kernel/bpf/core.c

   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_HIGHMEM | __GFP_ZERO |
  80                           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         kmemcheck_annotate_bitfield(fp, meta);
  90
  91         aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
  92         if (aux == NULL) {
  93                 vfree(fp);
  94                 return NULL;
  95         }
  96
  97         fp->pages = size / PAGE_SIZE;
  98         fp->aux = aux;
  99         fp->aux->prog = fp;
 100
 101         INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode);
 102
 103         return fp;
 104 }
 105 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
 106
 107 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
 108                                   gfp_t gfp_extra_flags)
 109 {
 110         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
 111                           gfp_extra_flags;
 112         struct bpf_prog *fp;
 113         u32 pages, delta;
 114         int ret;
 115
 116         BUG_ON(fp_old == NULL);
 117
 118         size = round_up(size, PAGE_SIZE);
 119         pages = size / PAGE_SIZE;
 120         if (pages <= fp_old->pages)
 121                 return fp_old;
 122
 123         delta = pages - fp_old->pages;
 124         ret = __bpf_prog_charge(fp_old->aux->user, delta);
 125         if (ret)
 126                 return NULL;
 127
 128         fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
 129         if (fp == NULL) {
 130                 __bpf_prog_uncharge(fp_old->aux->user, delta);
 131         } else {
 132                 kmemcheck_annotate_bitfield(fp, meta);
 133
 134                 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
 135                 fp->pages = pages;
 136                 fp->aux->prog = fp;
 137
 138                 /* We keep fp->aux from fp_old around in the new
 139                  * reallocated structure.
 140                  */
 141                 fp_old->aux = NULL;
 142                 __bpf_prog_free(fp_old);
 143         }
 144
 145         return fp;
 146 }
 147
 148 void __bpf_prog_free(struct bpf_prog *fp)
 149 {
 150         kfree(fp->aux);
 151         vfree(fp);
 152 }
 153
 154 int bpf_prog_calc_tag(struct bpf_prog *fp)
 155 {
 156         const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
 157         u32 raw_size = bpf_prog_tag_scratch_size(fp);
 158         u32 digest[SHA_DIGEST_WORDS];
 159         u32 ws[SHA_WORKSPACE_WORDS];
 160         u32 i, bsize, psize, blocks;
 161         struct bpf_insn *dst;
 162         bool was_ld_map;
 163         u8 *raw, *todo;
 164         __be32 *result;
 165         __be64 *bits;
 166
 167         raw = vmalloc(raw_size);
 168         if (!raw)
 169                 return -ENOMEM;
 170
 171         sha_init(digest);
 172         memset(ws, 0, sizeof(ws));
 173
 174         /* We need to take out the map fd for the digest calculation
 175          * since they are unstable from user space side.
 176          */
 177         dst = (void *)raw;
 178         for (i = 0, was_ld_map = false; i < fp->len; i++) {
 179                 dst[i] = fp->insnsi[i];
 180                 if (!was_ld_map &&
 181                     dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
 182                     dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
 183                         was_ld_map = true;
 184                         dst[i].imm = 0;
 185                 } else if (was_ld_map &&
 186                            dst[i].code == 0 &&
 187                            dst[i].dst_reg == 0 &&
 188                            dst[i].src_reg == 0 &&
 189                            dst[i].off == 0) {
 190                         was_ld_map = false;
 191                         dst[i].imm = 0;
 192                 } else {
 193                         was_ld_map = false;
 194                 }
 195         }
 196
 197         psize = bpf_prog_insn_size(fp);
 198         memset(&raw[psize], 0, raw_size - psize);
 199         raw[psize++] = 0x80;
 200
 201         bsize  = round_up(psize, SHA_MESSAGE_BYTES);
 202         blocks = bsize / SHA_MESSAGE_BYTES;
 203         todo   = raw;
 204         if (bsize - psize >= sizeof(__be64)) {
 205                 bits = (__be64 *)(todo + bsize - sizeof(__be64));
 206         } else {
 207                 bits = (__be64 *)(todo + bsize + bits_offset);
 208                 blocks++;
 209         }
 210         *bits = cpu_to_be64((psize - 1) << 3);
 211
 212         while (blocks--) {
 213                 sha_transform(digest, todo, ws);
 214                 todo += SHA_MESSAGE_BYTES;
 215         }
 216
 217         result = (__force __be32 *)digest;
 218         for (i = 0; i < SHA_DIGEST_WORDS; i++)
 219                 result[i] = cpu_to_be32(digest[i]);
 220         memcpy(fp->tag, result, sizeof(fp->tag));
 221
 222         vfree(raw);
 223         return 0;
 224 }
 225
 226 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
 227 {
 228         return BPF_CLASS(insn->code) == BPF_JMP  &&
 229                /* Call and Exit are both special jumps with no
 230                 * target inside the BPF instruction image.
