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/bpf.h> 26 #include <linux/btf.h> 27 #include <linux/objtool.h> 28 #include <linux/overflow.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 #include <linux/execmem.h> 41 42 #include <asm/barrier.h> 43 #include <asm/unaligned.h> 44 45 /* Registers */ 46 #define BPF_R0 regs[BPF_REG_0] 47 #define BPF_R1 regs[BPF_REG_1] 48 #define BPF_R2 regs[BPF_REG_2] 49 #define BPF_R3 regs[BPF_REG_3] 50 #define BPF_R4 regs[BPF_REG_4] 51 #define BPF_R5 regs[BPF_REG_5] 52 #define BPF_R6 regs[BPF_REG_6] 53 #define BPF_R7 regs[BPF_REG_7] 54 #define BPF_R8 regs[BPF_REG_8] 55 #define BPF_R9 regs[BPF_REG_9] 56 #define BPF_R10 regs[BPF_REG_10] 57 58 /* Named registers */ 59 #define DST regs[insn->dst_reg] 60 #define SRC regs[insn->src_reg] 61 #define FP regs[BPF_REG_FP] 62 #define AX regs[BPF_REG_AX] 63 #define ARG1 regs[BPF_REG_ARG1] 64 #define CTX regs[BPF_REG_CTX] 65 #define OFF insn->off 66 #define IMM insn->imm 67 68 struct bpf_mem_alloc bpf_global_ma; 69 bool bpf_global_ma_set; 70 71 /* No hurry in this branch 72 * 73 * Exported for the bpf jit load helper. 74 */ 75 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 76 { 77 u8 *ptr = NULL; 78 79 if (k >= SKF_NET_OFF) { 80 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 81 } else if (k >= SKF_LL_OFF) { 82 if (unlikely(!skb_mac_header_was_set(skb))) 83 return NULL; 84 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 85 } 86 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 87 return ptr; 88 89 return NULL; 90 } 91 92 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 93 enum page_size_enum { 94 __PAGE_SIZE = PAGE_SIZE 95 }; 96 97 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 98 { 99 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 100 struct bpf_prog_aux *aux; 101 struct bpf_prog *fp; 102 103 size = round_up(size, __PAGE_SIZE); 104 fp = __vmalloc(size, gfp_flags); 105 if (fp == NULL) 106 return NULL; 107 108 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 109 if (aux == NULL) { 110 vfree(fp); 111 return NULL; 112 } 113 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 114 if (!fp->active) { 115 vfree(fp); 116 kfree(aux); 117 return NULL; 118 } 119 120 fp->pages = size / PAGE_SIZE; 121 fp->aux = aux; 122 fp->aux->prog = fp; 123 fp->jit_requested = ebpf_jit_enabled(); 124 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 125 #ifdef CONFIG_CGROUP_BPF 126 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 127 #endif 128 129 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 130 #ifdef CONFIG_FINEIBT 131 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 132 #endif 133 mutex_init(&fp->aux->used_maps_mutex); 134 mutex_init(&fp->aux->dst_mutex); 135 136 return fp; 137 } 138 139 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 140 { 141 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 142 struct bpf_prog *prog; 143 int cpu; 144 145 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 146 if (!prog) 147 return NULL; 148 149 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 150 if (!prog->stats) { 151 free_percpu(prog->active); 152 kfree(prog->aux); 153 vfree(prog); 154 return NULL; 155 } 156 157 for_each_possible_cpu(cpu) { 158 struct bpf_prog_stats *pstats; 159 160 pstats = per_cpu_ptr(prog->stats, cpu); 161 u64_stats_init(&pstats->syncp); 162 } 163 return prog; 164 } 165 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 166 167 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 168 { 169 if (!prog->aux->nr_linfo || !prog->jit_requested) 170 return 0; 171 172 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 173 sizeof(*prog->aux->jited_linfo), 174 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 175 if (!prog->aux->jited_linfo) 176 return -ENOMEM; 177 178 return 0; 179 } 180 181 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 182 { 183 if (prog->aux->jited_linfo && 184 (!prog->jited || !prog->aux->jited_linfo[0])) { 185 kvfree(prog->aux->jited_linfo); 186 prog->aux->jited_linfo = NULL; 187 } 188 189 kfree(prog->aux->kfunc_tab); 190 prog->aux->kfunc_tab = NULL; 191 } 192 193 /* The jit engine is responsible to provide an array 194 * for insn_off to the jited_off mapping (insn_to_jit_off). 195 * 196 * The idx to this array is the insn_off. Hence, the insn_off 197 * here is relative to the prog itself instead of the main prog. 198 * This array has one entry for each xlated bpf insn. 199 * 200 * jited_off is the byte off to the end of the jited insn. 201 * 202 * Hence, with 203 * insn_start: 204 * The first bpf insn off of the prog. The insn off 205 * here is relative to the main prog. 206 * e.g. if prog is a subprog, insn_start > 0 207 * linfo_idx: 208 * The prog's idx to prog->aux->linfo and jited_linfo 209 * 210 * jited_linfo[linfo_idx] = prog->bpf_func 211 * 212 * For i > linfo_idx, 213 * 214 * jited_linfo[i] = prog->bpf_func + 215 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 216 */ 217 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 218 const u32 *insn_to_jit_off) 219 { 220 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 221 const struct bpf_line_info *linfo; 222 void **jited_linfo; 223 224 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 225 /* Userspace did not provide linfo */ 226 return; 227 228 linfo_idx = prog->aux->linfo_idx; 229 linfo = &prog->aux->linfo[linfo_idx]; 230 insn_start = linfo[0].insn_off; 231 insn_end = insn_start + prog->len; 232 233 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 234 jited_linfo[0] = prog->bpf_func; 235 236 nr_linfo = prog->aux->nr_linfo - linfo_idx; 237 238 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 239 /* The verifier ensures that linfo[i].insn_off is 240 * strictly increasing 241 */ 242 jited_linfo[i] = prog->bpf_func + 243 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 244 } 245 246 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 247 gfp_t gfp_extra_flags) 248 { 249 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 250 struct bpf_prog *fp; 251 u32 pages; 252 253 size = round_up(size, PAGE_SIZE); 254 pages = size / PAGE_SIZE; 255 if (pages <= fp_old->pages) 256 return fp_old; 257 258 fp = __vmalloc(size, gfp_flags); 259 if (fp) { 260 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 261 fp->pages = pages; 262 fp->aux->prog = fp; 263 264 /* We keep fp->aux from fp_old around in the new 265 * reallocated structure. 266 */ 267 fp_old->aux = NULL; 268 fp_old->stats = NULL; 269 fp_old->active = NULL; 270 __bpf_prog_free(fp_old); 271 } 272 273 return fp; 274 } 275 276 void __bpf_prog_free(struct bpf_prog *fp) 277 { 278 if (fp->aux) { 279 mutex_destroy(&fp->aux->used_maps_mutex); 280 mutex_destroy(&fp->aux->dst_mutex); 281 kfree(fp->aux->poke_tab); 282 kfree(fp->aux); 283 } 284 free_percpu(fp->stats); 285 free_percpu(fp->active); 286 vfree(fp); 287 } 288 289 int bpf_prog_calc_tag(struct bpf_prog *fp) 290 { 291 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 292 u32 raw_size = bpf_prog_tag_scratch_size(fp); 293 u32 digest[SHA1_DIGEST_WORDS]; 294 u32 ws[SHA1_WORKSPACE_WORDS]; 295 u32 i, bsize, psize, blocks; 296 struct bpf_insn *dst; 297 bool was_ld_map; 298 u8 *raw, *todo; 299 __be32 *result; 300 __be64 *bits; 301 302 raw = vmalloc(raw_size); 303 if (!raw) 304 return -ENOMEM; 305 306 sha1_init(digest); 307 memset(ws, 0, sizeof(ws)); 308 309 /* We need to take out the map fd for the digest calculation 310 * since they are unstable from user space side. 311 */ 312 dst = (void *)raw; 313 for (i = 0, was_ld_map = false; i < fp->len; i++) { 314 dst[i] = fp->insnsi[i]; 315 if (!was_ld_map && 316 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 317 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 318 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 319 was_ld_map = true; 320 dst[i].imm = 0; 321 } else if (was_ld_map && 322 dst[i].code == 0 && 323 dst[i].dst_reg == 0 && 324 dst[i].src_reg == 0 && 325 dst[i].off == 0) { 326 was_ld_map = false; 327 dst[i].imm = 0; 328 } else { 329 was_ld_map = false; 330 } 331 } 332 333 psize = bpf_prog_insn_size(fp); 334 memset(&raw[psize], 0, raw_size - psize); 335 raw[psize++] = 0x80; 336 337 bsize = round_up(psize, SHA1_BLOCK_SIZE); 338 blocks = bsize / SHA1_BLOCK_SIZE; 339 todo = raw; 340 if (bsize - psize >= sizeof(__be64)) { 341 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 342 } else { 343 bits = (__be64 *)(todo + bsize + bits_offset); 344 blocks++; 345 } 346 *bits = cpu_to_be64((psize - 1) << 3); 347 348 while (blocks--) { 349 sha1_transform(digest, todo, ws); 350 todo += SHA1_BLOCK_SIZE; 351 } 352 353 result = (__force __be32 *)digest; 354 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 355 result[i] = cpu_to_be32(digest[i]); 356 memcpy(fp->tag, result, sizeof(fp->tag)); 357 358 vfree(raw); 359 return 0; 360 } 361 362 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 363 s32 end_new, s32 curr, const bool probe_pass) 364 { 365 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 366 s32 delta = end_new - end_old; 367 s64 imm = insn->imm; 368 369 if (curr < pos && curr + imm + 1 >= end_old) 370 imm += delta; 371 else if (curr >= end_new && curr + imm + 1 < end_new) 372 imm -= delta; 373 if (imm < imm_min || imm > imm_max) 374 return -ERANGE; 375 if (!probe_pass) 376 insn->imm = imm; 377 return 0; 378 } 379 380 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 381 s32 end_new, s32 curr, const bool probe_pass) 382 { 383 s64 off_min, off_max, off; 384 s32 delta = end_new - end_old; 385 386 if (insn->code == (BPF_JMP32 | BPF_JA)) { 387 off = insn->imm; 388 off_min = S32_MIN; 389 off_max = S32_MAX; 390 } else { 391 off = insn->off; 392 off_min = S16_MIN; 393 off_max = S16_MAX; 394 } 395 396 if (curr < pos && curr + off + 1 >= end_old) 397 off += delta; 398 else if (curr >= end_new && curr + off + 1 < end_new) 399 off -= delta; 400 if (off < off_min || off > off_max) 401 return -ERANGE; 402 if (!probe_pass) { 403 if (insn->code == (BPF_JMP32 | BPF_JA)) 404 insn->imm = off; 405 else 406 insn->off = off; 407 } 408 return 0; 409 } 410 411 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 412 s32 end_new, const bool probe_pass) 413 { 414 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 415 struct bpf_insn *insn = prog->insnsi; 416 int ret = 0; 417 418 for (i = 0; i < insn_cnt; i++, insn++) { 419 u8 code; 420 421 /* In the probing pass we still operate on the original, 422 * unpatched image in order to check overflows before we 423 * do any other adjustments. Therefore skip the patchlet. 