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
  32 #include <asm/unaligned.h>
  33
  34 /* Registers */
  35 #define BPF_R0  regs[BPF_REG_0]
  36 #define BPF_R1  regs[BPF_REG_1]
  37 #define BPF_R2  regs[BPF_REG_2]
  38 #define BPF_R3  regs[BPF_REG_3]
  39 #define BPF_R4  regs[BPF_REG_4]
  40 #define BPF_R5  regs[BPF_REG_5]
  41 #define BPF_R6  regs[BPF_REG_6]
  42 #define BPF_R7  regs[BPF_REG_7]
  43 #define BPF_R8  regs[BPF_REG_8]
  44 #define BPF_R9  regs[BPF_REG_9]
  45 #define BPF_R10 regs[BPF_REG_10]
  46
  47 /* Named registers */
  48 #define DST     regs[insn->dst_reg]
  49 #define SRC     regs[insn->src_reg]
  50 #define FP      regs[BPF_REG_FP]
  51 #define ARG1    regs[BPF_REG_ARG1]
  52 #define CTX     regs[BPF_REG_CTX]
  53 #define IMM     insn->imm
  54
  55 /* No hurry in this branch
  56  *
  57  * Exported for the bpf jit load helper.
  58  */
  59 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
  60 {
  61         u8 *ptr = NULL;
  62
  63         if (k >= SKF_NET_OFF)
  64                 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
  65         else if (k >= SKF_LL_OFF)
  66                 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
  67
  68         if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
  69                 return ptr;
  70
  71         return NULL;
  72 }
  73
  74 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
  75 {
  76         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
  77                           gfp_extra_flags;
  78         struct bpf_prog_aux *aux;
  79         struct bpf_prog *fp;
  80
  81         size = round_up(size, PAGE_SIZE);
  82         fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
  83         if (fp == NULL)
  84                 return NULL;
  85
  86         kmemcheck_annotate_bitfield(fp, meta);
  87
  88         aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
  89         if (aux == NULL) {
  90                 vfree(fp);
  91                 return NULL;
  92         }
  93
  94         fp->pages = size / PAGE_SIZE;
  95         fp->aux = aux;
  96         fp->aux->prog = fp;
  97
  98         return fp;
  99 }
 100 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
 101
 102 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
 103                                   gfp_t gfp_extra_flags)
 104 {
 105         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
 106                           gfp_extra_flags;
 107         struct bpf_prog *fp;
 108         u32 pages, delta;
 109         int ret;
 110
 111         BUG_ON(fp_old == NULL);
 112
 113         size = round_up(size, PAGE_SIZE);
 114         pages = size / PAGE_SIZE;
 115         if (pages <= fp_old->pages)
 116                 return fp_old;
 117
 118         delta = pages - fp_old->pages;
 119         ret = __bpf_prog_charge(fp_old->aux->user, delta);
 120         if (ret)
 121                 return NULL;
 122
 123         fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
 124         if (fp == NULL) {
 125                 __bpf_prog_uncharge(fp_old->aux->user, delta);
 126         } else {
 127                 kmemcheck_annotate_bitfield(fp, meta);
 128
 129                 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
 130                 fp->pages = pages;
 131                 fp->aux->prog = fp;
 132
 133                 /* We keep fp->aux from fp_old around in the new
 134                  * reallocated structure.
 135                  */
 136                 fp_old->aux = NULL;
 137                 __bpf_prog_free(fp_old);
 138         }
 139
 140         return fp;
 141 }
 142
 143 void __bpf_prog_free(struct bpf_prog *fp)
 144 {
 145         kfree(fp->aux);
 146         vfree(fp);
 147 }
 148
 149 int bpf_prog_calc_tag(struct bpf_prog *fp)
 150 {
 151         const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
 152         u32 raw_size = bpf_prog_tag_scratch_size(fp);
 153         u32 digest[SHA_DIGEST_WORDS];
 154         u32 ws[SHA_WORKSPACE_WORDS];
 155         u32 i, bsize, psize, blocks;
 156         struct bpf_insn *dst;
 157         bool was_ld_map;
 158         u8 *raw, *todo;
 159         __be32 *result;
 160         __be64 *bits;
 161
 162         raw = vmalloc(raw_size);
 163         if (!raw)
 164                 return -ENOMEM;
 165
 166         sha_init(digest);
 167         memset(ws, 0, sizeof(ws));
 168
 169         /* We need to take out the map fd for the digest calculation
 170          * since they are unstable from user space side.
