1 /* 2 * Linux Socket Filter - Kernel level socket filtering 3 * 4 * Based on the design of the Berkeley Packet Filter. The new 5 * internal format has been designed by PLUMgrid: 6 * 7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 8 * 9 * Authors: 10 * 11 * Jay Schulist <jschlst@samba.org> 12 * Alexei Starovoitov <ast@plumgrid.com> 13 * Daniel Borkmann <dborkman@redhat.com> 14 * 15 * This program is free software; you can redistribute it and/or 16 * modify it under the terms of the GNU General Public License 17 * as published by the Free Software Foundation; either version 18 * 2 of the License, or (at your option) any later version. 19 * 20 * Andi Kleen - Fix a few bad bugs and races. 21 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 22 */ 23 24 #include <linux/filter.h> 25 #include <linux/skbuff.h> 26 #include <linux/vmalloc.h> 27 #include <linux/random.h> 28 #include <linux/moduleloader.h> 29 #include <linux/bpf.h> 30 #include <linux/frame.h> 31 #include <linux/rbtree_latch.h> 32 #include <linux/kallsyms.h> 33 #include <linux/rcupdate.h> 34 35 #include <asm/unaligned.h> 36 37 /* Registers */ 38 #define BPF_R0 regs[BPF_REG_0] 39 #define BPF_R1 regs[BPF_REG_1] 40 #define BPF_R2 regs[BPF_REG_2] 41 #define BPF_R3 regs[BPF_REG_3] 42 #define BPF_R4 regs[BPF_REG_4] 43 #define BPF_R5 regs[BPF_REG_5] 44 #define BPF_R6 regs[BPF_REG_6] 45 #define BPF_R7 regs[BPF_REG_7] 46 #define BPF_R8 regs[BPF_REG_8] 47 #define BPF_R9 regs[BPF_REG_9] 48 #define BPF_R10 regs[BPF_REG_10] 49 50 /* Named registers */ 51 #define DST regs[insn->dst_reg] 52 #define SRC regs[insn->src_reg] 53 #define FP regs[BPF_REG_FP] 54 #define ARG1 regs[BPF_REG_ARG1] 55 #define CTX regs[BPF_REG_CTX] 56 #define IMM insn->imm 57 58 /* No hurry in this branch 59 * 60 * Exported for the bpf jit load helper. 61 */ 62 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 63 { 64 u8 *ptr = NULL; 65 66 if (k >= SKF_NET_OFF) 67 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 68 else if (k >= SKF_LL_OFF) 69 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 70 71 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 72 return ptr; 73 74 return NULL; 75 } 76 77 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 78 { 79 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 80 struct bpf_prog_aux *aux; 81 struct bpf_prog *fp; 82 83 size = round_up(size, PAGE_SIZE); 84 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 85 if (fp == NULL) 86 return NULL; 87 88 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags); 89 if (aux == NULL) { 90 vfree(fp); 91 return NULL; 92 } 93 94 fp->pages = size / PAGE_SIZE; 95 fp->aux = aux; 96 fp->aux->prog = fp; 97 fp->jit_requested = ebpf_jit_enabled(); 98 99 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode); 100 101 return fp; 102 } 103 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 104 105 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 106 gfp_t gfp_extra_flags) 107 { 108 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 109 struct bpf_prog *fp; 110 u32 pages, delta; 111 int ret; 112 113 BUG_ON(fp_old == NULL); 114 115 size = round_up(size, PAGE_SIZE); 116 pages = size / PAGE_SIZE; 117 if (pages <= fp_old->pages) 118 return fp_old; 119 120 delta = pages - fp_old->pages; 121 ret = __bpf_prog_charge(fp_old->aux->user, delta); 122 if (ret) 123 return NULL; 124 125 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 126 if (fp == NULL) { 127 __bpf_prog_uncharge(fp_old->aux->user, delta); 128 } else { 129 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 130 fp->pages = pages; 131 fp->aux->prog = fp; 132 133 /* We keep fp->aux from fp_old around in the new 134 * reallocated structure. 135 */ 136 fp_old->aux = NULL; 137 __bpf_prog_free(fp_old); 138 } 139 140 return fp; 141 } 142 143 void __bpf_prog_free(struct bpf_prog *fp) 144 { 145 kfree(fp->aux); 146 vfree(fp); 147 } 148 149 int bpf_prog_calc_tag(struct bpf_prog *fp) 150 { 151 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64); 152 u32 raw_size = bpf_prog_tag_scratch_size(fp); 153 u32 digest[SHA_DIGEST_WORDS]; 154 u32 ws[SHA_WORKSPACE_WORDS]; 155 u32 i, bsize, psize, blocks; 156 struct bpf_insn *dst; 157 bool was_ld_map; 158 u8 *raw, *todo; 159 __be32 *result; 160 __be64 *bits; 161 162 raw = vmalloc(raw_size); 163 if (!raw) 164 return -ENOMEM; 165 166 sha_init(digest); 167 memset(ws, 0, sizeof(ws)); 168 169 /* We need to take out the map fd for the digest calculation 170 * since they are unstable from user space side. 171 */ 172 dst = (void *)raw; 173 for (i = 0, was_ld_map = false; i < fp->len; i++) { 174 dst[i] = fp->insnsi[i]; 175 if (!was_ld_map && 176 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 177 dst[i].src_reg == BPF_PSEUDO_MAP_FD) { 178 was_ld_map = true; 179 dst[i].imm = 0; 180 } else if (was_ld_map && 181 dst[i].code == 0 && 182 dst[i].dst_reg == 0 && 183 dst[i].src_reg == 0 && 184 dst[i].off == 0) { 185 was_ld_map = false; 186 dst[i].imm = 0; 187 } else { 188 was_ld_map = false; 189 } 190 } 191 192 psize = bpf_prog_insn_size(fp); 193 memset(&raw[psize], 0, raw_size - psize); 194 raw[psize++] = 0x80; 195 196 bsize = round_up(psize, SHA_MESSAGE_BYTES); 197 blocks = bsize / SHA_MESSAGE_BYTES; 198 todo = raw; 199 if (bsize - psize >= sizeof(__be64)) { 200 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 201 } else { 202 bits = (__be64 *)(todo + bsize + bits_offset); 203 blocks++; 204 } 205 *bits = cpu_to_be64((psize - 1) << 3); 206 207 while (blocks--) { 208 sha_transform(digest, todo, ws); 209 todo += SHA_MESSAGE_BYTES; 210 } 211 212 result = (__force __be32 *)digest; 213 for (i = 0; i < SHA_DIGEST_WORDS; i++) 214 result[i] = cpu_to_be32(digest[i]); 215 memcpy(fp->tag, result, sizeof(fp->tag)); 216 217 vfree(raw); 218 return 0; 219 } 220 221 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta) 222 { 223 struct bpf_insn *insn = prog->insnsi; 224 u32 i, insn_cnt = prog->len; 225 bool pseudo_call; 226 u8 code; 227 int off; 228 229 for (i = 0; i < insn_cnt; i++, insn++) { 230 code = insn->code; 231 if (BPF_CLASS(code) != BPF_JMP) 232 continue; 233 if (BPF_OP(code) == BPF_EXIT) 234 continue; 235 if (BPF_OP(code) == BPF_CALL) { 236 if (insn->src_reg == BPF_PSEUDO_CALL) 237 pseudo_call = true; 238 else 239 continue; 240 } else { 241 pseudo_call = false; 242 } 243 off = pseudo_call ? insn->imm : insn->off; 244 245 /* Adjust offset of jmps if we cross boundaries. */ 246 if (i < pos && i + off + 1 > pos) 247 off += delta; 248 else if (i > pos + delta && i + off + 1 <= pos + delta) 249 off -= delta; 250 251 if (pseudo_call) 252 insn->imm = off; 253 else 254 insn->off = off; 255 } 256 } 257 258 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 259 const struct bpf_insn *patch, u32 len) 260 { 261 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 262 struct bpf_prog *prog_adj; 263 264 /* Since our patchlet doesn't expand the image, we're done. */ 265 if (insn_delta == 0) { 266 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 267 return prog; 268 } 269 270 insn_adj_cnt = prog->len + insn_delta; 271 272 /* Several new instructions need to be inserted. Make room 273 * for them. Likely, there's no need for a new allocation as 274 * last page could have large enough tailroom. 