 231                 */
 232                BPF_OP(insn->code) != BPF_CALL &&
 233                BPF_OP(insn->code) != BPF_EXIT;
 234 }
 235
 236 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
 237 {
 238         struct bpf_insn *insn = prog->insnsi;
 239         u32 i, insn_cnt = prog->len;
 240
 241         for (i = 0; i < insn_cnt; i++, insn++) {
 242                 if (!bpf_is_jmp_and_has_target(insn))
 243                         continue;
 244
 245                 /* Adjust offset of jmps if we cross boundaries. */
 246                 if (i < pos && i + insn->off + 1 > pos)
 247                         insn->off += delta;
 248                 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
 249                         insn->off -= delta;
 250         }
 251 }
 252
 253 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
 254                                        const struct bpf_insn *patch, u32 len)
 255 {
 256         u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
 257         struct bpf_prog *prog_adj;
 258
 259         /* Since our patchlet doesn't expand the image, we're done. */
 260         if (insn_delta == 0) {
 261                 memcpy(prog->insnsi + off, patch, sizeof(*patch));
 262                 return prog;
 263         }
 264
 265         insn_adj_cnt = prog->len + insn_delta;
 266
 267         /* Several new instructions need to be inserted. Make room
 268          * for them. Likely, there's no need for a new allocation as
 269          * last page could have large enough tailroom.
 270          */
 271         prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
 272                                     GFP_USER);
 273         if (!prog_adj)
 274                 return NULL;
 275
 276         prog_adj->len = insn_adj_cnt;
 277
 278         /* Patching happens in 3 steps:
 279          *
 280          * 1) Move over tail of insnsi from next instruction onwards,
 281          *    so we can patch the single target insn with one or more
 282          *    new ones (patching is always from 1 to n insns, n > 0).
 283          * 2) Inject new instructions at the target location.
 284          * 3) Adjust branch offsets if necessary.
 285          */
 286         insn_rest = insn_adj_cnt - off - len;
 287
 288         memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
 289                 sizeof(*patch) * insn_rest);
 290         memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
 291
 292         bpf_adj_branches(prog_adj, off, insn_delta);
 293
 294         return prog_adj;
 295 }
 296
 297 #ifdef CONFIG_BPF_JIT
 298 static __always_inline void
 299 bpf_get_prog_addr_region(const struct bpf_prog *prog,
 300                          unsigned long *symbol_start,
 301                          unsigned long *symbol_end)
 302 {
 303         const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
 304         unsigned long addr = (unsigned long)hdr;
 305
 306         WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
 307
 308         *symbol_start = addr;
 309         *symbol_end   = addr + hdr->pages * PAGE_SIZE;
 310 }
 311
 312 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym)
 313 {
 314         BUILD_BUG_ON(sizeof("bpf_prog_") +
 315                      sizeof(prog->tag) * 2 + 1 > KSYM_NAME_LEN);
 316
 317         sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
 318         sym  = bin2hex(sym, prog->tag, sizeof(prog->tag));
 319         *sym = 0;
 320 }
 321
 322 static __always_inline unsigned long
 323 bpf_get_prog_addr_start(struct latch_tree_node *n)
 324 {
 325         unsigned long symbol_start, symbol_end;
 326         const struct bpf_prog_aux *aux;
 327
 328         aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
 329         bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
 330
 331         return symbol_start;
 332 }
 333
 334 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
 335                                           struct latch_tree_node *b)
 336 {
 337         return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b);
 338 }
 339
 340 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
 341 {
 342         unsigned long val = (unsigned long)key;
 343         unsigned long symbol_start, symbol_end;
 344         const struct bpf_prog_aux *aux;
 345
 346         aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
 347         bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
 348
 349         if (val < symbol_start)
 350                 return -1;
 351         if (val >= symbol_end)
 352                 return  1;
 353
 354         return 0;
 355 }
 356
 357 static const struct latch_tree_ops bpf_tree_ops = {
 358         .less   = bpf_tree_less,
 359         .comp   = bpf_tree_comp,
 360 };
 361
 362 static DEFINE_SPINLOCK(bpf_lock);
 363 static LIST_HEAD(bpf_kallsyms);
 364 static struct latch_tree_root bpf_tree __cacheline_aligned;
 365
 366 int bpf_jit_kallsyms __read_mostly;
 367
 368 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux)
 369 {
 370         WARN_ON_ONCE(!list_empty(&aux->ksym_lnode));
 371         list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms);
 372         latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
 373 }
 374
 375 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux)
 376 {
 377         if (list_empty(&aux->ksym_lnode))
 378                 return;
 379
 380         latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
 381         list_del_rcu(&aux->ksym_lnode);
 382 }
 383
 384 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
 385 {
 386         return fp->jited && !bpf_prog_was_classic(fp);
 387 }
 388
 389 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
 390 {
 391         return list_empty(&fp->aux->ksym_lnode) ||
 392                fp->aux->ksym_lnode.prev == LIST_POISON2;
 393 }
 394
 395 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
 396 {
 397         if (!bpf_prog_kallsyms_candidate(fp) ||
 398             !capable(CAP_SYS_ADMIN))
 399                 return;
 400
 401         spin_lock_bh(&bpf_lock);
 402         bpf_prog_ksym_node_add(fp->aux);
 403         spin_unlock_bh(&bpf_lock);
 404 }
 405
 406 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
 407 {
 408         if (!bpf_prog_kallsyms_candidate(fp))
 409                 return;
 410
 411         spin_lock_bh(&bpf_lock);
 412         bpf_prog_ksym_node_del(fp->aux);
 413         spin_unlock_bh(&bpf_lock);
 414 }
 415
 416 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr)
 417 {
 418         struct latch_tree_node *n;
 419
 420         if (!bpf_jit_kallsyms_enabled())
 421                 return NULL;
 422
 423         n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
 424         return n ?
