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