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         kmemcheck_annotate_bitfield(fp, meta);
  86
  87         aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
  88         if (aux == NULL) {
  89                 vfree(fp);
  90                 return NULL;
  91         }
  92
  93         fp->pages = size / PAGE_SIZE;
  94         fp->aux = aux;
  95
  96         return fp;
  97 }
  98 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
  99
 100 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
 101                                   gfp_t gfp_extra_flags)
 102 {
 103         gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
 104                           gfp_extra_flags;
 105         struct bpf_prog *fp;
 106
 107         BUG_ON(fp_old == NULL);
 108
 109         size = round_up(size, PAGE_SIZE);
 110         if (size <= fp_old->pages * PAGE_SIZE)
 111                 return fp_old;
 112
 113         fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
 114         if (fp != NULL) {
 115                 kmemcheck_annotate_bitfield(fp, meta);
 116
 117                 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
 118                 fp->pages = size / PAGE_SIZE;
 119
 120                 /* We keep fp->aux from fp_old around in the new
 121                  * reallocated structure.
 122                  */
 123                 fp_old->aux = NULL;
 124                 __bpf_prog_free(fp_old);
 125         }
 126
 127         return fp;
 128 }
 129 EXPORT_SYMBOL_GPL(bpf_prog_realloc);
 130
 131 void __bpf_prog_free(struct bpf_prog *fp)
 132 {
 133         kfree(fp->aux);
 134         vfree(fp);
 135 }
 136 EXPORT_SYMBOL_GPL(__bpf_prog_free);
 137
 138 #ifdef CONFIG_BPF_JIT
 139 struct bpf_binary_header *
 140 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
 141                      unsigned int alignment,
 142                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
 143 {
 144         struct bpf_binary_header *hdr;
 145         unsigned int size, hole, start;
 146
 147         /* Most of BPF filters are really small, but if some of them
 148          * fill a page, allow at least 128 extra bytes to insert a
 149          * random section of illegal instructions.
 150          */
 151         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
 152         hdr = module_alloc(size);
 153         if (hdr == NULL)
 154                 return NULL;
 155
 156         /* Fill space with illegal/arch-dep instructions. */
 157         bpf_fill_ill_insns(hdr, size);
 158
 159         hdr->pages = size / PAGE_SIZE;
 160         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
 161                      PAGE_SIZE - sizeof(*hdr));
 162         start = (prandom_u32() % hole) & ~(alignment - 1);
 163
 164         /* Leave a random number of instructions before BPF code. */
 165         *image_ptr = &hdr->image[start];
 166
 167         return hdr;
 168 }
 169
 170 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
 171 {
 172         module_memfree(hdr);
 173 }
 174 #endif /* CONFIG_BPF_JIT */
 175
 176 /* Base function for offset calculation. Needs to go into .text section,
 177  * therefore keeping it non-static as well; will also be used by JITs
 178  * anyway later on, so do not let the compiler omit it.
 179  */
 180 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 181 {
 182         return 0;
 183 }
 184 EXPORT_SYMBOL_GPL(__bpf_call_base);
 185
 186 /**
 187  *      __bpf_prog_run - run eBPF program on a given context
 188  *      @ctx: is the data we are operating on
 189  *      @insn: is the array of eBPF instructions
 190  *
 191  * Decode and execute eBPF instructions.
