linux/arch/mips/net/bpf_jit.c
<<
>>
Prefs
   1/*
   2 * Just-In-Time compiler for BPF filters on MIPS
   3 *
   4 * Copyright (c) 2014 Imagination Technologies Ltd.
   5 * Author: Markos Chandras <markos.chandras@imgtec.com>
   6 *
   7 * This program is free software; you can redistribute it and/or modify it
   8 * under the terms of the GNU General Public License as published by the
   9 * Free Software Foundation; version 2 of the License.
  10 */
  11
  12#include <linux/bitops.h>
  13#include <linux/compiler.h>
  14#include <linux/errno.h>
  15#include <linux/filter.h>
  16#include <linux/if_vlan.h>
  17#include <linux/moduleloader.h>
  18#include <linux/netdevice.h>
  19#include <linux/string.h>
  20#include <linux/slab.h>
  21#include <linux/types.h>
  22#include <asm/asm.h>
  23#include <asm/bitops.h>
  24#include <asm/cacheflush.h>
  25#include <asm/cpu-features.h>
  26#include <asm/uasm.h>
  27
  28#include "bpf_jit.h"
  29
  30/* ABI
  31 * r_skb_hl     SKB header length
  32 * r_data       SKB data pointer
  33 * r_off        Offset
  34 * r_A          BPF register A
  35 * r_X          BPF register X
  36 * r_skb        *skb
  37 * r_M          *scratch memory
  38 * r_skb_len    SKB length
  39 *
  40 * On entry (*bpf_func)(*skb, *filter)
  41 * a0 = MIPS_R_A0 = skb;
  42 * a1 = MIPS_R_A1 = filter;
  43 *
  44 * Stack
  45 * ...
  46 * M[15]
  47 * M[14]
  48 * M[13]
  49 * ...
  50 * M[0] <-- r_M
  51 * saved reg k-1
  52 * saved reg k-2
  53 * ...
  54 * saved reg 0 <-- r_sp
  55 * <no argument area>
  56 *
  57 *                     Packet layout
  58 *
  59 * <--------------------- len ------------------------>
  60 * <--skb-len(r_skb_hl)-->< ----- skb->data_len ------>
  61 * ----------------------------------------------------
  62 * |                  skb->data                       |
  63 * ----------------------------------------------------
  64 */
  65
  66#define ptr typeof(unsigned long)
  67
  68#define SCRATCH_OFF(k)          (4 * (k))
  69
  70/* JIT flags */
  71#define SEEN_CALL               (1 << BPF_MEMWORDS)
  72#define SEEN_SREG_SFT           (BPF_MEMWORDS + 1)
  73#define SEEN_SREG_BASE          (1 << SEEN_SREG_SFT)
  74#define SEEN_SREG(x)            (SEEN_SREG_BASE << (x))
  75#define SEEN_OFF                SEEN_SREG(2)
  76#define SEEN_A                  SEEN_SREG(3)
  77#define SEEN_X                  SEEN_SREG(4)
  78#define SEEN_SKB                SEEN_SREG(5)
  79#define SEEN_MEM                SEEN_SREG(6)
  80/* SEEN_SK_DATA also implies skb_hl an skb_len */
  81#define SEEN_SKB_DATA           (SEEN_SREG(7) | SEEN_SREG(1) | SEEN_SREG(0))
  82
  83/* Arguments used by JIT */
  84#define ARGS_USED_BY_JIT        2 /* only applicable to 64-bit */
  85
  86#define SBIT(x)                 (1 << (x)) /* Signed version of BIT() */
  87
  88/**
  89 * struct jit_ctx - JIT context
  90 * @skf:                The sk_filter
  91 * @prologue_bytes:     Number of bytes for prologue
  92 * @idx:                Instruction index
  93 * @flags:              JIT flags
  94 * @offsets:            Instruction offsets
  95 * @target:             Memory location for the compiled filter
  96 */
  97struct jit_ctx {
  98        const struct bpf_prog *skf;
  99        unsigned int prologue_bytes;
 100        u32 idx;
 101        u32 flags;
 102        u32 *offsets;
 103        u32 *target;
 104};
 105
 106
 107static inline int optimize_div(u32 *k)
 108{
 109        /* power of 2 divides can be implemented with right shift */
 110        if (!(*k & (*k-1))) {
 111                *k = ilog2(*k);
 112                return 1;
 113        }
 114
 115        return 0;
 116}
 117
 118static inline void emit_jit_reg_move(ptr dst, ptr src, struct jit_ctx *ctx);
 119
 120/* Simply emit the instruction if the JIT memory space has been allocated */
 121#define emit_instr(ctx, func, ...)                      \
 122do {                                                    \
 123        if ((ctx)->target != NULL) {                    \
 124                u32 *p = &(ctx)->target[ctx->idx];      \
 125                uasm_i_##func(&p, ##__VA_ARGS__);       \
 126        }                                               \
 127        (ctx)->idx++;                                   \
 128} while (0)
 129
 130/*
 131 * Similar to emit_instr but it must be used when we need to emit
 132 * 32-bit or 64-bit instructions
 133 */
 134#define emit_long_instr(ctx, func, ...)                 \
 135do {                                                    \
 136        if ((ctx)->target != NULL) {                    \
 137                u32 *p = &(ctx)->target[ctx->idx];      \
 138                UASM_i_##func(&p, ##__VA_ARGS__);       \
 139        }                                               \
 140        (ctx)->idx++;                                   \
 141} while (0)
 142
 143/* Determine if immediate is within the 16-bit signed range */
 144static inline bool is_range16(s32 imm)
 145{
 146        return !(imm >= SBIT(15) || imm < -SBIT(15));
 147}
 148
 149static inline void emit_addu(unsigned int dst, unsigned int src1,
 150                             unsigned int src2, struct jit_ctx *ctx)
 151{
 152        emit_instr(ctx, addu, dst, src1, src2);
 153}
 154
 155static inline void emit_nop(struct jit_ctx *ctx)
 156{
 157        emit_instr(ctx, nop);
 158}
 159
 160/* Load a u32 immediate to a register */
 161static inline void emit_load_imm(unsigned int dst, u32 imm, struct jit_ctx *ctx)
 162{
 163        if (ctx->target != NULL) {
 164                /* addiu can only handle s16 */
 165                if (!is_range16(imm)) {
 166                        u32 *p = &ctx->target[ctx->idx];
 167                        uasm_i_lui(&p, r_tmp_imm, (s32)imm >> 16);
 168                        p = &ctx->target[ctx->idx + 1];
 169                        uasm_i_ori(&p, dst, r_tmp_imm, imm & 0xffff);
 170                } else {
 171                        u32 *p = &ctx->target[ctx->idx];
 172                        uasm_i_addiu(&p, dst, r_zero, imm);
 173                }
 174        }
 175        ctx->idx++;
 176
 177        if (!is_range16(imm))
 178                ctx->idx++;
 179}
 180
 181static inline void emit_or(unsigned int dst, unsigned int src1,
 182                           unsigned int src2, struct jit_ctx *ctx)
 183{
 184        emit_instr(ctx, or, dst, src1, src2);
 185}
 186
 187static inline void emit_ori(unsigned int dst, unsigned src, u32 imm,
 188                            struct jit_ctx *ctx)
 189{
 190        if (imm >= BIT(16)) {
 191                emit_load_imm(r_tmp, imm, ctx);
 192                emit_or(dst, src, r_tmp, ctx);
 193        } else {
 194                emit_instr(ctx, ori, dst, src, imm);
 195        }
 196}
 197
 198static inline void emit_daddiu(unsigned int dst, unsigned int src,