 171          */
 172         dst = (void *)raw;
 173         for (i = 0, was_ld_map = false; i < fp->len; i++) {
 174                 dst[i] = fp->insnsi[i];
 175                 if (!was_ld_map &&
 176                     dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
 177                     dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
 178                         was_ld_map = true;
 179                         dst[i].imm = 0;
 180                 } else if (was_ld_map &&
 181                            dst[i].code == 0 &&
 182                            dst[i].dst_reg == 0 &&
 183                            dst[i].src_reg == 0 &&
 184                            dst[i].off == 0) {
 185                         was_ld_map = false;
 186                         dst[i].imm = 0;
 187                 } else {
 188                         was_ld_map = false;
 189                 }
 190         }
 191
 192         psize = bpf_prog_insn_size(fp);
 193         memset(&raw[psize], 0, raw_size - psize);
 194         raw[psize++] = 0x80;
 195
 196         bsize  = round_up(psize, SHA_MESSAGE_BYTES);
 197         blocks = bsize / SHA_MESSAGE_BYTES;
 198         todo   = raw;
 199         if (bsize - psize >= sizeof(__be64)) {
 200                 bits = (__be64 *)(todo + bsize - sizeof(__be64));
 201         } else {
 202                 bits = (__be64 *)(todo + bsize + bits_offset);
 203                 blocks++;
 204         }
 205         *bits = cpu_to_be64((psize - 1) << 3);
 206
 207         while (blocks--) {
 208                 sha_transform(digest, todo, ws);
 209                 todo += SHA_MESSAGE_BYTES;
 210         }
 211
 212         result = (__force __be32 *)digest;
 213         for (i = 0; i < SHA_DIGEST_WORDS; i++)
 214                 result[i] = cpu_to_be32(digest[i]);
 215         memcpy(fp->tag, result, sizeof(fp->tag));
 216
 217         vfree(raw);
 218         return 0;
 219 }
 220
 221 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
 222 {
 223         return BPF_CLASS(insn->code) == BPF_JMP  &&
 224                /* Call and Exit are both special jumps with no
 225                 * target inside the BPF instruction image.
 226                 */
 227                BPF_OP(insn->code) != BPF_CALL &&
 228                BPF_OP(insn->code) != BPF_EXIT;
 229 }
 230
 231 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
 232 {
 233         struct bpf_insn *insn = prog->insnsi;
 234         u32 i, insn_cnt = prog->len;
 235
 236         for (i = 0; i < insn_cnt; i++, insn++) {
 237                 if (!bpf_is_jmp_and_has_target(insn))
 238                         continue;
 239
 240                 /* Adjust offset of jmps if we cross boundaries. */
 241                 if (i < pos && i + insn->off + 1 > pos)
 242                         insn->off += delta;
 243                 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
 244                         insn->off -= delta;
 245         }
 246 }
 247
 248 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
 249                                        const struct bpf_insn *patch, u32 len)
 250 {
 251         u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
 252         struct bpf_prog *prog_adj;
 253
 254         /* Since our patchlet doesn't expand the image, we're done. */
 255         if (insn_delta == 0) {
 256                 memcpy(prog->insnsi + off, patch, sizeof(*patch));
 257                 return prog;
 258         }
 259
 260         insn_adj_cnt = prog->len + insn_delta;
 261
 262         /* Several new instructions need to be inserted. Make room
 263          * for them. Likely, there's no need for a new allocation as
 264          * last page could have large enough tailroom.
 265          */
 266         prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
 267                                     GFP_USER);
 268         if (!prog_adj)
 269                 return NULL;
 270
 271         prog_adj->len = insn_adj_cnt;
 272
 273         /* Patching happens in 3 steps:
 274          *
 275          * 1) Move over tail of insnsi from next instruction onwards,
 276          *    so we can patch the single target insn with one or more
 277          *    new ones (patching is always from 1 to n insns, n > 0).
 278          * 2) Inject new instructions at the target location.
 279          * 3) Adjust branch offsets if necessary.
 280          */
 281         insn_rest = insn_adj_cnt - off - len;
 282
 283         memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
 284                 sizeof(*patch) * insn_rest);
 285         memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
 286
 287         bpf_adj_branches(prog_adj, off, insn_delta);
 288
 289         return prog_adj;
 290 }
 291
 292 #ifdef CONFIG_BPF_JIT
 293 struct bpf_binary_header *
 294 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
 295                      unsigned int alignment,
 296                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
 297 {
 298         struct bpf_binary_header *hdr;
 299         unsigned int size, hole, start;
 300
 301         /* Most of BPF filters are really small, but if some of them
 302          * fill a page, allow at least 128 extra bytes to insert a
 303          * random section of illegal instructions.
