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; 911 912 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 913 GFP_KERNEL); 914 if (!pack) 915 return NULL; 916 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 917 if (!pack->ptr) { 918 kfree(pack); 919 return NULL; 920 } 921 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 922 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 923 list_add_tail(&pack->list, &pack_list); 924 925 set_vm_flush_reset_perms(pack->ptr); 926 set_memory_rox((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE); 927 return pack; 928 } 929 930 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 931 { 932 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 933 struct bpf_prog_pack *pack; 934 unsigned long pos; 935 void *ptr = NULL; 936 937 mutex_lock(&pack_mutex); 938 if (size > BPF_PROG_PACK_SIZE) { 939 size = round_up(size, PAGE_SIZE); 940 ptr = bpf_jit_alloc_exec(size); 941 if (ptr) { 942 bpf_fill_ill_insns(ptr, size); 943 set_vm_flush_reset_perms(ptr); 944 set_memory_rox((unsigned long)ptr, size / PAGE_SIZE); 945 } 946 goto out; 947 } 948 list_for_each_entry(pack, &pack_list, list) { 949 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 950 nbits, 0); 951 if (pos < BPF_PROG_CHUNK_COUNT) 952 goto found_free_area; 953 } 954 955 pack = alloc_new_pack(bpf_fill_ill_insns); 956 if (!pack) 957 goto out; 958 959 pos = 0; 960 961 found_free_area: 962 bitmap_set(pack->bitmap, pos, nbits); 963 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 964 965 out: 966 mutex_unlock(&pack_mutex); 967 return ptr; 968 } 969 970 void bpf_prog_pack_free(void *ptr, u32 size) 971 { 972 struct bpf_prog_pack *pack = NULL, *tmp; 973 unsigned int nbits; 974 unsigned long pos; 975 976 mutex_lock(&pack_mutex); 977 if (size > BPF_PROG_PACK_SIZE) { 978 bpf_jit_free_exec(ptr); 979 goto out; 980 } 981 982 list_for_each_entry(tmp, &pack_list, list) { 983 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 984 pack = tmp; 985 break; 986 } 987 } 988 989 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 990 goto out; 991 992 nbits = BPF_PROG_SIZE_TO_NBITS(size); 993 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 994 995 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 996 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 997 998 bitmap_clear(pack->bitmap, pos, nbits); 999 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1000 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1001 list_del(&pack->list); 1002 bpf_jit_free_exec(pack->ptr); 1003 kfree(pack); 1004 } 1005 out: 1006 mutex_unlock(&pack_mutex); 1007 } 1008 1009 static atomic_long_t bpf_jit_current; 1010 1011 /* Can be overridden by an arch's JIT compiler if it has a custom, 1012 * dedicated BPF backend memory area, or if neither of the two 1013 * below apply. 1014 */ 1015 u64 __weak bpf_jit_alloc_exec_limit(void) 1016 { 1017 #if defined(MODULES_VADDR) 1018 return MODULES_END - MODULES_VADDR; 1019 #else 1020 return VMALLOC_END - VMALLOC_START; 1021 #endif 1022 } 1023 1024 static int __init bpf_jit_charge_init(void) 1025 { 1026 /* Only used as heuristic here to derive limit. */ 1027 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1028 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1029 PAGE_SIZE), LONG_MAX); 1030 return 0; 1031 } 1032 pure_initcall(bpf_jit_charge_init); 1033 1034 int bpf_jit_charge_modmem(u32 size) 1035 { 1036 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1037 if (!bpf_capable()) { 1038 atomic_long_sub(size, &bpf_jit_current); 1039 return -EPERM; 1040 } 1041 } 1042 1043 return 0; 1044 } 1045 1046 void bpf_jit_uncharge_modmem(u32 size) 1047 { 1048 atomic_long_sub(size, &bpf_jit_current); 1049 } 1050 1051 void *__weak bpf_jit_alloc_exec(unsigned long size) 1052 { 1053 return module_alloc(size); 1054 } 1055 1056 void __weak bpf_jit_free_exec(void *addr) 1057 { 1058 module_memfree(addr); 1059 } 1060 1061 struct bpf_binary_header * 1062 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1063 unsigned int alignment, 1064 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1065 { 1066 struct bpf_binary_header *hdr; 1067 u32 size, hole, start; 1068 1069 WARN_ON_ONCE(!is_power_of_2(alignment) || 1070 alignment > BPF_IMAGE_ALIGNMENT); 1071 1072 /* Most of BPF filters are really small, but if some of them 1073 * fill a page, allow at least 128 extra bytes to insert a 1074 * random section of illegal instructions. 1075 */ 1076 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1077 1078 if (bpf_jit_charge_modmem(size)) 1079 return NULL; 1080 hdr = bpf_jit_alloc_exec(size); 1081 if (!hdr) { 1082 bpf_jit_uncharge_modmem(size); 1083 return NULL; 1084 } 1085 1086 /* Fill space with illegal/arch-dep instructions. */ 1087 bpf_fill_ill_insns(hdr, size); 1088 1089 hdr->size = size; 1090 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1091 PAGE_SIZE - sizeof(*hdr)); 1092 start = get_random_u32_below(hole) & ~(alignment - 1); 1093 1094 /* Leave a random number of instructions before BPF code. */ 1095 *image_ptr = &hdr->image[start]; 1096 1097 return hdr; 1098 } 1099 1100 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1101 { 1102 u32 size = hdr->size; 1103 1104 bpf_jit_free_exec(hdr); 1105 bpf_jit_uncharge_modmem(size); 1106 } 1107 1108 /* Allocate jit binary from bpf_prog_pack allocator. 1109 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1110 * to the memory. To solve this problem, a RW buffer is also allocated at 1111 * as the same time. The JIT engine should calculate offsets based on the 1112 * RO memory address, but write JITed program to the RW buffer. Once the 1113 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1114 * the JITed program to the RO memory. 1115 */ 1116 struct bpf_binary_header * 1117 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1118 unsigned int alignment, 1119 struct bpf_binary_header **rw_header, 1120 u8 **rw_image, 1121 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1122 { 1123 struct bpf_binary_header *ro_header; 1124 u32 size, hole, start; 1125 1126 WARN_ON_ONCE(!is_power_of_2(alignment) || 1127 alignment > BPF_IMAGE_ALIGNMENT); 1128 1129 /* add 16 bytes for a random section of illegal instructions */ 1130 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1131 1132 if (bpf_jit_charge_modmem(size)) 1133 return NULL; 1134 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1135 if (!ro_header) { 1136 bpf_jit_uncharge_modmem(size); 1137 return NULL; 1138 } 1139 1140 *rw_header = kvmalloc(size, GFP_KERNEL); 1141 if (!*rw_header) { 1142 bpf_prog_pack_free(ro_header, size); 1143 bpf_jit_uncharge_modmem(size); 1144 return NULL; 1145 } 1146 1147 /* Fill space with illegal/arch-dep instructions. */ 1148 bpf_fill_ill_insns(*rw_header, size); 1149 (*rw_header)->size = size; 1150 1151 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1152 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1153 start = get_random_u32_below(hole) & ~(alignment - 1); 1154 1155 *image_ptr = &ro_header->image[start]; 1156 *rw_image = &(*rw_header)->image[start]; 1157 1158 return ro_header; 1159 } 1160 1161 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1162 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog, 1163 struct bpf_binary_header *ro_header, 1164 struct bpf_binary_header *rw_header) 1165 { 1166 void *ptr; 1167 1168 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1169 1170 kvfree(rw_header); 1171 1172 if (IS_ERR(ptr)) { 1173 bpf_prog_pack_free(ro_header, ro_header->size); 1174 return PTR_ERR(ptr); 1175 } 1176 return 0; 1177 } 1178 1179 /* bpf_jit_binary_pack_free is called in two different scenarios: 1180 * 1) when the program is freed after; 1181 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1182 * For case 2), we need to free both the RO memory and the RW buffer. 