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 static __always_inline void 304 bpf_get_prog_addr_region(const struct bpf_prog *prog, 305 unsigned long *symbol_start, 306 unsigned long *symbol_end) 307 { 308 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 309 unsigned long addr = (unsigned long)hdr; 310 311 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 312 313 *symbol_start = addr; 314 *symbol_end = addr + hdr->pages * PAGE_SIZE; 315 } 316 317 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym) 318 { 319 const char *end = sym + KSYM_NAME_LEN; 320 321 BUILD_BUG_ON(sizeof("bpf_prog_") + 322 sizeof(prog->tag) * 2 + 323 /* name has been null terminated. 324 * We should need +1 for the '_' preceding 325 * the name. However, the null character 326 * is double counted between the name and the 327 * sizeof("bpf_prog_") above, so we omit 328 * the +1 here. 329 */ 330 sizeof(prog->aux->name) > KSYM_NAME_LEN); 331 332 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 333 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 334 if (prog->aux->name[0]) 335 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 336 else 337 *sym = 0; 338 } 339 340 static __always_inline unsigned long 341 bpf_get_prog_addr_start(struct latch_tree_node *n) 342 { 343 unsigned long symbol_start, symbol_end; 344 const struct bpf_prog_aux *aux; 345 346 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 347 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 348 349 return symbol_start; 350 } 351 352 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 353 struct latch_tree_node *b) 354 { 355 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b); 356 } 357 358 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 359 { 360 unsigned long val = (unsigned long)key; 361 unsigned long symbol_start, symbol_end; 362 const struct bpf_prog_aux *aux; 363 364 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 365 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 366 367 if (val < symbol_start) 368 return -1; 369 if (val >= symbol_end) 370 return 1; 371 372 return 0; 373 } 374 375 static const struct latch_tree_ops bpf_tree_ops = { 376 .less = bpf_tree_less, 377 .comp = bpf_tree_comp, 378 }; 379 380 static DEFINE_SPINLOCK(bpf_lock); 381 static LIST_HEAD(bpf_kallsyms); 382 static struct latch_tree_root bpf_tree __cacheline_aligned; 383 384 int bpf_jit_kallsyms __read_mostly; 385 386 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux) 387 { 388 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode)); 389 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms); 390 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 391 } 392 393 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux) 394 { 395 if (list_empty(&aux->ksym_lnode)) 396 return; 397 398 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 399 list_del_rcu(&aux->ksym_lnode); 400 } 401 402 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 403 { 404 return fp->jited && !bpf_prog_was_classic(fp); 405 } 406 407 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 408 { 409 return list_empty(&fp->aux->ksym_lnode) || 410 fp->aux->ksym_lnode.prev == LIST_POISON2; 411 } 412 413 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 414 { 415 if (!bpf_prog_kallsyms_candidate(fp) || 416 !capable(CAP_SYS_ADMIN)) 417 return; 418 419 spin_lock_bh(&bpf_lock); 420 bpf_prog_ksym_node_add(fp->aux); 421 spin_unlock_bh(&bpf_lock); 422 } 423 424 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 425 { 426 if (!bpf_prog_kallsyms_candidate(fp)) 427 return; 428 429 spin_lock_bh(&bpf_lock); 430 bpf_prog_ksym_node_del(fp->aux); 431 spin_unlock_bh(&bpf_lock); 432 } 433 434 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr) 435 { 436 struct latch_tree_node *n; 437 438 if (!bpf_jit_kallsyms_enabled()) 439 return NULL; 440 441 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 442 return n ? 443 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 444 NULL; 445 } 446 447 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 448 unsigned long *off, char *sym) 449 { 450 unsigned long symbol_start, symbol_end; 451 struct bpf_prog *prog; 452 char *ret = NULL; 453 454 rcu_read_lock(); 455 prog = bpf_prog_kallsyms_find(addr); 456 if (prog) { 457 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 458 bpf_get_prog_name(prog, sym); 459 460 ret = sym; 461 if (size) 462 *size = symbol_end - symbol_start; 463 if (off) 464 *off = addr - symbol_start; 465 } 466 rcu_read_unlock(); 467 468 return ret; 469 } 470 471 bool is_bpf_text_address(unsigned long addr) 472 { 473 bool ret; 474 475 rcu_read_lock(); 476 ret = bpf_prog_kallsyms_find(addr) != NULL; 477 rcu_read_unlock(); 478 479 return ret; 480 } 481 482 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 483 char *sym) 484 { 485 unsigned long symbol_start, symbol_end; 486 struct bpf_prog_aux *aux; 487 unsigned int it = 0; 488 int ret = -ERANGE; 489 490 if (!bpf_jit_kallsyms_enabled()) 491 return ret; 492 493 rcu_read_lock(); 494 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 495 if (it++ != symnum) 496 continue; 497 498 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 499 bpf_get_prog_name(aux->prog, sym); 500 501 *value = symbol_start; 502 *type = BPF_SYM_ELF_TYPE; 503 504 ret = 0; 505 break; 506 } 507 rcu_read_unlock(); 508 509 return ret; 510 } 511 512 struct bpf_binary_header * 513 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 514 unsigned int alignment, 515 bpf_jit_fill_hole_t bpf_fill_ill_insns) 516 { 517 struct bpf_binary_header *hdr; 518 unsigned int size, hole, start; 519 520 /* Most of BPF filters are really small, but if some of them 521 * fill a page, allow at least 128 extra bytes to insert a 522 * random section of illegal instructions. 