 304          */
 305         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
 306         hdr = module_alloc(size);
 307         if (hdr == NULL)
 308                 return NULL;
 309
 310         /* Fill space with illegal/arch-dep instructions. */
 311         bpf_fill_ill_insns(hdr, size);
 312
 313         hdr->pages = size / PAGE_SIZE;
 314         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
 315                      PAGE_SIZE - sizeof(*hdr));
 316         start = (get_random_int() % hole) & ~(alignment - 1);
 317
 318         /* Leave a random number of instructions before BPF code. */
 319         *image_ptr = &hdr->image[start];
 320
 321         return hdr;
 322 }
 323
 324 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
 325 {
 326         module_memfree(hdr);
 327 }
 328
 329 int bpf_jit_harden __read_mostly;
 330
 331 static int bpf_jit_blind_insn(const struct bpf_insn *from,
 332                               const struct bpf_insn *aux,
 333                               struct bpf_insn *to_buff)
 334 {
 335         struct bpf_insn *to = to_buff;
 336         u32 imm_rnd = get_random_int();
 337         s16 off;
 338
 339         BUILD_BUG_ON(BPF_REG_AX  + 1 != MAX_BPF_JIT_REG);
 340         BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
 341
 342         if (from->imm == 0 &&
 343             (from->code == (BPF_ALU   | BPF_MOV | BPF_K) ||
 344              from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
 345                 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
 346                 goto out;
 347         }
 348
 349         switch (from->code) {
 350         case BPF_ALU | BPF_ADD | BPF_K:
 351         case BPF_ALU | BPF_SUB | BPF_K:
 352         case BPF_ALU | BPF_AND | BPF_K:
 353         case BPF_ALU | BPF_OR  | BPF_K:
 354         case BPF_ALU | BPF_XOR | BPF_K:
 355         case BPF_ALU | BPF_MUL | BPF_K:
 356         case BPF_ALU | BPF_MOV | BPF_K:
 357         case BPF_ALU | BPF_DIV | BPF_K:
 358         case BPF_ALU | BPF_MOD | BPF_K:
 359                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 360                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 361                 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
 362                 break;
 363
 364         case BPF_ALU64 | BPF_ADD | BPF_K:
 365         case BPF_ALU64 | BPF_SUB | BPF_K:
 366         case BPF_ALU64 | BPF_AND | BPF_K:
 367         case BPF_ALU64 | BPF_OR  | BPF_K:
 368         case BPF_ALU64 | BPF_XOR | BPF_K:
 369         case BPF_ALU64 | BPF_MUL | BPF_K:
 370         case BPF_ALU64 | BPF_MOV | BPF_K:
 371         case BPF_ALU64 | BPF_DIV | BPF_K:
 372         case BPF_ALU64 | BPF_MOD | BPF_K:
 373                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 374                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 375                 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
 376                 break;
 377
 378         case BPF_JMP | BPF_JEQ  | BPF_K:
 379         case BPF_JMP | BPF_JNE  | BPF_K:
 380         case BPF_JMP | BPF_JGT  | BPF_K:
 381         case BPF_JMP | BPF_JGE  | BPF_K:
 382         case BPF_JMP | BPF_JSGT | BPF_K:
 383         case BPF_JMP | BPF_JSGE | BPF_K:
 384         case BPF_JMP | BPF_JSET | BPF_K:
 385                 /* Accommodate for extra offset in case of a backjump. */
 386                 off = from->off;
 387                 if (off < 0)
 388                         off -= 2;
 389                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 390                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 391                 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
 392                 break;
 393
 394         case BPF_LD | BPF_ABS | BPF_W:
 395         case BPF_LD | BPF_ABS | BPF_H:
 396         case BPF_LD | BPF_ABS | BPF_B:
 397                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 398                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 399                 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
 400                 break;
 401
 402         case BPF_LD | BPF_IND | BPF_W:
 403         case BPF_LD | BPF_IND | BPF_H:
 404         case BPF_LD | BPF_IND | BPF_B:
 405                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 406                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 407                 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
 408                 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
 409                 break;
 410
 411         case BPF_LD | BPF_IMM | BPF_DW:
 412                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
 413                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 414                 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
 415                 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
 416                 break;
 417         case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
 418                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
 419                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 420                 *to++ = BPF_ALU64_REG(BPF_OR,  aux[0].dst_reg, BPF_REG_AX);
 421                 break;
 422
 423         case BPF_ST | BPF_MEM | BPF_DW:
 424         case BPF_ST | BPF_MEM | BPF_W:
 425         case BPF_ST | BPF_MEM | BPF_H:
 426         case BPF_ST | BPF_MEM | BPF_B:
 427                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
 428                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
 429                 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
 430                 break;
 431         }
 432 out:
 433         return to - to_buff;
 434 }
 435
 436 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
 437                                               gfp_t gfp_extra_flags)
 438 {
 439         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
 440                           gfp_extra_flags;
 441         struct bpf_prog *fp;
 442
 443         fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
 444         if (fp != NULL) {
 445                 kmemcheck_annotate_bitfield(fp, meta);
 446
 447                 /* aux->prog still points to the fp_other one, so
 448                  * when promoting the clone to the real program,
 449                  * this still needs to be adapted.