275 */ 276 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 277 GFP_USER); 278 if (!prog_adj) 279 return NULL; 280 281 prog_adj->len = insn_adj_cnt; 282 283 /* Patching happens in 3 steps: 284 * 285 * 1) Move over tail of insnsi from next instruction onwards, 286 * so we can patch the single target insn with one or more 287 * new ones (patching is always from 1 to n insns, n > 0). 288 * 2) Inject new instructions at the target location. 289 * 3) Adjust branch offsets if necessary. 290 */ 291 insn_rest = insn_adj_cnt - off - len; 292 293 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 294 sizeof(*patch) * insn_rest); 295 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 296 297 bpf_adj_branches(prog_adj, off, insn_delta); 298 299 return prog_adj; 300 } 301 302 #ifdef CONFIG_BPF_JIT 303 /* All BPF JIT sysctl knobs here. */ 304 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_ALWAYS_ON); 305 int bpf_jit_harden __read_mostly; 306 int bpf_jit_kallsyms __read_mostly; 307 308 static __always_inline void 309 bpf_get_prog_addr_region(const struct bpf_prog *prog, 310 unsigned long *symbol_start, 311 unsigned long *symbol_end) 312 { 313 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 314 unsigned long addr = (unsigned long)hdr; 315 316 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 317 318 *symbol_start = addr; 319 *symbol_end = addr + hdr->pages * PAGE_SIZE; 320 } 321 322 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym) 323 { 324 const char *end = sym + KSYM_NAME_LEN; 325 326 BUILD_BUG_ON(sizeof("bpf_prog_") + 327 sizeof(prog->tag) * 2 + 328 /* name has been null terminated. 329 * We should need +1 for the '_' preceding 330 * the name. However, the null character 331 * is double counted between the name and the 332 * sizeof("bpf_prog_") above, so we omit 333 * the +1 here. 334 */ 335 sizeof(prog->aux->name) > KSYM_NAME_LEN); 336 337 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 338 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 339 if (prog->aux->name[0]) 340 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 341 else 342 *sym = 0; 343 } 344 345 static __always_inline unsigned long 346 bpf_get_prog_addr_start(struct latch_tree_node *n) 347 { 348 unsigned long symbol_start, symbol_end; 349 const struct bpf_prog_aux *aux; 350 351 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 352 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 353 354 return symbol_start; 355 } 356 357 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 358 struct latch_tree_node *b) 359 { 360 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b); 361 } 362 363 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 364 { 365 unsigned long val = (unsigned long)key; 366 unsigned long symbol_start, symbol_end; 367 const struct bpf_prog_aux *aux; 368 369 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 370 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 371 372 if (val < symbol_start) 373 return -1; 374 if (val >= symbol_end) 375 return 1; 376 377 return 0; 378 } 379 380 static const struct latch_tree_ops bpf_tree_ops = { 381 .less = bpf_tree_less, 382 .comp = bpf_tree_comp, 383 }; 384 385 static DEFINE_SPINLOCK(bpf_lock); 386 static LIST_HEAD(bpf_kallsyms); 387 static struct latch_tree_root bpf_tree __cacheline_aligned; 388 389 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux) 390 { 391 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode)); 392 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms); 393 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 394 } 395 396 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux) 397 { 398 if (list_empty(&aux->ksym_lnode)) 399 return; 400 401 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 402 list_del_rcu(&aux->ksym_lnode); 403 } 404 405 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 406 { 407 return fp->jited && !bpf_prog_was_classic(fp); 408 } 409 410 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 411 { 412 return list_empty(&fp->aux->ksym_lnode) || 413 fp->aux->ksym_lnode.prev == LIST_POISON2; 414 } 415 416 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 417 { 418 if (!bpf_prog_kallsyms_candidate(fp) || 419 !capable(CAP_SYS_ADMIN)) 420 return; 421 422 spin_lock_bh(&bpf_lock); 423 bpf_prog_ksym_node_add(fp->aux); 424 spin_unlock_bh(&bpf_lock); 425 } 426 427 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 428 { 429 if (!bpf_prog_kallsyms_candidate(fp)) 430 return; 431 432 spin_lock_bh(&bpf_lock); 433 bpf_prog_ksym_node_del(fp->aux); 434 spin_unlock_bh(&bpf_lock); 435 } 436 437 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr) 438 { 439 struct latch_tree_node *n; 440 441 if (!bpf_jit_kallsyms_enabled()) 442 return NULL; 443 444 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 445 return n ? 