 425                container_of(n, struct bpf_prog_aux, ksym_tnode)->prog :
 426                NULL;
 427 }
 428
 429 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
 430                                  unsigned long *off, char *sym)
 431 {
 432         unsigned long symbol_start, symbol_end;
 433         struct bpf_prog *prog;
 434         char *ret = NULL;
 435
 436         rcu_read_lock();
 437         prog = bpf_prog_kallsyms_find(addr);
 438         if (prog) {
 439                 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end);
 440                 bpf_get_prog_name(prog, sym);
 441
 442                 ret = sym;
 443                 if (size)
 444                         *size = symbol_end - symbol_start;
 445                 if (off)
 446                         *off  = addr - symbol_start;
 447         }
 448         rcu_read_unlock();
 449
 450         return ret;
 451 }
 452
 453 bool is_bpf_text_address(unsigned long addr)
 454 {
 455         bool ret;
 456
 457         rcu_read_lock();
 458         ret = bpf_prog_kallsyms_find(addr) != NULL;
 459         rcu_read_unlock();
 460
 461         return ret;
 462 }
 463
 464 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
 465                     char *sym)
 466 {
 467         unsigned long symbol_start, symbol_end;
 468         struct bpf_prog_aux *aux;
 469         unsigned int it = 0;
 470         int ret = -ERANGE;
 471
 472         if (!bpf_jit_kallsyms_enabled())
 473                 return ret;
 474
 475         rcu_read_lock();
 476         list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) {
 477                 if (it++ != symnum)
 478                         continue;
 479
 480                 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
 481                 bpf_get_prog_name(aux->prog, sym);
 482
 483                 *value = symbol_start;
 484                 *type  = BPF_SYM_ELF_TYPE;
 485
 486                 ret = 0;
 487                 break;
 488         }
 489         rcu_read_unlock();
 490
 491         return ret;
 492 }
 493
 494 struct bpf_binary_header *
 495 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
 496                      unsigned int alignment,
 497                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
 498 {
 499         struct bpf_binary_header *hdr;
 500         unsigned int size, hole, start;
 501
 502         /* Most of BPF filters are really small, but if some of them
 503          * fill a page, allow at least 128 extra bytes to insert a
 504          * random section of illegal instructions.
 505          */
 506         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
 507         hdr = module_alloc(size);
 508         if (hdr == NULL)
 509                 return NULL;
 510
 511         /* Fill space with illegal/arch-dep instructions. */
 512         bpf_fill_ill_insns(hdr, size);
 513
 514         hdr->pages = size / PAGE_SIZE;
 515         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
 516                      PAGE_SIZE - sizeof(*hdr));
 517         start = (get_random_int() % hole) & ~(alignment - 1);
 518
 519         /* Leave a random number of instructions before BPF code. */
 520         *image_ptr = &hdr->image[start];
 521
 522         return hdr;
 523 }
 524
 525 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
 526 {
 527         module_memfree(hdr);
 528 }
 529
 530 /* This symbol is only overridden by archs that have different
 531  * requirements than the usual eBPF JITs, f.e. when they only
 532  * implement cBPF JIT, do not set images read-only, etc.
