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 kmemcheck_annotate_bitfield(fp, meta); 89 90 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags); 91 if (aux == NULL) { 92 vfree(fp); 93 return NULL; 94 } 95 96 fp->pages = size / PAGE_SIZE; 97 fp->aux = aux; 98 fp->aux->prog = fp; 99 100 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode); 101 102 return fp; 103 } 104 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 105 106 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 107 gfp_t gfp_extra_flags) 108 { 109 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 110 struct bpf_prog *fp; 111 u32 pages, delta; 112 int ret; 113 114 BUG_ON(fp_old == NULL); 115 116 size = round_up(size, PAGE_SIZE); 117 pages = size / PAGE_SIZE; 118 if (pages <= fp_old->pages) 119 return fp_old; 120 121 delta = pages - fp_old->pages; 122 ret = __bpf_prog_charge(fp_old->aux->user, delta); 123 if (ret) 124 return NULL; 125 126 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL); 127 if (fp == NULL) { 128 __bpf_prog_uncharge(fp_old->aux->user, delta); 129 } else { 130 kmemcheck_annotate_bitfield(fp, meta); 131 132 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 133 fp->pages = pages; 134 fp->aux->prog = fp; 135 136 /* We keep fp->aux from fp_old around in the new 137 * reallocated structure. 138 */ 139 fp_old->aux = NULL; 140 __bpf_prog_free(fp_old); 141 } 142 143 return fp; 144 } 145 146 void __bpf_prog_free(struct bpf_prog *fp) 147 { 148 kfree(fp->aux); 149 vfree(fp); 150 } 151 152 int bpf_prog_calc_tag(struct bpf_prog *fp) 153 { 154 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64); 155 u32 raw_size = bpf_prog_tag_scratch_size(fp); 156 u32 digest[SHA_DIGEST_WORDS]; 157 u32 ws[SHA_WORKSPACE_WORDS]; 158 u32 i, bsize, psize, blocks; 159 struct bpf_insn *dst; 160 bool was_ld_map; 161 u8 *raw, *todo; 162 __be32 *result; 163 __be64 *bits; 164 165 raw = vmalloc(raw_size); 166 if (!raw) 167 return -ENOMEM; 168 169 sha_init(digest); 170 memset(ws, 0, sizeof(ws)); 171 172 /* We need to take out the map fd for the digest calculation 173 * since they are unstable from user space side. 174 */ 175 dst = (void *)raw; 176 for (i = 0, was_ld_map = false; i < fp->len; i++) { 177 dst[i] = fp->insnsi[i]; 178 if (!was_ld_map && 179 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 180 dst[i].src_reg == BPF_PSEUDO_MAP_FD) { 181 was_ld_map = true; 182 dst[i].imm = 0; 183 } else if (was_ld_map && 184 dst[i].code == 0 && 185 dst[i].dst_reg == 0 && 186 dst[i].src_reg == 0 && 187 dst[i].off == 0) { 188 was_ld_map = false; 189 dst[i].imm = 0; 190 } else { 191 was_ld_map = false; 192 } 193 } 194 195 psize = bpf_prog_insn_size(fp); 196 memset(&raw[psize], 0, raw_size - psize); 197 raw[psize++] = 0x80; 198 199 bsize = round_up(psize, SHA_MESSAGE_BYTES); 200 blocks = bsize / SHA_MESSAGE_BYTES; 201 todo = raw; 202 if (bsize - psize >= sizeof(__be64)) { 203 bits = (__be64 *)(todo + bsize - sizeof(__be64)); 204 } else { 205 bits = (__be64 *)(todo + bsize + bits_offset); 206 blocks++; 207 } 208 *bits = cpu_to_be64((psize - 1) << 3); 209 210 while (blocks--) { 211 sha_transform(digest, todo, ws); 212 todo += SHA_MESSAGE_BYTES; 213 } 214 215 result = (__force __be32 *)digest; 216 for (i = 0; i < SHA_DIGEST_WORDS; i++) 217 result[i] = cpu_to_be32(digest[i]); 218 memcpy(fp->tag, result, sizeof(fp->tag)); 219 220 vfree(raw); 221 return 0; 222 } 223 224 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn) 225 { 226 return BPF_CLASS(insn->code) == BPF_JMP && 227 /* Call and Exit are both special jumps with no 228 * target inside the BPF instruction image. 229 */ 230 BPF_OP(insn->code) != BPF_CALL && 231 BPF_OP(insn->code) != BPF_EXIT; 232 } 233 234 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta) 235 { 236 struct bpf_insn *insn = prog->insnsi; 237 u32 i, insn_cnt = prog->len; 238 239 for (i = 0; i < insn_cnt; i++, insn++) { 240 if (!bpf_is_jmp_and_has_target(insn)) 241 continue; 242 243 /* Adjust offset of jmps if we cross boundaries. */ 244 if (i < pos && i + insn->off + 1 > pos) 245 insn->off += delta; 246 else if (i > pos + delta && i + insn->off + 1 <= pos + delta) 247 insn->off -= delta; 248 } 249 } 250 251 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 252 const struct bpf_insn *patch, u32 len) 253 { 254 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 255 struct bpf_prog *prog_adj; 256 257 /* Since our patchlet doesn't expand the image, we're done. */ 258 if (insn_delta == 0) { 259 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 260 return prog; 261 } 262 263 insn_adj_cnt = prog->len + insn_delta; 264 265 /* Several new instructions need to be inserted. Make room 266 * for them. Likely, there's no need for a new allocation as 267 * last page could have large enough tailroom. 268 */ 269 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 270 GFP_USER); 271 if (!prog_adj) 272 return NULL; 273 274 prog_adj->len = insn_adj_cnt; 275 276 /* Patching happens in 3 steps: 277 * 278 * 1) Move over tail of insnsi from next instruction onwards, 279 * so we can patch the single target insn with one or more 280 * new ones (patching is always from 1 to n insns, n > 0). 281 * 2) Inject new instructions at the target location. 282 * 3) Adjust branch offsets if necessary. 