1183 * 1184 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1185 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1186 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1187 * bpf_arch_text_copy (when jit fails). 1188 */ 1189 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1190 struct bpf_binary_header *rw_header) 1191 { 1192 u32 size = ro_header->size; 1193 1194 bpf_prog_pack_free(ro_header, size); 1195 kvfree(rw_header); 1196 bpf_jit_uncharge_modmem(size); 1197 } 1198 1199 struct bpf_binary_header * 1200 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1201 { 1202 unsigned long real_start = (unsigned long)fp->bpf_func; 1203 unsigned long addr; 1204 1205 addr = real_start & BPF_PROG_CHUNK_MASK; 1206 return (void *)addr; 1207 } 1208 1209 static inline struct bpf_binary_header * 1210 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1211 { 1212 unsigned long real_start = (unsigned long)fp->bpf_func; 1213 unsigned long addr; 1214 1215 addr = real_start & PAGE_MASK; 1216 return (void *)addr; 1217 } 1218 1219 /* This symbol is only overridden by archs that have different 1220 * requirements than the usual eBPF JITs, f.e. when they only 1221 * implement cBPF JIT, do not set images read-only, etc. 1222 */ 1223 void __weak bpf_jit_free(struct bpf_prog *fp) 1224 { 1225 if (fp->jited) { 1226 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1227 1228 bpf_jit_binary_free(hdr); 1229 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1230 } 1231 1232 bpf_prog_unlock_free(fp); 1233 } 1234 1235 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1236 const struct bpf_insn *insn, bool extra_pass, 1237 u64 *func_addr, bool *func_addr_fixed) 1238 { 1239 s16 off = insn->off; 1240 s32 imm = insn->imm; 1241 u8 *addr; 1242 int err; 1243 1244 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1245 if (!*func_addr_fixed) { 1246 /* Place-holder address till the last pass has collected 1247 * all addresses for JITed subprograms in which case we 1248 * can pick them up from prog->aux. 1249 */ 1250 if (!extra_pass) 1251 addr = NULL; 1252 else if (prog->aux->func && 1253 off >= 0 && off < prog->aux->real_func_cnt) 1254 addr = (u8 *)prog->aux->func[off]->bpf_func; 1255 else 1256 return -EINVAL; 1257 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1258 bpf_jit_supports_far_kfunc_call()) { 1259 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1260 if (err) 1261 return err; 1262 } else { 1263 /* Address of a BPF helper call. Since part of the core 1264 * kernel, it's always at a fixed location. __bpf_call_base 1265 * and the helper with imm relative to it are both in core 1266 * kernel. 1267 */ 1268 addr = (u8 *)__bpf_call_base + imm; 1269 } 1270 1271 *func_addr = (unsigned long)addr; 1272 return 0; 1273 } 1274 1275 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1276 const struct bpf_insn *aux, 1277 struct bpf_insn *to_buff, 1278 bool emit_zext) 1279 { 1280 struct bpf_insn *to = to_buff; 1281 u32 imm_rnd = get_random_u32(); 1282 s16 off; 1283 1284 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1285 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1286 1287 /* Constraints on AX register: 1288 * 1289 * AX register is inaccessible from user space. It is mapped in 1290 * all JITs, and used here for constant blinding rewrites. It is 1291 * typically "stateless" meaning its contents are only valid within 1292 * the executed instruction, but not across several instructions. 1293 * There are a few exceptions however which are further detailed 1294 * below. 1295 * 1296 * Constant blinding is only used by JITs, not in the interpreter. 1297 * The interpreter uses AX in some occasions as a local temporary 1298 * register e.g. in DIV or MOD instructions. 1299 * 1300 * In restricted circumstances, the verifier can also use the AX 1301 * register for rewrites as long as they do not interfere with 1302 * the above cases! 1303 */ 1304 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1305 goto out; 1306 1307 if (from->imm == 0 && 1308 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1309 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1310 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1311 goto out; 1312 } 1313 1314 switch (from->code) { 1315 case BPF_ALU | BPF_ADD | BPF_K: 1316 case BPF_ALU | BPF_SUB | BPF_K: 1317 case BPF_ALU | BPF_AND | BPF_K: 1318 case BPF_ALU | BPF_OR | BPF_K: 1319 case BPF_ALU | BPF_XOR | BPF_K: 1320 case BPF_ALU | BPF_MUL | BPF_K: 1321 case BPF_ALU | BPF_MOV | BPF_K: 1322 case BPF_ALU | BPF_DIV | BPF_K: 1323 case BPF_ALU | BPF_MOD | BPF_K: 1324 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1325 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1326 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1327 break; 1328 1329 case BPF_ALU64 | BPF_ADD | BPF_K: 1330 case BPF_ALU64 | BPF_SUB | BPF_K: 1331 case BPF_ALU64 | BPF_AND | BPF_K: 1332 case BPF_ALU64 | BPF_OR | BPF_K: 1333 case BPF_ALU64 | BPF_XOR | BPF_K: 1334 case BPF_ALU64 | BPF_MUL | BPF_K: 1335 case BPF_ALU64 | BPF_MOV | BPF_K: 1336 case BPF_ALU64 | BPF_DIV | BPF_K: 1337 case BPF_ALU64 | BPF_MOD | BPF_K: 1338 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1339 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1340 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1341 break; 1342 1343 case BPF_JMP | BPF_JEQ | BPF_K: 1344 case BPF_JMP | BPF_JNE | BPF_K: 1345 case BPF_JMP | BPF_JGT | BPF_K: 1346 case BPF_JMP | BPF_JLT | BPF_K: 1347 case BPF_JMP | BPF_JGE | BPF_K: 1348 case BPF_JMP | BPF_JLE | BPF_K: 1349 case BPF_JMP | BPF_JSGT | BPF_K: 1350 case BPF_JMP | BPF_JSLT | BPF_K: 1351 case BPF_JMP | BPF_JSGE | BPF_K: 1352 case BPF_JMP | BPF_JSLE | BPF_K: 1353 case BPF_JMP | BPF_JSET | BPF_K: 1354 /* Accommodate for extra offset in case of a backjump. */ 1355 off = from->off; 1356 if (off < 0) 1357 off -= 2; 1358 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1359 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1360 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1361 break; 1362 1363 case BPF_JMP32 | BPF_JEQ | BPF_K: 1364 case BPF_JMP32 | BPF_JNE | BPF_K: 1365 case BPF_JMP32 | BPF_JGT | BPF_K: 1366 case BPF_JMP32 | BPF_JLT | BPF_K: 1367 case BPF_JMP32 | BPF_JGE | BPF_K: 1368 case BPF_JMP32 | BPF_JLE | BPF_K: 1369 case BPF_JMP32 | BPF_JSGT | BPF_K: 1370 case BPF_JMP32 | BPF_JSLT | BPF_K: 1371 case BPF_JMP32 | BPF_JSGE | BPF_K: 1372 case BPF_JMP32 | BPF_JSLE | BPF_K: 1373 case BPF_JMP32 | BPF_JSET | BPF_K: 1374 /* Accommodate for extra offset in case of a backjump. */ 1375 off = from->off; 1376 if (off < 0) 1377 off -= 2; 1378 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1379 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1380 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1381 off); 1382 break; 1383 1384 case BPF_LD | BPF_IMM | BPF_DW: 1385 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1386 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1387 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1388 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1389 break; 1390 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1391 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1392 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1393 if (emit_zext) 1394 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1395 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1396 break; 1397 1398 case BPF_ST | BPF_MEM | BPF_DW: 1399 case BPF_ST | BPF_MEM | BPF_W: 1400 case BPF_ST | BPF_MEM | BPF_H: 1401 case BPF_ST | BPF_MEM | BPF_B: 1402 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1403 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1404 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1405 break; 1406 } 1407 out: 1408 return to - to_buff; 1409 } 1410 1411 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1412 gfp_t gfp_extra_flags) 1413 { 1414 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1415 struct bpf_prog *fp; 1416 1417 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1418 if (fp != NULL) { 1419 /* aux->prog still points to the fp_other one, so 1420 * when promoting the clone to the real program, 1421 * this still needs to be adapted. 