 533  */
 534 void __weak bpf_jit_free(struct bpf_prog *fp)
 535 {
 536         if (fp->jited) {
 537                 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
 538
 539                 bpf_jit_binary_unlock_ro(hdr);
 540                 bpf_jit_binary_free(hdr);
 541
 542                 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
 543         }
 544
 545         bpf_prog_unlock_free(fp);
 546 }
 547
 548 int bpf_jit_harden __read_mostly;
 549
 550 static int bpf_jit_blind_insn(const struct bpf_insn *from,
 551                               const struct bpf_insn *aux,
 552                               struct bpf_insn *to_buff)
 553 {
 554         struct bpf_insn *to = to_buff;
 555         u32 imm_rnd = get_random_int();
 556         s16 off;
 557
 558         BUILD_BUG_ON(BPF_REG_AX  + 1 != MAX_BPF_JIT_REG);
 559         BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
 560
 561         if (from->imm == 0 &&
 562             (from->code == (BPF_ALU   | BPF_MOV | BPF_K) ||
 563              from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
 564                 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
 565                 goto out;
 566         }
 567
 568         switch (from->code) {
 569         case BPF_ALU | BPF_ADD | BPF_K:
 570         case BPF_ALU | BPF_SUB | BPF_K:
 571         case BPF_ALU | BPF_AND | BPF_K:
 572         case BPF_ALU | BPF_OR  | BPF_K:
 573         case BPF_ALU | BPF_XOR | BPF_K:
 574         case BPF_ALU | BPF_MUL | BPF_K:
 575         case BPF_ALU | BPF_MOV | BPF_K:
 576         case BPF_ALU | BPF_DIV | BPF_K:
 577         case BPF_ALU | BPF_MOD | BPF_K:
 578                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 579                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 580                 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
 581                 break;
 582
 583         case BPF_ALU64 | BPF_ADD | BPF_K:
 584         case BPF_ALU64 | BPF_SUB | BPF_K:
 585         case BPF_ALU64 | BPF_AND | BPF_K:
 586         case BPF_ALU64 | BPF_OR  | BPF_K:
 587         case BPF_ALU64 | BPF_XOR | BPF_K:
 588         case BPF_ALU64 | BPF_MUL | BPF_K:
 589         case BPF_ALU64 | BPF_MOV | BPF_K:
 590         case BPF_ALU64 | BPF_DIV | BPF_K:
 591         case BPF_ALU64 | BPF_MOD | BPF_K:
 592                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 593                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 594                 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
 595                 break;
 596
 597         case BPF_JMP | BPF_JEQ  | BPF_K:
 598         case BPF_JMP | BPF_JNE  | BPF_K:
 599         case BPF_JMP | BPF_JGT  | BPF_K:
 600         case BPF_JMP | BPF_JGE  | BPF_K:
 601         case BPF_JMP | BPF_JSGT | BPF_K:
 602         case BPF_JMP | BPF_JSGE | BPF_K:
 603         case BPF_JMP | BPF_JSET | BPF_K:
 604                 /* Accommodate for extra offset in case of a backjump. */
 605                 off = from->off;
 606                 if (off < 0)
 607                         off -= 2;
 608                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 609                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 610                 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
 611                 break;
 612
 613         case BPF_LD | BPF_ABS | BPF_W:
 614         case BPF_LD | BPF_ABS | BPF_H:
 615         case BPF_LD | BPF_ABS | BPF_B:
 616                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 617                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 618                 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
 619                 break;
 620
 621         case BPF_LD | BPF_IND | BPF_W:
 622         case BPF_LD | BPF_IND | BPF_H:
 623         case BPF_LD | BPF_IND | BPF_B:
 624                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 625                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 626                 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
 627                 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
 628                 break;
 629
 630         case BPF_LD | BPF_IMM | BPF_DW:
 631                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
 632                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 633                 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
 634                 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
 635                 break;
 636         case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
 637                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
 638                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 639                 *to++ = BPF_ALU64_REG(BPF_OR,  aux[0].dst_reg, BPF_REG_AX);
 640                 break;
 641
 642         case BPF_ST | BPF_MEM | BPF_DW:
 643         case BPF_ST | BPF_MEM | BPF_W:
 644         case BPF_ST | BPF_MEM | BPF_H:
 645         case BPF_ST | BPF_MEM | BPF_B:
 646                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 647                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 648                 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
 649                 break;
 650         }
 651 out:
 652         return to - to_buff;
 653 }
 654
 655 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
 656                                               gfp_t gfp_extra_flags)
 657 {
 658         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
 659                           gfp_extra_flags;
 660         struct bpf_prog *fp;
 661
 662         fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
 663         if (fp != NULL) {
 664                 kmemcheck_annotate_bitfield(fp, meta);
 665
 666                 /* aux->prog still points to the fp_other one, so
 667                  * when promoting the clone to the real program,
 668                  * this still needs to be adapted.
 669                  */
 670                 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
 671         }
 672
 673         return fp;
 674 }
 675
 676 static void bpf_prog_clone_free(struct bpf_prog *fp)
 677 {
 678         /* aux was stolen by the other clone, so we cannot free
 679          * it from this path! It will be freed eventually by the
 680          * other program on release.
 681          *
 682          * At this point, we don't need a deferred release since
 683          * clone is guaranteed to not be locked.
 684          */
 685         fp->aux = NULL;
 686         __bpf_prog_free(fp);
 687 }
 688
 689 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
 690 {
 691         /* We have to repoint aux->prog to self, as we don't
 692          * know whether fp here is the clone or the original.
 693          */
 694         fp->aux->prog = fp;
 695         bpf_prog_clone_free(fp_other);
 696 }
 697
 698 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
 699 {
 700         struct bpf_insn insn_buff[16], aux[2];
 701         struct bpf_prog *clone, *tmp;
 702         int insn_delta, insn_cnt;
 703         struct bpf_insn *insn;
 704         int i, rewritten;
 705
 706         if (!bpf_jit_blinding_enabled())
 707                 return prog;
 708
 709         clone = bpf_prog_clone_create(prog, GFP_USER);
 710         if (!clone)
 711                 return ERR_PTR(-ENOMEM);
 712
 713         insn_cnt = clone->len;
 714         insn = clone->insnsi;
 715
 716         for (i = 0; i < insn_cnt; i++, insn++) {
 717                 /* We temporarily need to hold the original ld64 insn
 718                  * so that we can still access the first part in the
 719                  * second blinding run.
