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         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         /* Registers used in classic BPF programs need to be reset first. */
 310         regs[BPF_REG_A] = 0;
 311         regs[BPF_REG_X] = 0;
 312
 313 select_insn:
 314         goto *jumptable[insn->code];
 315
 316         /* ALU */
 317 #define ALU(OPCODE, OP)                 \
 318         ALU64_##OPCODE##_X:             \
 319                 DST = DST OP SRC;       \
 320                 CONT;                   \
 321         ALU_##OPCODE##_X:               \
 322                 DST = (u32) DST OP (u32) SRC;   \
 323                 CONT;                   \
 324         ALU64_##OPCODE##_K:             \
 325                 DST = DST OP IMM;               \
 326                 CONT;                   \
 327         ALU_##OPCODE##_K:               \
 328                 DST = (u32) DST OP (u32) IMM;   \
 329                 CONT;
 330
 331         ALU(ADD,  +)
 332         ALU(SUB,  -)
 333         ALU(AND,  &)
 334         ALU(OR,   |)
 335         ALU(LSH, <<)
 336         ALU(RSH, >>)
 337         ALU(XOR,  ^)
 338         ALU(MUL,  *)
 339 #undef ALU
 340         ALU_NEG:
 341                 DST = (u32) -DST;
 342                 CONT;
 343         ALU64_NEG:
 344                 DST = -DST;
 345                 CONT;
 346         ALU_MOV_X:
 347                 DST = (u32) SRC;
 348                 CONT;
 349         ALU_MOV_K:
 350                 DST = (u32) IMM;
 351                 CONT;
 352         ALU64_MOV_X:
 353                 DST = SRC;
 354                 CONT;
 355         ALU64_MOV_K:
 356                 DST = IMM;
 357                 CONT;
 358         LD_IMM_DW:
 359                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
 360                 insn++;
 361                 CONT;
 362         ALU64_ARSH_X:
 363                 (*(s64 *) &DST) >>= SRC;
 364                 CONT;
 365         ALU64_ARSH_K:
 366                 (*(s64 *) &DST) >>= IMM;
 367                 CONT;
 368         ALU64_MOD_X:
 369                 if (unlikely(SRC == 0))
 370                         return 0;
 371                 div64_u64_rem(DST, SRC, &tmp);
 372                 DST = tmp;
 373                 CONT;
 374         ALU_MOD_X:
 375                 if (unlikely(SRC == 0))
 376                         return 0;
 377                 tmp = (u32) DST;
 378                 DST = do_div(tmp, (u32) SRC);
 379                 CONT;
 380         ALU64_MOD_K:
 381                 div64_u64_rem(DST, IMM, &tmp);
 382                 DST = tmp;
 383                 CONT;
 384         ALU_MOD_K:
 385                 tmp = (u32) DST;
 386                 DST = do_div(tmp, (u32) IMM);
 387                 CONT;
 388         ALU64_DIV_X:
 389                 if (unlikely(SRC == 0))
 390                         return 0;
 391                 DST = div64_u64(DST, SRC);
 392                 CONT;
 393         ALU_DIV_X:
 394                 if (unlikely(SRC == 0))
 395                         return 0;
 396                 tmp = (u32) DST;
 397                 do_div(tmp, (u32) SRC);
 398                 DST = (u32) tmp;
 399                 CONT;
 400         ALU64_DIV_K:
 401                 DST = div64_u64(DST, IMM);
 402                 CONT;
 403         ALU_DIV_K:
 404                 tmp = (u32) DST;
 405                 do_div(tmp, (u32) IMM);
 406                 DST = (u32) tmp;
 407                 CONT;
 408         ALU_END_TO_BE:
 409                 switch (IMM) {
 410                 case 16:
 411                         DST = (__force u16) cpu_to_be16(DST);
 412                         break;
 413                 case 32:
 414                         DST = (__force u32) cpu_to_be32(DST);
 415                         break;
 416                 case 64:
 417                         DST = (__force u64) cpu_to_be64(DST);
 418                         break;
 419                 }
 420                 CONT;
 421         ALU_END_TO_LE:
 422                 switch (IMM) {
 423                 case 16:
 424                         DST = (__force u16) cpu_to_le16(DST);
 425                         break;
 426                 case 32:
 427                         DST = (__force u32) cpu_to_le32(DST);
 428                         break;
 429                 case 64:
 430                         DST = (__force u64) cpu_to_le64(DST);
 431                         break;
 432                 }
 433                 CONT;
 434
 435         /* CALL */
 436         JMP_CALL:
 437                 /* Function call scratches BPF_R1-BPF_R5 registers,
 438                  * preserves BPF_R6-BPF_R9, and stores return value
 439                  * into BPF_R0.
