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/prandom.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 <linux/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->ext_mutex); 135 mutex_init(&fp->aux->dst_mutex); 136 137 return fp; 138 } 139 140 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 141 { 142 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 143 struct bpf_prog *prog; 144 int cpu; 145 146 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 147 if (!prog) 148 return NULL; 149 150 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 151 if (!prog->stats) { 152 free_percpu(prog->active); 153 kfree(prog->aux); 154 vfree(prog); 155 return NULL; 156 } 157 158 for_each_possible_cpu(cpu) { 159 struct bpf_prog_stats *pstats; 160 161 pstats = per_cpu_ptr(prog->stats, cpu); 162 u64_stats_init(&pstats->syncp); 163 } 164 return prog; 165 } 166 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 167 168 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 169 { 170 if (!prog->aux->nr_linfo || !prog->jit_requested) 171 return 0; 172 173 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 174 sizeof(*prog->aux->jited_linfo), 175 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 176 if (!prog->aux->jited_linfo) 177 return -ENOMEM; 178 179 return 0; 180 } 181 182 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 183 { 184 if (prog->aux->jited_linfo && 185 (!prog->jited || !prog->aux->jited_linfo[0])) { 186 kvfree(prog->aux->jited_linfo); 187 prog->aux->jited_linfo = NULL; 188 } 189 190 kfree(prog->aux->kfunc_tab); 191 prog->aux->kfunc_tab = NULL; 192 } 193 194 /* The jit engine is responsible to provide an array 195 * for insn_off to the jited_off mapping (insn_to_jit_off). 196 * 197 * The idx to this array is the insn_off. Hence, the insn_off 198 * here is relative to the prog itself instead of the main prog. 199 * This array has one entry for each xlated bpf insn. 200 * 201 * jited_off is the byte off to the end of the jited insn. 202 * 203 * Hence, with 204 * insn_start: 205 * The first bpf insn off of the prog. The insn off 206 * here is relative to the main prog. 207 * e.g. if prog is a subprog, insn_start > 0 208 * linfo_idx: 209 * The prog's idx to prog->aux->linfo and jited_linfo 210 * 211 * jited_linfo[linfo_idx] = prog->bpf_func 212 * 213 * For i > linfo_idx, 214 * 215 * jited_linfo[i] = prog->bpf_func + 216 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 217 */ 218 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 219 const u32 *insn_to_jit_off) 220 { 221 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 222 const struct bpf_line_info *linfo; 223 void **jited_linfo; 224 225 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 226 /* Userspace did not provide linfo */ 227 return; 228 229 linfo_idx = prog->aux->linfo_idx; 230 linfo = &prog->aux->linfo[linfo_idx]; 231 insn_start = linfo[0].insn_off; 232 insn_end = insn_start + prog->len; 233 234 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 235 jited_linfo[0] = prog->bpf_func; 236 237 nr_linfo = prog->aux->nr_linfo - linfo_idx; 238 239 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 240 /* The verifier ensures that linfo[i].insn_off is 241 * strictly increasing 242 */ 243 jited_linfo[i] = prog->bpf_func + 244 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 245 } 246 247 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 248 gfp_t gfp_extra_flags) 249 { 250 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 251 struct bpf_prog *fp; 252 u32 pages; 253 254 size = round_up(size, PAGE_SIZE); 255 pages = size / PAGE_SIZE; 256 if (pages <= fp_old->pages) 257 return fp_old; 258 259 fp = __vmalloc(size, gfp_flags); 260 if (fp) { 261 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 262 fp->pages = pages; 263 fp->aux->prog = fp; 264 265 /* We keep fp->aux from fp_old around in the new 266 * reallocated structure. 267 */ 268 fp_old->aux = NULL; 269 fp_old->stats = NULL; 270 fp_old->active = NULL; 271 __bpf_prog_free(fp_old); 272 } 273 274 return fp; 275 } 276 277 void __bpf_prog_free(struct bpf_prog *fp) 278 { 279 if (fp->aux) { 280 mutex_destroy(&fp->aux->used_maps_mutex); 281 mutex_destroy(&fp->aux->dst_mutex); 282 kfree(fp->aux->poke_tab); 283 kfree(fp->aux); 284 } 285 free_percpu(fp->stats); 286 free_percpu(fp->active); 287 vfree(fp); 288 } 289 290 int bpf_prog_calc_tag(struct bpf_prog *fp) 291 { 292 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64); 293 u32 raw_size = bpf_prog_tag_scratch_size(fp); 294 u32 digest[SHA1_DIGEST_WORDS]; 295 u32 ws[SHA1_WORKSPACE_WORDS]; 296 u32 i, bsize, psize, blocks; 297 struct bpf_insn *dst; 298 bool was_ld_map; 299 u8 *raw, *todo; 300 __be32 *result; 301 __be64 *bits; 302 303 raw = vmalloc(raw_size); 304 if (!raw) 305 return -ENOMEM; 306 307 sha1_init(digest); 308 memset(ws, 0, sizeof(ws)); 309 310 /* We need to take out the map fd for the digest calculation 311 * since they are unstable from user space side. 312 */ 313 dst = (void *)raw; 314 for (i = 0, was_ld_map = false; i < fp->len; i++) { 315 dst[i] = fp->insnsi[i]; 316 if (!was_ld_map && 317 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 318 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 319 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 320 was_ld_map = true; 321 dst[i].imm = 0; 322 } else if (was_ld_map && 323 dst[i].code == 0 && 324 dst[i].dst_reg == 0 && 325 dst[i].src_reg == 0 && 326 dst[i].off == 0) { 327 was_ld_map = false; 328 dst[i].imm = 0; 329 } else { 330 was_ld_map = false; 331 } 332 } 333 334 psize = bpf_prog_insn_size(fp); 335 memset(&raw[psize], 0, raw_size - psize); 336 raw[psize++] = 0x80; 337 338 bsize = round_up(psize, SHA1_BLOCK_SIZE); 339 blocks = bsize / SHA1_BLOCK_SIZE; 340 todo = raw; 341 if (bsize - psize >= sizeof(__be64)) { 342 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 343 } else { 344 bits = (__be64 *)(todo + bsize + bits_offset); 345 blocks++; 346 } 347 *bits = cpu_to_be64((psize - 1) << 3); 348 349 while (blocks--) { 350 sha1_transform(digest, todo, ws); 351 todo += SHA1_BLOCK_SIZE; 352 } 353 354 result = (__force __be32 *)digest; 355 for (i = 0; i < SHA1_DIGEST_WORDS; i++) 356 result[i] = cpu_to_be32(digest[i]); 357 memcpy(fp->tag, result, sizeof(fp->tag)); 358 359 vfree(raw); 360 return 0; 361 } 362 363 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 364 s32 end_new, s32 curr, const bool probe_pass) 365 { 366 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 367 s32 delta = end_new - end_old; 368 s64 imm = insn->imm; 369 370 if (curr < pos && curr + imm + 1 >= end_old) 371 imm += delta; 372 else if (curr >= end_new && curr + imm + 1 < end_new) 373 imm -= delta; 374 if (imm < imm_min || imm > imm_max) 375 return -ERANGE; 376 if (!probe_pass) 377 insn->imm = imm; 378 return 0; 379 } 380 381 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 382 s32 end_new, s32 curr, const bool probe_pass) 383 { 384 s64 off_min, off_max, off; 385 s32 delta = end_new - end_old; 386 387 if (insn->code == (BPF_JMP32 | BPF_JA)) { 388 off = insn->imm; 389 off_min = S32_MIN; 390 off_max = S32_MAX; 391 } else { 392 off = insn->off; 393 off_min = S16_MIN; 394 off_max = S16_MAX; 395 } 396 397 if (curr < pos && curr + off + 1 >= end_old) 398 off += delta; 399 else if (curr >= end_new && curr + off + 1 < end_new) 400 off -= delta; 401 if (off < off_min || off > off_max) 402 return -ERANGE; 403 if (!probe_pass) { 404 if (insn->code == (BPF_JMP32 | BPF_JA)) 405 insn->imm = off; 406 else 407 insn->off = off; 408 } 409 return 0; 410 } 411 412 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 413 s32 end_new, const bool probe_pass) 414 { 415 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 416 struct bpf_insn *insn = prog->insnsi; 417 int ret = 0; 418 419 for (i = 0; i < insn_cnt; i++, insn++) { 420 u8 code; 421 422 /* In the probing pass we still operate on the original, 423 * unpatched image in order to check overflows before we 424 * do any other adjustments. Therefore skip the patchlet. 425 */ 426 if (probe_pass && i == pos) { 427 i = end_new; 428 insn = prog->insnsi + end_old; 429 } 430 if (bpf_pseudo_func(insn)) { 431 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 432 end_new, i, probe_pass); 433 if (ret) 434 return ret; 435 continue; 436 } 437 code = insn->code; 438 if ((BPF_CLASS(code) != BPF_JMP && 439 BPF_CLASS(code) != BPF_JMP32) || 440 BPF_OP(code) == BPF_EXIT) 441 continue; 442 /* Adjust offset of jmps if we cross patch boundaries. */ 443 if (BPF_OP(code) == BPF_CALL) { 444 if (insn->src_reg != BPF_PSEUDO_CALL) 445 continue; 446 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 447 end_new, i, probe_pass); 448 } else { 449 ret = bpf_adj_delta_to_off(insn, pos, end_old, 450 end_new, i, probe_pass); 451 } 452 if (ret) 453 break; 454 } 455 456 return ret; 457 } 458 459 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 460 { 461 struct bpf_line_info *linfo; 462 u32 i, nr_linfo; 463 464 nr_linfo = prog->aux->nr_linfo; 465 if (!nr_linfo || !delta) 466 return; 467 468 linfo = prog->aux->linfo; 469 470 for (i = 0; i < nr_linfo; i++) 471 if (off < linfo[i].insn_off) 472 break; 473 474 /* Push all off < linfo[i].insn_off by delta */ 475 for (; i < nr_linfo; i++) 476 linfo[i].insn_off += delta; 477 } 478 479 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 480 const struct bpf_insn *patch, u32 len) 481 { 482 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 483 const u32 cnt_max = S16_MAX; 484 struct bpf_prog *prog_adj; 485 int err; 486 487 /* Since our patchlet doesn't expand the image, we're done. */ 488 if (insn_delta == 0) { 489 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 490 return prog; 491 } 492 493 insn_adj_cnt = prog->len + insn_delta; 494 495 /* Reject anything that would potentially let the insn->off 496 * target overflow when we have excessive program expansions. 497 * We need to probe here before we do any reallocation where 498 * we afterwards may not fail anymore. 499 */ 500 if (insn_adj_cnt > cnt_max && 501 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 502 return ERR_PTR(err); 503 504 /* Several new instructions need to be inserted. Make room 505 * for them. Likely, there's no need for a new allocation as 506 * last page could have large enough tailroom. 