283 */ 284 insn_rest = insn_adj_cnt - off - len; 285 286 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 287 sizeof(*patch) * insn_rest); 288 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 289 290 bpf_adj_branches(prog_adj, off, insn_delta); 291 292 return prog_adj; 293 } 294 295 #ifdef CONFIG_BPF_JIT 296 static __always_inline void 297 bpf_get_prog_addr_region(const struct bpf_prog *prog, 298 unsigned long *symbol_start, 299 unsigned long *symbol_end) 300 { 301 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog); 302 unsigned long addr = (unsigned long)hdr; 303 304 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 305 306 *symbol_start = addr; 307 *symbol_end = addr + hdr->pages * PAGE_SIZE; 308 } 309 310 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym) 311 { 312 const char *end = sym + KSYM_NAME_LEN; 313 314 BUILD_BUG_ON(sizeof("bpf_prog_") + 315 sizeof(prog->tag) * 2 + 316 /* name has been null terminated. 317 * We should need +1 for the '_' preceding 318 * the name. However, the null character 319 * is double counted between the name and the 320 * sizeof("bpf_prog_") above, so we omit 321 * the +1 here. 322 */ 323 sizeof(prog->aux->name) > KSYM_NAME_LEN); 324 325 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 326 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 327 if (prog->aux->name[0]) 328 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 329 else 330 *sym = 0; 331 } 332 333 static __always_inline unsigned long 334 bpf_get_prog_addr_start(struct latch_tree_node *n) 335 { 336 unsigned long symbol_start, symbol_end; 337 const struct bpf_prog_aux *aux; 338 339 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 340 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 341 342 return symbol_start; 343 } 344 345 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 346 struct latch_tree_node *b) 347 { 348 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b); 349 } 350 351 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 352 { 353 unsigned long val = (unsigned long)key; 354 unsigned long symbol_start, symbol_end; 355 const struct bpf_prog_aux *aux; 356 357 aux = container_of(n, struct bpf_prog_aux, ksym_tnode); 358 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 359 360 if (val < symbol_start) 361 return -1; 362 if (val >= symbol_end) 363 return 1; 364 365 return 0; 366 } 367 368 static const struct latch_tree_ops bpf_tree_ops = { 369 .less = bpf_tree_less, 370 .comp = bpf_tree_comp, 371 }; 372 373 static DEFINE_SPINLOCK(bpf_lock); 374 static LIST_HEAD(bpf_kallsyms); 375 static struct latch_tree_root bpf_tree __cacheline_aligned; 376 377 int bpf_jit_kallsyms __read_mostly; 378 379 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux) 380 { 381 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode)); 382 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms); 383 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 384 } 385 386 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux) 387 { 388 if (list_empty(&aux->ksym_lnode)) 389 return; 390 391 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops); 392 list_del_rcu(&aux->ksym_lnode); 393 } 394 395 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 396 { 397 return fp->jited && !bpf_prog_was_classic(fp); 398 } 399 400 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp) 401 { 402 return list_empty(&fp->aux->ksym_lnode) || 403 fp->aux->ksym_lnode.prev == LIST_POISON2; 404 } 405 406 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 407 { 408 if (!bpf_prog_kallsyms_candidate(fp) || 409 !capable(CAP_SYS_ADMIN)) 410 return; 411 412 spin_lock_bh(&bpf_lock); 413 bpf_prog_ksym_node_add(fp->aux); 414 spin_unlock_bh(&bpf_lock); 415 } 416 417 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 418 { 419 if (!bpf_prog_kallsyms_candidate(fp)) 420 return; 421 422 spin_lock_bh(&bpf_lock); 423 bpf_prog_ksym_node_del(fp->aux); 424 spin_unlock_bh(&bpf_lock); 425 } 426 427 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr) 428 { 429 struct latch_tree_node *n; 430 431 if (!bpf_jit_kallsyms_enabled()) 432 return NULL; 433 434 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 435 return n ? 436 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog : 437 NULL; 438 } 439 440 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size, 441 unsigned long *off, char *sym) 442 { 443 unsigned long symbol_start, symbol_end; 444 struct bpf_prog *prog; 445 char *ret = NULL; 446 447 rcu_read_lock(); 448 prog = bpf_prog_kallsyms_find(addr); 449 if (prog) { 450 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end); 451 bpf_get_prog_name(prog, sym); 452 453 ret = sym; 454 if (size) 455 *size = symbol_end - symbol_start; 456 if (off) 457 *off = addr - symbol_start; 458 } 459 rcu_read_unlock(); 460 461 return ret; 462 } 463 464 bool is_bpf_text_address(unsigned long addr) 465 { 466 bool ret; 467 468 rcu_read_lock(); 469 ret = bpf_prog_kallsyms_find(addr) != NULL; 470 rcu_read_unlock(); 471 472 return ret; 473 } 474 475 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 476 char *sym) 477 { 478 unsigned long symbol_start, symbol_end; 479 struct bpf_prog_aux *aux; 480 unsigned int it = 0; 481 int ret = -ERANGE; 482 483 if (!bpf_jit_kallsyms_enabled()) 484 return ret; 485 486 rcu_read_lock(); 487 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) { 488 if (it++ != symnum) 489 continue; 490 491 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end); 492 bpf_get_prog_name(aux->prog, sym); 493 494 *value = symbol_start; 495 *type = BPF_SYM_ELF_TYPE; 496 497 ret = 0; 498 break; 499 } 500 rcu_read_unlock(); 501 502 return ret; 503 } 504 505 struct bpf_binary_header * 506 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 507 unsigned int alignment, 508 bpf_jit_fill_hole_t bpf_fill_ill_insns) 509 { 510 struct bpf_binary_header *hdr; 511 unsigned int size, hole, start; 512 513 /* Most of BPF filters are really small, but if some of them 514 * fill a page, allow at least 128 extra bytes to insert a 515 * random section of illegal instructions. 