424 */ 425 if (probe_pass && i == pos) { 426 i = end_new; 427 insn = prog->insnsi + end_old; 428 } 429 if (bpf_pseudo_func(insn)) { 430 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 431 end_new, i, probe_pass); 432 if (ret) 433 return ret; 434 continue; 435 } 436 code = insn->code; 437 if ((BPF_CLASS(code) != BPF_JMP && 438 BPF_CLASS(code) != BPF_JMP32) || 439 BPF_OP(code) == BPF_EXIT) 440 continue; 441 /* Adjust offset of jmps if we cross patch boundaries. */ 442 if (BPF_OP(code) == BPF_CALL) { 443 if (insn->src_reg != BPF_PSEUDO_CALL) 444 continue; 445 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 446 end_new, i, probe_pass); 447 } else { 448 ret = bpf_adj_delta_to_off(insn, pos, end_old, 449 end_new, i, probe_pass); 450 } 451 if (ret) 452 break; 453 } 454 455 return ret; 456 } 457 458 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 459 { 460 struct bpf_line_info *linfo; 461 u32 i, nr_linfo; 462 463 nr_linfo = prog->aux->nr_linfo; 464 if (!nr_linfo || !delta) 465 return; 466 467 linfo = prog->aux->linfo; 468 469 for (i = 0; i < nr_linfo; i++) 470 if (off < linfo[i].insn_off) 471 break; 472 473 /* Push all off < linfo[i].insn_off by delta */ 474 for (; i < nr_linfo; i++) 475 linfo[i].insn_off += delta; 476 } 477 478 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 479 const struct bpf_insn *patch, u32 len) 480 { 481 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 482 const u32 cnt_max = S16_MAX; 483 struct bpf_prog *prog_adj; 484 int err; 485 486 /* Since our patchlet doesn't expand the image, we're done. */ 487 if (insn_delta == 0) { 488 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 489 return prog; 490 } 491 492 insn_adj_cnt = prog->len + insn_delta; 493 494 /* Reject anything that would potentially let the insn->off 495 * target overflow when we have excessive program expansions. 496 * We need to probe here before we do any reallocation where 497 * we afterwards may not fail anymore. 498 */ 499 if (insn_adj_cnt > cnt_max && 500 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 501 return ERR_PTR(err); 502 503 /* Several new instructions need to be inserted. Make room 504 * for them. Likely, there's no need for a new allocation as 505 * last page could have large enough tailroom. 506 */ 507 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 508 GFP_USER); 509 if (!prog_adj) 510 return ERR_PTR(-ENOMEM); 511 512 prog_adj->len = insn_adj_cnt; 513 514 /* Patching happens in 3 steps: 515 * 516 * 1) Move over tail of insnsi from next instruction onwards, 517 * so we can patch the single target insn with one or more 518 * new ones (patching is always from 1 to n insns, n > 0). 519 * 2) Inject new instructions at the target location. 520 * 3) Adjust branch offsets if necessary. 521 */ 522 insn_rest = insn_adj_cnt - off - len; 523 524 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 525 sizeof(*patch) * insn_rest); 526 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 527 528 /* We are guaranteed to not fail at this point, otherwise 529 * the ship has sailed to reverse to the original state. An 530 * overflow cannot happen at this point. 531 */ 532 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 533 534 bpf_adj_linfo(prog_adj, off, insn_delta); 535 536 return prog_adj; 537 } 538 539 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 540 { 541 /* Branch offsets can't overflow when program is shrinking, no need 542 * to call bpf_adj_branches(..., true) here 543 */ 544 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 545 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 546 prog->len -= cnt; 547 548 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false)); 549 } 550 551 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 552 { 553 int i; 554 555 for (i = 0; i < fp->aux->real_func_cnt; i++) 556 bpf_prog_kallsyms_del(fp->aux->func[i]); 557 } 558 559 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 560 { 561 bpf_prog_kallsyms_del_subprogs(fp); 562 bpf_prog_kallsyms_del(fp); 563 } 564 565 #ifdef CONFIG_BPF_JIT 566 /* All BPF JIT sysctl knobs here. */ 567 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 568 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 569 int bpf_jit_harden __read_mostly; 570 long bpf_jit_limit __read_mostly; 571 long bpf_jit_limit_max __read_mostly; 572 573 static void 574 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 575 { 576 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 577 578 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 579 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 580 } 581 582 static void 583 bpf_prog_ksym_set_name(struct bpf_prog *prog) 584 { 585 char *sym = prog->aux->ksym.name; 586 const char *end = sym + KSYM_NAME_LEN; 587 const struct btf_type *type; 588 const char *func_name; 589 590 BUILD_BUG_ON(sizeof("bpf_prog_") + 591 sizeof(prog->tag) * 2 + 592 /* name has been null terminated. 593 * We should need +1 for the '_' preceding 594 * the name. However, the null character 595 * is double counted between the name and the 596 * sizeof("bpf_prog_") above, so we omit 597 * the +1 here. 598 */ 599 sizeof(prog->aux->name) > KSYM_NAME_LEN); 600 601 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 602 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 603 604 /* prog->aux->name will be ignored if full btf name is available */ 605 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 606 type = btf_type_by_id(prog->aux->btf, 607 prog->aux->func_info[prog->aux->func_idx].type_id); 608 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 609 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 610 return; 611 } 612 613 if (prog->aux->name[0]) 614 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 615 else 616 *sym = 0; 617 } 618 619 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 620 { 621 return container_of(n, struct bpf_ksym, tnode)->start; 622 } 623 624 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 625 struct latch_tree_node *b) 626 { 627 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 628 } 629 630 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 631 { 632 unsigned long val = (unsigned long)key; 633 const struct bpf_ksym *ksym; 634 635 ksym = container_of(n, struct bpf_ksym, tnode); 636 637 if (val < ksym->start) 638 return -1; 639 /* Ensure that we detect return addresses as part of the program, when 640 * the final instruction is a call for a program part of the stack 641 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 642 */ 643 if (val > ksym->end) 644 return 1; 645 646 return 0; 647 } 648 649 static const struct latch_tree_ops bpf_tree_ops = { 650 .less = bpf_tree_less, 651 .comp = bpf_tree_comp, 652 }; 653 654 static DEFINE_SPINLOCK(bpf_lock); 655 static LIST_HEAD(bpf_kallsyms); 656 static struct latch_tree_root bpf_tree __cacheline_aligned; 657 658 void bpf_ksym_add(struct bpf_ksym *ksym) 659 { 660 spin_lock_bh(&bpf_lock); 661 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 662 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 663 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 664 spin_unlock_bh(&bpf_lock); 665 } 666 667 static void __bpf_ksym_del(struct bpf_ksym *ksym) 668 { 669 if (list_empty(&ksym->lnode)) 670 return; 671 672 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 673 list_del_rcu(&ksym->lnode); 674 } 675 676 void bpf_ksym_del(struct bpf_ksym *ksym) 677 { 678 spin_lock_bh(&bpf_lock); 679 __bpf_ksym_del(ksym); 680 spin_unlock_bh(&bpf_lock); 681 } 682 683 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 684 { 685 return fp->jited && !bpf_prog_was_classic(fp); 686 } 687 688 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 689 { 690 if (!bpf_prog_kallsyms_candidate(fp) || 691 !bpf_token_capable(fp->aux->token, CAP_BPF)) 692 return; 693 694 bpf_prog_ksym_set_addr(fp); 695 bpf_prog_ksym_set_name(fp); 696 fp->aux->ksym.prog = true; 697 698 bpf_ksym_add(&fp->aux->ksym); 699 700 #ifdef CONFIG_FINEIBT 701 /* 702 * When FineIBT, code in the __cfi_foo() symbols can get executed 703 * and hence unwinder needs help. 704 */ 705 if (cfi_mode != CFI_FINEIBT) 706 return; 707 708 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 709 "__cfi_%s", fp->aux->ksym.name); 710 711 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 712 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 713 714 bpf_ksym_add(&fp->aux->ksym_prefix); 715 #endif 716 } 717 718 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 719 { 720 if (!bpf_prog_kallsyms_candidate(fp)) 721 return; 722 723 bpf_ksym_del(&fp->aux->ksym); 724 #ifdef CONFIG_FINEIBT 725 if (cfi_mode != CFI_FINEIBT) 726 return; 727 bpf_ksym_del(&fp->aux->ksym_prefix); 728 #endif 729 } 730 731 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 732 { 733 struct latch_tree_node *n; 734 735 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 736 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 737 } 738 739 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 740 unsigned long *off, char *sym) 741 { 742 struct bpf_ksym *ksym; 743 char *ret = NULL; 744 745 rcu_read_lock(); 746 ksym = bpf_ksym_find(addr); 747 if (ksym) { 748 unsigned long symbol_start = ksym->start; 749 unsigned long symbol_end = ksym->end; 750 751 strscpy(sym, ksym->name, KSYM_NAME_LEN); 752 753 ret = sym; 754 if (size) 755 *size = symbol_end - symbol_start; 756 if (off) 757 *off = addr - symbol_start; 758 } 759 rcu_read_unlock(); 760 761 return ret; 762 } 763 764 bool is_bpf_text_address(unsigned long addr) 765 { 766 bool ret; 767 768 rcu_read_lock(); 769 ret = bpf_ksym_find(addr) != NULL; 770 rcu_read_unlock(); 771 772 return ret; 773 } 774 775 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 776 { 777 struct bpf_ksym *ksym = bpf_ksym_find(addr); 778 779 return ksym && ksym->prog ? 780 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 781 NULL; 782 } 783 784 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 785 { 786 const struct exception_table_entry *e = NULL; 787 struct bpf_prog *prog; 788 789 rcu_read_lock(); 790 prog = bpf_prog_ksym_find(addr); 791 if (!prog) 792 goto out; 793 if (!prog->aux->num_exentries) 794 goto out; 795 796 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 797 out: 798 rcu_read_unlock(); 799 return e; 800 } 801 802 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 803 char *sym) 804 { 805 struct bpf_ksym *ksym; 806 unsigned int it = 0; 807 int ret = -ERANGE; 808 809 if (!bpf_jit_kallsyms_enabled()) 810 return ret; 811 812 rcu_read_lock(); 813 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 814 if (it++ != symnum) 815 continue; 816 817 strscpy(sym, ksym->name, KSYM_NAME_LEN); 818 819 *value = ksym->start; 820 *type = BPF_SYM_ELF_TYPE; 821 822 ret = 0; 823 break; 824 } 825 rcu_read_unlock(); 826 827 return ret; 828 } 829 830 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 831 struct bpf_jit_poke_descriptor *poke) 832 { 833 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 834 static const u32 poke_tab_max = 1024; 835 u32 slot = prog->aux->size_poke_tab; 836 u32 size = slot + 1; 837 838 if (size > poke_tab_max) 839 return -ENOSPC; 840 if (poke->tailcall_target || poke->tailcall_target_stable || 841 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 842 return -EINVAL; 843 844 switch (poke->reason) { 845 case BPF_POKE_REASON_TAIL_CALL: 846 if (!poke->tail_call.map) 847 return -EINVAL; 848 break; 849 default: 850 return -EINVAL; 851 } 852 853 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 854 if (!tab) 855 return -ENOMEM; 856 857 memcpy(&tab[slot], poke, sizeof(*poke)); 858 prog->aux->size_poke_tab = size; 859 prog->aux->poke_tab = tab; 860 861 return slot; 862 } 863 864 /* 865 * BPF program pack allocator. 