507 */ 508 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 509 GFP_USER); 510 if (!prog_adj) 511 return ERR_PTR(-ENOMEM); 512 513 prog_adj->len = insn_adj_cnt; 514 515 /* Patching happens in 3 steps: 516 * 517 * 1) Move over tail of insnsi from next instruction onwards, 518 * so we can patch the single target insn with one or more 519 * new ones (patching is always from 1 to n insns, n > 0). 520 * 2) Inject new instructions at the target location. 521 * 3) Adjust branch offsets if necessary. 522 */ 523 insn_rest = insn_adj_cnt - off - len; 524 525 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 526 sizeof(*patch) * insn_rest); 527 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 528 529 /* We are guaranteed to not fail at this point, otherwise 530 * the ship has sailed to reverse to the original state. An 531 * overflow cannot happen at this point. 532 */ 533 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 534 535 bpf_adj_linfo(prog_adj, off, insn_delta); 536 537 return prog_adj; 538 } 539 540 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 541 { 542 int err; 543 544 /* Branch offsets can't overflow when program is shrinking, no need 545 * to call bpf_adj_branches(..., true) here 546 */ 547 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 548 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 549 prog->len -= cnt; 550 551 err = bpf_adj_branches(prog, off, off + cnt, off, false); 552 WARN_ON_ONCE(err); 553 return err; 554 } 555 556 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 557 { 558 int i; 559 560 for (i = 0; i < fp->aux->real_func_cnt; i++) 561 bpf_prog_kallsyms_del(fp->aux->func[i]); 562 } 563 564 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 565 { 566 bpf_prog_kallsyms_del_subprogs(fp); 567 bpf_prog_kallsyms_del(fp); 568 } 569 570 #ifdef CONFIG_BPF_JIT 571 /* All BPF JIT sysctl knobs here. */ 572 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 573 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 574 int bpf_jit_harden __read_mostly; 575 long bpf_jit_limit __read_mostly; 576 long bpf_jit_limit_max __read_mostly; 577 578 static void 579 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 580 { 581 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 582 583 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 584 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 585 } 586 587 static void 588 bpf_prog_ksym_set_name(struct bpf_prog *prog) 589 { 590 char *sym = prog->aux->ksym.name; 591 const char *end = sym + KSYM_NAME_LEN; 592 const struct btf_type *type; 593 const char *func_name; 594 595 BUILD_BUG_ON(sizeof("bpf_prog_") + 596 sizeof(prog->tag) * 2 + 597 /* name has been null terminated. 598 * We should need +1 for the '_' preceding 599 * the name. However, the null character 600 * is double counted between the name and the 601 * sizeof("bpf_prog_") above, so we omit 602 * the +1 here. 603 */ 604 sizeof(prog->aux->name) > KSYM_NAME_LEN); 605 606 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 607 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 608 609 /* prog->aux->name will be ignored if full btf name is available */ 610 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 611 type = btf_type_by_id(prog->aux->btf, 612 prog->aux->func_info[prog->aux->func_idx].type_id); 613 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 614 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 615 return; 616 } 617 618 if (prog->aux->name[0]) 619 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 620 else 621 *sym = 0; 622 } 623 624 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 625 { 626 return container_of(n, struct bpf_ksym, tnode)->start; 627 } 628 629 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 630 struct latch_tree_node *b) 631 { 632 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 633 } 634 635 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 636 { 637 unsigned long val = (unsigned long)key; 638 const struct bpf_ksym *ksym; 639 640 ksym = container_of(n, struct bpf_ksym, tnode); 641 642 if (val < ksym->start) 643 return -1; 644 /* Ensure that we detect return addresses as part of the program, when 645 * the final instruction is a call for a program part of the stack 646 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 647 */ 648 if (val > ksym->end) 649 return 1; 650 651 return 0; 652 } 653 654 static const struct latch_tree_ops bpf_tree_ops = { 655 .less = bpf_tree_less, 656 .comp = bpf_tree_comp, 657 }; 658 659 static DEFINE_SPINLOCK(bpf_lock); 660 static LIST_HEAD(bpf_kallsyms); 661 static struct latch_tree_root bpf_tree __cacheline_aligned; 662 663 void bpf_ksym_add(struct bpf_ksym *ksym) 664 { 665 spin_lock_bh(&bpf_lock); 666 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 667 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 668 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 669 spin_unlock_bh(&bpf_lock); 670 } 671 672 static void __bpf_ksym_del(struct bpf_ksym *ksym) 673 { 674 if (list_empty(&ksym->lnode)) 675 return; 676 677 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 678 list_del_rcu(&ksym->lnode); 679 } 680 681 void bpf_ksym_del(struct bpf_ksym *ksym) 682 { 683 spin_lock_bh(&bpf_lock); 684 __bpf_ksym_del(ksym); 685 spin_unlock_bh(&bpf_lock); 686 } 687 688 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 689 { 690 return fp->jited && !bpf_prog_was_classic(fp); 691 } 692 693 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 694 { 695 if (!bpf_prog_kallsyms_candidate(fp) || 696 !bpf_token_capable(fp->aux->token, CAP_BPF)) 697 return; 698 699 bpf_prog_ksym_set_addr(fp); 700 bpf_prog_ksym_set_name(fp); 701 fp->aux->ksym.prog = true; 702 703 bpf_ksym_add(&fp->aux->ksym); 704 705 #ifdef CONFIG_FINEIBT 706 /* 707 * When FineIBT, code in the __cfi_foo() symbols can get executed 708 * and hence unwinder needs help. 709 */ 710 if (cfi_mode != CFI_FINEIBT) 711 return; 712 713 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 714 "__cfi_%s", fp->aux->ksym.name); 715 716 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 717 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 718 719 bpf_ksym_add(&fp->aux->ksym_prefix); 720 #endif 721 } 722 723 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 724 { 725 if (!bpf_prog_kallsyms_candidate(fp)) 726 return; 727 728 bpf_ksym_del(&fp->aux->ksym); 729 #ifdef CONFIG_FINEIBT 730 if (cfi_mode != CFI_FINEIBT) 731 return; 732 bpf_ksym_del(&fp->aux->ksym_prefix); 733 #endif 734 } 735 736 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 737 { 738 struct latch_tree_node *n; 739 740 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 741 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 742 } 743 744 int __bpf_address_lookup(unsigned long addr, unsigned long *size, 745 unsigned long *off, char *sym) 746 { 747 struct bpf_ksym *ksym; 748 int ret = 0; 749 750 rcu_read_lock(); 751 ksym = bpf_ksym_find(addr); 752 if (ksym) { 753 unsigned long symbol_start = ksym->start; 754 unsigned long symbol_end = ksym->end; 755 756 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 757 758 if (size) 759 *size = symbol_end - symbol_start; 760 if (off) 761 *off = addr - symbol_start; 762 } 763 rcu_read_unlock(); 764 765 return ret; 766 } 767 768 bool is_bpf_text_address(unsigned long addr) 769 { 770 bool ret; 771 772 rcu_read_lock(); 773 ret = bpf_ksym_find(addr) != NULL; 774 rcu_read_unlock(); 775 776 return ret; 777 } 778 779 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 780 { 781 struct bpf_ksym *ksym = bpf_ksym_find(addr); 782 783 return ksym && ksym->prog ? 784 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 785 NULL; 786 } 787 788 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 789 { 790 const struct exception_table_entry *e = NULL; 791 struct bpf_prog *prog; 792 793 rcu_read_lock(); 794 prog = bpf_prog_ksym_find(addr); 795 if (!prog) 796 goto out; 797 if (!prog->aux->num_exentries) 798 goto out; 799 800 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 801 out: 802 rcu_read_unlock(); 803 return e; 804 } 805 806 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 807 char *sym) 808 { 809 struct bpf_ksym *ksym; 810 unsigned int it = 0; 811 int ret = -ERANGE; 812 813 if (!bpf_jit_kallsyms_enabled()) 814 return ret; 815 816 rcu_read_lock(); 817 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 818 if (it++ != symnum) 819 continue; 820 821 strscpy(sym, ksym->name, KSYM_NAME_LEN); 822 823 *value = ksym->start; 824 *type = BPF_SYM_ELF_TYPE; 825 826 ret = 0; 827 break; 828 } 829 rcu_read_unlock(); 830 831 return ret; 832 } 833 834 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 835 struct bpf_jit_poke_descriptor *poke) 836 { 837 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 838 static const u32 poke_tab_max = 1024; 839 u32 slot = prog->aux->size_poke_tab; 840 u32 size = slot + 1; 841 842 if (size > poke_tab_max) 843 return -ENOSPC; 844 if (poke->tailcall_target || poke->tailcall_target_stable || 845 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 846 return -EINVAL; 847 848 switch (poke->reason) { 849 case BPF_POKE_REASON_TAIL_CALL: 850 if (!poke->tail_call.map) 851 return -EINVAL; 852 break; 853 default: 854 return -EINVAL; 855 } 856 857 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 858 if (!tab) 859 return -ENOMEM; 860 861 memcpy(&tab[slot], poke, sizeof(*poke)); 862 prog->aux->size_poke_tab = size; 863 prog->aux->poke_tab = tab; 864 865 return slot; 866 } 867 868 /* 869 * BPF program pack allocator. 870 * 871 * Most BPF programs are pretty small. Allocating a hole page for each 872 * program is sometime a waste. Many small bpf program also adds pressure 873 * to instruction TLB. To solve this issue, we introduce a BPF program pack 874 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 875 * to host BPF programs. 