523 */ 524 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 525 hdr = module_alloc(size); 526 if (hdr == NULL) 527 return NULL; 528 529 /* Fill space with illegal/arch-dep instructions. */ 530 bpf_fill_ill_insns(hdr, size); 531 532 hdr->pages = size / PAGE_SIZE; 533 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 534 PAGE_SIZE - sizeof(*hdr)); 535 start = (get_random_int() % hole) & ~(alignment - 1); 536 537 /* Leave a random number of instructions before BPF code. */ 538 *image_ptr = &hdr->image[start]; 539 540 return hdr; 541 } 542 543 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 544 { 545 module_memfree(hdr); 546 } 547 548 /* This symbol is only overridden by archs that have different 549 * requirements than the usual eBPF JITs, f.e. when they only 550 * implement cBPF JIT, do not set images read-only, etc. 551 */ 552 void __weak bpf_jit_free(struct bpf_prog *fp) 553 { 554 if (fp->jited) { 555 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 556 557 bpf_jit_binary_unlock_ro(hdr); 558 bpf_jit_binary_free(hdr); 559 560 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 561 } 562 563 bpf_prog_unlock_free(fp); 564 } 565 566 int bpf_jit_harden __read_mostly; 567 568 static int bpf_jit_blind_insn(const struct bpf_insn *from, 569 const struct bpf_insn *aux, 570 struct bpf_insn *to_buff) 571 { 572 struct bpf_insn *to = to_buff; 573 u32 imm_rnd = get_random_int(); 574 s16 off; 575 576 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 577 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 578 579 if (from->imm == 0 && 580 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 581 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 582 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 583 goto out; 584 } 585 586 switch (from->code) { 587 case BPF_ALU | BPF_ADD | BPF_K: 588 case BPF_ALU | BPF_SUB | BPF_K: 589 case BPF_ALU | BPF_AND | BPF_K: 590 case BPF_ALU | BPF_OR | BPF_K: 591 case BPF_ALU | BPF_XOR | BPF_K: 592 case BPF_ALU | BPF_MUL | BPF_K: 593 case BPF_ALU | BPF_MOV | BPF_K: 594 case BPF_ALU | BPF_DIV | BPF_K: 595 case BPF_ALU | BPF_MOD | BPF_K: 596 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 597 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 598 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 599 break; 600 601 case BPF_ALU64 | BPF_ADD | BPF_K: 602 case BPF_ALU64 | BPF_SUB | BPF_K: 603 case BPF_ALU64 | BPF_AND | BPF_K: 604 case BPF_ALU64 | BPF_OR | BPF_K: 605 case BPF_ALU64 | BPF_XOR | BPF_K: 606 case BPF_ALU64 | BPF_MUL | BPF_K: 607 case BPF_ALU64 | BPF_MOV | BPF_K: 608 case BPF_ALU64 | BPF_DIV | BPF_K: 609 case BPF_ALU64 | BPF_MOD | BPF_K: 610 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 611 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 612 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 613 break; 614 615 case BPF_JMP | BPF_JEQ | BPF_K: 616 case BPF_JMP | BPF_JNE | BPF_K: 617 case BPF_JMP | BPF_JGT | BPF_K: 618 case BPF_JMP | BPF_JLT | BPF_K: 619 case BPF_JMP | BPF_JGE | BPF_K: 620 case BPF_JMP | BPF_JLE | BPF_K: 621 case BPF_JMP | BPF_JSGT | BPF_K: 622 case BPF_JMP | BPF_JSLT | BPF_K: 623 case BPF_JMP | BPF_JSGE | BPF_K: 624 case BPF_JMP | BPF_JSLE | BPF_K: 625 case BPF_JMP | BPF_JSET | BPF_K: 626 /* Accommodate for extra offset in case of a backjump. */ 627 off = from->off; 628 if (off < 0) 629 off -= 2; 630 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 631 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 632 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 633 break; 634 635 case BPF_LD | BPF_ABS | BPF_W: 636 case BPF_LD | BPF_ABS | BPF_H: 637 case BPF_LD | BPF_ABS | BPF_B: 638 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 639 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 640 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0); 641 break; 642 643 case BPF_LD | BPF_IND | BPF_W: 644 case BPF_LD | BPF_IND | BPF_H: 645 case BPF_LD | BPF_IND | BPF_B: 646 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 647 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 648 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg); 649 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0); 650 break; 651 652 case BPF_LD | BPF_IMM | BPF_DW: 653 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 654 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 655 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 656 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 657 break; 658 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 659 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 660 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 661 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 662 break; 663 664 case BPF_ST | BPF_MEM | BPF_DW: 665 case BPF_ST | BPF_MEM | BPF_W: 666 case BPF_ST | BPF_MEM | BPF_H: 667 case BPF_ST | BPF_MEM | BPF_B: 668 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 669 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 670 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 671 break; 672 } 673 out: 674 return to - to_buff; 675 } 676 677 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 678 gfp_t gfp_extra_flags) 679 { 680 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 681 struct bpf_prog *fp; 682 683 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 684 if (fp != NULL) { 685 /* aux->prog still points to the fp_other one, so 686 * when promoting the clone to the real program, 687 * this still needs to be adapted. 