866 * 867 * Most BPF programs are pretty small. Allocating a hole page for each 868 * program is sometime a waste. Many small bpf program also adds pressure 869 * to instruction TLB. To solve this issue, we introduce a BPF program pack 870 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 871 * to host BPF programs. 872 */ 873 #define BPF_PROG_CHUNK_SHIFT 6 874 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 875 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 876 877 struct bpf_prog_pack { 878 struct list_head list; 879 void *ptr; 880 unsigned long bitmap[]; 881 }; 882 883 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 884 { 885 memset(area, 0, size); 886 } 887 888 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 889 890 static DEFINE_MUTEX(pack_mutex); 891 static LIST_HEAD(pack_list); 892 893 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 894 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 895 */ 896 #ifdef PMD_SIZE 897 /* PMD_SIZE is really big for some archs. It doesn't make sense to 898 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 899 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 900 * greater than or equal to 2MB. 901 */ 902 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 903 #else 904 #define BPF_PROG_PACK_SIZE PAGE_SIZE 905 #endif 906 907 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 908 909 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 910 { 911 struct bpf_prog_pack *pack; 912 int err; 913 914 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 915 GFP_KERNEL); 916 if (!pack) 917 return NULL; 918 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 919 if (!pack->ptr) 920 goto out; 921 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 922 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 923 924 set_vm_flush_reset_perms(pack->ptr); 925 err = set_memory_rox((unsigned long)pack->ptr, 926 BPF_PROG_PACK_SIZE / PAGE_SIZE); 927 if (err) 928 goto out; 929 list_add_tail(&pack->list, &pack_list); 930 return pack; 931 932 out: 933 bpf_jit_free_exec(pack->ptr); 934 kfree(pack); 935 return NULL; 936 } 937 938 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 939 { 940 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 941 struct bpf_prog_pack *pack; 942 unsigned long pos; 943 void *ptr = NULL; 944 945 mutex_lock(&pack_mutex); 946 if (size > BPF_PROG_PACK_SIZE) { 947 size = round_up(size, PAGE_SIZE); 948 ptr = bpf_jit_alloc_exec(size); 949 if (ptr) { 950 int err; 951 952 bpf_fill_ill_insns(ptr, size); 953 set_vm_flush_reset_perms(ptr); 954 err = set_memory_rox((unsigned long)ptr, 955 size / PAGE_SIZE); 956 if (err) { 957 bpf_jit_free_exec(ptr); 958 ptr = NULL; 959 } 960 } 961 goto out; 962 } 963 list_for_each_entry(pack, &pack_list, list) { 964 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 965 nbits, 0); 966 if (pos < BPF_PROG_CHUNK_COUNT) 967 goto found_free_area; 968 } 969 970 pack = alloc_new_pack(bpf_fill_ill_insns); 971 if (!pack) 972 goto out; 973 974 pos = 0; 975 976 found_free_area: 977 bitmap_set(pack->bitmap, pos, nbits); 978 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 979 980 out: 981 mutex_unlock(&pack_mutex); 982 return ptr; 983 } 984 985 void bpf_prog_pack_free(void *ptr, u32 size) 986 { 987 struct bpf_prog_pack *pack = NULL, *tmp; 988 unsigned int nbits; 989 unsigned long pos; 990 991 mutex_lock(&pack_mutex); 992 if (size > BPF_PROG_PACK_SIZE) { 993 bpf_jit_free_exec(ptr); 994 goto out; 995 } 996 997 list_for_each_entry(tmp, &pack_list, list) { 998 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 999 pack = tmp; 1000 break; 1001 } 1002 } 1003 1004 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1005 goto out; 1006 1007 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1008 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1009 1010 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1011 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1012 1013 bitmap_clear(pack->bitmap, pos, nbits); 1014 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1015 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1016 list_del(&pack->list); 1017 bpf_jit_free_exec(pack->ptr); 1018 kfree(pack); 1019 } 1020 out: 1021 mutex_unlock(&pack_mutex); 1022 } 1023 1024 static atomic_long_t bpf_jit_current; 1025 1026 /* Can be overridden by an arch's JIT compiler if it has a custom, 1027 * dedicated BPF backend memory area, or if neither of the two 1028 * below apply. 1029 */ 1030 u64 __weak bpf_jit_alloc_exec_limit(void) 1031 { 1032 #if defined(MODULES_VADDR) 1033 return MODULES_END - MODULES_VADDR; 1034 #else 1035 return VMALLOC_END - VMALLOC_START; 1036 #endif 1037 } 1038 1039 static int __init bpf_jit_charge_init(void) 1040 { 1041 /* Only used as heuristic here to derive limit. */ 1042 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1043 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1044 PAGE_SIZE), LONG_MAX); 1045 return 0; 1046 } 1047 pure_initcall(bpf_jit_charge_init); 1048 1049 int bpf_jit_charge_modmem(u32 size) 1050 { 1051 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1052 if (!bpf_capable()) { 1053 atomic_long_sub(size, &bpf_jit_current); 1054 return -EPERM; 1055 } 1056 } 1057 1058 return 0; 1059 } 1060 1061 void bpf_jit_uncharge_modmem(u32 size) 1062 { 1063 atomic_long_sub(size, &bpf_jit_current); 1064 } 1065 1066 void *__weak bpf_jit_alloc_exec(unsigned long size) 1067 { 1068 return execmem_alloc(EXECMEM_BPF, size); 1069 } 1070 1071 void __weak bpf_jit_free_exec(void *addr) 1072 { 1073 execmem_free(addr); 1074 } 1075 1076 struct bpf_binary_header * 1077 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1078 unsigned int alignment, 1079 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1080 { 1081 struct bpf_binary_header *hdr; 1082 u32 size, hole, start; 1083 1084 WARN_ON_ONCE(!is_power_of_2(alignment) || 1085 alignment > BPF_IMAGE_ALIGNMENT); 1086 1087 /* Most of BPF filters are really small, but if some of them 1088 * fill a page, allow at least 128 extra bytes to insert a 1089 * random section of illegal instructions. 1090 */ 1091 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1092 1093 if (bpf_jit_charge_modmem(size)) 1094 return NULL; 1095 hdr = bpf_jit_alloc_exec(size); 1096 if (!hdr) { 1097 bpf_jit_uncharge_modmem(size); 1098 return NULL; 1099 } 1100 1101 /* Fill space with illegal/arch-dep instructions. */ 1102 bpf_fill_ill_insns(hdr, size); 1103 1104 hdr->size = size; 1105 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1106 PAGE_SIZE - sizeof(*hdr)); 1107 start = get_random_u32_below(hole) & ~(alignment - 1); 1108 1109 /* Leave a random number of instructions before BPF code. */ 1110 *image_ptr = &hdr->image[start]; 1111 1112 return hdr; 1113 } 1114 1115 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1116 { 1117 u32 size = hdr->size; 1118 1119 bpf_jit_free_exec(hdr); 1120 bpf_jit_uncharge_modmem(size); 1121 } 1122 1123 /* Allocate jit binary from bpf_prog_pack allocator. 1124 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1125 * to the memory. To solve this problem, a RW buffer is also allocated at 1126 * as the same time. The JIT engine should calculate offsets based on the 1127 * RO memory address, but write JITed program to the RW buffer. Once the 1128 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1129 * the JITed program to the RO memory. 1130 */ 1131 struct bpf_binary_header * 1132 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1133 unsigned int alignment, 1134 struct bpf_binary_header **rw_header, 1135 u8 **rw_image, 1136 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1137 { 1138 struct bpf_binary_header *ro_header; 1139 u32 size, hole, start; 1140 1141 WARN_ON_ONCE(!is_power_of_2(alignment) || 1142 alignment > BPF_IMAGE_ALIGNMENT); 1143 1144 /* add 16 bytes for a random section of illegal instructions */ 1145 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1146 1147 if (bpf_jit_charge_modmem(size)) 1148 return NULL; 1149 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1150 if (!ro_header) { 1151 bpf_jit_uncharge_modmem(size); 1152 return NULL; 1153 } 1154 1155 *rw_header = kvmalloc(size, GFP_KERNEL); 1156 if (!*rw_header) { 1157 bpf_prog_pack_free(ro_header, size); 1158 bpf_jit_uncharge_modmem(size); 1159 return NULL; 1160 } 1161 1162 /* Fill space with illegal/arch-dep instructions. */ 1163 bpf_fill_ill_insns(*rw_header, size); 1164 (*rw_header)->size = size; 1165 1166 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1167 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1168 start = get_random_u32_below(hole) & ~(alignment - 1); 1169 1170 *image_ptr = &ro_header->image[start]; 1171 *rw_image = &(*rw_header)->image[start]; 1172 1173 return ro_header; 1174 } 1175 1176 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1177 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog, 1178 struct bpf_binary_header *ro_header, 1179 struct bpf_binary_header *rw_header) 1180 { 1181 void *ptr; 1182 1183 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1184 1185 kvfree(rw_header); 1186 1187 if (IS_ERR(ptr)) { 1188 bpf_prog_pack_free(ro_header, ro_header->size); 1189 return PTR_ERR(ptr); 1190 } 1191 return 0; 1192 } 1193 1194 /* bpf_jit_binary_pack_free is called in two different scenarios: 1195 * 1) when the program is freed after; 1196 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1197 * For case 2), we need to free both the RO memory and the RW buffer. 1198 * 1199 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1200 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1201 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1202 * bpf_arch_text_copy (when jit fails). 