876 */ 877 #define BPF_PROG_CHUNK_SHIFT 6 878 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 879 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 880 881 struct bpf_prog_pack { 882 struct list_head list; 883 void *ptr; 884 unsigned long bitmap[]; 885 }; 886 887 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 888 { 889 memset(area, 0, size); 890 } 891 892 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 893 894 static DEFINE_MUTEX(pack_mutex); 895 static LIST_HEAD(pack_list); 896 897 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 898 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 899 */ 900 #ifdef PMD_SIZE 901 /* PMD_SIZE is really big for some archs. It doesn't make sense to 902 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 903 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 904 * greater than or equal to 2MB. 905 */ 906 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 907 #else 908 #define BPF_PROG_PACK_SIZE PAGE_SIZE 909 #endif 910 911 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 912 913 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 914 { 915 struct bpf_prog_pack *pack; 916 int err; 917 918 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 919 GFP_KERNEL); 920 if (!pack) 921 return NULL; 922 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 923 if (!pack->ptr) 924 goto out; 925 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 926 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 927 928 set_vm_flush_reset_perms(pack->ptr); 929 err = set_memory_rox((unsigned long)pack->ptr, 930 BPF_PROG_PACK_SIZE / PAGE_SIZE); 931 if (err) 932 goto out; 933 list_add_tail(&pack->list, &pack_list); 934 return pack; 935 936 out: 937 bpf_jit_free_exec(pack->ptr); 938 kfree(pack); 939 return NULL; 940 } 941 942 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 943 { 944 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 945 struct bpf_prog_pack *pack; 946 unsigned long pos; 947 void *ptr = NULL; 948 949 mutex_lock(&pack_mutex); 950 if (size > BPF_PROG_PACK_SIZE) { 951 size = round_up(size, PAGE_SIZE); 952 ptr = bpf_jit_alloc_exec(size); 953 if (ptr) { 954 int err; 955 956 bpf_fill_ill_insns(ptr, size); 957 set_vm_flush_reset_perms(ptr); 958 err = set_memory_rox((unsigned long)ptr, 959 size / PAGE_SIZE); 960 if (err) { 961 bpf_jit_free_exec(ptr); 962 ptr = NULL; 963 } 964 } 965 goto out; 966 } 967 list_for_each_entry(pack, &pack_list, list) { 968 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 969 nbits, 0); 970 if (pos < BPF_PROG_CHUNK_COUNT) 971 goto found_free_area; 972 } 973 974 pack = alloc_new_pack(bpf_fill_ill_insns); 975 if (!pack) 976 goto out; 977 978 pos = 0; 979 980 found_free_area: 981 bitmap_set(pack->bitmap, pos, nbits); 982 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 983 984 out: 985 mutex_unlock(&pack_mutex); 986 return ptr; 987 } 988 989 void bpf_prog_pack_free(void *ptr, u32 size) 990 { 991 struct bpf_prog_pack *pack = NULL, *tmp; 992 unsigned int nbits; 993 unsigned long pos; 994 995 mutex_lock(&pack_mutex); 996 if (size > BPF_PROG_PACK_SIZE) { 997 bpf_jit_free_exec(ptr); 998 goto out; 999 } 1000 1001 list_for_each_entry(tmp, &pack_list, list) { 1002 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 1003 pack = tmp; 1004 break; 1005 } 1006 } 1007 1008 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1009 goto out; 1010 1011 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1012 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1013 1014 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1015 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1016 1017 bitmap_clear(pack->bitmap, pos, nbits); 1018 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1019 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1020 list_del(&pack->list); 1021 bpf_jit_free_exec(pack->ptr); 1022 kfree(pack); 1023 } 1024 out: 1025 mutex_unlock(&pack_mutex); 1026 } 1027 1028 static atomic_long_t bpf_jit_current; 1029 1030 /* Can be overridden by an arch's JIT compiler if it has a custom, 1031 * dedicated BPF backend memory area, or if neither of the two 1032 * below apply. 1033 */ 1034 u64 __weak bpf_jit_alloc_exec_limit(void) 1035 { 1036 #if defined(MODULES_VADDR) 1037 return MODULES_END - MODULES_VADDR; 1038 #else 1039 return VMALLOC_END - VMALLOC_START; 1040 #endif 1041 } 1042 1043 static int __init bpf_jit_charge_init(void) 1044 { 1045 /* Only used as heuristic here to derive limit. */ 1046 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1047 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1048 PAGE_SIZE), LONG_MAX); 1049 return 0; 1050 } 1051 pure_initcall(bpf_jit_charge_init); 1052 1053 int bpf_jit_charge_modmem(u32 size) 1054 { 1055 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1056 if (!bpf_capable()) { 1057 atomic_long_sub(size, &bpf_jit_current); 1058 return -EPERM; 1059 } 1060 } 1061 1062 return 0; 1063 } 1064 1065 void bpf_jit_uncharge_modmem(u32 size) 1066 { 1067 atomic_long_sub(size, &bpf_jit_current); 1068 } 1069 1070 void *__weak bpf_jit_alloc_exec(unsigned long size) 1071 { 1072 return execmem_alloc(EXECMEM_BPF, size); 1073 } 1074 1075 void __weak bpf_jit_free_exec(void *addr) 1076 { 1077 execmem_free(addr); 1078 } 1079 1080 struct bpf_binary_header * 1081 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1082 unsigned int alignment, 1083 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1084 { 1085 struct bpf_binary_header *hdr; 1086 u32 size, hole, start; 1087 1088 WARN_ON_ONCE(!is_power_of_2(alignment) || 1089 alignment > BPF_IMAGE_ALIGNMENT); 1090 1091 /* Most of BPF filters are really small, but if some of them 1092 * fill a page, allow at least 128 extra bytes to insert a 1093 * random section of illegal instructions. 1094 */ 1095 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1096 1097 if (bpf_jit_charge_modmem(size)) 1098 return NULL; 1099 hdr = bpf_jit_alloc_exec(size); 1100 if (!hdr) { 1101 bpf_jit_uncharge_modmem(size); 1102 return NULL; 1103 } 1104 1105 /* Fill space with illegal/arch-dep instructions. */ 1106 bpf_fill_ill_insns(hdr, size); 1107 1108 hdr->size = size; 1109 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1110 PAGE_SIZE - sizeof(*hdr)); 1111 start = get_random_u32_below(hole) & ~(alignment - 1); 1112 1113 /* Leave a random number of instructions before BPF code. */ 1114 *image_ptr = &hdr->image[start]; 1115 1116 return hdr; 1117 } 1118 1119 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1120 { 1121 u32 size = hdr->size; 1122 1123 bpf_jit_free_exec(hdr); 1124 bpf_jit_uncharge_modmem(size); 1125 } 1126 1127 /* Allocate jit binary from bpf_prog_pack allocator. 1128 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1129 * to the memory. To solve this problem, a RW buffer is also allocated at 1130 * as the same time. The JIT engine should calculate offsets based on the 1131 * RO memory address, but write JITed program to the RW buffer. Once the 1132 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1133 * the JITed program to the RO memory. 1134 */ 1135 struct bpf_binary_header * 1136 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1137 unsigned int alignment, 1138 struct bpf_binary_header **rw_header, 1139 u8 **rw_image, 1140 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1141 { 1142 struct bpf_binary_header *ro_header; 1143 u32 size, hole, start; 1144 1145 WARN_ON_ONCE(!is_power_of_2(alignment) || 1146 alignment > BPF_IMAGE_ALIGNMENT); 1147 1148 /* add 16 bytes for a random section of illegal instructions */ 1149 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1150 1151 if (bpf_jit_charge_modmem(size)) 1152 return NULL; 1153 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1154 if (!ro_header) { 1155 bpf_jit_uncharge_modmem(size); 1156 return NULL; 1157 } 1158 1159 *rw_header = kvmalloc(size, GFP_KERNEL); 1160 if (!*rw_header) { 1161 bpf_prog_pack_free(ro_header, size); 1162 bpf_jit_uncharge_modmem(size); 1163 return NULL; 1164 } 1165 1166 /* Fill space with illegal/arch-dep instructions. */ 1167 bpf_fill_ill_insns(*rw_header, size); 1168 (*rw_header)->size = size; 1169 1170 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1171 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1172 start = get_random_u32_below(hole) & ~(alignment - 1); 1173 1174 *image_ptr = &ro_header->image[start]; 1175 *rw_image = &(*rw_header)->image[start]; 1176 1177 return ro_header; 1178 } 1179 1180 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1181 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1182 struct bpf_binary_header *rw_header) 1183 { 1184 void *ptr; 1185 1186 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1187 1188 kvfree(rw_header); 1189 1190 if (IS_ERR(ptr)) { 1191 bpf_prog_pack_free(ro_header, ro_header->size); 1192 return PTR_ERR(ptr); 1193 } 1194 return 0; 1195 } 1196 1197 /* bpf_jit_binary_pack_free is called in two different scenarios: 1198 * 1) when the program is freed after; 1199 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1200 * For case 2), we need to free both the RO memory and the RW buffer. 1201 * 1202 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1203 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1204 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1205 * bpf_arch_text_copy (when jit fails). 1206 */ 1207 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1208 struct bpf_binary_header *rw_header) 1209 { 1210 u32 size = ro_header->size; 1211 1212 bpf_prog_pack_free(ro_header, size); 1213 kvfree(rw_header); 1214 bpf_jit_uncharge_modmem(size); 1215 } 1216 1217 struct bpf_binary_header * 1218 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1219 { 1220 unsigned long real_start = (unsigned long)fp->bpf_func; 1221 unsigned long addr; 1222 1223 addr = real_start & BPF_PROG_CHUNK_MASK; 1224 return (void *)addr; 1225 } 1226 1227 static inline struct bpf_binary_header * 1228 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1229 { 1230 unsigned long real_start = (unsigned long)fp->bpf_func; 1231 unsigned long addr; 1232 1233 addr = real_start & PAGE_MASK; 1234 return (void *)addr; 1235 } 1236 1237 /* This symbol is only overridden by archs that have different 1238 * requirements than the usual eBPF JITs, f.e. when they only 1239 * implement cBPF JIT, do not set images read-only, etc. 1240 */ 1241 void __weak bpf_jit_free(struct bpf_prog *fp) 1242 { 1243 if (fp->jited) { 1244 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1245 1246 bpf_jit_binary_free(hdr); 1247 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1248 } 1249 1250 bpf_prog_unlock_free(fp); 1251 } 1252 1253 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1254 const struct bpf_insn *insn, bool extra_pass, 1255 u64 *func_addr, bool *func_addr_fixed) 1256 { 1257 s16 off = insn->off; 1258 s32 imm = insn->imm; 1259 u8 *addr; 1260 int err; 1261 1262 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1263 if (!