 720                  */
 721                 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
 722                     insn[1].code == 0)
 723                         memcpy(aux, insn, sizeof(aux));
 724
 725                 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
 726                 if (!rewritten)
 727                         continue;
 728
 729                 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
 730                 if (!tmp) {
 731                         /* Patching may have repointed aux->prog during
 732                          * realloc from the original one, so we need to
 733                          * fix it up here on error.
 734                          */
 735                         bpf_jit_prog_release_other(prog, clone);
 736                         return ERR_PTR(-ENOMEM);
 737                 }
 738
 739                 clone = tmp;
 740                 insn_delta = rewritten - 1;
 741
 742                 /* Walk new program and skip insns we just inserted. */
 743                 insn = clone->insnsi + i + insn_delta;
 744                 insn_cnt += insn_delta;
 745                 i        += insn_delta;
 746         }
 747
 748         return clone;
 749 }
 750 #endif /* CONFIG_BPF_JIT */
 751
 752 /* Base function for offset calculation. Needs to go into .text section,
 753  * therefore keeping it non-static as well; will also be used by JITs
 754  * anyway later on, so do not let the compiler omit it.
 755  */
 756 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 757 {
 758         return 0;
 759 }
 760 EXPORT_SYMBOL_GPL(__bpf_call_base);
 761
 762 /**
 763  *      __bpf_prog_run - run eBPF program on a given context
 764  *      @ctx: is the data we are operating on
 765  *      @insn: is the array of eBPF instructions
 766  *
 767  * Decode and execute eBPF instructions.
 768  */
 769 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
 770 {
 771         u64 stack[MAX_BPF_STACK / sizeof(u64)];
 772         u64 regs[MAX_BPF_REG], tmp;
 773         static const void *jumptable[256] = {
 774                 [0 ... 255] = &&default_label,
 775                 /* Now overwrite non-defaults ... */
 776                 /* 32 bit ALU operations */
 777                 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
 778                 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
 779                 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
 780                 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
 781                 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
 782                 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
 783                 [BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
 784                 [BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
 785                 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
 786                 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
 787                 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
 788                 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
 789                 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
 790                 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
 791                 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
 792                 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
 793                 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
 794                 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
 795                 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
 796                 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
 797                 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
 798                 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
 799                 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
 800                 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
 801                 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
 802                 /* 64 bit ALU operations */
 803                 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
 804                 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
 805                 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
 806                 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
 807                 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
 808                 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
 809                 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
 810                 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
 811                 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
 812                 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
 813                 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
 814                 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
 815                 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
 816                 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
 817                 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
 818                 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
 819                 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
 820                 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
 821                 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
 822                 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
 823                 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
 824                 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
 825                 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
 826                 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
 827                 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
 828                 /* Call instruction */
 829                 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
 830                 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
 831                 /* Jumps */
 832                 [BPF_JMP | BPF_JA] = &&JMP_JA,
 833                 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
 834                 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
 835                 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
 836                 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
 837                 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
 838                 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
 839                 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
 840                 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
 841                 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
 842                 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
 843                 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
 844                 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
 845                 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
 846                 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
 847                 /* Program return */
 848                 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
 849                 /* Store instructions */
 850                 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