1422 */ 1423 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1424 } 1425 1426 return fp; 1427 } 1428 1429 static void bpf_prog_clone_free(struct bpf_prog *fp) 1430 { 1431 /* aux was stolen by the other clone, so we cannot free 1432 * it from this path! It will be freed eventually by the 1433 * other program on release. 1434 * 1435 * At this point, we don't need a deferred release since 1436 * clone is guaranteed to not be locked. 1437 */ 1438 fp->aux = NULL; 1439 fp->stats = NULL; 1440 fp->active = NULL; 1441 __bpf_prog_free(fp); 1442 } 1443 1444 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1445 { 1446 /* We have to repoint aux->prog to self, as we don't 1447 * know whether fp here is the clone or the original. 1448 */ 1449 fp->aux->prog = fp; 1450 bpf_prog_clone_free(fp_other); 1451 } 1452 1453 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1454 { 1455 struct bpf_insn insn_buff[16], aux[2]; 1456 struct bpf_prog *clone, *tmp; 1457 int insn_delta, insn_cnt; 1458 struct bpf_insn *insn; 1459 int i, rewritten; 1460 1461 if (!prog->blinding_requested || prog->blinded) 1462 return prog; 1463 1464 clone = bpf_prog_clone_create(prog, GFP_USER); 1465 if (!clone) 1466 return ERR_PTR(-ENOMEM); 1467 1468 insn_cnt = clone->len; 1469 insn = clone->insnsi; 1470 1471 for (i = 0; i < insn_cnt; i++, insn++) { 1472 if (bpf_pseudo_func(insn)) { 1473 /* ld_imm64 with an address of bpf subprog is not 1474 * a user controlled constant. Don't randomize it, 1475 * since it will conflict with jit_subprogs() logic. 1476 */ 1477 insn++; 1478 i++; 1479 continue; 1480 } 1481 1482 /* We temporarily need to hold the original ld64 insn 1483 * so that we can still access the first part in the 1484 * second blinding run. 1485 */ 1486 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1487 insn[1].code == 0) 1488 memcpy(aux, insn, sizeof(aux)); 1489 1490 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1491 clone->aux->verifier_zext); 1492 if (!rewritten) 1493 continue; 1494 1495 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1496 if (IS_ERR(tmp)) { 1497 /* Patching may have repointed aux->prog during 1498 * realloc from the original one, so we need to 1499 * fix it up here on error. 1500 */ 1501 bpf_jit_prog_release_other(prog, clone); 1502 return tmp; 1503 } 1504 1505 clone = tmp; 1506 insn_delta = rewritten - 1; 1507 1508 /* Walk new program and skip insns we just inserted. */ 1509 insn = clone->insnsi + i + insn_delta; 1510 insn_cnt += insn_delta; 1511 i += insn_delta; 1512 } 1513 1514 clone->blinded = 1; 1515 return clone; 1516 } 1517 #endif /* CONFIG_BPF_JIT */ 1518 1519 /* Base function for offset calculation. Needs to go into .text section, 1520 * therefore keeping it non-static as well; will also be used by JITs 1521 * anyway later on, so do not let the compiler omit it. This also needs 1522 * to go into kallsyms for correlation from e.g. bpftool, so naming 1523 * must not change. 1524 */ 1525 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1526 { 1527 return 0; 1528 } 1529 EXPORT_SYMBOL_GPL(__bpf_call_base); 1530 1531 /* All UAPI available opcodes. */ 1532 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1533 /* 32 bit ALU operations. */ \ 1534 /* Register based. */ \ 1535 INSN_3(ALU, ADD, X), \ 1536 INSN_3(ALU, SUB, X), \ 1537 INSN_3(ALU, AND, X), \ 1538 INSN_3(ALU, OR, X), \ 1539 INSN_3(ALU, LSH, X), \ 1540 INSN_3(ALU, RSH, X), \ 1541 INSN_3(ALU, XOR, X), \ 1542 INSN_3(ALU, MUL, X), \ 1543 INSN_3(ALU, MOV, X), \ 1544 INSN_3(ALU, ARSH, X), \ 1545 INSN_3(ALU, DIV, X), \ 1546 INSN_3(ALU, MOD, X), \ 1547 INSN_2(ALU, NEG), \ 1548 INSN_3(ALU, END, TO_BE), \ 1549 INSN_3(ALU, END, TO_LE), \ 1550 /* Immediate based. */ \ 1551 INSN_3(ALU, ADD, K), \ 1552 INSN_3(ALU, SUB, K), \ 1553 INSN_3(ALU, AND, K), \ 1554 INSN_3(ALU, OR, K), \ 1555 INSN_3(ALU, LSH, K), \ 1556 INSN_3(ALU, RSH, K), \ 1557 INSN_3(ALU, XOR, K), \ 1558 INSN_3(ALU, MUL, K), \ 1559 INSN_3(ALU, MOV, K), \ 1560 INSN_3(ALU, ARSH, K), \ 1561 INSN_3(ALU, DIV, K), \ 1562 INSN_3(ALU, MOD, K), \ 1563 /* 64 bit ALU operations. */ \ 1564 /* Register based. */ \ 1565 INSN_3(ALU64, ADD, X), \ 1566 INSN_3(ALU64, SUB, X), \ 1567 INSN_3(ALU64, AND, X), \ 1568 INSN_3(ALU64, OR, X), \ 1569 INSN_3(ALU64, LSH, X), \ 1570 INSN_3(ALU64, RSH, X), \ 1571 INSN_3(ALU64, XOR, X), \ 1572 INSN_3(ALU64, MUL, X), \ 1573 INSN_3(ALU64, MOV, X), \ 1574 INSN_3(ALU64, ARSH, X), \ 1575 INSN_3(ALU64, DIV, X), \ 1576 INSN_3(ALU64, MOD, X), \ 1577 INSN_2(ALU64, NEG), \ 1578 INSN_3(ALU64, END, TO_LE), \ 1579 /* Immediate based. */ \ 1580 INSN_3(ALU64, ADD, K), \ 1581 INSN_3(ALU64, SUB, K), \ 1582 INSN_3(ALU64, AND, K), \ 1583 INSN_3(ALU64, OR, K), \ 1584 INSN_3(ALU64, LSH, K), \ 1585 INSN_3(ALU64, RSH, K), \ 1586 INSN_3(ALU64, XOR, K), \ 1587 INSN_3(ALU64, MUL, K), \ 1588 INSN_3(ALU64, MOV, K), \ 1589 INSN_3(ALU64, ARSH, K), \ 1590 INSN_3(ALU64, DIV, K), \ 1591 INSN_3(ALU64, MOD, K), \ 1592 /* Call instruction. */ \ 1593 INSN_2(JMP, CALL), \ 1594 /* Exit instruction. */ \ 1595 INSN_2(JMP, EXIT), \ 1596 /* 32-bit Jump instructions. */ \ 1597 /* Register based. */ \ 1598 INSN_3(JMP32, JEQ, X), \ 1599 INSN_3(JMP32, JNE, X), \ 1600 INSN_3(JMP32, JGT, X), \ 1601 INSN_3(JMP32, JLT, X), \ 1602 INSN_3(JMP32, JGE, X), \ 1603 INSN_3(JMP32, JLE, X), \ 1604 INSN_3(JMP32, JSGT, X), \ 1605 INSN_3(JMP32, JSLT, X), \ 1606 INSN_3(JMP32, JSGE, X), \ 1607 INSN_3(JMP32, JSLE, X), \ 1608 INSN_3(JMP32, JSET, X), \ 1609 /* Immediate based. */ \ 1610 INSN_3(JMP32, JEQ, K), \ 1611 INSN_3(JMP32, JNE, K), \ 1612 INSN_3(JMP32, JGT, K), \ 1613 INSN_3(JMP32, JLT, K), \ 1614 INSN_3(JMP32, JGE, K), \ 1615 INSN_3(JMP32, JLE, K), \ 1616 INSN_3(JMP32, JSGT, K), \ 1617 INSN_3(JMP32, JSLT, K), \ 1618 INSN_3(JMP32, JSGE, K), \ 1619 INSN_3(JMP32, JSLE, K), \ 1620 INSN_3(JMP32, JSET, K), \ 1621 /* Jump instructions. */ \ 1622 /* Register based. */ \ 1623 INSN_3(JMP, JEQ, X), \ 1624 INSN_3(JMP, JNE, X), \ 1625 INSN_3(JMP, JGT, X), \ 1626 INSN_3(JMP, JLT, X), \ 1627 INSN_3(JMP, JGE, X), \ 1628 INSN_3(JMP, JLE, X), \ 1629 INSN_3(JMP, JSGT, X), \ 1630 INSN_3(JMP, JSLT, X), \ 1631 INSN_3(JMP, JSGE, X), \ 1632 INSN_3(JMP, JSLE, X), \ 1633 INSN_3(JMP, JSET, X), \ 1634 /* Immediate based. */ \ 1635 INSN_3(JMP, JEQ, K), \ 1636 INSN_3(JMP, JNE, K), \ 1637 INSN_3(JMP, JGT, K), \ 1638 INSN_3(JMP, JLT, K), \ 1639 INSN_3(JMP, JGE, K), \ 1640 INSN_3(JMP, JLE, K), \ 1641 INSN_3(JMP, JSGT, K), \ 1642 INSN_3(JMP, JSLT, K), \ 1643 INSN_3(JMP, JSGE, K), \ 1644 INSN_3(JMP, JSLE, K), \ 1645 INSN_3(JMP, JSET, K), \ 1646 INSN_2(JMP, JA), \ 1647 INSN_2(JMP32, JA), \ 1648 /* Store instructions. */ \ 1649 /* Register based. */ \ 1650 INSN_3(STX, MEM, B), \ 1651 INSN_3(STX, MEM, H), \ 1652 INSN_3(STX, MEM, W), \ 1653 INSN_3(STX, MEM, DW), \ 1654 INSN_3(STX, ATOMIC, W), \ 1655 INSN_3(STX, ATOMIC, DW), \ 1656 /* Immediate based. */ \ 1657 INSN_3(ST, MEM, B), \ 1658 INSN_3(ST, MEM, H), \ 1659 INSN_3(ST, MEM, W), \ 1660 INSN_3(ST, MEM, DW), \ 1661 /* Load instructions. */ \ 1662 /* Register based. */ \ 1663 INSN_3(LDX, MEM, B), \ 1664 INSN_3(LDX, MEM, H), \ 1665 INSN_3(LDX, MEM, W), \ 1666 INSN_3(LDX, MEM, DW), \ 1667 INSN_3(LDX, MEMSX, B), \ 1668 INSN_3(LDX, MEMSX, H), \ 1669 INSN_3(LDX, MEMSX, W), \ 1670 /* Immediate based. */ \ 1671 INSN_3(LD, IMM, DW) 1672 1673 bool bpf_opcode_in_insntable(u8 code) 1674 { 1675 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1676 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1677 static const bool public_insntable[256] = { 1678 [0 ... 