 450                  */
 451                 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
 452         }
 453
 454         return fp;
 455 }
 456
 457 static void bpf_prog_clone_free(struct bpf_prog *fp)
 458 {
 459         /* aux was stolen by the other clone, so we cannot free
 460          * it from this path! It will be freed eventually by the
 461          * other program on release.
 462          *
 463          * At this point, we don't need a deferred release since
 464          * clone is guaranteed to not be locked.
 465          */
 466         fp->aux = NULL;
 467         __bpf_prog_free(fp);
 468 }
 469
 470 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
 471 {
 472         /* We have to repoint aux->prog to self, as we don't
 473          * know whether fp here is the clone or the original.
 474          */
 475         fp->aux->prog = fp;
 476         bpf_prog_clone_free(fp_other);
 477 }
 478
 479 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
 480 {
 481         struct bpf_insn insn_buff[16], aux[2];
 482         struct bpf_prog *clone, *tmp;
 483         int insn_delta, insn_cnt;
 484         struct bpf_insn *insn;
 485         int i, rewritten;
 486
 487         if (!bpf_jit_blinding_enabled())
 488                 return prog;
 489
 490         clone = bpf_prog_clone_create(prog, GFP_USER);
 491         if (!clone)
 492                 return ERR_PTR(-ENOMEM);
 493
 494         insn_cnt = clone->len;
 495         insn = clone->insnsi;
 496
 497         for (i = 0; i < insn_cnt; i++, insn++) {
 498                 /* We temporarily need to hold the original ld64 insn
 499                  * so that we can still access the first part in the
 500                  * second blinding run.
 501                  */
 502                 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
 503                     insn[1].code == 0)
 504                         memcpy(aux, insn, sizeof(aux));
 505
 506                 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
 507                 if (!rewritten)
 508                         continue;
 509
 510                 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
 511                 if (!tmp) {
 512                         /* Patching may have repointed aux->prog during
 513                          * realloc from the original one, so we need to
 514                          * fix it up here on error.
 515                          */
 516                         bpf_jit_prog_release_other(prog, clone);
 517                         return ERR_PTR(-ENOMEM);
 518                 }
 519
 520                 clone = tmp;
 521                 insn_delta = rewritten - 1;
 522
 523                 /* Walk new program and skip insns we just inserted. */
 524                 insn = clone->insnsi + i + insn_delta;
 525                 insn_cnt += insn_delta;
 526                 i        += insn_delta;
 527         }
 528
 529         return clone;
 530 }
 531 #endif /* CONFIG_BPF_JIT */
 532
 533 /* Base function for offset calculation. Needs to go into .text section,
 534  * therefore keeping it non-static as well; will also be used by JITs
 535  * anyway later on, so do not let the compiler omit it.
 536  */
 537 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 538 {
 539         return 0;
 540 }
 541 EXPORT_SYMBOL_GPL(__bpf_call_base);
 542
 543 /**
 544  *      __bpf_prog_run - run eBPF program on a given context
 545  *      @ctx: is the data we are operating on
 546  *      @insn: is the array of eBPF instructions
 547  *
 548  * Decode and execute eBPF instructions.
