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 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
 141 {
 142         return BPF_CLASS(insn->code) == BPF_JMP  &&
 143                /* Call and Exit are both special jumps with no
 144                 * target inside the BPF instruction image.
 145                 */
 146                BPF_OP(insn->code) != BPF_CALL &&
 147                BPF_OP(insn->code) != BPF_EXIT;
 148 }
 149
 150 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
 151 {
 152         struct bpf_insn *insn = prog->insnsi;
 153         u32 i, insn_cnt = prog->len;
 154
 155         for (i = 0; i < insn_cnt; i++, insn++) {
 156                 if (!bpf_is_jmp_and_has_target(insn))
 157                         continue;
 158
 159                 /* Adjust offset of jmps if we cross boundaries. */
 160                 if (i < pos && i + insn->off + 1 > pos)
 161                         insn->off += delta;
 162                 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
 163                         insn->off -= delta;
 164         }
 165 }
 166
 167 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
 168                                        const struct bpf_insn *patch, u32 len)
 169 {
 170         u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
 171         struct bpf_prog *prog_adj;
 172
 173         /* Since our patchlet doesn't expand the image, we're done. */
 174         if (insn_delta == 0) {
 175                 memcpy(prog->insnsi + off, patch, sizeof(*patch));
 176                 return prog;
 177         }
 178
 179         insn_adj_cnt = prog->len + insn_delta;
 180
 181         /* Several new instructions need to be inserted. Make room
 182          * for them. Likely, there's no need for a new allocation as
 183          * last page could have large enough tailroom.
 184          */
 185         prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
 186                                     GFP_USER);
 187         if (!prog_adj)
 188                 return NULL;
 189
 190         prog_adj->len = insn_adj_cnt;
 191
 192         /* Patching happens in 3 steps:
 193          *
 194          * 1) Move over tail of insnsi from next instruction onwards,
 195          *    so we can patch the single target insn with one or more
 196          *    new ones (patching is always from 1 to n insns, n > 0).
 197          * 2) Inject new instructions at the target location.
 198          * 3) Adjust branch offsets if necessary.
 199          */
 200         insn_rest = insn_adj_cnt - off - len;
 201
 202         memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
 203                 sizeof(*patch) * insn_rest);
 204         memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
 205
 206         bpf_adj_branches(prog_adj, off, insn_delta);
 207
 208         return prog_adj;
 209 }
 210
 211 #ifdef CONFIG_BPF_JIT
 212 struct bpf_binary_header *
 213 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
 214                      unsigned int alignment,
 215                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
 216 {
 217         struct bpf_binary_header *hdr;
 218         unsigned int size, hole, start;
 219
 220         /* Most of BPF filters are really small, but if some of them
 221          * fill a page, allow at least 128 extra bytes to insert a
 222          * random section of illegal instructions.
 223          */
 224         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
 225         hdr = module_alloc(size);
 226         if (hdr == NULL)
 227                 return NULL;
 228
 229         /* Fill space with illegal/arch-dep instructions. */
 230         bpf_fill_ill_insns(hdr, size);
 231
 232         hdr->pages = size / PAGE_SIZE;
 233         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
 234                      PAGE_SIZE - sizeof(*hdr));
 235         start = (prandom_u32() % hole) & ~(alignment - 1);
 236
 237         /* Leave a random number of instructions before BPF code. */
 238         *image_ptr = &hdr->image[start];
 239
 240         return hdr;
 241 }
 242
 243 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
 244 {
 245         module_memfree(hdr);
 246 }
 247 #endif /* CONFIG_BPF_JIT */
 248
 249 /* Base function for offset calculation. Needs to go into .text section,
 250  * therefore keeping it non-static as well; will also be used by JITs
 251  * anyway later on, so do not let the compiler omit it.
 252  */
 253 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 254 {
 255         return 0;
 256 }
 257 EXPORT_SYMBOL_GPL(__bpf_call_base);
 258
 259 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
 260 /**
 261  *      __bpf_prog_run - run eBPF program on a given context
 262  *      @ctx: is the data we are operating on
 263  *      @insn: is the array of eBPF instructions
 264  *
 265  * Decode and execute eBPF instructions.