446 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 447 NULL; 448 } 449 450 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 451 unsigned long *off, char *sym) 452 { 453 unsigned long symbol_start, symbol_end; 454 struct bpf_prog *prog; 455 char *ret = NULL; 456 457 rcu_read_lock(); 458 prog = bpf_prog_kallsyms_find(addr); 459 if (prog) { 460 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 461 bpf_get_prog_name(prog, sym); 462 463 ret = sym; 464 if (size) 465 *size = symbol_end - symbol_start; 466 if (off) 467 *off = addr - symbol_start; 468 } 469 rcu_read_unlock(); 470 471 return ret; 472 } 473 474 bool is_bpf_text_address(unsigned long addr) 475 { 476 bool ret; 477 478 rcu_read_lock(); 479 ret = bpf_prog_kallsyms_find(addr) != NULL; 480 rcu_read_unlock(); 481 482 return ret; 483 } 484 485 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 486 char *sym) 487 { 488 unsigned long symbol_start, symbol_end; 489 struct bpf_prog_aux *aux; 490 unsigned int it = 0; 491 int ret = -ERANGE; 492 493 if (!bpf_jit_kallsyms_enabled()) 494 return ret; 495 496 rcu_read_lock(); 497 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 498 if (it++ != symnum) 499 continue; 500 501 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 502 bpf_get_prog_name(aux->prog, sym); 503 504 *value = symbol_start; 505 *type = BPF_SYM_ELF_TYPE; 506 507 ret = 0; 508 break; 509 } 510 rcu_read_unlock(); 511 512 return ret; 513 } 514 515 struct bpf_binary_header * 516 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 517 unsigned int alignment, 518 bpf_jit_fill_hole_t bpf_fill_ill_insns) 519 { 520 struct bpf_binary_header *hdr; 521 unsigned int size, hole, start; 522 523 /* Most of BPF filters are really small, but if some of them 524 * fill a page, allow at least 128 extra bytes to insert a 525 * random section of illegal instructions. 526 */ 527 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 528 hdr = module_alloc(size); 529 if (hdr == NULL) 530 return NULL; 531 532 /* Fill space with illegal/arch-dep instructions. */ 533 bpf_fill_ill_insns(hdr, size); 534 535 hdr->pages = size / PAGE_SIZE; 536 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 537 PAGE_SIZE - sizeof(*hdr)); 538 start = (get_random_int() % hole) & ~(alignment - 1); 539 540 /* Leave a random number of instructions before BPF code. */ 541 *image_ptr = &hdr->image[start]; 542 543 return hdr; 544 } 545 546 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 547 { 548 module_memfree(hdr); 549 } 550 551 /* This symbol is only overridden by archs that have different 552 * requirements than the usual eBPF JITs, f.e. when they only 553 * implement cBPF JIT, do not set images read-only, etc. 554 */ 555 void __weak bpf_jit_free(struct bpf_prog *fp) 556 { 557 if (fp->jited) { 558 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 559 560 bpf_jit_binary_unlock_ro(hdr); 561 bpf_jit_binary_free(hdr); 562 563 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 564 } 565 566 bpf_prog_unlock_free(fp); 567 } 568 569 static int bpf_jit_blind_insn(const struct bpf_insn *from, 570 const struct bpf_insn *aux, 571 struct bpf_insn *to_buff) 572 { 573 struct bpf_insn *to = to_buff; 574 u32 imm_rnd = get_random_int(); 575 s16 off; 576 577 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 578 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 579 580 if (from->imm == 0 && 581 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 582 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 583 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 584 goto out; 585 } 586 587 switch (from->code) { 588 case BPF_ALU | BPF_ADD | BPF_K: 589 case BPF_ALU | BPF_SUB | BPF_K: 590 case BPF_ALU | BPF_AND | BPF_K: 591 case BPF_ALU | BPF_OR | BPF_K: 592 case BPF_ALU | BPF_XOR | BPF_K: 593 case BPF_ALU | BPF_MUL | BPF_K: 594 case BPF_ALU | BPF_MOV | BPF_K: 595 case BPF_ALU | BPF_DIV | BPF_K: 596 case BPF_ALU | BPF_MOD | BPF_K: 597 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 598 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 599 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 600 break; 601 602 case BPF_ALU64 | BPF_ADD | BPF_K: 603 case BPF_ALU64 | BPF_SUB | BPF_K: 604 case BPF_ALU64 | BPF_AND | BPF_K: 605 case BPF_ALU64 | BPF_OR | BPF_K: 606 case BPF_ALU64 | BPF_XOR | BPF_K: 607 case BPF_ALU64 | BPF_MUL | BPF_K: 608 case BPF_ALU64 | BPF_MOV | BPF_K: 609 case BPF_ALU64 | BPF_DIV | BPF_K: 610 case BPF_ALU64 | BPF_MOD | BPF_K: 611 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 612 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 613 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 614 break; 615 616 case BPF_JMP | BPF_JEQ | BPF_K: 617 case BPF_JMP | BPF_JNE | BPF_K: 618 case BPF_JMP | BPF_JGT | BPF_K: 619 case BPF_JMP | BPF_JLT | BPF_K: 620 case BPF_JMP | BPF_JGE | BPF_K: 621 case BPF_JMP | BPF_JLE | BPF_K: 622 case BPF_JMP | BPF_JSGT | BPF_K: 623 case BPF_JMP | BPF_JSLT | BPF_K: 624 case BPF_JMP | BPF_JSGE | BPF_K: 625 case BPF_JMP | BPF_JSLE | BPF_K: 626 case BPF_JMP | BPF_JSET | BPF_K: 627 /* Accommodate for extra offset in case of a backjump. */ 628 off = from->off; 629 if (off < 0) 630 off -= 2; 631 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 632 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 633 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 634 break; 635 636 case BPF_LD | BPF_ABS | BPF_W: 637 case BPF_LD | BPF_ABS | BPF_H: 638 case BPF_LD | BPF_ABS | BPF_B: 639 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 640 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 641 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0); 642 break; 643 644 case BPF_LD | BPF_IND | BPF_W: 645 case BPF_LD | BPF_IND | BPF_H: 646 case BPF_LD | BPF_IND | BPF_B: 647 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 648 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 649 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg); 650 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0); 651 break; 652 653 case BPF_LD | BPF_IMM | BPF_DW: 654 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 655 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 656 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 657 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 658 break; 659 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 660 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 661 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 662 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 663 break; 664 665 case BPF_ST | BPF_MEM | BPF_DW: 666 case BPF_ST | BPF_MEM | BPF_W: 667 case BPF_ST | BPF_MEM | BPF_H: 668 case BPF_ST | BPF_MEM | BPF_B: 669 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 670 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 671 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 672 break; 673 } 674 out: 675 return to - to_buff; 676 } 677 678 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 679 gfp_t gfp_extra_flags) 680 { 681 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 682 struct bpf_prog *fp; 683 684 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 685 if (fp != NULL) { 686 /* aux->prog still points to the fp_other one, so 687 * when promoting the clone to the real program, 688 * this still needs to be adapted. 