 199                               int imm, struct jit_ctx *ctx)
 200{
 201        /*
 202         * Only used for stack, so the imm is relatively small
 203         * and it fits in 15-bits
 204         */
 205        emit_instr(ctx, daddiu, dst, src, imm);
 206}
 207
 208static inline void emit_addiu(unsigned int dst, unsigned int src,
 209                              u32 imm, struct jit_ctx *ctx)
 210{
 211        if (!is_range16(imm)) {
 212                emit_load_imm(r_tmp, imm, ctx);
 213                emit_addu(dst, r_tmp, src, ctx);
 214        } else {
 215                emit_instr(ctx, addiu, dst, src, imm);
 216        }
 217}
 218
 219static inline void emit_and(unsigned int dst, unsigned int src1,
 220                            unsigned int src2, struct jit_ctx *ctx)
 221{
 222        emit_instr(ctx, and, dst, src1, src2);
 223}
 224
 225static inline void emit_andi(unsigned int dst, unsigned int src,
 226                             u32 imm, struct jit_ctx *ctx)
 227{
 228        /* If imm does not fit in u16 then load it to register */
 229        if (imm >= BIT(16)) {
 230                emit_load_imm(r_tmp, imm, ctx);
 231                emit_and(dst, src, r_tmp, ctx);
 232        } else {
 233                emit_instr(ctx, andi, dst, src, imm);
 234        }
 235}
 236
 237static inline void emit_xor(unsigned int dst, unsigned int src1,
 238                            unsigned int src2, struct jit_ctx *ctx)
 239{
 240        emit_instr(ctx, xor, dst, src1, src2);
 241}
 242
 243static inline void emit_xori(ptr dst, ptr src, u32 imm, struct jit_ctx *ctx)
 244{
 245        /* If imm does not fit in u16 then load it to register */
 246        if (imm >= BIT(16)) {
 247                emit_load_imm(r_tmp, imm, ctx);
 248                emit_xor(dst, src, r_tmp, ctx);
 249        } else {
 250                emit_instr(ctx, xori, dst, src, imm);
 251        }
 252}
 253
 254static inline void emit_stack_offset(int offset, struct jit_ctx *ctx)
 255{
 256        emit_long_instr(ctx, ADDIU, r_sp, r_sp, offset);
 257}
 258
 259static inline void emit_subu(unsigned int dst, unsigned int src1,
 260                             unsigned int src2, struct jit_ctx *ctx)
 261{
 262        emit_instr(ctx, subu, dst, src1, src2);
 263}
 264
 265static inline void emit_neg(unsigned int reg, struct jit_ctx *ctx)
 266{
 267        emit_subu(reg, r_zero, reg, ctx);
 268}
 269
 270static inline void emit_sllv(unsigned int dst, unsigned int src,
 271                             unsigned int sa, struct jit_ctx *ctx)
 272{
 273        emit_instr(ctx, sllv, dst, src, sa);
 274}
 275
 276static inline void emit_sll(unsigned int dst, unsigned int src,
 277                            unsigned int sa, struct jit_ctx *ctx)
 278{
 279        /* sa is 5-bits long */
 280        if (sa >= BIT(5))
 281                /* Shifting >= 32 results in zero */
 282                emit_jit_reg_move(dst, r_zero, ctx);
 283        else
 284                emit_instr(ctx, sll, dst, src, sa);
 285}
 286
 287static inline void emit_srlv(unsigned int dst, unsigned int src,
 288                             unsigned int sa, struct jit_ctx *ctx)
 289{
 290        emit_instr(ctx, srlv, dst, src, sa);
 291}
 292
 293static inline void emit_srl(unsigned int dst, unsigned int src,
 294                            unsigned int sa, struct jit_ctx *ctx)
 295{
 296        /* sa is 5-bits long */
 297        if (sa >= BIT(5))
 298                /* Shifting >= 32 results in zero */
 299                emit_jit_reg_move(dst, r_zero, ctx);
 300        else
 301                emit_instr(ctx, srl, dst, src, sa);
 302}
 303
 304static inline void emit_slt(unsigned int dst, unsigned int src1,
 305                            unsigned int src2, struct jit_ctx *ctx)
 306{
 307        emit_instr(ctx, slt, dst, src1, src2);
 308}
 309
 310static inline void emit_sltu(unsigned int dst, unsigned int src1,
 311                             unsigned int src2, struct jit_ctx *ctx)
 312{
 313        emit_instr(ctx, sltu, dst, src1, src2);
 314}
 315
 316static inline void emit_sltiu(unsigned dst, unsigned int src,
 317                              unsigned int imm, struct jit_ctx *ctx)
 318{
 319        /* 16 bit immediate */
 320        if (!is_range16((s32)imm)) {
 321                emit_load_imm(r_tmp, imm, ctx);
 322                emit_sltu(dst, src, r_tmp, ctx);
 323        } else {
 324                emit_instr(ctx, sltiu, dst, src, imm);
 325        }
 326
 327}
 328
 329/* Store register on the stack */
 330static inline void emit_store_stack_reg(ptr reg, ptr base,
 331                                        unsigned int offset,
 332                                        struct jit_ctx *ctx)
 333{
 334        emit_long_instr(ctx, SW, reg, offset, base);
 335}
 336
 337static inline void emit_store(ptr reg, ptr base, unsigned int offset,
 338                              struct jit_ctx *ctx)
 339{
 340        emit_instr(ctx, sw, reg, offset, base);
 341}
 342
 343static inline void emit_load_stack_reg(ptr reg, ptr base,
 344                                       unsigned int offset,
 345                                       struct jit_ctx *ctx)
 346{
 347        emit_long_instr(ctx, LW, reg, offset, base);
 348}
 349
 350static inline void emit_load(unsigned int reg, unsigned int base,
 351                             unsigned int offset, struct jit_ctx *ctx)
 352{
 353        emit_instr(ctx, lw, reg, offset, base);
 354}
 355
 356static inline void emit_load_byte(unsigned int reg, unsigned int base,
 357                                  unsigned int offset, struct jit_ctx *ctx)
 358{
 359        emit_instr(ctx, lb, reg, offset, base);
 360}
 361
 362static inline void emit_half_load(unsigned int reg, unsigned int base,
 363                                  unsigned int offset, struct jit_ctx *ctx)
 364{
 365        emit_instr(ctx, lh, reg, offset, base);
 366}
 367
 368static inline void emit_half_load_unsigned(unsigned int reg, unsigned int base,
 369                                           unsigned int offset, struct jit_ctx *ctx)
 370{
 371        emit_instr(ctx, lhu, reg, offset, base);
 372}
 373
 374static inline void emit_mul(unsigned int dst, unsigned int src1,
 375                            unsigned int src2, struct jit_ctx *ctx)
 376{
 377        emit_instr(ctx, mul, dst, src1, src2);
 378}
 379
 380static inline void emit_div(unsigned int dst, unsigned int src,
 381                            struct jit_ctx *ctx)
 382{
 383        if (ctx->target != NULL) {
 384                u32 *p = &ctx->target[ctx->idx];
 385                uasm_i_divu(&p, dst, src);
 386                p = &ctx->target[ctx->idx + 1];
 387                uasm_i_mflo(&p, dst);
 388        }
 389        ctx->idx += 2; /* 2 insts */
 390}
 391
 392static inline void emit_mod(unsigned int dst, unsigned int src,
 393                            struct jit_ctx *ctx)
 394{
 395        if (ctx->target != NULL) {
 396                u32 *p = &ctx->target[ctx->idx];
 397                uasm_i_divu(&p, dst, src);
 398                p = &ctx->target[ctx->idx + 1];
 399                uasm_i_mfhi(&p, dst);
 400        }
 401        ctx->idx += 2; /* 2 insts */
 402}
 403
 404static inline void emit_dsll(unsigned int dst, unsigned int src,
 405                             unsigned int sa, struct jit_ctx *ctx)
 406{
 407        emit_instr(ctx, dsll, dst, src, sa);
 408}
 409
 410static inline void emit_dsrl32(unsigned int dst, unsigned int src,
 411                               unsigned int sa, struct jit_ctx *ctx)
 412{
 413        emit_instr(ctx, dsrl32, dst, src, sa);
 414}
 415
 416static inline void emit_wsbh(unsigned int dst, unsigned int src,
 417                             struct jit_ctx *ctx)
 418{
 419        emit_instr(ctx, wsbh, dst, src);
 420}
 421
 422/* load pointer to register */
 423static inline void emit_load_ptr(unsigned int dst, unsigned int src,
 424                                     int imm, struct jit_ctx *ctx)
 425{
 426        /* src contains the base addr of the 32/64-pointer */
 427        emit_long_instr(ctx, LW, dst, imm, src);
 428}
 429
 430/* load a function pointer to register */
 431static inline void emit_load_func(unsigned int reg, ptr imm,
 432                                  struct jit_ctx *ctx)
 433{
 434        if (IS_ENABLED(CONFIG_64BIT)) {
 435                /* At this point imm is always 64-bit */
 436                emit_load_imm(r_tmp, (u64)imm >> 32, ctx);
 437                emit_dsll(r_tmp_imm, r_tmp, 16, ctx); /* left shift by 16 */
 438                emit_ori(r_tmp, r_tmp_imm, (imm >> 16) & 0xffff, ctx);
 439                emit_dsll(r_tmp_imm, r_tmp, 16, ctx); /* left shift by 16 */
 440                emit_ori(reg, r_tmp_imm, imm & 0xffff, ctx);
 441        } else {
 442                emit_load_imm(reg, imm, ctx);
 443        }
 444}
 445
 446/* Move to real MIPS register */
 447static inline void emit_reg_move(ptr dst, ptr src, struct jit_ctx *ctx)
 448{
 449        emit_long_instr(ctx, ADDU, dst, src, r_zero);
 450}
 451
 452/* Move to JIT (32-bit) register */
 453static inline void emit_jit_reg_move(ptr dst, ptr src, struct jit_ctx *ctx)
 454{
 455        emit_addu(dst, src, r_zero, ctx);
 456}
 457
 458/* Compute the immediate value for PC-relative branches. */
 459static inline u32 b_imm(unsigned int tgt, struct jit_ctx *ctx)
 460{
 461        if (ctx->target == NULL)
 462                return 0;
 463
 464        /*
 465         * We want a pc-relative branch. We only do forward branches
 466         * so tgt is always after pc. tgt is the instruction offset
 467         * we want to jump to.
 468
 469         * Branch on MIPS:
 470         * I: target_offset <- sign_extend(offset)
 471         * I+1: PC += target_offset (delay slot)
 472         *
 473         * ctx->idx currently points to the branch instruction
 474         * but the offset is added to the delay slot so we need
 475         * to subtract 4.
 476         */
 477        return ctx->offsets[tgt] -
 478                (ctx->idx * 4 - ctx->prologue_bytes) - 4;
 479}
 480
 481static inline void emit_bcond(int cond, unsigned int reg1, unsigned int reg2,
 482                             unsigned int imm, struct jit_ctx *ctx)
 483{
 484        if (ctx->target != NULL) {
 485                u32 *p = &ctx->target[ctx->idx];
 486
 487                switch (cond) {
 488                case MIPS_COND_EQ:
 489                        uasm_i_beq(&p, reg1, reg2, imm);
 490                        break;
 491                case MIPS_COND_NE:
 492                        uasm_i_bne(&p, reg1, reg2, imm);
 493                        break;
 494                case MIPS_COND_ALL:
 495                        uasm_i_b(&p, imm);
 496                        break;
 497                default:
 498                        pr_warn("%s: Unhandled branch conditional: %d\n",
 499                                __func__, cond);
 500                }
 501        }
 502        ctx->idx++;
 503}
 504
 505static inline void emit_b(unsigned int imm, struct jit_ctx *ctx)
 506{
 507        emit_bcond(MIPS_COND_ALL, r_zero, r_zero, imm, ctx);
 508}
 509
 510static inline void emit_jalr(unsigned int link, unsigned int reg,
 511                             struct jit_ctx *ctx)
 512{
 513        emit_instr(ctx, jalr, link, reg);
 514}
 515
 516static inline void emit_jr(unsigned int reg, struct jit_ctx *ctx)
 517{
 518        emit_instr(ctx, jr, reg);
 519}
 520
 521static inline u16 align_sp(unsigned int num)
 522{
 523        /* Double word alignment for 32-bit, quadword for 64-bit */
 524        unsigned int align = IS_ENABLED(CONFIG_64BIT) ? 16 : 8;
 525        num = (num + (align - 1)) & -align;
 526        return num;
 527}
 528
 529static void save_bpf_jit_regs(struct jit_ctx *ctx, unsigned offset)
 530{
 531        int i = 0, real_off = 0;
 532        u32 sflags, tmp_flags;
 533
 534        /* Adjust the stack pointer */
 535        if (offset)
 536                emit_stack_offset(-align_sp(offset), ctx);
 537
 538        tmp_flags = sflags = ctx->flags >> SEEN_SREG_SFT;
 539        /* sflags is essentially a bitmap */
 540        while (tmp_flags) {
 541                if ((sflags >> i) & 0x1) {
 542                        emit_store_stack_reg(MIPS_R_S0 + i, r_sp, real_off,
 543                                             ctx);
 544                        real_off += SZREG;
 545                }
 546                i++;
 547                tmp_flags >>= 1;
 548        }
 549
 550        /* save return address */
 551        if (ctx->flags & SEEN_CALL) {
 552                emit_store_stack_reg(r_ra, r_sp, real_off, ctx);
 553                real_off += SZREG;
 554        }
 555
 556        /* Setup r_M leaving the alignment gap if necessary */
 557        if (ctx->flags & SEEN_MEM) {
 558                if (real_off % (SZREG * 2))
 559                        real_off += SZREG;
 560                emit_long_instr(ctx, ADDIU, r_M, r_sp, real_off);
 561        }
 562}
 563
 564static void restore_bpf_jit_regs(struct jit_ctx *ctx,
 565                                 unsigned int offset)
 566{
 567        int i, real_off = 0;
 568        u32 sflags, tmp_flags;
 569
 570        tmp_flags = sflags = ctx->flags >> SEEN_SREG_SFT;
 571        /* sflags is a bitmap */
 572        i = 0;
 573        while (tmp_flags) {
 574                if ((sflags >> i) & 0x1) {
 575                        emit_load_stack_reg(MIPS_R_S0 + i, r_sp, real_off,
 576                                            ctx);
 577                        real_off += SZREG;
 578                }
 579                i++;
 580                tmp_flags >>= 1;
 581        }
 582
 583        /* restore return address */
 584        if (ctx->flags & SEEN_CALL)
 585                emit_load_stack_reg(r_ra, r_sp, real_off, ctx);
 586
 587        /* Restore the sp and discard the scrach memory */
 588        if (offset)
 589                emit_stack_offset(align_sp(offset), ctx);
 590}
 591
 592static unsigned int get_stack_depth(struct jit_ctx *ctx)
 593{
 594        int sp_off = 0;
 595
 596
 597        /* How may s* regs do we need to preserved? */
 598        sp_off += hweight32(ctx->flags >> SEEN_SREG_SFT) * SZREG;
 599
 600        if (ctx->flags & SEEN_MEM)
 601                sp_off += 4 * BPF_MEMWORDS; /* BPF_MEMWORDS are 32-bit */
 602
 603        if (ctx->flags & SEEN_CALL)
 604                sp_off += SZREG; /* Space for our ra register */
 605
 606        return sp_off;
 607}
 608
 609static void build_prologue(struct jit_ctx *ctx)
 610{
 611        int sp_off;
 612
 613        /* Calculate the total offset for the stack pointer */
 614        sp_off = get_stack_depth(ctx);
 615        save_bpf_jit_regs(ctx, sp_off);
 616
 617        if (ctx->flags & SEEN_SKB)
 618                emit_reg_move(r_skb, MIPS_R_A0, ctx);
 619
 620        if (ctx->flags & SEEN_SKB_DATA) {
 621                /* Load packet length */
 622                emit_load(r_skb_len, r_skb, offsetof(struct sk_buff, len),
 623                          ctx);
 624                emit_load(r_tmp, r_skb, offsetof(struct sk_buff, data_len),
 625                          ctx);
 626                /* Load the data pointer */
 627                emit_load_ptr(r_skb_data, r_skb,
 628                              offsetof(struct sk_buff, data), ctx);
 629                /* Load the header length */
 630                emit_subu(r_skb_hl, r_skb_len, r_tmp, ctx);
 631        }
 632
 633        if (ctx->flags & SEEN_X)
 634                emit_jit_reg_move(r_X, r_zero, ctx);
 635
 636        /*
 637         * Do not leak kernel data to userspace, we only need to clear
 638         * r_A if it is ever used.  In fact if it is never used, we
 639         * will not save/restore it, so clearing it in this case would
 640         * corrupt the state of the caller.
 641         */
 642        if (bpf_needs_clear_a(&ctx->skf->insns[0]) &&
 643            (ctx->flags & SEEN_A))
 644                emit_jit_reg_move(r_A, r_zero, ctx);
 645}
 646
 647static void build_epilogue(struct jit_ctx *ctx)
 648{
 649        unsigned int sp_off;
 650
 651        /* Calculate the total offset for the stack pointer */
 652
 653        sp_off = get_stack_depth(ctx);
 654        restore_bpf_jit_regs(ctx, sp_off);
 655
 656        /* Return */
 657        emit_jr(r_ra, ctx);
 658        emit_nop(ctx);
 659}
 660
 661#define CHOOSE_LOAD_FUNC(K, func) \
 662        ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative : func) : \
 663         func##_positive)
 664
 665static int build_body(struct jit_ctx *ctx)
 666{
 667        const struct bpf_prog *prog = ctx->skf;
 668        const struct sock_filter *inst;
 669        unsigned int i, off, condt;
 670        u32 k, b_off __maybe_unused;
 671        u8 (*sk_load_func)(unsigned long *skb, int offset);
 672
 673        for (i = 0; i < prog->len; i++) {
 674                u16 code;
 675
 676                inst = &(prog->insns[i]);
 677                pr_debug("%s: code->0x%02x, jt->0x%x, jf->0x%x, k->0x%x\n",
 678                         __func__, inst->code, inst->jt, inst->jf, inst->k);
 679                k = inst->k;
 680                code = bpf_anc_helper(inst);
 681
 682                if (ctx->target == NULL)
 683                        ctx->offsets[i] = ctx->idx * 4;
 684
 685                switch (code) {
 686                case BPF_LD | BPF_IMM:
 687                        /* A <- k ==> li r_A, k */
 688                        ctx->flags |= SEEN_A;
 689                        emit_load_imm(r_A, k, ctx);
 690                        break;
 691                case BPF_LD | BPF_W | BPF_LEN:
 692                        BUILD_BUG_ON(sizeof_field(struct sk_buff, len) != 4);
 693                        /* A <- len ==> lw r_A, offset(skb) */
 694                        ctx->flags |= SEEN_SKB | SEEN_A;
 695                        off = offsetof(struct sk_buff, len);
 696                        emit_load(r_A, r_skb, off, ctx);
 697                        break;
 698                case BPF_LD | BPF_MEM:
 699                        /* A <- M[k] ==> lw r_A, offset(M) */
 700                        ctx->flags |= SEEN_MEM | SEEN_A;
 701                        emit_load(r_A, r_M, SCRATCH_OFF(k), ctx);
 702                        break;
 703                case BPF_LD | BPF_W | BPF_ABS:
 704                        /* A <- P[k:4] */
 705                        sk_load_func = CHOOSE_LOAD_FUNC(k, sk_load_word);
 706                        goto load;
 707                case BPF_LD | BPF_H | BPF_ABS:
 708                        /* A <- P[k:2] */
 709                        sk_load_func = CHOOSE_LOAD_FUNC(k, sk_load_half);
 710                        goto load;
 711                case BPF_LD | BPF_B | BPF_ABS:
 712                        /* A <- P[k:1] */
 713                        sk_load_func = CHOOSE_LOAD_FUNC(k, sk_load_byte);
 714load:
 715                        emit_load_imm(r_off, k, ctx);
 716load_common:
 717                        ctx->flags |= SEEN_CALL | SEEN_OFF |
 718                                SEEN_SKB | SEEN_A | SEEN_SKB_DATA;
 719
 720                        emit_load_func(r_s0, (ptr)sk_load_func, ctx);
 721                        emit_reg_move(MIPS_R_A0, r_skb, ctx);
 722                        emit_jalr(MIPS_R_RA, r_s0, ctx);
 723                        /* Load second argument to delay slot */
 724                        emit_reg_move(MIPS_R_A1, r_off, ctx);
 725                        /* Check the error value */
 726                        emit_bcond(MIPS_COND_EQ, r_ret, 0, b_imm(i + 1, ctx),
 727                                   ctx);
 728                        /* Load return register on DS for failures */
 729                        emit_reg_move(r_ret, r_zero, ctx);
 730                        /* Return with error */
 731                        emit_b(b_imm(prog->len, ctx), ctx);
 732                        emit_nop(ctx);
 733                        break;
 734                case BPF_LD | BPF_W | BPF_IND:
 735                        /* A <- P[X + k:4] */
 736                        sk_load_func = sk_load_word;
 737                        goto load_ind;
 738                case BPF_LD | BPF_H | BPF_IND:
 739                        /* A <- P[X + k:2] */
 740                        sk_load_func = sk_load_half;
 741                        goto load_ind;
 742                case BPF_LD | BPF_B | BPF_IND:
 743                        /* A <- P[X + k:1] */
 744                        sk_load_func = sk_load_byte;
 745load_ind:
 746                        ctx->flags |= SEEN_OFF | SEEN_X;
 747                        emit_addiu(r_off, r_X, k, ctx);
 748                        goto load_common;
 749                case BPF_LDX | BPF_IMM:
 750                        /* X <- k */
 751                        ctx->flags |= SEEN_X;
 752                        emit_load_imm(r_X, k, ctx);
 753                        break;
 754                case BPF_LDX | BPF_MEM:
 755                        /* X <- M[k] */
 756                        ctx->flags |= SEEN_X | SEEN_MEM;
 757                        emit_load(r_X, r_M, SCRATCH_OFF(k), ctx);
 758                        break;
 759                case BPF_LDX | BPF_W | BPF_LEN:
 760                        /* X <- len */
 761                        ctx->flags |= SEEN_X | SEEN_SKB;
 762                        off = offsetof(struct sk_buff, len);
 763                        emit_load(r_X, r_skb, off, ctx);
 764                        break;
 765                case BPF_LDX | BPF_B | BPF_MSH:
 766                        /* X <- 4 * (P[k:1] & 0xf) */
 767                        ctx->flags |= SEEN_X | SEEN_CALL | SEEN_SKB;
 768                        /* Load offset to a1 */
 769                        emit_load_func(r_s0, (ptr)sk_load_byte, ctx);
 770                        /*
 771                         * This may emit two instructions so it may not fit
 772                         * in the delay slot. So use a0 in the delay slot.
 773                         */
 774                        emit_load_imm(MIPS_R_A1, k, ctx);
 775                        emit_jalr(MIPS_R_RA, r_s0, ctx);
 776                        emit_reg_move(MIPS_R_A0, r_skb, ctx); /* delay slot */
 777                        /* Check the error value */
 778                        emit_bcond(MIPS_COND_NE, r_ret, 0,
 779                                   b_imm(prog->len, ctx), ctx);
 780                        emit_reg_move(r_ret, r_zero, ctx);
 781                        /* We are good */
 782                        /* X <- P[1:K] & 0xf */
 783                        emit_andi(r_X, r_A, 0xf, ctx);
 784                        /* X << 2 */
 785                        emit_b(b_imm(i + 1, ctx), ctx);
 786                        emit_sll(r_X, r_X, 2, ctx); /* delay slot */
 787                        break;
 788                case BPF_ST:
 789                        /* M[k] <- A */
 790                        ctx->flags |= SEEN_MEM | SEEN_A;
 791                        emit_store(r_A, r_M, SCRATCH_OFF(k), ctx);
 792                        break;
 793                case BPF_STX:
 794                        /* M[k] <- X */
 795                        ctx->flags |= SEEN_MEM | SEEN_X;
 796                        emit_store(r_X, r_M, SCRATCH_OFF(k), ctx);
 797                        break;
 798                case BPF_ALU | BPF_ADD | BPF_K:
 799                        /* A += K */
 800                        ctx->flags |= SEEN_A;
 801                        emit_addiu(r_A, r_A, k, ctx);
 802                        break;
 803                case BPF_ALU | BPF_ADD | BPF_X:
 804                        /* A += X */
 805                        ctx->flags |= SEEN_A | SEEN_X;
 806                        emit_addu(r_A, r_A, r_X, ctx);
 807                        break;
 808                case BPF_ALU | BPF_SUB | BPF_K:
 809                        /* A -= K */
 810                        ctx->flags |= SEEN_A;
 811                        emit_addiu(r_A, r_A, -k, ctx);
 812                        break;
 813                case BPF_ALU | BPF_SUB | BPF_X:
 814                        /* A -= X */
 815                        ctx->flags |= SEEN_A | SEEN_X;
 816                        emit_subu(r_A, r_A, r_X, ctx);
 817                        break;
 818                case BPF_ALU | BPF_MUL | BPF_K:
 819                        /* A *= K */
 820                        /* Load K to scratch register before MUL */
 821                        ctx->flags |= SEEN_A;
 822                        emit_load_imm(r_s0, k, ctx);
 823                        emit_mul(r_A, r_A, r_s0, ctx);
 824                        break;
 825                case BPF_ALU | BPF_MUL | BPF_X:
 826                        /* A *= X */
 827                        ctx->flags |= SEEN_A | SEEN_X;
 828                        emit_mul(r_A, r_A, r_X, ctx);
 829                        break;
 830                case BPF_ALU | BPF_DIV | BPF_K:
 831                        /* A /= k */
 832                        if (k == 1)
 833                                break;
 834                        if (optimize_div(&k)) {
 835                                ctx->flags |= SEEN_A;
 836                                emit_srl(r_A, r_A, k, ctx);
 837                                break;
 838                        }
 839                        ctx->flags |= SEEN_A;
 840                        emit_load_imm(r_s0, k, ctx);
 841                        emit_div(r_A, r_s0, ctx);
 842                        break;
 843                case BPF_ALU | BPF_MOD | BPF_K:
 844                        /* A %= k */
 845                        if (k == 1) {
 846                                ctx->flags |= SEEN_A;
 847                                emit_jit_reg_move(r_A, r_zero, ctx);
 848                        } else {
 849                                ctx->flags |= SEEN_A;
 850                                emit_load_imm(r_s0, k, ctx);
 851                                emit_mod(r_A, r_s0, ctx);
 852                        }
 853                        break;
 854                case BPF_ALU | BPF_DIV | BPF_X:
 855                        /* A /= X */
 856                        ctx->flags |= SEEN_X | SEEN_A;
 857                        /* Check if r_X is zero */
 858                        emit_bcond(MIPS_COND_EQ, r_X, r_zero,
 859                                   b_imm(prog->len, ctx), ctx);
 860                        emit_load_imm(r_ret, 0, ctx); /* delay slot */
 861                        emit_div(r_A, r_X, ctx);
 862                        break;
 863                case BPF_ALU | BPF_MOD | BPF_X:
 864                        /* A %= X */
 865                        ctx->flags |= SEEN_X | SEEN_A;
 866                        /* Check if r_X is zero */
 867                        emit_bcond(MIPS_COND_EQ, r_X, r_zero,
 868                                   b_imm(prog->len, ctx), ctx);
 869                        emit_load_imm(r_ret, 0, ctx); /* delay slot */
 870                        emit_mod(r_A, r_X, ctx);
 871                        break;
 872                case BPF_ALU | BPF_OR | BPF_K:
 873                        /* A |= K */