1203 */ 1204 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1205 struct bpf_binary_header *rw_header) 1206 { 1207 u32 size = ro_header->size; 1208 1209 bpf_prog_pack_free(ro_header, size); 1210 kvfree(rw_header); 1211 bpf_jit_uncharge_modmem(size); 1212 } 1213 1214 struct bpf_binary_header * 1215 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1216 { 1217 unsigned long real_start = (unsigned long)fp->bpf_func; 1218 unsigned long addr; 1219 1220 addr = real_start & BPF_PROG_CHUNK_MASK; 1221 return (void *)addr; 1222 } 1223 1224 static inline struct bpf_binary_header * 1225 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1226 { 1227 unsigned long real_start = (unsigned long)fp->bpf_func; 1228 unsigned long addr; 1229 1230 addr = real_start & PAGE_MASK; 1231 return (void *)addr; 1232 } 1233 1234 /* This symbol is only overridden by archs that have different 1235 * requirements than the usual eBPF JITs, f.e. when they only 1236 * implement cBPF JIT, do not set images read-only, etc. 1237 */ 1238 void __weak bpf_jit_free(struct bpf_prog *fp) 1239 { 1240 if (fp->jited) { 1241 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1242 1243 bpf_jit_binary_free(hdr); 1244 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1245 } 1246 1247 bpf_prog_unlock_free(fp); 1248 } 1249 1250 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1251 const struct bpf_insn *insn, bool extra_pass, 1252 u64 *func_addr, bool *func_addr_fixed) 1253 { 1254 s16 off = insn->off; 1255 s32 imm = insn->imm; 1256 u8 *addr; 1257 int err; 1258 1259 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1260 if (!*func_addr_fixed) { 1261 /* Place-holder address till the last pass has collected 1262 * all addresses for JITed subprograms in which case we 1263 * can pick them up from prog->aux. 1264 */ 1265 if (!extra_pass) 1266 addr = NULL; 1267 else if (prog->aux->func && 1268 off >= 0 && off < prog->aux->real_func_cnt) 1269 addr = (u8 *)prog->aux->func[off]->bpf_func; 1270 else 1271 return -EINVAL; 1272 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1273 bpf_jit_supports_far_kfunc_call()) { 1274 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1275 if (err) 1276 return err; 1277 } else { 1278 /* Address of a BPF helper call. Since part of the core 1279 * kernel, it's always at a fixed location. __bpf_call_base 1280 * and the helper with imm relative to it are both in core 1281 * kernel. 1282 */ 1283 addr = (u8 *)__bpf_call_base + imm; 1284 } 1285 1286 *func_addr = (unsigned long)addr; 1287 return 0; 1288 } 1289 1290 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1291 const struct bpf_insn *aux, 1292 struct bpf_insn *to_buff, 1293 bool emit_zext) 1294 { 1295 struct bpf_insn *to = to_buff; 1296 u32 imm_rnd = get_random_u32(); 1297 s16 off; 1298 1299 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1300 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1301 1302 /* Constraints on AX register: 1303 * 1304 * AX register is inaccessible from user space. It is mapped in 1305 * all JITs, and used here for constant blinding rewrites. It is 1306 * typically "stateless" meaning its contents are only valid within 1307 * the executed instruction, but not across several instructions. 1308 * There are a few exceptions however which are further detailed 1309 * below. 1310 * 1311 * Constant blinding is only used by JITs, not in the interpreter. 1312 * The interpreter uses AX in some occasions as a local temporary 1313 * register e.g. in DIV or MOD instructions. 1314 * 1315 * In restricted circumstances, the verifier can also use the AX 1316 * register for rewrites as long as they do not interfere with 1317 * the above cases! 1318 */ 1319 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1320 goto out; 1321 1322 if (from->imm == 0 && 1323 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1324 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1325 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1326 goto out; 1327 } 1328 1329 switch (from->code) { 1330 case BPF_ALU | BPF_ADD | BPF_K: 1331 case BPF_ALU | BPF_SUB | BPF_K: 1332 case BPF_ALU | BPF_AND | BPF_K: 1333 case BPF_ALU | BPF_OR | BPF_K: 1334 case BPF_ALU | BPF_XOR | BPF_K: 1335 case BPF_ALU | BPF_MUL | BPF_K: 1336 case BPF_ALU | BPF_MOV | BPF_K: 1337 case BPF_ALU | BPF_DIV | BPF_K: 1338 case BPF_ALU | BPF_MOD | BPF_K: 1339 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1340 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1341 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1342 break; 1343 1344 case BPF_ALU64 | BPF_ADD | BPF_K: 1345 case BPF_ALU64 | BPF_SUB | BPF_K: 1346 case BPF_ALU64 | BPF_AND | BPF_K: 1347 case BPF_ALU64 | BPF_OR | BPF_K: 1348 case BPF_ALU64 | BPF_XOR | BPF_K: 1349 case BPF_ALU64 | BPF_MUL | BPF_K: 1350 case BPF_ALU64 | BPF_MOV | BPF_K: 1351 case BPF_ALU64 | BPF_DIV | BPF_K: 1352 case BPF_ALU64 | BPF_MOD | BPF_K: 1353 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1354 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1355 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1356 break; 1357 1358 case BPF_JMP | BPF_JEQ | BPF_K: 1359 case BPF_JMP | BPF_JNE | BPF_K: 1360 case BPF_JMP | BPF_JGT | BPF_K: 1361 case BPF_JMP | BPF_JLT | BPF_K: 1362 case BPF_JMP | BPF_JGE | BPF_K: 1363 case BPF_JMP | BPF_JLE | BPF_K: 1364 case BPF_JMP | BPF_JSGT | BPF_K: 1365 case BPF_JMP | BPF_JSLT | BPF_K: 1366 case BPF_JMP | BPF_JSGE | BPF_K: 1367 case BPF_JMP | BPF_JSLE | BPF_K: 1368 case BPF_JMP | BPF_JSET | BPF_K: 1369 /* Accommodate for extra offset in case of a backjump. */ 1370 off = from->off; 1371 if (off < 0) 1372 off -= 2; 1373 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1374 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1375 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1376 break; 1377 1378 case BPF_JMP32 | BPF_JEQ | BPF_K: 1379 case BPF_JMP32 | BPF_JNE | BPF_K: 1380 case BPF_JMP32 | BPF_JGT | BPF_K: 1381 case BPF_JMP32 | BPF_JLT | BPF_K: 1382 case BPF_JMP32 | BPF_JGE | BPF_K: 1383 case BPF_JMP32 | BPF_JLE | BPF_K: 1384 case BPF_JMP32 | BPF_JSGT | BPF_K: 1385 case BPF_JMP32 | BPF_JSLT | BPF_K: 1386 case BPF_JMP32 | BPF_JSGE | BPF_K: 1387 case BPF_JMP32 | BPF_JSLE | BPF_K: 1388 case BPF_JMP32 | BPF_JSET | BPF_K: 1389 /* Accommodate for extra offset in case of a backjump. */ 1390 off = from->off; 1391 if (off < 0) 1392 off -= 2; 1393 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1394 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1395 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1396 off); 1397 break; 1398 1399 case BPF_LD | BPF_IMM | BPF_DW: 1400 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1401 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1402 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1403 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1404 break; 1405 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1406 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1407 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1408 if (emit_zext) 1409 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1410 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1411 break; 1412 1413 case BPF_ST | BPF_MEM | BPF_DW: 1414 case BPF_ST | BPF_MEM | BPF_W: 1415 case BPF_ST | BPF_MEM | BPF_H: 1416 case BPF_ST | BPF_MEM | BPF_B: 1417 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1418 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1419 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1420 break; 1421 } 1422 out: 1423 return to - to_buff; 1424 } 1425 1426 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1427 gfp_t gfp_extra_flags) 1428 { 1429 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1430 struct bpf_prog *fp; 1431 1432 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1433 if (fp != NULL) { 1434 /* aux->prog still points to the fp_other one, so 1435 * when promoting the clone to the real program, 1436 * this still needs to be adapted. 1437 */ 1438 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1439 } 1440 1441 return fp; 1442 } 1443 1444 static void bpf_prog_clone_free(struct bpf_prog *fp) 1445 { 1446 /* aux was stolen by the other clone, so we cannot free 1447 * it from this path! It will be freed eventually by the 1448 * other program on release. 1449 * 1450 * At this point, we don't need a deferred release since 1451 * clone is guaranteed to not be locked. 1452 */ 1453 fp->aux = NULL; 1454 fp->stats = NULL; 1455 fp->active = NULL; 1456 __bpf_prog_free(fp); 1457 } 1458 1459 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1460 { 1461 /* We have to repoint aux->prog to self, as we don't 1462 * know whether fp here is the clone or the original. 1463 */ 1464 fp->aux->prog = fp; 1465 bpf_prog_clone_free(fp_other); 1466 } 1467 1468 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1469 { 1470 struct bpf_insn insn_buff[16], aux[2]; 1471 struct bpf_prog *clone, *tmp; 1472 int insn_delta, insn_cnt; 1473 struct bpf_insn *insn; 1474 int i, rewritten; 1475 1476 if (!prog->blinding_requested || prog->blinded) 1477 return prog; 1478 1479 clone = bpf_prog_clone_create(prog, GFP_USER); 1480 if (!clone) 1481 return ERR_PTR(-ENOMEM); 1482 1483 insn_cnt = clone->len; 1484 insn = clone->insnsi; 1485 1486 for (i = 0; i < insn_cnt; i++, insn++) { 1487 if (bpf_pseudo_func(insn)) { 1488 /* ld_imm64 with an address of bpf subprog is not 1489 * a user controlled constant. Don't randomize it, 1490 * since it will conflict with jit_subprogs() logic. 1491 */ 1492 insn++; 1493 i++; 1494 continue; 1495 } 1496 1497 /* We temporarily need to hold the original ld64 insn 1498 * so that we can still access the first part in the 1499 * second blinding run. 1500 */ 1501 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1502 insn[1].code == 0) 1503 memcpy(aux, insn, sizeof(aux)); 1504 1505 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1506 clone->aux->verifier_zext); 1507 if (!rewritten) 1508 continue; 1509 1510 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1511 if (IS_ERR(tmp)) { 1512 /* Patching may have repointed aux->prog during 1513 * realloc from the original one, so we need to 1514 * fix it up here on error. 1515 */ 1516 bpf_jit_prog_release_other(prog, clone); 1517 return tmp; 1518 } 1519 1520 clone = tmp; 1521 insn_delta = rewritten - 1; 1522 1523 /* Walk new program and skip insns we just inserted. */ 1524 insn = clone->insnsi + i + insn_delta; 1525 insn_cnt += insn_delta; 1526 i += insn_delta; 1527 } 1528 1529 clone->blinded = 1; 1530 return clone; 1531 } 1532 #endif /* CONFIG_BPF_JIT */ 1533 1534 /* Base function for offset calculation. Needs to go into .text section, 1535 * therefore keeping it non-static as well; will also be used by JITs 1536 * anyway later on, so do not let the compiler omit it. This also needs 1537 * to go into kallsyms for correlation from e.g. bpftool, so naming 1538 * must not change. 