255] = false, 1679 /* Now overwrite non-defaults ... */ 1680 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1681 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1682 [BPF_LD | BPF_ABS | BPF_B] = true, 1683 [BPF_LD | BPF_ABS | BPF_H] = true, 1684 [BPF_LD | BPF_ABS | BPF_W] = true, 1685 [BPF_LD | BPF_IND | BPF_B] = true, 1686 [BPF_LD | BPF_IND | BPF_H] = true, 1687 [BPF_LD | BPF_IND | BPF_W] = true, 1688 [BPF_JMP | BPF_JCOND] = true, 1689 }; 1690 #undef BPF_INSN_3_TBL 1691 #undef BPF_INSN_2_TBL 1692 return public_insntable[code]; 1693 } 1694 1695 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1696 /** 1697 * ___bpf_prog_run - run eBPF program on a given context 1698 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1699 * @insn: is the array of eBPF instructions 1700 * 1701 * Decode and execute eBPF instructions. 1702 * 1703 * Return: whatever value is in %BPF_R0 at program exit 1704 */ 1705 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1706 { 1707 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1708 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1709 static const void * const jumptable[256] __annotate_jump_table = { 1710 [0 ... 255] = &&default_label, 1711 /* Now overwrite non-defaults ... */ 1712 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1713 /* Non-UAPI available opcodes. */ 1714 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1715 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1716 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1717 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1718 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1719 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1720 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1721 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1722 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1723 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1724 }; 1725 #undef BPF_INSN_3_LBL 1726 #undef BPF_INSN_2_LBL 1727 u32 tail_call_cnt = 0; 1728 1729 #define CONT ({ insn++; goto select_insn; }) 1730 #define CONT_JMP ({ insn++; goto select_insn; }) 1731 1732 select_insn: 1733 goto *jumptable[insn->code]; 1734 1735 /* Explicitly mask the register-based shift amounts with 63 or 31 1736 * to avoid undefined behavior. Normally this won't affect the 1737 * generated code, for example, in case of native 64 bit archs such 1738 * as x86-64 or arm64, the compiler is optimizing the AND away for 1739 * the interpreter. In case of JITs, each of the JIT backends compiles 1740 * the BPF shift operations to machine instructions which produce 1741 * implementation-defined results in such a case; the resulting 1742 * contents of the register may be arbitrary, but program behaviour 1743 * as a whole remains defined. In other words, in case of JIT backends, 1744 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1745 */ 1746 /* ALU (shifts) */ 1747 #define SHT(OPCODE, OP) \ 1748 ALU64_##OPCODE##_X: \ 1749 DST = DST OP (SRC & 63); \ 1750 CONT; \ 1751 ALU_##OPCODE##_X: \ 1752 DST = (u32) DST OP ((u32) SRC & 31); \ 1753 CONT; \ 1754 ALU64_##OPCODE##_K: \ 1755 DST = DST OP IMM; \ 1756 CONT; \ 1757 ALU_##OPCODE##_K: \ 1758 DST = (u32) DST OP (u32) IMM; \ 1759 CONT; 1760 /* ALU (rest) */ 1761 #define ALU(OPCODE, OP) \ 1762 ALU64_##OPCODE##_X: \ 1763 DST = DST OP SRC; \ 1764 CONT; \ 1765 ALU_##OPCODE##_X: \ 1766 DST = (u32) DST OP (u32) SRC; \ 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(ADD, +) 1775 ALU(SUB, -) 1776 ALU(AND, &) 1777 ALU(OR, |) 1778 ALU(XOR, ^) 1779 ALU(MUL, *) 1780 SHT(LSH, <<) 1781 SHT(RSH, >>) 1782 #undef SHT 1783 #undef ALU 1784 ALU_NEG: 1785 DST = (u32) -DST; 1786 CONT; 1787 ALU64_NEG: 1788 DST = -DST; 1789 CONT; 1790 ALU_MOV_X: 1791 switch (OFF) { 1792 case 0: 1793 DST = (u32) SRC; 1794 break; 1795 case 8: 1796 DST = (u32)(s8) SRC; 1797 break; 1798 case 16: 1799 DST = (u32)(s16) SRC; 1800 break; 1801 } 1802 CONT; 1803 ALU_MOV_K: 1804 DST = (u32) IMM; 1805 CONT; 1806 ALU64_MOV_X: 1807 switch (OFF) { 1808 case 0: 1809 DST = SRC; 1810 break; 1811 case 8: 1812 DST = (s8) SRC; 1813 break; 1814 case 16: 1815 DST = (s16) SRC; 1816 break; 1817 case 32: 1818 DST = (s32) SRC; 1819 break; 1820 } 1821 CONT; 1822 ALU64_MOV_K: 1823 DST = IMM; 1824 CONT; 1825 LD_IMM_DW: 1826 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1827 insn++; 1828 CONT; 1829 ALU_ARSH_X: 1830 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1831 CONT; 1832 ALU_ARSH_K: 1833 DST = (u64) (u32) (((s32) DST) >> IMM); 1834 CONT; 1835 ALU64_ARSH_X: 1836 (*(s64 *) &DST) >>= (SRC & 63); 1837 CONT; 1838 ALU64_ARSH_K: 1839 (*(s64 *) &DST) >>= IMM; 1840 CONT; 1841 ALU64_MOD_X: 1842 switch (OFF) { 1843 case 0: 1844 div64_u64_rem(DST, SRC, &AX); 1845 DST = AX; 1846 break; 1847 case 1: 1848 AX = div64_s64(DST, SRC); 1849 DST = DST - AX * SRC; 1850 break; 1851 } 1852 CONT; 1853 ALU_MOD_X: 1854 switch (OFF) { 1855 case 0: 1856 AX = (u32) DST; 1857 DST = do_div(AX, (u32) SRC); 1858 break; 1859 case 1: 1860 AX = abs((s32)DST); 1861 AX = do_div(AX, abs((s32)SRC)); 1862 if ((s32)DST < 0) 1863 DST = (u32)-AX; 1864 else 1865 DST = (u32)AX; 1866 break; 1867 } 1868 CONT; 1869 ALU64_MOD_K: 1870 switch (OFF) { 1871 case 0: 1872 div64_u64_rem(DST, IMM, &AX); 1873 DST = AX; 1874 break; 1875 case 1: 1876 AX = div64_s64(DST, IMM); 1877 DST = DST - AX * IMM; 1878 break; 1879 } 1880 CONT; 1881 ALU_MOD_K: 1882 switch (OFF) { 1883 case 0: 1884 AX = (u32) DST; 1885 DST = do_div(AX, (u32) IMM); 1886 break; 1887 case 1: 1888 AX = abs((s32)DST); 1889 AX = do_div(AX, abs((s32)IMM)); 1890 if ((s32)DST < 0) 1891 DST = (u32)-AX; 1892 else 1893 DST = (u32)AX; 1894 break; 1895 } 1896 CONT; 1897 ALU64_DIV_X: 1898 switch (OFF) { 1899 case 0: 1900 DST = div64_u64(DST, SRC); 1901 break; 1902 case 1: 1903 DST = div64_s64(DST, SRC); 1904 break; 1905 } 1906 CONT; 1907 ALU_DIV_X: 1908 switch (OFF) { 1909 case 0: 1910 AX = (u32) DST; 1911 do_div(AX, (u32) SRC); 1912 DST = (u32) AX; 1913 break; 1914 case 1: 1915 AX = abs((s32)DST); 1916 do_div(AX, abs((s32)SRC)); 1917 if (((s32)DST < 0) == ((s32)SRC < 0)) 1918 DST = (u32)AX; 1919 else 1920 DST = (u32)-AX; 1921 break; 1922 } 1923 CONT; 1924 ALU64_DIV_K: 1925 switch (OFF) { 1926 case 0: 1927 DST = div64_u64(DST, IMM); 1928 break; 1929 case 1: 1930 DST = div64_s64(DST, IMM); 1931 break; 1932 } 1933 CONT; 1934 ALU_DIV_K: 1935 switch (OFF) { 1936 case 0: 1937 AX = (u32) DST; 1938 do_div(AX, (u32) IMM); 1939 DST = (u32) AX; 1940 break; 1941 case 1: 1942 AX = abs((s32)DST); 1943 do_div(AX, abs((s32)IMM)); 1944 if (((s32)DST < 0) == ((s32)IMM < 0)) 1945 DST = (u32)AX; 1946 else 1947 DST = (u32)-AX; 1948 break; 1949 } 1950 CONT; 1951 ALU_END_TO_BE: 1952 switch (IMM) { 1953 case 16: 1954 DST = (__force u16) cpu_to_be16(DST); 1955 break; 1956 case 32: 1957 DST = (__force u32) cpu_to_be32(DST); 1958 break; 1959 case 64: 1960 DST = (__force u64) cpu_to_be64(DST); 1961 break; 1962 } 1963 CONT; 1964 ALU_END_TO_LE: 1965 switch (IMM) { 1966 case 16: 1967 DST = (__force u16) cpu_to_le16(DST); 1968 break; 1969 case 32: 1970 DST = (__force u32) cpu_to_le32(DST); 1971 break; 1972 case 64: 1973 DST = (__force u64) cpu_to_le64(DST); 1974 break; 1975 } 1976 CONT; 1977 ALU64_END_TO_LE: 1978 switch (IMM) { 1979 case 16: 1980 DST = (__force u16) __swab16(DST); 1981 break; 1982 case 32: 1983 DST = (__force u32) __swab32(DST); 1984 break; 1985 case 64: 1986 DST = (__force u64) __swab64(DST); 1987 break; 1988 } 1989 CONT; 1990 1991 /* CALL */ 1992 JMP_CALL: 1993 /* Function call scratches BPF_R1-BPF_R5 registers, 1994 * preserves BPF_R6-BPF_R9, and stores return value 1995 * into BPF_R0. 