NetBSD/sys/net/bpfjit.c

2311 lines
51 KiB
C

/* $NetBSD: bpfjit.c,v 1.48 2020/02/01 02:54:02 riastradh Exp $ */
/*-
* Copyright (c) 2011-2015 Alexander Nasonov.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
* FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
* COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
* AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/cdefs.h>
#ifdef _KERNEL
__KERNEL_RCSID(0, "$NetBSD: bpfjit.c,v 1.48 2020/02/01 02:54:02 riastradh Exp $");
#else
__RCSID("$NetBSD: bpfjit.c,v 1.48 2020/02/01 02:54:02 riastradh Exp $");
#endif
#include <sys/types.h>
#include <sys/queue.h>
#ifndef _KERNEL
#include <assert.h>
#define BJ_ASSERT(c) assert(c)
#else
#define BJ_ASSERT(c) KASSERT(c)
#endif
#ifndef _KERNEL
#include <stdlib.h>
#define BJ_ALLOC(sz) malloc(sz)
#define BJ_FREE(p, sz) free(p)
#else
#include <sys/kmem.h>
#define BJ_ALLOC(sz) kmem_alloc(sz, KM_SLEEP)
#define BJ_FREE(p, sz) kmem_free(p, sz)
#endif
#ifndef _KERNEL
#include <limits.h>
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#include <string.h>
#else
#include <sys/atomic.h>
#include <sys/module.h>
#endif
#define __BPF_PRIVATE
#include <net/bpf.h>
#include <net/bpfjit.h>
#include <sljitLir.h>
#if !defined(_KERNEL) && defined(SLJIT_VERBOSE) && SLJIT_VERBOSE
#include <stdio.h> /* for stderr */
#endif
/*
* Number of saved registers to pass to sljit_emit_enter() function.
*/
#define NSAVEDS 3
/*
* Arguments of generated bpfjit_func_t.
* The first argument is reassigned upon entry
* to a more frequently used buf argument.
*/
#define BJ_CTX_ARG SLJIT_S0
#define BJ_ARGS SLJIT_S1
/*
* Permanent register assignments.
*/
#define BJ_BUF SLJIT_S0
//#define BJ_ARGS SLJIT_S1
#define BJ_BUFLEN SLJIT_S2
#define BJ_AREG SLJIT_R0
#define BJ_TMP1REG SLJIT_R1
#define BJ_TMP2REG SLJIT_R2
#define BJ_XREG SLJIT_R3
#define BJ_TMP3REG SLJIT_R4
#ifdef _KERNEL
#define MAX_MEMWORDS BPF_MAX_MEMWORDS
#else
#define MAX_MEMWORDS BPF_MEMWORDS
#endif
#define BJ_INIT_NOBITS ((bpf_memword_init_t)0)
#define BJ_INIT_MBIT(k) BPF_MEMWORD_INIT(k)
#define BJ_INIT_ABIT BJ_INIT_MBIT(MAX_MEMWORDS)
#define BJ_INIT_XBIT BJ_INIT_MBIT(MAX_MEMWORDS + 1)
/*
* Get a number of memwords and external memwords from a bpf_ctx object.
*/
#define GET_EXTWORDS(bc) ((bc) ? (bc)->extwords : 0)
#define GET_MEMWORDS(bc) (GET_EXTWORDS(bc) ? GET_EXTWORDS(bc) : BPF_MEMWORDS)
/*
* Optimization hints.
*/
typedef unsigned int bpfjit_hint_t;
#define BJ_HINT_ABS 0x01 /* packet read at absolute offset */
#define BJ_HINT_IND 0x02 /* packet read at variable offset */
#define BJ_HINT_MSH 0x04 /* BPF_MSH instruction */
#define BJ_HINT_COP 0x08 /* BPF_COP or BPF_COPX instruction */
#define BJ_HINT_COPX 0x10 /* BPF_COPX instruction */
#define BJ_HINT_XREG 0x20 /* BJ_XREG is needed */
#define BJ_HINT_LDX 0x40 /* BPF_LDX instruction */
#define BJ_HINT_PKT (BJ_HINT_ABS|BJ_HINT_IND|BJ_HINT_MSH)
/*
* Datatype for Array Bounds Check Elimination (ABC) pass.
*/
typedef uint64_t bpfjit_abc_length_t;
#define MAX_ABC_LENGTH (UINT32_MAX + UINT64_C(4)) /* max. width is 4 */
struct bpfjit_stack
{
bpf_ctx_t *ctx;
uint32_t *extmem; /* pointer to external memory store */
uint32_t reg; /* saved A or X register */
#ifdef _KERNEL
int err; /* 3rd argument for m_xword/m_xhalf/m_xbyte function call */
#endif
uint32_t mem[BPF_MEMWORDS]; /* internal memory store */
};
/*
* Data for BPF_JMP instruction.
* Forward declaration for struct bpfjit_jump.
*/
struct bpfjit_jump_data;
/*
* Node of bjumps list.
*/
struct bpfjit_jump {
struct sljit_jump *sjump;
SLIST_ENTRY(bpfjit_jump) entries;
struct bpfjit_jump_data *jdata;
};
/*
* Data for BPF_JMP instruction.
*/
struct bpfjit_jump_data {
/*
* These entries make up bjumps list:
* jtf[0] - when coming from jt path,
* jtf[1] - when coming from jf path.
*/
struct bpfjit_jump jtf[2];
/*
* Length calculated by Array Bounds Check Elimination (ABC) pass.
*/
bpfjit_abc_length_t abc_length;
/*
* Length checked by the last out-of-bounds check.
*/
bpfjit_abc_length_t checked_length;
};
/*
* Data for "read from packet" instructions.
* See also read_pkt_insn() function below.
*/
struct bpfjit_read_pkt_data {
/*
* Length calculated by Array Bounds Check Elimination (ABC) pass.
*/
bpfjit_abc_length_t abc_length;
/*
* If positive, emit "if (buflen < check_length) return 0"
* out-of-bounds check.
* Values greater than UINT32_MAX generate unconditional "return 0".
*/
bpfjit_abc_length_t check_length;
};
/*
* Additional (optimization-related) data for bpf_insn.
*/
struct bpfjit_insn_data {
/* List of jumps to this insn. */
SLIST_HEAD(, bpfjit_jump) bjumps;
union {
struct bpfjit_jump_data jdata;
struct bpfjit_read_pkt_data rdata;
} u;
bpf_memword_init_t invalid;
bool unreachable;
};
#ifdef _KERNEL
uint32_t m_xword(const struct mbuf *, uint32_t, int *);
uint32_t m_xhalf(const struct mbuf *, uint32_t, int *);
uint32_t m_xbyte(const struct mbuf *, uint32_t, int *);
MODULE(MODULE_CLASS_MISC, bpfjit, "sljit")
static int
bpfjit_modcmd(modcmd_t cmd, void *arg)
{
switch (cmd) {
case MODULE_CMD_INIT:
bpfjit_module_ops.bj_free_code = &bpfjit_free_code;
atomic_store_release(&bpfjit_module_ops.bj_generate_code,
&bpfjit_generate_code);
return 0;
case MODULE_CMD_FINI:
return EOPNOTSUPP;
default:
return ENOTTY;
}
}
#endif
/*
* Return a number of scratch registers to pass
* to sljit_emit_enter() function.