516 */ 517 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 518 hdr = module_alloc(size); 519 if (hdr == NULL) 520 return NULL; 521 522 /* Fill space with illegal/arch-dep instructions. */ 523 bpf_fill_ill_insns(hdr, size); 524 525 hdr->pages = size / PAGE_SIZE; 526 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 527 PAGE_SIZE - sizeof(*hdr)); 528 start = (get_random_int() % hole) & ~(alignment - 1); 529 530 /* Leave a random number of instructions before BPF code. */ 531 *image_ptr = &hdr->image[start]; 532 533 return hdr; 534 } 535 536 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 537 { 538 module_memfree(hdr); 539 } 540 541 /* This symbol is only overridden by archs that have different 542 * requirements than the usual eBPF JITs, f.e. when they only 543 * implement cBPF JIT, do not set images read-only, etc. 544 */ 545 void __weak bpf_jit_free(struct bpf_prog *fp) 546 { 547 if (fp->jited) { 548 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 549 550 bpf_jit_binary_unlock_ro(hdr); 551 bpf_jit_binary_free(hdr); 552 553 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 554 } 555 556 bpf_prog_unlock_free(fp); 557 } 558 559 int bpf_jit_harden __read_mostly; 560 561 static int bpf_jit_blind_insn(const struct bpf_insn *from, 562 const struct bpf_insn *aux, 563 struct bpf_insn *to_buff) 564 { 565 struct bpf_insn *to = to_buff; 566 u32 imm_rnd = get_random_int(); 567 s16 off; 568 569 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 570 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 571 572 if (from->imm == 0 && 573 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 574 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 575 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 576 goto out; 577 } 578 579 switch (from->code) { 580 case BPF_ALU | BPF_ADD | BPF_K: 581 case BPF_ALU | BPF_SUB | BPF_K: 582 case BPF_ALU | BPF_AND | BPF_K: 583 case BPF_ALU | BPF_OR | BPF_K: 584 case BPF_ALU | BPF_XOR | BPF_K: 585 case BPF_ALU | BPF_MUL | BPF_K: 586 case BPF_ALU | BPF_MOV | BPF_K: 587 case BPF_ALU | BPF_DIV | BPF_K: 588 case BPF_ALU | BPF_MOD | BPF_K: 589 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 590 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 591 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX); 592 break; 593 594 case BPF_ALU64 | BPF_ADD | BPF_K: 595 case BPF_ALU64 | BPF_SUB | BPF_K: 596 case BPF_ALU64 | BPF_AND | BPF_K: 597 case BPF_ALU64 | BPF_OR | BPF_K: 598 case BPF_ALU64 | BPF_XOR | BPF_K: 599 case BPF_ALU64 | BPF_MUL | BPF_K: 600 case BPF_ALU64 | BPF_MOV | BPF_K: 601 case BPF_ALU64 | BPF_DIV | BPF_K: 602 case BPF_ALU64 | BPF_MOD | BPF_K: 603 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 604 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 605 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX); 606 break; 607 608 case BPF_JMP | BPF_JEQ | BPF_K: 609 case BPF_JMP | BPF_JNE | BPF_K: 610 case BPF_JMP | BPF_JGT | BPF_K: 611 case BPF_JMP | BPF_JLT | BPF_K: 612 case BPF_JMP | BPF_JGE | BPF_K: 613 case BPF_JMP | BPF_JLE | BPF_K: 614 case BPF_JMP | BPF_JSGT | BPF_K: 615 case BPF_JMP | BPF_JSLT | BPF_K: 616 case BPF_JMP | BPF_JSGE | BPF_K: 617 case BPF_JMP | BPF_JSLE | BPF_K: 618 case BPF_JMP | BPF_JSET | BPF_K: 619 /* Accommodate for extra offset in case of a backjump. */ 620 off = from->off; 621 if (off < 0) 622 off -= 2; 623 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 624 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 625 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 626 break; 627 628 case BPF_LD | BPF_ABS | BPF_W: 629 case BPF_LD | BPF_ABS | BPF_H: 630 case BPF_LD | BPF_ABS | BPF_B: 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_LD_IND(from->code, BPF_REG_AX, 0); 634 break; 635 636 case BPF_LD | BPF_IND | BPF_W: 637 case BPF_LD | BPF_IND | BPF_H: 638 case BPF_LD | BPF_IND | 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_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg); 642 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0); 643 break; 644 645 case BPF_LD | BPF_IMM | BPF_DW: 646 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 647 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 648 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 649 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 650 break; 651 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 652 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 653 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 654 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 655 break; 656 657 case BPF_ST | BPF_MEM | BPF_DW: 658 case BPF_ST | BPF_MEM | BPF_W: 659 case BPF_ST | BPF_MEM | BPF_H: 660 case BPF_ST | BPF_MEM | BPF_B: 661 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 662 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 663 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 664 break; 665 } 666 out: 667 return to - to_buff; 668 } 669 670 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 671 gfp_t gfp_extra_flags) 672 { 673 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 674 struct bpf_prog *fp; 675 676 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL); 677 if (fp != NULL) { 678 kmemcheck_annotate_bitfield(fp, meta); 679 680 /* aux->prog still points to the fp_other one, so 681 * when promoting the clone to the real program, 682 * this still needs to be adapted. 683 */ 684 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 685 } 686 687 return fp; 688 } 689 690 static void bpf_prog_clone_free(struct bpf_prog *fp) 691 { 692 /* aux was stolen by the other clone, so we cannot free 693 * it from this path! It will be freed eventually by the 694 * other program on release. 695 * 696 * At this point, we don't need a deferred release since 697 * clone is guaranteed to not be locked. 