1996 */ 1997 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1998 BPF_R4, BPF_R5); 1999 CONT; 2000 2001 JMP_CALL_ARGS: 2002 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2003 BPF_R3, BPF_R4, 2004 BPF_R5, 2005 insn + insn->off + 1); 2006 CONT; 2007 2008 JMP_TAIL_CALL: { 2009 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2010 struct bpf_array *array = container_of(map, struct bpf_array, map); 2011 struct bpf_prog *prog; 2012 u32 index = BPF_R3; 2013 2014 if (unlikely(index >= array->map.max_entries)) 2015 goto out; 2016 2017 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2018 goto out; 2019 2020 tail_call_cnt++; 2021 2022 prog = READ_ONCE(array->ptrs[index]); 2023 if (!prog) 2024 goto out; 2025 2026 /* ARG1 at this point is guaranteed to point to CTX from 2027 * the verifier side due to the fact that the tail call is 2028 * handled like a helper, that is, bpf_tail_call_proto, 2029 * where arg1_type is ARG_PTR_TO_CTX. 2030 */ 2031 insn = prog->insnsi; 2032 goto select_insn; 2033 out: 2034 CONT; 2035 } 2036 JMP_JA: 2037 insn += insn->off; 2038 CONT; 2039 JMP32_JA: 2040 insn += insn->imm; 2041 CONT; 2042 JMP_EXIT: 2043 return BPF_R0; 2044 /* JMP */ 2045 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2046 JMP_##OPCODE##_X: \ 2047 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2048 insn += insn->off; \ 2049 CONT_JMP; \ 2050 } \ 2051 CONT; \ 2052 JMP32_##OPCODE##_X: \ 2053 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2054 insn += insn->off; \ 2055 CONT_JMP; \ 2056 } \ 2057 CONT; \ 2058 JMP_##OPCODE##_K: \ 2059 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2060 insn += insn->off; \ 2061 CONT_JMP; \ 2062 } \ 2063 CONT; \ 2064 JMP32_##OPCODE##_K: \ 2065 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2066 insn += insn->off; \ 2067 CONT_JMP; \ 2068 } \ 2069 CONT; 2070 COND_JMP(u, JEQ, ==) 2071 COND_JMP(u, JNE, !=) 2072 COND_JMP(u, JGT, >) 2073 COND_JMP(u, JLT, <) 2074 COND_JMP(u, JGE, >=) 2075 COND_JMP(u, JLE, <=) 2076 COND_JMP(u, JSET, &) 2077 COND_JMP(s, JSGT, >) 2078 COND_JMP(s, JSLT, <) 2079 COND_JMP(s, JSGE, >=) 2080 COND_JMP(s, JSLE, <=) 2081 #undef COND_JMP 2082 /* ST, STX and LDX*/ 2083 ST_NOSPEC: 2084 /* Speculation barrier for mitigating Speculative Store Bypass. 2085 * In case of arm64, we rely on the firmware mitigation as 2086 * controlled via the ssbd kernel parameter. Whenever the 2087 * mitigation is enabled, it works for all of the kernel code 2088 * with no need to provide any additional instructions here. 2089 * In case of x86, we use 'lfence' insn for mitigation. We 2090 * reuse preexisting logic from Spectre v1 mitigation that 2091 * happens to produce the required code on x86 for v4 as well. 2092 */ 2093 barrier_nospec(); 2094 CONT; 2095 #define LDST(SIZEOP, SIZE) \ 2096 STX_MEM_##SIZEOP: \ 2097 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2098 CONT; \ 2099 ST_MEM_##SIZEOP: \ 2100 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2101 CONT; \ 2102 LDX_MEM_##SIZEOP: \ 2103 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2104 CONT; \ 2105 LDX_PROBE_MEM_##SIZEOP: \ 2106 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2107 (const void *)(long) (SRC + insn->off)); \ 2108 DST = *((SIZE *)&DST); \ 2109 CONT; 2110 2111 LDST(B, u8) 2112 LDST(H, u16) 2113 LDST(W, u32) 2114 LDST(DW, u64) 2115 #undef LDST 2116 2117 #define LDSX(SIZEOP, SIZE) \ 2118 LDX_MEMSX_##SIZEOP: \ 2119 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2120 CONT; \ 2121 LDX_PROBE_MEMSX_##SIZEOP: \ 2122 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2123 (const void *)(long) (SRC + insn->off)); \ 2124 DST = *((SIZE *)&DST); \ 2125 CONT; 2126 2127 LDSX(B, s8) 2128 LDSX(H, s16) 2129 LDSX(W, s32) 2130 #undef LDSX 2131 2132 #define ATOMIC_ALU_OP(BOP, KOP) \ 2133 case BOP: \ 2134 if (BPF_SIZE(insn->code) == BPF_W) \ 2135 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2136 (DST + insn->off)); \ 2137 else \ 2138 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2139 (DST + insn->off)); \ 2140 break; \ 2141 case BOP | BPF_FETCH: \ 2142 if (BPF_SIZE(insn->code) == BPF_W) \ 2143 SRC = (u32) atomic_fetch_##KOP( \ 2144 (u32) SRC, \ 2145 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2146 else \ 2147 SRC = (u64) atomic64_fetch_##KOP( \ 2148 (u64) SRC, \ 2149 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2150 break; 2151 2152 STX_ATOMIC_DW: 2153 STX_ATOMIC_W: 2154 switch (IMM) { 2155 ATOMIC_ALU_OP(BPF_ADD, add) 2156 ATOMIC_ALU_OP(BPF_AND, and) 2157 ATOMIC_ALU_OP(BPF_OR, or) 2158 ATOMIC_ALU_OP(BPF_XOR, xor) 2159 #undef ATOMIC_ALU_OP 2160 2161 case BPF_XCHG: 2162 if (BPF_SIZE(insn->code) == BPF_W) 2163 SRC = (u32) atomic_xchg( 2164 (atomic_t *)(unsigned long) (DST + insn->off), 2165 (u32) SRC); 2166 else 2167 SRC = (u64) atomic64_xchg( 2168 (atomic64_t *)(unsigned long) (DST + insn->off), 2169 (u64) SRC); 2170 break; 2171 case BPF_CMPXCHG: 2172 if (BPF_SIZE(insn->code) == BPF_W) 2173 BPF_R0 = (u32) atomic_cmpxchg( 2174 (atomic_t *)(unsigned long) (DST + insn->off), 2175 (u32) BPF_R0, (u32) SRC); 2176 else 2177 BPF_R0 = (u64) atomic64_cmpxchg( 2178 (atomic64_t *)(unsigned long) (DST + insn->off), 2179 (u64) BPF_R0, (u64) SRC); 2180 break; 2181 2182 default: 2183 goto default_label; 2184 } 2185 CONT; 2186 2187 default_label: 2188 /* If we ever reach this, we have a bug somewhere. Die hard here 2189 * instead of just returning 0; we could be somewhere in a subprog, 2190 * so execution could continue otherwise which we do /not/ want. 2191 * 2192 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2193 */ 2194 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2195 insn->code, insn->imm); 2196 BUG_ON(1); 2197 return 0; 2198 } 2199 2200 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2201 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2202 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2203 { \ 2204 u64 stack[stack_size / sizeof(u64)]; \ 2205 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2206 \ 2207 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2208 ARG1 = (u64) (unsigned long) ctx; \ 2209 return ___bpf_prog_run(regs, insn); \ 2210 } 2211 2212 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2213 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2214 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2215 const struct bpf_insn *insn) \ 2216 { \ 2217 u64 stack[stack_size / sizeof(u64)]; \ 2218 u64 regs[MAX_BPF_EXT_REG]; \ 2219 \ 2220 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2221 BPF_R1 = r1; \ 2222 BPF_R2 = r2; \ 2223 BPF_R3 = r3; \ 2224 BPF_R4 = r4; \ 2225 BPF_R5 = r5; \ 2226 return ___bpf_prog_run(regs, insn); \ 2227 } 2228 2229 #define EVAL1(FN, X) FN(X) 2230 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2231 