*func_addr_fixed) { 1264 /* Place-holder address till the last pass has collected 1265 * all addresses for JITed subprograms in which case we 1266 * can pick them up from prog->aux. 1267 */ 1268 if (!extra_pass) 1269 addr = NULL; 1270 else if (prog->aux->func && 1271 off >= 0 && off < prog->aux->real_func_cnt) 1272 addr = (u8 *)prog->aux->func[off]->bpf_func; 1273 else 1274 return -EINVAL; 1275 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1276 bpf_jit_supports_far_kfunc_call()) { 1277 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1278 if (err) 1279 return err; 1280 } else { 1281 /* Address of a BPF helper call. Since part of the core 1282 * kernel, it's always at a fixed location. __bpf_call_base 1283 * and the helper with imm relative to it are both in core 1284 * kernel. 1285 */ 1286 addr = (u8 *)__bpf_call_base + imm; 1287 } 1288 1289 *func_addr = (unsigned long)addr; 1290 return 0; 1291 } 1292 1293 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1294 const struct bpf_insn *aux, 1295 struct bpf_insn *to_buff, 1296 bool emit_zext) 1297 { 1298 struct bpf_insn *to = to_buff; 1299 u32 imm_rnd = get_random_u32(); 1300 s16 off; 1301 1302 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1303 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1304 1305 /* Constraints on AX register: 1306 * 1307 * AX register is inaccessible from user space. It is mapped in 1308 * all JITs, and used here for constant blinding rewrites. It is 1309 * typically "stateless" meaning its contents are only valid within 1310 * the executed instruction, but not across several instructions. 1311 * There are a few exceptions however which are further detailed 1312 * below. 1313 * 1314 * Constant blinding is only used by JITs, not in the interpreter. 1315 * The interpreter uses AX in some occasions as a local temporary 1316 * register e.g. in DIV or MOD instructions. 1317 * 1318 * In restricted circumstances, the verifier can also use the AX 1319 * register for rewrites as long as they do not interfere with 1320 * the above cases! 1321 */ 1322 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1323 goto out; 1324 1325 if (from->imm == 0 && 1326 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1327 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1328 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1329 goto out; 1330 } 1331 1332 switch (from->code) { 1333 case BPF_ALU | BPF_ADD | BPF_K: 1334 case BPF_ALU | BPF_SUB | BPF_K: 1335 case BPF_ALU | BPF_AND | BPF_K: 1336 case BPF_ALU | BPF_OR | BPF_K: 1337 case BPF_ALU | BPF_XOR | BPF_K: 1338 case BPF_ALU | BPF_MUL | BPF_K: 1339 case BPF_ALU | BPF_MOV | BPF_K: 1340 case BPF_ALU | BPF_DIV | BPF_K: 1341 case BPF_ALU | BPF_MOD | BPF_K: 1342 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1343 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1344 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1345 break; 1346 1347 case BPF_ALU64 | BPF_ADD | BPF_K: 1348 case BPF_ALU64 | BPF_SUB | BPF_K: 1349 case BPF_ALU64 | BPF_AND | BPF_K: 1350 case BPF_ALU64 | BPF_OR | BPF_K: 1351 case BPF_ALU64 | BPF_XOR | BPF_K: 1352 case BPF_ALU64 | BPF_MUL | BPF_K: 1353 case BPF_ALU64 | BPF_MOV | BPF_K: 1354 case BPF_ALU64 | BPF_DIV | BPF_K: 1355 case BPF_ALU64 | BPF_MOD | BPF_K: 1356 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1357 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1358 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1359 break; 1360 1361 case BPF_JMP | BPF_JEQ | BPF_K: 1362 case BPF_JMP | BPF_JNE | BPF_K: 1363 case BPF_JMP | BPF_JGT | BPF_K: 1364 case BPF_JMP | BPF_JLT | BPF_K: 1365 case BPF_JMP | BPF_JGE | BPF_K: 1366 case BPF_JMP | BPF_JLE | BPF_K: 1367 case BPF_JMP | BPF_JSGT | BPF_K: 1368 case BPF_JMP | BPF_JSLT | BPF_K: 1369 case BPF_JMP | BPF_JSGE | BPF_K: 1370 case BPF_JMP | BPF_JSLE | BPF_K: 1371 case BPF_JMP | BPF_JSET | BPF_K: 1372 /* Accommodate for extra offset in case of a backjump. */ 1373 off = from->off; 1374 if (off < 0) 1375 off -= 2; 1376 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1377 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1378 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1379 break; 1380 1381 case BPF_JMP32 | BPF_JEQ | BPF_K: 1382 case BPF_JMP32 | BPF_JNE | BPF_K: 1383 case BPF_JMP32 | BPF_JGT | BPF_K: 1384 case BPF_JMP32 | BPF_JLT | BPF_K: 1385 case BPF_JMP32 | BPF_JGE | BPF_K: 1386 case BPF_JMP32 | BPF_JLE | BPF_K: 1387 case BPF_JMP32 | BPF_JSGT | BPF_K: 1388 case BPF_JMP32 | BPF_JSLT | BPF_K: 1389 case BPF_JMP32 | BPF_JSGE | BPF_K: 1390 case BPF_JMP32 | BPF_JSLE | BPF_K: 1391 case BPF_JMP32 | BPF_JSET | BPF_K: 1392 /* Accommodate for extra offset in case of a backjump. */ 1393 off = from->off; 1394 if (off < 0) 1395 off -= 2; 1396 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1397 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1398 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1399 off); 1400 break; 1401 1402 case BPF_LD | BPF_IMM | BPF_DW: 1403 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1404 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1405 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1406 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1407 break; 1408 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1409 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1410 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1411 if (emit_zext) 1412 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1413 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1414 break; 1415 1416 case BPF_ST | BPF_MEM | BPF_DW: 1417 case BPF_ST | BPF_MEM | BPF_W: 1418 case BPF_ST | BPF_MEM | BPF_H: 1419 case BPF_ST | BPF_MEM | BPF_B: 1420 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1421 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1422 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1423 break; 1424 } 1425 out: 1426 return to - to_buff; 1427 } 1428 1429 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1430 gfp_t gfp_extra_flags) 1431 { 1432 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1433 struct bpf_prog *fp; 1434 1435 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1436 if (fp != NULL) { 1437 /* aux->prog still points to the fp_other one, so 1438 * when promoting the clone to the real program, 1439 * this still needs to be adapted. 1440 */ 1441 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1442 } 1443 1444 return fp; 1445 } 1446 1447 static void bpf_prog_clone_free(struct bpf_prog *fp) 1448 { 1449 /* aux was stolen by the other clone, so we cannot free 1450 * it from this path! It will be freed eventually by the 1451 * other program on release. 1452 * 1453 * At this point, we don't need a deferred release since 1454 * clone is guaranteed to not be locked. 1455 */ 1456 fp->aux = NULL; 1457 fp->stats = NULL; 1458 fp->active = NULL; 1459 __bpf_prog_free(fp); 1460 } 1461 1462 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1463 { 1464 /* We have to repoint aux->prog to self, as we don't 1465 * know whether fp here is the clone or the original. 1466 */ 1467 fp->aux->prog = fp; 1468 bpf_prog_clone_free(fp_other); 1469 } 1470 1471 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1472 { 1473 struct bpf_insn insn_buff[16], aux[2]; 1474 struct bpf_prog *clone, *tmp; 1475 int insn_delta, insn_cnt; 1476 struct bpf_insn *insn; 1477 int i, rewritten; 1478 1479 if (!prog->blinding_requested || prog->blinded) 1480 return prog; 1481 1482 clone = bpf_prog_clone_create(prog, GFP_USER); 1483 if (!clone) 1484 return ERR_PTR(-ENOMEM); 1485 1486 insn_cnt = clone->len; 1487 insn = clone->insnsi; 1488 1489 for (i = 0; i < insn_cnt; i++, insn++) { 1490 if (bpf_pseudo_func(insn)) { 1491 /* ld_imm64 with an address of bpf subprog is not 1492 * a user controlled constant. Don't randomize it, 1493 * since it will conflict with jit_subprogs() logic. 1494 */ 1495 insn++; 1496 i++; 1497 continue; 1498 } 1499 1500 /* We temporarily need to hold the original ld64 insn 1501 * so that we can still access the first part in the 1502 * second blinding run. 1503 */ 1504 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1505 insn[1].code == 0) 1506 memcpy(aux, insn, sizeof(aux)); 1507 1508 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1509 clone->aux->verifier_zext); 1510 if (!rewritten) 1511 continue; 1512 1513 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1514 if (IS_ERR(tmp)) { 1515 /* Patching may have repointed aux->prog during 1516 * realloc from the original one, so we need to 1517 * fix it up here on error. 1518 */ 1519 bpf_jit_prog_release_other(prog, clone); 1520 return tmp; 1521 } 1522 1523 clone = tmp; 1524 insn_delta = rewritten - 1; 1525 1526 /* Walk new program and skip insns we just inserted. */ 1527 insn = clone->insnsi + i + insn_delta; 1528 insn_cnt += insn_delta; 1529 i += insn_delta; 1530 } 1531 1532 clone->blinded = 1; 1533 return clone; 1534 } 1535 #endif /* CONFIG_BPF_JIT */ 1536 1537 /* Base function for offset calculation. Needs to go into .text section, 1538 * therefore keeping it non-static as well; will also be used by JITs 1539 * anyway later on, so do not let the compiler omit it. This also needs 1540 * to go into kallsyms for correlation from e.g. bpftool, so naming 1541 * must not change. 