698 */ 699 fp->aux = NULL; 700 __bpf_prog_free(fp); 701 } 702 703 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 704 { 705 /* We have to repoint aux->prog to self, as we don't 706 * know whether fp here is the clone or the original. 707 */ 708 fp->aux->prog = fp; 709 bpf_prog_clone_free(fp_other); 710 } 711 712 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 713 { 714 struct bpf_insn insn_buff[16], aux[2]; 715 struct bpf_prog *clone, *tmp; 716 int insn_delta, insn_cnt; 717 struct bpf_insn *insn; 718 int i, rewritten; 719 720 if (!bpf_jit_blinding_enabled()) 721 return prog; 722 723 clone = bpf_prog_clone_create(prog, GFP_USER); 724 if (!clone) 725 return ERR_PTR(-ENOMEM); 726 727 insn_cnt = clone->len; 728 insn = clone->insnsi; 729 730 for (i = 0; i < insn_cnt; i++, insn++) { 731 /* We temporarily need to hold the original ld64 insn 732 * so that we can still access the first part in the 733 * second blinding run. 734 */ 735 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 736 insn[1].code == 0) 737 memcpy(aux, insn, sizeof(aux)); 738 739 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff); 740 if (!rewritten) 741 continue; 742 743 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 744 if (!tmp) { 745 /* Patching may have repointed aux->prog during 746 * realloc from the original one, so we need to 747 * fix it up here on error. 748 */ 749 bpf_jit_prog_release_other(prog, clone); 750 return ERR_PTR(-ENOMEM); 751 } 752 753 clone = tmp; 754 insn_delta = rewritten - 1; 755 756 /* Walk new program and skip insns we just inserted. */ 757 insn = clone->insnsi + i + insn_delta; 758 insn_cnt += insn_delta; 759 i += insn_delta; 760 } 761 762 return clone; 763 } 764 #endif /* CONFIG_BPF_JIT */ 765 766 /* Base function for offset calculation. Needs to go into .text section, 767 * therefore keeping it non-static as well; will also be used by JITs 768 * anyway later on, so do not let the compiler omit it. 769 */ 770 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 771 { 772 return 0; 773 } 774 EXPORT_SYMBOL_GPL(__bpf_call_base); 775 776 /** 777 * __bpf_prog_run - run eBPF program on a given context 778 * @ctx: is the data we are operating on 779 * @insn: is the array of eBPF instructions 780 * 781 * Decode and execute eBPF instructions. 782 */ 783 static unsigned int ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, 784 u64 *stack) 785 { 786 u64 tmp; 787 static const void *jumptable[256] = { 788 [0 ... 255] = &&default_label, 789 /* Now overwrite non-defaults ... */ 790 /* 32 bit ALU operations */ 791 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X, 792 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K, 793 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X, 794 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K, 795 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X, 796 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K, 797 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X, 798 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K, 799 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X, 800 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K, 801 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X, 802 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K, 803 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X, 804 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K, 805 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X, 806 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K, 807 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X, 808 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K, 809 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X, 810 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K, 811 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X, 812 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K, 813 [BPF_ALU | BPF_NEG] = &&ALU_NEG, 814 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE, 815 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE, 816 /* 64 bit ALU operations */ 817 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X, 818 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K, 819 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X, 820 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K, 821 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X, 822 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K, 823 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X, 824 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K, 825 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X, 826 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K, 827 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X, 828 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K, 829 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X, 830 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K, 831 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X, 832 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K, 833 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X, 834 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K, 835 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X, 836 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K, 837 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X, 838 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K, 839 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X, 840 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K, 841 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG, 842 /* Call instruction */ 843 [BPF_JMP | BPF_CALL] = &&JMP_CALL, 844 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 845 /* Jumps */ 846 [BPF_JMP | BPF_JA] = &&JMP_JA, 847 