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2232 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2233 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2234 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2235 2236 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2237 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2238 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2239 2240 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2241 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2242 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2243 2244 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2245 2246 static unsigned int (*interpreters[])(const void *ctx, 2247 const struct bpf_insn *insn) = { 2248 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2249 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2250 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2251 }; 2252 #undef PROG_NAME_LIST 2253 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2254 static __maybe_unused 2255 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2256 const struct bpf_insn *insn) = { 2257 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2258 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2259 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2260 }; 2261 #undef PROG_NAME_LIST 2262 2263 #ifdef CONFIG_BPF_SYSCALL 2264 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2265 { 2266 stack_depth = max_t(u32, stack_depth, 1); 2267 insn->off = (s16) insn->imm; 2268 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2269 __bpf_call_base_args; 2270 insn->code = BPF_JMP | BPF_CALL_ARGS; 2271 } 2272 #endif 2273 #else 2274 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2275 const struct bpf_insn *insn) 2276 { 2277 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2278 * is not working properly, so warn about it! 2279 */ 2280 WARN_ON_ONCE(1); 2281 return 0; 2282 } 2283 #endif 2284 2285 bool bpf_prog_map_compatible(struct bpf_map *map, 2286 const struct bpf_prog *fp) 2287 { 2288 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2289 bool ret; 2290 2291 if (fp->kprobe_override) 2292 return false; 2293 2294 /* XDP programs inserted into maps are not guaranteed to run on 2295 * a particular netdev (and can run outside driver context entirely 2296 * in the case of devmap and cpumap). Until device checks 2297 * are implemented, prohibit adding dev-bound programs to program maps. 2298 */ 2299 if (bpf_prog_is_dev_bound(fp->aux)) 2300 return false; 2301 2302 spin_lock(&map->owner.lock); 2303 if (!map->owner.type) { 2304 /* There's no owner yet where we could check for 2305 * compatibility. 2306 */ 2307 map->owner.type = prog_type; 2308 map->owner.jited = fp->jited; 2309 map->owner.xdp_has_frags = fp->aux->xdp_has_frags; 2310 ret = true; 2311 } else { 2312 ret = map->owner.type == prog_type && 2313 map->owner.jited == fp->jited && 2314 map->owner.xdp_has_frags == fp->aux->xdp_has_frags; 2315 } 2316 spin_unlock(&map->owner.lock); 2317 2318 return ret; 2319 } 2320 2321 static int bpf_check_tail_call(const struct bpf_prog *fp) 2322 { 2323 struct bpf_prog_aux *aux = fp->aux; 2324 int i, ret = 0; 2325 2326 mutex_lock(&aux->used_maps_mutex); 2327 for (i = 0; i < aux->used_map_cnt; i++) { 2328 struct bpf_map *map = aux->used_maps[i]; 2329 2330 if (!map_type_contains_progs(map)) 2331 continue; 2332 2333 if (!bpf_prog_map_compatible(map, fp)) { 2334 ret = -EINVAL; 2335 goto out; 2336 } 2337 } 2338 2339 out: 2340 mutex_unlock(&aux->used_maps_mutex); 2341 return ret; 2342 } 2343 2344 static void bpf_prog_select_func(struct bpf_prog *fp) 2345 { 2346 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2347 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2348 2349 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2350 #else 2351 fp->bpf_func = __bpf_prog_ret0_warn; 2352 #endif 2353 } 2354 2355 /** 2356 * bpf_prog_select_runtime - select exec runtime for BPF program 2357 * @fp: bpf_prog populated with BPF program 2358 * @err: pointer to error variable 2359 * 2360 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2361 * The BPF program will be executed via bpf_prog_run() function. 2362 * 2363 * Return: the &fp argument along with &err set to 0 for success or 2364 * a negative errno code on failure 2365 */ 2366 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2367 { 2368 /* In case of BPF to BPF calls, verifier did all the prep 2369 * work with regards to JITing, etc. 2370 */ 2371 bool jit_needed = false; 2372 2373 if (fp->bpf_func) 2374 goto finalize; 2375 2376 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2377 bpf_prog_has_kfunc_call(fp)) 2378 jit_needed = true; 2379 2380 bpf_prog_select_func(fp); 2381 2382 /* eBPF JITs can rewrite the program in case constant 2383 * blinding is active. However, in case of error during 2384 * blinding, bpf_int_jit_compile() must always return a 2385 * valid program, which in this case would simply not 2386 * be JITed, but falls back to the interpreter. 2387 */ 2388 if (!bpf_prog_is_offloaded(fp->aux)) { 2389 *err = bpf_prog_alloc_jited_linfo(fp); 2390 if (*err) 2391 return fp; 2392 2393 fp = bpf_int_jit_compile(fp); 2394 bpf_prog_jit_attempt_done(fp); 2395 if (!fp->jited && jit_needed) { 2396 *err = -ENOTSUPP; 2397 return fp; 2398 } 2399 } else { 2400 *err = bpf_prog_offload_compile(fp); 2401 if (*err) 2402 return fp; 2403 } 2404 2405 finalize: 2406 bpf_prog_lock_ro(fp); 2407 2408 /* The tail call compatibility check can only be done at 2409 * this late stage as we need to determine, if we deal 2410 * with JITed or non JITed program concatenations and not 2411 * all eBPF JITs might immediately support all features. 2412 */ 2413 *err = bpf_check_tail_call(fp); 2414 2415 return fp; 2416 } 2417 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2418 2419 static unsigned int __bpf_prog_ret1(const void *ctx, 2420 const struct bpf_insn *insn) 2421 { 2422 return 1; 2423 } 2424 2425 static struct bpf_prog_dummy { 2426 struct bpf_prog prog; 2427 } dummy_bpf_prog = { 2428 .prog = { 2429 .bpf_func = __bpf_prog_ret1, 2430 }, 2431 }; 2432 2433 struct bpf_empty_prog_array bpf_empty_prog_array = { 2434 .null_prog = NULL, 2435 }; 2436 EXPORT_SYMBOL(bpf_empty_prog_array); 2437 2438 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2439 { 2440 if (prog_cnt) 2441 return kzalloc(sizeof(struct bpf_prog_array) + 2442 sizeof(struct bpf_prog_array_item) * 2443 (prog_cnt + 1), 2444 flags); 2445 2446 return &bpf_empty_prog_array.hdr; 2447 } 2448 2449 void bpf_prog_array_free(struct bpf_prog_array *progs) 2450 { 2451 if (!progs || progs == &bpf_empty_prog_array.hdr) 2452 return; 2453 kfree_rcu(progs, rcu); 2454 } 2455 2456 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2457 { 2458 struct bpf_prog_array *progs; 2459 2460 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2461 * no need to call kfree_rcu(), just call kfree() directly. 