 192  */
 193 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
 194 {
 195         u64 stack[MAX_BPF_STACK / sizeof(u64)];
 196         u64 regs[MAX_BPF_REG], tmp;
 197         static const void *jumptable[256] = {
 198                 [0 ... 255] = &&default_label,
 199                 /* Now overwrite non-defaults ... */
 200                 /* 32 bit ALU operations */
 201                 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
 202                 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
 203                 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
 204                 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
 205                 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
 206                 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
 207                 [BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
 208                 [BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
 209                 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
 210                 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
 211                 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
 212                 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
 213                 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
 214                 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
 215                 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
 216                 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
 217                 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
 218                 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
 219                 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
 220                 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
 221                 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
 222                 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
 223                 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
 224                 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
 225                 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
 226                 /* 64 bit ALU operations */
 227                 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
 228                 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
 229                 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
 230                 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
 231                 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
 232                 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
 233                 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
 234                 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
 235                 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
 236                 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
 237                 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
 238                 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
 239                 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
 240                 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
 241                 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
 242                 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
 243                 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
 244                 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
 245                 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
 246                 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
 247                 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
 248                 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
 249                 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
 250                 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
 251                 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
 252                 /* Call instruction */
 253                 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
 254                 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
 255                 /* Jumps */
 256                 [BPF_JMP | BPF_JA] = &&JMP_JA,
 257                 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
 258                 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
 259                 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
 260                 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
 261                 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
 262                 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
 263                 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
 264                 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
 265                 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
 266                 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
 267                 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
 268                 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
 269                 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
 270                 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
 271                 /* Program return */
 272                 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
 273                 /* Store instructions */
 274                 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
 275                 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
 276                 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
 277                 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
 278                 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
 279                 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
 280                 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
 281                 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
 282                 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
 283                 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
 284                 /* Load instructions */
 285                 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
 286                 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
 287                 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
 288                 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
 289                 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
 290                 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
 291                 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
 292                 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