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X, 848 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K, 849 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X, 850 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K, 851 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X, 852 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K, 853 [BPF_JMP | BPF_JLT | BPF_X] = &&JMP_JLT_X, 854 [BPF_JMP | BPF_JLT | BPF_K] = &&JMP_JLT_K, 855 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X, 856 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K, 857 [BPF_JMP | BPF_JLE | BPF_X] = &&JMP_JLE_X, 858 [BPF_JMP | BPF_JLE | BPF_K] = &&JMP_JLE_K, 859 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X, 860 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K, 861 [BPF_JMP | BPF_JSLT | BPF_X] = &&JMP_JSLT_X, 862 [BPF_JMP | BPF_JSLT | BPF_K] = &&JMP_JSLT_K, 863 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X, 864 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K, 865 [BPF_JMP | BPF_JSLE | BPF_X] = &&JMP_JSLE_X, 866 [BPF_JMP | BPF_JSLE | BPF_K] = &&JMP_JSLE_K, 867 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X, 868 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K, 869 /* Program return */ 870 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT, 871 /* Store instructions */ 872 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B, 873 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H, 874 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W, 875 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW, 876 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W, 877 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW, 878 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B, 879 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H, 880 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W, 881 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW, 882 /* Load instructions */ 883 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B, 884 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H, 885 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W, 886 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW, 887 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W, 888 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H, 889 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B, 890 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W, 891 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H, 892 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B, 893 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW, 894 }; 895 u32 tail_call_cnt = 0; 896 void *ptr; 897 int off; 898 899 #define CONT ({ insn++; goto select_insn; }) 900 #define CONT_JMP ({ insn++; goto select_insn; }) 901 902 select_insn: 903 goto *jumptable[insn->code]; 904 905 /* ALU */ 906 #define ALU(OPCODE, OP) \ 907 ALU64_##OPCODE##_X: \ 908 DST = DST OP SRC; \ 909 CONT; \ 910 ALU_##OPCODE##_X: \ 911 DST = (u32) DST OP (u32) SRC; \ 912 CONT; \ 913 ALU64_##OPCODE##_K: \ 914 DST = DST OP IMM; \ 915 CONT; \ 916 ALU_##OPCODE##_K: \ 917 DST = (u32) DST OP (u32) IMM; \ 918 CONT; 919 920 ALU(ADD, +) 921 ALU(SUB, -) 922 ALU(AND, &) 923 ALU(OR, |) 924 ALU(LSH, <<) 925 ALU(RSH, >>) 926 ALU(XOR, ^) 927 ALU(MUL, *) 928 #undef ALU 929 ALU_NEG: 930 DST = (u32) -DST; 931 CONT; 932 ALU64_NEG: 933 DST = -DST; 934 CONT; 935 ALU_MOV_X: 936 DST = (u32) SRC; 937 CONT; 938 ALU_MOV_K: 939 DST = (u32) IMM; 940 CONT; 941 ALU64_MOV_X: 942 DST = SRC; 943 CONT; 944 ALU64_MOV_K: 945 DST = IMM; 946 CONT; 947 LD_IMM_DW: 948 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 949 insn++; 950 CONT; 951 ALU64_ARSH_X: 952 (*(s64 *) &DST) >>= SRC; 953 CONT; 954 ALU64_ARSH_K: 955 (*(s64 *) &DST) >>= IMM; 956 CONT; 957 ALU64_MOD_X: 958 if (unlikely(SRC == 0)) 959 return 0; 960 div64_u64_rem(DST, SRC, &tmp); 961 DST = tmp; 962 CONT; 963 ALU_MOD_X: 964 if (unlikely(SRC == 0)) 965 return 0; 966 tmp = (u32) DST; 967 DST = do_div(tmp, (u32) SRC); 968 CONT; 969 ALU64_MOD_K: 970 div64_u64_rem(DST, IMM, &tmp); 971 DST = tmp; 972 CONT; 973 ALU_MOD_K: 974 tmp = (u32) DST; 975 DST = do_div(tmp, (u32) IMM); 976 CONT; 977 ALU64_DIV_X: 978 if (unlikely(SRC == 0)) 979 return 0; 980 DST = div64_u64(DST, SRC); 981 CONT; 982 ALU_DIV_X: 983 if (unlikely(SRC == 0)) 984 return 0; 985 tmp = (u32) DST; 986 do_div(tmp, (u32) SRC); 987 DST = (u32) tmp; 988 CONT; 989 ALU64_DIV_K: 990 DST = div64_u64(DST, IMM); 991 CONT; 992 ALU_DIV_K: 993 tmp = (u32) DST; 994 do_div(tmp, (u32) IMM); 995 DST = (u32) tmp; 996 CONT; 997 ALU_END_TO_BE: 998 switch (IMM) { 999 case 16: 1000 DST = (__force u16) cpu_to_be16(DST); 1001 break; 1002 case 32: 1003 DST = (__force u32) cpu_to_be32(DST); 1004 break; 1005 case 64: 1006 DST = (__force u64) cpu_to_be64(DST); 1007 break; 1008 } 1009 CONT; 1010 ALU_END_TO_LE: 1011 switch (IMM) { 1012 case 16: 1013 DST = (__force u16) cpu_to_le16(DST); 1014 break; 1015 case 32: 1016 DST = (__force u32) cpu_to_le32(DST); 1017 break; 1018 case 64: 1019 DST = (__force u64) cpu_to_le64(DST); 1020 break; 1021 } 1022 CONT; 1023 1024 /* CALL */ 1025 JMP_CALL: 1026 /* Function call scratches BPF_R1-BPF_R5 registers, 1027 * preserves BPF_R6-BPF_R9, and stores return value 1028 * into BPF_R0. 