Home | History | Annotate | Line # | Download | only in net
bpfjit.c revision 1.28
      1 /*	$NetBSD: bpfjit.c,v 1.28 2014/07/13 21:54:46 alnsn Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2011-2014 Alexander Nasonov.
      5  * All rights reserved.
      6  *
      7  * Redistribution and use in source and binary forms, with or without
      8  * modification, are permitted provided that the following conditions
      9  * are met:
     10  *
     11  * 1. Redistributions of source code must retain the above copyright
     12  *    notice, this list of conditions and the following disclaimer.
     13  * 2. Redistributions in binary form must reproduce the above copyright
     14  *    notice, this list of conditions and the following disclaimer in
     15  *    the documentation and/or other materials provided with the
     16  *    distribution.
     17  *
     18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
     19  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
     20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
     21  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
     22  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
     23  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
     24  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
     25  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
     26  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
     27  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
     28  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     29  * SUCH DAMAGE.
     30  */
     31 
     32 #include <sys/cdefs.h>
     33 #ifdef _KERNEL
     34 __KERNEL_RCSID(0, "$NetBSD: bpfjit.c,v 1.28 2014/07/13 21:54:46 alnsn Exp $");
     35 #else
     36 __RCSID("$NetBSD: bpfjit.c,v 1.28 2014/07/13 21:54:46 alnsn Exp $");
     37 #endif
     38 
     39 #include <sys/types.h>
     40 #include <sys/queue.h>
     41 
     42 #ifndef _KERNEL
     43 #include <assert.h>
     44 #define BJ_ASSERT(c) assert(c)
     45 #else
     46 #define BJ_ASSERT(c) KASSERT(c)
     47 #endif
     48 
     49 #ifndef _KERNEL
     50 #include <stdlib.h>
     51 #define BJ_ALLOC(sz) malloc(sz)
     52 #define BJ_FREE(p, sz) free(p)
     53 #else
     54 #include <sys/kmem.h>
     55 #define BJ_ALLOC(sz) kmem_alloc(sz, KM_SLEEP)
     56 #define BJ_FREE(p, sz) kmem_free(p, sz)
     57 #endif
     58 
     59 #ifndef _KERNEL
     60 #include <limits.h>
     61 #include <stdbool.h>
     62 #include <stddef.h>
     63 #include <stdint.h>
     64 #else
     65 #include <sys/atomic.h>
     66 #include <sys/module.h>
     67 #endif
     68 
     69 #define	__BPF_PRIVATE
     70 #include <net/bpf.h>
     71 #include <net/bpfjit.h>
     72 #include <sljitLir.h>
     73 
     74 #if !defined(_KERNEL) && defined(SLJIT_VERBOSE) && SLJIT_VERBOSE
     75 #include <stdio.h> /* for stderr */
     76 #endif
     77 
     78 /*
     79  * Arguments of generated bpfjit_func_t.
     80  * The first argument is reassigned upon entry
     81  * to a more frequently used buf argument.
     82  */
     83 #define BJ_CTX_ARG	SLJIT_SAVED_REG1
     84 #define BJ_ARGS		SLJIT_SAVED_REG2
     85 
     86 /*
     87  * Permanent register assignments.
     88  */
     89 #define BJ_BUF		SLJIT_SAVED_REG1
     90 //#define BJ_ARGS	SLJIT_SAVED_REG2
     91 #define BJ_BUFLEN	SLJIT_SAVED_REG3
     92 #define BJ_AREG		SLJIT_SCRATCH_REG1
     93 #define BJ_TMP1REG	SLJIT_SCRATCH_REG2
     94 #define BJ_TMP2REG	SLJIT_SCRATCH_REG3
     95 #define BJ_XREG		SLJIT_TEMPORARY_EREG1
     96 #define BJ_TMP3REG	SLJIT_TEMPORARY_EREG2
     97 
     98 #ifdef _KERNEL
     99 #define MAX_MEMWORDS BPF_MAX_MEMWORDS
    100 #else
    101 #define MAX_MEMWORDS BPF_MEMWORDS
    102 #endif
    103 
    104 #define BJ_INIT_NOBITS  ((bpf_memword_init_t)0)
    105 #define BJ_INIT_MBIT(k) BPF_MEMWORD_INIT(k)
    106 #define BJ_INIT_ABIT    BJ_INIT_MBIT(MAX_MEMWORDS)
    107 #define BJ_INIT_XBIT    BJ_INIT_MBIT(MAX_MEMWORDS + 1)
    108 
    109 /*
    110  * Get a number of memwords and external memwords from a bpf_ctx object.
    111  */
    112 #define GET_EXTWORDS(bc) ((bc) ? (bc)->extwords : 0)
    113 #define GET_MEMWORDS(bc) (GET_EXTWORDS(bc) ? GET_EXTWORDS(bc) : BPF_MEMWORDS)
    114 
    115 /*
    116  * Optimization hints.
    117  */
    118 typedef unsigned int bpfjit_hint_t;
    119 #define BJ_HINT_ABS  0x01 /* packet read at absolute offset   */
    120 #define BJ_HINT_IND  0x02 /* packet read at variable offset   */
    121 #define BJ_HINT_COP  0x04 /* BPF_COP or BPF_COPX instruction  */
    122 #define BJ_HINT_COPX 0x08 /* BPF_COPX instruction             */
    123 #define BJ_HINT_XREG 0x10 /* BJ_XREG is needed                */
    124 #define BJ_HINT_LDX  0x20 /* BPF_LDX instruction              */
    125 #define BJ_HINT_PKT  (BJ_HINT_ABS|BJ_HINT_IND) /* packet read */
    126 
    127 /*
    128  * Datatype for Array Bounds Check Elimination (ABC) pass.