688 */ 689 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 690 } 691 692 return fp; 693 } 694 695 static void bpf_prog_clone_free(struct bpf_prog *fp) 696 { 697 /* aux was stolen by the other clone, so we cannot free 698 * it from this path! It will be freed eventually by the 699 * other program on release. 700 * 701 * At this point, we don't need a deferred release since 702 * clone is guaranteed to not be locked. 703 */ 704 fp->aux = NULL; 705 __bpf_prog_free(fp); 706 } 707 708 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 709 { 710 /* We have to repoint aux->prog to self, as we don't 711 * know whether fp here is the clone or the original. 712 */ 713 fp->aux->prog = fp; 714 bpf_prog_clone_free(fp_other); 715 } 716 717 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 718 { 719 struct bpf_insn insn_buff[16], aux[2]; 720 struct bpf_prog *clone, *tmp; 721 int insn_delta, insn_cnt; 722 struct bpf_insn *insn; 723 int i, rewritten; 724 725 if (!bpf_jit_blinding_enabled(prog) || prog->blinded) 726 return prog; 727 728 clone = bpf_prog_clone_create(prog, GFP_USER); 729 if (!clone) 730 return ERR_PTR(-ENOMEM); 731 732 insn_cnt = clone->len; 733 insn = clone->insnsi; 734 735 for (i = 0; i < insn_cnt; i++, insn++) { 736 /* We temporarily need to hold the original ld64 insn 737 * so that we can still access the first part in the 738 * second blinding run. 739 */ 740 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 741 insn[1].code == 0) 742 memcpy(aux, insn, sizeof(aux)); 743 744 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff); 745 if (!rewritten) 746 continue; 747 748 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 749 if (!tmp) { 750 /* Patching may have repointed aux->prog during 751 * realloc from the original one, so we need to 752 * fix it up here on error. 753 */ 754 bpf_jit_prog_release_other(prog, clone); 755 return ERR_PTR(-ENOMEM); 756 } 757 758 clone = tmp; 759 insn_delta = rewritten - 1; 760 761 /* Walk new program and skip insns we just inserted. */ 762 insn = clone->insnsi + i + insn_delta; 763 insn_cnt += insn_delta; 764 i += insn_delta; 765 } 766 767 clone->blinded = 1; 768 return clone; 769 } 770 #endif /* CONFIG_BPF_JIT */ 771 772 /* Base function for offset calculation. Needs to go into .text section, 773 * therefore keeping it non-static as well; will also be used by JITs 774 * anyway later on, so do not let the compiler omit it. This also needs 775 * to go into kallsyms for correlation from e.g. bpftool, so naming 776 * must not change. 777 */ 778 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 779 { 780 return 0; 781 } 782 EXPORT_SYMBOL_GPL(__bpf_call_base); 783 784 /** 785 * __bpf_prog_run - run eBPF program on a given context 786 * @ctx: is the data we are operating on 787 * @insn: is the array of eBPF instructions 788 * 789 * Decode and execute eBPF instructions. 790 */ 791 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack) 792 { 793 u64 tmp; 794 static const void *jumptable[256] = { 795 [0 ... 255] = &&default_label, 796 /* Now overwrite non-defaults ... */ 797 /* 32 bit ALU operations */ 798 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X, 799 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K, 800 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X, 801 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K, 802 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X, 803 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K, 804 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X, 805 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K, 806 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X, 807 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K, 808 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X, 809 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K, 810 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X, 811 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K, 812 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X, 813 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K, 814 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X, 815 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K, 816 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X, 817 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K, 818 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X, 819 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K, 820 [BPF_ALU | BPF_NEG] = &&ALU_NEG, 821 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE, 822 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE, 823 /* 64 bit ALU operations */ 824 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X, 825 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K, 826 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X, 827 