1542 */ 1543 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1544 { 1545 return 0; 1546 } 1547 EXPORT_SYMBOL_GPL(__bpf_call_base); 1548 1549 /* All UAPI available opcodes. */ 1550 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1551 /* 32 bit ALU operations. */ \ 1552 /* Register based. */ \ 1553 INSN_3(ALU, ADD, X), \ 1554 INSN_3(ALU, SUB, X), \ 1555 INSN_3(ALU, AND, X), \ 1556 INSN_3(ALU, OR, X), \ 1557 INSN_3(ALU, LSH, X), \ 1558 INSN_3(ALU, RSH, X), \ 1559 INSN_3(ALU, XOR, X), \ 1560 INSN_3(ALU, MUL, X), \ 1561 INSN_3(ALU, MOV, X), \ 1562 INSN_3(ALU, ARSH, X), \ 1563 INSN_3(ALU, DIV, X), \ 1564 INSN_3(ALU, MOD, X), \ 1565 INSN_2(ALU, NEG), \ 1566 INSN_3(ALU, END, TO_BE), \ 1567 INSN_3(ALU, END, TO_LE), \ 1568 /* Immediate based. */ \ 1569 INSN_3(ALU, ADD, K), \ 1570 INSN_3(ALU, SUB, K), \ 1571 INSN_3(ALU, AND, K), \ 1572 INSN_3(ALU, OR, K), \ 1573 INSN_3(ALU, LSH, K), \ 1574 INSN_3(ALU, RSH, K), \ 1575 INSN_3(ALU, XOR, K), \ 1576 INSN_3(ALU, MUL, K), \ 1577 INSN_3(ALU, MOV, K), \ 1578 INSN_3(ALU, ARSH, K), \ 1579 INSN_3(ALU, DIV, K), \ 1580 INSN_3(ALU, MOD, K), \ 1581 /* 64 bit ALU operations. */ \ 1582 /* Register based. */ \ 1583 INSN_3(ALU64, ADD, X), \ 1584 INSN_3(ALU64, SUB, X), \ 1585 INSN_3(ALU64, AND, X), \ 1586 INSN_3(ALU64, OR, X), \ 1587 INSN_3(ALU64, LSH, X), \ 1588 INSN_3(ALU64, RSH, X), \ 1589 INSN_3(ALU64, XOR, X), \ 1590 INSN_3(ALU64, MUL, X), \ 1591 INSN_3(ALU64, MOV, X), \ 1592 INSN_3(ALU64, ARSH, X), \ 1593 INSN_3(ALU64, DIV, X), \ 1594 INSN_3(ALU64, MOD, X), \ 1595 INSN_2(ALU64, NEG), \ 1596 INSN_3(ALU64, END, TO_LE), \ 1597 /* Immediate based. */ \ 1598 INSN_3(ALU64, ADD, K), \ 1599 INSN_3(ALU64, SUB, K), \ 1600 INSN_3(ALU64, AND, K), \ 1601 INSN_3(ALU64, OR, K), \ 1602 INSN_3(ALU64, LSH, K), \ 1603 INSN_3(ALU64, RSH, K), \ 1604 INSN_3(ALU64, XOR, K), \ 1605 INSN_3(ALU64, MUL, K), \ 1606 INSN_3(ALU64, MOV, K), \ 1607 INSN_3(ALU64, ARSH, K), \ 1608 INSN_3(ALU64, DIV, K), \ 1609 INSN_3(ALU64, MOD, K), \ 1610 /* Call instruction. */ \ 1611 INSN_2(JMP, CALL), \ 1612 /* Exit instruction. */ \ 1613 INSN_2(JMP, EXIT), \ 1614 /* 32-bit Jump instructions. */ \ 1615 /* Register based. */ \ 1616 INSN_3(JMP32, JEQ, X), \ 1617 INSN_3(JMP32, JNE, X), \ 1618 INSN_3(JMP32, JGT, X), \ 1619 INSN_3(JMP32, JLT, X), \ 1620 INSN_3(JMP32, JGE, X), \ 1621 INSN_3(JMP32, JLE, X), \ 1622 INSN_3(JMP32, JSGT, X), \ 1623 INSN_3(JMP32, JSLT, X), \ 1624 INSN_3(JMP32, JSGE, X), \ 1625 INSN_3(JMP32, JSLE, X), \ 1626 INSN_3(JMP32, JSET, X), \ 1627 /* Immediate based. */ \ 1628 INSN_3(JMP32, JEQ, K), \ 1629 INSN_3(JMP32, JNE, K), \ 1630 INSN_3(JMP32, JGT, K), \ 1631 INSN_3(JMP32, JLT, K), \ 1632 INSN_3(JMP32, JGE, K), \ 1633 INSN_3(JMP32, JLE, K), \ 1634 INSN_3(JMP32, JSGT, K), \ 1635 INSN_3(JMP32, JSLT, K), \ 1636 INSN_3(JMP32, JSGE, K), \ 1637 INSN_3(JMP32, JSLE, K), \ 1638 INSN_3(JMP32, JSET, K), \ 1639 /* Jump instructions. */ \ 1640 /* Register based. */ \ 1641 INSN_3(JMP, JEQ, X), \ 1642 INSN_3(JMP, JNE, X), \ 1643 INSN_3(JMP, JGT, X), \ 1644 INSN_3(JMP, JLT, X), \ 1645 INSN_3(JMP, JGE, X), \ 1646 INSN_3(JMP, JLE, X), \ 1647 INSN_3(JMP, JSGT, X), \ 1648 INSN_3(JMP, JSLT, X), \ 1649 INSN_3(JMP, JSGE, X), \ 1650 INSN_3(JMP, JSLE, X), \ 1651 INSN_3(JMP, JSET, X), \ 1652 /* Immediate based. */ \ 1653 INSN_3(JMP, JEQ, K), \ 1654 INSN_3(JMP, JNE, K), \ 1655 INSN_3(JMP, JGT, K), \ 1656 INSN_3(JMP, JLT, K), \ 1657 INSN_3(JMP, JGE, K), \ 1658 INSN_3(JMP, JLE, K), \ 1659 INSN_3(JMP, JSGT, K), \ 1660 INSN_3(JMP, JSLT, K), \ 1661 INSN_3(JMP, JSGE, K), \ 1662 INSN_3(JMP, JSLE, K), \ 1663 INSN_3(JMP, JSET, K), \ 1664 INSN_2(JMP, JA), \ 1665 INSN_2(JMP32, JA), \ 1666 /* Store instructions. */ \ 1667 /* Register based. */ \ 1668 INSN_3(STX, MEM, B), \ 1669 INSN_3(STX, MEM, H), \ 1670 INSN_3(STX, MEM, W), \ 1671 INSN_3(STX, MEM, DW), \ 1672 INSN_3(STX, ATOMIC, W), \ 1673 INSN_3(STX, ATOMIC, DW), \ 1674 /* Immediate based. */ \ 1675 INSN_3(ST, MEM, B), \ 1676 INSN_3(ST, MEM, H), \ 1677 INSN_3(ST, MEM, W), \ 1678 INSN_3(ST, MEM, DW), \ 1679 /* Load instructions. */ \ 1680 /* Register based. */ \ 1681 INSN_3(LDX, MEM, B), \ 1682 INSN_3(LDX, MEM, H), \ 1683 INSN_3(LDX, MEM, W), \ 1684 INSN_3(LDX, MEM, DW), \ 1685 INSN_3(LDX, MEMSX, B), \ 1686 INSN_3(LDX, MEMSX, H), \ 1687 INSN_3(LDX, MEMSX, W), \ 1688 /* Immediate based. */ \ 1689 INSN_3(LD, IMM, DW) 1690 1691 bool bpf_opcode_in_insntable(u8 code) 1692 { 1693 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1694 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1695 static const bool public_insntable[256] = { 1696 [0 ... 255] = false, 1697 /* Now overwrite non-defaults ... */ 1698 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1699 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1700 [BPF_LD | BPF_ABS | BPF_B] = true, 1701 [BPF_LD | BPF_ABS | BPF_H] = true, 1702 [BPF_LD | BPF_ABS | BPF_W] = true, 1703 [BPF_LD | BPF_IND | BPF_B] = true, 1704 [BPF_LD | BPF_IND | BPF_H] = true, 1705 [BPF_LD | BPF_IND | BPF_W] = true, 1706 [BPF_JMP | BPF_JCOND] = true, 1707 }; 1708 #undef BPF_INSN_3_TBL 1709 #undef BPF_INSN_2_TBL 1710 return public_insntable[code]; 1711 } 1712 1713 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1714 /** 1715 * ___bpf_prog_run - run eBPF program on a given context 1716 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1717 * @insn: is the array of eBPF instructions 1718 * 1719 * Decode and execute eBPF instructions. 1720 * 1721 * Return: whatever value is in %BPF_R0 at program exit 1722 */ 1723 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1724 { 1725 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1726 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1727 static const void * const jumptable[256] __annotate_jump_table = { 1728 [0 ... 255] = &&default_label, 1729 /* Now overwrite non-defaults ... */ 1730 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1731 /* Non-UAPI available opcodes. */ 1732 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1733 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1734 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1735 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1736 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1737 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1738 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1739 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1740 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1741 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1742 }; 1743 #undef BPF_INSN_3_LBL 1744 #undef BPF_INSN_2_LBL 1745 u32 tail_call_cnt = 0; 1746 1747 #define CONT ({ insn++; goto select_insn; }) 1748 #define CONT_JMP ({ insn++; goto select_insn; }) 1749 1750 select_insn: 1751 goto *jumptable[insn->code]; 1752 1753 /* Explicitly mask the register-based shift amounts with 63 or 31 1754 * to avoid undefined behavior. Normally this won't affect the 1755 * generated code, for example, in case of native 64 bit archs such 1756 * as x86-64 or arm64, the compiler is optimizing the AND away for 1757 * the interpreter. In case of JITs, each of the JIT backends compiles 1758 * the BPF shift operations to machine instructions which produce 1759 * implementation-defined results in such a case; the resulting 1760 * contents of the register may be arbitrary, but program behaviour 1761 * as a whole remains defined. In other words, in case of JIT backends, 1762 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1763 */ 1764 /* ALU (shifts) */ 1765 #define SHT(OPCODE, OP) \ 1766 ALU64_##OPCODE##_X: \ 1767 DST = DST OP (SRC & 63); \ 1768 CONT; \ 1769 ALU_##OPCODE##_X: \ 1770 DST = (u32) DST OP ((u32) SRC & 31); \ 1771 CONT; \ 1772 ALU64_##OPCODE##_K: \ 1773 DST = DST OP IMM; \ 1774 CONT; \ 1775 ALU_##OPCODE##_K: \ 1776 DST = (u32) DST OP (u32) IMM; \ 1777 CONT; 1778 /* ALU (rest) */ 1779 #define ALU(OPCODE, OP) \ 1780 ALU64_##OPCODE##_X: \ 1781 DST = DST OP SRC; \ 1782 CONT; \ 1783 ALU_##OPCODE##_X: \ 1784 DST = (u32) DST OP (u32) SRC; \ 1785 CONT; \ 1786 ALU64_##OPCODE##_K: \ 1787 DST = DST OP IMM; \ 1788 CONT; \ 1789 ALU_##OPCODE##_K: \ 1790 DST = (u32) DST OP (u32) IMM; \ 1791 CONT; 1792 ALU(ADD, +) 1793 ALU(SUB, -) 1794 ALU(AND, &) 1795 ALU(OR, |) 1796 ALU(XOR, ^) 1797 ALU(MUL, *) 1798 SHT(LSH, <<) 1799 SHT(RSH, >>) 1800 #undef SHT 1801 #undef ALU 1802 ALU_NEG: 1803 DST = (u32) -DST; 1804 CONT; 1805 ALU64_NEG: 1806 DST = -DST; 1807 CONT; 1808 ALU_MOV_X: 1809 switch (OFF) { 1810 case 0: 1811 DST = (u32) SRC; 1812 break; 1813 case 8: 1814 DST = (u32)(s8) SRC; 1815 break; 1816 case 16: 1817 DST = (u32)(s16) SRC; 1818 break; 1819 } 1820 CONT; 1821 ALU_MOV_K: 1822 DST = (u32) IMM; 1823 CONT; 1824 ALU64_MOV_X: 1825 switch (OFF) { 1826 case 0: 1827 DST = SRC; 1828 break; 1829 case 8: 1830 DST = (s8) SRC; 1831 break; 1832 case 16: 1833 DST = (s16) SRC; 1834 break; 1835 case 32: 1836 DST = (s32) SRC; 1837 break; 1838 } 1839 CONT; 1840 ALU64_MOV_K: 1841 DST = IMM; 1842 CONT; 1843 LD_IMM_DW: 1844 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1845 insn++; 1846 CONT; 1847 ALU_ARSH_X: 1848 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1849 CONT; 1850 ALU_ARSH_K: 1851 DST = (u64) (u32) (((s32) DST) >> IMM); 1852 CONT; 1853 ALU64_ARSH_X: 1854 (*(s64 *) &DST) >>= (SRC & 63); 1855 CONT; 1856 ALU64_ARSH_K: 1857 (*(s64 *) &DST) >>= IMM; 1858 CONT; 1859 ALU64_MOD_X: 1860 switch (OFF) { 1861 case 0: 1862 div64_u64_rem(DST, SRC, &AX); 1863 DST = AX; 1864 break; 1865 case 1: 1866 AX = div64_s64(DST, SRC); 1867 DST = DST - AX * SRC; 1868 break; 1869 } 1870 CONT; 1871 ALU_MOD_X: 1872 switch (OFF) { 1873 case 0: 1874 AX = (u32) DST; 1875 DST = do_div(AX, (u32) SRC); 1876 break; 1877 case 1: 1878 AX = abs((s32)DST); 1879 AX = do_div(AX, abs((s32)SRC)); 1880 if ((s32)DST < 0) 1881 DST = (u32)-AX; 1882 else 1883 DST = (u32)AX; 1884 break; 1885 } 1886 CONT; 1887 ALU64_MOD_K: 1888 switch (OFF) { 1889 case 0: 1890 div64_u64_rem(DST, IMM, &AX); 1891 DST = AX; 1892 break; 1893 case 1: 1894 AX = div64_s64(DST, IMM); 1895 DST = DST - AX * IMM; 1896 break; 1897 } 1898 CONT; 1899 ALU_MOD_K: 1900 switch (OFF) { 1901 case 0: 1902 AX = (u32) DST; 1903 DST = do_div(AX, (u32) IMM); 1904 break; 1905 case 1: 1906 AX = abs((s32)DST); 1907 AX = do_div(AX, abs((s32)IMM)); 1908 if ((s32)DST < 0) 1909 DST = (u32)-AX; 1910 else 1911 DST = (u32)AX; 1912 break; 1913 } 1914 CONT; 1915 ALU64_DIV_X: 1916 switch (OFF) { 1917 case 0: 1918 DST = div64_u64(DST, SRC); 1919 break; 1920 case 1: 1921 DST = div64_s64(DST, SRC); 1922 break; 1923 } 1924 CONT; 1925 ALU_DIV_X: 1926 switch (OFF) { 1927 case 0: 1928 AX = (u32) DST; 1929 do_div(AX, (u32) SRC); 1930 DST = (u32) AX; 1931 break; 1932 case 1: 1933 AX = abs((s32)DST); 1934 do_div(AX, abs((s32)SRC)); 1935 if (((s32)DST < 0) == ((s32)SRC < 0)) 1936 DST = (u32)AX; 1937 else 1938 DST = (u32)-AX; 1939 break; 1940 } 1941 CONT; 1942 ALU64_DIV_K: 1943 switch (OFF) { 1944 case 0: 1945 DST = div64_u64(DST, IMM); 1946 break; 1947 case 1: 1948 DST = div64_s64(DST, IMM); 1949 break; 1950 } 1951 CONT; 1952 ALU_DIV_K: 1953 switch (OFF) { 1954 case 0: 1955 AX = (u32) DST; 1956 do_div(AX, (u32) IMM); 1957 DST = (u32) AX; 1958 break; 1959 case 1: 1960 AX = abs((s32)DST); 1961 do_div(AX, abs((s32)IMM)); 1962 if (((s32)DST < 0) == ((s32)IMM < 0)) 1963 DST = (u32)AX; 1964 else 1965 DST = (u32)-AX; 1966 break; 1967 } 1968 CONT; 1969 ALU_END_TO_BE: 1970 switch (IMM) { 1971 case 16: 1972 DST = (__force u16) cpu_to_be16(DST); 1973 break; 1974 case 32: 1975 DST = (__force u32) cpu_to_be32(DST); 1976 break; 1977 case 64: 1978 DST = (__force u64) cpu_to_be64(DST); 1979 break; 1980 } 1981 CONT; 1982 ALU_END_TO_LE: 1983 switch (IMM) { 1984 case 16: 1985 DST = (__force u16) cpu_to_le16(DST); 1986 break; 1987 case 32: 1988 DST = (__force u32) cpu_to_le32(DST); 1989 break; 1990 case 64: 1991 DST = (__force u64) cpu_to_le64(DST); 1992 break; 1993 } 1994 CONT; 1995 ALU64_END_TO_LE: 1996 switch (IMM) { 1997 case 16: 1998 DST = (__force u16) __swab16(DST); 1999 break; 2000 case 32: 2001 DST = (__force u32) __swab32(DST); 2002 break; 2003 case 64: 2004 DST = (__force u64) __swab64(DST); 2005 break; 2006 } 2007 CONT; 2008 2009 /* CALL */ 2010 JMP_CALL: 2011 /* Function call scratches BPF_R1-BPF_R5 registers, 2012 * preserves BPF_R6-BPF_R9, and stores return value 2013 * into BPF_R0. 