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
  31 #include <asm/unaligned.h>
  32
  33 /* Registers */
  34 #define BPF_R0  regs[BPF_REG_0]
  35 #define BPF_R1  regs[BPF_REG_1]
  36 #define BPF_R2  regs[BPF_REG_2]
  37 #define BPF_R3  regs[BPF_REG_3]
  38 #define BPF_R4  regs[BPF_REG_4]
  39 #define BPF_R5  regs[BPF_REG_5]
  40 #define BPF_R6  regs[BPF_REG_6]
  41 #define BPF_R7  regs[BPF_REG_7]
  42 #define BPF_R8  regs[BPF_REG_8]
  43 #define BPF_R9  regs[BPF_REG_9]
  44 #define BPF_R10 regs[BPF_REG_10]
  45
  46 /* Named registers */
  47 #define DST     regs[insn->dst_reg]
  48 #define SRC     regs[insn->src_reg]
  49 #define FP      regs[BPF_REG_FP]
  50 #define ARG1    regs[BPF_REG_ARG1]
  51 #define CTX     regs[BPF_REG_CTX]
  52 #define IMM     insn->imm
  53
  54 /* No hurry in this branch
  55  *
  56  * Exported for the bpf jit load helper.
  57  */
  58 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
  59 {
  60         u8 *ptr = NULL;
  61
  62         if (k >= SKF_NET_OFF)
  63                 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
  64         else if (k >= SKF_LL_OFF)
  65                 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
  66
  67         if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
  68                 return ptr;
  69
  70         return NULL;
  71 }
  72
  73 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
  74 {
  75         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
  76                           gfp_extra_flags;
  77         struct bpf_prog_aux *aux;
  78         struct bpf_prog *fp;
  79
  80         size = round_up(size, PAGE_SIZE);
  81         fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
  82         if (fp == NULL)
  83                 return NULL;
  84
  85         aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
  86         if (aux == NULL) {
  87                 vfree(fp);
  88                 return NULL;
  89         }
  90
  91         fp->pages = size / PAGE_SIZE;
  92         fp->aux = aux;
  93
  94         return fp;
  95 }
  96 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
  97
  98 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
  99                                   gfp_t gfp_extra_flags)
 100 {
 101         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
 102                           gfp_extra_flags;
 103         struct bpf_prog *fp;
 104
 105         BUG_ON(fp_old == NULL);
 106
 107         size = round_up(size, PAGE_SIZE);
 108         if (size <= fp_old->pages * PAGE_SIZE)
 109                 return fp_old;
 110
 111         fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
 112         if (fp != NULL) {
 113                 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
 114                 fp->pages = size / PAGE_SIZE;
 115
 116                 /* We keep fp->aux from fp_old around in the new
 117                  * reallocated structure.
 118                  */
 119                 fp_old->aux = NULL;
 120                 __bpf_prog_free(fp_old);
 121         }
 122
 123         return fp;
 124 }
 125 EXPORT_SYMBOL_GPL(bpf_prog_realloc);
 126
 127 void __bpf_prog_free(struct bpf_prog *fp)
 128 {
 129         kfree(fp->aux);
 130         vfree(fp);
 131 }
 132 EXPORT_SYMBOL_GPL(__bpf_prog_free);
 133
 134 #ifdef CONFIG_BPF_JIT
 135 struct bpf_binary_header *
 136 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
 137                      unsigned int alignment,
 138                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
 139 {
 140         struct bpf_binary_header *hdr;
 141         unsigned int size, hole, start;
 142
 143         /* Most of BPF filters are really small, but if some of them
 144          * fill a page, allow at least 128 extra bytes to insert a
 145          * random section of illegal instructions.
 146          */
 147         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
 148         hdr = module_alloc(size);
 149         if (hdr == NULL)
 150                 return NULL;
 151
 152         /* Fill space with illegal/arch-dep instructions. */
 153         bpf_fill_ill_insns(hdr, size);
 154
 155         hdr->pages = size / PAGE_SIZE;
 156         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
 157                      PAGE_SIZE - sizeof(*hdr));
 158         start = (prandom_u32() % hole) & ~(alignment - 1);
 159
 160         /* Leave a random number of instructions before BPF code. */
 161         *image_ptr = &hdr->image[start];
 162
 163         return hdr;
 164 }
 165
 166 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
 167 {
 168         module_memfree(hdr);
 169 }
 170 #endif /* CONFIG_BPF_JIT */
 171
 172 /* Base function for offset calculation. Needs to go into .text section,
 173  * therefore keeping it non-static as well; will also be used by JITs
 174  * anyway later on, so do not let the compiler omit it.