*/
static sljit_s32
nscratches(bpfjit_hint_t hints)
{
sljit_s32 rv = 2;
#ifdef _KERNEL
if (hints & BJ_HINT_PKT)
rv = 3; /* xcall with three arguments */
#endif
if (hints & BJ_HINT_IND)
rv = 3; /* uses BJ_TMP2REG */
if (hints & BJ_HINT_COP)
rv = 3; /* calls copfunc with three arguments */
if (hints & BJ_HINT_XREG)
rv = 4; /* uses BJ_XREG */
#ifdef _KERNEL
if (hints & BJ_HINT_LDX)
rv = 5; /* uses BJ_TMP3REG */
#endif
if (hints & BJ_HINT_COPX)
rv = 5; /* uses BJ_TMP3REG */
return rv;
}
static uint32_t
read_width(const struct bpf_insn *pc)
{
switch (BPF_SIZE(pc->code)) {
case BPF_W: return 4;
case BPF_H: return 2;
case BPF_B: return 1;
default: return 0;
}
}
/*
* Copy buf and buflen members of bpf_args from BJ_ARGS
* pointer to BJ_BUF and BJ_BUFLEN registers.
*/
static int
load_buf_buflen(struct sljit_compiler *compiler)
{
int status;
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
BJ_BUF, 0,
SLJIT_MEM1(BJ_ARGS),
offsetof(struct bpf_args, pkt));
if (status != SLJIT_SUCCESS)
return status;
status = sljit_emit_op1(compiler,
SLJIT_MOV, /* size_t source */
BJ_BUFLEN, 0,
SLJIT_MEM1(BJ_ARGS),
offsetof(struct bpf_args, buflen));
return status;
}
static bool
grow_jumps(struct sljit_jump ***jumps, size_t *size)
{
struct sljit_jump **newptr;
const size_t elemsz = sizeof(struct sljit_jump *);
size_t old_size = *size;
size_t new_size = 2 * old_size;
if (new_size < old_size || new_size > SIZE_MAX / elemsz)
return false;
newptr = BJ_ALLOC(new_size * elemsz);
if (newptr == NULL)
return false;
memcpy(newptr, *jumps, old_size * elemsz);
BJ_FREE(*jumps, old_size * elemsz);
*jumps = newptr;
*size = new_size;
return true;
}
static bool
append_jump(struct sljit_jump *jump, struct sljit_jump ***jumps,
size_t *size, size_t *max_size)
{
if (*size == *max_size && !grow_jumps(jumps, max_size))
return false;
(*jumps)[(*size)++] = jump;
return true;
}
/*
* Emit code for BPF_LD+BPF_B+BPF_ABS A <- P[k:1].
*/
static int
emit_read8(struct sljit_compiler *compiler, sljit_s32 src, uint32_t k)
{
return sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_AREG, 0,
SLJIT_MEM1(src), k);
}
/*
* Emit code for BPF_LD+BPF_H+BPF_ABS A <- P[k:2].
*/
static int
emit_read16(struct sljit_compiler *compiler, sljit_s32 src, uint32_t k)
{
int status;
BJ_ASSERT(k <= UINT32_MAX - 1);
/* A = buf[k]; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_AREG, 0,
SLJIT_MEM1(src), k);
if (status != SLJIT_SUCCESS)
return status;
/* tmp1 = buf[k+1]; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_TMP1REG, 0,
SLJIT_MEM1(src), k+1);
if (status != SLJIT_SUCCESS)
return status;
/* A = A << 8; */
status = sljit_emit_op2(compiler,
SLJIT_SHL,
BJ_AREG, 0,
BJ_AREG, 0,
SLJIT_IMM, 8);
if (status != SLJIT_SUCCESS)
return status;
/* A = A + tmp1; */
status = sljit_emit_op2(compiler,
SLJIT_ADD,
BJ_AREG, 0,
BJ_AREG, 0,
BJ_TMP1REG, 0);
return status;
}
/*
* Emit code for BPF_LD+BPF_W+BPF_ABS A <- P[k:4].
*/
static int
emit_read32(struct sljit_compiler *compiler, sljit_s32 src, uint32_t k)
{
int status;
BJ_ASSERT(k <= UINT32_MAX - 3);
/* A = buf[k]; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_AREG, 0,
SLJIT_MEM1(src), k);
if (status != SLJIT_SUCCESS)
return status;
/* tmp1 = buf[k+1]; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_TMP1REG, 0,
SLJIT_MEM1(src), k+1);
if (status != SLJIT_SUCCESS)
return status;
/* A = A << 8; */
status = sljit_emit_op2(compiler,
SLJIT_SHL,
BJ_AREG, 0,
BJ_AREG, 0,
SLJIT_IMM, 8);
if (status != SLJIT_SUCCESS)
return status;
/* A = A + tmp1; */
status = sljit_emit_op2(compiler,
SLJIT_ADD,
BJ_AREG, 0,
BJ_AREG, 0,
BJ_TMP1REG, 0);
if (status != SLJIT_SUCCESS)
return status;
/* tmp1 = buf[k+2]; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_TMP1REG, 0,
SLJIT_MEM1(src), k+2);
if (status != SLJIT_SUCCESS)
return status;
/* A = A << 8; */
status = sljit_emit_op2(compiler,
SLJIT_SHL,
BJ_AREG, 0,
BJ_AREG, 0,
SLJIT_IMM, 8);
if (status != SLJIT_SUCCESS)
return status;
/* A = A + tmp1; */
status = sljit_emit_op2(compiler,
SLJIT_ADD,
BJ_AREG, 0,
BJ_AREG, 0,
BJ_TMP1REG, 0);
if (status != SLJIT_SUCCESS)
return status;
/* tmp1 = buf[k+3]; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_TMP1REG, 0,
SLJIT_MEM1(src), k+3);
if (status != SLJIT_SUCCESS)
return status;
/* A = A << 8; */
status = sljit_emit_op2(compiler,
SLJIT_SHL,
BJ_AREG, 0,
BJ_AREG, 0,
SLJIT_IMM, 8);
if (status != SLJIT_SUCCESS)
return status;
/* A = A + tmp1; */
status = sljit_emit_op2(compiler,
SLJIT_ADD,
BJ_AREG, 0,
BJ_AREG, 0,
BJ_TMP1REG, 0);
return status;
}
#ifdef _KERNEL
/*
* Emit code for m_xword/m_xhalf/m_xbyte call.
*
* @pc BPF_LD+BPF_W+BPF_ABS A <- P[k:4]
* BPF_LD+BPF_H+BPF_ABS A <- P[k:2]
* BPF_LD+BPF_B+BPF_ABS A <- P[k:1]
* BPF_LD+BPF_W+BPF_IND A <- P[X+k:4]
* BPF_LD+BPF_H+BPF_IND A <- P[X+k:2]
* BPF_LD+BPF_B+BPF_IND A <- P[X+k:1]
* BPF_LDX+BPF_B+BPF_MSH X <- 4*(P[k:1]&0xf)
*/
static int
emit_xcall(struct sljit_compiler *compiler, bpfjit_hint_t hints,
const struct bpf_insn *pc, int dst, struct sljit_jump ***ret0,
size_t *ret0_size, size_t *ret0_maxsize,
uint32_t (*fn)(const struct mbuf *, uint32_t, int *))
{
#if BJ_XREG == SLJIT_RETURN_REG || \
BJ_XREG == SLJIT_R0 || \
BJ_XREG == SLJIT_R1 || \
BJ_XREG == SLJIT_R2
#error "Not supported assignment of registers."
#endif
struct sljit_jump *jump;
sljit_s32 save_reg;
int status;
save_reg = (BPF_CLASS(pc->code) == BPF_LDX) ? BJ_AREG : BJ_XREG;
if (save_reg == BJ_AREG || (hints & BJ_HINT_XREG)) {
/* save A or X */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, reg),
save_reg, 0);
if (status != SLJIT_SUCCESS)
return status;
}
/*
* Prepare registers for fn(mbuf, k, &err) call.