1539 */ 1540 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1541 { 1542 return 0; 1543 } 1544 EXPORT_SYMBOL_GPL(__bpf_call_base); 1545 1546 /* All UAPI available opcodes. */ 1547 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1548 /* 32 bit ALU operations. */ \ 1549 /* Register based. */ \ 1550 INSN_3(ALU, ADD, X), \ 1551 INSN_3(ALU, SUB, X), \ 1552 INSN_3(ALU, AND, X), \ 1553 INSN_3(ALU, OR, X), \ 1554 INSN_3(ALU, LSH, X), \ 1555 INSN_3(ALU, RSH, X), \ 1556 INSN_3(ALU, XOR, X), \ 1557 INSN_3(ALU, MUL, X), \ 1558 INSN_3(ALU, MOV, X), \ 1559 INSN_3(ALU, ARSH, X), \ 1560 INSN_3(ALU, DIV, X), \ 1561 INSN_3(ALU, MOD, X), \ 1562 INSN_2(ALU, NEG), \ 1563 INSN_3(ALU, END, TO_BE), \ 1564 INSN_3(ALU, END, TO_LE), \ 1565 /* Immediate based. */ \ 1566 INSN_3(ALU, ADD, K), \ 1567 INSN_3(ALU, SUB, K), \ 1568 INSN_3(ALU, AND, K), \ 1569 INSN_3(ALU, OR, K), \ 1570 INSN_3(ALU, LSH, K), \ 1571 INSN_3(ALU, RSH, K), \ 1572 INSN_3(ALU, XOR, K), \ 1573 INSN_3(ALU, MUL, K), \ 1574 INSN_3(ALU, MOV, K), \ 1575 INSN_3(ALU, ARSH, K), \ 1576 INSN_3(ALU, DIV, K), \ 1577 INSN_3(ALU, MOD, K), \ 1578 /* 64 bit ALU operations. */ \ 1579 /* Register based. */ \ 1580 INSN_3(ALU64, ADD, X), \ 1581 INSN_3(ALU64, SUB, X), \ 1582 INSN_3(ALU64, AND, X), \ 1583 INSN_3(ALU64, OR, X), \ 1584 INSN_3(ALU64, LSH, X), \ 1585 INSN_3(ALU64, RSH, X), \ 1586 INSN_3(ALU64, XOR, X), \ 1587 INSN_3(ALU64, MUL, X), \ 1588 INSN_3(ALU64, MOV, X), \ 1589 INSN_3(ALU64, ARSH, X), \ 1590 INSN_3(ALU64, DIV, X), \ 1591 INSN_3(ALU64, MOD, X), \ 1592 INSN_2(ALU64, NEG), \ 1593 INSN_3(ALU64, END, TO_LE), \ 1594 /* Immediate based. */ \ 1595 INSN_3(ALU64, ADD, K), \ 1596 INSN_3(ALU64, SUB, K), \ 1597 INSN_3(ALU64, AND, K), \ 1598 INSN_3(ALU64, OR, K), \ 1599 INSN_3(ALU64, LSH, K), \ 1600 INSN_3(ALU64, RSH, K), \ 1601 INSN_3(ALU64, XOR, K), \ 1602 INSN_3(ALU64, MUL, K), \ 1603 INSN_3(ALU64, MOV, K), \ 1604 INSN_3(ALU64, ARSH, K), \ 1605 INSN_3(ALU64, DIV, K), \ 1606 INSN_3(ALU64, MOD, K), \ 1607 /* Call instruction. */ \ 1608 INSN_2(JMP, CALL), \ 1609 /* Exit instruction. */ \ 1610 INSN_2(JMP, EXIT), \ 1611 /* 32-bit Jump instructions. */ \ 1612 /* Register based. */ \ 1613 INSN_3(JMP32, JEQ, X), \ 1614 INSN_3(JMP32, JNE, X), \ 1615 INSN_3(JMP32, JGT, X), \ 1616 INSN_3(JMP32, JLT, X), \ 1617 INSN_3(JMP32, JGE, X), \ 1618 INSN_3(JMP32, JLE, X), \ 1619 INSN_3(JMP32, JSGT, X), \ 1620 INSN_3(JMP32, JSLT, X), \ 1621 INSN_3(JMP32, JSGE, X), \ 1622 INSN_3(JMP32, JSLE, X), \ 1623 INSN_3(JMP32, JSET, X), \ 1624 /* Immediate based. */ \ 1625 INSN_3(JMP32, JEQ, K), \ 1626 INSN_3(JMP32, JNE, K), \ 1627 INSN_3(JMP32, JGT, K), \ 1628 INSN_3(JMP32, JLT, K), \ 1629 INSN_3(JMP32, JGE, K), \ 1630 INSN_3(JMP32, JLE, K), \ 1631 INSN_3(JMP32, JSGT, K), \ 1632 INSN_3(JMP32, JSLT, K), \ 1633 INSN_3(JMP32, JSGE, K), \ 1634 INSN_3(JMP32, JSLE, K), \ 1635 INSN_3(JMP32, JSET, K), \ 1636 /* Jump instructions. */ \ 1637 /* Register based. */ \ 1638 INSN_3(JMP, JEQ, X), \ 1639 INSN_3(JMP, JNE, X), \ 1640 INSN_3(JMP, JGT, X), \ 1641 INSN_3(JMP, JLT, X), \ 1642 INSN_3(JMP, JGE, X), \ 1643 INSN_3(JMP, JLE, X), \ 1644 INSN_3(JMP, JSGT, X), \ 1645 INSN_3(JMP, JSLT, X), \ 1646 INSN_3(JMP, JSGE, X), \ 1647 INSN_3(JMP, JSLE, X), \ 1648 INSN_3(JMP, JSET, X), \ 1649 /* Immediate based. */ \ 1650 INSN_3(JMP, JEQ, K), \ 1651 INSN_3(JMP, JNE, K), \ 1652 INSN_3(JMP, JGT, K), \ 1653 INSN_3(JMP, JLT, K), \ 1654 INSN_3(JMP, JGE, K), \ 1655 INSN_3(JMP, JLE, K), \ 1656 INSN_3(JMP, JSGT, K), \ 1657 INSN_3(JMP, JSLT, K), \ 1658 INSN_3(JMP, JSGE, K), \ 1659 INSN_3(JMP, JSLE, K), \ 1660 INSN_3(JMP, JSET, K), \ 1661 INSN_2(JMP, JA), \ 1662 INSN_2(JMP32, JA), \ 1663 /* Store instructions. */ \ 1664 /* Register based. */ \ 1665 INSN_3(STX, MEM, B), \ 1666 INSN_3(STX, MEM, H), \ 1667 INSN_3(STX, MEM, W), \ 1668 INSN_3(STX, MEM, DW), \ 1669 INSN_3(STX, ATOMIC, W), \ 1670 INSN_3(STX, ATOMIC, DW), \ 1671 /* Immediate based. */ \ 1672 INSN_3(ST, MEM, B), \ 1673 INSN_3(ST, MEM, H), \ 1674 INSN_3(ST, MEM, W), \ 1675 INSN_3(ST, MEM, DW), \ 1676 /* Load instructions. */ \ 1677 /* Register based. */ \ 1678 INSN_3(LDX, MEM, B), \ 1679 INSN_3(LDX, MEM, H), \ 1680 INSN_3(LDX, MEM, W), \ 1681 INSN_3(LDX, MEM, DW), \ 1682 INSN_3(LDX, MEMSX, B), \ 1683 INSN_3(LDX, MEMSX, H), \ 1684 INSN_3(LDX, MEMSX, W), \ 1685 /* Immediate based. */ \ 1686 INSN_3(LD, IMM, DW) 1687 1688 bool bpf_opcode_in_insntable(u8 code) 1689 { 1690 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1691 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1692 static const bool public_insntable[256] = { 1693 [0 ... 255] = false, 1694 /* Now overwrite non-defaults ... */ 1695 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1696 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1697 [BPF_LD | BPF_ABS | BPF_B] = true, 1698 [BPF_LD | BPF_ABS | BPF_H] = true, 1699 [BPF_LD | BPF_ABS | BPF_W] = true, 1700 [BPF_LD | BPF_IND | BPF_B] = true, 1701 [BPF_LD | BPF_IND | BPF_H] = true, 1702 [BPF_LD | BPF_IND | BPF_W] = true, 1703 [BPF_JMP | BPF_JCOND] = true, 1704 }; 1705 #undef BPF_INSN_3_TBL 1706 #undef BPF_INSN_2_TBL 1707 return public_insntable[code]; 1708 } 1709 1710 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1711 /** 1712 * ___bpf_prog_run - run eBPF program on a given context 1713 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1714 * @insn: is the array of eBPF instructions 1715 * 1716 * Decode and execute eBPF instructions. 1717 * 1718 * Return: whatever value is in %BPF_R0 at program exit 1719 */ 1720 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1721 { 1722 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1723 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1724 static const void * const jumptable[256] __annotate_jump_table = { 1725 [0 ... 255] = &&default_label, 1726 /* Now overwrite non-defaults ... */ 1727 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1728 /* Non-UAPI available opcodes. */ 1729 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1730 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1731 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1732 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1733 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1734 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1735 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1736 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1737 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1738 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1739 }; 1740 #undef BPF_INSN_3_LBL 1741 #undef BPF_INSN_2_LBL 1742 u32 tail_call_cnt = 0; 1743 1744 #define CONT ({ insn++; goto select_insn; }) 1745 #define CONT_JMP ({ insn++; goto select_insn; }) 1746 1747 select_insn: 1748 goto *jumptable[insn->code]; 1749 1750 /* Explicitly mask the register-based shift amounts with 63 or 31 1751 * to avoid undefined behavior. Normally this won't affect the 1752 * generated code, for example, in case of native 64 bit archs such 1753 * as x86-64 or arm64, the compiler is optimizing the AND away for 1754 * the interpreter. In case of JITs, each of the JIT backends compiles 1755 * the BPF shift operations to machine instructions which produce 1756 * implementation-defined results in such a case; the resulting 1757 * contents of the register may be arbitrary, but program behaviour 1758 * as a whole remains defined. In other words, in case of JIT backends, 1759 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1760 */ 1761 /* ALU (shifts) */ 1762 #define SHT(OPCODE, OP) \ 1763 ALU64_##OPCODE##_X: \ 1764 DST = DST OP (SRC & 63); \ 1765 CONT; \ 1766 ALU_##OPCODE##_X: \ 1767 DST = (u32) DST OP ((u32) SRC & 31); \ 1768 CONT; \ 1769 ALU64_##OPCODE##_K: \ 1770 DST = DST OP IMM; \ 1771 CONT; \ 1772 ALU_##OPCODE##_K: \ 1773 DST = (u32) DST OP (u32) IMM; \ 1774 CONT; 1775 /* ALU (rest) */ 1776 #define ALU(OPCODE, OP) \ 1777 ALU64_##OPCODE##_X: \ 1778 DST = DST OP SRC; \ 1779 CONT; \ 1780 ALU_##OPCODE##_X: \ 1781 DST = (u32) DST OP (u32) SRC; \ 1782 CONT; \ 1783 ALU64_##OPCODE##_K: \ 1784 DST = DST OP IMM; \ 1785 CONT; \ 1786 ALU_##OPCODE##_K: \ 1787 DST = (u32) DST OP (u32) IMM; \ 1788 CONT; 1789 ALU(ADD, +) 1790 ALU(SUB, -) 1791 ALU(AND, &) 1792 ALU(OR, |) 1793 ALU(XOR, ^) 1794 ALU(MUL, *) 1795 SHT(LSH, <<) 1796 SHT(RSH, >>) 1797 #undef SHT 1798 #undef ALU 1799 ALU_NEG: 1800 DST = (u32) -DST; 1801 CONT; 1802 ALU64_NEG: 1803 DST = -DST; 1804 CONT; 1805 ALU_MOV_X: 1806 switch (OFF) { 1807 case 0: 1808 DST = (u32) SRC; 1809 break; 1810 case 8: 1811 DST = (u32)(s8) SRC; 1812 break; 1813 case 16: 1814 DST = (u32)(s16) SRC; 1815 break; 1816 } 1817 CONT; 1818 ALU_MOV_K: 1819 DST = (u32) IMM; 1820 CONT; 1821 ALU64_MOV_X: 1822 switch (OFF) { 1823 case 0: 1824 DST = SRC; 1825 break; 1826 case 8: 1827 DST = (s8) SRC; 1828 break; 1829 case 16: 1830 DST = (s16) SRC; 1831 break; 1832 case 32: 1833 DST = (s32) SRC; 1834 break; 1835 } 1836 CONT; 1837 ALU64_MOV_K: 1838 DST = IMM; 1839 CONT; 1840 LD_IMM_DW: 1841 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1842 insn++; 1843 CONT; 1844 ALU_ARSH_X: 1845 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1846 CONT; 1847 ALU_ARSH_K: 1848 DST = (u64) (u32) (((s32) DST) >> IMM); 1849 CONT; 1850 ALU64_ARSH_X: 1851 (*(s64 *) &DST) >>= (SRC & 63); 1852 CONT; 1853 ALU64_ARSH_K: 1854 (*(s64 *) &DST) >>= IMM; 1855 CONT; 1856 ALU64_MOD_X: 1857 switch (OFF) { 1858 case 0: 1859 div64_u64_rem(DST, SRC, &AX); 1860 DST = AX; 1861 break; 1862 case 1: 1863 AX = div64_s64(DST, SRC); 1864 DST = DST - AX * SRC; 1865 break; 1866 } 1867 CONT; 1868 ALU_MOD_X: 1869 switch (OFF) { 1870 case 0: 1871 AX = (u32) DST; 1872 DST = do_div(AX, (u32) SRC); 1873 break; 1874 case 1: 1875 AX = abs((s32)DST); 1876 AX = do_div(AX, abs((s32)SRC)); 1877 if ((s32)DST < 0) 1878 DST = (u32)-AX; 1879 else 1880 DST = (u32)AX; 1881 break; 1882 } 1883 CONT; 1884 ALU64_MOD_K: 1885 switch (OFF) { 1886 case 0: 1887 div64_u64_rem(DST, IMM, &AX); 1888 DST = AX; 1889 break; 1890 case 1: 1891 AX = div64_s64(DST, IMM); 1892 DST = DST - AX * IMM; 1893 break; 1894 } 1895 CONT; 1896 ALU_MOD_K: 1897 switch (OFF) { 1898 case 0: 1899 AX = (u32) DST; 1900 DST = do_div(AX, (u32) IMM); 1901 break; 1902 case 1: 1903 AX = abs((s32)DST); 1904 AX = do_div(AX, abs((s32)IMM)); 1905 if ((s32)DST < 0) 1906 DST = (u32)-AX; 1907 else 1908 DST = (u32)AX; 1909 break; 1910 } 1911 CONT; 1912 ALU64_DIV_X: 1913 switch (OFF) { 1914 case 0: 1915 DST = div64_u64(DST, SRC); 1916 break; 1917 case 1: 1918 DST = div64_s64(DST, SRC); 1919 break; 1920 } 1921 CONT; 1922 ALU_DIV_X: 1923 switch (OFF) { 1924 case 0: 1925 AX = (u32) DST; 1926 do_div(AX, (u32) SRC); 1927 DST = (u32) AX; 1928 break; 1929 case 1: 1930 AX = abs((s32)DST); 1931 do_div(AX, abs((s32)SRC)); 1932 if (((s32)DST < 0) == ((s32)SRC < 0)) 1933 DST = (u32)AX; 1934 else 1935 DST = (u32)-AX; 1936 break; 1937 } 1938 CONT; 1939 ALU64_DIV_K: 1940 switch (OFF) { 1941 case 0: 1942 DST = div64_u64(DST, IMM); 1943 break; 1944 case 1: 1945 DST = div64_s64(DST, IMM); 1946 break; 1947 } 1948 CONT; 1949 ALU_DIV_K: 1950 switch (OFF) { 1951 case 0: 1952 AX = (u32) DST; 1953 do_div(AX, (u32) IMM); 1954 DST = (u32) AX; 1955 break; 1956 case 1: 1957 AX = abs((s32)DST); 1958 do_div(AX, abs((s32)IMM)); 1959 if (((s32)DST < 0) == ((s32)IMM < 0)) 1960 DST = (u32)AX; 1961 else 1962 DST = (u32)-AX; 1963 break; 1964 } 1965 CONT; 1966 ALU_END_TO_BE: 1967 switch (IMM) { 1968 case 16: 1969 DST = (__force u16) cpu_to_be16(DST); 1970 break; 1971 case 32: 1972 DST = (__force u32) cpu_to_be32(DST); 1973 break; 1974 case 64: 1975 DST = (__force u64) cpu_to_be64(DST); 1976 break; 1977 } 1978 CONT; 1979 ALU_END_TO_LE: 1980 switch (IMM) { 1981 case 16: 1982 DST = (__force u16) cpu_to_le16(DST); 1983 break; 1984 case 32: 1985 DST = (__force u32) cpu_to_le32(DST); 1986 break; 1987 case 64: 1988 DST = (__force u64) cpu_to_le64(DST); 1989 break; 1990 } 1991 CONT; 1992 ALU64_END_TO_LE: 1993 switch (IMM) { 1994 case 16: 1995 DST = (__force u16) __swab16(DST); 1996 break; 1997 case 32: 1998 DST = (__force u32) __swab32(DST); 1999 break; 2000 case 64: 2001 DST = (__force u64) __swab64(DST); 2002 break; 2003 } 2004 CONT; 2005 2006 /* CALL */ 2007 JMP_CALL: 2008 /* Function call scratches BPF_R1-BPF_R5 registers, 2009 * preserves BPF_R6-BPF_R9, and stores return value 2010 * into BPF_R0. 2011 */ 2012 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2013 BPF_R4, BPF_R5); 2014 CONT; 2015 2016 JMP_CALL_ARGS: 2017 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2018 BPF_R3, BPF_R4, 2019 BPF_R5, 2020 insn + insn->off + 1); 2021 CONT; 2022 2023 JMP_TAIL_CALL: { 2024 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2025 struct bpf_array *array = container_of(map, struct bpf_array, map); 2026 struct bpf_prog *prog; 2027 u32 index = BPF_R3; 2028 2029 if (unlikely(index >= array->map.max_entries)) 2030 goto out; 2031 2032 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2033 goto out; 2034 2035 tail_call_cnt++; 2036 2037 prog = READ_ONCE(array->ptrs[index]); 2038 if (!prog) 2039 goto out; 2040 2041 /* ARG1 at this point is guaranteed to point to CTX from 2042 * the verifier side due to the fact that the tail call is 2043 * handled like a helper, that is, bpf_tail_call_proto, 2044 * where arg1_type is ARG_PTR_TO_CTX. 2045 */ 2046 insn = prog->insnsi; 2047 goto select_insn; 2048 out: 2049 CONT; 2050 } 2051 JMP_JA: 2052 insn += insn->off; 2053 CONT; 2054 JMP32_JA: 2055 insn += insn->imm; 2056 CONT; 2057 JMP_EXIT: 2058 return BPF_R0; 2059 /* JMP */ 2060 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2061 JMP_##OPCODE##_X: \ 2062 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2063 insn += insn->off; \ 2064 CONT_JMP; \ 2065 } \ 2066 CONT; \ 2067 JMP32_##OPCODE##_X: \ 2068 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2069 insn += insn->off; \ 2070 CONT_JMP; \ 2071 } \ 2072 CONT; \ 2073 JMP_##OPCODE##_K: \ 2074 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2075 insn += insn->off; \ 2076 CONT_JMP; \ 2077 } \ 2078 CONT; \ 2079 JMP32_##OPCODE##_K: \ 2080 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2081 insn += insn->off; \ 2082 CONT_JMP; \ 2083 } \ 2084 CONT; 2085 COND_JMP(u, JEQ, ==) 2086 COND_JMP(u, JNE, !=) 2087 COND_JMP(u, JGT, >) 2088 COND_JMP(u, JLT, <) 2089 COND_JMP(u, JGE, >=) 2090 COND_JMP(u, JLE, <=) 2091 COND_JMP(u, JSET, &) 2092 COND_JMP(s, JSGT, >) 2093 COND_JMP(s, JSLT, <) 2094 COND_JMP(s, JSGE, >=) 2095 COND_JMP(s, JSLE, <=) 2096 #undef COND_JMP 2097 /* ST, STX and LDX*/ 2098 ST_NOSPEC: 2099 /* Speculation barrier for mitigating Speculative Store Bypass. 2100 * In case of arm64, we rely on the firmware mitigation as 2101 * controlled via the ssbd kernel parameter. Whenever the 2102 * mitigation is enabled, it works for all of the kernel code 2103 * with no need to provide any additional instructions here. 2104 * In case of x86, we use 'lfence' insn for mitigation. We 2105 * reuse preexisting logic from Spectre v1 mitigation that 2106 * happens to produce the required code on x86 for v4 as well. 2107 */ 2108 barrier_nospec(); 2109 CONT; 2110 #define LDST(SIZEOP, SIZE) \ 2111 STX_MEM_##SIZEOP: \ 2112 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2113 CONT; \ 2114 ST_MEM_##SIZEOP: \ 2115 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2116 CONT; \ 2117 LDX_MEM_##SIZEOP: \ 2118 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2119 CONT; \ 2120 LDX_PROBE_MEM_##SIZEOP: \ 2121 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2122 (const void *)(long) (SRC + insn->off)); \ 2123 DST = *((SIZE *)&DST); \ 2124 CONT; 2125 2126 LDST(B, u8) 2127 LDST(H, u16) 2128 LDST(W, u32) 2129 LDST(DW, u64) 2130 #undef LDST 2131 2132 #define LDSX(SIZEOP, SIZE) \ 2133 LDX_MEMSX_##SIZEOP: \ 2134 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2135 CONT; \ 2136 LDX_PROBE_MEMSX_##SIZEOP: \ 2137 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2138 (const void *)(long) (SRC + insn->off)); \ 2139 DST = *((SIZE *)&DST); \ 2140 CONT; 2141 2142 LDSX(B, s8) 2143 LDSX(H, s16) 2144 LDSX(W, s32) 2145 #undef LDSX 2146 2147 #define ATOMIC_ALU_OP(BOP, KOP) \ 2148 case BOP: \ 2149 if (BPF_SIZE(insn->code) == BPF_W) \ 2150 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2151 (DST + insn->off)); \ 2152 else \ 2153 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2154 (DST + insn->off)); \ 2155 break; \ 2156 case BOP | BPF_FETCH: \ 2157 if (BPF_SIZE(insn->code) == BPF_W) \ 2158 SRC = (u32) atomic_fetch_##KOP( \ 2159 (u32) SRC, \ 2160 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2161 else \ 2162 SRC = (u64) atomic64_fetch_##KOP( \ 2163 (u64) SRC, \ 2164 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2165 break; 2166 2167 STX_ATOMIC_DW: 2168 STX_ATOMIC_W: 2169 switch (IMM) { 2170 ATOMIC_ALU_OP(BPF_ADD, add) 2171 ATOMIC_ALU_OP(BPF_AND, and) 2172 ATOMIC_ALU_OP(BPF_OR, or) 2173 ATOMIC_ALU_OP(BPF_XOR, xor) 2174 #undef ATOMIC_ALU_OP 2175 2176 case BPF_XCHG: 2177 if (BPF_SIZE(insn->code) == BPF_W) 2178 SRC = (u32) atomic_xchg( 2179 (atomic_t *)(unsigned long) (DST + insn->off), 2180 (u32) SRC); 2181 else 2182 SRC = (u64) atomic64_xchg( 2183 (atomic64_t *)(unsigned long) (DST + insn->off), 2184 (u64) SRC); 2185 break; 2186 case BPF_CMPXCHG: 2187 if (BPF_SIZE(insn->code) == BPF_W) 2188 BPF_R0 = (u32) atomic_cmpxchg( 2189 (atomic_t *)(unsigned long) (DST + insn->off), 2190 (u32) BPF_R0, (u32) SRC); 2191 else 2192 BPF_R0 = (u64) atomic64_cmpxchg( 2193 (atomic64_t *)(unsigned long) (DST + insn->off), 2194 (u64) BPF_R0, (u64) SRC); 2195 break; 2196 2197 default: 2198 goto default_label; 2199 } 2200 CONT; 2201 2202 default_label: 2203 /* If we ever reach this, we have a bug somewhere. Die hard here 2204 * instead of just returning 0; we could be somewhere in a subprog, 2205 * so execution could continue otherwise which we do /not/ want. 2206 * 2207 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2208 */ 2209 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2210 insn->code, insn->imm); 2211 BUG_ON(1); 2212 return 0; 2213 } 2214 2215 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2216 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2217 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2218 { \ 2219 u64 stack[stack_size / sizeof(u64)]; \ 2220 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2221 \ 2222 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2223 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2224 ARG1 = (u64) (unsigned long) ctx; \ 2225 return ___bpf_prog_run(regs, insn); \ 2226 } 2227 2228 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2229 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2230 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2231 const struct bpf_insn *insn) \ 2232 { \ 2233 u64 stack[stack_size / sizeof(u64)]; \ 2234 u64 regs[MAX_BPF_EXT_REG]; \ 2235 \ 2236 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2237 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2238 BPF_R1 = r1; \ 2239 BPF_R2 = r2; \ 2240 BPF_R3 = r3; \ 2241 BPF_R4 = r4; \ 2242 BPF_R5 = r5; \ 2243 return ___bpf_prog_run(regs, insn); \ 2244 } 2245 2246 #define EVAL1(FN, X) FN(X) 2247 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2248 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2249 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2250 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2251 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2252 2253 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2254 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2255 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2256 2257 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2258 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2259 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2260 2261 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2262 2263 static unsigned int (*interpreters[])(const void *ctx, 2264 const struct bpf_insn *insn) = { 2265 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2266 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2267 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2268 }; 2269 #undef PROG_NAME_LIST 2270 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2271 static __maybe_unused 2272 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2273 const struct bpf_insn *insn) = { 2274 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2275 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2276 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2277 }; 2278 #undef PROG_NAME_LIST 2279 2280 #ifdef CONFIG_BPF_SYSCALL 2281 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2282 { 2283 stack_depth = max_t(u32, stack_depth, 1); 2284 insn->off = (s16) insn->imm; 2285 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2286 __bpf_call_base_args; 2287 insn->code = BPF_JMP | BPF_CALL_ARGS; 2288 } 2289 #endif 2290 #else 2291 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2292 const struct bpf_insn *insn) 2293 { 2294 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2295 * is not working properly, so warn about it! 2296 */ 2297 WARN_ON_ONCE(1); 2298 return 0; 2299 } 2300 #endif 2301 2302 bool bpf_prog_map_compatible(struct bpf_map *map, 2303 const struct bpf_prog *fp) 2304 { 2305 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2306 bool ret; 2307 2308 if (fp->kprobe_override) 2309 return false; 2310 2311 /* XDP programs inserted into maps are not guaranteed to run on 2312 * a particular netdev (and can run outside driver context entirely 2313 * in the case of devmap and cpumap). Until device checks 2314 * are implemented, prohibit adding dev-bound programs to program maps. 2315 */ 2316 if (bpf_prog_is_dev_bound(fp->aux)) 2317 return false; 2318 2319 spin_lock(&map->owner.lock); 2320 if (!map->owner.type) { 2321 /* There's no owner yet where we could check for 2322 * compatibility. 