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K, 828 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X, 829 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K, 830 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X, 831 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K, 832 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X, 833 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K, 834 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X, 835 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K, 836 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X, 837 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K, 838 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X, 839 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K, 840 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X, 841 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K, 842 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X, 843 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K, 844 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X, 845 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K, 846 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X, 847 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K, 848 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG, 849 /* Call instruction */ 850 [BPF_JMP | BPF_CALL] = &&JMP_CALL, 851 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 852 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 853 /* Jumps */ 854 [BPF_JMP | BPF_JA] = &&JMP_JA, 855 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X, 856 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K, 857 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X, 858 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K, 859 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X, 860 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K, 861 [BPF_JMP | BPF_JLT | BPF_X] = &&JMP_JLT_X, 862 [BPF_JMP | BPF_JLT | BPF_K] = &&JMP_JLT_K, 863 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X, 864 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K, 865 [BPF_JMP | BPF_JLE | BPF_X] = &&JMP_JLE_X, 866 [BPF_JMP | BPF_JLE | BPF_K] = &&JMP_JLE_K, 867 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X, 868 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K, 869 [BPF_JMP | BPF_JSLT | BPF_X] = &&JMP_JSLT_X, 870 [BPF_JMP | BPF_JSLT | BPF_K] = &&JMP_JSLT_K, 871 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X, 872 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K, 873 [BPF_JMP | BPF_JSLE | BPF_X] = &&JMP_JSLE_X, 874 [BPF_JMP | BPF_JSLE | BPF_K] = &&JMP_JSLE_K, 875 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X, 876 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K, 877 /* Program return */ 878 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT, 879 /* Store instructions */ 880 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B, 881 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H, 882 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W, 883 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW, 884 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W, 885 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW, 886 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B, 887 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H, 888 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W, 889 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW, 890 /* Load instructions */ 891 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B, 892 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H, 893 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W, 894 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW, 895 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W, 896 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H, 897 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B, 898 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W, 899 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H, 900 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B, 901 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW, 902 }; 903 u32 tail_call_cnt = 0; 904 void *ptr; 905 int off; 906 907 #define CONT ({ insn++; goto select_insn; }) 908 #define CONT_JMP ({ insn++; goto select_insn; }) 909 910 select_insn: 911 goto *jumptable[insn->code]; 912 913 /* ALU */ 914 #define ALU(OPCODE, OP) \ 915 ALU64_##OPCODE##_X: \ 916 DST = DST OP SRC; \ 917 CONT; \ 918 ALU_##OPCODE##_X: \ 919 DST = (u32) DST OP (u32) SRC; \ 920 CONT; \ 921 ALU64_##OPCODE##_K: \ 922 DST = DST OP IMM; \ 923 CONT; \ 924 ALU_##OPCODE##_K: \ 925 DST = (u32) DST OP (u32) IMM; \ 926 CONT; 927 928 ALU(ADD, +) 929 ALU(SUB, -) 930 ALU(AND, &) 931 ALU(OR, |) 932 ALU(LSH, <<) 933 ALU(RSH, >>) 934 ALU(XOR, ^) 935 ALU(MUL, *) 936 #undef ALU 937 ALU_NEG: 938 DST = (u32) -DST; 939 CONT; 940 ALU64_NEG: 941 DST = -DST; 942 CONT; 943 ALU_MOV_X: 944 DST = (u32) SRC; 945 CONT; 946 ALU_MOV_K: 947 DST = (u32) IMM; 948 CONT; 949 ALU64_MOV_X: 950 DST = SRC; 951 CONT; 952 ALU64_MOV_K: 953 DST = IMM; 954 CONT; 955 LD_IMM_DW: 956 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 957 insn++; 958 CONT; 959 ALU64_ARSH_X: 960 (*(s64 *) &DST) >>= SRC; 961 CONT; 962 ALU64_ARSH_K: 963 (*(s64 *) &DST) >>= IMM; 964 CONT; 965 ALU64_MOD_X: 966 if (unlikely(SRC == 0)) 