 874                        ctx->flags |= SEEN_A;
 875                        emit_ori(r_A, r_A, k, ctx);
 876                        break;
 877                case BPF_ALU | BPF_OR | BPF_X:
 878                        /* A |= X */
 879                        ctx->flags |= SEEN_A;
 880                        emit_ori(r_A, r_A, r_X, ctx);
 881                        break;
 882                case BPF_ALU | BPF_XOR | BPF_K:
 883                        /* A ^= k */
 884                        ctx->flags |= SEEN_A;
 885                        emit_xori(r_A, r_A, k, ctx);
 886                        break;
 887                case BPF_ANC | SKF_AD_ALU_XOR_X:
 888                case BPF_ALU | BPF_XOR | BPF_X:
 889                        /* A ^= X */
 890                        ctx->flags |= SEEN_A;
 891                        emit_xor(r_A, r_A, r_X, ctx);
 892                        break;
 893                case BPF_ALU | BPF_AND | BPF_K:
 894                        /* A &= K */
 895                        ctx->flags |= SEEN_A;
 896                        emit_andi(r_A, r_A, k, ctx);
 897                        break;
 898                case BPF_ALU | BPF_AND | BPF_X:
 899                        /* A &= X */
 900                        ctx->flags |= SEEN_A | SEEN_X;
 901                        emit_and(r_A, r_A, r_X, ctx);
 902                        break;
 903                case BPF_ALU | BPF_LSH | BPF_K:
 904                        /* A <<= K */
 905                        ctx->flags |= SEEN_A;
 906                        emit_sll(r_A, r_A, k, ctx);
 907                        break;
 908                case BPF_ALU | BPF_LSH | BPF_X:
 909                        /* A <<= X */
 910                        ctx->flags |= SEEN_A | SEEN_X;
 911                        emit_sllv(r_A, r_A, r_X, ctx);
 912                        break;
 913                case BPF_ALU | BPF_RSH | BPF_K:
 914                        /* A >>= K */
 915                        ctx->flags |= SEEN_A;
 916                        emit_srl(r_A, r_A, k, ctx);
 917                        break;
 918                case BPF_ALU | BPF_RSH | BPF_X:
 919                        ctx->flags |= SEEN_A | SEEN_X;
 920                        emit_srlv(r_A, r_A, r_X, ctx);
 921                        break;
 922                case BPF_ALU | BPF_NEG:
 923                        /* A = -A */
 924                        ctx->flags |= SEEN_A;
 925                        emit_neg(r_A, ctx);
 926                        break;
 927                case BPF_JMP | BPF_JA:
 928                        /* pc += K */
 929                        emit_b(b_imm(i + k + 1, ctx), ctx);
 930                        emit_nop(ctx);
 931                        break;
 932                case BPF_JMP | BPF_JEQ | BPF_K:
 933                        /* pc += ( A == K ) ? pc->jt : pc->jf */
 934                        condt = MIPS_COND_EQ | MIPS_COND_K;
 935                        goto jmp_cmp;
 936                case BPF_JMP | BPF_JEQ | BPF_X:
 937                        ctx->flags |= SEEN_X;
 938                        /* pc += ( A == X ) ? pc->jt : pc->jf */
 939                        condt = MIPS_COND_EQ | MIPS_COND_X;
 940                        goto jmp_cmp;
 941                case BPF_JMP | BPF_JGE | BPF_K:
 942                        /* pc += ( A >= K ) ? pc->jt : pc->jf */
 943                        condt = MIPS_COND_GE | MIPS_COND_K;
 944                        goto jmp_cmp;
 945                case BPF_JMP | BPF_JGE | BPF_X:
 946                        ctx->flags |= SEEN_X;
 947                        /* pc += ( A >= X ) ? pc->jt : pc->jf */
 948                        condt = MIPS_COND_GE | MIPS_COND_X;
 949                        goto jmp_cmp;
 950                case BPF_JMP | BPF_JGT | BPF_K:
 951                        /* pc += ( A > K ) ? pc->jt : pc->jf */
 952                        condt = MIPS_COND_GT | MIPS_COND_K;
 953                        goto jmp_cmp;
 954                case BPF_JMP | BPF_JGT | BPF_X:
 955                        ctx->flags |= SEEN_X;
 956                        /* pc += ( A > X ) ? pc->jt : pc->jf */
 957                        condt = MIPS_COND_GT | MIPS_COND_X;
 958jmp_cmp:
 959                        /* Greater or Equal */
 960                        if ((condt & MIPS_COND_GE) ||
 961                            (condt & MIPS_COND_GT)) {
 962                                if (condt & MIPS_COND_K) { /* K */
 963                                        ctx->flags |= SEEN_A;
 964                                        emit_sltiu(r_s0, r_A, k, ctx);
 965                                } else { /* X */
 966                                        ctx->flags |= SEEN_A |
 967                                                SEEN_X;
 968                                        emit_sltu(r_s0, r_A, r_X, ctx);
 969                                }
 970                                /* A < (K|X) ? r_scrach = 1 */
 971                                b_off = b_imm(i + inst->jf + 1, ctx);
 972                                emit_bcond(MIPS_COND_NE, r_s0, r_zero, b_off,
 973                                           ctx);
 974                                emit_nop(ctx);
 975                                /* A > (K|X) ? scratch = 0 */
 976                                if (condt & MIPS_COND_GT) {
 977                                        /* Checking for equality */
 978                                        ctx->flags |= SEEN_A | SEEN_X;
 979                                        if (condt & MIPS_COND_K)
 980                                                emit_load_imm(r_s0, k, ctx);
 981                                        else
 982                                                emit_jit_reg_move(r_s0, r_X,
 983                                                                  ctx);
 984                                        b_off = b_imm(i + inst->jf + 1, ctx);
 985                                        emit_bcond(MIPS_COND_EQ, r_A, r_s0,
 986                                                   b_off, ctx);
 987                                        emit_nop(ctx);
 988                                        /* Finally, A > K|X */
 989                                        b_off = b_imm(i + inst->jt + 1, ctx);
 990                                        emit_b(b_off, ctx);
 991                                        emit_nop(ctx);
 992                                } else {
 993                                        /* A >= (K|X) so jump */
 994                                        b_off = b_imm(i + inst->jt + 1, ctx);
 995                                        emit_b(b_off, ctx);
 996                                        emit_nop(ctx);
 997                                }
 998                        } else {