*/
status = sljit_emit_op1(compiler,
SLJIT_MOV,
SLJIT_R0, 0,
BJ_BUF, 0);
if (status != SLJIT_SUCCESS)
return status;
if (BPF_CLASS(pc->code) == BPF_LD && BPF_MODE(pc->code) == BPF_IND) {
if (pc->k == 0) {
/* k = X; */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
SLJIT_R1, 0,
BJ_XREG, 0);
if (status != SLJIT_SUCCESS)
return status;
} else {
/* if (X > UINT32_MAX - pc->k) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_GREATER,
BJ_XREG, 0,
SLJIT_IMM, UINT32_MAX - pc->k);
if (jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
return SLJIT_ERR_ALLOC_FAILED;
/* k = X + pc->k; */
status = sljit_emit_op2(compiler,
SLJIT_ADD,
SLJIT_R1, 0,
BJ_XREG, 0,
SLJIT_IMM, (uint32_t)pc->k);
if (status != SLJIT_SUCCESS)
return status;
}
} else {
/* k = pc->k */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
SLJIT_R1, 0,
SLJIT_IMM, (uint32_t)pc->k);
if (status != SLJIT_SUCCESS)
return status;
}
/*
* The third argument of fn is an address on stack.
*/
status = sljit_get_local_base(compiler,
SLJIT_R2, 0,
offsetof(struct bpfjit_stack, err));
if (status != SLJIT_SUCCESS)
return status;
/* fn(buf, k, &err); */
status = sljit_emit_ijump(compiler,
SLJIT_CALL3,
SLJIT_IMM, SLJIT_FUNC_OFFSET(fn));
if (status != SLJIT_SUCCESS)
return status;
if (dst != SLJIT_RETURN_REG) {
/* move return value to dst */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
dst, 0,
SLJIT_RETURN_REG, 0);
if (status != SLJIT_SUCCESS)
return status;
}
/* if (*err != 0) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_NOT_EQUAL|SLJIT_I32_OP,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, err),
SLJIT_IMM, 0);
if (jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
return SLJIT_ERR_ALLOC_FAILED;
if (save_reg == BJ_AREG || (hints & BJ_HINT_XREG)) {
/* restore A or X */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
save_reg, 0,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, reg));
if (status != SLJIT_SUCCESS)
return status;
}
return SLJIT_SUCCESS;
}
#endif
/*
* Emit code for BPF_COP and BPF_COPX instructions.
*/
static int
emit_cop(struct sljit_compiler *compiler, bpfjit_hint_t hints,
const bpf_ctx_t *bc, const struct bpf_insn *pc,
struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize)
{
#if BJ_XREG == SLJIT_RETURN_REG || \
BJ_XREG == SLJIT_R0 || \
BJ_XREG == SLJIT_R1 || \
BJ_XREG == SLJIT_R2 || \
BJ_TMP3REG == SLJIT_R0 || \
BJ_TMP3REG == SLJIT_R1 || \
BJ_TMP3REG == SLJIT_R2
#error "Not supported assignment of registers."
#endif
struct sljit_jump *jump;
sljit_s32 call_reg;
sljit_sw call_off;
int status;
BJ_ASSERT(bc != NULL && bc->copfuncs != NULL);
if (hints & BJ_HINT_LDX) {
/* save X */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, reg),
BJ_XREG, 0);
if (status != SLJIT_SUCCESS)
return status;
}
if (BPF_MISCOP(pc->code) == BPF_COP) {
call_reg = SLJIT_IMM;
call_off = SLJIT_FUNC_OFFSET(bc->copfuncs[pc->k]);
} else {
/* if (X >= bc->nfuncs) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_GREATER_EQUAL,
BJ_XREG, 0,
SLJIT_IMM, bc->nfuncs);
if (jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
return SLJIT_ERR_ALLOC_FAILED;
/* tmp1 = ctx; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
BJ_TMP1REG, 0,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, ctx));
if (status != SLJIT_SUCCESS)
return status;
/* tmp1 = ctx->copfuncs; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
BJ_TMP1REG, 0,
SLJIT_MEM1(BJ_TMP1REG),
offsetof(struct bpf_ctx, copfuncs));
if (status != SLJIT_SUCCESS)
return status;
/* tmp2 = X; */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_TMP2REG, 0,
BJ_XREG, 0);
if (status != SLJIT_SUCCESS)
return status;
/* tmp3 = ctx->copfuncs[tmp2]; */
call_reg = BJ_TMP3REG;
call_off = 0;
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
call_reg, call_off,
SLJIT_MEM2(BJ_TMP1REG, BJ_TMP2REG),
SLJIT_WORD_SHIFT);
if (status != SLJIT_SUCCESS)
return status;
}
/*
* Copy bpf_copfunc_t arguments to registers.
*/
#if BJ_AREG != SLJIT_R2
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
SLJIT_R2, 0,
BJ_AREG, 0);
if (status != SLJIT_SUCCESS)
return status;
#endif
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
SLJIT_R0, 0,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, ctx));
if (status != SLJIT_SUCCESS)
return status;
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
SLJIT_R1, 0,
BJ_ARGS, 0);
if (status != SLJIT_SUCCESS)
return status;
status = sljit_emit_ijump(compiler,
SLJIT_CALL3, call_reg, call_off);
if (status != SLJIT_SUCCESS)
return status;
#if BJ_AREG != SLJIT_RETURN_REG
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
SLJIT_RETURN_REG, 0);
if (status != SLJIT_SUCCESS)
return status;
#endif
if (hints & BJ_HINT_LDX) {
/* restore X */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
BJ_XREG, 0,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, reg));
if (status != SLJIT_SUCCESS)
return status;
}
return SLJIT_SUCCESS;
}
/*
* Generate code for
* BPF_LD+BPF_W+BPF_ABS A <- P[k:4]
* BPF_LD+BPF_H+BPF_ABS A <- P[k:2]
* BPF_LD+BPF_B+BPF_ABS A <- P[k:1]
* BPF_LD+BPF_W+BPF_IND A <- P[X+k:4]
* BPF_LD+BPF_H+BPF_IND A <- P[X+k:2]
* BPF_LD+BPF_B+BPF_IND A <- P[X+k:1]
*/
static int
emit_pkt_read(struct sljit_compiler *compiler, bpfjit_hint_t hints,
const struct bpf_insn *pc, struct sljit_jump *to_mchain_jump,
struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize)
{
int status = SLJIT_ERR_ALLOC_FAILED;
uint32_t width;
sljit_s32 ld_reg;
struct sljit_jump *jump;
#ifdef _KERNEL
struct sljit_label *label;
struct sljit_jump *over_mchain_jump;
const bool check_zero_buflen = (to_mchain_jump != NULL);
#endif
const uint32_t k = pc->k;
#ifdef _KERNEL
if (to_mchain_jump == NULL) {
to_mchain_jump = sljit_emit_cmp(compiler,
SLJIT_EQUAL,
BJ_BUFLEN, 0,
SLJIT_IMM, 0);
if (to_mchain_jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
}
#endif
ld_reg = BJ_BUF;
width = read_width(pc);
if (width == 0)
return SLJIT_ERR_ALLOC_FAILED;
if (BPF_MODE(pc->code) == BPF_IND) {
/* tmp1 = buflen - (pc->k + width); */
status = sljit_emit_op2(compiler,
SLJIT_SUB,
BJ_TMP1REG, 0,
BJ_BUFLEN, 0,
SLJIT_IMM, k + width);
if (status != SLJIT_SUCCESS)
return status;
/* ld_reg = buf + X; */
ld_reg = BJ_TMP2REG;
status = sljit_emit_op2(compiler,
SLJIT_ADD,
ld_reg, 0,
BJ_BUF, 0,
BJ_XREG, 0);
if (status != SLJIT_SUCCESS)
return status;
/* if (tmp1 < X) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_LESS,
BJ_TMP1REG, 0,
BJ_XREG, 0);
if (jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
return SLJIT_ERR_ALLOC_FAILED;
}
/*
* Don't emit wrapped-around reads. They're dead code but
* dead code elimination logic isn't smart enough to figure
* it out.