1029 */ 1030 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 1031 BPF_R4, BPF_R5); 1032 CONT; 1033 1034 JMP_TAIL_CALL: { 1035 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 1036 struct bpf_array *array = container_of(map, struct bpf_array, map); 1037 struct bpf_prog *prog; 1038 u32 index = BPF_R3; 1039 1040 if (unlikely(index >= array->map.max_entries)) 1041 goto out; 1042 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT)) 1043 goto out; 1044 1045 tail_call_cnt++; 1046 1047 prog = READ_ONCE(array->ptrs[index]); 1048 if (!prog) 1049 goto out; 1050 1051 /* ARG1 at this point is guaranteed to point to CTX from 1052 * the verifier side due to the fact that the tail call is 1053 * handeled like a helper, that is, bpf_tail_call_proto, 1054 * where arg1_type is ARG_PTR_TO_CTX. 1055 */ 1056 insn = prog->insnsi; 1057 goto select_insn; 1058 out: 1059 CONT; 1060 } 1061 /* JMP */ 1062 JMP_JA: 1063 insn += insn->off; 1064 CONT; 1065 JMP_JEQ_X: 1066 if (DST == SRC) { 1067 insn += insn->off; 1068 CONT_JMP; 1069 } 1070 CONT; 1071 JMP_JEQ_K: 1072 if (DST == IMM) { 1073 insn += insn->off; 1074 CONT_JMP; 1075 } 1076 CONT; 1077 JMP_JNE_X: 1078 if (DST != SRC) { 1079 insn += insn->off; 1080 CONT_JMP; 1081 } 1082 CONT; 1083 JMP_JNE_K: 1084 if (DST != IMM) { 1085 insn += insn->off; 1086 CONT_JMP; 1087 } 1088 CONT; 1089 JMP_JGT_X: 1090 if (DST > SRC) { 1091 insn += insn->off; 1092 CONT_JMP; 1093 } 1094 CONT; 1095 JMP_JGT_K: 1096 if (DST > IMM) { 1097 insn += insn->off; 1098 CONT_JMP; 1099 } 1100 CONT; 1101 JMP_JLT_X: 1102 if (DST < SRC) { 1103 insn += insn->off; 1104 CONT_JMP; 1105 } 1106 CONT; 1107 JMP_JLT_K: 1108 if (DST < IMM) { 1109 insn += insn->off; 1110 CONT_JMP; 1111 } 1112 CONT; 1113 JMP_JGE_X: 1114 if (DST >= SRC) { 1115 insn += insn->off; 1116 CONT_JMP; 1117 } 1118 CONT; 1119 JMP_JGE_K: 1120 if (DST >= IMM) { 1121 insn += insn->off; 1122 CONT_JMP; 1123 } 1124 CONT; 1125 JMP_JLE_X: 1126 if (DST <= SRC) { 1127 insn += insn->off; 1128 CONT_JMP; 1129 } 1130 CONT; 1131 JMP_JLE_K: 1132 if (DST <= IMM) { 1133 insn += insn->off; 1134 CONT_JMP; 1135 } 1136 CONT; 1137 JMP_JSGT_X: 1138 if (((s64) DST) > ((s64) SRC)) { 1139 insn += insn->off; 1140 CONT_JMP; 1141 } 1142 CONT; 1143 JMP_JSGT_K: 1144 if (((s64) DST) > ((s64) IMM)) { 1145 insn += insn->off; 1146 CONT_JMP; 1147 } 1148 CONT; 1149 JMP_JSLT_X: 1150 if (((s64) DST) < ((s64) SRC)) { 1151 insn += insn->off; 1152 CONT_JMP; 1153 } 1154 CONT; 1155 JMP_JSLT_K: 1156 if (((s64) DST) < ((s64) IMM)) { 1157 insn += insn->off; 1158 CONT_JMP; 1159 } 1160 CONT; 1161 JMP_JSGE_X: 1162 if (((s64) DST) >= ((s64) SRC)) { 1163 insn += insn->off; 1164 CONT_JMP; 1165 } 1166 CONT; 1167 JMP_JSGE_K: 1168 if (((s64) DST) >= ((s64) IMM)) { 1169 insn += insn->off; 1170 CONT_JMP; 1171 } 1172 CONT; 1173 JMP_JSLE_X: 1174 if (((s64) DST) <= ((s64) SRC)) { 1175 insn += insn->off; 1176 CONT_JMP; 1177 } 1178 CONT; 1179 JMP_JSLE_K: 1180 if (((s64) DST) <= ((s64) IMM)) { 1181 insn += insn->off; 1182 CONT_JMP; 1183 } 1184 CONT; 1185 JMP_JSET_X: 1186 if (DST & SRC) { 1187 insn += insn->off; 1188 CONT_JMP; 1189 } 1190 CONT; 1191 JMP_JSET_K: 1192 if (DST & IMM) { 1193 insn += insn->off; 1194 CONT_JMP; 1195 } 1196 CONT; 1197 JMP_EXIT: 1198 return BPF_R0; 1199 1200 /* STX and ST and LDX*/ 1201 #define LDST(SIZEOP, SIZE) \ 1202 STX_MEM_##SIZEOP: \ 1203 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 1204 CONT; \ 1205 ST_MEM_##SIZEOP: \ 1206 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 1207 CONT; \ 1208 LDX_MEM_##SIZEOP: \ 1209 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 1210 CONT; 1211 1212 LDST(B, u8) 1213 LDST(H, u16) 1214 LDST(W, u32) 1215 LDST(DW, u64) 1216 #undef LDST 1217 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */ 1218 atomic_add((u32) SRC, (atomic_t *)(unsigned long) 1219 (DST + insn->off)); 1220 CONT; 1221 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */ 1222 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long) 1223 (DST + insn->off)); 1224 CONT; 1225 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */ 1226 off = IMM; 1227 load_word: 1228 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only 1229 * appearing in the programs where ctx == skb 1230 * (see may_access_skb() in the verifier). All programs 1231 * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6, 1232 * bpf_convert_filter() saves it in BPF_R6, internal BPF 1233 * verifier will check that BPF_R6 == ctx. 1234 * 1235 * BPF_ABS and BPF_IND are wrappers of function calls, 1236 * so they scratch BPF_R1-BPF_R5 registers, preserve 1237 * BPF_R6-BPF_R9, and store return value into BPF_R0. 1238 * 1239 * Implicit input: 1240 * ctx == skb == BPF_R6 == CTX 1241 * 1242 * Explicit input: 1243 * SRC == any register 1244 * IMM == 32-bit immediate 1245 * 1246 * Output: 1247 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness 1248 */ 1249 1250 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp); 1251 if (likely(ptr != NULL)) { 1252 BPF_R0 = get_unaligned_be32(ptr); 1253 CONT; 1254 } 1255 1256 return 0; 1257 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */ 1258 off = IMM; 1259 load_half: 1260 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp); 1261 if (likely(ptr != NULL)) { 1262 BPF_R0 = get_unaligned_be16(ptr); 1263 CONT; 1264 } 1265 1266 return 0; 1267 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */ 1268 off = IMM; 1269 load_byte: 1270 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp); 1271 if (likely(ptr != NULL)) { 1272 BPF_R0 = *(u8 *)ptr; 1273 CONT; 1274 } 1275 1276 return 0; 1277 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */ 1278 off = IMM + SRC; 1279 goto load_word; 1280 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */ 1281 off = IMM + SRC; 1282 goto load_half; 1283 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */ 1284 off = IMM + SRC; 1285 goto load_byte; 1286 1287 default_label: 1288 /* If we ever reach this, we have a bug somewhere. */ 1289 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code); 1290 return 0; 1291 } 1292 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */ 1293 1294 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 1295 #define DEFINE_BPF_PROG_RUN(stack_size) \ 1296 