 851                 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
 852                 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
 853                 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
 854                 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
 855                 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
 856                 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
 857                 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
 858                 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
 859                 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
 860                 /* Load instructions */
 861                 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
 862                 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
 863                 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
 864                 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
 865                 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
 866                 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
 867                 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
 868                 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
 869                 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
 870                 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
 871                 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
 872         };
 873         u32 tail_call_cnt = 0;
 874         void *ptr;
 875         int off;
 876
 877 #define CONT     ({ insn++; goto select_insn; })
 878 #define CONT_JMP ({ insn++; goto select_insn; })
 879
 880         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
 881         ARG1 = (u64) (unsigned long) ctx;
 882
 883 select_insn:
 884         goto *jumptable[insn->code];
 885
 886         /* ALU */
 887 #define ALU(OPCODE, OP)                 \
 888         ALU64_##OPCODE##_X:             \
 889                 DST = DST OP SRC;       \
 890                 CONT;                   \
 891         ALU_##OPCODE##_X:               \
 892                 DST = (u32) DST OP (u32) SRC;   \
 893                 CONT;                   \
 894         ALU64_##OPCODE##_K:             \
 895                 DST = DST OP IMM;               \
 896                 CONT;                   \
 897         ALU_##OPCODE##_K:               \
 898                 DST = (u32) DST OP (u32) IMM;   \
 899                 CONT;
 900
 901         ALU(ADD,  +)
 902         ALU(SUB,  -)
 903         ALU(AND,  &)
 904         ALU(OR,   |)
 905         ALU(LSH, <<)
 906         ALU(RSH, >>)
 907         ALU(XOR,  ^)
 908         ALU(MUL,  *)
 909 #undef ALU
 910         ALU_NEG:
 911                 DST = (u32) -DST;
 912                 CONT;
 913         ALU64_NEG:
 914                 DST = -DST;
 915                 CONT;
 916         ALU_MOV_X:
 917                 DST = (u32) SRC;
 918                 CONT;
 919         ALU_MOV_K:
 920                 DST = (u32) IMM;
 921                 CONT;
 922         ALU64_MOV_X:
 923                 DST = SRC;
 924                 CONT;
 925         ALU64_MOV_K:
 926                 DST = IMM;
 927                 CONT;
 928         LD_IMM_DW:
 929                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
 930                 insn++;
 931                 CONT;
 932         ALU64_ARSH_X:
 933                 (*(s64 *) &DST) >>= SRC;
 934                 CONT;
 935         ALU64_ARSH_K:
 936                 (*(s64 *) &DST) >>= IMM;
 937                 CONT;
 938         ALU64_MOD_X:
 939                 if (unlikely(SRC == 0))
 940                         return 0;
 941                 div64_u64_rem(DST, SRC, &tmp);
 942                 DST = tmp;
 943                 CONT;
 944         ALU_MOD_X:
 945                 if (unlikely(SRC == 0))
 946                         return 0;
 947                 tmp = (u32) DST;
 948                 DST = do_div(tmp, (u32) SRC);
 949                 CONT;
 950         ALU64_MOD_K:
 951                 div64_u64_rem(DST, IMM, &tmp);
 952                 DST = tmp;
 953                 CONT;
 954         ALU_MOD_K:
 955                 tmp = (u32) DST;
 956                 DST = do_div(tmp, (u32) IMM);
 957                 CONT;
 958         ALU64_DIV_X:
 959                 if (unlikely(SRC == 0))
 960                         return 0;
 961                 DST = div64_u64(DST, SRC);
 962                 CONT;
 963         ALU_DIV_X:
 964                 if (unlikely(SRC == 0))
 965                         return 0;
 966                 tmp = (u32) DST;
 967                 do_div(tmp, (u32) SRC);
 968                 DST = (u32) tmp;
 969                 CONT;
 970         ALU64_DIV_K:
 971                 DST = div64_u64(DST, IMM);
 972                 CONT;
 973         ALU_DIV_K:
 974                 tmp = (u32) DST;
 975                 do_div(tmp, (u32) IMM);
 976                 DST = (u32) tmp;
 977                 CONT;
 978         ALU_END_TO_BE:
 979                 switch (IMM) {
 980                 case 16:
 981                         DST = (__force u16) cpu_to_be16(DST);
 982                         break;
 983                 case 32:
 984                         DST = (__force u32) cpu_to_be32(DST);
 985                         break;
 986                 case 64:
 987                         DST = (__force u64) cpu_to_be64(DST);
 988                         break;
 989                 }
 990                 CONT;
 991         ALU_END_TO_LE:
 992                 switch (IMM) {
 993                 case 16:
 994                         DST = (__force u16) cpu_to_le16(DST);
 995                         break;
 996                 case 32:
 997                         DST = (__force u32) cpu_to_le32(DST);
 998                         break;
 999                 case 64:
1000                         DST = (__force u64) cpu_to_le64(DST);
1001                         break;
1002                 }
1003                 CONT;
1004
1005         /* CALL */
1006         JMP_CALL:
1007                 /* Function call scratches BPF_R1-BPF_R5 registers,
1008                  * preserves BPF_R6-BPF_R9, and stores return value
1009                  * into BPF_R0.
1010                  */
1011                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1012                                                        BPF_R4, BPF_R5);
1013                 CONT;
1014
1015         JMP_TAIL_CALL: {
1016                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1017                 struct bpf_array *array = container_of(map, struct bpf_array, map);
1018                 struct bpf_prog *prog;
1019                 u64 index = BPF_R3;
1020
1021                 if (unlikely(index >= array->map.max_entries))
1022                         goto out;
1023                 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1024                         goto out;
1025
1026                 tail_call_cnt++;
1027
1028                 prog = READ_ONCE(array->ptrs[index]);
1029                 if (!prog)
1030                         goto out;
1031
1032                 /* ARG1 at this point is guaranteed to point to CTX from
1033                  * the verifier side due to the fact that the tail call is
1034                  * handeled like a helper, that is, bpf_tail_call_proto,
1035                  * where arg1_type is ARG_PTR_TO_CTX.