*/
if (k <= UINT32_MAX - width + 1) {
switch (width) {
case 4:
status = emit_read32(compiler, ld_reg, k);
break;
case 2:
status = emit_read16(compiler, ld_reg, k);
break;
case 1:
status = emit_read8(compiler, ld_reg, k);
break;
}
if (status != SLJIT_SUCCESS)
return status;
}
#ifdef _KERNEL
over_mchain_jump = sljit_emit_jump(compiler, SLJIT_JUMP);
if (over_mchain_jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
/* entry point to mchain handler */
label = sljit_emit_label(compiler);
if (label == NULL)
return SLJIT_ERR_ALLOC_FAILED;
sljit_set_label(to_mchain_jump, label);
if (check_zero_buflen) {
/* if (buflen != 0) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_NOT_EQUAL,
BJ_BUFLEN, 0,
SLJIT_IMM, 0);
if (jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
return SLJIT_ERR_ALLOC_FAILED;
}
switch (width) {
case 4:
status = emit_xcall(compiler, hints, pc, BJ_AREG,
ret0, ret0_size, ret0_maxsize, &m_xword);
break;
case 2:
status = emit_xcall(compiler, hints, pc, BJ_AREG,
ret0, ret0_size, ret0_maxsize, &m_xhalf);
break;
case 1:
status = emit_xcall(compiler, hints, pc, BJ_AREG,
ret0, ret0_size, ret0_maxsize, &m_xbyte);
break;
}
if (status != SLJIT_SUCCESS)
return status;
label = sljit_emit_label(compiler);
if (label == NULL)
return SLJIT_ERR_ALLOC_FAILED;
sljit_set_label(over_mchain_jump, label);
#endif
return SLJIT_SUCCESS;
}
static int
emit_memload(struct sljit_compiler *compiler,
sljit_s32 dst, uint32_t k, size_t extwords)
{
int status;
sljit_s32 src;
sljit_sw srcw;
srcw = k * sizeof(uint32_t);
if (extwords == 0) {
src = SLJIT_MEM1(SLJIT_SP);
srcw += offsetof(struct bpfjit_stack, mem);
} else {
/* copy extmem pointer to the tmp1 register */
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
BJ_TMP1REG, 0,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, extmem));
if (status != SLJIT_SUCCESS)
return status;
src = SLJIT_MEM1(BJ_TMP1REG);
}
return sljit_emit_op1(compiler, SLJIT_MOV_U32, dst, 0, src, srcw);
}
static int
emit_memstore(struct sljit_compiler *compiler,
sljit_s32 src, uint32_t k, size_t extwords)
{
int status;
sljit_s32 dst;
sljit_sw dstw;
dstw = k * sizeof(uint32_t);
if (extwords == 0) {
dst = SLJIT_MEM1(SLJIT_SP);
dstw += offsetof(struct bpfjit_stack, mem);
} else {
/* copy extmem pointer to the tmp1 register */
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
BJ_TMP1REG, 0,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, extmem));
if (status != SLJIT_SUCCESS)
return status;
dst = SLJIT_MEM1(BJ_TMP1REG);
}
return sljit_emit_op1(compiler, SLJIT_MOV_U32, dst, dstw, src, 0);
}
/*
* Emit code for BPF_LDX+BPF_B+BPF_MSH X <- 4*(P[k:1]&0xf).
*/
static int
emit_msh(struct sljit_compiler *compiler, bpfjit_hint_t hints,
const struct bpf_insn *pc, struct sljit_jump *to_mchain_jump,
struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize)
{
int status;
#ifdef _KERNEL
struct sljit_label *label;
struct sljit_jump *jump, *over_mchain_jump;
const bool check_zero_buflen = (to_mchain_jump != NULL);
#endif
const uint32_t k = pc->k;
#ifdef _KERNEL
if (to_mchain_jump == NULL) {
to_mchain_jump = sljit_emit_cmp(compiler,
SLJIT_EQUAL,
BJ_BUFLEN, 0,
SLJIT_IMM, 0);
if (to_mchain_jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
}
#endif
/* tmp1 = buf[k] */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U8,
BJ_TMP1REG, 0,
SLJIT_MEM1(BJ_BUF), k);
if (status != SLJIT_SUCCESS)
return status;
#ifdef _KERNEL
over_mchain_jump = sljit_emit_jump(compiler, SLJIT_JUMP);
if (over_mchain_jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
/* entry point to mchain handler */
label = sljit_emit_label(compiler);
if (label == NULL)
return SLJIT_ERR_ALLOC_FAILED;
sljit_set_label(to_mchain_jump, label);
if (check_zero_buflen) {
/* if (buflen != 0) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_NOT_EQUAL,
BJ_BUFLEN, 0,
SLJIT_IMM, 0);
if (jump == NULL)
return SLJIT_ERR_ALLOC_FAILED;
if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
return SLJIT_ERR_ALLOC_FAILED;
}
status = emit_xcall(compiler, hints, pc, BJ_TMP1REG,
ret0, ret0_size, ret0_maxsize, &m_xbyte);
if (status != SLJIT_SUCCESS)
return status;
label = sljit_emit_label(compiler);
if (label == NULL)
return SLJIT_ERR_ALLOC_FAILED;
sljit_set_label(over_mchain_jump, label);
#endif
/* tmp1 &= 0xf */
status = sljit_emit_op2(compiler,
SLJIT_AND,
BJ_TMP1REG, 0,
BJ_TMP1REG, 0,
SLJIT_IMM, 0xf);
if (status != SLJIT_SUCCESS)
return status;
/* X = tmp1 << 2 */
status = sljit_emit_op2(compiler,
SLJIT_SHL,
BJ_XREG, 0,
BJ_TMP1REG, 0,
SLJIT_IMM, 2);
if (status != SLJIT_SUCCESS)
return status;
return SLJIT_SUCCESS;
}
/*
* Emit code for A = A / k or A = A % k when k is a power of 2.
* @pc BPF_DIV or BPF_MOD instruction.
*/
static int
emit_pow2_moddiv(struct sljit_compiler *compiler, const struct bpf_insn *pc)
{
uint32_t k = pc->k;
int status = SLJIT_SUCCESS;
BJ_ASSERT(k != 0 && (k & (k - 1)) == 0);
if (BPF_OP(pc->code) == BPF_MOD) {
status = sljit_emit_op2(compiler,
SLJIT_AND,
BJ_AREG, 0,
BJ_AREG, 0,
SLJIT_IMM, k - 1);
} else {
int shift = 0;
/*
* Do shift = __builtin_ctz(k).
* The loop is slower, but that's ok.
*/
while (k > 1) {
k >>= 1;
shift++;
}
if (shift != 0) {
status = sljit_emit_op2(compiler,
SLJIT_LSHR|SLJIT_I32_OP,
BJ_AREG, 0,
BJ_AREG, 0,
SLJIT_IMM, shift);
}
}
return status;
}
#if !defined(BPFJIT_USE_UDIV)
static sljit_uw
divide(sljit_uw x, sljit_uw y)
{
return (uint32_t)x / (uint32_t)y;
}
static sljit_uw
modulus(sljit_uw x, sljit_uw y)
{
return (uint32_t)x % (uint32_t)y;
}
#endif
/*
* Emit code for A = A / div or A = A % div.
* @pc BPF_DIV or BPF_MOD instruction.
*/
static int
emit_moddiv(struct sljit_compiler *compiler, const struct bpf_insn *pc)
{
int status;
const bool xdiv = BPF_OP(pc->code) == BPF_DIV;
const bool xreg = BPF_SRC(pc->code) == BPF_X;
#if BJ_XREG == SLJIT_RETURN_REG || \
BJ_XREG == SLJIT_R0 || \
BJ_XREG == SLJIT_R1 || \
BJ_AREG == SLJIT_R1
#error "Not supported assignment of registers."