 440                  */
 441                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
 442                                                        BPF_R4, BPF_R5);
 443                 CONT;
 444
 445         JMP_TAIL_CALL: {
 446                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
 447                 struct bpf_array *array = container_of(map, struct bpf_array, map);
 448                 struct bpf_prog *prog;
 449                 u64 index = BPF_R3;
 450
 451                 if (unlikely(index >= array->map.max_entries))
 452                         goto out;
 453
 454                 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
 455                         goto out;
 456
 457                 tail_call_cnt++;
 458
 459                 prog = READ_ONCE(array->ptrs[index]);
 460                 if (unlikely(!prog))
 461                         goto out;
 462
 463                 /* ARG1 at this point is guaranteed to point to CTX from
 464                  * the verifier side due to the fact that the tail call is
 465                  * handeled like a helper, that is, bpf_tail_call_proto,
 466                  * where arg1_type is ARG_PTR_TO_CTX.
 467                  */
 468                 insn = prog->insnsi;
 469                 goto select_insn;
 470 out:
 471                 CONT;
 472         }
 473         /* JMP */
 474         JMP_JA:
 475                 insn += insn->off;
 476                 CONT;
 477         JMP_JEQ_X:
 478                 if (DST == SRC) {
 479                         insn += insn->off;
 480                         CONT_JMP;
 481                 }
 482                 CONT;
 483         JMP_JEQ_K:
 484                 if (DST == IMM) {
 485                         insn += insn->off;
 486                         CONT_JMP;
 487                 }
 488                 CONT;
 489         JMP_JNE_X:
 490                 if (DST != SRC) {
 491                         insn += insn->off;
 492                         CONT_JMP;
 493                 }
 494                 CONT;
 495         JMP_JNE_K:
 496                 if (DST != IMM) {
 497                         insn += insn->off;
 498                         CONT_JMP;
 499                 }
 500                 CONT;
 501         JMP_JGT_X:
 502                 if (DST > SRC) {
 503                         insn += insn->off;
 504                         CONT_JMP;
 505                 }
 506                 CONT;
 507         JMP_JGT_K:
 508                 if (DST > IMM) {
 509                         insn += insn->off;
 510                         CONT_JMP;
 511                 }
 512                 CONT;
 513         JMP_JGE_X:
 514                 if (DST >= SRC) {
 515                         insn += insn->off;
 516                         CONT_JMP;
 517                 }
 518                 CONT;
 519         JMP_JGE_K:
 520                 if (DST >= IMM) {
 521                         insn += insn->off;
 522                         CONT_JMP;
 523                 }
 524                 CONT;
 525         JMP_JSGT_X:
 526                 if (((s64) DST) > ((s64) SRC)) {
 527                         insn += insn->off;
 528                         CONT_JMP;
 529                 }
 530                 CONT;
 531         JMP_JSGT_K:
 532                 if (((s64) DST) > ((s64) IMM)) {
 533                         insn += insn->off;
 534                         CONT_JMP;
 535                 }
 536                 CONT;
 537         JMP_JSGE_X:
 538                 if (((s64) DST) >= ((s64) SRC)) {
 539                         insn += insn->off;
 540                         CONT_JMP;
 541                 }
 542                 CONT;
 543         JMP_JSGE_K:
 544                 if (((s64) DST) >= ((s64) IMM)) {
 545                         insn += insn->off;
 546                         CONT_JMP;
 547                 }
 548                 CONT;
 549         JMP_JSET_X:
 550                 if (DST & SRC) {
 551                         insn += insn->off;
 552                         CONT_JMP;
 553                 }
 554                 CONT;
 555         JMP_JSET_K:
 556                 if (DST & IMM) {
 557                         insn += insn->off;
 558                         CONT_JMP;
 559                 }
 560                 CONT;
 561         JMP_EXIT:
 562                 return BPF_R0;
 563
 564         /* STX and ST and LDX*/