689 */ 690 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 691 } 692 693 return fp; 694 } 695 696 static void bpf_prog_clone_free(struct bpf_prog *fp) 697 { 698 /* aux was stolen by the other clone, so we cannot free 699 * it from this path! It will be freed eventually by the 700 * other program on release. 701 * 702 * At this point, we don't need a deferred release since 703 * clone is guaranteed to not be locked. 704 */ 705 fp->aux = NULL; 706 __bpf_prog_free(fp); 707 } 708 709 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 710 { 711 /* We have to repoint aux->prog to self, as we don't 712 * know whether fp here is the clone or the original. 713 */ 714 fp->aux->prog = fp; 715 bpf_prog_clone_free(fp_other); 716 } 717 718 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 719 { 720 struct bpf_insn insn_buff[16], aux[2]; 721 struct bpf_prog *clone, *tmp; 722 int insn_delta, insn_cnt; 723 struct bpf_insn *insn; 724 int i, rewritten; 725 726 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 727 return prog; 728 729 clone = bpf_prog_clone_create(prog, GFP_USER); 730 if (!clone) 731 return ERR_PTR(-ENOMEM); 732 733 insn_cnt = clone->len; 734 insn = clone->insnsi; 735 736 for (i = 0; i < insn_cnt; i++, insn++) { 737 /* We temporarily need to hold the original ld64 insn 738 * so that we can still access the first part in the 739 * second blinding run. 740 */ 741 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 742 insn[1].code == 0) 743 memcpy(aux, insn, sizeof(aux)); 744 745 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff); 746 if (!rewritten) 747 continue; 748 749 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 750 if (!tmp) { 751 /* Patching may have repointed aux->prog during 752 * realloc from the original one, so we need to 753 * fix it up here on error. 754 */ 755 bpf_jit_prog_release_other(prog, clone); 756 return ERR_PTR(-ENOMEM); 757 } 758 759 clone = tmp; 760 insn_delta = rewritten - 1; 761 762 /* Walk new program and skip insns we just inserted. */ 763 insn = clone->insnsi + i + insn_delta; 764 insn_cnt += insn_delta; 765 i += insn_delta; 766 } 767 768 clone->blinded = 1; 769 return clone; 770 } 771 #endif /* CONFIG_BPF_JIT */ 772 773 /* Base function for offset calculation. Needs to go into .text section, 774 * therefore keeping it non-static as well; will also be used by JITs 775 * anyway later on, so do not let the compiler omit it. This also needs 776 * to go into kallsyms for correlation from e.g. bpftool, so naming 777 * must not change. 778 */ 779 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 780 { 781 return 0; 782 } 783 EXPORT_SYMBOL_GPL(__bpf_call_base); 784 785 /* All UAPI available opcodes. */ 786 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 787 /* 32 bit ALU operations. */ \ 788 /* Register based. */ \ 789 INSN_3(ALU, ADD, X), \ 790 INSN_3(ALU, SUB, X), \ 791 INSN_3(ALU, AND, X), \ 792 INSN_3(ALU, OR, X), \ 793 INSN_3(ALU, LSH, X), \ 794 INSN_3(ALU, RSH, X), \ 795 INSN_3(ALU, XOR, X), \ 796 INSN_3(ALU, MUL, X), \ 797 INSN_3(ALU, MOV, X), \ 798 INSN_3(ALU, DIV, X), \ 799 INSN_3(ALU, MOD, X), \ 800 INSN_2(ALU, NEG), \ 801 INSN_3(ALU, END, TO_BE), \ 802 INSN_3(ALU, END, TO_LE), \ 803 /* Immediate based. */ \ 804 INSN_3(ALU, ADD, K), \ 805 INSN_3(ALU, SUB, K), \ 806 INSN_3(ALU, AND, K), \ 807 INSN_3(ALU, OR, K), \ 808 INSN_3(ALU, LSH, K), \ 809 INSN_3(ALU, RSH, K), \ 810 INSN_3(ALU, XOR, K), \ 811 INSN_3(ALU, MUL, K), \ 812 INSN_3(ALU, MOV, K), \ 813 INSN_3(ALU, DIV, K), \ 814 INSN_3(ALU, MOD, K), \ 815 /* 64 bit ALU operations. */ \ 816 /* Register based. */ \ 817 INSN_3(ALU64, ADD, X), \ 818 INSN_3(ALU64, SUB, X), \ 819 INSN_3(ALU64, AND, X), \ 820 INSN_3(ALU64, OR, X), \ 821 INSN_3(ALU64, LSH, X), \ 822 INSN_3(ALU64, RSH, X), \ 823 INSN_3(ALU64, XOR, X), \ 824 INSN_3(ALU64, MUL, X), \ 825 INSN_3(ALU64, MOV, X), \ 826 INSN_3(ALU64, ARSH, X), \ 827 INSN_3(ALU64, DIV, X), \ 828 INSN_3(ALU64, MOD, X), \ 829 INSN_2(ALU64, NEG), \ 830 /* Immediate based. */ \ 831 INSN_3(ALU64, ADD, K), \ 832 INSN_3(ALU64, SUB, K), \ 833 INSN_3(ALU64, AND, K), \ 834 INSN_3(ALU64, OR, K), \ 835 INSN_3(ALU64, LSH, K), \ 836 INSN_3(ALU64, RSH, K), \ 837 INSN_3(ALU64, XOR, K), \ 838 INSN_3(ALU64, MUL, K), \ 839 INSN_3(ALU64, MOV, K), \ 840 INSN_3(ALU64, ARSH, K), \ 841 INSN_3(ALU64, DIV, K), \ 842 INSN_3(ALU64, MOD, K), \ 843 /* Call instruction. */ \ 844 INSN_2(JMP, CALL), \ 845 /* Exit instruction. */ \ 846 INSN_2(JMP, EXIT), \ 847 /* Jump instructions. */ \ 848 /* Register based. */ \ 849 INSN_3(JMP, JEQ, X), \ 850 INSN_3(JMP, JNE, X), \ 851 INSN_3(JMP, JGT, X), \ 852 INSN_3(JMP, JLT, X), \ 853 INSN_3(JMP, JGE, X), \ 854 INSN_3(JMP, JLE, X), \ 855 INSN_3(JMP, JSGT, X), \ 856 INSN_3(JMP, JSLT, X), \ 857 INSN_3(JMP, JSGE, X), \ 858 INSN_3(JMP, JSLE, X), \ 859 INSN_3(JMP, JSET, X), \ 860 /* Immediate based. */ \ 861 INSN_3(JMP, JEQ, K), \ 862 INSN_3(JMP, JNE, K), \ 863 INSN_3(JMP, JGT, K), \ 864 INSN_3(JMP, JLT, K), \ 865 INSN_3(JMP, JGE, K), \ 866 INSN_3(JMP, JLE, K), \ 867 INSN_3(JMP, JSGT, K), \ 868 INSN_3(JMP, JSLT, K), \ 869 INSN_3(JMP, JSGE, K), \ 870 INSN_3(JMP, JSLE, K), \ 871 INSN_3(JMP, JSET, K), \ 872 INSN_2(JMP, JA), \ 873 /* Store instructions. */ \ 874 /* Register based. */ \ 875 INSN_3(STX, MEM, B), \ 876 INSN_3(STX, MEM, H), \ 877 INSN_3(STX, MEM, W), \ 878 INSN_3(STX, MEM, DW), \ 879 INSN_3(STX, XADD, W), \ 880 INSN_3(STX, XADD, DW), \ 881 /* Immediate based. */ \ 882 INSN_3(ST, MEM, B), \ 883 INSN_3(ST, MEM, H), \ 884 INSN_3(ST, MEM, W), \ 885 INSN_3(ST, MEM, DW), \ 886 /* Load instructions. */ \ 887 /* Register based. */ \ 888 INSN_3(LDX, MEM, B), \ 889 INSN_3(LDX, MEM, H), \ 890 INSN_3(LDX, MEM, W), \ 891 INSN_3(LDX, MEM, DW), \ 892 /* Immediate based. */ \ 893 INSN_3(LD, IMM, DW), \ 894 /* Misc (old cBPF carry-over). */ \ 895 INSN_3(LD, ABS, B), \ 896 INSN_3(LD, ABS, H), \ 897 INSN_3(LD, ABS, W), \ 898 INSN_3(LD, IND, B), \ 899 INSN_3(LD, IND, H), \ 900 INSN_3(LD, IND, W) 901 902 bool bpf_opcode_in_insntable(u8 code) 903 { 904 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 905 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 906 static const bool public_insntable[256] = { 907 [0 ... 