#endif
#if BJ_AREG != SLJIT_R0
status = sljit_emit_op1(compiler,
SLJIT_MOV,
SLJIT_R0, 0,
BJ_AREG, 0);
if (status != SLJIT_SUCCESS)
return status;
#endif
status = sljit_emit_op1(compiler,
SLJIT_MOV,
SLJIT_R1, 0,
xreg ? BJ_XREG : SLJIT_IMM,
xreg ? 0 : (uint32_t)pc->k);
if (status != SLJIT_SUCCESS)
return status;
#if defined(BPFJIT_USE_UDIV)
status = sljit_emit_op0(compiler, SLJIT_UDIV|SLJIT_I32_OP);
if (BPF_OP(pc->code) == BPF_DIV) {
#if BJ_AREG != SLJIT_R0
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
SLJIT_R0, 0);
#endif
} else {
#if BJ_AREG != SLJIT_R1
/* Remainder is in SLJIT_R1. */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
SLJIT_R1, 0);
#endif
}
if (status != SLJIT_SUCCESS)
return status;
#else
status = sljit_emit_ijump(compiler,
SLJIT_CALL2,
SLJIT_IMM, xdiv ? SLJIT_FUNC_OFFSET(divide) :
SLJIT_FUNC_OFFSET(modulus));
#if BJ_AREG != SLJIT_RETURN_REG
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
SLJIT_RETURN_REG, 0);
if (status != SLJIT_SUCCESS)
return status;
#endif
#endif
return status;
}
/*
* Return true if pc is a "read from packet" instruction.
* If length is not NULL and return value is true, *length will
* be set to a safe length required to read a packet.
*/
static bool
read_pkt_insn(const struct bpf_insn *pc, bpfjit_abc_length_t *length)
{
bool rv;
bpfjit_abc_length_t width = 0; /* XXXuninit */
switch (BPF_CLASS(pc->code)) {
default:
rv = false;
break;
case BPF_LD:
rv = BPF_MODE(pc->code) == BPF_ABS ||
BPF_MODE(pc->code) == BPF_IND;
if (rv) {
width = read_width(pc);
rv = (width != 0);
}
break;
case BPF_LDX:
rv = BPF_MODE(pc->code) == BPF_MSH &&
BPF_SIZE(pc->code) == BPF_B;
width = 1;
break;
}
if (rv && length != NULL) {
/*
* Values greater than UINT32_MAX will generate
* unconditional "return 0".
*/
*length = (uint32_t)pc->k + width;
}
return rv;
}
static void
optimize_init(struct bpfjit_insn_data *insn_dat, size_t insn_count)
{
size_t i;
for (i = 0; i < insn_count; i++) {
SLIST_INIT(&insn_dat[i].bjumps);
insn_dat[i].invalid = BJ_INIT_NOBITS;
}
}
/*
* The function divides instructions into blocks. Destination of a jump
* instruction starts a new block. BPF_RET and BPF_JMP instructions
* terminate a block. Blocks are linear, that is, there are no jumps out
* from the middle of a block and there are no jumps in to the middle of
* a block.
*
* The function also sets bits in *initmask for memwords that
* need to be initialized to zero. Note that this set should be empty
* for any valid kernel filter program.
*/
static bool
optimize_pass1(const bpf_ctx_t *bc, const struct bpf_insn *insns,
struct bpfjit_insn_data *insn_dat, size_t insn_count,
bpf_memword_init_t *initmask, bpfjit_hint_t *hints)
{
struct bpfjit_jump *jtf;
size_t i;
uint32_t jt, jf;
bpfjit_abc_length_t length;
bpf_memword_init_t invalid; /* borrowed from bpf_filter() */
bool unreachable;
const size_t memwords = GET_MEMWORDS(bc);
*hints = 0;
*initmask = BJ_INIT_NOBITS;
unreachable = false;
invalid = ~BJ_INIT_NOBITS;
for (i = 0; i < insn_count; i++) {
if (!SLIST_EMPTY(&insn_dat[i].bjumps))
unreachable = false;
insn_dat[i].unreachable = unreachable;
if (unreachable)
continue;
invalid |= insn_dat[i].invalid;
if (read_pkt_insn(&insns[i], &length) && length > UINT32_MAX)
unreachable = true;
switch (BPF_CLASS(insns[i].code)) {
case BPF_RET:
if (BPF_RVAL(insns[i].code) == BPF_A)
*initmask |= invalid & BJ_INIT_ABIT;
unreachable = true;
continue;
case BPF_LD:
if (BPF_MODE(insns[i].code) == BPF_ABS)
*hints |= BJ_HINT_ABS;
if (BPF_MODE(insns[i].code) == BPF_IND) {
*hints |= BJ_HINT_IND | BJ_HINT_XREG;
*initmask |= invalid & BJ_INIT_XBIT;
}
if (BPF_MODE(insns[i].code) == BPF_MEM &&
(uint32_t)insns[i].k < memwords) {
*initmask |= invalid & BJ_INIT_MBIT(insns[i].k);
}
invalid &= ~BJ_INIT_ABIT;
continue;
case BPF_LDX:
*hints |= BJ_HINT_XREG | BJ_HINT_LDX;
if (BPF_MODE(insns[i].code) == BPF_MEM &&
(uint32_t)insns[i].k < memwords) {
*initmask |= invalid & BJ_INIT_MBIT(insns[i].k);
}
if (BPF_MODE(insns[i].code) == BPF_MSH &&
BPF_SIZE(insns[i].code) == BPF_B) {
*hints |= BJ_HINT_MSH;
}
invalid &= ~BJ_INIT_XBIT;
continue;
case BPF_ST:
*initmask |= invalid & BJ_INIT_ABIT;
if ((uint32_t)insns[i].k < memwords)
invalid &= ~BJ_INIT_MBIT(insns[i].k);
continue;
case BPF_STX:
*hints |= BJ_HINT_XREG;
*initmask |= invalid & BJ_INIT_XBIT;
if ((uint32_t)insns[i].k < memwords)
invalid &= ~BJ_INIT_MBIT(insns[i].k);
continue;
case BPF_ALU:
*initmask |= invalid & BJ_INIT_ABIT;
if (insns[i].code != (BPF_ALU|BPF_NEG) &&
BPF_SRC(insns[i].code) == BPF_X) {
*hints |= BJ_HINT_XREG;
*initmask |= invalid & BJ_INIT_XBIT;
}
invalid &= ~BJ_INIT_ABIT;
continue;
case BPF_MISC:
switch (BPF_MISCOP(insns[i].code)) {
case BPF_TAX: // X <- A
*hints |= BJ_HINT_XREG;
*initmask |= invalid & BJ_INIT_ABIT;
invalid &= ~BJ_INIT_XBIT;
continue;
case BPF_TXA: // A <- X
*hints |= BJ_HINT_XREG;
*initmask |= invalid & BJ_INIT_XBIT;
invalid &= ~BJ_INIT_ABIT;
continue;
case BPF_COPX:
*hints |= BJ_HINT_XREG | BJ_HINT_COPX;
/* FALLTHROUGH */
case BPF_COP:
*hints |= BJ_HINT_COP;
*initmask |= invalid & BJ_INIT_ABIT;
invalid &= ~BJ_INIT_ABIT;
continue;
}
continue;
case BPF_JMP:
/* Initialize abc_length for ABC pass. */
insn_dat[i].u.jdata.abc_length = MAX_ABC_LENGTH;
*initmask |= invalid & BJ_INIT_ABIT;
if (BPF_SRC(insns[i].code) == BPF_X) {
*hints |= BJ_HINT_XREG;
*initmask |= invalid & BJ_INIT_XBIT;
}
if (BPF_OP(insns[i].code) == BPF_JA) {
jt = jf = insns[i].k;
} else {
jt = insns[i].jt;
jf = insns[i].jf;
}
if (jt >= insn_count - (i + 1) ||
jf >= insn_count - (i + 1)) {
return false;
}
if (jt > 0 && jf > 0)
unreachable = true;
jt += i + 1;
jf += i + 1;
jtf = insn_dat[i].u.jdata.jtf;
jtf[0].jdata = &insn_dat[i].u.jdata;
SLIST_INSERT_HEAD(&insn_dat[jt].bjumps,
&jtf[0], entries);
if (jf != jt) {
jtf[1].jdata = &insn_dat[i].u.jdata;
SLIST_INSERT_HEAD(&insn_dat[jf].bjumps,
&jtf[1], entries);
}
insn_dat[jf].invalid |= invalid;
insn_dat[jt].invalid |= invalid;
invalid = 0;
continue;
}
}
return true;
}
/*
* Array Bounds Check Elimination (ABC) pass.