 266  */
 267 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
 268 {
 269         u64 stack[MAX_BPF_STACK / sizeof(u64)];
 270         u64 regs[MAX_BPF_REG], tmp;
 271         static const void *jumptable[256] = {
 272                 [0 ... 255] = &&default_label,
 273                 /* Now overwrite non-defaults ... */
 274                 /* 32 bit ALU operations */
 275                 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
 276                 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
 277                 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
 278                 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
 279                 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
 280                 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
 281                 [BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
 282                 [BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
 283                 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
 284                 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
 285                 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
 286                 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
 287                 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
 288                 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
 289                 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
 290                 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
 291                 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
 292                 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
 293                 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
 294                 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
 295                 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
 296                 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
 297                 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
 298                 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
 299                 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
 300                 /* 64 bit ALU operations */
 301                 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
 302                 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
 303                 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
 304                 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
 305                 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
 306                 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
 307                 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
 308                 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
 309                 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
 310                 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
 311                 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
 312                 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
 313                 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
 314                 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
 315                 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
 316                 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
 317                 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
 318                 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
 319                 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
 320                 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
 321                 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
 322                 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
 323                 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
 324                 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
 325                 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
 326                 /* Call instruction */
 327                 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
 328                 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
 329                 /* Jumps */
 330                 [BPF_JMP | BPF_JA] = &&JMP_JA,
 331                 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
 332                 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
 333                 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
 334                 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
 335                 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
 336                 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
 337                 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
 338                 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
 339                 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
 340                 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
 341                 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
 342                 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
 343                 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
 344                 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
 345                 /* Program return */
 346                 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
 347                 /* Store instructions */
 348                 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
 349                 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
 350                 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
 351                 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
 352                 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
 353                 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
 354                 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
 355                 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
 356                 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
 357                 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
 358                 /* Load instructions */
 359                 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
 360                 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
 361                 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
 362                 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
 363                 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
 364                 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
 365                 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
 366                 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
 367                 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