2014 */ 2015 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2016 BPF_R4, BPF_R5); 2017 CONT; 2018 2019 JMP_CALL_ARGS: 2020 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2021 BPF_R3, BPF_R4, 2022 BPF_R5, 2023 insn + insn->off + 1); 2024 CONT; 2025 2026 JMP_TAIL_CALL: { 2027 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2028 struct bpf_array *array = container_of(map, struct bpf_array, map); 2029 struct bpf_prog *prog; 2030 u32 index = BPF_R3; 2031 2032 if (unlikely(index >= array->map.max_entries)) 2033 goto out; 2034 2035 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2036 goto out; 2037 2038 tail_call_cnt++; 2039 2040 prog = READ_ONCE(array->ptrs[index]); 2041 if (!prog) 2042 goto out; 2043 2044 /* ARG1 at this point is guaranteed to point to CTX from 2045 * the verifier side due to the fact that the tail call is 2046 * handled like a helper, that is, bpf_tail_call_proto, 2047 * where arg1_type is ARG_PTR_TO_CTX. 2048 */ 2049 insn = prog->insnsi; 2050 goto select_insn; 2051 out: 2052 CONT; 2053 } 2054 JMP_JA: 2055 insn += insn->off; 2056 CONT; 2057 JMP32_JA: 2058 insn += insn->imm; 2059 CONT; 2060 JMP_EXIT: 2061 return BPF_R0; 2062 /* JMP */ 2063 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2064 JMP_##OPCODE##_X: \ 2065 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2066 insn += insn->off; \ 2067 CONT_JMP; \ 2068 } \ 2069 CONT; \ 2070 JMP32_##OPCODE##_X: \ 2071 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2072 insn += insn->off; \ 2073 CONT_JMP; \ 2074 } \ 2075 CONT; \ 2076 JMP_##OPCODE##_K: \ 2077 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2078 insn += insn->off; \ 2079 CONT_JMP; \ 2080 } \ 2081 CONT; \ 2082 JMP32_##OPCODE##_K: \ 2083 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2084 insn += insn->off; \ 2085 CONT_JMP; \ 2086 } \ 2087 CONT; 2088 COND_JMP(u, JEQ, ==) 2089 COND_JMP(u, JNE, !=) 2090 COND_JMP(u, JGT, >) 2091 COND_JMP(u, JLT, <) 2092 COND_JMP(u, JGE, >=) 2093 COND_JMP(u, JLE, <=) 2094 COND_JMP(u, JSET, &) 2095 COND_JMP(s, JSGT, >) 2096 COND_JMP(s, JSLT, <) 2097 COND_JMP(s, JSGE, >=) 2098 COND_JMP(s, JSLE, <=) 2099 #undef COND_JMP 2100 /* ST, STX and LDX*/ 2101 ST_NOSPEC: 2102 /* Speculation barrier for mitigating Speculative Store Bypass. 2103 * In case of arm64, we rely on the firmware mitigation as 2104 * controlled via the ssbd kernel parameter. Whenever the 2105 * mitigation is enabled, it works for all of the kernel code 2106 * with no need to provide any additional instructions here. 2107 * In case of x86, we use 'lfence' insn for mitigation. We 2108 * reuse preexisting logic from Spectre v1 mitigation that 2109 * happens to produce the required code on x86 for v4 as well. 2110 */ 2111 barrier_nospec(); 2112 CONT; 2113 #define LDST(SIZEOP, SIZE) \ 2114 STX_MEM_##SIZEOP: \ 2115 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2116 CONT; \ 2117 ST_MEM_##SIZEOP: \ 2118 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2119 CONT; \ 2120 LDX_MEM_##SIZEOP: \ 2121 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2122 CONT; \ 2123 LDX_PROBE_MEM_##SIZEOP: \ 2124 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2125 (const void *)(long) (SRC + insn->off)); \ 2126 DST = *((SIZE *)&DST); \ 2127 CONT; 2128 2129 LDST(B, u8) 2130 LDST(H, u16) 2131 LDST(W, u32) 2132 LDST(DW, u64) 2133 #undef LDST 2134 2135 #define LDSX(SIZEOP, SIZE) \ 2136 LDX_MEMSX_##SIZEOP: \ 2137 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2138 CONT; \ 2139 LDX_PROBE_MEMSX_##SIZEOP: \ 2140 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2141 (const void *)(long) (SRC + insn->off)); \ 2142 DST = *((SIZE *)&DST); \ 2143 CONT; 2144 2145 LDSX(B, s8) 2146 LDSX(H, s16) 2147 LDSX(W, s32) 2148 #undef LDSX 2149 2150 #define ATOMIC_ALU_OP(BOP, KOP) \ 2151 case BOP: \ 2152 if (BPF_SIZE(insn->code) == BPF_W) \ 2153 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2154 (DST + insn->off)); \ 2155 else \ 2156 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2157 (DST + insn->off)); \ 2158 break; \ 2159 case BOP | BPF_FETCH: \ 2160 if (BPF_SIZE(insn->code) == BPF_W) \ 2161 SRC = (u32) atomic_fetch_##KOP( \ 2162 (u32) SRC, \ 2163 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2164 else \ 2165 SRC = (u64) atomic64_fetch_##KOP( \ 2166 (u64) SRC, \ 2167 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2168 break; 2169 2170 STX_ATOMIC_DW: 2171 STX_ATOMIC_W: 2172 switch (IMM) { 2173 ATOMIC_ALU_OP(BPF_ADD, add) 2174 ATOMIC_ALU_OP(BPF_AND, and) 2175 ATOMIC_ALU_OP(BPF_OR, or) 2176 ATOMIC_ALU_OP(BPF_XOR, xor) 2177 #undef ATOMIC_ALU_OP 2178 2179 case BPF_XCHG: 2180 if (BPF_SIZE(insn->code) == BPF_W) 2181 SRC = (u32) atomic_xchg( 2182 (atomic_t *)(unsigned long) (DST + insn->off), 2183 (u32) SRC); 2184 else 2185 SRC = (u64) atomic64_xchg( 2186 (atomic64_t *)(unsigned long) (DST + insn->off), 2187 (u64) SRC); 2188 break; 2189 case BPF_CMPXCHG: 2190 if (BPF_SIZE(insn->code) == BPF_W) 2191 BPF_R0 = (u32) atomic_cmpxchg( 2192 (atomic_t *)(unsigned long) (DST + insn->off), 2193 (u32) BPF_R0, (u32) SRC); 2194 else 2195 BPF_R0 = (u64) atomic64_cmpxchg( 2196 (atomic64_t *)(unsigned long) (DST + insn->off), 2197 (u64) BPF_R0, (u64) SRC); 2198 break; 2199 2200 default: 2201 goto default_label; 2202 } 2203 CONT; 2204 2205 default_label: 2206 /* If we ever reach this, we have a bug somewhere. Die hard here 2207 * instead of just returning 0; we could be somewhere in a subprog, 2208 * so execution could continue otherwise which we do /not/ want. 2209 * 2210 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2211 */ 2212 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2213 insn->code, insn->imm); 2214 BUG_ON(1); 2215 return 0; 2216 } 2217 2218 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2219 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2220 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2221 { \ 2222 u64 stack[stack_size / sizeof(u64)]; \ 2223 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2224 \ 2225 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2226 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2227 ARG1 = (u64) (unsigned long) ctx; \ 2228 return ___bpf_prog_run(regs, insn); \ 2229 } 2230 2231 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2232 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2233 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2234 const struct bpf_insn *insn) \ 2235 { \ 2236 u64 stack[stack_size / sizeof(u64)]; \ 2237 u64 regs[MAX_BPF_EXT_REG]; \ 2238 \ 2239 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2240 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2241 BPF_R1 = r1; \ 2242 BPF_R2 = r2; \ 2243 BPF_R3 = r3; \ 2244 BPF_R4 = r4; \ 2245 BPF_R5 = r5; \ 2246 return ___bpf_prog_run(regs, insn); \ 2247 } 2248 2249 #define EVAL1(FN, X) FN(X) 2250 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2251 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2252 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2253 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2254 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2255 2256 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2257 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2258 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2259 2260 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2261 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2262 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2263 2264 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2265 2266 static unsigned int (*interpreters[])(const void *ctx, 2267 const struct bpf_insn *insn) = { 2268 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2269 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2270 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2271 }; 2272 #undef PROG_NAME_LIST 2273 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2274 static __maybe_unused 2275 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2276 const struct bpf_insn *insn) = { 2277 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2278 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2279 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2280 }; 2281 #undef PROG_NAME_LIST 2282 2283 #ifdef CONFIG_BPF_SYSCALL 2284 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2285 { 2286 stack_depth = max_t(u32, stack_depth, 1); 2287 insn->off = (s16) insn->imm; 2288 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2289 __bpf_call_base_args; 2290 insn->code = BPF_JMP | BPF_CALL_ARGS; 2291 } 2292 #endif 2293 #else 2294 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2295 const struct bpf_insn *insn) 2296 { 2297 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2298 * is not working properly, so warn about it! 2299 */ 2300 WARN_ON_ONCE(1); 2301 return 0; 2302 } 2303 #endif 2304 2305 bool bpf_prog_map_compatible(struct bpf_map *map, 2306 const struct bpf_prog *fp) 2307 { 2308 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2309 bool ret; 2310 struct bpf_prog_aux *aux = fp->aux; 2311 2312 if (fp->kprobe_override) 2313 return false; 2314 2315 /* XDP programs inserted into maps are not guaranteed to run on 2316 * a particular netdev (and can run outside driver context entirely 2317 * in the case of devmap and cpumap). Until device checks 2318 * are implemented, prohibit adding dev-bound programs to program maps. 2319 */ 2320 if (bpf_prog_is_dev_bound(aux)) 2321 return false; 2322 2323 spin_lock(&map->owner.lock); 2324 if (!map->owner.type) { 2325 /* There's no owner yet where we could check for 2326 * compatibility. 