*/
static void
optimize_pass2(const bpf_ctx_t *bc, const struct bpf_insn *insns,
struct bpfjit_insn_data *insn_dat, size_t insn_count)
{
struct bpfjit_jump *jmp;
const struct bpf_insn *pc;
struct bpfjit_insn_data *pd;
size_t i;
bpfjit_abc_length_t length, abc_length = 0;
const size_t extwords = GET_EXTWORDS(bc);
for (i = insn_count; i != 0; i--) {
pc = &insns[i-1];
pd = &insn_dat[i-1];
if (pd->unreachable)
continue;
switch (BPF_CLASS(pc->code)) {
case BPF_RET:
/*
* It's quite common for bpf programs to
* check packet bytes in increasing order
* and return zero if bytes don't match
* specified critetion. Such programs disable
* ABC optimization completely because for
* every jump there is a branch with no read
* instruction.
* With no side effects, BPF_STMT(BPF_RET+BPF_K, 0)
* is indistinguishable from out-of-bound load.
* Therefore, abc_length can be set to
* MAX_ABC_LENGTH and enable ABC for many
* bpf programs.
* If this optimization encounters any
* instruction with a side effect, it will
* reset abc_length.
*/
if (BPF_RVAL(pc->code) == BPF_K && pc->k == 0)
abc_length = MAX_ABC_LENGTH;
else
abc_length = 0;
break;
case BPF_MISC:
if (BPF_MISCOP(pc->code) == BPF_COP ||
BPF_MISCOP(pc->code) == BPF_COPX) {
/* COP instructions can have side effects. */
abc_length = 0;
}
break;
case BPF_ST:
case BPF_STX:
if (extwords != 0) {
/* Write to memory is visible after a call. */
abc_length = 0;
}
break;
case BPF_JMP:
abc_length = pd->u.jdata.abc_length;
break;
default:
if (read_pkt_insn(pc, &length)) {
if (abc_length < length)
abc_length = length;
pd->u.rdata.abc_length = abc_length;
}
break;
}
SLIST_FOREACH(jmp, &pd->bjumps, entries) {
if (jmp->jdata->abc_length > abc_length)
jmp->jdata->abc_length = abc_length;
}
}
}
static void
optimize_pass3(const struct bpf_insn *insns,
struct bpfjit_insn_data *insn_dat, size_t insn_count)
{
struct bpfjit_jump *jmp;
size_t i;
bpfjit_abc_length_t checked_length = 0;
for (i = 0; i < insn_count; i++) {
if (insn_dat[i].unreachable)
continue;
SLIST_FOREACH(jmp, &insn_dat[i].bjumps, entries) {
if (jmp->jdata->checked_length < checked_length)
checked_length = jmp->jdata->checked_length;
}
if (BPF_CLASS(insns[i].code) == BPF_JMP) {
insn_dat[i].u.jdata.checked_length = checked_length;
} else if (read_pkt_insn(&insns[i], NULL)) {
struct bpfjit_read_pkt_data *rdata =
&insn_dat[i].u.rdata;
rdata->check_length = 0;
if (checked_length < rdata->abc_length) {
checked_length = rdata->abc_length;
rdata->check_length = checked_length;
}
}
}
}
static bool
optimize(const bpf_ctx_t *bc, const struct bpf_insn *insns,
struct bpfjit_insn_data *insn_dat, size_t insn_count,
bpf_memword_init_t *initmask, bpfjit_hint_t *hints)
{
optimize_init(insn_dat, insn_count);
if (!optimize_pass1(bc, insns, insn_dat, insn_count, initmask, hints))
return false;
optimize_pass2(bc, insns, insn_dat, insn_count);
optimize_pass3(insns, insn_dat, insn_count);
return true;
}
/*
* Convert BPF_ALU operations except BPF_NEG and BPF_DIV to sljit operation.
*/
static bool
alu_to_op(const struct bpf_insn *pc, int *res)
{
const uint32_t k = pc->k;
/*
* Note: all supported 64bit arches have 32bit multiply
* instruction so SLJIT_I32_OP doesn't have any overhead.
*/
switch (BPF_OP(pc->code)) {
case BPF_ADD:
*res = SLJIT_ADD;
return true;
case BPF_SUB:
*res = SLJIT_SUB;
return true;
case BPF_MUL:
*res = SLJIT_MUL|SLJIT_I32_OP;
return true;
case BPF_OR:
*res = SLJIT_OR;
return true;
case BPF_XOR:
*res = SLJIT_XOR;
return true;
case BPF_AND:
*res = SLJIT_AND;
return true;
case BPF_LSH:
*res = SLJIT_SHL;
return k < 32;
case BPF_RSH:
*res = SLJIT_LSHR|SLJIT_I32_OP;
return k < 32;
default:
return false;
}
}
/*
* Convert BPF_JMP operations except BPF_JA to sljit condition.
*/
static bool
jmp_to_cond(const struct bpf_insn *pc, bool negate, int *res)
{
/*
* Note: all supported 64bit arches have 32bit comparison
* instructions so SLJIT_I32_OP doesn't have any overhead.
*/
*res = SLJIT_I32_OP;
switch (BPF_OP(pc->code)) {
case BPF_JGT:
*res |= negate ? SLJIT_LESS_EQUAL : SLJIT_GREATER;
return true;
case BPF_JGE:
*res |= negate ? SLJIT_LESS : SLJIT_GREATER_EQUAL;
return true;
case BPF_JEQ:
*res |= negate ? SLJIT_NOT_EQUAL : SLJIT_EQUAL;
return true;
case BPF_JSET:
*res |= negate ? SLJIT_EQUAL : SLJIT_NOT_EQUAL;
return true;
default:
return false;
}
}
/*
* Convert BPF_K and BPF_X to sljit register.
*/
static int
kx_to_reg(const struct bpf_insn *pc)
{
switch (BPF_SRC(pc->code)) {
case BPF_K: return SLJIT_IMM;
case BPF_X: return BJ_XREG;
default:
BJ_ASSERT(false);
return 0;
}
}
static sljit_sw
kx_to_reg_arg(const struct bpf_insn *pc)
{
switch (BPF_SRC(pc->code)) {
case BPF_K: return (uint32_t)pc->k; /* SLJIT_IMM, pc->k, */
case BPF_X: return 0; /* BJ_XREG, 0, */
default:
BJ_ASSERT(false);
return 0;
}
}
static bool
generate_insn_code(struct sljit_compiler *compiler, bpfjit_hint_t hints,
const bpf_ctx_t *bc, const struct bpf_insn *insns,
struct bpfjit_insn_data *insn_dat, size_t insn_count)
{
/* a list of jumps to out-of-bound return from a generated function */
struct sljit_jump **ret0;
size_t ret0_size, ret0_maxsize;
struct sljit_jump *jump;
struct sljit_label *label;
const struct bpf_insn *pc;
struct bpfjit_jump *bjump, *jtf;
struct sljit_jump *to_mchain_jump;
size_t i;
unsigned int rval, mode, src, op;
int branching, negate;
int status, cond, op2;
uint32_t jt, jf;
bool unconditional_ret;
bool rv;
const size_t extwords = GET_EXTWORDS(bc);
const size_t memwords = GET_MEMWORDS(bc);
ret0 = NULL;
rv = false;
ret0_size = 0;
ret0_maxsize = 64;
ret0 = BJ_ALLOC(ret0_maxsize * sizeof(ret0[0]));
if (ret0 == NULL)
goto fail;
/* reset sjump members of jdata */
for (i = 0; i < insn_count; i++) {
if (insn_dat[i].unreachable ||
BPF_CLASS(insns[i].code) != BPF_JMP) {
continue;
}
jtf = insn_dat[i].u.jdata.jtf;
jtf[0].sjump = jtf[1].sjump = NULL;
}
/* main loop */
for (i = 0; i < insn_count; i++) {
if (insn_dat[i].unreachable)
continue;
/*
* Resolve jumps to the current insn.
*/
label = NULL;
SLIST_FOREACH(bjump, &insn_dat[i].bjumps, entries) {
if (bjump->sjump != NULL) {
if (label == NULL)
label = sljit_emit_label(compiler);
if (label == NULL)
goto fail;
sljit_set_label(bjump->sjump, label);
}
}
to_mchain_jump = NULL;
unconditional_ret = false;
if (read_pkt_insn(&insns[i], NULL)) {
if (insn_dat[i].u.rdata.check_length > UINT32_MAX) {
/* Jump to "return 0" unconditionally. */
unconditional_ret = true;
jump = sljit_emit_jump(compiler, SLJIT_JUMP);
if (jump == NULL)
goto fail;
if (!append_jump(jump, &ret0,
&ret0_size, &ret0_maxsize))
goto fail;
} else if (insn_dat[i].u.rdata.check_length > 0) {
/* if (buflen < check_length) return 0; */
jump = sljit_emit_cmp(compiler,
SLJIT_LESS,
BJ_BUFLEN, 0,
SLJIT_IMM,
insn_dat[i].u.rdata.check_length);
if (jump == NULL)
goto fail;
#ifdef _KERNEL
to_mchain_jump = jump;
#else
if (!append_jump(jump, &ret0,
&ret0_size, &ret0_maxsize))
goto fail;
#endif
}
}
pc = &insns[i];
switch (BPF_CLASS(pc->code)) {
default:
goto fail;
case BPF_LD:
/* BPF_LD+BPF_IMM A <- k */
if (pc->code == (BPF_LD|BPF_IMM)) {
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
SLJIT_IMM, (uint32_t)pc->k);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
/* BPF_LD+BPF_MEM A <- M[k] */
if (pc->code == (BPF_LD|BPF_MEM)) {
if ((uint32_t)pc->k >= memwords)
goto fail;
status = emit_memload(compiler,
BJ_AREG, pc->k, extwords);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
/* BPF_LD+BPF_W+BPF_LEN A <- len */
if (pc->code == (BPF_LD|BPF_W|BPF_LEN)) {
status = sljit_emit_op1(compiler,
SLJIT_MOV, /* size_t source */
BJ_AREG, 0,
SLJIT_MEM1(BJ_ARGS),
offsetof(struct bpf_args, wirelen));
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
mode = BPF_MODE(pc->code);
if (mode != BPF_ABS && mode != BPF_IND)
goto fail;
if (unconditional_ret)
continue;
status = emit_pkt_read(compiler, hints, pc,
to_mchain_jump, &ret0, &ret0_size, &ret0_maxsize);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
case BPF_LDX:
mode = BPF_MODE(pc->code);
/* BPF_LDX+BPF_W+BPF_IMM X <- k */
if (mode == BPF_IMM) {
if (BPF_SIZE(pc->code) != BPF_W)
goto fail;
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_XREG, 0,
SLJIT_IMM, (uint32_t)pc->k);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
/* BPF_LDX+BPF_W+BPF_LEN X <- len */
if (mode == BPF_LEN) {
if (BPF_SIZE(pc->code) != BPF_W)
goto fail;
status = sljit_emit_op1(compiler,
SLJIT_MOV, /* size_t source */
BJ_XREG, 0,
SLJIT_MEM1(BJ_ARGS),
offsetof(struct bpf_args, wirelen));
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
/* BPF_LDX+BPF_W+BPF_MEM X <- M[k] */
if (mode == BPF_MEM) {
if (BPF_SIZE(pc->code) != BPF_W)
goto fail;
if ((uint32_t)pc->k >= memwords)
goto fail;
status = emit_memload(compiler,
BJ_XREG, pc->k, extwords);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
/* BPF_LDX+BPF_B+BPF_MSH X <- 4*(P[k:1]&0xf) */
if (mode != BPF_MSH || BPF_SIZE(pc->code) != BPF_B)
goto fail;
if (unconditional_ret)
continue;
status = emit_msh(compiler, hints, pc,
to_mchain_jump, &ret0, &ret0_size, &ret0_maxsize);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
case BPF_ST:
if (pc->code != BPF_ST ||
(uint32_t)pc->k >= memwords) {
goto fail;
}
status = emit_memstore(compiler,
BJ_AREG, pc->k, extwords);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
case BPF_STX:
if (pc->code != BPF_STX ||
(uint32_t)pc->k >= memwords) {
goto fail;
}
status = emit_memstore(compiler,
BJ_XREG, pc->k, extwords);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
case BPF_ALU:
if (pc->code == (BPF_ALU|BPF_NEG)) {
status = sljit_emit_op1(compiler,
SLJIT_NEG,
BJ_AREG, 0,
BJ_AREG, 0);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
op = BPF_OP(pc->code);
if (op != BPF_DIV && op != BPF_MOD) {
if (!alu_to_op(pc, &op2))
goto fail;
status = sljit_emit_op2(compiler,
op2, BJ_AREG, 0, BJ_AREG, 0,
kx_to_reg(pc), kx_to_reg_arg(pc));
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
/* BPF_DIV/BPF_MOD */
src = BPF_SRC(pc->code);
if (src != BPF_X && src != BPF_K)
goto fail;
/* division by zero? */
if (src == BPF_X) {
jump = sljit_emit_cmp(compiler,
SLJIT_EQUAL|SLJIT_I32_OP,
BJ_XREG, 0,
SLJIT_IMM, 0);
if (jump == NULL)
goto fail;
if (!append_jump(jump, &ret0,
&ret0_size, &ret0_maxsize))
goto fail;
} else if (pc->k == 0) {
jump = sljit_emit_jump(compiler, SLJIT_JUMP);
if (jump == NULL)
goto fail;
if (!append_jump(jump, &ret0,
&ret0_size, &ret0_maxsize))