255] = false, 908 /* Now overwrite non-defaults ... */ 909 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 910 }; 911 #undef BPF_INSN_3_TBL 912 #undef BPF_INSN_2_TBL 913 return public_insntable[code]; 914 } 915 916 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 917 /** 918 * __bpf_prog_run - run eBPF program on a given context 919 * @ctx: is the data we are operating on 920 * @insn: is the array of eBPF instructions 921 * 922 * Decode and execute eBPF instructions. 923 */ 924 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 925 { 926 u64 tmp; 927 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 928 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 929 static const void *jumptable[256] = { 930 [0 ... 255] = &&default_label, 931 /* Now overwrite non-defaults ... */ 932 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 933 /* Non-UAPI available opcodes. */ 934 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 935 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 936 }; 937 #undef BPF_INSN_3_LBL 938 #undef BPF_INSN_2_LBL 939 u32 tail_call_cnt = 0; 940 void *ptr; 941 int off; 942 943 #define CONT ({ insn++; goto select_insn; }) 944 #define CONT_JMP ({ insn++; goto select_insn; }) 945 946 select_insn: 947 goto *jumptable[insn->code]; 948 949 /* ALU */ 950 #define ALU(OPCODE, OP) \ 951 ALU64_##OPCODE##_X: \ 952 DST = DST OP SRC; \ 953 CONT; \ 954 ALU_##OPCODE##_X: \ 955 DST = (u32) DST OP (u32) SRC; \ 956 CONT; \ 957 ALU64_##OPCODE##_K: \ 958 DST = DST OP IMM; \ 959 CONT; \ 960 ALU_##OPCODE##_K: \ 961 DST = (u32) DST OP (u32) IMM; \ 962 CONT; 963 964 ALU(ADD, +) 965 ALU(SUB, -) 966 ALU(AND, &) 967 ALU(OR, |) 968 ALU(LSH, <<) 969 ALU(RSH, >>) 970 ALU(XOR, ^) 971 ALU(MUL, *) 972 #undef ALU 973 ALU_NEG: 974 DST = (u32) -DST; 975 CONT; 976 ALU64_NEG: 977 DST = -DST; 978 CONT; 979 ALU_MOV_X: 980 DST = (u32) SRC; 981 CONT; 982 ALU_MOV_K: 983 DST = (u32) IMM; 984 CONT; 985 ALU64_MOV_X: 986 DST = SRC; 987 CONT; 988 ALU64_MOV_K: 989 DST = IMM; 990 CONT; 991 LD_IMM_DW: 992 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 993 insn++; 994 CONT; 995 ALU64_ARSH_X: 996 (*(s64 *) &DST) >>= SRC; 997 CONT; 998 ALU64_ARSH_K: 999 (*(s64 *) &DST) >>= IMM; 1000 CONT; 1001 ALU64_MOD_X: 1002 div64_u64_rem(DST, SRC, &tmp); 1003 DST = tmp; 1004 CONT; 1005 ALU_MOD_X: 1006 tmp = (u32) DST; 1007 DST = do_div(tmp, (u32) SRC); 1008 CONT; 1009 ALU64_MOD_K: 1010 div64_u64_rem(DST, IMM, &tmp); 1011 DST = tmp; 1012 CONT; 1013 ALU_MOD_K: 1014 tmp = (u32) DST; 1015 DST = do_div(tmp, (u32) IMM); 1016 CONT; 1017 ALU64_DIV_X: 1018 DST = div64_u64(DST, SRC); 1019 CONT; 1020 ALU_DIV_X: 1021 tmp = (u32) DST; 1022 do_div(tmp, (u32) SRC); 1023 DST = (u32) tmp; 1024 CONT; 1025 ALU64_DIV_K: 1026 DST = div64_u64(DST, IMM); 1027 CONT; 1028 ALU_DIV_K: 1029 tmp = (u32) DST; 1030 do_div(tmp, (u32) IMM); 1031 DST = (u32) tmp; 1032 CONT; 1033 ALU_END_TO_BE: 1034 switch (IMM) { 1035 case 16: 1036 DST = (__force u16) cpu_to_be16(DST); 1037 break; 1038 case 32: 1039 DST = (__force u32) cpu_to_be32(DST); 1040 break; 1041 case 64: 1042 DST = (__force u64) cpu_to_be64(DST); 1043 break; 1044 } 1045 CONT; 1046 ALU_END_TO_LE: 1047 switch (IMM) { 1048 case 16: 1049 DST = (__force u16) cpu_to_le16(DST); 1050 break; 1051 case 32: 1052 DST = (__force u32) cpu_to_le32(DST); 1053 break; 1054 case 64: 1055 DST = (__force u64) cpu_to_le64(DST); 1056 break; 1057 } 1058 CONT; 1059 1060 /* CALL */ 1061 JMP_CALL: 1062 /* Function call scratches BPF_R1-BPF_R5 registers, 1063 * preserves BPF_R6-BPF_R9, and stores return value 1064 * into BPF_R0. 1065 */ 1066 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1067 BPF_R4, BPF_R5); 1068 CONT; 1069 1070 JMP_CALL_ARGS: 1071 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1072 BPF_R3, BPF_R4, 1073 BPF_R5, 1074 insn + insn->off + 1); 1075 CONT; 1076 1077 JMP_TAIL_CALL: { 1078 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1079 struct bpf_array *array = container_of(map, struct bpf_array, map); 1080 struct bpf_prog *prog; 1081 u32 index = BPF_R3; 1082 1083 if (unlikely(index >= array->map.max_entries)) 1084 goto out; 1085 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1086 goto out; 1087 1088 tail_call_cnt++; 1089 1090 prog = READ_ONCE(array->ptrs[index]); 1091 if (!prog) 1092 goto out; 1093 1094 /* ARG1 at this point is guaranteed to point to CTX from 1095 * the verifier side due to the fact that the tail call is 1096 * handeled like a helper, that is, bpf_tail_call_proto, 1097 * where arg1_type is ARG_PTR_TO_CTX. 1098 */ 1099 insn = prog->insnsi; 1100 goto select_insn; 1101 out: 1102 CONT; 1103 } 1104 /* JMP */ 1105 JMP_JA: 1106 insn += insn->off; 1107 CONT; 1108 JMP_JEQ_X: 1109 if (DST == SRC) { 1110 insn += insn->off; 1111 CONT_JMP; 1112 } 1113 CONT; 1114 JMP_JEQ_K: 1115 if (DST == IMM) { 1116 insn += insn->off; 1117 CONT_JMP; 1118 } 1119 CONT; 1120 JMP_JNE_X: 1121 if (DST != SRC) { 1122 insn += insn->off; 1123 CONT_JMP; 1124 } 1125 CONT; 1126 JMP_JNE_K: 1127 if (DST != IMM) { 1128 insn += insn->off; 1129 CONT_JMP; 1130 } 1131 CONT; 1132 JMP_JGT_X: 1133 if (DST > SRC) { 1134 insn += insn->off; 1135 CONT_JMP; 1136 } 1137 CONT; 1138 JMP_JGT_K: 1139 if (DST > IMM) { 1140 insn += insn->off; 1141 CONT_JMP; 1142 } 1143 CONT; 1144 JMP_JLT_X: 1145 if (DST < SRC) { 1146 insn += insn->off; 1147 CONT_JMP; 1148 } 1149 CONT; 1150 JMP_JLT_K: 1151 if (DST < IMM) { 1152 insn += insn->off; 1153 CONT_JMP; 1154 } 1155 CONT; 1156 JMP_JGE_X: 1157 if (DST >= SRC) { 1158 insn += insn->off; 1159 CONT_JMP; 1160 } 1161 CONT; 1162 JMP_JGE_K: 1163 if (DST >= IMM) { 1164 insn += insn->off; 1165 CONT_JMP; 1166 } 1167 CONT; 1168 JMP_JLE_X: 1169 if (DST <= SRC) { 1170 insn += insn->off; 1171 CONT_JMP; 1172 } 1173 CONT; 1174 JMP_JLE_K: 1175 if (DST <= IMM) { 1176 insn += insn->off; 1177 CONT_JMP; 1178 } 1179 CONT; 1180 JMP_JSGT_X: 1181 if (((s64) DST) > ((s64) SRC)) { 1182 insn += insn->off; 1183 CONT_JMP; 1184 } 1185 CONT; 1186 JMP_JSGT_K: 1187 if (((s64) DST) > ((s64) IMM)) { 1188 insn += insn->off; 1189 CONT_JMP; 1190 } 1191 CONT; 1192 JMP_JSLT_X: 1193 if (((s64) DST) < ((s64) SRC)) { 1194 insn += insn->off; 1195 CONT_JMP; 1196 } 1197 CONT; 1198 JMP_JSLT_K: 1199 if (((s64) DST) < ((s64) IMM)) { 