1036                  */
1037                 insn = prog->insnsi;
1038                 goto select_insn;
1039 out:
1040                 CONT;
1041         }
1042         /* JMP */
1043         JMP_JA:
1044                 insn += insn->off;
1045                 CONT;
1046         JMP_JEQ_X:
1047                 if (DST == SRC) {
1048                         insn += insn->off;
1049                         CONT_JMP;
1050                 }
1051                 CONT;
1052         JMP_JEQ_K:
1053                 if (DST == IMM) {
1054                         insn += insn->off;
1055                         CONT_JMP;
1056                 }
1057                 CONT;
1058         JMP_JNE_X:
1059                 if (DST != SRC) {
1060                         insn += insn->off;
1061                         CONT_JMP;
1062                 }
1063                 CONT;
1064         JMP_JNE_K:
1065                 if (DST != IMM) {
1066                         insn += insn->off;
1067                         CONT_JMP;
1068                 }
1069                 CONT;
1070         JMP_JGT_X:
1071                 if (DST > SRC) {
1072                         insn += insn->off;
1073                         CONT_JMP;
1074                 }
1075                 CONT;
1076         JMP_JGT_K:
1077                 if (DST > IMM) {
1078                         insn += insn->off;
1079                         CONT_JMP;
1080                 }
1081                 CONT;
1082         JMP_JGE_X:
1083                 if (DST >= SRC) {
1084                         insn += insn->off;
1085                         CONT_JMP;
1086                 }
1087                 CONT;
1088         JMP_JGE_K:
1089                 if (DST >= IMM) {
1090                         insn += insn->off;
1091                         CONT_JMP;
1092                 }
1093                 CONT;
1094         JMP_JSGT_X:
1095                 if (((s64) DST) > ((s64) SRC)) {
1096                         insn += insn->off;
1097                         CONT_JMP;
1098                 }
1099                 CONT;
1100         JMP_JSGT_K:
1101                 if (((s64) DST) > ((s64) IMM)) {
1102                         insn += insn->off;
1103                         CONT_JMP;
1104                 }
1105                 CONT;
1106         JMP_JSGE_X:
1107                 if (((s64) DST) >= ((s64) SRC)) {
1108                         insn += insn->off;
1109                         CONT_JMP;
1110                 }
1111                 CONT;
1112         JMP_JSGE_K:
1113                 if (((s64) DST) >= ((s64) IMM)) {
1114                         insn += insn->off;
1115                         CONT_JMP;
1116                 }
1117                 CONT;
1118         JMP_JSET_X:
1119                 if (DST & SRC) {
1120                         insn += insn->off;
1121                         CONT_JMP;
1122                 }
1123                 CONT;
1124         JMP_JSET_K:
1125                 if (DST & IMM) {
1126                         insn += insn->off;
1127                         CONT_JMP;
1128                 }
1129                 CONT;
1130         JMP_EXIT:
1131                 return BPF_R0;
1132
1133         /* STX and ST and LDX*/
1134 #define LDST(SIZEOP, SIZE)                                              \
1135         STX_MEM_##SIZEOP:                                               \
1136                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
1137                 CONT;                                                   \
1138         ST_MEM_##SIZEOP:                                                \
1139                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
1140                 CONT;                                                   \
1141         LDX_MEM_##SIZEOP:                                               \
1142                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
1143                 CONT;
1144
1145         LDST(B,   u8)
1146         LDST(H,  u16)
1147         LDST(W,  u32)
1148         LDST(DW, u64)
1149 #undef LDST
1150         STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1151                 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1152                            (DST + insn->off));
1153                 CONT;
1154         STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1155                 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1156                              (DST + insn->off));
1157                 CONT;
1158         LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
1159                 off = IMM;
1160 load_word:
1161                 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only
1162                  * appearing in the programs where ctx == skb
1163                  * (see may_access_skb() in the verifier). All programs
1164                  * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6,
1165                  * bpf_convert_filter() saves it in BPF_R6, internal BPF
1166                  * verifier will check that BPF_R6 == ctx.
1167                  *
1168                  * BPF_ABS and BPF_IND are wrappers of function calls,
1169                  * so they scratch BPF_R1-BPF_R5 registers, preserve
1170                  * BPF_R6-BPF_R9, and store return value into BPF_R0.
1171                  *
1172                  * Implicit input:
1173                  *   ctx == skb == BPF_R6 == CTX
1174                  *
1175                  * Explicit input:
1176                  *   SRC == any register
1177                  *   IMM == 32-bit immediate
1178                  *
1179                  * Output:
1180                  *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
1181                  */
1182
1183                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
1184                 if (likely(ptr != NULL)) {
1185                         BPF_R0 = get_unaligned_be32(ptr);
1186                         CONT;
1187                 }
1188
1189                 return 0;
1190         LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
1191                 off = IMM;
1192 load_half:
1193                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
1194                 if (likely(ptr != NULL)) {
1195                         BPF_R0 = get_unaligned_be16(ptr);
1196                         CONT;
1197                 }
1198
1199                 return 0;
1200         LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
1201                 off = IMM;
1202 load_byte:
1203                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
1204                 if (likely(ptr != NULL)) {
1205                         BPF_R0 = *(u8 *)ptr;
1206                         CONT;
1207                 }
1208
1209                 return 0;
1210         LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
1211                 off = IMM + SRC;
1212                 goto load_word;
1213         LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
1214                 off = IMM + SRC;
1215                 goto load_half;
1216         LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
1217                 off = IMM + SRC;
1218                 goto load_byte;
1219
1220         default_label:
1221                 /* If we ever reach this, we have a bug somewhere. */
1222                 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
1223                 return 0;
1224 }
1225 STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
1226
1227 bool bpf_prog_array_compatible(struct bpf_array *array,
1228                                const struct bpf_prog *fp)
1229 {
1230         if (!array->owner_prog_type) {
1231                 /* There's no owner yet where we could check for
1232                  * compatibility.