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 1297 { \ 1298 u64 stack[stack_size / sizeof(u64)]; \ 1299 u64 regs[MAX_BPF_REG]; \ 1300 \ 1301 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 1302 ARG1 = (u64) (unsigned long) ctx; \ 1303 return ___bpf_prog_run(regs, insn, stack); \ 1304 } 1305 1306 #define EVAL1(FN, X) FN(X) 1307 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 1308 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 1309 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 1310 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 1311 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 1312 1313 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 1314 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 1315 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 1316 1317 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 1318 1319 static unsigned int (*interpreters[])(const void *ctx, 1320 const struct bpf_insn *insn) = { 1321 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 1322 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 1323 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 1324 }; 1325 1326 bool bpf_prog_array_compatible(struct bpf_array *array, 1327 const struct bpf_prog *fp) 1328 { 1329 if (!array->owner_prog_type) { 1330 /* There's no owner yet where we could check for 1331 * compatibility. 1332 */ 1333 array->owner_prog_type = fp->type; 1334 array->owner_jited = fp->jited; 1335 1336 return true; 1337 } 1338 1339 return array->owner_prog_type == fp->type && 1340 array->owner_jited == fp->jited; 1341 } 1342 1343 static int bpf_check_tail_call(const struct bpf_prog *fp) 1344 { 1345 struct bpf_prog_aux *aux = fp->aux; 1346 int i; 1347 1348 for (i = 0; i < aux->used_map_cnt; i++) { 1349 struct bpf_map *map = aux->used_maps[i]; 1350 struct bpf_array *array; 1351 1352 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 1353 continue; 1354 1355 array = container_of(map, struct bpf_array, map); 1356 if (!bpf_prog_array_compatible(array, fp)) 1357 return -EINVAL; 1358 } 1359 1360 return 0; 1361 } 1362 1363 /** 1364 * bpf_prog_select_runtime - select exec runtime for BPF program 1365 * @fp: bpf_prog populated with internal BPF program 1366 * @err: pointer to error variable 1367 * 1368 * Try to JIT eBPF program, if JIT is not available, use interpreter. 1369 * The BPF program will be executed via BPF_PROG_RUN() macro. 1370 */ 1371 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 1372 { 1373 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 1374 1375 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1]; 1376 1377 /* eBPF JITs can rewrite the program in case constant 1378 * blinding is active. However, in case of error during 1379 * blinding, bpf_int_jit_compile() must always return a 1380 * valid program, which in this case would simply not 1381 * be JITed, but falls back to the interpreter. 1382 */ 1383 if (!bpf_prog_is_dev_bound(fp->aux)) { 1384 fp = bpf_int_jit_compile(fp); 1385 } else { 1386 *err = bpf_prog_offload_compile(fp); 1387 if (*err) 1388 return fp; 1389 } 1390 bpf_prog_lock_ro(fp); 1391 1392 /* The tail call compatibility check can only be done at 1393 * this late stage as we need to determine, if we deal 1394 * with JITed or non JITed program concatenations and not 1395 * all eBPF JITs might immediately support all features. 1396 */ 1397 *err = bpf_check_tail_call(fp); 1398 1399 return fp; 1400 } 1401 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 1402 1403 static unsigned int __bpf_prog_ret1(const void *ctx, 1404 const struct bpf_insn *insn) 1405 { 1406 return 1; 1407 } 1408 1409 static struct bpf_prog_dummy { 1410 struct bpf_prog prog; 1411 } dummy_bpf_prog = { 1412 .prog = { 1413 .bpf_func = __bpf_prog_ret1, 1414 }, 1415 }; 1416 1417 /* to avoid allocating empty bpf_prog_array for cgroups that 1418 * don't have bpf program attached use one global 'empty_prog_array' 1419 * It will not be modified the caller of bpf_prog_array_alloc() 1420 * (since caller requested prog_cnt == 0) 1421 * that pointer should be 'freed' by bpf_prog_array_free() 1422 */ 1423 static struct { 1424 struct bpf_prog_array hdr; 1425 struct bpf_prog *null_prog; 1426 } empty_prog_array = { 1427 .null_prog = NULL, 1428 }; 1429 1430 struct bpf_prog_array __rcu *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 1431 { 1432 if (prog_cnt) 1433 return kzalloc(sizeof(struct bpf_prog_array) + 1434 sizeof(struct bpf_prog *) * (prog_cnt + 1), 1435 flags); 1436 1437 return &empty_prog_array.hdr; 1438 } 1439 1440 void bpf_prog_array_free(struct bpf_prog_array __rcu *progs) 1441 { 1442 if (!progs || 1443 progs == (struct bpf_prog_array __rcu *)&empty_prog_array.hdr) 1444 return; 1445 kfree_rcu(progs, rcu); 1446 } 1447 1448 int bpf_prog_array_length(struct bpf_prog_array __rcu *progs) 1449 { 1450 struct bpf_prog **prog; 1451 u32 cnt = 0; 1452 1453 rcu_read_lock(); 1454 prog = rcu_dereference(progs)->progs; 1455 for (; *prog; prog++) 1456 cnt++; 1457 rcu_read_unlock(); 1458 return cnt; 1459 } 1460 1461 int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *progs, 1462 __u32 __user *prog_ids, u32 cnt) 1463 { 1464 struct bpf_prog **prog; 1465 u32 i = 0, id; 1466 1467 rcu_read_lock(); 1468 prog = rcu_dereference(progs)->progs; 1469 for (; *prog; prog++) { 1470 id = (*prog)->aux->id; 1471 if (copy_to_user(prog_ids + i, &id, sizeof(id))) { 1472 rcu_read_unlock(); 1473 return -EFAULT; 1474 } 1475 if (++i == cnt) { 1476 prog++; 1477 break; 1478 } 1479 } 1480 rcu_read_unlock(); 1481 if (*prog) 1482 return -ENOSPC; 1483 return 0; 1484 } 1485 1486 void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *progs, 1487 struct bpf_prog *old_prog) 1488 { 1489 struct bpf_prog **prog = progs->progs; 1490 1491 for (; *prog; prog++) 1492 if (*prog == old_prog) { 1493 WRITE_ONCE(*prog, &dummy_bpf_prog.prog); 1494 break; 1495 } 1496 } 1497 1498 int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array, 1499 struct bpf_prog *exclude_prog, 1500 struct bpf_prog *include_prog, 1501 struct bpf_prog_array **new_array) 1502 { 1503 int new_prog_cnt, carry_prog_cnt = 0; 1504 struct bpf_prog **existing_prog; 1505 struct bpf_prog_array *array; 1506 int new_prog_idx = 0; 1507 1508 /* Figure out how many existing progs we need to carry over to 1509 * the new array. 