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