1200 insn += insn->off; 1201 CONT_JMP; 1202 } 1203 CONT; 1204 JMP_JSGE_X: 1205 if (((s64) DST) >= ((s64) SRC)) { 1206 insn += insn->off; 1207 CONT_JMP; 1208 } 1209 CONT; 1210 JMP_JSGE_K: 1211 if (((s64) DST) >= ((s64) IMM)) { 1212 insn += insn->off; 1213 CONT_JMP; 1214 } 1215 CONT; 1216 JMP_JSLE_X: 1217 if (((s64) DST) <= ((s64) SRC)) { 1218 insn += insn->off; 1219 CONT_JMP; 1220 } 1221 CONT; 1222 JMP_JSLE_K: 1223 if (((s64) DST) <= ((s64) IMM)) { 1224 insn += insn->off; 1225 CONT_JMP; 1226 } 1227 CONT; 1228 JMP_JSET_X: 1229 if (DST & SRC) { 1230 insn += insn->off; 1231 CONT_JMP; 1232 } 1233 CONT; 1234 JMP_JSET_K: 1235 if (DST & IMM) { 1236 insn += insn->off; 1237 CONT_JMP; 1238 } 1239 CONT; 1240 JMP_EXIT: 1241 return BPF_R0; 1242 1243 /* STX and ST and LDX*/ 1244 #define LDST(SIZEOP, SIZE) \ 1245 STX_MEM_##SIZEOP: \ 1246 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1247 CONT; \ 1248 ST_MEM_##SIZEOP: \ 1249 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1250 CONT; \ 1251 LDX_MEM_##SIZEOP: \ 1252 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1253 CONT; 1254 1255 LDST(B, u8) 1256 LDST(H, u16) 1257 LDST(W, u32) 1258 LDST(DW, u64) 1259 #undef LDST 1260 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1261 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1262 (DST + insn->off)); 1263 CONT; 1264 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1265 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1266 (DST + insn->off)); 1267 CONT; 1268 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */ 1269 off = IMM; 1270 load_word: 1271 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only 1272 * appearing in the programs where ctx == skb 1273 * (see may_access_skb() in the verifier). All programs 1274 * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6, 1275 * bpf_convert_filter() saves it in BPF_R6, internal BPF 1276 * verifier will check that BPF_R6 == ctx. 1277 * 1278 * BPF_ABS and BPF_IND are wrappers of function calls, 1279 * so they scratch BPF_R1-BPF_R5 registers, preserve 1280 * BPF_R6-BPF_R9, and store return value into BPF_R0. 1281 * 1282 * Implicit input: 1283 * ctx == skb == BPF_R6 == CTX 1284 * 1285 * Explicit input: 1286 * SRC == any register 1287 * IMM == 32-bit immediate 1288 * 1289 * Output: 1290 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness 1291 */ 1292 1293 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp); 1294 if (likely(ptr != NULL)) { 1295 BPF_R0 = get_unaligned_be32(ptr); 1296 CONT; 1297 } 1298 1299 return 0; 1300 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */ 1301 off = IMM; 1302 load_half: 1303 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp); 1304 if (likely(ptr != NULL)) { 1305 BPF_R0 = get_unaligned_be16(ptr); 1306 CONT; 1307 } 1308 1309 return 0; 1310 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */ 1311 off = IMM; 1312 load_byte: 1313 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp); 1314 if (likely(ptr != NULL)) { 1315 BPF_R0 = *(u8 *)ptr; 1316 CONT; 1317 } 1318 1319 return 0; 1320 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */ 1321 off = IMM + SRC; 1322 goto load_word; 1323 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */ 1324 off = IMM + SRC; 1325 goto load_half; 1326 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */ 1327 off = IMM + SRC; 1328 goto load_byte; 1329 1330 default_label: 1331 /* If we ever reach this, we have a bug somewhere. Die hard here 1332 * instead of just returning 0; we could be somewhere in a subprog, 1333 * so execution could continue otherwise which we do /not/ want. 1334 * 1335 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 1336 */ 1337 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code); 1338 BUG_ON(1); 1339 return 0; 1340 } 1341 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */ 1342 1343 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1344 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1345 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1346 { \ 1347 u64 stack[stack_size / sizeof(u64)]; \ 1348 u64 regs[MAX_BPF_REG]; \ 1349 \ 1350 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1351 ARG1 = (u64) (unsigned long) ctx; \ 1352 return ___bpf_prog_run(regs, insn, stack); \ 1353 } 1354 1355 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1356 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1357 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1358 const struct bpf_insn *insn) \ 1359 { \ 1360 u64 stack[stack_size / sizeof(u64)]; \ 1361 u64 regs[MAX_BPF_REG]; \ 1362 \ 1363 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1364 BPF_R1 = r1; \ 1365 BPF_R2 = r2; \ 1366 BPF_R3 = r3; \ 1367 BPF_R4 = r4; \ 1368 BPF_R5 = r5; \ 1369 return ___bpf_prog_run(regs, insn, stack); \ 1370 } 1371 1372 #define EVAL1(FN, X) FN(X) 1373 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1374 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1375 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1376 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1377 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1378 1379 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1380 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1381 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1382 1383 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1384 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1385 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1386 1387 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1388 1389 static unsigned int (*interpreters[])(const void *ctx, 1390 const struct bpf_insn *insn) = { 1391 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1392 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1393 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1394 }; 1395 #undef PROG_NAME_LIST 1396 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1397 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1398 const struct bpf_insn *insn) = { 1399 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1400 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1401 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1402 }; 1403 #undef PROG_NAME_LIST 1404 1405 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1406 { 1407 stack_depth = max_t(u32, stack_depth, 1); 1408 insn->off = (s16) insn->imm; 1409 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1410 __bpf_call_base_args; 1411 insn->code = BPF_JMP | BPF_CALL_ARGS; 1412 } 1413 1414 #else 1415 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 1416 const struct bpf_insn *insn) 1417 { 1418 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 1419 * is not working properly, so warn about it! 1420 */ 1421 WARN_ON_ONCE(1); 1422 return 0; 1423 } 1424 #endif 1425 1426 bool bpf_prog_array_compatible(struct bpf_array *array, 1427 const struct bpf_prog *fp) 1428 { 1429 if (fp->kprobe_override) 1430 return false; 1431 1432 if (!array->owner_prog_type) { 1433 /* There's no owner yet where we could check for 1434 * compatibility. 