 999                                /* A == K|X */
1000                                if (condt & MIPS_COND_K) { /* K */
1001                                        ctx->flags |= SEEN_A;
1002                                        emit_load_imm(r_s0, k, ctx);
1003                                        /* jump true */
1004                                        b_off = b_imm(i + inst->jt + 1, ctx);
1005                                        emit_bcond(MIPS_COND_EQ, r_A, r_s0,
1006                                                   b_off, ctx);
1007                                        emit_nop(ctx);
1008                                        /* jump false */
1009                                        b_off = b_imm(i + inst->jf + 1,
1010                                                      ctx);
1011                                        emit_bcond(MIPS_COND_NE, r_A, r_s0,
1012                                                   b_off, ctx);
1013                                        emit_nop(ctx);
1014                                } else { /* X */
1015                                        /* jump true */
1016                                        ctx->flags |= SEEN_A | SEEN_X;
1017                                        b_off = b_imm(i + inst->jt + 1,
1018                                                      ctx);
1019                                        emit_bcond(MIPS_COND_EQ, r_A, r_X,
1020                                                   b_off, ctx);
1021                                        emit_nop(ctx);
1022                                        /* jump false */
1023                                        b_off = b_imm(i + inst->jf + 1, ctx);
1024                                        emit_bcond(MIPS_COND_NE, r_A, r_X,
1025                                                   b_off, ctx);
1026                                        emit_nop(ctx);
1027                                }
1028                        }
1029                        break;
1030                case BPF_JMP | BPF_JSET | BPF_K:
1031                        ctx->flags |= SEEN_A;
1032                        /* pc += (A & K) ? pc -> jt : pc -> jf */
1033                        emit_load_imm(r_s1, k, ctx);
1034                        emit_and(r_s0, r_A, r_s1, ctx);
1035                        /* jump true */
1036                        b_off = b_imm(i + inst->jt + 1, ctx);
1037                        emit_bcond(MIPS_COND_NE, r_s0, r_zero, b_off, ctx);
1038                        emit_nop(ctx);
1039                        /* jump false */
1040                        b_off = b_imm(i + inst->jf + 1, ctx);
1041                        emit_b(b_off, ctx);
1042                        emit_nop(ctx);
1043                        break;
1044                case BPF_JMP | BPF_JSET | BPF_X:
1045                        ctx->flags |= SEEN_X | SEEN_A;
1046                        /* pc += (A & X) ? pc -> jt : pc -> jf */
1047                        emit_and(r_s0, r_A, r_X, ctx);
1048                        /* jump true */
1049                        b_off = b_imm(i + inst->jt + 1, ctx);
1050                        emit_bcond(MIPS_COND_NE, r_s0, r_zero, b_off, ctx);
1051                        emit_nop(ctx);
1052                        /* jump false */
1053                        b_off = b_imm(i + inst->jf + 1, ctx);
1054                        emit_b(b_off, ctx);
1055                        emit_nop(ctx);
1056                        break;
1057                case BPF_RET | BPF_A:
1058                        ctx->flags |= SEEN_A;
1059                        if (i != prog->len - 1)
1060                                /*
1061                                 * If this is not the last instruction
1062                                 * then jump to the epilogue
1063                                 */
1064                                emit_b(b_imm(prog->len, ctx), ctx);
1065                        emit_reg_move(r_ret, r_A, ctx); /* delay slot */
1066                        break;
1067                case BPF_RET | BPF_K:
1068                        /*
1069                         * It can emit two instructions so it does not fit on
1070                         * the delay slot.
1071                         */
1072                        emit_load_imm(r_ret, k, ctx);
1073                        if (i != prog->len - 1) {
1074                                /*
1075                                 * If this is not the last instruction
1076                                 * then jump to the epilogue
1077                                 */
1078                                emit_b(b_imm(prog->len, ctx), ctx);
1079                                emit_nop(ctx);
1080                        }
1081                        break;
1082                case BPF_MISC | BPF_TAX:
1083                        /* X = A */
1084                        ctx->flags |= SEEN_X | SEEN_A;
1085                        emit_jit_reg_move(r_X, r_A, ctx);
1086                        break;
1087                case BPF_MISC | BPF_TXA:
1088                        /* A = X */
1089                        ctx->flags |= SEEN_A | SEEN_X;
1090                        emit_jit_reg_move(r_A, r_X, ctx);
1091                        break;
1092                /* AUX */
1093                case BPF_ANC | SKF_AD_PROTOCOL:
1094                        /* A = ntohs(skb->protocol */
1095                        ctx->flags |= SEEN_SKB | SEEN_OFF | SEEN_A;
1096                        BUILD_BUG_ON(sizeof_field(struct sk_buff,
1097                                                  protocol) != 2);
1098                        off = offsetof(struct sk_buff, protocol);
1099                        emit_half_load(r_A, r_skb, off, ctx);
1100#ifdef CONFIG_CPU_LITTLE_ENDIAN
1101                        /* This needs little endian fixup */
1102                        if (cpu_has_wsbh) {
1103                                /* R2 and later have the wsbh instruction */
1104                                emit_wsbh(r_A, r_A, ctx);
1105                        } else {
1106                                /* Get first byte */
1107                                emit_andi(r_tmp_imm, r_A, 0xff, ctx);
1108                                /* Shift it */
1109                                emit_sll(r_tmp, r_tmp_imm, 8, ctx);
1110                                /* Get second byte */
1111                                emit_srl(r_tmp_imm, r_A, 8, ctx);
1112                                emit_andi(r_tmp_imm, r_tmp_imm, 0xff, ctx);
1113                                /* Put everyting together in r_A */
1114                                emit_or(r_A, r_tmp, r_tmp_imm, ctx);
1115                        }
1116#endif
1117                        break;
1118                case BPF_ANC | SKF_AD_CPU:
1119                        ctx->flags |= SEEN_A | SEEN_OFF;