 565 #define LDST(SIZEOP, SIZE)                                              \
 566         STX_MEM_##SIZEOP:                                               \
 567                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
 568                 CONT;                                                   \
 569         ST_MEM_##SIZEOP:                                                \
 570                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
 571                 CONT;                                                   \
 572         LDX_MEM_##SIZEOP:                                               \
 573                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
 574                 CONT;
 575
 576         LDST(B,   u8)
 577         LDST(H,  u16)
 578         LDST(W,  u32)
 579         LDST(DW, u64)
 580 #undef LDST
 581         STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
 582                 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
 583                            (DST + insn->off));
 584                 CONT;
 585         STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
 586                 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
 587                              (DST + insn->off));
 588                 CONT;
 589         LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
 590                 off = IMM;
 591 load_word:
 592                 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
 593                  * only appearing in the programs where ctx ==
 594                  * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
 595                  * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
 596                  * internal BPF verifier will check that BPF_R6 ==
 597                  * ctx.
 598                  *
 599                  * BPF_ABS and BPF_IND are wrappers of function calls,
 600                  * so they scratch BPF_R1-BPF_R5 registers, preserve
 601                  * BPF_R6-BPF_R9, and store return value into BPF_R0.
 602                  *
 603                  * Implicit input:
 604                  *   ctx == skb == BPF_R6 == CTX
 605                  *
 606                  * Explicit input:
 607                  *   SRC == any register
 608                  *   IMM == 32-bit immediate
 609                  *
 610                  * Output:
 611                  *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
 612                  */
 613
 614                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
 615                 if (likely(ptr != NULL)) {
 616                         BPF_R0 = get_unaligned_be32(ptr);
 617                         CONT;
 618                 }
 619
 620                 return 0;
 621         LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
 622                 off = IMM;
 623 load_half:
 624                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
 625                 if (likely(ptr != NULL)) {
 626                         BPF_R0 = get_unaligned_be16(ptr);
 627                         CONT;
 628                 }
 629
 630                 return 0;
 631         LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
 632                 off = IMM;
 633 load_byte:
 634                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
 635                 if (likely(ptr != NULL)) {
 636                         BPF_R0 = *(u8 *)ptr;
 637                         CONT;
 638                 }
 639
 640                 return 0;
 641         LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
 642                 off = IMM + SRC;
 643                 goto load_word;
 644         LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
 645                 off = IMM + SRC;
 646                 goto load_half;
 647         LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
 648                 off = IMM + SRC;
 649                 goto load_byte;
 650
 651         default_label:
 652                 /* If we ever reach this, we have a bug somewhere. */
 653                 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
 654                 return 0;
 655 }
 656
 657 bool bpf_prog_array_compatible(struct bpf_array *array,
 658                                const struct bpf_prog *fp)
 659 {
 660         if (!array->owner_prog_type) {
 661                 /* There's no owner yet where we could check for
 662                  * compatibility.