 293                 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
 294                 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
 295                 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
 296         };
 297         u32 tail_call_cnt = 0;
 298         void *ptr;
 299         int off;
 300
 301 #define CONT     ({ insn++; goto select_insn; })
 302 #define CONT_JMP ({ insn++; goto select_insn; })
 303
 304         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
 305         ARG1 = (u64) (unsigned long) ctx;
 306
 307         /* Registers used in classic BPF programs need to be reset first. */
 308         regs[BPF_REG_A] = 0;
 309         regs[BPF_REG_X] = 0;
 310
 311 select_insn:
 312         goto *jumptable[insn->code];
 313
 314         /* ALU */
 315 #define ALU(OPCODE, OP)                 \
 316         ALU64_##OPCODE##_X:             \
 317                 DST = DST OP SRC;       \
 318                 CONT;                   \
 319         ALU_##OPCODE##_X:               \
 320                 DST = (u32) DST OP (u32) SRC;   \
 321                 CONT;                   \
 322         ALU64_##OPCODE##_K:             \
 323                 DST = DST OP IMM;               \
 324                 CONT;                   \
 325         ALU_##OPCODE##_K:               \
 326                 DST = (u32) DST OP (u32) IMM;   \
 327                 CONT;
 328
 329         ALU(ADD,  +)
 330         ALU(SUB,  -)
 331         ALU(AND,  &)
 332         ALU(OR,   |)
 333         ALU(LSH, <<)
 334         ALU(RSH, >>)
 335         ALU(XOR,  ^)
 336         ALU(MUL,  *)
 337 #undef ALU
 338         ALU_NEG:
 339                 DST = (u32) -DST;
 340                 CONT;
 341         ALU64_NEG:
 342                 DST = -DST;
 343                 CONT;
 344         ALU_MOV_X:
 345                 DST = (u32) SRC;
 346                 CONT;
 347         ALU_MOV_K:
 348                 DST = (u32) IMM;
 349                 CONT;
 350         ALU64_MOV_X:
 351                 DST = SRC;
 352                 CONT;
 353         ALU64_MOV_K:
 354                 DST = IMM;
 355                 CONT;
 356         LD_IMM_DW:
 357                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
 358                 insn++;
 359                 CONT;
 360         ALU64_ARSH_X:
 361                 (*(s64 *) &DST) >>= SRC;
 362                 CONT;
 363         ALU64_ARSH_K:
 364                 (*(s64 *) &DST) >>= IMM;
 365                 CONT;
 366         ALU64_MOD_X:
 367                 if (unlikely(SRC == 0))
 368                         return 0;
 369                 div64_u64_rem(DST, SRC, &tmp);
 370                 DST = tmp;
 371                 CONT;
 372         ALU_MOD_X:
 373                 if (unlikely(SRC == 0))
 374                         return 0;
 375                 tmp = (u32) DST;
 376                 DST = do_div(tmp, (u32) SRC);
 377                 CONT;
 378         ALU64_MOD_K:
 379                 div64_u64_rem(DST, IMM, &tmp);
 380                 DST = tmp;
 381                 CONT;
 382         ALU_MOD_K:
 383                 tmp = (u32) DST;
 384                 DST = do_div(tmp, (u32) IMM);
 385                 CONT;
 386         ALU64_DIV_X:
 387                 if (unlikely(SRC == 0))
 388                         return 0;
 389                 DST = div64_u64(DST, SRC);
 390                 CONT;
 391         ALU_DIV_X:
 392                 if (unlikely(SRC == 0))
 393                         return 0;
 394                 tmp = (u32) DST;
 395                 do_div(tmp, (u32) SRC);
 396                 DST = (u32) tmp;
 397                 CONT;
 398         ALU64_DIV_K:
 399                 DST = div64_u64(DST, IMM);
 400                 CONT;
 401         ALU_DIV_K:
 402                 tmp = (u32) DST;
 403                 do_div(tmp, (u32) IMM);
 404                 DST = (u32) tmp;
 405                 CONT;
 406         ALU_END_TO_BE:
 407                 switch (IMM) {
 408                 case 16:
 409                         DST = (__force u16) cpu_to_be16(DST);
 410                         break;
 411                 case 32:
 412                         DST = (__force u32) cpu_to_be32(DST);
 413                         break;
 414                 case 64:
 415                         DST = (__force u64) cpu_to_be64(DST);
 416                         break;
 417                 }
 418                 CONT;
 419         ALU_END_TO_LE:
 420                 switch (IMM) {
 421                 case 16:
 422                         DST = (__force u16) cpu_to_le16(DST);
 423                         break;
 424                 case 32:
 425                         DST = (__force u32) cpu_to_le32(DST);
 426                         break;
 427                 case 64:
 428                         DST = (__force u64) cpu_to_le64(DST);
 429                         break;
 430                 }
 431                 CONT;
 432
 433         /* CALL */
 434         JMP_CALL:
 435                 /* Function call scratches BPF_R1-BPF_R5 registers,
 436                  * preserves BPF_R6-BPF_R9, and stores return value
 437                  * into BPF_R0.
 438                  */
 439                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
 440                                                        BPF_R4, BPF_R5);
 441                 CONT;
 442
 443         JMP_TAIL_CALL: {
 444                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
 445                 struct bpf_array *array = container_of(map, struct bpf_array, map);
 446                 struct bpf_prog *prog;
 447                 u64 index = BPF_R3;
 448
 449                 if (unlikely(index >= array->map.max_entries))
 450                         goto out;
 451
 452                 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
 453                         goto out;
 454
 455                 tail_call_cnt++;
 456
 457                 prog = READ_ONCE(array->ptrs[index]);
 458                 if (unlikely(!prog))
 459                         goto out;
 460
 461                 /* ARG1 at this point is guaranteed to point to CTX from
 462                  * the verifier side due to the fact that the tail call is
 463                  * handeled like a helper, that is, bpf_tail_call_proto,
 464                  * where arg1_type is ARG_PTR_TO_CTX.
 465                  */
 466                 insn = prog->insnsi;
 467                 goto select_insn;
 468 out:
 469                 CONT;
 470         }
 471         /* JMP */
 472         JMP_JA:
 473                 insn += insn->off;
 474                 CONT;
 475         JMP_JEQ_X:
 476                 if (DST == SRC) {
 477                         insn += insn->off;
 478                         CONT_JMP;
 479                 }
 480                 CONT;
 481         JMP_JEQ_K:
 482                 if (DST == IMM) {
 483                         insn += insn->off;
 484                         CONT_JMP;
 485                 }
 486                 CONT;