1435 */ 1436 array->owner_prog_type = fp->type; 1437 array->owner_jited = fp->jited; 1438 1439 return true; 1440 } 1441 1442 return array->owner_prog_type == fp->type && 1443 array->owner_jited == fp->jited; 1444 } 1445 1446 static int bpf_check_tail_call(const struct bpf_prog *fp) 1447 { 1448 struct bpf_prog_aux *aux = fp->aux; 1449 int i; 1450 1451 for (i = 0; i < aux->used_map_cnt; i++) { 1452 struct bpf_map *map = aux->used_maps[i]; 1453 struct bpf_array *array; 1454 1455 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1456 continue; 1457 1458 array = container_of(map, struct bpf_array, map); 1459 if (!bpf_prog_array_compatible(array, fp)) 1460 return -EINVAL; 1461 } 1462 1463 return 0; 1464 } 1465 1466 /** 1467 * bpf_prog_select_runtime - select exec runtime for BPF program 1468 * @fp: bpf_prog populated with internal BPF program 1469 * @err: pointer to error variable 1470 * 1471 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1472 * The BPF program will be executed via BPF_PROG_RUN() macro. 1473 */ 1474 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1475 { 1476 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1477 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1478 1479 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1480 #else 1481 fp->bpf_func = __bpf_prog_ret0_warn; 1482 #endif 1483 1484 /* eBPF JITs can rewrite the program in case constant 1485 * blinding is active. However, in case of error during 1486 * blinding, bpf_int_jit_compile() must always return a 1487 * valid program, which in this case would simply not 1488 * be JITed, but falls back to the interpreter. 1489 */ 1490 if (!bpf_prog_is_dev_bound(fp->aux)) { 1491 fp = bpf_int_jit_compile(fp); 1492 #ifdef CONFIG_BPF_JIT_ALWAYS_ON 1493 if (!fp->jited) { 1494 *err = -ENOTSUPP; 1495 return fp; 1496 } 1497 #endif 1498 } else { 1499 *err = bpf_prog_offload_compile(fp); 1500 if (*err) 1501 return fp; 1502 } 1503 bpf_prog_lock_ro(fp); 1504 1505 /* The tail call compatibility check can only be done at 1506 * this late stage as we need to determine, if we deal 1507 * with JITed or non JITed program concatenations and not 1508 * all eBPF JITs might immediately support all features. 1509 */ 1510 *err = bpf_check_tail_call(fp); 1511 1512 return fp; 1513 } 1514 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1515 1516 static unsigned int __bpf_prog_ret1(const void *ctx, 1517 const struct bpf_insn *insn) 1518 { 1519 return 1; 1520 } 1521 1522 static struct bpf_prog_dummy { 1523 struct bpf_prog prog; 1524 } dummy_bpf_prog = { 1525 .prog = { 1526 .bpf_func = __bpf_prog_ret1, 1527 }, 1528 }; 1529 1530 /* to avoid allocating empty bpf_prog_array for cgroups that 1531 * don't have bpf program attached use one global 'empty_prog_array' 1532 * It will not be modified the caller of bpf_prog_array_alloc() 1533 * (since caller requested prog_cnt == 0) 1534 * that pointer should be 'freed' by bpf_prog_array_free() 1535 */ 1536 static struct { 1537 struct bpf_prog_array hdr; 1538 struct bpf_prog *null_prog; 1539 } empty_prog_array = { 1540 .null_prog = NULL, 1541 }; 1542 1543 struct bpf_prog_array __rcu *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1544 { 1545 if (prog_cnt) 1546 return kzalloc(sizeof(struct bpf_prog_array) + 1547 sizeof(struct bpf_prog *) * (prog_cnt + 1), 1548 flags); 1549 1550 return &empty_prog_array.hdr; 1551 } 1552 1553 void bpf_prog_array_free(struct bpf_prog_array __rcu *progs) 1554 { 1555 if (!progs || 1556 progs == (struct bpf_prog_array __rcu *)&empty_prog_array.hdr) 1557 return; 1558 kfree_rcu(progs, rcu); 1559 } 1560 1561 int bpf_prog_array_length(struct bpf_prog_array __rcu *progs) 1562 { 1563 struct bpf_prog **prog; 1564 u32 cnt = 0; 1565 1566 rcu_read_lock(); 1567 prog = rcu_dereference(progs)->progs; 1568 for (; *prog; prog++) 1569 if (*prog != &dummy_bpf_prog.prog) 1570 cnt++; 1571 rcu_read_unlock(); 1572 return cnt; 1573 } 1574 1575 static bool bpf_prog_array_copy_core(struct bpf_prog **prog, 1576 u32 *prog_ids, 1577 u32 request_cnt) 1578 { 1579 int i = 0; 1580 1581 for (; *prog; prog++) { 1582 if (*prog == &dummy_bpf_prog.prog) 1583 continue; 1584 prog_ids[i] = (*prog)->aux->id; 1585 if (++i == request_cnt) { 1586 prog++; 1587 break; 1588 } 1589 } 1590 1591 return !!(*prog); 1592 } 1593 1594 int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *progs, 1595 __u32 __user *prog_ids, u32 cnt) 1596 { 1597 struct bpf_prog **prog; 1598 unsigned long err = 0; 1599 bool nospc; 1600 u32 *ids; 1601 1602 /* users of this function are doing: 1603 * cnt = bpf_prog_array_length(); 1604 * if (cnt > 0) 1605 * bpf_prog_array_copy_to_user(..., cnt); 1606 * so below kcalloc doesn't need extra cnt > 0 check, but 1607 * bpf_prog_array_length() releases rcu lock and 1608 * prog array could have been swapped with empty or larger array, 1609 * so always copy 'cnt' prog_ids to the user. 1610 * In a rare race the user will see zero prog_ids 1611 */ 1612 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 1613 if (!ids) 1614 return -ENOMEM; 1615 rcu_read_lock(); 1616 prog = rcu_dereference(progs)->progs; 1617 nospc = bpf_prog_array_copy_core(prog, ids, cnt); 1618 rcu_read_unlock(); 1619 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 1620 kfree(ids); 1621 if (err) 1622 return -EFAULT; 1623 if (nospc) 1624 return -ENOSPC; 1625 return 0; 1626 } 1627 1628 void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *progs, 1629 struct bpf_prog *old_prog) 1630 { 1631 struct bpf_prog **prog = progs->progs; 1632 1633 for (; *prog; prog++) 1634 if (*prog == old_prog) { 1635 WRITE_ONCE(*prog, &dummy_bpf_prog.prog); 1636 break; 1637 } 1638 } 1639 1640 int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array, 1641 struct bpf_prog *exclude_prog, 1642 struct bpf_prog *include_prog, 1643 struct bpf_prog_array **new_array) 1644 { 1645 int new_prog_cnt, carry_prog_cnt = 0; 1646 struct bpf_prog **existing_prog; 1647 struct bpf_prog_array *array; 1648 int new_prog_idx = 0; 1649 1650 /* Figure out how many existing progs we need to carry over to 1651 * the new array. 