2462 */ 2463 progs = container_of(rcu, struct bpf_prog_array, rcu); 2464 if (rcu_trace_implies_rcu_gp()) 2465 kfree(progs); 2466 else 2467 kfree_rcu(progs, rcu); 2468 } 2469 2470 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2471 { 2472 if (!progs || progs == &bpf_empty_prog_array.hdr) 2473 return; 2474 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2475 } 2476 2477 int bpf_prog_array_length(struct bpf_prog_array *array) 2478 { 2479 struct bpf_prog_array_item *item; 2480 u32 cnt = 0; 2481 2482 for (item = array->items; item->prog; item++) 2483 if (item->prog != &dummy_bpf_prog.prog) 2484 cnt++; 2485 return cnt; 2486 } 2487 2488 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2489 { 2490 struct bpf_prog_array_item *item; 2491 2492 for (item = array->items; item->prog; item++) 2493 if (item->prog != &dummy_bpf_prog.prog) 2494 return false; 2495 return true; 2496 } 2497 2498 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2499 u32 *prog_ids, 2500 u32 request_cnt) 2501 { 2502 struct bpf_prog_array_item *item; 2503 int i = 0; 2504 2505 for (item = array->items; item->prog; item++) { 2506 if (item->prog == &dummy_bpf_prog.prog) 2507 continue; 2508 prog_ids[i] = item->prog->aux->id; 2509 if (++i == request_cnt) { 2510 item++; 2511 break; 2512 } 2513 } 2514 2515 return !!(item->prog); 2516 } 2517 2518 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2519 __u32 __user *prog_ids, u32 cnt) 2520 { 2521 unsigned long err = 0; 2522 bool nospc; 2523 u32 *ids; 2524 2525 /* users of this function are doing: 2526 * cnt = bpf_prog_array_length(); 2527 * if (cnt > 0) 2528 * bpf_prog_array_copy_to_user(..., cnt); 2529 * so below kcalloc doesn't need extra cnt > 0 check. 2530 */ 2531 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2532 if (!ids) 2533 return -ENOMEM; 2534 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2535 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2536 kfree(ids); 2537 if (err) 2538 return -EFAULT; 2539 if (nospc) 2540 return -ENOSPC; 2541 return 0; 2542 } 2543 2544 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2545 struct bpf_prog *old_prog) 2546 { 2547 struct bpf_prog_array_item *item; 2548 2549 for (item = array->items; item->prog; item++) 2550 if (item->prog == old_prog) { 2551 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2552 break; 2553 } 2554 } 2555 2556 /** 2557 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2558 * index into the program array with 2559 * a dummy no-op program. 2560 * @array: a bpf_prog_array 2561 * @index: the index of the program to replace 2562 * 2563 * Skips over dummy programs, by not counting them, when calculating 2564 * the position of the program to replace. 2565 * 2566 * Return: 2567 * * 0 - Success 2568 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2569 * * -ENOENT - Index out of range 2570 */ 2571 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2572 { 2573 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2574 } 2575 2576 /** 2577 * bpf_prog_array_update_at() - Updates the program at the given index 2578 * into the program array. 2579 * @array: a bpf_prog_array 2580 * @index: the index of the program to update 2581 * @prog: the program to insert into the array 2582 * 2583 * Skips over dummy programs, by not counting them, when calculating 2584 * the position of the program to update. 2585 * 2586 * Return: 2587 * * 0 - Success 2588 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2589 * * -ENOENT - Index out of range 2590 */ 2591 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2592 struct bpf_prog *prog) 2593 { 2594 struct bpf_prog_array_item *item; 2595 2596 if (unlikely(index < 0)) 2597 return -EINVAL; 2598 2599 for (item = array->items; item->prog; item++) { 2600 if (item->prog == &dummy_bpf_prog.prog) 2601 continue; 2602 if (!index) { 2603 WRITE_ONCE(item->prog, prog); 2604 return 0; 2605 } 2606 index--; 2607 } 2608 return -ENOENT; 2609 } 2610 2611 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2612 struct bpf_prog *exclude_prog, 2613 struct bpf_prog *include_prog, 2614 u64 bpf_cookie, 2615 struct bpf_prog_array **new_array) 2616 { 2617 int new_prog_cnt, carry_prog_cnt = 0; 2618 struct bpf_prog_array_item *existing, *new; 2619 struct bpf_prog_array *array; 2620 bool found_exclude = false; 2621 2622 /* Figure out how many existing progs we need to carry over to 2623 * the new array. 2624 */ 2625 if (old_array) { 2626 existing = old_array->items; 2627 for (; existing->prog; existing++) { 2628 if (existing->prog == exclude_prog) { 2629 found_exclude = true; 2630 continue; 2631 } 2632 if (existing->prog != &dummy_bpf_prog.prog) 2633 carry_prog_cnt++; 2634 if (existing->prog == include_prog) 2635 return -EEXIST; 2636 } 2637 } 2638 2639 if (exclude_prog && !found_exclude) 2640 return -ENOENT; 2641 2642 /* How many progs (not NULL) will be in the new array? */ 2643 new_prog_cnt = carry_prog_cnt; 2644 if (include_prog) 2645 new_prog_cnt += 1; 2646 2647 /* Do we have any prog (not NULL) in the new array? */ 2648 if (!new_prog_cnt) { 2649 *new_array = NULL; 2650 return 0; 2651 } 2652 2653 /* +1 as the end of prog_array is marked with NULL */ 2654 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2655 if (!array) 2656 return -ENOMEM; 2657 new = array->items; 2658 2659 /* Fill in the new prog array */ 2660 if (carry_prog_cnt) { 2661 existing = old_array->items; 2662 for (; existing->prog; existing++) { 2663 if (existing->prog == exclude_prog || 2664 existing->prog == &dummy_bpf_prog.prog) 2665 continue; 2666 2667 new->prog = existing->prog; 2668 new->bpf_cookie = existing->bpf_cookie; 2669 new++; 2670 } 2671 } 2672 if (include_prog) { 2673 new->prog = include_prog; 2674 new->bpf_cookie = bpf_cookie; 2675 new++; 2676 } 2677 new->prog = NULL; 2678 *new_array = array; 2679 return 0; 2680 } 2681 2682 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2683 u32 *prog_ids, u32 request_cnt, 2684 u32 *prog_cnt) 2685 { 2686 u32 cnt = 0; 2687 2688 if (array) 2689 cnt = bpf_prog_array_length(array); 2690 2691 *prog_cnt = cnt; 2692 2693 /* return early if user requested only program count or nothing to copy */ 2694 if (!request_cnt || !cnt) 2695 return 0; 2696 2697 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2698 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2699 : 0; 2700 } 2701 2702 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2703 struct bpf_map **used_maps, u32 len) 2704 { 2705 struct bpf_map *map; 2706 bool sleepable; 2707 u32 i; 2708 2709 sleepable = aux->prog->sleepable; 2710 for (i = 0; i < len; i++) { 2711 map = used_maps[i]; 2712 if (map->ops->map_poke_untrack) 2713 map->ops->map_poke_untrack(map, aux); 2714 if (sleepable) 2715 atomic64_dec(&map->sleepable_refcnt); 2716 bpf_map_put(map); 2717 } 2718 } 2719 2720 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2721 { 2722 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2723 kfree(aux->used_maps); 2724 } 2725 2726 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2727 struct btf_mod_pair *used_btfs, u32 len) 2728 { 2729 #ifdef CONFIG_BPF_SYSCALL 2730 struct btf_mod_pair *btf_mod; 2731 u32 i; 2732 2733 for (i = 0; i < len; i++) { 2734 btf_mod = &used_btfs[i]; 2735 if (btf_mod->module) 2736 module_put(btf_mod->module); 2737 btf_put(btf_mod->btf); 2738 } 2739 #endif 2740 } 2741 2742 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2743 { 2744 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2745 kfree(aux->used_btfs); 2746 } 2747 2748 static void bpf_prog_free_deferred(struct work_struct *work) 2749 { 2750 struct bpf_prog_aux *aux; 2751 int i; 2752 2753 aux = container_of(work, struct bpf_prog_aux, work); 2754 #ifdef CONFIG_BPF_SYSCALL 2755 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2756 #endif 2757 #ifdef CONFIG_CGROUP_BPF 2758 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2759 bpf_cgroup_atype_put(aux->cgroup_atype); 2760 #endif 2761 bpf_free_used_maps(aux); 2762 bpf_free_used_btfs(aux); 2763 if (bpf_prog_is_dev_bound(aux)) 2764 bpf_prog_dev_bound_destroy(aux->prog); 2765 #ifdef CONFIG_PERF_EVENTS 2766 if (aux->prog->has_callchain_buf) 2767 put_callchain_buffers(); 2768 #endif 2769 if (aux->dst_trampoline) 