 487         JMP_JNE_X:
 488                 if (DST != SRC) {
 489                         insn += insn->off;
 490                         CONT_JMP;
 491                 }
 492                 CONT;
 493         JMP_JNE_K:
 494                 if (DST != IMM) {
 495                         insn += insn->off;
 496                         CONT_JMP;
 497                 }
 498                 CONT;
 499         JMP_JGT_X:
 500                 if (DST > SRC) {
 501                         insn += insn->off;
 502                         CONT_JMP;
 503                 }
 504                 CONT;
 505         JMP_JGT_K:
 506                 if (DST > IMM) {
 507                         insn += insn->off;
 508                         CONT_JMP;
 509                 }
 510                 CONT;
 511         JMP_JGE_X:
 512                 if (DST >= SRC) {
 513                         insn += insn->off;
 514                         CONT_JMP;
 515                 }
 516                 CONT;
 517         JMP_JGE_K:
 518                 if (DST >= IMM) {
 519                         insn += insn->off;
 520                         CONT_JMP;
 521                 }
 522                 CONT;
 523         JMP_JSGT_X:
 524                 if (((s64) DST) > ((s64) SRC)) {
 525                         insn += insn->off;
 526                         CONT_JMP;
 527                 }
 528                 CONT;
 529         JMP_JSGT_K:
 530                 if (((s64) DST) > ((s64) IMM)) {
 531                         insn += insn->off;
 532                         CONT_JMP;
 533                 }
 534                 CONT;
 535         JMP_JSGE_X:
 536                 if (((s64) DST) >= ((s64) SRC)) {
 537                         insn += insn->off;
 538                         CONT_JMP;
 539                 }
 540                 CONT;
 541         JMP_JSGE_K:
 542                 if (((s64) DST) >= ((s64) IMM)) {
 543                         insn += insn->off;
 544                         CONT_JMP;
 545                 }
 546                 CONT;
 547         JMP_JSET_X:
 548                 if (DST & SRC) {
 549                         insn += insn->off;
 550                         CONT_JMP;
 551                 }
 552                 CONT;
 553         JMP_JSET_K:
 554                 if (DST & IMM) {
 555                         insn += insn->off;
 556                         CONT_JMP;
 557                 }
 558                 CONT;
 559         JMP_EXIT:
 560                 return BPF_R0;
 561
 562         /* STX and ST and LDX*/
 563 #define LDST(SIZEOP, SIZE)                                              \
 564         STX_MEM_##SIZEOP:                                               \
 565                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
 566                 CONT;                                                   \
 567         ST_MEM_##SIZEOP:                                                \
 568                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
 569                 CONT;                                                   \
 570         LDX_MEM_##SIZEOP:                                               \
 571                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
 572                 CONT;
 573
 574         LDST(B,   u8)
 575         LDST(H,  u16)
 576         LDST(W,  u32)
 577         LDST(DW, u64)
 578 #undef LDST
 579         STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
 580                 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
 581                            (DST + insn->off));
 582                 CONT;
 583         STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
 584                 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
 585                              (DST + insn->off));
 586                 CONT;
 587         LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
 588                 off = IMM;
 589 load_word:
 590                 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
 591                  * only appearing in the programs where ctx ==
 592                  * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
 593                  * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
 594                  * internal BPF verifier will check that BPF_R6 ==
 595                  * ctx.
 596                  *
 597                  * BPF_ABS and BPF_IND are wrappers of function calls,
 598                  * so they scratch BPF_R1-BPF_R5 registers, preserve
 599                  * BPF_R6-BPF_R9, and store return value into BPF_R0.
 600                  *
 601                  * Implicit input:
 602                  *   ctx == skb == BPF_R6 == CTX
 603                  *
 604                  * Explicit input:
 605                  *   SRC == any register
 606                  *   IMM == 32-bit immediate
 607                  *
 608                  * Output:
 609                  *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
 610                  */
 611
 612                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
 613                 if (likely(ptr != NULL)) {
 614                         BPF_R0 = get_unaligned_be32(ptr);
 615                         CONT;
 616                 }
 617
 618                 return 0;
 619         LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
 620                 off = IMM;
 621 load_half:
 622                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
 623                 if (likely(ptr != NULL)) {
 624                         BPF_R0 = get_unaligned_be16(ptr);
 625                         CONT;
 626                 }
 627
 628                 return 0;
 629         LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
 630                 off = IMM;
 631 load_byte:
 632                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
 633                 if (likely(ptr != NULL)) {
 634                         BPF_R0 = *(u8 *)ptr;
 635                         CONT;
 636                 }
 637
 638                 return 0;
 639         LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
 640                 off = IMM + SRC;
 641                 goto load_word;
 642         LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
 643                 off = IMM + SRC;
 644                 goto load_half;
 645         LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
 646                 off = IMM + SRC;
 647                 goto load_byte;
 648
 649         default_label:
 650                 /* If we ever reach this, we have a bug somewhere. */
 651                 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
 652                 return 0;
 653 }
 654
 655 bool bpf_prog_array_compatible(struct bpf_array *array,
 656                                const struct bpf_prog *fp)
 657 {
 658         if (!array->owner_prog_type) {
 659                 /* There's no owner yet where we could check for
 660                  * compatibility.