 368                 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
 369                 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
 370         };
 371         u32 tail_call_cnt = 0;
 372         void *ptr;
 373         int off;
 374
 375 #define CONT     ({ insn++; goto select_insn; })
 376 #define CONT_JMP ({ insn++; goto select_insn; })
 377
 378         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
 379         ARG1 = (u64) (unsigned long) ctx;
 380
 381         /* Registers used in classic BPF programs need to be reset first. */
 382         regs[BPF_REG_A] = 0;
 383         regs[BPF_REG_X] = 0;
 384
 385 select_insn:
 386         goto *jumptable[insn->code];
 387
 388         /* ALU */
 389 #define ALU(OPCODE, OP)                 \
 390         ALU64_##OPCODE##_X:             \
 391                 DST = DST OP SRC;       \
 392                 CONT;                   \
 393         ALU_##OPCODE##_X:               \
 394                 DST = (u32) DST OP (u32) SRC;   \
 395                 CONT;                   \
 396         ALU64_##OPCODE##_K:             \
 397                 DST = DST OP IMM;               \
 398                 CONT;                   \
 399         ALU_##OPCODE##_K:               \
 400                 DST = (u32) DST OP (u32) IMM;   \
 401                 CONT;
 402
 403         ALU(ADD,  +)
 404         ALU(SUB,  -)
 405         ALU(AND,  &)
 406         ALU(OR,   |)
 407         ALU(LSH, <<)
 408         ALU(RSH, >>)
 409         ALU(XOR,  ^)
 410         ALU(MUL,  *)
 411 #undef ALU
 412         ALU_NEG:
 413                 DST = (u32) -DST;
 414                 CONT;
 415         ALU64_NEG:
 416                 DST = -DST;
 417                 CONT;
 418         ALU_MOV_X:
 419                 DST = (u32) SRC;
 420                 CONT;
 421         ALU_MOV_K:
 422                 DST = (u32) IMM;
 423                 CONT;
 424         ALU64_MOV_X:
 425                 DST = SRC;
 426                 CONT;
 427         ALU64_MOV_K:
 428                 DST = IMM;
 429                 CONT;
 430         LD_IMM_DW:
 431                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
 432                 insn++;
 433                 CONT;
 434         ALU64_ARSH_X:
 435                 (*(s64 *) &DST) >>= SRC;
 436                 CONT;
 437         ALU64_ARSH_K:
 438                 (*(s64 *) &DST) >>= IMM;
 439                 CONT;
 440         ALU64_MOD_X:
 441                 if (unlikely(SRC == 0))
 442                         return 0;
 443                 div64_u64_rem(DST, SRC, &tmp);
 444                 DST = tmp;
 445                 CONT;
 446         ALU_MOD_X:
 447                 if (unlikely((u32)SRC == 0))
 448                         return 0;
 449                 tmp = (u32) DST;
 450                 DST = do_div(tmp, (u32) SRC);
 451                 CONT;
 452         ALU64_MOD_K:
 453                 div64_u64_rem(DST, IMM, &tmp);
 454                 DST = tmp;
 455                 CONT;
 456         ALU_MOD_K:
 457                 tmp = (u32) DST;
 458                 DST = do_div(tmp, (u32) IMM);
 459                 CONT;
 460         ALU64_DIV_X:
 461                 if (unlikely(SRC == 0))
 462                         return 0;
 463                 DST = div64_u64(DST, SRC);
 464                 CONT;
 465         ALU_DIV_X:
 466                 if (unlikely((u32)SRC == 0))
 467                         return 0;
 468                 tmp = (u32) DST;
 469                 do_div(tmp, (u32) SRC);
 470                 DST = (u32) tmp;
 471                 CONT;
 472         ALU64_DIV_K:
 473                 DST = div64_u64(DST, IMM);
 474                 CONT;
 475         ALU_DIV_K:
 476                 tmp = (u32) DST;
 477                 do_div(tmp, (u32) IMM);
 478                 DST = (u32) tmp;
 479                 CONT;
 480         ALU_END_TO_BE:
 481                 switch (IMM) {
 482                 case 16:
 483                         DST = (__force u16) cpu_to_be16(DST);
 484                         break;
 485                 case 32:
 486                         DST = (__force u32) cpu_to_be32(DST);
 487                         break;
 488                 case 64:
 489                         DST = (__force u64) cpu_to_be64(DST);
 490                         break;
 491                 }
 492                 CONT;
 493         ALU_END_TO_LE:
 494                 switch (IMM) {
 495                 case 16:
 496                         DST = (__force u16) cpu_to_le16(DST);
 497                         break;
 498                 case 32:
 499                         DST = (__force u32) cpu_to_le32(DST);
 500                         break;
 501                 case 64:
 502                         DST = (__force u64) cpu_to_le64(DST);
 503                         break;
 504                 }
 505                 CONT;
 506
 507         /* CALL */
 508         JMP_CALL:
 509                 /* Function call scratches BPF_R1-BPF_R5 registers,
 510                  * preserves BPF_R6-BPF_R9, and stores return value
 511                  * into BPF_R0.
 512                  */
 513                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
 514                                                        BPF_R4, BPF_R5);
 515                 CONT;
 516
 517         JMP_TAIL_CALL: {
 518                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
 519                 struct bpf_array *array = container_of(map, struct bpf_array, map);
 520                 struct bpf_prog *prog;
 521                 u32 index = BPF_R3;
 522
 523                 if (unlikely(index >= array->map.max_entries))
 524                         goto out;
 525
 526                 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
 527                         goto out;
 528
 529                 tail_call_cnt++;
 530
 531                 prog = READ_ONCE(array->ptrs[index]);
 532                 if (unlikely(!prog))
 533                         goto out;
 534
 535                 /* ARG1 at this point is guaranteed to point to CTX from
 536                  * the verifier side due to the fact that the tail call is
 537                  * handeled like a helper, that is, bpf_tail_call_proto,
 538                  * where arg1_type is ARG_PTR_TO_CTX.
 539                  */
 540                 insn = prog->insnsi;
 541                 goto select_insn;
 542 out:
 543                 CONT;
 544         }
 545         /* JMP */
 546         JMP_JA:
 547                 insn += insn->off;
 548                 CONT;
 549         JMP_JEQ_X:
 550                 if (DST == SRC) {
 551                         insn += insn->off;
 552                         CONT_JMP;
 553                 }
 554                 CONT;
 555         JMP_JEQ_K:
 556                 if (DST == IMM) {
 557                         insn += insn->off;
 558                         CONT_JMP;
 559                 }
 560                 CONT;
 561         JMP_JNE_X:
 562                 if (DST != SRC) {