2327 */ 2328 map->owner.type = prog_type; 2329 map->owner.jited = fp->jited; 2330 map->owner.xdp_has_frags = aux->xdp_has_frags; 2331 map->owner.attach_func_proto = aux->attach_func_proto; 2332 ret = true; 2333 } else { 2334 ret = map->owner.type == prog_type && 2335 map->owner.jited == fp->jited && 2336 map->owner.xdp_has_frags == aux->xdp_has_frags; 2337 if (ret && 2338 map->owner.attach_func_proto != aux->attach_func_proto) { 2339 switch (prog_type) { 2340 case BPF_PROG_TYPE_TRACING: 2341 case BPF_PROG_TYPE_LSM: 2342 case BPF_PROG_TYPE_EXT: 2343 case BPF_PROG_TYPE_STRUCT_OPS: 2344 ret = false; 2345 break; 2346 default: 2347 break; 2348 } 2349 } 2350 } 2351 spin_unlock(&map->owner.lock); 2352 2353 return ret; 2354 } 2355 2356 static int bpf_check_tail_call(const struct bpf_prog *fp) 2357 { 2358 struct bpf_prog_aux *aux = fp->aux; 2359 int i, ret = 0; 2360 2361 mutex_lock(&aux->used_maps_mutex); 2362 for (i = 0; i < aux->used_map_cnt; i++) { 2363 struct bpf_map *map = aux->used_maps[i]; 2364 2365 if (!map_type_contains_progs(map)) 2366 continue; 2367 2368 if (!bpf_prog_map_compatible(map, fp)) { 2369 ret = -EINVAL; 2370 goto out; 2371 } 2372 } 2373 2374 out: 2375 mutex_unlock(&aux->used_maps_mutex); 2376 return ret; 2377 } 2378 2379 static void bpf_prog_select_func(struct bpf_prog *fp) 2380 { 2381 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2382 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2383 2384 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 2385 #else 2386 fp->bpf_func = __bpf_prog_ret0_warn; 2387 #endif 2388 } 2389 2390 /** 2391 * bpf_prog_select_runtime - select exec runtime for BPF program 2392 * @fp: bpf_prog populated with BPF program 2393 * @err: pointer to error variable 2394 * 2395 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2396 * The BPF program will be executed via bpf_prog_run() function. 2397 * 2398 * Return: the &fp argument along with &err set to 0 for success or 2399 * a negative errno code on failure 2400 */ 2401 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2402 { 2403 /* In case of BPF to BPF calls, verifier did all the prep 2404 * work with regards to JITing, etc. 2405 */ 2406 bool jit_needed = false; 2407 2408 if (fp->bpf_func) 2409 goto finalize; 2410 2411 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2412 bpf_prog_has_kfunc_call(fp)) 2413 jit_needed = true; 2414 2415 bpf_prog_select_func(fp); 2416 2417 /* eBPF JITs can rewrite the program in case constant 2418 * blinding is active. However, in case of error during 2419 * blinding, bpf_int_jit_compile() must always return a 2420 * valid program, which in this case would simply not 2421 * be JITed, but falls back to the interpreter. 2422 */ 2423 if (!bpf_prog_is_offloaded(fp->aux)) { 2424 *err = bpf_prog_alloc_jited_linfo(fp); 2425 if (*err) 2426 return fp; 2427 2428 fp = bpf_int_jit_compile(fp); 2429 bpf_prog_jit_attempt_done(fp); 2430 if (!fp->jited && jit_needed) { 2431 *err = -ENOTSUPP; 2432 return fp; 2433 } 2434 } else { 2435 *err = bpf_prog_offload_compile(fp); 2436 if (*err) 2437 return fp; 2438 } 2439 2440 finalize: 2441 *err = bpf_prog_lock_ro(fp); 2442 if (*err) 2443 return fp; 2444 2445 /* The tail call compatibility check can only be done at 2446 * this late stage as we need to determine, if we deal 2447 * with JITed or non JITed program concatenations and not 2448 * all eBPF JITs might immediately support all features. 2449 */ 2450 *err = bpf_check_tail_call(fp); 2451 2452 return fp; 2453 } 2454 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2455 2456 static unsigned int __bpf_prog_ret1(const void *ctx, 2457 const struct bpf_insn *insn) 2458 { 2459 return 1; 2460 } 2461 2462 static struct bpf_prog_dummy { 2463 struct bpf_prog prog; 2464 } dummy_bpf_prog = { 2465 .prog = { 2466 .bpf_func = __bpf_prog_ret1, 2467 }, 2468 }; 2469 2470 struct bpf_empty_prog_array bpf_empty_prog_array = { 2471 .null_prog = NULL, 2472 }; 2473 EXPORT_SYMBOL(bpf_empty_prog_array); 2474 2475 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2476 { 2477 struct bpf_prog_array *p; 2478 2479 if (prog_cnt) 2480 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags); 2481 else 2482 p = &bpf_empty_prog_array.hdr; 2483 2484 return p; 2485 } 2486 2487 void bpf_prog_array_free(struct bpf_prog_array *progs) 2488 { 2489 if (!progs || progs == &bpf_empty_prog_array.hdr) 2490 return; 2491 kfree_rcu(progs, rcu); 2492 } 2493 2494 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2495 { 2496 struct bpf_prog_array *progs; 2497 2498 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2499 * no need to call kfree_rcu(), just call kfree() directly. 2500 */ 2501 progs = container_of(rcu, struct bpf_prog_array, rcu); 2502 if (rcu_trace_implies_rcu_gp()) 2503 kfree(progs); 2504 else 2505 kfree_rcu(progs, rcu); 2506 } 2507 2508 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2509 { 2510 if (!progs || progs == &bpf_empty_prog_array.hdr) 2511 return; 2512 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2513 } 2514 2515 int bpf_prog_array_length(struct bpf_prog_array *array) 2516 { 2517 struct bpf_prog_array_item *item; 2518 u32 cnt = 0; 2519 2520 for (item = array->items; item->prog; item++) 2521 if (item->prog != &dummy_bpf_prog.prog) 2522 cnt++; 2523 return cnt; 2524 } 2525 2526 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2527 { 2528 struct bpf_prog_array_item *item; 2529 2530 for (item = array->items; item->prog; item++) 2531 if (item->prog != &dummy_bpf_prog.prog) 2532 return false; 2533 return true; 2534 } 2535 2536 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2537 u32 *prog_ids, 2538 u32 request_cnt) 2539 { 2540 struct bpf_prog_array_item *item; 2541 int i = 0; 2542 2543 for (item = array->items; item->prog; item++) { 2544 if (item->prog == &dummy_bpf_prog.prog) 2545 continue; 2546 prog_ids[i] = item->prog->aux->id; 2547 if (++i == request_cnt) { 2548 item++; 2549 break; 2550 } 2551 } 2552 2553 return !!(item->prog); 2554 } 2555 2556 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2557 __u32 __user *prog_ids, u32 cnt) 2558 { 2559 unsigned long err = 0; 2560 bool nospc; 2561 u32 *ids; 2562 2563 /* users of this function are doing: 2564 * cnt = bpf_prog_array_length(); 2565 * if (cnt > 0) 2566 * bpf_prog_array_copy_to_user(..., cnt); 2567 * so below kcalloc doesn't need extra cnt > 0 check. 2568 */ 2569 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2570 if (!ids) 2571 return -ENOMEM; 2572 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2573 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2574 kfree(ids); 2575 if (err) 2576 return -EFAULT; 2577 if (nospc) 2578 return -ENOSPC; 2579 return 0; 2580 } 2581 2582 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2583 struct bpf_prog *old_prog) 2584 { 2585 struct bpf_prog_array_item *item; 2586 2587 for (item = array->items; item->prog; item++) 2588 if (item->prog == old_prog) { 2589 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2590 break; 2591 } 2592 } 2593 2594 /** 2595 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2596 * index into the program array with 2597 * a dummy no-op program. 2598 * @array: a bpf_prog_array 2599 * @index: the index of the program to replace 2600 * 2601 * Skips over dummy programs, by not counting them, when calculating 2602 * the position of the program to replace. 2603 * 2604 * Return: 2605 * * 0 - Success 2606 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2607 * * -ENOENT - Index out of range 2608 */ 2609 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2610 { 2611 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2612 } 2613 2614 /** 2615 * bpf_prog_array_update_at() - Updates the program at the given index 2616 * into the program array. 2617 * @array: a bpf_prog_array 2618 * @index: the index of the program to update 2619 * @prog: the program to insert into the array 2620 * 2621 * Skips over dummy programs, by not counting them, when calculating 2622 * the position of the program to update. 2623 * 2624 * Return: 2625 * * 0 - Success 2626 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2627 * * -ENOENT - Index out of range 2628 */ 2629 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2630 struct bpf_prog *prog) 2631 { 2632 struct bpf_prog_array_item *item; 2633 2634 if (unlikely(index < 0)) 2635 return -EINVAL; 2636 2637 for (item = array->items; item->prog; item++) { 2638 if (item->prog == &dummy_bpf_prog.prog) 2639 continue; 2640 if (!index) { 2641 WRITE_ONCE(item->prog, prog); 2642 return 0; 2643 } 2644 index--; 2645 } 2646 return -ENOENT; 2647 } 2648 2649 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2650 struct bpf_prog *exclude_prog, 2651 struct bpf_prog *include_prog, 2652 u64 bpf_cookie, 2653 struct bpf_prog_array **new_array) 2654 { 2655 int new_prog_cnt, carry_prog_cnt = 0; 2656 struct bpf_prog_array_item *existing, *new; 2657 struct bpf_prog_array *array; 2658 bool found_exclude = false; 2659 2660 /* Figure out how many existing progs we need to carry over to 2661 * the new array. 2662 */ 2663 if (old_array) { 2664 existing = old_array->items; 2665 for (; existing->prog; existing++) { 2666 if (existing->prog == exclude_prog) { 2667 found_exclude = true; 2668 continue; 2669 } 2670 if (existing->prog != &dummy_bpf_prog.prog) 2671 carry_prog_cnt++; 2672 if (existing->prog == include_prog) 2673 return -EEXIST; 2674 } 2675 } 2676 2677 if (exclude_prog && !found_exclude) 2678 return -ENOENT; 2679 2680 /* How many progs (not NULL) will be in the new array? */ 2681 new_prog_cnt = carry_prog_cnt; 2682 if (include_prog) 2683 new_prog_cnt += 1; 2684 2685 /* Do we have any prog (not NULL) in the new array? */ 2686 if (!new_prog_cnt) { 2687 *new_array = NULL; 2688 return 0; 2689 } 2690 2691 /* +1 as the end of prog_array is marked with NULL */ 2692 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2693 if (!array) 2694 return -ENOMEM; 2695 new = array->items; 2696 2697 /* Fill in the new prog array */ 2698 if (carry_prog_cnt) { 2699 existing = old_array->items; 2700 for (; existing->prog; existing++) { 2701 if (existing->prog == exclude_prog || 2702 existing->prog == &dummy_bpf_prog.prog) 2703 continue; 2704 2705 new->prog = existing->prog; 2706 new->bpf_cookie = existing->bpf_cookie; 2707 new++; 2708 } 2709 } 2710 if (include_prog) { 2711 new->prog = include_prog; 2712 new->bpf_cookie = bpf_cookie; 2713 new++; 2714 } 2715 new->prog = NULL; 2716 *new_array = array; 2717 return 0; 2718 } 2719 2720 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2721 u32 *prog_ids, u32 request_cnt, 2722 u32 *prog_cnt) 2723 { 2724 u32 cnt = 0; 2725 2726 if (array) 2727 cnt = bpf_prog_array_length(array); 2728 2729 *prog_cnt = cnt; 2730 2731 /* return