 549  */
 550 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
 551 {
 552         u64 stack[MAX_BPF_STACK / sizeof(u64)];
 553         u64 regs[MAX_BPF_REG], tmp;
 554         static const void *jumptable[256] = {
 555                 [0 ... 255] = &&default_label,
 556                 /* Now overwrite non-defaults ... */
 557                 /* 32 bit ALU operations */
 558                 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
 559                 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
 560                 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
 561                 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
 562                 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
 563                 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
 564                 [BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
 565                 [BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
 566                 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
 567                 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
 568                 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
 569                 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
 570                 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
 571                 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
 572                 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
 573                 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
 574                 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
 575                 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
 576                 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
 577                 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
 578                 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
 579                 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
 580                 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
 581                 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
 582                 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
 583                 /* 64 bit ALU operations */
 584                 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
 585                 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
 586                 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
 587                 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
 588                 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
 589                 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
 590                 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
 591                 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
 592                 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
 593                 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
 594                 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
 595                 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
 596                 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
 597                 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
 598                 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
 599                 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
 600                 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
 601                 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
 602                 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
 603                 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
 604                 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
 605                 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
 606                 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
 607                 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
 608                 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
 609                 /* Call instruction */
 610                 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
 611                 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
 612                 /* Jumps */
 613                 [BPF_JMP | BPF_JA] = &&JMP_JA,
 614                 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
 615                 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
 616                 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
 617                 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
 618                 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
 619                 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
 620                 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
 621                 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
 622                 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
 623                 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
 624                 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
 625                 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
 626                 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
 627                 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
 628                 /* Program return */
 629                 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
 630                 /* Store instructions */
 631                 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
 632                 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
 633                 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
 634                 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
 635                 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
 636                 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
 637                 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
 638                 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
 639                 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
 640                 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
 641                 /* Load instructions */
 642                 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
 643                 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
 644                 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
 645                 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
 646                 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
 647                 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
 648                 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
 649                 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