1233                  */
1234                 array->owner_prog_type = fp->type;
1235                 array->owner_jited = fp->jited;
1236
1237                 return true;
1238         }
1239
1240         return array->owner_prog_type == fp->type &&
1241                array->owner_jited == fp->jited;
1242 }
1243
1244 static int bpf_check_tail_call(const struct bpf_prog *fp)
1245 {
1246         struct bpf_prog_aux *aux = fp->aux;
1247         int i;
1248
1249         for (i = 0; i < aux->used_map_cnt; i++) {
1250                 struct bpf_map *map = aux->used_maps[i];
1251                 struct bpf_array *array;
1252
1253                 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1254                         continue;
1255
1256                 array = container_of(map, struct bpf_array, map);
1257                 if (!bpf_prog_array_compatible(array, fp))
1258                         return -EINVAL;
1259         }
1260
1261         return 0;
1262 }
1263
1264 /**
1265  *      bpf_prog_select_runtime - select exec runtime for BPF program
1266  *      @fp: bpf_prog populated with internal BPF program
1267  *      @err: pointer to error variable
1268  *
1269  * Try to JIT eBPF program, if JIT is not available, use interpreter.
1270  * The BPF program will be executed via BPF_PROG_RUN() macro.
1271  */
1272 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1273 {
1274         fp->bpf_func = (void *) __bpf_prog_run;
1275
1276         /* eBPF JITs can rewrite the program in case constant
1277          * blinding is active. However, in case of error during
1278          * blinding, bpf_int_jit_compile() must always return a
1279          * valid program, which in this case would simply not
1280          * be JITed, but falls back to the interpreter.
1281          */
1282         fp = bpf_int_jit_compile(fp);
1283         bpf_prog_lock_ro(fp);
1284
1285         /* The tail call compatibility check can only be done at
1286          * this late stage as we need to determine, if we deal
1287          * with JITed or non JITed program concatenations and not
1288          * all eBPF JITs might immediately support all features.
1289          */
1290         *err = bpf_check_tail_call(fp);
1291
1292         return fp;
1293 }
1294 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1295
1296 static void bpf_prog_free_deferred(struct work_struct *work)
1297 {
1298         struct bpf_prog_aux *aux;
1299
1300         aux = container_of(work, struct bpf_prog_aux, work);
1301         bpf_jit_free(aux->prog);
1302 }
1303
1304 /* Free internal BPF program */
1305 void bpf_prog_free(struct bpf_prog *fp)
1306 {
1307         struct bpf_prog_aux *aux = fp->aux;
1308
1309         INIT_WORK(&aux->work, bpf_prog_free_deferred);
1310         schedule_work(&aux->work);
1311 }
1312 EXPORT_SYMBOL_GPL(bpf_prog_free);
1313
1314 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1315 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1316
1317 void bpf_user_rnd_init_once(void)
1318 {
1319         prandom_init_once(&bpf_user_rnd_state);
1320 }
1321
1322 BPF_CALL_0(bpf_user_rnd_u32)
1323 {
1324         /* Should someone ever have the rather unwise idea to use some
1325          * of the registers passed into this function, then note that
1326          * this function is called from native eBPF and classic-to-eBPF
1327          * transformations. Register assignments from both sides are
1328          * different, f.e. classic always sets fn(ctx, A, X) here.
1329          */
1330         struct rnd_state *state;
1331         u32 res;
1332
1333         state = &get_cpu_var(bpf_user_rnd_state);
1334         res = prandom_u32_state(state);
1335         put_cpu_var(bpf_user_rnd_state);
1336
1337         return res;
1338 }
1339
1340 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1341 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1342 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1343 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1344
1345 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1346 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1347 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1348 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1349
1350 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1351 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1352 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1353
1354 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1355 {
1356         return NULL;
1357 }
1358
1359 u64 __weak
1360 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1361                  void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1362 {
1363         return -ENOTSUPP;
1364 }
1365
1366 /* Always built-in helper functions. */
1367 const struct bpf_func_proto bpf_tail_call_proto = {
1368         .func           = NULL,
1369         .gpl_only       = false,
1370         .ret_type       = RET_VOID,
1371         .arg1_type      = ARG_PTR_TO_CTX,
1372         .arg2_type      = ARG_CONST_MAP_PTR,
1373         .arg3_type      = ARG_ANYTHING,
1374 };
1375
1376 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
1377  * It is encouraged to implement bpf_int_jit_compile() instead, so that
1378  * eBPF and implicitly also cBPF can get JITed!
1379  */
1380 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1381 {
1382         return prog;
1383 }
1384
1385 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
1386  * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
1387  */
1388 void __weak bpf_jit_compile(struct bpf_prog *prog)
1389 {
1390 }
1391
1392 bool __weak bpf_helper_changes_pkt_data(void *func)
1393 {
1394         return false;
1395 }
1396
1397 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1398  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1399  */
1400 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1401                          int len)
1402 {
1403         return -EFAULT;
1404 }
1405
1406 /* All definitions of tracepoints related to BPF. */
1407 #define CREATE_TRACE_POINTS
1408 #include <linux/bpf_trace.h>
1409
1410 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
1411
1412 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type);
1413 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu);