arch/arm/net/bpf_jit_32.c

   1 /*
   2  * Just-In-Time compiler for BPF filters on 32bit ARM
   3  *
   4  * Copyright (c) 2011 Mircea Gherzan <mgherzan@gmail.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License as published by the
   8  * Free Software Foundation; version 2 of the License.
   9  */
  10
  11 #include <linux/bitops.h>
  12 #include <linux/compiler.h>
  13 #include <linux/errno.h>
  14 #include <linux/filter.h>
  15 #include <linux/moduleloader.h>
  16 #include <linux/netdevice.h>
  17 #include <linux/string.h>
  18 #include <linux/slab.h>
  19 #include <linux/if_vlan.h>
  20 #include <asm/cacheflush.h>
  21 #include <asm/hwcap.h>
  22
  23 #include "bpf_jit_32.h"
  24
  25 /*
  26  * ABI:
  27  *
  28  * r0   scratch register
  29  * r4   BPF register A
  30  * r5   BPF register X
  31  * r6   pointer to the skb
  32  * r7   skb->data
  33  * r8   skb_headlen(skb)
  34  */
  35
  36 #define r_scratch       ARM_R0
  37 /* r1-r3 are (also) used for the unaligned loads on the non-ARMv7 slowpath */
  38 #define r_off           ARM_R1
  39 #define r_A             ARM_R4
  40 #define r_X             ARM_R5
  41 #define r_skb           ARM_R6
  42 #define r_skb_data      ARM_R7
  43 #define r_skb_hl        ARM_R8
  44
  45 #define SCRATCH_SP_OFFSET       0
  46 #define SCRATCH_OFF(k)          (SCRATCH_SP_OFFSET + 4 * (k))
  47
  48 #define SEEN_MEM                ((1 << BPF_MEMWORDS) - 1)
  49 #define SEEN_MEM_WORD(k)        (1 << (k))
  50 #define SEEN_X                  (1 << BPF_MEMWORDS)
  51 #define SEEN_CALL               (1 << (BPF_MEMWORDS + 1))
  52 #define SEEN_SKB                (1 << (BPF_MEMWORDS + 2))
  53 #define SEEN_DATA               (1 << (BPF_MEMWORDS + 3))
  54
  55 #define FLAG_NEED_X_RESET       (1 << 0)
  56
  57 struct jit_ctx {
  58         const struct sk_filter *skf;
  59         unsigned idx;
  60         unsigned prologue_bytes;
  61         int ret0_fp_idx;
  62         u32 seen;
  63         u32 flags;
  64         u32 *offsets;
  65         u32 *target;
  66 #if __LINUX_ARM_ARCH__ < 7
  67         u16 epilogue_bytes;
  68         u16 imm_count;
  69         u32 *imms;
  70 #endif
  71 };
  72
  73 int bpf_jit_enable __read_mostly;
  74
  75 static u64 jit_get_skb_b(struct sk_buff *skb, unsigned offset)
  76 {
  77         u8 ret;
  78         int err;
  79
  80         err = skb_copy_bits(skb, offset, &ret, 1);
  81
  82         return (u64)err << 32 | ret;
  83 }
  84
  85 static u64 jit_get_skb_h(struct sk_buff *skb, unsigned offset)
  86 {
  87         u16 ret;
  88         int err;
  89
  90         err = skb_copy_bits(skb, offset, &ret, 2);
  91
  92         return (u64)err << 32 | ntohs(ret);
  93 }
  94
  95 static u64 jit_get_skb_w(struct sk_buff *skb, unsigned offset)
  96 {
  97         u32 ret;
  98         int err;
  99
 100         err = skb_copy_bits(skb, offset, &ret, 4);
 101
 102         return (u64)err << 32 | ntohl(ret);
 103 }
 104
 105 /*
 106  * Wrapper that handles both OABI and EABI and assures Thumb2 interworking
 107  * (where the assembly routines like __aeabi_uidiv could cause problems).
 108  */
 109 static u32 jit_udiv(u32 dividend, u32 divisor)
 110 {
 111         return dividend / divisor;
 112 }
 113
 114 static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx)
 115 {
 116         if (ctx->target != NULL)
 117                 ctx->target[ctx->idx] = inst | (cond << 28);
 118
 119         ctx->idx++;
 120 }
 121
 122 /*
 123  * Emit an instruction that will be executed unconditionally.
 124  */
 125 static inline void emit(u32 inst, struct jit_ctx *ctx)
 126 {
 127         _emit(ARM_COND_AL, inst, ctx);
 128 }
 129
 130 static u16 saved_regs(struct jit_ctx *ctx)
 131 {
 132         u16 ret = 0;
 133
 134         if ((ctx->skf->len > 1) ||
 135             (ctx->skf->insns[0].code == BPF_S_RET_A))
 136                 ret |= 1 << r_A;
 137
 138 #ifdef CONFIG_FRAME_POINTER
 139         ret |= (1 << ARM_FP) | (1 << ARM_IP) | (1 << ARM_LR) | (1 << ARM_PC);
 140 #else
 141         if (ctx->seen & SEEN_CALL)
 142                 ret |= 1 << ARM_LR;
 143 #endif
 144         if (ctx->seen & (SEEN_DATA | SEEN_SKB))
 145                 ret |= 1 << r_skb;
 146         if (ctx->seen & SEEN_DATA)
 147                 ret |= (1 << r_skb_data) | (1 << r_skb_hl);
 148         if (ctx->seen & SEEN_X)
 149                 ret |= 1 << r_X;
 150
 151         return ret;
 152 }
 153
 154 static inline int mem_words_used(struct jit_ctx *ctx)
 155 {
 156         /* yes, we do waste some stack space IF there are "holes" in the set" */
 157         return fls(ctx->seen & SEEN_MEM);
 158 }
 159
 160 static inline bool is_load_to_a(u16 inst)
 161 {
 162         switch (inst) {
 163         case BPF_S_LD_W_LEN:
 164         case BPF_S_LD_W_ABS:
 165         case BPF_S_LD_H_ABS:
 166         case BPF_S_LD_B_ABS:
 167         case BPF_S_ANC_CPU:
 168         case BPF_S_ANC_IFINDEX:
 169         case BPF_S_ANC_MARK:
 170         case BPF_S_ANC_PROTOCOL:
 171         case BPF_S_ANC_RXHASH:
 172         case BPF_S_ANC_VLAN_TAG:
 173         case BPF_S_ANC_VLAN_TAG_PRESENT:
 174         case BPF_S_ANC_QUEUE:
 175                 return true;
 176         default:
 177                 return false;
 178         }
 179 }
 180
 181 static void build_prologue(struct jit_ctx *ctx)
 182 {
 183         u16 reg_set = saved_regs(ctx);
 184         u16 first_inst = ctx->skf->insns[0].code;
 185         u16 off;
 186
 187 #ifdef CONFIG_FRAME_POINTER
 188         emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
 189         emit(ARM_PUSH(reg_set), ctx);
 190         emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
 191 #else
 192         if (reg_set)
 193                 emit(ARM_PUSH(reg_set), ctx);
 194 #endif
 195
 196         if (ctx->seen & (SEEN_DATA | SEEN_SKB))
 197                 emit(ARM_MOV_R(r_skb, ARM_R0), ctx);
 198
 199         if (ctx->seen & SEEN_DATA) {
 200                 off = offsetof(struct sk_buff, data);
 201                 emit(ARM_LDR_I(r_skb_data, r_skb, off), ctx);
 202                 /* headlen = len - data_len */
 203                 off = offsetof(struct sk_buff, len);
 204                 emit(ARM_LDR_I(r_skb_hl, r_skb, off), ctx);
 205                 off = offsetof(struct sk_buff, data_len);
 206                 emit(ARM_LDR_I(r_scratch, r_skb, off), ctx);
 207                 emit(ARM_SUB_R(r_skb_hl, r_skb_hl, r_scratch), ctx);
 208         }
 209
 210         if (ctx->flags & FLAG_NEED_X_RESET)
 211                 emit(ARM_MOV_I(r_X, 0), ctx);
 212
 213         /* do not leak kernel data to userspace */
 214         if ((first_inst != BPF_S_RET_K) && !(is_load_to_a(first_inst)))
 215                 emit(ARM_MOV_I(r_A, 0), ctx);
 216
 217         /* stack space for the BPF_MEM words */
 218         if (ctx->seen & SEEN_MEM)
 219                 emit(ARM_SUB_I(ARM_SP, ARM_SP, mem_words_used(ctx) * 4), ctx);
 220 }
 221
 222 static void build_epilogue(struct jit_ctx *ctx)
 223 {
 224         u16 reg_set = saved_regs(ctx);
 225
 226         if (ctx->seen & SEEN_MEM)
 227                 emit(ARM_ADD_I(ARM_SP, ARM_SP, mem_words_used(ctx) * 4), ctx);
 228
 229         reg_set &= ~(1 << ARM_LR);
 230
 231 #ifdef CONFIG_FRAME_POINTER
 232         /* the first instruction of the prologue was: mov ip, sp */
 233         reg_set &= ~(1 << ARM_IP);
 234         reg_set |= (1 << ARM_SP);
 235         emit(ARM_LDM(ARM_SP, reg_set), ctx);
 236 #else
 237         if (reg_set) {
 238                 if (ctx->seen & SEEN_CALL)
 239                         reg_set |= 1 << ARM_PC;
 240                 emit(ARM_POP(reg_set), ctx);
 241         }
 242
 243         if (!(ctx->seen & SEEN_CALL))
 244                 emit(ARM_BX(ARM_LR), ctx);
 245 #endif
 246 }
 247
 248 static int16_t imm8m(u32 x)
 249 {
 250         u32 rot;
 251
 252         for (rot = 0; rot < 16; rot++)
 253                 if ((x & ~ror32(0xff, 2 * rot)) == 0)
 254                         return rol32(x, 2 * rot) | (rot << 8);
 255
 256         return -1;
 257 }
 258
 259 #if __LINUX_ARM_ARCH__ < 7
 260
 261 static u16 imm_offset(u32 k, struct jit_ctx *ctx)
 262 {
 263         unsigned i = 0, offset;
 264         u16 imm;
 265
 266         /* on the "fake" run we just count them (duplicates included) */
 267         if (ctx->target == NULL) {
 268                 ctx->imm_count++;
 269                 return 0;
 270         }
 271
 272         while ((i < ctx->imm_count) && ctx->imms[i]) {
 273                 if (ctx->imms[i] == k)
 274                         break;
 275                 i++;
 276         }
 277
 278         if (ctx->imms[i] == 0)
 279                 ctx->imms[i] = k;
 280
 281         /* constants go just after the epilogue */
 282         offset =  ctx->offsets[ctx->skf->len];
 283         offset += ctx->prologue_bytes;
 284         offset += ctx->epilogue_bytes;
 285         offset += i * 4;
 286
 287         ctx->target[offset / 4] = k;
 288
 289         /* PC in ARM mode == address of the instruction + 8 */
 290         imm = offset - (8 + ctx->idx * 4);
 291
 292         return imm;
 293 }
 294
 295 #endif /* __LINUX_ARM_ARCH__ */
 296
 297 /*
 298  * Move an immediate that's not an imm8m to a core register.
 299  */
 300 static inline void emit_mov_i_no8m(int rd, u32 val, struct jit_ctx *ctx)
 301 {
 302 #if __LINUX_ARM_ARCH__ < 7
 303         emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx);
 304 #else
 305         emit(ARM_MOVW(rd, val & 0xffff), ctx);
 306         if (val > 0xffff)
 307                 emit(ARM_MOVT(rd, val >> 16), ctx);
 308 #endif
 309 }
 310
 311 static inline void emit_mov_i(int rd, u32 val, struct jit_ctx *ctx)
 312 {
 313         int imm12 = imm8m(val);
 314
 315         if (imm12 >= 0)
 316                 emit(ARM_MOV_I(rd, imm12), ctx);
 317         else
 318                 emit_mov_i_no8m(rd, val, ctx);
 319 }
 320
 321 #if __LINUX_ARM_ARCH__ < 6
 322
 323 static void emit_load_be32(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
 324 {
 325         _emit(cond, ARM_LDRB_I(ARM_R3, r_addr, 1), ctx);
 326         _emit(cond, ARM_LDRB_I(ARM_R1, r_addr, 0), ctx);
 327         _emit(cond, ARM_LDRB_I(ARM_R2, r_addr, 3), ctx);
 328         _emit(cond, ARM_LSL_I(ARM_R3, ARM_R3, 16), ctx);
 329         _emit(cond, ARM_LDRB_I(ARM_R0, r_addr, 2), ctx);
 330         _emit(cond, ARM_ORR_S(ARM_R3, ARM_R3, ARM_R1, SRTYPE_LSL, 24), ctx);
 331         _emit(cond, ARM_ORR_R(ARM_R3, ARM_R3, ARM_R2), ctx);
 332         _emit(cond, ARM_ORR_S(r_res, ARM_R3, ARM_R0, SRTYPE_LSL, 8), ctx);
 333 }
 334
 335 static void emit_load_be16(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
 336 {
 337         _emit(cond, ARM_LDRB_I(ARM_R1, r_addr, 0), ctx);
 338         _emit(cond, ARM_LDRB_I(ARM_R2, r_addr, 1), ctx);
 339         _emit(cond, ARM_ORR_S(r_res, ARM_R2, ARM_R1, SRTYPE_LSL, 8), ctx);
 340 }
 341
 342 static inline void emit_swap16(u8 r_dst, u8 r_src, struct jit_ctx *ctx)
 343 {
 344         /* r_dst = (r_src << 8) | (r_src >> 8) */
 345         emit(ARM_LSL_I(ARM_R1, r_src, 8), ctx);
 346         emit(ARM_ORR_S(r_dst, ARM_R1, r_src, SRTYPE_LSR, 8), ctx);
 347
 348         /*
 349          * we need to mask out the bits set in r_dst[23:16] due to
 350          * the first shift instruction.
 351          *
 352          * note that 0x8ff is the encoded immediate 0x00ff0000.
 353          */
 354         emit(ARM_BIC_I(r_dst, r_dst, 0x8ff), ctx);
 355 }
 356
 357 #else  /* ARMv6+ */
 358
 359 static void emit_load_be32(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
 360 {
 361         _emit(cond, ARM_LDR_I(r_res, r_addr, 0), ctx);
 362 #ifdef __LITTLE_ENDIAN
 363         _emit(cond, ARM_REV(r_res, r_res), ctx);
 364 #endif
 365 }
 366
 367 static void emit_load_be16(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
 368 {
 369         _emit(cond, ARM_LDRH_I(r_res, r_addr, 0), ctx);
 370 #ifdef __LITTLE_ENDIAN
 371         _emit(cond, ARM_REV16(r_res, r_res), ctx);
 372 #endif
 373 }
 374
 375 static inline void emit_swap16(u8 r_dst __maybe_unused,
 376                                u8 r_src __maybe_unused,
 377                                struct jit_ctx *ctx __maybe_unused)
 378 {
 379 #ifdef __LITTLE_ENDIAN
 380         emit(ARM_REV16(r_dst, r_src), ctx);
 381 #endif
 382 }
 383
 384 #endif /* __LINUX_ARM_ARCH__ < 6 */
 385
 386
 387 /* Compute the immediate value for a PC-relative branch. */
 388 static inline u32 b_imm(unsigned tgt, struct jit_ctx *ctx)
 389 {
 390         u32 imm;
 391
 392         if (ctx->target == NULL)
 393                 return 0;
 394         /*
 395          * BPF allows only forward jumps and the offset of the target is
 396          * still the one computed during the first pass.
 397          */
 398         imm  = ctx->offsets[tgt] + ctx->prologue_bytes - (ctx->idx * 4 + 8);
 399
 400         return imm >> 2;
 401 }
 402
 403 #define OP_IMM3(op, r1, r2, imm_val, ctx)                               \
 404         do {                                                            \
 405                 imm12 = imm8m(imm_val);                                 \
 406                 if (imm12 < 0) {                                        \
 407                         emit_mov_i_no8m(r_scratch, imm_val, ctx);       \
 408                         emit(op ## _R((r1), (r2), r_scratch), ctx);     \
 409                 } else {                                                \
 410                         emit(op ## _I((r1), (r2), imm12), ctx);         \
 411                 }                                                       \
 412         } while (0)
 413
 414 static inline void emit_err_ret(u8 cond, struct jit_ctx *ctx)
 415 {
 416         if (ctx->ret0_fp_idx >= 0) {
 417                 _emit(cond, ARM_B(b_imm(ctx->ret0_fp_idx, ctx)), ctx);
 418                 /* NOP to keep the size constant between passes */
 419                 emit(ARM_MOV_R(ARM_R0, ARM_R0), ctx);
 420         } else {
 421                 _emit(cond, ARM_MOV_I(ARM_R0, 0), ctx);
 422                 _emit(cond, ARM_B(b_imm(ctx->skf->len, ctx)), ctx);
 423         }
 424 }
 425
 426 static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx)
 427 {
 428 #if __LINUX_ARM_ARCH__ < 5
 429         emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx);
 430
 431         if (elf_hwcap & HWCAP_THUMB)
 432                 emit(ARM_BX(tgt_reg), ctx);
 433         else
 434                 emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx);
 435 #else
 436         emit(ARM_BLX_R(tgt_reg), ctx);
 437 #endif
 438 }
 439
 440 static inline void emit_udiv(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx)
 441 {
 442 #if __LINUX_ARM_ARCH__ == 7
 443         if (elf_hwcap & HWCAP_IDIVA) {
 444                 emit(ARM_UDIV(rd, rm, rn), ctx);
 445                 return;
 446         }
 447 #endif
 448         if (rm != ARM_R0)
 449                 emit(ARM_MOV_R(ARM_R0, rm), ctx);
 450         if (rn != ARM_R1)
 451                 emit(ARM_MOV_R(ARM_R1, rn), ctx);
 452
 453         ctx->seen |= SEEN_CALL;
 454         emit_mov_i(ARM_R3, (u32)jit_udiv, ctx);
 455         emit_blx_r(ARM_R3, ctx);
 456
 457         if (rd != ARM_R0)
 458                 emit(ARM_MOV_R(rd, ARM_R0), ctx);
 459 }
 460
 461 static inline void update_on_xread(struct jit_ctx *ctx)
 462 {
 463         if (!(ctx->seen & SEEN_X))
 464                 ctx->flags |= FLAG_NEED_X_RESET;
 465
 466         ctx->seen |= SEEN_X;
 467 }
 468
 469 static int build_body(struct jit_ctx *ctx)
 470 {
 471         void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w};
 472         const struct sk_filter *prog = ctx->skf;
 473         const struct sock_filter *inst;
 474         unsigned i, load_order, off, condt;
 475         int imm12;
 476         u32 k;
 477
 478         for (i = 0; i < prog->len; i++) {
 479                 inst = &(prog->insns[i]);
 480                 /* K as an immediate value operand */
 481                 k = inst->k;
 482
 483                 /* compute offsets only in the fake pass */
 484                 if (ctx->target == NULL)
 485                         ctx->offsets[i] = ctx->idx * 4;
 486
 487                 switch (inst->code) {
 488                 case BPF_S_LD_IMM:
 489                         emit_mov_i(r_A, k, ctx);
 490                         break;
 491                 case BPF_S_LD_W_LEN:
 492                         ctx->seen |= SEEN_SKB;
 493                         BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
 494                         emit(ARM_LDR_I(r_A, r_skb,
 495                                        offsetof(struct sk_buff, len)), ctx);
 496                         break;
 497                 case BPF_S_LD_MEM:
 498                         /* A = scratch[k] */
 499                         ctx->seen |= SEEN_MEM_WORD(k);
 500                         emit(ARM_LDR_I(r_A, ARM_SP, SCRATCH_OFF(k)), ctx);
 501                         break;
 502                 case BPF_S_LD_W_ABS:
 503                         load_order = 2;
 504                         goto load;
 505                 case BPF_S_LD_H_ABS:
 506                         load_order = 1;
 507                         goto load;
 508                 case BPF_S_LD_B_ABS:
 509                         load_order = 0;
 510 load:
 511                         /* the interpreter will deal with the negative K */
 512                         if ((int)k < 0)
 513                                 return -ENOTSUPP;
 514                         emit_mov_i(r_off, k, ctx);
 515 load_common:
 516                         ctx->seen |= SEEN_DATA | SEEN_CALL;
 517
 518                         if (load_order > 0) {
 519                                 emit(ARM_SUB_I(r_scratch, r_skb_hl,
 520                                                1 << load_order), ctx);
 521                                 emit(ARM_CMP_R(r_scratch, r_off), ctx);
 522                                 condt = ARM_COND_HS;
 523                         } else {
 524                                 emit(ARM_CMP_R(r_skb_hl, r_off), ctx);
 525                                 condt = ARM_COND_HI;
 526                         }
 527
 528                         _emit(condt, ARM_ADD_R(r_scratch, r_off, r_skb_data),
 529                               ctx);
 530
 531                         if (load_order == 0)
 532                                 _emit(condt, ARM_LDRB_I(r_A, r_scratch, 0),
 533                                       ctx);
 534                         else if (load_order == 1)
 535                                 emit_load_be16(condt, r_A, r_scratch, ctx);
 536                         else if (load_order == 2)
 537                                 emit_load_be32(condt, r_A, r_scratch, ctx);
 538
 539                         _emit(condt, ARM_B(b_imm(i + 1, ctx)), ctx);
 540
 541                         /* the slowpath */
 542                         emit_mov_i(ARM_R3, (u32)load_func[load_order], ctx);
 543                         emit(ARM_MOV_R(ARM_R0, r_skb), ctx);
 544                         /* the offset is already in R1 */
 545                         emit_blx_r(ARM_R3, ctx);
 546                         /* check the result of skb_copy_bits */
 547                         emit(ARM_CMP_I(ARM_R1, 0), ctx);
 548                         emit_err_ret(ARM_COND_NE, ctx);
 549                         emit(ARM_MOV_R(r_A, ARM_R0), ctx);
 550                         break;
 551                 case BPF_S_LD_W_IND:
 552                         load_order = 2;
 553                         goto load_ind;
 554                 case BPF_S_LD_H_IND:
 555                         load_order = 1;
 556                         goto load_ind;
 557                 case BPF_S_LD_B_IND:
 558                         load_order = 0;
 559 load_ind:
 560                         OP_IMM3(ARM_ADD, r_off, r_X, k, ctx);
 561                         goto load_common;
 562                 case BPF_S_LDX_IMM:
 563                         ctx->seen |= SEEN_X;
 564                         emit_mov_i(r_X, k, ctx);
 565                         break;
 566                 case BPF_S_LDX_W_LEN:
 567                         ctx->seen |= SEEN_X | SEEN_SKB;
 568                         emit(ARM_LDR_I(r_X, r_skb,
 569                                        offsetof(struct sk_buff, len)), ctx);
 570                         break;
 571                 case BPF_S_LDX_MEM:
 572                         ctx->seen |= SEEN_X | SEEN_MEM_WORD(k);
 573                         emit(ARM_LDR_I(r_X, ARM_SP, SCRATCH_OFF(k)), ctx);
 574                         break;
 575                 case BPF_S_LDX_B_MSH:
 576                         /* x = ((*(frame + k)) & 0xf) << 2; */
 577                         ctx->seen |= SEEN_X | SEEN_DATA | SEEN_CALL;
 578                         /* the interpreter should deal with the negative K */
 579                         if (k < 0)
 580                                 return -1;
 581                         /* offset in r1: we might have to take the slow path */
 582                         emit_mov_i(r_off, k, ctx);
 583                         emit(ARM_CMP_R(r_skb_hl, r_off), ctx);
 584
 585                         /* load in r0: common with the slowpath */
 586                         _emit(ARM_COND_HI, ARM_LDRB_R(ARM_R0, r_skb_data,
 587                                                       ARM_R1), ctx);
 588                         /*
 589                          * emit_mov_i() might generate one or two instructions,
 590                          * the same holds for emit_blx_r()
 591                          */
 592                         _emit(ARM_COND_HI, ARM_B(b_imm(i + 1, ctx) - 2), ctx);
 593
 594                         emit(ARM_MOV_R(ARM_R0, r_skb), ctx);
 595                         /* r_off is r1 */
 596                         emit_mov_i(ARM_R3, (u32)jit_get_skb_b, ctx);
 597                         emit_blx_r(ARM_R3, ctx);
 598                         /* check the return value of skb_copy_bits */
 599                         emit(ARM_CMP_I(ARM_R1, 0), ctx);
 600                         emit_err_ret(ARM_COND_NE, ctx);
 601
 602                         emit(ARM_AND_I(r_X, ARM_R0, 0x00f), ctx);
 603                         emit(ARM_LSL_I(r_X, r_X, 2), ctx);
 604                         break;
 605                 case BPF_S_ST:
 606                         ctx->seen |= SEEN_MEM_WORD(k);
 607                         emit(ARM_STR_I(r_A, ARM_SP, SCRATCH_OFF(k)), ctx);
 608                         break;
 609                 case BPF_S_STX:
 610                         update_on_xread(ctx);
 611                         ctx->seen |= SEEN_MEM_WORD(k);
 612                         emit(ARM_STR_I(r_X, ARM_SP, SCRATCH_OFF(k)), ctx);
 613                         break;
 614                 case BPF_S_ALU_ADD_K:
 615                         /* A += K */
 616                         OP_IMM3(ARM_ADD, r_A, r_A, k, ctx);
 617                         break;
 618                 case BPF_S_ALU_ADD_X:
 619                         update_on_xread(ctx);
 620                         emit(ARM_ADD_R(r_A, r_A, r_X), ctx);
 621                         break;
 622                 case BPF_S_ALU_SUB_K:
 623                         /* A -= K */
 624                         OP_IMM3(ARM_SUB, r_A, r_A, k, ctx);
 625                         break;
 626                 case BPF_S_ALU_SUB_X:
 627                         update_on_xread(ctx);
 628                         emit(ARM_SUB_R(r_A, r_A, r_X), ctx);
 629                         break;
 630                 case BPF_S_ALU_MUL_K:
 631                         /* A *= K */
 632                         emit_mov_i(r_scratch, k, ctx);
 633                         emit(ARM_MUL(r_A, r_A, r_scratch), ctx);
 634                         break;
 635                 case BPF_S_ALU_MUL_X:
 636                         update_on_xread(ctx);
 637                         emit(ARM_MUL(r_A, r_A, r_X), ctx);
 638                         break;
 639                 case BPF_S_ALU_DIV_K:
 640                         /* current k == reciprocal_value(userspace k) */
 641                         emit_mov_i(r_scratch, k, ctx);
 642                         /* A = top 32 bits of the product */
 643                         emit(ARM_UMULL(r_scratch, r_A, r_A, r_scratch), ctx);
 644                         break;
 645                 case BPF_S_ALU_DIV_X:
 646                         update_on_xread(ctx);
 647                         emit(ARM_CMP_I(r_X, 0), ctx);
 648                         emit_err_ret(ARM_COND_EQ, ctx);
 649                         emit_udiv(r_A, r_A, r_X, ctx);
 650                         break;
 651                 case BPF_S_ALU_OR_K:
 652                         /* A |= K */
 653                         OP_IMM3(ARM_ORR, r_A, r_A, k, ctx);
 654                         break;
 655                 case BPF_S_ALU_OR_X:
 656                         update_on_xread(ctx);
 657                         emit(ARM_ORR_R(r_A, r_A, r_X), ctx);
 658                         break;
 659                 case BPF_S_ALU_XOR_K:
 660                         /* A ^= K; */
 661                         OP_IMM3(ARM_EOR, r_A, r_A, k, ctx);
 662                         break;
 663                 case BPF_S_ANC_ALU_XOR_X:
 664                 case BPF_S_ALU_XOR_X:
 665                         /* A ^= X */
 666                         update_on_xread(ctx);
 667                         emit(ARM_EOR_R(r_A, r_A, r_X), ctx);
 668                         break;
 669                 case BPF_S_ALU_AND_K:
 670                         /* A &= K */
 671                         OP_IMM3(ARM_AND, r_A, r_A, k, ctx);
 672                         break;
 673                 case BPF_S_ALU_AND_X:
 674                         update_on_xread(ctx);
 675                         emit(ARM_AND_R(r_A, r_A, r_X), ctx);
 676                         break;
 677                 case BPF_S_ALU_LSH_K:
 678                         if (unlikely(k > 31))
 679                                 return -1;
 680                         emit(ARM_LSL_I(r_A, r_A, k), ctx);
 681                         break;
 682                 case BPF_S_ALU_LSH_X:
 683                         update_on_xread(ctx);
 684                         emit(ARM_LSL_R(r_A, r_A, r_X), ctx);
 685                         break;
 686                 case BPF_S_ALU_RSH_K:
 687                         if (unlikely(k > 31))
 688                                 return -1;
 689                         emit(ARM_LSR_I(r_A, r_A, k), ctx);
 690                         break;
 691                 case BPF_S_ALU_RSH_X:
 692                         update_on_xread(ctx);
 693                         emit(ARM_LSR_R(r_A, r_A, r_X), ctx);
 694                         break;
 695                 case BPF_S_ALU_NEG:
 696                         /* A = -A */
 697                         emit(ARM_RSB_I(r_A, r_A, 0), ctx);
 698                         break;
 699                 case BPF_S_JMP_JA:
 700                         /* pc += K */
 701                         emit(ARM_B(b_imm(i + k + 1, ctx)), ctx);
 702                         break;
 703                 case BPF_S_JMP_JEQ_K:
 704                         /* pc += (A == K) ? pc->jt : pc->jf */
 705                         condt  = ARM_COND_EQ;
 706                         goto cmp_imm;
 707                 case BPF_S_JMP_JGT_K:
 708                         /* pc += (A > K) ? pc->jt : pc->jf */
 709                         condt  = ARM_COND_HI;
 710                         goto cmp_imm;
 711                 case BPF_S_JMP_JGE_K:
 712                         /* pc += (A >= K) ? pc->jt : pc->jf */
 713                         condt  = ARM_COND_HS;
 714 cmp_imm:
 715                         imm12 = imm8m(k);
 716                         if (imm12 < 0) {
 717                                 emit_mov_i_no8m(r_scratch, k, ctx);
 718                                 emit(ARM_CMP_R(r_A, r_scratch), ctx);
 719                         } else {
 720                                 emit(ARM_CMP_I(r_A, imm12), ctx);
 721                         }
 722 cond_jump:
 723                         if (inst->jt)
 724                                 _emit(condt, ARM_B(b_imm(i + inst->jt + 1,
 725                                                    ctx)), ctx);
 726                         if (inst->jf)
 727                                 _emit(condt ^ 1, ARM_B(b_imm(i + inst->jf + 1,
 728                                                              ctx)), ctx);
 729                         break;
 730                 case BPF_S_JMP_JEQ_X:
 731                         /* pc += (A == X) ? pc->jt : pc->jf */
 732                         condt   = ARM_COND_EQ;
 733                         goto cmp_x;
 734                 case BPF_S_JMP_JGT_X:
 735                         /* pc += (A > X) ? pc->jt : pc->jf */
 736                         condt   = ARM_COND_HI;
 737                         goto cmp_x;
 738                 case BPF_S_JMP_JGE_X:
 739                         /* pc += (A >= X) ? pc->jt : pc->jf */
 740                         condt   = ARM_COND_CS;
 741 cmp_x:
 742                         update_on_xread(ctx);
 743                         emit(ARM_CMP_R(r_A, r_X), ctx);
 744                         goto cond_jump;
 745                 case BPF_S_JMP_JSET_K:
 746                         /* pc += (A & K) ? pc->jt : pc->jf */
 747                         condt  = ARM_COND_NE;
 748                         /* not set iff all zeroes iff Z==1 iff EQ */
 749
 750                         imm12 = imm8m(k);
 751                         if (imm12 < 0) {
 752                                 emit_mov_i_no8m(r_scratch, k, ctx);
 753                                 emit(ARM_TST_R(r_A, r_scratch), ctx);
 754                         } else {
 755                                 emit(ARM_TST_I(r_A, imm12), ctx);
 756                         }
 757                         goto cond_jump;
 758                 case BPF_S_JMP_JSET_X:
 759                         /* pc += (A & X) ? pc->jt : pc->jf */
 760                         update_on_xread(ctx);
 761                         condt  = ARM_COND_NE;
 762                         emit(ARM_TST_R(r_A, r_X), ctx);
 763                         goto cond_jump;
 764                 case BPF_S_RET_A:
 765                         emit(ARM_MOV_R(ARM_R0, r_A), ctx);
 766                         goto b_epilogue;
 767                 case BPF_S_RET_K:
 768                         if ((k == 0) && (ctx->ret0_fp_idx < 0))
 769                                 ctx->ret0_fp_idx = i;
 770                         emit_mov_i(ARM_R0, k, ctx);
 771 b_epilogue:
 772                         if (i != ctx->skf->len - 1)
 773                                 emit(ARM_B(b_imm(prog->len, ctx)), ctx);
 774                         break;
 775                 case BPF_S_MISC_TAX:
 776                         /* X = A */
 777                         ctx->seen |= SEEN_X;
 778                         emit(ARM_MOV_R(r_X, r_A), ctx);
 779                         break;
 780                 case BPF_S_MISC_TXA:
 781                         /* A = X */
 782                         update_on_xread(ctx);
 783                         emit(ARM_MOV_R(r_A, r_X), ctx);
 784                         break;
 785                 case BPF_S_ANC_PROTOCOL:
 786                         /* A = ntohs(skb->protocol) */
 787                         ctx->seen |= SEEN_SKB;
 788                         BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
 789                                                   protocol) != 2);
 790                         off = offsetof(struct sk_buff, protocol);
 791                         emit(ARM_LDRH_I(r_scratch, r_skb, off), ctx);
 792                         emit_swap16(r_A, r_scratch, ctx);
 793                         break;
 794                 case BPF_S_ANC_CPU:
 795                         /* r_scratch = current_thread_info() */
 796                         OP_IMM3(ARM_BIC, r_scratch, ARM_SP, THREAD_SIZE - 1, ctx);
 797                         /* A = current_thread_info()->cpu */
 798                         BUILD_BUG_ON(FIELD_SIZEOF(struct thread_info, cpu) != 4);
 799                         off = offsetof(struct thread_info, cpu);
 800                         emit(ARM_LDR_I(r_A, r_scratch, off), ctx);
 801                         break;
 802                 case BPF_S_ANC_IFINDEX:
 803                         /* A = skb->dev->ifindex */
 804                         ctx->seen |= SEEN_SKB;
 805                         off = offsetof(struct sk_buff, dev);
 806                         emit(ARM_LDR_I(r_scratch, r_skb, off), ctx);
 807
 808                         emit(ARM_CMP_I(r_scratch, 0), ctx);
 809                         emit_err_ret(ARM_COND_EQ, ctx);
 810
 811                         BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
 812                                                   ifindex) != 4);
 813                         off = offsetof(struct net_device, ifindex);