1120                        /* A = current_thread_info()->cpu */
1121                        BUILD_BUG_ON(sizeof_field(struct thread_info,
1122                                                  cpu) != 4);
1123                        off = offsetof(struct thread_info, cpu);
1124                        /* $28/gp points to the thread_info struct */
1125                        emit_load(r_A, 28, off, ctx);
1126                        break;
1127                case BPF_ANC | SKF_AD_IFINDEX:
1128                        /* A = skb->dev->ifindex */
1129                case BPF_ANC | SKF_AD_HATYPE:
1130                        /* A = skb->dev->type */
1131                        ctx->flags |= SEEN_SKB | SEEN_A;
1132                        off = offsetof(struct sk_buff, dev);
1133                        /* Load *dev pointer */
1134                        emit_load_ptr(r_s0, r_skb, off, ctx);
1135                        /* error (0) in the delay slot */
1136                        emit_bcond(MIPS_COND_EQ, r_s0, r_zero,
1137                                   b_imm(prog->len, ctx), ctx);
1138                        emit_reg_move(r_ret, r_zero, ctx);
1139                        if (code == (BPF_ANC | SKF_AD_IFINDEX)) {
1140                                BUILD_BUG_ON(sizeof_field(struct net_device, ifindex) != 4);
1141                                off = offsetof(struct net_device, ifindex);
1142                                emit_load(r_A, r_s0, off, ctx);
1143                        } else { /* (code == (BPF_ANC | SKF_AD_HATYPE) */
1144                                BUILD_BUG_ON(sizeof_field(struct net_device, type) != 2);
1145                                off = offsetof(struct net_device, type);
1146                                emit_half_load_unsigned(r_A, r_s0, off, ctx);
1147                        }
1148                        break;
1149                case BPF_ANC | SKF_AD_MARK:
1150                        ctx->flags |= SEEN_SKB | SEEN_A;
1151                        BUILD_BUG_ON(sizeof_field(struct sk_buff, mark) != 4);
1152                        off = offsetof(struct sk_buff, mark);
1153                        emit_load(r_A, r_skb, off, ctx);
1154                        break;
1155                case BPF_ANC | SKF_AD_RXHASH:
1156                        ctx->flags |= SEEN_SKB | SEEN_A;
1157                        BUILD_BUG_ON(sizeof_field(struct sk_buff, hash) != 4);
1158                        off = offsetof(struct sk_buff, hash);
1159                        emit_load(r_A, r_skb, off, ctx);
1160                        break;
1161                case BPF_ANC | SKF_AD_VLAN_TAG:
1162                        ctx->flags |= SEEN_SKB | SEEN_A;
1163                        BUILD_BUG_ON(sizeof_field(struct sk_buff,
1164                                                  vlan_tci) != 2);
1165                        off = offsetof(struct sk_buff, vlan_tci);
1166                        emit_half_load_unsigned(r_A, r_skb, off, ctx);
1167                        break;
1168                case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
1169                        ctx->flags |= SEEN_SKB | SEEN_A;
1170                        emit_load_byte(r_A, r_skb, PKT_VLAN_PRESENT_OFFSET(), ctx);
1171                        if (PKT_VLAN_PRESENT_BIT)
1172                                emit_srl(r_A, r_A, PKT_VLAN_PRESENT_BIT, ctx);
1173                        if (PKT_VLAN_PRESENT_BIT < 7)
1174                                emit_andi(r_A, r_A, 1, ctx);
1175                        break;
1176                case BPF_ANC | SKF_AD_PKTTYPE:
1177                        ctx->flags |= SEEN_SKB;
1178
1179                        emit_load_byte(r_tmp, r_skb, PKT_TYPE_OFFSET(), ctx);
1180                        /* Keep only the last 3 bits */
1181                        emit_andi(r_A, r_tmp, PKT_TYPE_MAX, ctx);
1182#ifdef __BIG_ENDIAN_BITFIELD
1183                        /* Get the actual packet type to the lower 3 bits */
1184                        emit_srl(r_A, r_A, 5, ctx);
1185#endif
1186                        break;
1187                case BPF_ANC | SKF_AD_QUEUE:
1188                        ctx->flags |= SEEN_SKB | SEEN_A;
1189                        BUILD_BUG_ON(sizeof_field(struct sk_buff,
1190                                                  queue_mapping) != 2);
1191                        BUILD_BUG_ON(offsetof(struct sk_buff,
1192                                              queue_mapping) > 0xff);
1193                        off = offsetof(struct sk_buff, queue_mapping);
1194                        emit_half_load_unsigned(r_A, r_skb, off, ctx);
1195                        break;
1196                default:
1197                        pr_debug("%s: Unhandled opcode: 0x%02x\n", __FILE__,
1198                                 inst->code);
1199                        return -1;
1200                }
1201        }
1202
1203        /* compute offsets only during the first pass */
1204        if (ctx->target == NULL)
1205                ctx->offsets[i] = ctx->idx * 4;
1206
1207        return 0;
1208}
1209
1210void bpf_jit_compile(struct bpf_prog *fp)
1211{
1212        struct jit_ctx ctx;
1213        unsigned int alloc_size, tmp_idx;
1214
1215        if (!bpf_jit_enable)
1216                return;
1217
1218        memset(&ctx, 0, sizeof(ctx));
1219
1220        ctx.offsets = kcalloc(fp->len + 1, sizeof(*ctx.offsets), GFP_KERNEL);
1221        if (ctx.offsets == NULL)
1222                return;
1223
1224        ctx.skf = fp;
1225
1226        if (build_body(&ctx))
1227                goto out;
1228
1229        tmp_idx = ctx.idx;
1230        build_prologue(&ctx);
1231        ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1232        /* just to complete the ctx.idx count */
1233        build_epilogue(&ctx);
1234
1235        alloc_size = 4 * ctx.idx;
1236        ctx.target = module_alloc(alloc_size);
1237        if (ctx.target == NULL)
1238                goto out;
1239
1240        /* Clean it */
1241        memset(ctx.target, 0, alloc_size);
1242
1243        ctx.idx = 0;
1244
1245        /* Generate the actual JIT code */
1246        build_prologue(&ctx);
1247        build_body(&ctx);
1248        build_epilogue(&ctx);
1249
1250        /* Update the icache */
1251        flush_icache_range((ptr)ctx.target, (ptr)(ctx.target + ctx.idx));
1252
1253        if (bpf_jit_enable > 1)
1254                /* Dump JIT code */
1255                bpf_jit_dump(fp->len, alloc_size, 2, ctx.target);
1256
1257        fp->bpf_func = (void *)ctx.target;
1258        fp->jited = 1;
1259
1260out:
1261        kfree(ctx.offsets);
1262}
1263
1264void bpf_jit_free(struct bpf_prog *fp)
1265{
1266        if (fp->jited)
1267                module_memfree(fp->bpf_func);
1268
1269        bpf_prog_unlock_free(fp);
1270}
1271