrack(map, aux); 2730 if (sleepable) 2731 atomic64_dec(&map->sleepable_refcnt); 2732 bpf_map_put(map); 2733 } 2734 } 2735 2736 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2737 { 2738 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2739 kfree(aux->used_maps); 2740 } 2741 2742 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2743 struct btf_mod_pair *used_btfs, u32 len) 2744 { 2745 #ifdef CONFIG_BPF_SYSCALL 2746 struct btf_mod_pair *btf_mod; 2747 u32 i; 2748 2749 for (i = 0; i < len; i++) { 2750 btf_mod = &used_btfs[i]; 2751 if (btf_mod->module) 2752 module_put(btf_mod->module); 2753 btf_put(btf_mod->btf); 2754 } 2755 #endif 2756 } 2757 2758 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2759 { 2760 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2761 kfree(aux->used_btfs); 2762 } 2763 2764 static void bpf_prog_free_deferred(struct work_struct *work) 2765 { 2766 struct bpf_prog_aux *aux; 2767 int i; 2768 2769 aux = container_of(work, struct bpf_prog_aux, work); 2770 #ifdef CONFIG_BPF_SYSCALL 2771 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2772 #endif 2773 #ifdef CONFIG_CGROUP_BPF 2774 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2775 bpf_cgroup_atype_put(aux->cgroup_atype); 2776 #endif 2777 bpf_free_used_maps(aux); 2778 bpf_free_used_btfs(aux); 2779 if (bpf_prog_is_dev_bound(aux)) 2780 bpf_prog_dev_bound_destroy(aux->prog); 2781 #ifdef CONFIG_PERF_EVENTS 2782 if (aux->prog->has_callchain_buf) 2783 put_callchain_buffers(); 2784 #endif 2785 if (aux->dst_trampoline) 2786 bpf_trampoline_put(aux->dst_trampoline); 2787 for (i = 0; i < aux->real_func_cnt; i++) { 2788 /* We can just unlink the subprog poke descriptor table as 2789 * it was originally linked to the main program and is also 2790 * released along with it. 2791 */ 2792 aux->func[i]->aux->poke_tab = NULL; 2793 bpf_jit_free(aux->func[i]); 2794 } 2795 if (aux->real_func_cnt) { 2796 kfree(aux->func); 2797 bpf_prog_unlock_free(aux->prog); 2798 } else { 2799 bpf_jit_free(aux->prog); 2800 } 2801 } 2802 2803 void bpf_prog_free(struct bpf_prog *fp) 2804 { 2805 struct bpf_prog_aux *aux = fp->aux; 2806 2807 if (aux->dst_prog) 2808 bpf_prog_put(aux->dst_prog); 2809 bpf_token_put(aux->token); 2810 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2811 schedule_work(&aux->work); 2812 } 2813 EXPORT_SYMBOL_GPL(bpf_prog_free); 2814 2815 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2816 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2817 2818 void bpf_user_rnd_init_once(void) 2819 { 2820 prandom_init_once(&bpf_user_rnd_state); 2821 } 2822 2823 BPF_CALL_0(bpf_user_rnd_u32) 2824 { 2825 /* Should someone ever have the rather unwise idea to use some 2826 * of the registers passed into this function, then note that 2827 * this function is called from native eBPF and classic-to-eBPF 2828 * transformations. Register assignments from both sides are 2829 * different, f.e. classic always sets fn(ctx, A, X) here. 2830 */ 2831 struct rnd_state *state; 2832 u32 res; 2833 2834 state = &get_cpu_var(bpf_user_rnd_state); 2835 res = prandom_u32_state(state); 2836 put_cpu_var(bpf_user_rnd_state); 2837 2838 return res; 2839 } 2840 2841 BPF_CALL_0(bpf_get_raw_cpu_id) 2842 { 2843 return raw_smp_processor_id(); 2844 } 2845 2846 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2847 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2848 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2849 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2850 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2851 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2852 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2853 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2854 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2855 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2856 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2857 2858 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2859 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2860 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2861 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2862 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2863 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2864 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 2865 2866 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2867 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2868 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2869 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2870 