 661                  */
 662                 array->owner_prog_type = fp->type;
 663                 array->owner_jited = fp->jited;
 664
 665                 return true;
 666         }
 667
 668         return array->owner_prog_type == fp->type &&
 669                array->owner_jited == fp->jited;
 670 }
 671
 672 static int bpf_check_tail_call(const struct bpf_prog *fp)
 673 {
 674         struct bpf_prog_aux *aux = fp->aux;
 675         int i;
 676
 677         for (i = 0; i < aux->used_map_cnt; i++) {
 678                 struct bpf_map *map = aux->used_maps[i];
 679                 struct bpf_array *array;
 680
 681                 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
 682                         continue;
 683
 684                 array = container_of(map, struct bpf_array, map);
 685                 if (!bpf_prog_array_compatible(array, fp))
 686                         return -EINVAL;
 687         }
 688
 689         return 0;
 690 }
 691
 692 /**
 693  *      bpf_prog_select_runtime - select exec runtime for BPF program
 694  *      @fp: bpf_prog populated with internal BPF program
 695  *
 696  * Try to JIT eBPF program, if JIT is not available, use interpreter.
 697  * The BPF program will be executed via BPF_PROG_RUN() macro.
 698  */
 699 int bpf_prog_select_runtime(struct bpf_prog *fp)
 700 {
 701         fp->bpf_func = (void *) __bpf_prog_run;
 702
 703         bpf_int_jit_compile(fp);
 704         bpf_prog_lock_ro(fp);
 705
 706         /* The tail call compatibility check can only be done at
 707          * this late stage as we need to determine, if we deal
 708          * with JITed or non JITed program concatenations and not
 709          * all eBPF JITs might immediately support all features.
 710          */
 711         return bpf_check_tail_call(fp);
 712 }
 713 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
 714
 715 static void bpf_prog_free_deferred(struct work_struct *work)
 716 {
 717         struct bpf_prog_aux *aux;
 718
 719         aux = container_of(work, struct bpf_prog_aux, work);
 720         bpf_jit_free(aux->prog);
 721 }
 722
 723 /* Free internal BPF program */
 724 void bpf_prog_free(struct bpf_prog *fp)
 725 {
 726         struct bpf_prog_aux *aux = fp->aux;
 727
 728         INIT_WORK(&aux->work, bpf_prog_free_deferred);
 729         aux->prog = fp;
 730         schedule_work(&aux->work);
 731 }
 732 EXPORT_SYMBOL_GPL(bpf_prog_free);
 733
 734 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
 735 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
 736
 737 void bpf_user_rnd_init_once(void)
 738 {
 739         prandom_init_once(&bpf_user_rnd_state);
 740 }
 741
 742 u64 bpf_user_rnd_u32(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 743 {
 744         /* Should someone ever have the rather unwise idea to use some
 745          * of the registers passed into this function, then note that
 746          * this function is called from native eBPF and classic-to-eBPF
 747          * transformations. Register assignments from both sides are
 748          * different, f.e. classic always sets fn(ctx, A, X) here.
 749          */
 750         struct rnd_state *state;
 751         u32 res;
 752
 753         state = &get_cpu_var(bpf_user_rnd_state);
 754         res = prandom_u32_state(state);
 755         put_cpu_var(state);
 756
 757         return res;
 758 }
 759
 760 /* Weak definitions of helper functions in case we don't have bpf syscall. */
 761 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
 762 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
 763 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
 764
 765 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
 766 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
 767 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
 768 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
 769 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
 770 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
 771 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
 772 {
 773         return NULL;
 774 }
 775
 776 /* Always built-in helper functions. */
 777 const struct bpf_func_proto bpf_tail_call_proto = {
 778         .func           = NULL,
 779         .gpl_only       = false,
 780         .ret_type       = RET_VOID,
 781         .arg1_type      = ARG_PTR_TO_CTX,
 782         .arg2_type      = ARG_CONST_MAP_PTR,
 783         .arg3_type      = ARG_ANYTHING,
 784 };
 785
 786 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
 787 void __weak bpf_int_jit_compile(struct bpf_prog *prog)
 788 {
 789 }
 790
 791 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
 792  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
 793  */
 794 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
 795                          int len)
 796 {
 797         return -EFAULT;
 798 }