 650                 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
 651                 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
 652                 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
 653         };
 654         u32 tail_call_cnt = 0;
 655         void *ptr;
 656         int off;
 657
 658 #define CONT     ({ insn++; goto select_insn; })
 659 #define CONT_JMP ({ insn++; goto select_insn; })
 660
 661         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
 662         ARG1 = (u64) (unsigned long) ctx;
 663
 664 select_insn:
 665         goto *jumptable[insn->code];
 666
 667         /* ALU */
 668 #define ALU(OPCODE, OP)                 \
 669         ALU64_##OPCODE##_X:             \
 670                 DST = DST OP SRC;       \
 671                 CONT;                   \
 672         ALU_##OPCODE##_X:               \
 673                 DST = (u32) DST OP (u32) SRC;   \
 674                 CONT;                   \
 675         ALU64_##OPCODE##_K:             \
 676                 DST = DST OP IMM;               \
 677                 CONT;                   \
 678         ALU_##OPCODE##_K:               \
 679                 DST = (u32) DST OP (u32) IMM;   \
 680                 CONT;
 681
 682         ALU(ADD,  +)
 683         ALU(SUB,  -)
 684         ALU(AND,  &)
 685         ALU(OR,   |)
 686         ALU(LSH, <<)
 687         ALU(RSH, >>)
 688         ALU(XOR,  ^)
 689         ALU(MUL,  *)
 690 #undef ALU
 691         ALU_NEG:
 692                 DST = (u32) -DST;
 693                 CONT;
 694         ALU64_NEG:
 695                 DST = -DST;
 696                 CONT;
 697         ALU_MOV_X:
 698                 DST = (u32) SRC;
 699                 CONT;
 700         ALU_MOV_K:
 701                 DST = (u32) IMM;
 702                 CONT;
 703         ALU64_MOV_X:
 704                 DST = SRC;
 705                 CONT;
 706         ALU64_MOV_K:
 707                 DST = IMM;
 708                 CONT;
 709         LD_IMM_DW:
 710                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
 711                 insn++;
 712                 CONT;
 713         ALU64_ARSH_X:
 714                 (*(s64 *) &DST) >>= SRC;
 715                 CONT;
 716         ALU64_ARSH_K:
 717                 (*(s64 *) &DST) >>= IMM;
 718                 CONT;
 719         ALU64_MOD_X:
 720                 if (unlikely(SRC == 0))
 721                         return 0;
 722                 div64_u64_rem(DST, SRC, &tmp);
 723                 DST = tmp;
 724                 CONT;
 725         ALU_MOD_X:
 726                 if (unlikely(SRC == 0))
 727                         return 0;
 728                 tmp = (u32) DST;
 729                 DST = do_div(tmp, (u32) SRC);
 730                 CONT;
 731         ALU64_MOD_K:
 732                 div64_u64_rem(DST, IMM, &tmp);
 733                 DST = tmp;
 734                 CONT;
 735         ALU_MOD_K:
 736                 tmp = (u32) DST;
 737                 DST = do_div(tmp, (u32) IMM);
 738                 CONT;
 739         ALU64_DIV_X:
 740                 if (unlikely(SRC == 0))
 741                         return 0;
 742                 DST = div64_u64(DST, SRC);
 743                 CONT;
 744         ALU_DIV_X:
 745                 if (unlikely(SRC == 0))
 746                         return 0;
 747                 tmp = (u32) DST;
 748                 do_div(tmp, (u32) SRC);
 749                 DST = (u32) tmp;
 750                 CONT;
 751         ALU64_DIV_K:
 752                 DST = div64_u64(DST, IMM);
 753                 CONT;
 754         ALU_DIV_K:
 755                 tmp = (u32) DST;
 756                 do_div(tmp, (u32) IMM);
 757                 DST = (u32) tmp;
 758                 CONT;
 759         ALU_END_TO_BE:
 760                 switch (IMM) {
 761                 case 16:
 762                         DST = (__force u16) cpu_to_be16(DST);
 763                         break;
 764                 case 32:
 765                         DST = (__force u32) cpu_to_be32(DST);
 766                         break;
 767                 case 64:
 768                         DST = (__force u64) cpu_to_be64(DST);
 769                         break;
 770                 }
 771                 CONT;
 772         ALU_END_TO_LE:
 773                 switch (IMM) {
 774                 case 16:
 775                         DST = (__force u16) cpu_to_le16(DST);
 776                         break;
 777                 case 32:
 778                         DST = (__force u32) cpu_to_le32(DST);
 779                         break;
 780                 case 64:
 781                         DST = (__force u64) cpu_to_le64(DST);
 782                         break;
 783                 }
 784                 CONT;
 785
 786         /* CALL */
 787         JMP_CALL:
 788                 /* Function call scratches BPF_R1-BPF_R5 registers,
 789                  * preserves BPF_R6-BPF_R9, and stores return value
 790                  * into BPF_R0.
 791                  */
 792                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
 793                                                        BPF_R4, BPF_R5);
 794                 CONT;
 795
 796         JMP_TAIL_CALL: {
 797                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
 798                 struct bpf_array *array = container_of(map, struct bpf_array, map);
 799                 struct bpf_prog *prog;
 800                 u64 index = BPF_R3;
 801
 802                 if (unlikely(index >= array->map.max_entries))
 803                         goto out;
 804                 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
 805                         goto out;
 806
 807                 tail_call_cnt++;
 808
 809                 prog = READ_ONCE(array->ptrs[index]);
 810                 if (!prog)
 811                         goto out;
 812
 813                 /* ARG1 at this point is guaranteed to point to CTX from
 814                  * the verifier side due to the fact that the tail call is
 815                  * handeled like a helper, that is, bpf_tail_call_proto,
 816                  * where arg1_type is ARG_PTR_TO_CTX.
 817                  */
 818                 insn = prog->insnsi;
 819                 goto select_insn;
 820 out:
 821                 CONT;
 822         }
 823         /* JMP */
 824         JMP_JA:
 825                 insn += insn->off;
 826                 CONT;
 827         JMP_JEQ_X:
 828                 if (DST == SRC) {
 829                         insn += insn->off;
 830                         CONT_JMP;
 831                 }
 832                 CONT;
 833         JMP_JEQ_K:
 834                 if (DST == IMM) {
 835                         insn += insn->off;
 836                         CONT_JMP;
 837                 }
 838                 CONT;
 839         JMP_JNE_X:
 840                 if (DST != SRC) {