2323 */ 2324 map->owner.type = prog_type; 2325 map->owner.jited = fp->jited; 2326 map->owner.xdp_has_frags = fp->aux->xdp_has_frags; 2327 ret = true; 2328 } else { 2329 ret = map->owner.type == prog_type && 2330 map->owner.jited == fp->jited && 2331 map->owner.xdp_has_frags == fp->aux->xdp_has_frags; 2332 } 2333 spin_unlock(&map->owner.lock); 2334 2335 return ret; 2336 } 2337 2338 static int bpf_check_tail_call(const struct bpf_prog *fp) 2339 { 2340 struct bpf_prog_aux *aux = fp->aux; 2341 int i, ret = 0; 2342 2343 mutex_lock(&aux->used_maps_mutex); 2344 for (i = 0; i < aux->used_map_cnt; i++) { 2345 struct bpf_map *map = aux->used_maps[i]; 2346 2347 if (!map_type_contains_progs(map)) 2348 continue; 2349 2350 if (!bpf_prog_map_compatible(map, fp)) { 2351 ret = -EINVAL; 2352 goto out; 2353 } 2354 } 2355 2356 out: 2357 mutex_unlock(&aux->used_maps_mutex); 2358 return ret; 2359 } 2360 2361 static void bpf_prog_select_func(struct bpf_prog *fp) 2362 { 2363 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2364 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2365 2366 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2367 #else 2368 fp->bpf_func = __bpf_prog_ret0_warn; 2369 #endif 2370 } 2371 2372 /** 2373 * bpf_prog_select_runtime - select exec runtime for BPF program 2374 * @fp: bpf_prog populated with BPF program 2375 * @err: pointer to error variable 2376 * 2377 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2378 * The BPF program will be executed via bpf_prog_run() function. 2379 * 2380 * Return: the &fp argument along with &err set to 0 for success or 2381 * a negative errno code on failure 2382 */ 2383 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2384 { 2385 /* In case of BPF to BPF calls, verifier did all the prep 2386 * work with regards to JITing, etc. 2387 */ 2388 bool jit_needed = false; 2389 2390 if (fp->bpf_func) 2391 goto finalize; 2392 2393 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2394 bpf_prog_has_kfunc_call(fp)) 2395 jit_needed = true; 2396 2397 bpf_prog_select_func(fp); 2398 2399 /* eBPF JITs can rewrite the program in case constant 2400 * blinding is active. However, in case of error during 2401 * blinding, bpf_int_jit_compile() must always return a 2402 * valid program, which in this case would simply not 2403 * be JITed, but falls back to the interpreter. 2404 */ 2405 if (!bpf_prog_is_offloaded(fp->aux)) { 2406 *err = bpf_prog_alloc_jited_linfo(fp); 2407 if (*err) 2408 return fp; 2409 2410 fp = bpf_int_jit_compile(fp); 2411 bpf_prog_jit_attempt_done(fp); 2412 if (!fp->jited && jit_needed) { 2413 *err = -ENOTSUPP; 2414 return fp; 2415 } 2416 } else { 2417 *err = bpf_prog_offload_compile(fp); 2418 if (*err) 2419 return fp; 2420 } 2421 2422 finalize: 2423 *err = bpf_prog_lock_ro(fp); 2424 if (*err) 2425 return fp; 2426 2427 /* The tail call compatibility check can only be done at 2428 * this late stage as we need to determine, if we deal 2429 * with JITed or non JITed program concatenations and not 2430 * all eBPF JITs might immediately support all features. 2431 */ 2432 *err = bpf_check_tail_call(fp); 2433 2434 return fp; 2435 } 2436 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2437 2438 static unsigned int __bpf_prog_ret1(const void *ctx, 2439 const struct bpf_insn *insn) 2440 { 2441 return 1; 2442 } 2443 2444 static struct bpf_prog_dummy { 2445 struct bpf_prog prog; 2446 } dummy_bpf_prog = { 2447 .prog = { 2448 .bpf_func = __bpf_prog_ret1, 2449 }, 2450 }; 2451 2452 struct bpf_empty_prog_array bpf_empty_prog_array = { 2453 .null_prog = NULL, 2454 }; 2455 EXPORT_SYMBOL(bpf_empty_prog_array); 2456 2457 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2458 { 2459 struct bpf_prog_array *p; 2460 2461 if (prog_cnt) 2462 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags); 2463 else 2464 p = &bpf_empty_prog_array.hdr; 2465 2466 return p; 2467 } 2468 2469 void bpf_prog_array_free(struct bpf_prog_array *progs) 2470 { 2471 if (!progs || progs == &bpf_empty_prog_array.hdr) 2472 return; 2473 kfree_rcu(progs, rcu); 2474 } 2475 2476 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2477 { 2478 struct bpf_prog_array *progs; 2479 2480 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2481 * no need to call kfree_rcu(), just call kfree() directly. 2482 */ 2483 progs = container_of(rcu, struct bpf_prog_array, rcu); 2484 if (rcu_trace_implies_rcu_gp()) 2485 kfree(progs); 2486 else 2487 kfree_rcu(progs, rcu); 2488 } 2489 2490 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2491 { 2492 if (!progs || progs == &bpf_empty_prog_array.hdr) 2493 return; 2494 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2495 } 2496 2497 int bpf_prog_array_length(struct bpf_prog_array *array) 2498 { 2499 struct bpf_prog_array_item *item; 2500 u32 cnt = 0; 2501 2502 for (item = array->items; item->prog; item++) 2503 if (item->prog != &dummy_bpf_prog.prog) 2504 cnt++; 2505 return cnt; 2506 } 2507 2508 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2509 { 2510 struct bpf_prog_array_item *item; 2511 2512 for (item = array->items; item->prog; item++) 2513 if (item->prog != &dummy_bpf_prog.prog) 2514 return false; 2515 return true; 2516 } 2517 2518 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2519 u32 *prog_ids, 2520 u32 request_cnt) 2521 { 2522 struct bpf_prog_array_item *item; 2523 int i = 0; 2524 2525 for (item = array->items; item->prog; item++) { 2526 if (item->prog == &dummy_bpf_prog.prog) 2527 continue; 2528 prog_ids[i] = item->prog->aux->id; 2529 if (++i == request_cnt) { 2530 item++; 2531 break; 2532 } 2533 } 2534 2535 return !!(item->prog); 2536 } 2537 2538 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2539 __u32 __user *prog_ids, u32 cnt) 2540 { 2541 unsigned long err = 0; 2542 bool nospc; 2543 u32 *ids; 2544 2545 /* users of this function are doing: 2546 * cnt = bpf_prog_array_length(); 2547 * if (cnt > 0) 2548 * bpf_prog_array_copy_to_user(..., cnt); 2549 * so below kcalloc doesn't need extra cnt > 0 check. 2550 */ 2551 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2552 if (!ids) 2553 return -ENOMEM; 2554 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2555 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2556 kfree(ids); 2557 if (err) 2558 return -EFAULT; 2559 if (nospc) 2560 return -ENOSPC; 2561 return 0; 2562 } 2563 2564 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2565 struct bpf_prog *old_prog) 2566 { 2567 struct bpf_prog_array_item *item; 2568 2569 for (item = array->items; item->prog; item++) 2570 if (item->prog == old_prog) { 2571 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2572 break; 2573 } 2574 } 2575 2576 /** 2577 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2578 * index into the program array with 2579 * a dummy no-op program. 2580 * @array: a bpf_prog_array 2581 * @index: the index of the program to replace 2582 * 2583 * Skips over dummy programs, by not counting them, when calculating 2584 * the position of the program to replace. 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_delete_safe_at(struct bpf_prog_array *array, int index) 2592 { 2593 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2594 } 2595 2596 /** 2597 * bpf_prog_array_update_at() - Updates the program at the given index 2598 * into the program array. 2599 * @array: a bpf_prog_array 2600 * @index: the index of the program to update 2601 * @prog: the program to insert into the array 2602 * 2603 * Skips over dummy programs, by not counting them, when calculating 2604 * the position of the program to update. 2605 * 2606 * Return: 2607 * * 0 - Success 2608 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2609 * * -ENOENT - Index out of range 2610 */ 2611 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2612 struct bpf_prog *prog) 2613 { 2614 struct bpf_prog_array_item *item; 2615 2616 if (unlikely(index < 0)) 2617 return -EINVAL; 2618 2619 for (item = array->items; item->prog; item++) { 2620 if (item->prog == &dummy_bpf_prog.prog) 2621 continue; 2622 if (!index) { 2623 WRITE_ONCE(item->prog, prog); 2624 return 0; 2625 } 2626 index--; 2627 } 2628 return -ENOENT; 2629 } 2630 2631 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2632 struct bpf_prog *exclude_prog, 2633 struct bpf_prog *include_prog, 2634 u64 bpf_cookie, 2635 struct bpf_prog_array **new_array) 2636 { 2637 int new_prog_cnt, carry_prog_cnt = 0; 2638 struct bpf_prog_array_item *existing, *new; 2639 struct bpf_prog_array *array; 2640 bool found_exclude = false; 2641 2642 /* Figure out how many existing progs we need to carry over to 2643 * the new array. 