1510 */ 1511 if (old_array) { 1512 existing_prog = old_array->progs; 1513 for (; *existing_prog; existing_prog++) { 1514 if (*existing_prog != exclude_prog && 1515 *existing_prog != &dummy_bpf_prog.prog) 1516 carry_prog_cnt++; 1517 if (*existing_prog == include_prog) 1518 return -EEXIST; 1519 } 1520 } 1521 1522 /* How many progs (not NULL) will be in the new array? */ 1523 new_prog_cnt = carry_prog_cnt; 1524 if (include_prog) 1525 new_prog_cnt += 1; 1526 1527 /* Do we have any prog (not NULL) in the new array? */ 1528 if (!new_prog_cnt) { 1529 *new_array = NULL; 1530 return 0; 1531 } 1532 1533 /* +1 as the end of prog_array is marked with NULL */ 1534 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 1535 if (!array) 1536 return -ENOMEM; 1537 1538 /* Fill in the new prog array */ 1539 if (carry_prog_cnt) { 1540 existing_prog = old_array->progs; 1541 for (; *existing_prog; existing_prog++) 1542 if (*existing_prog != exclude_prog && 1543 *existing_prog != &dummy_bpf_prog.prog) 1544 array->progs[new_prog_idx++] = *existing_prog; 1545 } 1546 if (include_prog) 1547 array->progs[new_prog_idx++] = include_prog; 1548 array->progs[new_prog_idx] = NULL; 1549 *new_array = array; 1550 return 0; 1551 } 1552 1553 static void bpf_prog_free_deferred(struct work_struct *work) 1554 { 1555 struct bpf_prog_aux *aux; 1556 1557 aux = container_of(work, struct bpf_prog_aux, work); 1558 if (bpf_prog_is_dev_bound(aux)) 1559 bpf_prog_offload_destroy(aux->prog); 1560 bpf_jit_free(aux->prog); 1561 } 1562 1563 /* Free internal BPF program */ 1564 void bpf_prog_free(struct bpf_prog *fp) 1565 { 1566 struct bpf_prog_aux *aux = fp->aux; 1567 1568 INIT_WORK(&aux->work, bpf_prog_free_deferred); 1569 schedule_work(&aux->work); 1570 } 1571 EXPORT_SYMBOL_GPL(bpf_prog_free); 1572 1573 /* RNG for unpriviledged user space with separated state from prandom_u32(). */ 1574 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 1575 1576 void bpf_user_rnd_init_once(void) 1577 { 1578 prandom_init_once(&bpf_user_rnd_state); 1579 } 1580 1581 BPF_CALL_0(bpf_user_rnd_u32) 1582 { 1583 /* Should someone ever have the rather unwise idea to use some 1584 * of the registers passed into this function, then note that 1585 * this function is called from native eBPF and classic-to-eBPF 1586 * transformations. Register assignments from both sides are 1587 * different, f.e. classic always sets fn(ctx, A, X) here. 1588 */ 1589 struct rnd_state *state; 1590 u32 res; 1591 1592 state = &get_cpu_var(bpf_user_rnd_state); 1593 res = prandom_u32_state(state); 1594 put_cpu_var(bpf_user_rnd_state); 1595 1596 return res; 1597 } 1598 1599 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 1600 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 1601 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 1602 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 1603 1604 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 1605 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 1606 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 1607 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 1608 1609 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 1610 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 1611 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 1612 const struct bpf_func_proto bpf_sock_map_update_proto __weak; 1613 1614 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 1615 { 1616 return NULL; 1617 } 1618 1619 u64 __weak 1620 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 1621 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 1622 { 1623 return -ENOTSUPP; 1624 } 1625 1626 /* Always built-in helper functions. */ 1627 const struct bpf_func_proto bpf_tail_call_proto = { 1628 .func = NULL, 1629 .gpl_only = false, 1630 .ret_type = RET_VOID, 1631 .arg1_type = ARG_PTR_TO_CTX, 1632 .arg2_type = ARG_CONST_MAP_PTR, 1633 .arg3_type = ARG_ANYTHING, 1634 }; 1635 1636 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 1637 * It is encouraged to implement bpf_int_jit_compile() instead, so that 1638 * eBPF and implicitly also cBPF can get JITed! 1639 */ 1640 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 1641 { 1642 return prog; 1643 } 1644 1645 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 1646 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 1647 */ 1648 void __weak bpf_jit_compile(struct bpf_prog *prog) 1649 { 1650 } 1651 1652 bool __weak bpf_helper_changes_pkt_data(void *func) 1653 { 1654 return false; 1655 } 1656 1657 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 1658 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 1659 */ 1660 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 1661 int len) 1662 { 1663 return -EFAULT; 1664 } 1665 1666 /* All definitions of tracepoints related to BPF. */ 1667 #define CREATE_TRACE_POINTS 1668 #include <linux/bpf_trace.h> 1669 1670 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 1671 1672 /* These are only used within the BPF_SYSCALL code */ 1673 #ifdef CONFIG_BPF_SYSCALL 1674 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type); 1675 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu); 1676 #endif 1677