 841                         insn += insn->off;
 842                         CONT_JMP;
 843                 }
 844                 CONT;
 845         JMP_JNE_K:
 846                 if (DST != IMM) {
 847                         insn += insn->off;
 848                         CONT_JMP;
 849                 }
 850                 CONT;
 851         JMP_JGT_X:
 852                 if (DST > SRC) {
 853                         insn += insn->off;
 854                         CONT_JMP;
 855                 }
 856                 CONT;
 857         JMP_JGT_K:
 858                 if (DST > IMM) {
 859                         insn += insn->off;
 860                         CONT_JMP;
 861                 }
 862                 CONT;
 863         JMP_JGE_X:
 864                 if (DST >= SRC) {
 865                         insn += insn->off;
 866                         CONT_JMP;
 867                 }
 868                 CONT;
 869         JMP_JGE_K:
 870                 if (DST >= IMM) {
 871                         insn += insn->off;
 872                         CONT_JMP;
 873                 }
 874                 CONT;
 875         JMP_JSGT_X:
 876                 if (((s64) DST) > ((s64) SRC)) {
 877                         insn += insn->off;
 878                         CONT_JMP;
 879                 }
 880                 CONT;
 881         JMP_JSGT_K:
 882                 if (((s64) DST) > ((s64) IMM)) {
 883                         insn += insn->off;
 884                         CONT_JMP;
 885                 }
 886                 CONT;
 887         JMP_JSGE_X:
 888                 if (((s64) DST) >= ((s64) SRC)) {
 889                         insn += insn->off;
 890                         CONT_JMP;
 891                 }
 892                 CONT;
 893         JMP_JSGE_K:
 894                 if (((s64) DST) >= ((s64) IMM)) {
 895                         insn += insn->off;
 896                         CONT_JMP;
 897                 }
 898                 CONT;
 899         JMP_JSET_X:
 900                 if (DST & SRC) {
 901                         insn += insn->off;
 902                         CONT_JMP;
 903                 }
 904                 CONT;
 905         JMP_JSET_K:
 906                 if (DST & IMM) {
 907                         insn += insn->off;
 908                         CONT_JMP;
 909                 }
 910                 CONT;
 911         JMP_EXIT:
 912                 return BPF_R0;
 913
 914         /* STX and ST and LDX*/
 915 #define LDST(SIZEOP, SIZE)                                              \
 916         STX_MEM_##SIZEOP:                                               \
 917                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
 918                 CONT;                                                   \
 919         ST_MEM_##SIZEOP:                                                \
 920                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
 921                 CONT;                                                   \
 922         LDX_MEM_##SIZEOP:                                               \
 923                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
 924                 CONT;
 925
 926         LDST(B,   u8)
 927         LDST(H,  u16)
 928         LDST(W,  u32)
 929         LDST(DW, u64)
 930 #undef LDST
 931         STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
 932                 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
 933                            (DST + insn->off));
 934                 CONT;
 935         STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
 936                 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
 937                              (DST + insn->off));
 938                 CONT;
 939         LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
 940                 off = IMM;
 941 load_word:
 942                 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
 943                  * only appearing in the programs where ctx ==
 944                  * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
 945                  * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
 946                  * internal BPF verifier will check that BPF_R6 ==
 947                  * ctx.
 948                  *
 949                  * BPF_ABS and BPF_IND are wrappers of function calls,
 950                  * so they scratch BPF_R1-BPF_R5 registers, preserve
 951                  * BPF_R6-BPF_R9, and store return value into BPF_R0.
 952                  *
 953                  * Implicit input:
 954                  *   ctx == skb == BPF_R6 == CTX
 955                  *
 956                  * Explicit input:
 957                  *   SRC == any register
 958                  *   IMM == 32-bit immediate
 959                  *
 960                  * Output:
 961                  *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
 962                  */
 963
 964                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
 965                 if (likely(ptr != NULL)) {
 966                         BPF_R0 = get_unaligned_be32(ptr);
 967                         CONT;
 968                 }
 969
 970                 return 0;
 971         LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
 972                 off = IMM;
 973 load_half:
 974                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
 975                 if (likely(ptr != NULL)) {
 976                         BPF_R0 = get_unaligned_be16(ptr);
 977                         CONT;
 978                 }
 979
 980                 return 0;
 981         LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
 982                 off = IMM;
 983 load_byte:
 984                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
 985                 if (likely(ptr != NULL)) {
 986                         BPF_R0 = *(u8 *)ptr;
 987                         CONT;
 988                 }
 989
 990                 return 0;
 991         LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
 992                 off = IMM + SRC;
 993                 goto load_word;
 994         LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
 995                 off = IMM + SRC;
 996                 goto load_half;
 997         LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
 998                 off = IMM + SRC;
 999                 goto load_byte;
1000
1001         default_label:
1002                 /* If we ever reach this, we have a bug somewhere. */
1003                 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
1004                 return 0;
1005 }
1006 STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
1007
1008 bool bpf_prog_array_compatible(struct bpf_array *array,
1009                                const struct bpf_prog *fp)
1010 {
1011         if (!array->owner_prog_type) {
1012                 /* There's no owner yet where we could check for
1013                  * compatibility.