1652 */ 1653 if (old_array) { 1654 existing_prog = old_array->progs; 1655 for (; *existing_prog; existing_prog++) { 1656 if (*existing_prog != exclude_prog && 1657 *existing_prog != &dummy_bpf_prog.prog) 1658 carry_prog_cnt++; 1659 if (*existing_prog == include_prog) 1660 return -EEXIST; 1661 } 1662 } 1663 1664 /* How many progs (not NULL) will be in the new array? */ 1665 new_prog_cnt = carry_prog_cnt; 1666 if (include_prog) 1667 new_prog_cnt += 1; 1668 1669 /* Do we have any prog (not NULL) in the new array? */ 1670 if (!new_prog_cnt) { 1671 *new_array = NULL; 1672 return 0; 1673 } 1674 1675 /* +1 as the end of prog_array is marked with NULL */ 1676 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1677 if (!array) 1678 return -ENOMEM; 1679 1680 /* Fill in the new prog array */ 1681 if (carry_prog_cnt) { 1682 existing_prog = old_array->progs; 1683 for (; *existing_prog; existing_prog++) 1684 if (*existing_prog != exclude_prog && 1685 *existing_prog != &dummy_bpf_prog.prog) 1686 array->progs[new_prog_idx++] = *existing_prog; 1687 } 1688 if (include_prog) 1689 array->progs[new_prog_idx++] = include_prog; 1690 array->progs[new_prog_idx] = NULL; 1691 *new_array = array; 1692 return 0; 1693 } 1694 1695 int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array, 1696 u32 *prog_ids, u32 request_cnt, 1697 u32 *prog_cnt) 1698 { 1699 struct bpf_prog **prog; 1700 u32 cnt = 0; 1701 1702 if (array) 1703 cnt = bpf_prog_array_length(array); 1704 1705 *prog_cnt = cnt; 1706 1707 /* return early if user requested only program count or nothing to copy */ 1708 if (!request_cnt || !cnt) 1709 return 0; 1710 1711 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 1712 prog = rcu_dereference_check(array, 1)->progs; 1713 return bpf_prog_array_copy_core(prog, prog_ids, request_cnt) ? -ENOSPC 1714 : 0; 1715 } 1716 1717 static void bpf_prog_free_deferred(struct work_struct *work) 1718 { 1719 struct bpf_prog_aux *aux; 1720 int i; 1721 1722 aux = container_of(work, struct bpf_prog_aux, work); 1723 if (bpf_prog_is_dev_bound(aux)) 1724 bpf_prog_offload_destroy(aux->prog); 1725 for (i = 0; i < aux->func_cnt; i++) 1726 bpf_jit_free(aux->func[i]); 1727 if (aux->func_cnt) { 1728 kfree(aux->func); 1729 bpf_prog_unlock_free(aux->prog); 1730 } else { 1731 bpf_jit_free(aux->prog); 1732 } 1733 } 1734 1735 /* Free internal BPF program */ 1736 void bpf_prog_free(struct bpf_prog *fp) 1737 { 1738 struct bpf_prog_aux *aux = fp->aux; 1739 1740 INIT_WORK(&aux->work, bpf_prog_free_deferred); 1741 schedule_work(&aux->work); 1742 } 1743 EXPORT_SYMBOL_GPL(bpf_prog_free); 1744 1745 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 1746 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 1747 1748 void bpf_user_rnd_init_once(void) 1749 { 1750 prandom_init_once(&bpf_user_rnd_state); 1751 } 1752 1753 BPF_CALL_0(bpf_user_rnd_u32) 1754 { 1755 /* Should someone ever have the rather unwise idea to use some 1756 * of the registers passed into this function, then note that 1757 * this function is called from native eBPF and classic-to-eBPF 1758 * transformations. Register assignments from both sides are 1759 * different, f.e. classic always sets fn(ctx, A, X) here. 1760 */ 1761 struct rnd_state *state; 1762 u32 res; 1763 1764 state = &get_cpu_var(bpf_user_rnd_state); 1765 res = prandom_u32_state(state); 1766 put_cpu_var(bpf_user_rnd_state); 1767 1768 return res; 1769 } 1770 1771 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 1772 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 1773 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 1774 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 1775 1776 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 1777 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 1778 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 1779 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 1780 1781 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 1782 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 1783 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 1784 const struct bpf_func_proto bpf_sock_map_update_proto __weak; 1785 1786 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 1787 { 1788 return NULL; 1789 } 1790 1791 u64 __weak 1792 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 1793 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 1794 { 1795 return -ENOTSUPP; 1796 } 1797 1798 /* Always built-in helper functions. */ 1799 const struct bpf_func_proto bpf_tail_call_proto = { 1800 .func = NULL, 1801 .gpl_only = false, 1802 .ret_type = RET_VOID, 1803 .arg1_type = ARG_PTR_TO_CTX, 1804 .arg2_type = ARG_CONST_MAP_PTR, 1805 .arg3_type = ARG_ANYTHING, 1806 }; 1807 1808 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 1809 * It is encouraged to implement bpf_int_jit_compile() instead, so that 1810 * eBPF and implicitly also cBPF can get JITed! 1811 */ 1812 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 1813 { 1814 return prog; 1815 } 1816 1817 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 1818 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 1819 */ 1820 void __weak bpf_jit_compile(struct bpf_prog *prog) 1821 { 1822 } 1823 1824 bool __weak bpf_helper_changes_pkt_data(void *func) 1825 { 1826 return false; 1827 } 1828 1829 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 1830 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 1831 */ 1832 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 1833 int len) 1834 { 1835 return -EFAULT; 1836 } 1837 1838 /* All definitions of tracepoints related to BPF. */ 1839 #define CREATE_TRACE_POINTS 1840 #include <linux/bpf_trace.h> 1841 1842 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 1843 1844 /* These are only used within the BPF_SYSCALL code */ 1845 #ifdef CONFIG_BPF_SYSCALL 1846 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type); 1847 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu); 1848 #endif 1849