early if user requested only program count or nothing to copy */ 2732 if (!request_cnt || !cnt) 2733 return 0; 2734 2735 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2736 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2737 : 0; 2738 } 2739 2740 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2741 struct bpf_map **used_maps, u32 len) 2742 { 2743 struct bpf_map *map; 2744 bool sleepable; 2745 u32 i; 2746 2747 sleepable = aux->prog->sleepable; 2748 for (i = 0; i < len; i++) { 2749 map = used_maps[i]; 2750 if (map->ops->map_poke_untrack) 2751 map->ops->map_poke_untrack(map, aux); 2752 if (sleepable) 2753 atomic64_dec(&map->sleepable_refcnt); 2754 bpf_map_put(map); 2755 } 2756 } 2757 2758 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2759 { 2760 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2761 kfree(aux->used_maps); 2762 } 2763 2764 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 2765 { 2766 #ifdef CONFIG_BPF_SYSCALL 2767 struct btf_mod_pair *btf_mod; 2768 u32 i; 2769 2770 for (i = 0; i < len; i++) { 2771 btf_mod = &used_btfs[i]; 2772 if (btf_mod->module) 2773 module_put(btf_mod->module); 2774 btf_put(btf_mod->btf); 2775 } 2776 #endif 2777 } 2778 2779 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2780 { 2781 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 2782 kfree(aux->used_btfs); 2783 } 2784 2785 static void bpf_prog_free_deferred(struct work_struct *work) 2786 { 2787 struct bpf_prog_aux *aux; 2788 int i; 2789 2790 aux = container_of(work, struct bpf_prog_aux, work); 2791 #ifdef CONFIG_BPF_SYSCALL 2792 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2793 #endif 2794 #ifdef CONFIG_CGROUP_BPF 2795 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2796 bpf_cgroup_atype_put(aux->cgroup_atype); 2797 #endif 2798 bpf_free_used_maps(aux); 2799 bpf_free_used_btfs(aux); 2800 if (bpf_prog_is_dev_bound(aux)) 2801 bpf_prog_dev_bound_destroy(aux->prog); 2802 #ifdef CONFIG_PERF_EVENTS 2803 if (aux->prog->has_callchain_buf) 2804 put_callchain_buffers(); 2805 #endif 2806 if (aux->dst_trampoline) 2807 bpf_trampoline_put(aux->dst_trampoline); 2808 for (i = 0; i < aux->real_func_cnt; i++) { 2809 /* We can just unlink the subprog poke descriptor table as 2810 * it was originally linked to the main program and is also 2811 * released along with it. 2812 */ 2813 aux->func[i]->aux->poke_tab = NULL; 2814 bpf_jit_free(aux->func[i]); 2815 } 2816 if (aux->real_func_cnt) { 2817 kfree(aux->func); 2818 bpf_prog_unlock_free(aux->prog); 2819 } else { 2820 bpf_jit_free(aux->prog); 2821 } 2822 } 2823 2824 void bpf_prog_free(struct bpf_prog *fp) 2825 { 2826 struct bpf_prog_aux *aux = fp->aux; 2827 2828 if (aux->dst_prog) 2829 bpf_prog_put(aux->dst_prog); 2830 bpf_token_put(aux->token); 2831 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2832 schedule_work(&aux->work); 2833 } 2834 EXPORT_SYMBOL_GPL(bpf_prog_free); 2835 2836 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 2837 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2838 2839 void bpf_user_rnd_init_once(void) 2840 { 2841 prandom_init_once(&bpf_user_rnd_state); 2842 } 2843 2844 BPF_CALL_0(bpf_user_rnd_u32) 2845 { 2846 /* Should someone ever have the rather unwise idea to use some 2847 * of the registers passed into this function, then note that 2848 * this function is called from native eBPF and classic-to-eBPF 2849 * transformations. Register assignments from both sides are 2850 * different, f.e. classic always sets fn(ctx, A, X) here. 2851 */ 2852 struct rnd_state *state; 2853 u32 res; 2854 2855 state = &get_cpu_var(bpf_user_rnd_state); 2856 res = prandom_u32_state(state); 2857 put_cpu_var(bpf_user_rnd_state); 2858 2859 return res; 2860 } 2861 2862 BPF_CALL_0(bpf_get_raw_cpu_id) 2863 { 2864 return raw_smp_processor_id(); 2865 } 2866 2867 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2868 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2869 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2870 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2871 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2872 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2873 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2874 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2875 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2876 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2877 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2878 2879 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2880 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2881 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2882 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2883 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2884 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2885 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 2886 2887 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2888 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2889 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2890 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2891 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2892 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2893 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2894 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2895 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2896 const struct bpf_func_proto bpf_set_retval_proto __weak; 2897 const struct bpf_func_proto bpf_get_retval_proto __weak; 2898 2899 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2900 { 2901 return NULL; 2902 } 2903 2904 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2905 { 2906 return NULL; 2907 } 2908 2909 u64 __weak 2910 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2911 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2912 { 2913 return -ENOTSUPP; 2914 } 2915 EXPORT_SYMBOL_GPL(bpf_event_output); 2916 2917 /* Always built-in helper functions. */ 2918 const struct bpf_func_proto bpf_tail_call_proto = { 2919 .func = NULL, 2920 .gpl_only = false, 2921 .ret_type = RET_VOID, 2922 .arg1_type = ARG_PTR_TO_CTX, 2923 .arg2_type = ARG_CONST_MAP_PTR, 2924 .arg3_type = ARG_ANYTHING, 2925 }; 2926 2927 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 2928 * It is encouraged to implement bpf_int_jit_compile() instead, so that 2929 * eBPF and implicitly also cBPF can get JITed! 2930 */ 2931 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 2932 { 2933 return prog; 2934 } 2935 2936 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 2937 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 2938 */ 2939 void __weak bpf_jit_compile(struct bpf_prog *prog) 2940 { 2941 } 2942 2943 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 2944 { 2945 return false; 2946 } 2947 2948 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 2949 * analysis code and wants explicit zero extension inserted by verifier. 2950 * Otherwise, return FALSE. 2951 * 2952 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 2953 * you don't override this. JITs that don't want these extra insns can detect 2954 * them using insn_is_zext. 2955 */ 2956 bool __weak bpf_jit_needs_zext(void) 2957 { 2958 return false; 2959 } 2960 2961 /* Return true if the JIT inlines the call to the helper corresponding to 2962 * the imm. 2963 * 2964 * The verifier will not patch the insn->imm for the call to the helper if 2965 * this returns true. 2966 */ 2967 bool __weak bpf_jit_inlines_helper_call(s32 imm) 2968 { 2969 return false; 2970 } 2971 2972 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 2973 bool __weak bpf_jit_supports_subprog_tailcalls(void) 2974 { 2975 return false; 2976 } 2977 2978 bool __weak bpf_jit_supports_percpu_insn(void) 2979 { 2980 return false; 2981 } 2982 2983 bool __weak bpf_jit_supports_kfunc_call(void) 2984 { 2985 return false; 2986 } 2987 2988 bool __weak bpf_jit_supports_far_kfunc_call(void) 2989 { 2990 return false; 2991 } 2992 2993 bool __weak bpf_jit_supports_arena(void) 2994 { 2995 return false; 2996 } 2997 2998 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 2999 { 3000 return false; 3001 } 3002 3003 u64 __weak bpf_arch_uaddress_limit(void) 3004 { 3005 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3006 return TASK_SIZE; 3007 #else 3008 return 0; 3009 #endif 3010 } 3011 3012 /* Return TRUE if the JIT backend satisfies the following two conditions: 3013 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3014 * 2) Under the specific arch, the implementation of xchg() is the same 3015 * as atomic_xchg() on pointer-sized words. 3016 */ 3017 bool __weak bpf_jit_supports_ptr_xchg(void) 3018 { 3019 return false; 3020 } 3021 3022 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3023 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3024 */ 3025 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3026 int len) 3027 { 3028 return -EFAULT; 3029 } 3030 3031 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 3032 void *addr1, void *addr2) 3033 { 3034 return -ENOTSUPP; 3035 } 3036 3037 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3038 { 3039 return ERR_PTR(-ENOTSUPP); 3040 } 3041 3042 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3043 { 3044 return -ENOTSUPP; 3045 } 3046 3047 bool __weak bpf_jit_supports_exceptions(void) 3048 { 3049 return false; 3050 } 3051 3052 bool __weak bpf_jit_supports_private_stack(void) 3053 { 3054 return false; 3055 } 3056 3057 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3058 { 3059 } 3060 3061 /* for configs without MMU or 32-bit */ 3062 __weak const struct bpf_map_ops arena_map_ops; 3063 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3064 { 3065 return 0; 3066 } 3067 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3068 { 3069 return 0; 3070 } 3071 3072 #ifdef CONFIG_BPF_SYSCALL 3073 static int __init bpf_global_ma_init(void) 3074 { 3075 int ret; 3076 3077 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3078 bpf_global_ma_set = !ret; 3079 return ret; 3080 } 3081 late_initcall(bpf_global_ma_init); 3082 #endif 3083 3084 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3085 EXPORT_SYMBOL(bpf_stats_enabled_key); 3086 3087 /* All definitions of tracepoints related to BPF. */ 3088 #define CREATE_TRACE_POINTS 3089 #include <linux/bpf_trace.h> 3090 3091 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3092 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3093