967 return 0; 968 div64_u64_rem(DST, SRC, &tmp); 969 DST = tmp; 970 CONT; 971 ALU_MOD_X: 972 if (unlikely(SRC == 0)) 973 return 0; 974 tmp = (u32) DST; 975 DST = do_div(tmp, (u32) SRC); 976 CONT; 977 ALU64_MOD_K: 978 div64_u64_rem(DST, IMM, &tmp); 979 DST = tmp; 980 CONT; 981 ALU_MOD_K: 982 tmp = (u32) DST; 983 DST = do_div(tmp, (u32) IMM); 984 CONT; 985 ALU64_DIV_X: 986 if (unlikely(SRC == 0)) 987 return 0; 988 DST = div64_u64(DST, SRC); 989 CONT; 990 ALU_DIV_X: 991 if (unlikely(SRC == 0)) 992 return 0; 993 tmp = (u32) DST; 994 do_div(tmp, (u32) SRC); 995 DST = (u32) tmp; 996 CONT; 997 ALU64_DIV_K: 998 DST = div64_u64(DST, IMM); 999 CONT; 1000 ALU_DIV_K: 1001 tmp = (u32) DST; 1002 do_div(tmp, (u32) IMM); 1003 DST = (u32) tmp; 1004 CONT; 1005 ALU_END_TO_BE: 1006 switch (IMM) { 1007 case 16: 1008 DST = (__force u16) cpu_to_be16(DST); 1009 break; 1010 case 32: 1011 DST = (__force u32) cpu_to_be32(DST); 1012 break; 1013 case 64: 1014 DST = (__force u64) cpu_to_be64(DST); 1015 break; 1016 } 1017 CONT; 1018 ALU_END_TO_LE: 1019 switch (IMM) { 1020 case 16: 1021 DST = (__force u16) cpu_to_le16(DST); 1022 break; 1023 case 32: 1024 DST = (__force u32) cpu_to_le32(DST); 1025 break; 1026 case 64: 1027 DST = (__force u64) cpu_to_le64(DST); 1028 break; 1029 } 1030 CONT; 1031 1032 /* CALL */ 1033 JMP_CALL: 1034 /* Function call scratches BPF_R1-BPF_R5 registers, 1035 * preserves BPF_R6-BPF_R9, and stores return value 1036 * into BPF_R0. 1037 */ 1038 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1039 BPF_R4, BPF_R5); 1040 CONT; 1041 1042 JMP_CALL_ARGS: 1043 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 1044 BPF_R3, BPF_R4, 1045 BPF_R5, 1046 insn + insn->off + 1); 1047 CONT; 1048 1049 JMP_TAIL_CALL: { 1050 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1051 struct bpf_array *array = container_of(map, struct bpf_array, map); 1052 struct bpf_prog *prog; 1053 u32 index = BPF_R3; 1054 1055 if (unlikely(index >= array->map.max_entries)) 1056 goto out; 1057 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1058 goto out; 1059 1060 tail_call_cnt++; 1061 1062 prog = READ_ONCE(array->ptrs[index]); 1063 if (!prog) 1064 goto out; 1065 1066 /* ARG1 at this point is guaranteed to point to CTX from 1067 * the verifier side due to the fact that the tail call is 1068 * handeled like a helper, that is, bpf_tail_call_proto, 1069 * where arg1_type is ARG_PTR_TO_CTX. 1070 */ 1071 insn = prog->insnsi; 1072 goto select_insn; 1073 out: 1074 CONT; 1075 } 1076 /* JMP */ 1077 JMP_JA: 1078 insn += insn->off; 1079 CONT; 1080 JMP_JEQ_X: 1081 if (DST == SRC) { 1082 insn += insn->off; 1083 CONT_JMP; 1084 } 1085 CONT; 1086 JMP_JEQ_K: 1087 if (DST == IMM) { 1088 insn += insn->off; 1089 CONT_JMP; 1090 } 1091 CONT; 1092 JMP_JNE_X: 1093 if (DST != SRC) { 1094 insn += insn->off; 1095 CONT_JMP; 1096 } 1097 CONT; 1098 JMP_JNE_K: 1099 if (DST != IMM) { 1100 insn += insn->off; 1101 CONT_JMP; 1102 } 1103 CONT; 1104 JMP_JGT_X: 1105 if (DST > SRC) { 1106 insn += insn->off; 1107 CONT_JMP; 1108 } 1109 CONT; 1110 JMP_JGT_K: 1111 if (DST > IMM) { 1112 insn += insn->off; 1113 CONT_JMP; 1114 } 1115 CONT; 1116 JMP_JLT_X: 1117 if (DST < SRC) { 1118 insn += insn->off; 1119 CONT_JMP; 1120 } 1121 CONT; 1122 JMP_JLT_K: 1123 if (DST < IMM) { 1124 insn += insn->off; 1125 CONT_JMP; 1126 } 1127 CONT; 1128 JMP_JGE_X: 1129 if (DST >= SRC) { 1130 insn += insn->off; 1131 CONT_JMP; 1132 } 1133 CONT; 1134 JMP_JGE_K: 1135 if (DST >= IMM) { 1136 insn += insn->off; 1137 CONT_JMP; 1138 } 1139 CONT; 1140 JMP_JLE_X: 1141 if (DST <= SRC) { 1142 insn += insn->off; 1143 CONT_JMP; 1144 } 1145 CONT; 1146 JMP_JLE_K: 1147 if (DST <= IMM) { 1148 insn += insn->off; 1149 CONT_JMP; 1150 } 1151 CONT; 1152 JMP_JSGT_X: 1153 if (((s64) DST) > ((s64) SRC)) { 1154 insn += insn->off; 1155 CONT_JMP; 1156 } 1157 CONT; 1158 JMP_JSGT_K: 1159 if (((s64) DST) > ((s64) IMM)) { 1160 insn += insn->off; 1161 CONT_JMP; 1162 } 1163 CONT; 1164 JMP_JSLT_X: 1165 if (((s64) DST) < ((s64) SRC)) { 1166 insn += insn->off; 1167 CONT_JMP; 1168 } 1169 CONT; 1170 JMP_JSLT_K: 1171 if (((s64) DST) < ((s64) IMM)) { 1172 insn += insn->off; 1173 CONT_JMP; 1174 } 1175 CONT; 1176 JMP_JSGE_X: 1177 if (((s64) DST) >= ((s64) SRC)) { 1178 insn += insn->off; 1179 CONT_JMP; 1180 } 1181 CONT; 1182 JMP_JSGE_K: 1183 if (((s64) DST) >= ((s64) IMM)) { 1184 insn += insn->off; 1185 CONT_JMP; 1186 } 1187 CONT; 1188 JMP_JSLE_X: 1189 if (((s64) DST) <= ((s64) SRC)) { 1190 insn += insn->off; 1191 CONT_JMP; 1192 } 1193 CONT; 1194 JMP_JSLE_K: 1195 if (((s64) DST) <= ((s64) IMM)) { 1196 insn += insn->off; 1197 CONT_JMP; 1198 } 1199 CONT; 1200 JMP_JSET_X: 1201 if (DST & SRC) { 1202 insn += insn->off; 1203 CONT_JMP; 1204 } 1205 CONT; 1206 JMP_JSET_K: 1207 if (DST & IMM) { 1208 insn += insn->off; 1209 CONT_JMP; 1210 } 1211 CONT; 1212 JMP_EXIT: 1213 return BPF_R0; 1214 1215 /* STX and ST and LDX*/ 1216 #define LDST(SIZEOP, SIZE) \ 1217 STX_MEM_##SIZEOP: \ 1218 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1219 CONT; \ 1220 ST_MEM_##SIZEOP: \ 1221 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1222 CONT; \ 1223 LDX_MEM_##SIZEOP: \ 1224 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1225 CONT; 1226 1227 LDST(B, u8) 1228 LDST(H, u16) 1229 LDST(W, u32) 1230 LDST(DW, u64) 1231 #undef LDST 1232 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1233 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1234 (DST + insn->off)); 1235 CONT; 1236 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1237 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1238 (DST + insn->off)); 1239 CONT; 1240 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */ 1241 off = IMM; 1242 load_word: 1243 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only 1244 * appearing in the programs where ctx == skb 1245 * (see may_access_skb() in the verifier). All programs 1246 * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6, 1247 * bpf_convert_filter() saves it in BPF_R6, internal BPF 1248 * verifier will check that BPF_R6 == ctx. 1249 * 1250 * BPF_ABS and BPF_IND are wrappers of function calls, 1251 * so they scratch BPF_R1-BPF_R5 registers, preserve 1252 * BPF_R6-BPF_R9, and store return value into BPF_R0. 1253 * 1254 * Implicit input: 1255 * ctx == skb == BPF_R6 == CTX 1256 * 1257 * Explicit input: 1258 * SRC == any register 1259 * IMM == 32-bit immediate 1260 * 1261 * Output: 1262 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness 1263 */ 1264 1265 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp); 1266 if (likely(ptr != NULL)) { 1267 BPF_R0 = get_unaligned_be32(ptr); 1268 CONT; 1269 } 1270 1271 return 0; 1272 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */ 1273 off = IMM; 1274 load_half: 1275 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp); 1276 if (likely(ptr != NULL)) { 1277 BPF_R0 = get_unaligned_be16(ptr); 1278 CONT; 1279 } 1280 1281 return 0; 1282 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */ 1283 off = IMM; 1284 load_byte: 1285 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp); 1286 if (likely(ptr != NULL)) { 1287 BPF_R0 = *(u8 *)ptr; 1288 CONT; 1289 } 1290 1291 return 0; 1292 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */ 1293 off = IMM + SRC; 1294 goto load_word; 1295 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */ 1296 off = IMM + SRC; 1297 goto load_half; 1298 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */ 1299 off = IMM + SRC; 1300 goto load_byte; 1301 1302 default_label: 1303 /* If we ever reach this, we have a bug somewhere. */ 1304 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code); 1305 return 0; 1306 } 1307 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */ 1308 1309 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1310 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1311 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1312 { \ 1313 u64 stack[stack_size / sizeof(u64)]; \ 1314 u64 regs[MAX_BPF_REG]; \ 1315 \ 1316 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1317 ARG1 = (u64) (unsigned long) ctx; \ 1318 return ___bpf_prog_run(regs, insn, stack); \ 1319 } 1320 1321 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 1322 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 1323 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 1324 const struct bpf_insn *insn) \ 1325 { \ 1326 u64 stack[stack_size / sizeof(u64)]; \ 1327 u64 regs[MAX_BPF_REG]; \ 1328 \ 1329 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1330 BPF_R1 = r1; \ 1331 BPF_R2 = r2; \ 1332 BPF_R3 = r3; \ 1333 BPF_R4 = r4; \ 1334 BPF_R5 = r5; \ 1335 return ___bpf_prog_run(regs, insn, stack); \ 1336 } 1337 1338 #define EVAL1(FN, X) FN(X) 1339 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1340 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1341 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1342 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1343 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1344 1345 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1346 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1347 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1348 1349 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 1350 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 1351 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 1352 1353 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1354 1355 static unsigned int (*interpreters[])(const void *ctx, 1356 const struct bpf_insn *insn) = { 1357 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1358 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1359 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1360 }; 1361 #undef PROG_NAME_LIST 1362 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 1363 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1364 const struct bpf_insn *insn) = { 1365 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1366 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1367 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1368 }; 1369 #undef PROG_NAME_LIST 1370 1371 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 1372 { 1373 stack_depth = max_t(u32, stack_depth, 1); 1374 insn->off = (s16) insn->imm; 1375 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 1376 __bpf_call_base_args; 1377 insn->code = BPF_JMP | BPF_CALL_ARGS; 1378 } 1379 1380 bool bpf_prog_array_compatible(struct bpf_array *array, 1381 const struct bpf_prog *fp) 1382 { 1383 if (fp->kprobe_override) 1384 return false; 1385 1386 if (!array->owner_prog_type) { 1387 /* There's no owner yet where we could check for 1388 * compatibility. 