 175  */
 176 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 177 {
 178         return 0;
 179 }
 180 EXPORT_SYMBOL_GPL(__bpf_call_base);
 181
 182 /**
 183  *      __bpf_prog_run - run eBPF program on a given context
 184  *      @ctx: is the data we are operating on
 185  *      @insn: is the array of eBPF instructions
 186  *
 187  * Decode and execute eBPF instructions.
 188  */
 189 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
 190 {
 191         u64 stack[MAX_BPF_STACK / sizeof(u64)];
 192         u64 regs[MAX_BPF_REG], tmp;
 193         static const void *jumptable[256] = {
 194                 [0 ... 255] = &&default_label,
 195                 /* Now overwrite non-defaults ... */
 196                 /* 32 bit ALU operations */
 197                 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
 198                 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
 199                 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
 200                 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
 201                 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
 202                 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
 203                 [BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
 204                 [BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
 205                 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
 206                 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
 207                 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
 208                 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
 209                 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
 210                 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
 211                 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
 212                 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
 213                 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
 214                 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
 215                 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
 216                 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
 217                 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
 218                 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
 219                 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
 220                 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
 221                 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
 222                 /* 64 bit ALU operations */
 223                 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
 224                 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
 225                 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
 226                 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
 227                 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
 228                 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
 229                 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
 230                 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
 231                 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
 232                 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
 233                 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
 234                 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
 235                 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
 236                 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
 237                 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
 238                 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
 239                 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
 240                 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
 241                 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
 242                 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
 243                 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
 244                 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
 245                 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
 246                 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
 247                 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
 248                 /* Call instruction */
 249                 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
 250                 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
 251                 /* Jumps */
 252                 [BPF_JMP | BPF_JA] = &&JMP_JA,
 253                 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
 254                 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
 255                 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
 256                 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
 257                 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
 258                 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
 259                 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
 260                 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
 261                 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
 262                 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
 263                 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
 264                 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
 265                 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
 266                 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
 267                 /* Program return */
 268                 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
 269                 /* Store instructions */
 270                 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
 271                 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
 272                 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
 273                 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
 274                 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
 275                 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
 276                 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
 277                 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
 278                 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
 279                 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
 280                 /* Load instructions */
 281                 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