 563                         insn += insn->off;
 564                         CONT_JMP;
 565                 }
 566                 CONT;
 567         JMP_JNE_K:
 568                 if (DST != IMM) {
 569                         insn += insn->off;
 570                         CONT_JMP;
 571                 }
 572                 CONT;
 573         JMP_JGT_X:
 574                 if (DST > SRC) {
 575                         insn += insn->off;
 576                         CONT_JMP;
 577                 }
 578                 CONT;
 579         JMP_JGT_K:
 580                 if (DST > IMM) {
 581                         insn += insn->off;
 582                         CONT_JMP;
 583                 }
 584                 CONT;
 585         JMP_JGE_X:
 586                 if (DST >= SRC) {
 587                         insn += insn->off;
 588                         CONT_JMP;
 589                 }
 590                 CONT;
 591         JMP_JGE_K:
 592                 if (DST >= IMM) {
 593                         insn += insn->off;
 594                         CONT_JMP;
 595                 }
 596                 CONT;
 597         JMP_JSGT_X:
 598                 if (((s64) DST) > ((s64) SRC)) {
 599                         insn += insn->off;
 600                         CONT_JMP;
 601                 }
 602                 CONT;
 603         JMP_JSGT_K:
 604                 if (((s64) DST) > ((s64) IMM)) {
 605                         insn += insn->off;
 606                         CONT_JMP;
 607                 }
 608                 CONT;
 609         JMP_JSGE_X:
 610                 if (((s64) DST) >= ((s64) SRC)) {
 611                         insn += insn->off;
 612                         CONT_JMP;
 613                 }
 614                 CONT;
 615         JMP_JSGE_K:
 616                 if (((s64) DST) >= ((s64) IMM)) {
 617                         insn += insn->off;
 618                         CONT_JMP;
 619                 }
 620                 CONT;
 621         JMP_JSET_X:
 622                 if (DST & SRC) {
 623                         insn += insn->off;
 624                         CONT_JMP;
 625                 }
 626                 CONT;
 627         JMP_JSET_K:
 628                 if (DST & IMM) {
 629                         insn += insn->off;
 630                         CONT_JMP;
 631                 }
 632                 CONT;
 633         JMP_EXIT:
 634                 return BPF_R0;
 635
 636         /* STX and ST and LDX*/
 637 #define LDST(SIZEOP, SIZE)                                              \
 638         STX_MEM_##SIZEOP:                                               \
 639                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
 640                 CONT;                                                   \
 641         ST_MEM_##SIZEOP:                                                \
 642                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
 643                 CONT;                                                   \
 644         LDX_MEM_##SIZEOP:                                               \
 645                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
 646                 CONT;
 647
 648         LDST(B,   u8)
 649         LDST(H,  u16)
 650         LDST(W,  u32)
 651         LDST(DW, u64)
 652 #undef LDST
 653         STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
 654                 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
 655                            (DST + insn->off));
 656                 CONT;
 657         STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
 658                 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
 659                              (DST + insn->off));
 660                 CONT;
 661         LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
 662                 off = IMM;
 663 load_word:
 664                 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
 665                  * only appearing in the programs where ctx ==
 666                  * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
 667                  * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
 668                  * internal BPF verifier will check that BPF_R6 ==
 669                  * ctx.
 670                  *
 671                  * BPF_ABS and BPF_IND are wrappers of function calls,
 672                  * so they scratch BPF_R1-BPF_R5 registers, preserve
 673                  * BPF_R6-BPF_R9, and store return value into BPF_R0.
 674                  *
 675                  * Implicit input:
 676                  *   ctx == skb == BPF_R6 == CTX
 677                  *
 678                  * Explicit input:
 679                  *   SRC == any register
 680                  *   IMM == 32-bit immediate
 681                  *
 682                  * Output:
 683                  *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
 684                  */
 685
 686                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
 687                 if (likely(ptr != NULL)) {
 688                         BPF_R0 = get_unaligned_be32(ptr);
 689                         CONT;
 690                 }
 691
 692                 return 0;
 693         LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
 694                 off = IMM;
 695 load_half:
 696                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
 697                 if (likely(ptr != NULL)) {
 698                         BPF_R0 = get_unaligned_be16(ptr);
 699                         CONT;
 700                 }
 701
 702                 return 0;
 703         LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
 704                 off = IMM;
 705 load_byte:
 706                 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
 707                 if (likely(ptr != NULL)) {
 708                         BPF_R0 = *(u8 *)ptr;
 709                         CONT;
 710                 }
 711
 712                 return 0;
 713         LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
 714                 off = IMM + SRC;
 715                 goto load_word;
 716         LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
 717                 off = IMM + SRC;
 718                 goto load_half;
 719         LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
 720                 off = IMM + SRC;
 721                 goto load_byte;
 722
 723         default_label:
 724                 /* If we ever reach this, we have a bug somewhere. */
 725                 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
 726                 return 0;
 727 }
 728
 729 #else
 730 static unsigned int __bpf_prog_ret0(void *ctx, const struct bpf_insn *insn)
 731 {
 732         return 0;
 733 }
 734 #endif
 735
 736 bool bpf_prog_array_compatible(struct bpf_array *array,
 737                                const struct bpf_prog *fp)
 738 {
 739         if (!array->owner_prog_type) {
 740                 /* There's no owner yet where we could check for
 741                  * compatibility.