2770 bpf_trampoline_put(aux->dst_trampoline); 2771 for (i = 0; i < aux->real_func_cnt; i++) { 2772 /* We can just unlink the subprog poke descriptor table as 2773 * it was originally linked to the main program and is also 2774 * released along with it. 2775 */ 2776 aux->func[i]->aux->poke_tab = NULL; 2777 bpf_jit_free(aux->func[i]); 2778 } 2779 if (aux->real_func_cnt) { 2780 kfree(aux->func); 2781 bpf_prog_unlock_free(aux->prog); 2782 } else { 2783 bpf_jit_free(aux->prog); 2784 } 2785 } 2786 2787 void bpf_prog_free(struct bpf_prog *fp) 2788 { 2789 struct bpf_prog_aux *aux = fp->aux; 2790 2791 if (aux->dst_prog) 2792 bpf_prog_put(aux->dst_prog); 2793 bpf_token_put(aux->token); 2794 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2795 schedule_work(&aux->work); 2796 } 2797 EXPORT_SYMBOL_GPL(bpf_prog_free); 2798 2799 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 2800 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2801 2802 void bpf_user_rnd_init_once(void) 2803 { 2804 prandom_init_once(&bpf_user_rnd_state); 2805 } 2806 2807 BPF_CALL_0(bpf_user_rnd_u32) 2808 { 2809 /* Should someone ever have the rather unwise idea to use some 2810 * of the registers passed into this function, then note that 2811 * this function is called from native eBPF and classic-to-eBPF 2812 * transformations. Register assignments from both sides are 2813 * different, f.e. classic always sets fn(ctx, A, X) here. 2814 */ 2815 struct rnd_state *state; 2816 u32 res; 2817 2818 state = &get_cpu_var(bpf_user_rnd_state); 2819 res = prandom_u32_state(state); 2820 put_cpu_var(bpf_user_rnd_state); 2821 2822 return res; 2823 } 2824 2825 BPF_CALL_0(bpf_get_raw_cpu_id) 2826 { 2827 return raw_smp_processor_id(); 2828 } 2829 2830 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2831 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2832 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2833 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2834 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2835 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2836 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2837 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2838 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2839 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2840 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2841 2842 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2843 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2844 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2845 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2846 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2847 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2848 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 2849 2850 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2851 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2852 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2853 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2854 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2855 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2856 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2857 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2858 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2859 const struct bpf_func_proto bpf_set_retval_proto __weak; 2860 const struct bpf_func_proto bpf_get_retval_proto __weak; 2861 2862 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2863 { 2864 return NULL; 2865 } 2866 2867 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2868 { 2869 return NULL; 2870 } 2871 2872 u64 __weak 2873 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2874 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2875 { 2876 return -ENOTSUPP; 2877 } 2878 EXPORT_SYMBOL_GPL(bpf_event_output); 2879 2880 /* Always built-in helper functions. */ 2881 const struct bpf_func_proto bpf_tail_call_proto = { 2882 .func = NULL, 2883 .gpl_only = false, 2884 .ret_type = RET_VOID, 2885 .arg1_type = ARG_PTR_TO_CTX, 2886 .arg2_type = ARG_CONST_MAP_PTR, 2887 .arg3_type = ARG_ANYTHING, 2888 }; 2889 2890 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2891 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2892 * eBPF and implicitly also cBPF can get JITed! 2893 */ 2894 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2895 { 2896 return prog; 2897 } 2898 2899 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2900 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2901 */ 2902 void __weak bpf_jit_compile(struct bpf_prog *prog) 2903 { 2904 } 2905 2906 bool __weak bpf_helper_changes_pkt_data(void *func) 2907 { 2908 return false; 2909 } 2910 2911 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2912 * analysis code and wants explicit zero extension inserted by verifier. 2913 * Otherwise, return FALSE. 2914 * 2915 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2916 * you don't override this. JITs that don't want these extra insns can detect 2917 * them using insn_is_zext. 2918 */ 2919 bool __weak bpf_jit_needs_zext(void) 2920 { 2921 return false; 2922 } 2923 2924 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 2925 bool __weak bpf_jit_supports_subprog_tailcalls(void) 2926 { 2927 return false; 2928 } 2929 2930 bool __weak bpf_jit_supports_kfunc_call(void) 2931 { 2932 return false; 2933 } 2934 2935 bool __weak bpf_jit_supports_far_kfunc_call(void) 2936 { 2937 return false; 2938 } 2939 2940 bool __weak bpf_jit_supports_arena(void) 2941 { 2942 return false; 2943 } 2944 2945 /* Return TRUE if the JIT backend satisfies the following two conditions: 2946 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 2947 * 2) Under the specific arch, the implementation of xchg() is the same 2948 * as atomic_xchg() on pointer-sized words. 2949 */ 2950 bool __weak bpf_jit_supports_ptr_xchg(void) 2951 { 2952 return false; 2953 } 2954 2955 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 2956 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 2957 */ 2958 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 2959 int len) 2960 { 2961 return -EFAULT; 2962 } 2963 2964 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2965 void *addr1, void *addr2) 2966 { 2967 return -ENOTSUPP; 2968 } 2969 2970 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 2971 { 2972 return ERR_PTR(-ENOTSUPP); 2973 } 2974 2975 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 2976 { 2977 return -ENOTSUPP; 2978 } 2979 2980 bool __weak bpf_jit_supports_exceptions(void) 2981 { 2982 return false; 2983 } 2984 2985 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 2986 { 2987 } 2988 2989 /* for configs without MMU or 32-bit */ 2990 __weak const struct bpf_map_ops arena_map_ops; 2991 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 2992 { 2993 return 0; 2994 } 2995 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 2996 { 2997 return 0; 2998 } 2999 3000 #ifdef CONFIG_BPF_SYSCALL 3001 static int __init bpf_global_ma_init(void) 3002 { 3003 int ret; 3004 3005 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3006 bpf_global_ma_set = !ret; 3007 return ret; 3008 } 3009 late_initcall(bpf_global_ma_init); 3010 #endif 3011 3012 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3013 EXPORT_SYMBOL(bpf_stats_enabled_key); 3014 3015 /* All definitions of tracepoints related to BPF. */ 3016 #define CREATE_TRACE_POINTS 3017 #include <linux/bpf_trace.h> 3018 3019 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3020 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3021