 282                 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
 283                 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
 284                 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
 285                 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
 286                 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
 287                 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
 288                 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
 289                 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
 290                 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
 291                 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
 292         };
 293         u32 tail_call_cnt = 0;
 294         void *ptr;
 295         int off;
 296
 297 #define CONT     ({ insn++; goto select_insn; })
 298 #define CONT_JMP ({ insn++; goto select_insn; })
 299
 300         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
 301         ARG1 = (u64) (unsigned long) ctx;
 302
 303         /* Registers used in classic BPF programs need to be reset first. */
 304         regs[BPF_REG_A] = 0;
 305         regs[BPF_REG_X] = 0;
 306
 307 select_insn:
 308         goto *jumptable[insn->code];
 309
 310         /* ALU */
 311 #define ALU(OPCODE, OP)                 \
 312         ALU64_##OPCODE##_X:             \
 313                 DST = DST OP SRC;       \
 314                 CONT;                   \
 315         ALU_##OPCODE##_X:               \
 316                 DST = (u32) DST OP (u32) SRC;   \
 317                 CONT;                   \
 318         ALU64_##OPCODE##_K:             \
 319                 DST = DST OP IMM;               \
 320                 CONT;                   \
 321         ALU_##OPCODE##_K:               \
 322                 DST = (u32) DST OP (u32) IMM;   \
 323                 CONT;
 324
 325         ALU(ADD,  +)
 326         ALU(SUB,  -)
 327         ALU(AND,  &)
 328         ALU(OR,   |)
 329         ALU(LSH, <<)
 330         ALU(RSH, >>)
 331         ALU(XOR,  ^)
 332         ALU(MUL,  *)
 333 #undef ALU
 334         ALU_NEG:
 335                 DST = (u32) -DST;
 336                 CONT;
 337         ALU64_NEG:
 338                 DST = -DST;
 339                 CONT;
 340         ALU_MOV_X:
 341                 DST = (u32) SRC;
 342                 CONT;
 343         ALU_MOV_K:
 344                 DST = (u32) IMM;
 345                 CONT;
 346         ALU64_MOV_X:
 347                 DST = SRC;
 348                 CONT;
 349         ALU64_MOV_K:
 350                 DST = IMM;
 351                 CONT;
 352         LD_IMM_DW:
 353                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
 354                 insn++;
 355                 CONT;
 356         ALU64_ARSH_X:
 357                 (*(s64 *) &DST) >>= SRC;
 358                 CONT;
 359         ALU64_ARSH_K:
 360                 (*(s64 *) &DST) >>= IMM;
 361                 CONT;
 362         ALU64_MOD_X:
 363                 if (unlikely(SRC == 0))
 364                         return 0;
 365                 div64_u64_rem(DST, SRC, &tmp);
 366                 DST = tmp;
 367                 CONT;
 368         ALU_MOD_X:
 369                 if (unlikely(SRC == 0))
 370                         return 0;
 371                 tmp = (u32) DST;
 372                 DST = do_div(tmp, (u32) SRC);
 373                 CONT;
 374         ALU64_MOD_K:
 375                 div64_u64_rem(DST, IMM, &tmp);
 376                 DST = tmp;
 377                 CONT;
 378         ALU_MOD_K:
 379                 tmp = (u32) DST;
 380                 DST = do_div(tmp, (u32) IMM);
 381                 CONT;
 382         ALU64_DIV_X:
 383                 if (unlikely(SRC == 0))
 384                         return 0;
 385                 DST = div64_u64(DST, SRC);
 386                 CONT;
 387         ALU_DIV_X:
 388                 if (unlikely(SRC == 0))
 389                         return 0;
 390                 tmp = (u32) DST;
 391                 do_div(tmp, (u32) SRC);
 392                 DST = (u32) tmp;
 393                 CONT;
 394         ALU64_DIV_K:
 395                 DST = div64_u64(DST, IMM);
 396                 CONT;
 397         ALU_DIV_K:
 398                 tmp = (u32) DST;
 399                 do_div(tmp, (u32) IMM);
 400                 DST = (u32) tmp;
 401                 CONT;
 402         ALU_END_TO_BE:
 403                 switch (IMM) {
 404                 case 16:
 405                         DST = (__force u16) cpu_to_be16(DST);
 406                         break;
 407                 case 32:
 408                         DST = (__force u32) cpu_to_be32(DST);
 409                         break;
 410                 case 64:
 411                         DST = (__force u64) cpu_to_be64(DST);
 412                         break;
 413                 }
 414                 CONT;
 415         ALU_END_TO_LE:
 416                 switch (IMM) {
 417                 case 16:
 418                         DST = (__force u16) cpu_to_le16(DST);
 419                         break;
 420                 case 32:
 421                         DST = (__force u32) cpu_to_le32(DST);
 422                         break;
 423                 case 64:
 424                         DST = (__force u64) cpu_to_le64(DST);
 425                         break;
 426                 }
 427                 CONT;
 428
 429         /* CALL */
 430         JMP_CALL:
 431                 /* Function call scratches BPF_R1-BPF_R5 registers,
 432                  * preserves BPF_R6-BPF_R9, and stores return value
 433                  * into BPF_R0.
 434                  */
 435                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
 436                                                        BPF_R4, BPF_R5);
 437                 CONT;
 438
 439         JMP_TAIL_CALL: {
 440                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
 441                 struct bpf_array *array = container_of(map, struct bpf_array, map);
 442                 struct bpf_prog *prog;
 443                 u64 index = BPF_R3;
 444
 445                 if (unlikely(index >= array->map.max_entries))
 446                         goto out;
 447
 448                 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
 449                         goto out;
 450
 451                 tail_call_cnt++;
 452
 453                 prog = READ_ONCE(array->ptrs[index]);
 454                 if (unlikely(!prog))
 455                         goto out;
 456
 457                 /* ARG1 at this point is guaranteed to point to CTX from
 458                  * the verifier side due to the fact that the tail call is
 459                  * handeled like a helper, that is, bpf_tail_call_proto,
 460                  * where arg1_type is ARG_PTR_TO_CTX.