 663                  */
 664                 array->owner_prog_type = fp->type;
 665                 array->owner_jited = fp->jited;
 666
 667                 return true;
 668         }
 669
 670         return array->owner_prog_type == fp->type &&
 671                array->owner_jited == fp->jited;
 672 }
 673
 674 static int bpf_check_tail_call(const struct bpf_prog *fp)
 675 {
 676         struct bpf_prog_aux *aux = fp->aux;
 677         int i;
 678
 679         for (i = 0; i < aux->used_map_cnt; i++) {
 680                 struct bpf_map *map = aux->used_maps[i];
 681                 struct bpf_array *array;
 682
 683                 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
 684                         continue;
 685
 686                 array = container_of(map, struct bpf_array, map);
 687                 if (!bpf_prog_array_compatible(array, fp))
 688                         return -EINVAL;
 689         }
 690
 691         return 0;
 692 }
 693
 694 /**
 695  *      bpf_prog_select_runtime - select exec runtime for BPF program
 696  *      @fp: bpf_prog populated with internal BPF program
 697  *
 698  * Try to JIT eBPF program, if JIT is not available, use interpreter.
 699  * The BPF program will be executed via BPF_PROG_RUN() macro.
 700  */
 701 int bpf_prog_select_runtime(struct bpf_prog *fp)
 702 {
 703         fp->bpf_func = (void *) __bpf_prog_run;
 704
 705         bpf_int_jit_compile(fp);
 706         bpf_prog_lock_ro(fp);
 707
 708         /* The tail call compatibility check can only be done at
 709          * this late stage as we need to determine, if we deal
 710          * with JITed or non JITed program concatenations and not
 711          * all eBPF JITs might immediately support all features.
 712          */
 713         return bpf_check_tail_call(fp);
 714 }
 715 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
 716
 717 static void bpf_prog_free_deferred(struct work_struct *work)
 718 {
 719         struct bpf_prog_aux *aux;
 720
 721         aux = container_of(work, struct bpf_prog_aux, work);
 722         bpf_jit_free(aux->prog);
 723 }
 724
 725 /* Free internal BPF program */
 726 void bpf_prog_free(struct bpf_prog *fp)
 727 {
 728         struct bpf_prog_aux *aux = fp->aux;
 729
 730         INIT_WORK(&aux->work, bpf_prog_free_deferred);
 731         schedule_work(&aux->work);
 732 }
 733 EXPORT_SYMBOL_GPL(bpf_prog_free);
 734
 735 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
 736 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
 737
 738 void bpf_user_rnd_init_once(void)
 739 {
 740         prandom_init_once(&bpf_user_rnd_state);
 741 }
 742
 743 u64 bpf_user_rnd_u32(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 744 {
 745         /* Should someone ever have the rather unwise idea to use some
 746          * of the registers passed into this function, then note that
 747          * this function is called from native eBPF and classic-to-eBPF
 748          * transformations. Register assignments from both sides are
 749          * different, f.e. classic always sets fn(ctx, A, X) here.
 750          */
 751         struct rnd_state *state;
 752         u32 res;
 753
 754         state = &get_cpu_var(bpf_user_rnd_state);
 755         res = prandom_u32_state(state);
 756         put_cpu_var(state);
 757
 758         return res;
 759 }
 760
 761 /* Weak definitions of helper functions in case we don't have bpf syscall. */
 762 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
 763 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
 764 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
 765
 766 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
 767 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
 768 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
 769 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
 770 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
 771 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
 772 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
 773 {
 774         return NULL;
 775 }
 776
 777 /* Always built-in helper functions. */
 778 const struct bpf_func_proto bpf_tail_call_proto = {
 779         .func           = NULL,
 780         .gpl_only       = false,
 781         .ret_type       = RET_VOID,
 782         .arg1_type      = ARG_PTR_TO_CTX,
 783         .arg2_type      = ARG_CONST_MAP_PTR,
 784         .arg3_type      = ARG_ANYTHING,
 785 };
 786
 787 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
 788 void __weak bpf_int_jit_compile(struct bpf_prog *prog)
 789 {
 790 }
 791
 792 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
 793  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
 794  */
 795 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
 796                          int len)
 797 {
 798         return -EFAULT;
 799 }