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2871 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2872 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2873 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2874 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2875 const struct bpf_func_proto bpf_set_retval_proto __weak; 2876 const struct bpf_func_proto bpf_get_retval_proto __weak; 2877 2878 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2879 { 2880 return NULL; 2881 } 2882 2883 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2884 { 2885 return NULL; 2886 } 2887 2888 u64 __weak 2889 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2890 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2891 { 2892 return -ENOTSUPP; 2893 } 2894 EXPORT_SYMBOL_GPL(bpf_event_output); 2895 2896 /* Always built-in helper functions. */ 2897 const struct bpf_func_proto bpf_tail_call_proto = { 2898 .func = NULL, 2899 .gpl_only = false, 2900 .ret_type = RET_VOID, 2901 .arg1_type = ARG_PTR_TO_CTX, 2902 .arg2_type = ARG_CONST_MAP_PTR, 2903 .arg3_type = ARG_ANYTHING, 2904 }; 2905 2906 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2907 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2908 * eBPF and implicitly also cBPF can get JITed! 2909 */ 2910 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2911 { 2912 return prog; 2913 } 2914 2915 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2916 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2917 */ 2918 void __weak bpf_jit_compile(struct bpf_prog *prog) 2919 { 2920 } 2921 2922 bool __weak bpf_helper_changes_pkt_data(void *func) 2923 { 2924 return false; 2925 } 2926 2927 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2928 * analysis code and wants explicit zero extension inserted by verifier. 2929 * Otherwise, return FALSE. 2930 * 2931 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2932 * you don't override this. JITs that don't want these extra insns can detect 2933 * them using insn_is_zext. 2934 */ 2935 bool __weak bpf_jit_needs_zext(void) 2936 { 2937 return false; 2938 } 2939 2940 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 2941 bool __weak bpf_jit_supports_subprog_tailcalls(void) 2942 { 2943 return false; 2944 } 2945 2946 bool __weak bpf_jit_supports_kfunc_call(void) 2947 { 2948 return false; 2949 } 2950 2951 bool __weak bpf_jit_supports_far_kfunc_call(void) 2952 { 2953 return false; 2954 } 2955 2956 bool __weak bpf_jit_supports_arena(void) 2957 { 2958 return false; 2959 } 2960 2961 /* Return TRUE if the JIT backend satisfies the following two conditions: 2962 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 2963 * 2) Under the specific arch, the implementation of xchg() is the same 2964 * as atomic_xchg() on pointer-sized words. 2965 */ 2966 bool __weak bpf_jit_supports_ptr_xchg(void) 2967 { 2968 return false; 2969 } 2970 2971 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2972 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2973 */ 2974 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2975 int len) 2976 { 2977 return -EFAULT; 2978 } 2979 2980 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2981 void *addr1, void *addr2) 2982 { 2983 return -ENOTSUPP; 2984 } 2985 2986 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 2987 { 2988 return ERR_PTR(-ENOTSUPP); 2989 } 2990 2991 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 2992 { 2993 return -ENOTSUPP; 2994 } 2995 2996 bool __weak bpf_jit_supports_exceptions(void) 2997 { 2998 return false; 2999 } 3000 3001 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3002 { 3003 } 3004 3005 /* for configs without MMU or 32-bit */ 3006 __weak const struct bpf_map_ops arena_map_ops; 3007 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3008 { 3009 return 0; 3010 } 3011 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3012 { 3013 return 0; 3014 } 3015 3016 #ifdef CONFIG_BPF_SYSCALL 3017 static int __init bpf_global_ma_init(void) 3018 { 3019 int ret; 3020 3021 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3022 bpf_global_ma_set = !ret; 3023 return ret; 3024 } 3025 late_initcall(bpf_global_ma_init); 3026 #endif 3027 3028 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3029 EXPORT_SYMBOL(bpf_stats_enabled_key); 3030 3031 /* All definitions of tracepoints related to BPF. */ 3032 #define CREATE_TRACE_POINTS 3033 #include <linux/bpf_trace.h> 3034 3035 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3036 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3037