 461                  */
 462                 insn = prog->insnsi;
 463                 goto select_insn;
 464 out:
 465                 CONT;
 466         }
 467         /* JMP */
 468         JMP_JA:
 469                 insn += insn->off;
 470                 CONT;
 471         JMP_JEQ_X:
 472                 if (DST == SRC) {
 473                         insn += insn->off;
 474                         CONT_JMP;
 475                 }
 476                 CONT;
 477         JMP_JEQ_K:
 478                 if (DST == IMM) {
 479                         insn += insn->off;
 480                         CONT_JMP;
 481                 }
 482                 CONT;
 483         JMP_JNE_X:
 484                 if (DST != SRC) {
 485                         insn += insn->off;
 486                         CONT_JMP;
 487                 }
 488                 CONT;
 489         JMP_JNE_K:
 490                 if (DST != IMM) {
 491                         insn += insn->off;
 492                         CONT_JMP;
 493                 }
 494                 CONT;
 495         JMP_JGT_X:
 496                 if (DST > SRC) {
 497                         insn += insn->off;
 498                         CONT_JMP;
 499                 }
 500                 CONT;
 501         JMP_JGT_K:
 502                 if (DST > IMM) {
 503                         insn += insn->off;
 504                         CONT_JMP;
 505                 }
 506                 CONT;
 507         JMP_JGE_X:
 508                 if (DST >= SRC) {
 509                         insn += insn->off;
 510                         CONT_JMP;
 511                 }
 512                 CONT;
 513         JMP_JGE_K:
 514                 if (DST >= IMM) {
 515                         insn += insn->off;
 516                         CONT_JMP;
 517                 }
 518                 CONT;
 519         JMP_JSGT_X:
 520                 if (((s64) DST) > ((s64) SRC)) {
 521                         insn += insn->off;
 522                         CONT_JMP;
 523                 }
 524                 CONT;
 525         JMP_JSGT_K:
 526                 if (((s64) DST) > ((s64) IMM)) {
 527                         insn += insn->off;
 528                         CONT_JMP;
 529                 }
 530                 CONT;
 531         JMP_JSGE_X:
 532                 if (((s64) DST) >= ((s64) SRC)) {
 533                         insn += insn->off;
 534                         CONT_JMP;
 535                 }
 536                 CONT;
 537         JMP_JSGE_K:
 538                 if (((s64) DST) >= ((s64) IMM)) {
 539                         insn += insn->off;
 540                         CONT_JMP;
 541                 }
 542                 CONT;
 543         JMP_JSET_X:
 544                 if (DST & SRC) {
 545                         insn += insn->off;
 546                         CONT_JMP;
 547                 }
 548                 CONT;
 549         JMP_JSET_K:
 550                 if (DST & IMM) {
 551                         insn += insn->off;
 552                         CONT_JMP;
 553                 }
 554                 CONT;
 555         JMP_EXIT:
 556                 return BPF_R0;
 557
 558         /* STX and ST and LDX*/
 559 #define LDST(SIZEOP, SIZE)                                              \
 560         STX_MEM_##SIZEOP:                                               \
 561                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
 562                 CONT;                                                   \
 563         ST_MEM_##SIZEOP:                                                \
 564                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
 565                 CONT;                                                   \
 566         LDX_MEM_##SIZEOP:                                               \
 567                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
 568                 CONT;
 569
 570         LDST(B,   u8)
 571         LDST(H,  u16)
 572         LDST(W,  u32)
 573         LDST(DW, u64)
 574 #undef LDST
 575         STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
 576                 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
 577                            (DST + insn->off));
 578                 CONT;
 579         STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
 580                 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
 581                              (DST + insn->off));
 582                 CONT;
 583         LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
 584                 off = IMM;
 585 load_word:
 586                 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
 587                  * only appearing in the programs where ctx ==
 588                  * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
 589                  * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
 590                  * internal BPF verifier will check that BPF_R6 ==
 591                  * ctx.
 592                  *
 593                  * BPF_ABS and BPF_IND are wrappers of function calls,
 594                  * so they scratch BPF_R1-BPF_R5 registers, preserve
 595                  * BPF_R6-BPF_R9, and store return value into BPF_R0.