 742                  */
 743                 array->owner_prog_type = fp->type;
 744                 array->owner_jited = fp->jited;
 745
 746                 return true;
 747         }
 748
 749         return array->owner_prog_type == fp->type &&
 750                array->owner_jited == fp->jited;
 751 }
 752
 753 static int bpf_check_tail_call(const struct bpf_prog *fp)
 754 {
 755         struct bpf_prog_aux *aux = fp->aux;
 756         int i;
 757
 758         for (i = 0; i < aux->used_map_cnt; i++) {
 759                 struct bpf_map *map = aux->used_maps[i];
 760                 struct bpf_array *array;
 761
 762                 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
 763                         continue;
 764
 765                 array = container_of(map, struct bpf_array, map);
 766                 if (!bpf_prog_array_compatible(array, fp))
 767                         return -EINVAL;
 768         }
 769
 770         return 0;
 771 }
 772
 773 /**
 774  *      bpf_prog_select_runtime - select exec runtime for BPF program
 775  *      @fp: bpf_prog populated with internal BPF program
 776  *
 777  * Try to JIT eBPF program, if JIT is not available, use interpreter.
 778  * The BPF program will be executed via BPF_PROG_RUN() macro.
 779  */
 780 int bpf_prog_select_runtime(struct bpf_prog *fp)
 781 {
 782 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
 783         fp->bpf_func = (void *) __bpf_prog_run;
 784 #else
 785         fp->bpf_func = (void *) __bpf_prog_ret0;
 786 #endif
 787
 788         /* eBPF JITs can rewrite the program in case constant
 789          * blinding is active. However, in case of error during
 790          * blinding, bpf_int_jit_compile() must always return a
 791          * valid program, which in this case would simply not
 792          * be JITed, but falls back to the interpreter.
 793          */
 794         bpf_int_jit_compile(fp);
 795 #ifdef CONFIG_BPF_JIT_ALWAYS_ON
 796         if (!fp->jited)
 797                 return -ENOTSUPP;
 798 #endif
 799         bpf_prog_lock_ro(fp);
 800
 801         /* The tail call compatibility check can only be done at
 802          * this late stage as we need to determine, if we deal
 803          * with JITed or non JITed program concatenations and not
 804          * all eBPF JITs might immediately support all features.
 805          */
 806         return bpf_check_tail_call(fp);
 807 }
 808 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
 809
 810 static void bpf_prog_free_deferred(struct work_struct *work)
 811 {
 812         struct bpf_prog_aux *aux;
 813
 814         aux = container_of(work, struct bpf_prog_aux, work);
 815         bpf_jit_free(aux->prog);
 816 }
 817
 818 /* Free internal BPF program */
 819 void bpf_prog_free(struct bpf_prog *fp)
 820 {
 821         struct bpf_prog_aux *aux = fp->aux;
 822
 823         INIT_WORK(&aux->work, bpf_prog_free_deferred);
 824         schedule_work(&aux->work);
 825 }
 826 EXPORT_SYMBOL_GPL(bpf_prog_free);
 827
 828 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
 829 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
 830
 831 void bpf_user_rnd_init_once(void)
 832 {
 833         prandom_init_once(&bpf_user_rnd_state);
 834 }
 835
 836 u64 bpf_user_rnd_u32(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
 837 {
 838         /* Should someone ever have the rather unwise idea to use some
 839          * of the registers passed into this function, then note that
 840          * this function is called from native eBPF and classic-to-eBPF
 841          * transformations. Register assignments from both sides are
 842          * different, f.e. classic always sets fn(ctx, A, X) here.
 843          */
 844         struct rnd_state *state;
 845         u32 res;
 846
 847         state = &get_cpu_var(bpf_user_rnd_state);
 848         res = prandom_u32_state(state);
 849         put_cpu_var(state);
 850
 851         return res;
 852 }
 853
 854 /* Weak definitions of helper functions in case we don't have bpf syscall. */
 855 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
 856 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
 857 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
 858
 859 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
 860 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
 861 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
 862 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
 863 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
 864 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
 865 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
 866 {
 867         return NULL;
 868 }
 869
 870 /* Always built-in helper functions. */
 871 const struct bpf_func_proto bpf_tail_call_proto = {
 872         .func           = NULL,
 873         .gpl_only       = false,
 874         .ret_type       = RET_VOID,
 875         .arg1_type      = ARG_PTR_TO_CTX,
 876         .arg2_type      = ARG_CONST_MAP_PTR,
 877         .arg3_type      = ARG_ANYTHING,
 878 };
 879
 880 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
 881 void __weak bpf_int_jit_compile(struct bpf_prog *prog)
 882 {
 883 }
 884
 885 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
 886  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
 887  */
 888 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
 889                          int len)
 890 {
 891         return -EFAULT;
 892 }