2644 */ 2645 if (old_array) { 2646 existing = old_array->items; 2647 for (; existing->prog; existing++) { 2648 if (existing->prog == exclude_prog) { 2649 found_exclude = true; 2650 continue; 2651 } 2652 if (existing->prog != &dummy_bpf_prog.prog) 2653 carry_prog_cnt++; 2654 if (existing->prog == include_prog) 2655 return -EEXIST; 2656 } 2657 } 2658 2659 if (exclude_prog && !found_exclude) 2660 return -ENOENT; 2661 2662 /* How many progs (not NULL) will be in the new array? */ 2663 new_prog_cnt = carry_prog_cnt; 2664 if (include_prog) 2665 new_prog_cnt += 1; 2666 2667 /* Do we have any prog (not NULL) in the new array? */ 2668 if (!new_prog_cnt) { 2669 *new_array = NULL; 2670 return 0; 2671 } 2672 2673 /* +1 as the end of prog_array is marked with NULL */ 2674 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2675 if (!array) 2676 return -ENOMEM; 2677 new = array->items; 2678 2679 /* Fill in the new prog array */ 2680 if (carry_prog_cnt) { 2681 existing = old_array->items; 2682 for (; existing->prog; existing++) { 2683 if (existing->prog == exclude_prog || 2684 existing->prog == &dummy_bpf_prog.prog) 2685 continue; 2686 2687 new->prog = existing->prog; 2688 new->bpf_cookie = existing->bpf_cookie; 2689 new++; 2690 } 2691 } 2692 if (include_prog) { 2693 new->prog = include_prog; 2694 new->bpf_cookie = bpf_cookie; 2695 new++; 2696 } 2697 new->prog = NULL; 2698 *new_array = array; 2699 return 0; 2700 } 2701 2702 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2703 u32 *prog_ids, u32 request_cnt, 2704 u32 *prog_cnt) 2705 { 2706 u32 cnt = 0; 2707 2708 if (array) 2709 cnt = bpf_prog_array_length(array); 2710 2711 *prog_cnt = cnt; 2712 2713 /* return early if user requested only program count or nothing to copy */ 2714 if (!request_cnt || !cnt) 2715 return 0; 2716 2717 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2718 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2719 : 0; 2720 } 2721 2722 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2723 struct bpf_map **used_maps, u32 len) 2724 { 2725 struct bpf_map *map; 2726 bool sleepable; 2727 u32 i; 2728 2729 sleepable = aux->prog->sleepable; 2730 for (i = 0; i < len; i++) { 2731 map = used_maps[i]; 2732 if (map->ops->map_poke_untrack) 2733 map->ops->map_poke_untrack(map, aux); 2734 if (sleepable) 2735 atomic64_dec(&map->sleepable_refcnt); 2736 bpf_map_put(map); 2737 } 2738 } 2739 2740 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2741 { 2742 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2743 kfree(aux->used_maps); 2744 } 2745 2746 void __bpf_free_used_btfs(struct bpf_prog_aux *aux, 2747 struct btf_mod_pair *used_btfs, u32 len) 2748 { 2749 #ifdef CONFIG_BPF_SYSCALL 2750 struct btf_mod_pair *btf_mod; 2751 u32 i; 2752 2753 for (i = 0; i < len; i++) { 2754 btf_mod = &used_btfs[i]; 2755 if (btf_mod->module) 2756 module_put(btf_mod->module); 2757 btf_put(btf_mod->btf); 2758 } 2759 #endif 2760 } 2761 2762 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2763 { 2764 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt); 2765 kfree(aux->used_btfs); 2766 } 2767 2768 static void bpf_prog_free_deferred(struct work_struct *work) 2769 { 2770 struct bpf_prog_aux *aux; 2771 int i; 2772 2773 aux = container_of(work, struct bpf_prog_aux, work); 2774 #ifdef CONFIG_BPF_SYSCALL 2775 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2776 #endif 2777 #ifdef CONFIG_CGROUP_BPF 2778 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2779 bpf_cgroup_atype_put(aux->cgroup_atype); 2780 #endif 2781 bpf_free_used_maps(aux); 2782 bpf_free_used_btfs(aux); 2783 if (bpf_prog_is_dev_bound(aux)) 2784 bpf_prog_dev_bound_destroy(aux->prog); 2785 #ifdef CONFIG_PERF_EVENTS 2786 if (aux->prog->has_callchain_buf) 2787 put_callchain_buffers(); 2788 #endif 2789 if (aux->dst_trampoline) 2790 bpf_trampoline_put(aux->dst_trampoline); 2791 for (i = 0; i < aux->real_func_cnt; i++) { 2792 /* We can just unlink the subprog poke descriptor table as 2793 * it was originally linked to the main program and is also 2794 * released along with it. 2795 */ 2796 aux->func[i]->aux->poke_tab = NULL; 2797 bpf_jit_free(aux->func[i]); 2798 } 2799 if (aux->real_func_cnt) { 2800 kfree(aux->func); 2801 bpf_prog_unlock_free(aux->prog); 2802 } else { 2803 bpf_jit_free(aux->prog); 2804 } 2805 } 2806 2807 void bpf_prog_free(struct bpf_prog *fp) 2808 { 2809 struct bpf_prog_aux *aux = fp->aux; 2810 2811 if (aux->dst_prog) 2812 bpf_prog_put(aux->dst_prog); 2813 bpf_token_put(aux->token); 2814 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2815 schedule_work(&aux->work); 2816 } 2817 EXPORT_SYMBOL_GPL(bpf_prog_free); 2818 2819 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 2820 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2821 2822 void bpf_user_rnd_init_once(void) 2823 { 2824 prandom_init_once(&bpf_user_rnd_state); 2825 } 2826 2827 BPF_CALL_0(bpf_user_rnd_u32) 2828 { 2829 /* Should someone ever have the rather unwise idea to use some 2830 * of the registers passed into this function, then note that 2831 * this function is called from native eBPF and classic-to-eBPF 2832 * transformations. Register assignments from both sides are 2833 * different, f.e. classic always sets fn(ctx, A, X) here. 2834 */ 2835 struct rnd_state *state; 2836 u32 res; 2837 2838 state = &get_cpu_var(bpf_user_rnd_state); 2839 res = prandom_u32_state(state); 2840 put_cpu_var(bpf_user_rnd_state); 2841 2842 return res; 2843 } 2844 2845 BPF_CALL_0(bpf_get_raw_cpu_id) 2846 { 2847 return raw_smp_processor_id(); 2848 } 2849 2850 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2851 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2852 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2853 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2854 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2855 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2856 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2857 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2858 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2859 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2860 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2861 2862 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2863 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2864 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2865 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2866 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2867 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2868 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 2869 2870 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2871 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2872 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2873 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2874 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2875 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2876 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2877 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2878 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2879 const struct bpf_func_proto bpf_set_retval_proto __weak; 2880 const struct bpf_func_proto bpf_get_retval_proto __weak; 2881 2882 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2883 { 2884 return NULL; 2885 } 2886 2887 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2888 { 2889 return NULL; 2890 } 2891 2892 u64 __weak 2893 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2894 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2895 { 2896 return -ENOTSUPP; 2897 } 2898 EXPORT_SYMBOL_GPL(bpf_event_output); 2899 2900 /* Always built-in helper functions. */ 2901 const struct bpf_func_proto bpf_tail_call_proto = { 2902 .func = NULL, 2903 .gpl_only = false, 2904 .ret_type = RET_VOID, 2905 .arg1_type = ARG_PTR_TO_CTX, 2906 .arg2_type = ARG_CONST_MAP_PTR, 2907 .arg3_type = ARG_ANYTHING, 2908 }; 2909 2910 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2911 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2912 * eBPF and implicitly also cBPF can get JITed! 2913 */ 2914 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2915 { 2916 return prog; 2917 } 2918 2919 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2920 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2921 */ 2922 void __weak bpf_jit_compile(struct bpf_prog *prog) 2923 { 2924 } 2925 2926 bool __weak bpf_helper_changes_pkt_data(void *func) 2927 { 2928 return false; 2929 } 2930 2931 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2932 * analysis code and wants explicit zero extension inserted by verifier. 2933 * Otherwise, return FALSE. 2934 * 2935 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2936 * you don't override this. JITs that don't want these extra insns can detect 2937 * them using insn_is_zext. 2938 */ 2939 bool __weak bpf_jit_needs_zext(void) 2940 { 2941 return false; 2942 } 2943 2944 /* Return true if the JIT inlines the call to the helper corresponding to 2945 * the imm. 2946 * 2947 * The verifier will not patch the insn->imm for the call to the helper if 2948 * this returns true. 2949 */ 2950 bool __weak bpf_jit_inlines_helper_call(s32 imm) 2951 { 2952 return false; 2953 } 2954 2955 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 2956 bool __weak bpf_jit_supports_subprog_tailcalls(void) 2957 { 2958 return false; 2959 } 2960 2961 bool __weak bpf_jit_supports_percpu_insn(void) 2962 { 2963 return false; 2964 } 2965 2966 bool __weak bpf_jit_supports_kfunc_call(void) 2967 { 2968 return false; 2969 } 2970 2971 bool __weak bpf_jit_supports_far_kfunc_call(void) 2972 { 2973 return false; 2974 } 2975 2976 bool __weak bpf_jit_supports_arena(void) 2977 { 2978 return false; 2979 } 2980 2981 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 2982 { 2983 return false; 2984 } 2985 2986 u64 __weak bpf_arch_uaddress_limit(void) 2987 { 2988 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 2989 return TASK_SIZE; 2990 #else 2991 return 0; 2992 #endif 2993 } 2994 2995 /* Return TRUE if the JIT backend satisfies the following two conditions: 2996 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 2997 * 2) Under the specific arch, the implementation of xchg() is the same 2998 * as atomic_xchg() on pointer-sized words. 2999 */ 3000 bool __weak bpf_jit_supports_ptr_xchg(void) 3001 { 3002 return false; 3003 } 3004 3005 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3006 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3007 */ 3008 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3009 int len) 3010 { 3011 return -EFAULT; 3012 } 3013 3014 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 3015 void *addr1, void *addr2) 3016 { 3017 return -ENOTSUPP; 3018 } 3019 3020 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3021 { 3022 return ERR_PTR(-ENOTSUPP); 3023 } 3024 3025 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3026 { 3027 return -ENOTSUPP; 3028 } 3029 3030 bool __weak bpf_jit_supports_exceptions(void) 3031 { 3032 return false; 3033 } 3034 3035 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3036 { 3037 } 3038 3039 /* for configs without MMU or 32-bit */ 3040 __weak const struct bpf_map_ops arena_map_ops; 3041 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3042 { 3043 return 0; 3044 } 3045 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3046 { 3047 return 0; 3048 } 3049 3050 #ifdef CONFIG_BPF_SYSCALL 3051 static int __init bpf_global_ma_init(void) 3052 { 3053 int ret; 3054 3055 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3056 bpf_global_ma_set = !ret; 3057 return ret; 3058 } 3059 late_initcall(bpf_global_ma_init); 3060 #endif 3061 3062 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3063 EXPORT_SYMBOL(bpf_stats_enabled_key); 3064 3065 /* All definitions of tracepoints related to BPF. */ 3066 #define CREATE_TRACE_POINTS 3067 #include <linux/bpf_trace.h> 3068 3069 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3070 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3071