 596                  *
 597                  * Implicit input:
 598                  *   ctx == skb == BPF_R6 == CTX
 599                  *
 600                  * Explicit input:
 601                  *   SRC == any register
 602                  *   IMM == 32-bit immediate
 603                  *
 604                  * Output:
 605                  *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
 606                  */
 607
 608                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
 609                 if (likely(ptr != NULL)) {
 610                         BPF_R0 = get_unaligned_be32(ptr);
 611                         CONT;
 612                 }
 613
 614                 return 0;
 615         LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
 616                 off = IMM;
 617 load_half:
 618                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
 619                 if (likely(ptr != NULL)) {
 620                         BPF_R0 = get_unaligned_be16(ptr);
 621                         CONT;
 622                 }
 623
 624                 return 0;
 625         LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
 626                 off = IMM;
 627 load_byte:
 628                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
 629                 if (likely(ptr != NULL)) {
 630                         BPF_R0 = *(u8 *)ptr;
 631                         CONT;
 632                 }
 633
 634                 return 0;
 635         LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
 636                 off = IMM + SRC;
 637                 goto load_word;
 638         LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
 639                 off = IMM + SRC;
 640                 goto load_half;
 641         LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
 642                 off = IMM + SRC;
 643                 goto load_byte;
 644
 645         default_label:
 646                 /* If we ever reach this, we have a bug somewhere. */
 647                 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
 648                 return 0;
 649 }
 650
 651 bool bpf_prog_array_compatible(struct bpf_array *array,
 652                                const struct bpf_prog *fp)
 653 {
 654         if (!array->owner_prog_type) {
 655                 /* There's no owner yet where we could check for
 656                  * compatibility.
 657                  */
 658                 array->owner_prog_type = fp->type;
 659                 array->owner_jited = fp->jited;
 660
 661                 return true;
 662         }
 663
 664         return array->owner_prog_type == fp->type &&
 665                array->owner_jited == fp->jited;
 666 }
 667
 668 static int bpf_check_tail_call(const struct bpf_prog *fp)
 669 {
 670         struct bpf_prog_aux *aux = fp->aux;
 671         int i;
 672
 673         for (i = 0; i < aux->used_map_cnt; i++) {
 674                 struct bpf_map *map = aux->used_maps[i];
 675                 struct bpf_array *array;
 676
 677                 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
 678                         continue;
 679
 680                 array = container_of(map, struct bpf_array, map);
 681                 if (!bpf_prog_array_compatible(array, fp))
 682                         return -EINVAL;
 683         }
 684
 685         return 0;
 686 }
 687
 688 /**
 689  *      bpf_prog_select_runtime - select exec runtime for BPF program
 690  *      @fp: bpf_prog populated with internal BPF program
 691  *
 692  * Try to JIT eBPF program, if JIT is not available, use interpreter.
 693  * The BPF program will be executed via BPF_PROG_RUN() macro.
 694  */
 695 int bpf_prog_select_runtime(struct bpf_prog *fp)
 696 {
 697         fp->bpf_func = (void *) __bpf_prog_run;
 698
 699         bpf_int_jit_compile(fp);
 700         bpf_prog_lock_ro(fp);
 701
 702         /* The tail call compatibility check can only be done at
 703          * this late stage as we need to determine, if we deal
 704          * with JITed or non JITed program concatenations and not
 705          * all eBPF JITs might immediately support all features.
 706          */
 707         return bpf_check_tail_call(fp);
 708 }
 709 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
 710
 711 static void bpf_prog_free_deferred(struct work_struct *work)
 712 {
 713         struct bpf_prog_aux *aux;
 714
 715         aux = container_of(work, struct bpf_prog_aux, work);
 716         bpf_jit_free(aux->prog);
 717 }
 718
 719 /* Free internal BPF program */
 720 void bpf_prog_free(struct bpf_prog *fp)
 721 {
 722         struct bpf_prog_aux *aux = fp->aux;
 723
 724         INIT_WORK(&aux->work, bpf_prog_free_deferred);
 725         aux->prog = fp;
 726         schedule_work(&aux->work);
 727 }
 728 EXPORT_SYMBOL_GPL(bpf_prog_free);
 729
 730 /* Weak definitions of helper functions in case we don't have bpf syscall. */
 731 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
 732 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
 733 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
 734
 735 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
 736 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
 737 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
 738 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
 739 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
 740 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
 741 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
 742 {
 743         return NULL;
 744 }
 745
 746 /* Always built-in helper functions. */
 747 const struct bpf_func_proto bpf_tail_call_proto = {
 748         .func           = NULL,
 749         .gpl_only       = false,
 750         .ret_type       = RET_VOID,
 751         .arg1_type      = ARG_PTR_TO_CTX,
 752         .arg2_type      = ARG_CONST_MAP_PTR,
 753         .arg3_type      = ARG_ANYTHING,
 754 };
 755
 756 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
 757 void __weak bpf_int_jit_compile(struct bpf_prog *prog)
 758 {
 759 }
 760
 761 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
 762  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
 763  */
 764 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
 765                          int len)
 766 {
 767         return -EFAULT;
 768 }