 814                         emit(ARM_LDR_I(r_A, r_scratch, off), ctx);
 815                         break;
 816                 case BPF_S_ANC_MARK:
 817                         ctx->seen |= SEEN_SKB;
 818                         BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
 819                         off = offsetof(struct sk_buff, mark);
 820                         emit(ARM_LDR_I(r_A, r_skb, off), ctx);
 821                         break;
 822                 case BPF_S_ANC_RXHASH:
 823                         ctx->seen |= SEEN_SKB;
 824                         BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
 825                         off = offsetof(struct sk_buff, rxhash);
 826                         emit(ARM_LDR_I(r_A, r_skb, off), ctx);
 827                         break;
 828                 case BPF_S_ANC_VLAN_TAG:
 829                 case BPF_S_ANC_VLAN_TAG_PRESENT:
 830                         ctx->seen |= SEEN_SKB;
 831                         BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
 832                         off = offsetof(struct sk_buff, vlan_tci);
 833                         emit(ARM_LDRH_I(r_A, r_skb, off), ctx);
 834                         if (inst->code == BPF_S_ANC_VLAN_TAG)
 835                                 OP_IMM3(ARM_AND, r_A, r_A, VLAN_VID_MASK, ctx);
 836                         else
 837                                 OP_IMM3(ARM_AND, r_A, r_A, VLAN_TAG_PRESENT, ctx);
 838                         break;
 839                 case BPF_S_ANC_QUEUE:
 840                         ctx->seen |= SEEN_SKB;
 841                         BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
 842                                                   queue_mapping) != 2);
 843                         BUILD_BUG_ON(offsetof(struct sk_buff,
 844                                               queue_mapping) > 0xff);
 845                         off = offsetof(struct sk_buff, queue_mapping);
 846                         emit(ARM_LDRH_I(r_A, r_skb, off), ctx);
 847                         break;
 848                 default:
 849                         return -1;
 850                 }
 851         }
 852
 853         /* compute offsets only during the first pass */
 854         if (ctx->target == NULL)
 855                 ctx->offsets[i] = ctx->idx * 4;
 856
 857         return 0;
 858 }
 859
 860
 861 void bpf_jit_compile(struct sk_filter *fp)
 862 {
 863         struct jit_ctx ctx;
 864         unsigned tmp_idx;
 865         unsigned alloc_size;
 866
 867         if (!bpf_jit_enable)
 868                 return;
 869
 870         memset(&ctx, 0, sizeof(ctx));
 871         ctx.skf         = fp;
 872         ctx.ret0_fp_idx = -1;
 873
 874         ctx.offsets = kzalloc(4 * (ctx.skf->len + 1), GFP_KERNEL);
 875         if (ctx.offsets == NULL)
 876                 return;
 877
 878         /* fake pass to fill in the ctx->seen */
 879         if (unlikely(build_body(&ctx)))
 880                 goto out;
 881
 882         tmp_idx = ctx.idx;
 883         build_prologue(&ctx);
 884         ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
 885
 886 #if __LINUX_ARM_ARCH__ < 7
 887         tmp_idx = ctx.idx;
 888         build_epilogue(&ctx);
 889         ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
 890
 891         ctx.idx += ctx.imm_count;
 892         if (ctx.imm_count) {
 893                 ctx.imms = kzalloc(4 * ctx.imm_count, GFP_KERNEL);
 894                 if (ctx.imms == NULL)
 895                         goto out;
 896         }
 897 #else
 898         /* there's nothing after the epilogue on ARMv7 */
 899         build_epilogue(&ctx);
 900 #endif
 901
 902         alloc_size = 4 * ctx.idx;
 903         ctx.target = module_alloc(max(sizeof(struct work_struct),
 904                                       alloc_size));
 905         if (unlikely(ctx.target == NULL))
 906                 goto out;
 907
 908         ctx.idx = 0;
 909         build_prologue(&ctx);
 910         build_body(&ctx);
 911         build_epilogue(&ctx);
 912
 913         flush_icache_range((u32)ctx.target, (u32)(ctx.target + ctx.idx));
 914
 915 #if __LINUX_ARM_ARCH__ < 7
 916         if (ctx.imm_count)
 917                 kfree(ctx.imms);
 918 #endif
 919
 920         if (bpf_jit_enable > 1)
 921                 print_hex_dump(KERN_INFO, "BPF JIT code: ",
 922                                DUMP_PREFIX_ADDRESS, 16, 4, ctx.target,
 923                                alloc_size, false);
 924
 925         fp->bpf_func = (void *)ctx.target;
 926 out:
 927         kfree(ctx.offsets);
 928         return;
 929 }
 930
 931 static void bpf_jit_free_worker(struct work_struct *work)
 932 {
 933         module_free(NULL, work);
 934 }
 935
 936 void bpf_jit_free(struct sk_filter *fp)
 937 {
 938         struct work_struct *work;
 939
 940         if (fp->bpf_func != sk_run_filter) {
 941                 work = (struct work_struct *)fp->bpf_func;
 942
 943                 INIT_WORK(work, bpf_jit_free_worker);
 944                 schedule_work(work);
 945         }
 946 }