1014                  */
1015                 array->owner_prog_type = fp->type;
1016                 array->owner_jited = fp->jited;
1017
1018                 return true;
1019         }
1020
1021         return array->owner_prog_type == fp->type &&
1022                array->owner_jited == fp->jited;
1023 }
1024
1025 static int bpf_check_tail_call(const struct bpf_prog *fp)
1026 {
1027         struct bpf_prog_aux *aux = fp->aux;
1028         int i;
1029
1030         for (i = 0; i < aux->used_map_cnt; i++) {
1031                 struct bpf_map *map = aux->used_maps[i];
1032                 struct bpf_array *array;
1033
1034                 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1035                         continue;
1036
1037                 array = container_of(map, struct bpf_array, map);
1038                 if (!bpf_prog_array_compatible(array, fp))
1039                         return -EINVAL;
1040         }
1041
1042         return 0;
1043 }
1044
1045 /**
1046  *      bpf_prog_select_runtime - select exec runtime for BPF program
1047  *      @fp: bpf_prog populated with internal BPF program
1048  *      @err: pointer to error variable
1049  *
1050  * Try to JIT eBPF program, if JIT is not available, use interpreter.
1051  * The BPF program will be executed via BPF_PROG_RUN() macro.
1052  */
1053 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1054 {
1055         fp->bpf_func = (void *) __bpf_prog_run;
1056
1057         /* eBPF JITs can rewrite the program in case constant
1058          * blinding is active. However, in case of error during
1059          * blinding, bpf_int_jit_compile() must always return a
1060          * valid program, which in this case would simply not
1061          * be JITed, but falls back to the interpreter.
1062          */
1063         fp = bpf_int_jit_compile(fp);
1064         bpf_prog_lock_ro(fp);
1065
1066         /* The tail call compatibility check can only be done at
1067          * this late stage as we need to determine, if we deal
1068          * with JITed or non JITed program concatenations and not
1069          * all eBPF JITs might immediately support all features.
1070          */
1071         *err = bpf_check_tail_call(fp);
1072
1073         return fp;
1074 }
1075 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1076
1077 static void bpf_prog_free_deferred(struct work_struct *work)
1078 {
1079         struct bpf_prog_aux *aux;
1080
1081         aux = container_of(work, struct bpf_prog_aux, work);
1082         bpf_jit_free(aux->prog);
1083 }
1084
1085 /* Free internal BPF program */
1086 void bpf_prog_free(struct bpf_prog *fp)
1087 {
1088         struct bpf_prog_aux *aux = fp->aux;
1089
1090         INIT_WORK(&aux->work, bpf_prog_free_deferred);
1091         schedule_work(&aux->work);
1092 }
1093 EXPORT_SYMBOL_GPL(bpf_prog_free);
1094
1095 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1096 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1097
1098 void bpf_user_rnd_init_once(void)
1099 {
1100         prandom_init_once(&bpf_user_rnd_state);
1101 }
1102
1103 BPF_CALL_0(bpf_user_rnd_u32)
1104 {
1105         /* Should someone ever have the rather unwise idea to use some
1106          * of the registers passed into this function, then note that
1107          * this function is called from native eBPF and classic-to-eBPF
1108          * transformations. Register assignments from both sides are
1109          * different, f.e. classic always sets fn(ctx, A, X) here.
1110          */
1111         struct rnd_state *state;
1112         u32 res;
1113
1114         state = &get_cpu_var(bpf_user_rnd_state);
1115         res = prandom_u32_state(state);
1116         put_cpu_var(bpf_user_rnd_state);
1117
1118         return res;
1119 }
1120
1121 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1122 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1123 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1124 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1125
1126 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1127 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1128 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1129 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1130
1131 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1132 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1133 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1134
1135 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1136 {
1137         return NULL;
1138 }
1139
1140 u64 __weak
1141 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1142                  void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1143 {
1144         return -ENOTSUPP;
1145 }
1146
1147 /* Always built-in helper functions. */
1148 const struct bpf_func_proto bpf_tail_call_proto = {
1149         .func           = NULL,
1150         .gpl_only       = false,
1151         .ret_type       = RET_VOID,
1152         .arg1_type      = ARG_PTR_TO_CTX,
1153         .arg2_type      = ARG_CONST_MAP_PTR,
1154         .arg3_type      = ARG_ANYTHING,
1155 };
1156
1157 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
1158 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1159 {
1160         return prog;
1161 }
1162
1163 bool __weak bpf_helper_changes_pkt_data(void *func)
1164 {
1165         return false;
1166 }
1167
1168 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1169  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1170  */
1171 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1172                          int len)
1173 {
1174         return -EFAULT;
1175 }