1389 */ 1390 array->owner_prog_type = fp->type; 1391 array->owner_jited = fp->jited; 1392 1393 return true; 1394 } 1395 1396 return array->owner_prog_type == fp->type && 1397 array->owner_jited == fp->jited; 1398 } 1399 1400 static int bpf_check_tail_call(const struct bpf_prog *fp) 1401 { 1402 struct bpf_prog_aux *aux = fp->aux; 1403 int i; 1404 1405 for (i = 0; i < aux->used_map_cnt; i++) { 1406 struct bpf_map *map = aux->used_maps[i]; 1407 struct bpf_array *array; 1408 1409 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1410 continue; 1411 1412 array = container_of(map, struct bpf_array, map); 1413 if (!bpf_prog_array_compatible(array, fp)) 1414 return -EINVAL; 1415 } 1416 1417 return 0; 1418 } 1419 1420 /** 1421 * bpf_prog_select_runtime - select exec runtime for BPF program 1422 * @fp: bpf_prog populated with internal BPF program 1423 * @err: pointer to error variable 1424 * 1425 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1426 * The BPF program will be executed via BPF_PROG_RUN() macro. 1427 */ 1428 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1429 { 1430 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1431 1432 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1433 1434 /* eBPF JITs can rewrite the program in case constant 1435 * blinding is active. However, in case of error during 1436 * blinding, bpf_int_jit_compile() must always return a 1437 * valid program, which in this case would simply not 1438 * be JITed, but falls back to the interpreter. 1439 */ 1440 if (!bpf_prog_is_dev_bound(fp->aux)) { 1441 fp = bpf_int_jit_compile(fp); 1442 } else { 1443 *err = bpf_prog_offload_compile(fp); 1444 if (*err) 1445 return fp; 1446 } 1447 bpf_prog_lock_ro(fp); 1448 1449 /* The tail call compatibility check can only be done at 1450 * this late stage as we need to determine, if we deal 1451 * with JITed or non JITed program concatenations and not 1452 * all eBPF JITs might immediately support all features. 1453 */ 1454 *err = bpf_check_tail_call(fp); 1455 1456 return fp; 1457 } 1458 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1459 1460 static unsigned int __bpf_prog_ret1(const void *ctx, 1461 const struct bpf_insn *insn) 1462 { 1463 return 1; 1464 } 1465 1466 static struct bpf_prog_dummy { 1467 struct bpf_prog prog; 1468 } dummy_bpf_prog = { 1469 .prog = { 1470 .bpf_func = __bpf_prog_ret1, 1471 }, 1472 }; 1473 1474 /* to avoid allocating empty bpf_prog_array for cgroups that 1475 * don't have bpf program attached use one global 'empty_prog_array' 1476 * It will not be modified the caller of bpf_prog_array_alloc() 1477 * (since caller requested prog_cnt == 0) 1478 * that pointer should be 'freed' by bpf_prog_array_free() 1479 */ 1480 static struct { 1481 struct bpf_prog_array hdr; 1482 struct bpf_prog *null_prog; 1483 } empty_prog_array = { 1484 .null_prog = NULL, 1485 }; 1486 1487 struct bpf_prog_array __rcu *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1488 { 1489 if (prog_cnt) 1490 return kzalloc(sizeof(struct bpf_prog_array) + 1491 sizeof(struct bpf_prog *) * (prog_cnt + 1), 1492 flags); 1493 1494 return &empty_prog_array.hdr; 1495 } 1496 1497 void bpf_prog_array_free(struct bpf_prog_array __rcu *progs) 1498 { 1499 if (!progs || 1500 progs == (struct bpf_prog_array __rcu *)&empty_prog_array.hdr) 1501 return; 1502 kfree_rcu(progs, rcu); 1503 } 1504 1505 int bpf_prog_array_length(struct bpf_prog_array __rcu *progs) 1506 { 1507 struct bpf_prog **prog; 1508 u32 cnt = 0; 1509 1510 rcu_read_lock(); 1511 prog = rcu_dereference(progs)->progs; 1512 for (; *prog; prog++) 1513 if (*prog != &dummy_bpf_prog.prog) 1514 cnt++; 1515 rcu_read_unlock(); 1516 return cnt; 1517 } 1518 1519 int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *progs, 1520 __u32 __user *prog_ids, u32 cnt) 1521 { 1522 struct bpf_prog **prog; 1523 u32 i = 0, id; 1524 1525 rcu_read_lock(); 1526 prog = rcu_dereference(progs)->progs; 1527 for (; *prog; prog++) { 1528 if (*prog == &dummy_bpf_prog.prog) 1529 continue; 1530 id = (*prog)->aux->id; 1531 if (copy_to_user(prog_ids + i, &id, sizeof(id))) { 1532 rcu_read_unlock(); 1533 return -EFAULT; 1534 } 1535 if (++i == cnt) { 1536 prog++; 1537 break; 1538 } 1539 } 1540 rcu_read_unlock(); 1541 if (*prog) 1542 return -ENOSPC; 1543 return 0; 1544 } 1545 1546 void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *progs, 1547 struct bpf_prog *old_prog) 1548 { 1549 struct bpf_prog **prog = progs->progs; 1550 1551 for (; *prog; prog++) 1552 if (*prog == old_prog) { 1553 WRITE_ONCE(*prog, &dummy_bpf_prog.prog); 1554 break; 1555 } 1556 } 1557 1558 int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array, 1559 struct bpf_prog *exclude_prog, 1560 struct bpf_prog *include_prog, 1561 struct bpf_prog_array **new_array) 1562 { 1563 int new_prog_cnt, carry_prog_cnt = 0; 1564 struct bpf_prog **existing_prog; 1565 struct bpf_prog_array *array; 1566 int new_prog_idx = 0; 1567 1568 /* Figure out how many existing progs we need to carry over to 1569 * the new array. 1570 */ 1571 if (old_array) { 1572 existing_prog = old_array->progs; 1573 for (; *existing_prog; existing_prog++) { 1574 if (*existing_prog != exclude_prog && 1575 *existing_prog != &dummy_bpf_prog.prog) 1576 carry_prog_cnt++; 1577 if (*existing_prog == include_prog) 1578 return -EEXIST; 1579 } 1580 } 1581 1582 /* How many progs (not NULL) will be in the new array? */ 1583 new_prog_cnt = carry_prog_cnt; 1584 if (include_prog) 1585 new_prog_cnt += 1; 1586 1587 /* Do we have any prog (not NULL) in the new array? */ 1588 if (!new_prog_cnt) { 1589 *new_array = NULL; 1590 return 0; 1591 } 1592 1593 /* +1 as the end of prog_array is marked with NULL */ 1594 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1595 if (!array) 1596 return -ENOMEM; 1597 1598 /* Fill in the new prog array */ 1599 if (carry_prog_cnt) { 1600 existing_prog = old_array->progs; 1601 for (; *existing_prog; existing_prog++) 1602 if (*existing_prog != exclude_prog && 1603 *existing_prog != &dummy_bpf_prog.prog) 1604 array->progs[new_prog_idx++] = *existing_prog; 1605 } 1606 if (include_prog) 1607 array->progs[new_prog_idx++] = include_prog; 1608 array->progs[new_prog_idx] = NULL; 1609 *new_array = array; 1610 return 0; 1611 } 1612 1613 int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array, 1614 __u32 __user *prog_ids, u32 request_cnt, 1615 __u32 __user *prog_cnt) 1616 { 1617 u32 cnt = 0; 1618 1619 if (array) 1620 cnt = bpf_prog_array_length(array); 1621 1622 if (copy_to_user(prog_cnt, &cnt, sizeof(cnt))) 1623 return -EFAULT; 1624 1625 /* return early if user requested only program count or nothing to copy */ 1626 if (!request_cnt || !cnt) 1627 return 0; 1628 1629 return bpf_prog_array_copy_to_user(array, prog_ids, request_cnt); 1630 } 1631 1632 static void bpf_prog_free_deferred(struct work_struct *work) 1633 { 1634 struct bpf_prog_aux *aux; 1635 int i; 1636 1637 aux = container_of(work, struct bpf_prog_aux, work); 1638 if (bpf_prog_is_dev_bound(aux)) 1639 bpf_prog_offload_destroy(aux->prog); 1640 for (i = 0; i < aux->func_cnt; i++) 1641 bpf_jit_free(aux->func[i]); 1642 if (aux->func_cnt) { 1643 kfree(aux->func); 1644 bpf_prog_unlock_free(aux->prog); 1645 } else { 1646 bpf_jit_free(aux->prog); 1647 } 1648 } 1649 1650 /* Free internal BPF program */ 1651 void bpf_prog_free(struct bpf_prog *fp) 1652 { 1653 struct bpf_prog_aux *aux = fp->aux; 1654 1655 INIT_WORK(&aux->work, bpf_prog_free_deferred); 1656 schedule_work(&aux->work); 1657 } 1658 EXPORT_SYMBOL_GPL(bpf_prog_free); 1659 1660 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 1661 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 1662 1663 void bpf_user_rnd_init_once(void) 1664 { 1665 prandom_init_once(&bpf_user_rnd_state); 1666 } 1667 1668 BPF_CALL_0(bpf_user_rnd_u32) 1669 { 1670 /* Should someone ever have the rather unwise idea to use some 1671 * of the registers passed into this function, then note that 1672 * this function is called from native eBPF and classic-to-eBPF 1673 * transformations. Register assignments from both sides are 1674 * different, f.e. classic always sets fn(ctx, A, X) here. 1675 */ 1676 struct rnd_state *state; 1677 u32 res; 1678 1679 state = &get_cpu_var(bpf_user_rnd_state); 1680 res = prandom_u32_state(state); 1681 put_cpu_var(bpf_user_rnd_state); 1682 1683 return res; 1684 } 1685 1686 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 1687 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 1688 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 1689 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 1690 1691 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 1692 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 1693 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 1694 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 1695 1696 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 1697 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 1698 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 1699 const struct bpf_func_proto bpf_sock_map_update_proto __weak; 1700 1701 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 1702 { 1703 return NULL; 1704 } 1705 1706 u64 __weak 1707 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 1708 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 1709 { 1710 return -ENOTSUPP; 1711 } 1712 1713 /* Always built-in helper functions. */ 1714 const struct bpf_func_proto bpf_tail_call_proto = { 1715 .func = NULL, 1716 .gpl_only = false, 1717 .ret_type = RET_VOID, 1718 .arg1_type = ARG_PTR_TO_CTX, 1719 .arg2_type = ARG_CONST_MAP_PTR, 1720 .arg3_type = ARG_ANYTHING, 1721 }; 1722 1723 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 1724 * It is encouraged to implement bpf_int_jit_compile() instead, so that 1725 * eBPF and implicitly also cBPF can get JITed! 1726 */ 1727 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 1728 { 1729 return prog; 1730 } 1731 1732 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 1733 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 1734 */ 1735 void __weak bpf_jit_compile(struct bpf_prog *prog) 1736 { 1737 } 1738 1739 bool __weak bpf_helper_changes_pkt_data(void *func) 1740 { 1741 return false; 1742 } 1743 1744 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 1745 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 1746 */ 1747 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 1748 int len) 1749 { 1750 return -EFAULT; 1751 } 1752 1753 /* All definitions of tracepoints related to BPF. */ 1754 #define CREATE_TRACE_POINTS 1755 #include <linux/bpf_trace.h> 1756 1757 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 1758 1759 /* These are only used within the BPF_SYSCALL code */ 1760 #ifdef CONFIG_BPF_SYSCALL 1761 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type); 1762 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu); 1763 #endif 1764