goto fail;
}
if (src == BPF_X) {
status = emit_moddiv(compiler, pc);
if (status != SLJIT_SUCCESS)
goto fail;
} else if (pc->k != 0) {
if (pc->k & (pc->k - 1)) {
status = emit_moddiv(compiler, pc);
} else {
status = emit_pow2_moddiv(compiler, pc);
}
if (status != SLJIT_SUCCESS)
goto fail;
}
continue;
case BPF_JMP:
op = BPF_OP(pc->code);
if (op == BPF_JA) {
jt = jf = pc->k;
} else {
jt = pc->jt;
jf = pc->jf;
}
negate = (jt == 0) ? 1 : 0;
branching = (jt == jf) ? 0 : 1;
jtf = insn_dat[i].u.jdata.jtf;
if (branching) {
if (op != BPF_JSET) {
if (!jmp_to_cond(pc, negate, &cond))
goto fail;
jump = sljit_emit_cmp(compiler,
cond, BJ_AREG, 0,
kx_to_reg(pc), kx_to_reg_arg(pc));
} else {
status = sljit_emit_op2(compiler,
SLJIT_AND,
BJ_TMP1REG, 0,
BJ_AREG, 0,
kx_to_reg(pc), kx_to_reg_arg(pc));
if (status != SLJIT_SUCCESS)
goto fail;
if (!jmp_to_cond(pc, negate, &cond))
goto fail;
jump = sljit_emit_cmp(compiler,
cond, BJ_TMP1REG, 0, SLJIT_IMM, 0);
}
if (jump == NULL)
goto fail;
BJ_ASSERT(jtf[negate].sjump == NULL);
jtf[negate].sjump = jump;
}
if (!branching || (jt != 0 && jf != 0)) {
jump = sljit_emit_jump(compiler, SLJIT_JUMP);
if (jump == NULL)
goto fail;
BJ_ASSERT(jtf[branching].sjump == NULL);
jtf[branching].sjump = jump;
}
continue;
case BPF_RET:
rval = BPF_RVAL(pc->code);
if (rval == BPF_X)
goto fail;
/* BPF_RET+BPF_K accept k bytes */
if (rval == BPF_K) {
status = sljit_emit_return(compiler,
SLJIT_MOV_U32,
SLJIT_IMM, (uint32_t)pc->k);
if (status != SLJIT_SUCCESS)
goto fail;
}
/* BPF_RET+BPF_A accept A bytes */
if (rval == BPF_A) {
status = sljit_emit_return(compiler,
SLJIT_MOV_U32,
BJ_AREG, 0);
if (status != SLJIT_SUCCESS)
goto fail;
}
continue;
case BPF_MISC:
switch (BPF_MISCOP(pc->code)) {
case BPF_TAX:
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
BJ_XREG, 0,
BJ_AREG, 0);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
case BPF_TXA:
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
BJ_XREG, 0);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
case BPF_COP:
case BPF_COPX:
if (bc == NULL || bc->copfuncs == NULL)
goto fail;
if (BPF_MISCOP(pc->code) == BPF_COP &&
(uint32_t)pc->k >= bc->nfuncs) {
goto fail;
}
status = emit_cop(compiler, hints, bc, pc,
&ret0, &ret0_size, &ret0_maxsize);
if (status != SLJIT_SUCCESS)
goto fail;
continue;
}
goto fail;
} /* switch */
} /* main loop */
BJ_ASSERT(ret0_size <= ret0_maxsize);
if (ret0_size > 0) {
label = sljit_emit_label(compiler);
if (label == NULL)
goto fail;
for (i = 0; i < ret0_size; i++)
sljit_set_label(ret0[i], label);
}
status = sljit_emit_return(compiler,
SLJIT_MOV_U32,
SLJIT_IMM, 0);
if (status != SLJIT_SUCCESS)
goto fail;
rv = true;
fail:
if (ret0 != NULL)
BJ_FREE(ret0, ret0_maxsize * sizeof(ret0[0]));
return rv;
}
bpfjit_func_t
bpfjit_generate_code(const bpf_ctx_t *bc,
const struct bpf_insn *insns, size_t insn_count)
{
void *rv;
struct sljit_compiler *compiler;
size_t i;
int status;
/* optimization related */
bpf_memword_init_t initmask;
bpfjit_hint_t hints;
/* memory store location for initial zero initialization */
sljit_s32 mem_reg;
sljit_sw mem_off;
struct bpfjit_insn_data *insn_dat;
const size_t extwords = GET_EXTWORDS(bc);
const size_t memwords = GET_MEMWORDS(bc);
const bpf_memword_init_t preinited = extwords ? bc->preinited : 0;
rv = NULL;
compiler = NULL;
insn_dat = NULL;
if (memwords > MAX_MEMWORDS)
goto fail;
if (insn_count == 0 || insn_count > SIZE_MAX / sizeof(insn_dat[0]))
goto fail;
insn_dat = BJ_ALLOC(insn_count * sizeof(insn_dat[0]));
if (insn_dat == NULL)
goto fail;
if (!optimize(bc, insns, insn_dat, insn_count, &initmask, &hints))
goto fail;
compiler = sljit_create_compiler(NULL);
if (compiler == NULL)
goto fail;
#if !defined(_KERNEL) && defined(SLJIT_VERBOSE) && SLJIT_VERBOSE
sljit_compiler_verbose(compiler, stderr);
#endif
status = sljit_emit_enter(compiler, 0, 2, nscratches(hints),
NSAVEDS, 0, 0, sizeof(struct bpfjit_stack));
if (status != SLJIT_SUCCESS)
goto fail;
if (hints & BJ_HINT_COP) {
/* save ctx argument */
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, ctx),
BJ_CTX_ARG, 0);
if (status != SLJIT_SUCCESS)
goto fail;
}
if (extwords == 0) {
mem_reg = SLJIT_MEM1(SLJIT_SP);
mem_off = offsetof(struct bpfjit_stack, mem);
} else {
/* copy "mem" argument from bpf_args to bpfjit_stack */
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
BJ_TMP1REG, 0,
SLJIT_MEM1(BJ_ARGS), offsetof(struct bpf_args, mem));
if (status != SLJIT_SUCCESS)
goto fail;
status = sljit_emit_op1(compiler,
SLJIT_MOV_P,
SLJIT_MEM1(SLJIT_SP),
offsetof(struct bpfjit_stack, extmem),
BJ_TMP1REG, 0);
if (status != SLJIT_SUCCESS)
goto fail;
mem_reg = SLJIT_MEM1(BJ_TMP1REG);
mem_off = 0;
}
/*
* Exclude pre-initialised external memory words but keep
* initialization statuses of A and X registers in case
* bc->preinited wrongly sets those two bits.
*/
initmask &= ~preinited | BJ_INIT_ABIT | BJ_INIT_XBIT;
#if defined(_KERNEL)
/* bpf_filter() checks initialization of memwords. */
BJ_ASSERT((initmask & (BJ_INIT_MBIT(memwords) - 1)) == 0);
#endif
for (i = 0; i < memwords; i++) {
if (initmask & BJ_INIT_MBIT(i)) {
/* M[i] = 0; */
status = sljit_emit_op1(compiler,
SLJIT_MOV_U32,
mem_reg, mem_off + i * sizeof(uint32_t),
SLJIT_IMM, 0);
if (status != SLJIT_SUCCESS)
goto fail;
}
}
if (initmask & BJ_INIT_ABIT) {
/* A = 0; */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_AREG, 0,
SLJIT_IMM, 0);
if (status != SLJIT_SUCCESS)
goto fail;
}
if (initmask & BJ_INIT_XBIT) {
/* X = 0; */
status = sljit_emit_op1(compiler,
SLJIT_MOV,
BJ_XREG, 0,
SLJIT_IMM, 0);
if (status != SLJIT_SUCCESS)
goto fail;
}
status = load_buf_buflen(compiler);
if (status != SLJIT_SUCCESS)
goto fail;
if (!generate_insn_code(compiler, hints,
bc, insns, insn_dat, insn_count)) {
goto fail;
}
rv = sljit_generate_code(compiler);
fail:
if (compiler != NULL)
sljit_free_compiler(compiler);
if (insn_dat != NULL)
BJ_FREE(insn_dat, insn_count * sizeof(insn_dat[0]));
return (bpfjit_func_t)rv;
}
void
bpfjit_free_code(bpfjit_func_t code)
{
sljit_free_code((void *)code);
}