    129  */
    130 typedef uint64_t bpfjit_abc_length_t;
    131 #define MAX_ABC_LENGTH (UINT32_MAX + UINT64_C(4)) /* max. width is 4 */
    132 
    133 struct bpfjit_stack
    134 {
    135 	bpf_ctx_t *ctx;
    136 	uint32_t *extmem; /* pointer to external memory store */
    137 #ifdef _KERNEL
    138 	int err; /* 3rd argument for m_xword/m_xhalf/m_xbyte function call */
    139 #endif
    140 	uint32_t mem[BPF_MEMWORDS]; /* internal memory store */
    141 };
    142 
    143 /*
    144  * Data for BPF_JMP instruction.
    145  * Forward declaration for struct bpfjit_jump.
    146  */
    147 struct bpfjit_jump_data;
    148 
    149 /*
    150  * Node of bjumps list.
    151  */
    152 struct bpfjit_jump {
    153 	struct sljit_jump *sjump;
    154 	SLIST_ENTRY(bpfjit_jump) entries;
    155 	struct bpfjit_jump_data *jdata;
    156 };
    157 
    158 /*
    159  * Data for BPF_JMP instruction.
    160  */
    161 struct bpfjit_jump_data {
    162 	/*
    163 	 * These entries make up bjumps list:
    164 	 * jtf[0] - when coming from jt path,
    165 	 * jtf[1] - when coming from jf path.
    166 	 */
    167 	struct bpfjit_jump jtf[2];
    168 	/*
    169 	 * Length calculated by Array Bounds Check Elimination (ABC) pass.
    170 	 */
    171 	bpfjit_abc_length_t abc_length;
    172 	/*
    173 	 * Length checked by the last out-of-bounds check.
    174 	 */
    175 	bpfjit_abc_length_t checked_length;
    176 };
    177 
    178 /*
    179  * Data for "read from packet" instructions.
    180  * See also read_pkt_insn() function below.
    181  */
    182 struct bpfjit_read_pkt_data {
    183 	/*
    184 	 * Length calculated by Array Bounds Check Elimination (ABC) pass.
    185 	 */
    186 	bpfjit_abc_length_t abc_length;
    187 	/*
    188 	 * If positive, emit "if (buflen < check_length) return 0"
    189 	 * out-of-bounds check.
    190 	 * Values greater than UINT32_MAX generate unconditional "return 0".
    191 	 */
    192 	bpfjit_abc_length_t check_length;
    193 };
    194 
    195 /*
    196  * Additional (optimization-related) data for bpf_insn.
    197  */
    198 struct bpfjit_insn_data {
    199 	/* List of jumps to this insn. */
    200 	SLIST_HEAD(, bpfjit_jump) bjumps;
    201 
    202 	union {
    203 		struct bpfjit_jump_data     jdata;
    204 		struct bpfjit_read_pkt_data rdata;
    205 	} u;
    206 
    207 	bpf_memword_init_t invalid;
    208 	bool unreachable;
    209 };
    210 
    211 #ifdef _KERNEL
    212 
    213 uint32_t m_xword(const struct mbuf *, uint32_t, int *);
    214 uint32_t m_xhalf(const struct mbuf *, uint32_t, int *);
    215 uint32_t m_xbyte(const struct mbuf *, uint32_t, int *);
    216 
    217 MODULE(MODULE_CLASS_MISC, bpfjit, "sljit")
    218 
    219 static int
    220 bpfjit_modcmd(modcmd_t cmd, void *arg)
    221 {
    222 
    223 	switch (cmd) {
    224 	case MODULE_CMD_INIT:
    225 		bpfjit_module_ops.bj_free_code = &bpfjit_free_code;
    226 		membar_producer();
    227 		bpfjit_module_ops.bj_generate_code = &bpfjit_generate_code;
    228 		membar_producer();
    229 		return 0;
    230 
    231 	case MODULE_CMD_FINI:
    232 		return EOPNOTSUPP;
    233 
    234 	default:
    235 		return ENOTTY;
    236 	}
    237 }
    238 #endif
    239 
    240 /*
    241  * Return a number of scratch registers to pass
    242  * to sljit_emit_enter() function.
    243  */
    244 static sljit_si
    245 nscratches(bpfjit_hint_t hints)
    246 {
    247 	sljit_si rv = 2;
    248 
    249 #ifdef _KERNEL
    250 	if (hints & BJ_HINT_PKT)
    251 		rv = 3; /* xcall with three arguments */
    252 #endif
    253 
    254 	if (hints & BJ_HINT_IND)
    255 		rv = 3; /* uses BJ_TMP2REG */
    256 
    257 	if (hints & BJ_HINT_COP)
    258 		rv = 3; /* calls copfunc with three arguments */
    259 
    260 	if (hints & BJ_HINT_XREG)
    261 		rv = 4; /* uses BJ_XREG */
    262 
    263 #ifdef _KERNEL
    264 	if (hints & BJ_HINT_LDX)
    265 		rv = 5; /* uses BJ_TMP3REG */
    266 #endif
    267 
    268 	if (hints & BJ_HINT_COPX)
    269 		rv = 5; /* uses BJ_TMP3REG */
    270 
    271 	return rv;
    272 }
    273 
    274 static uint32_t
    275 read_width(const struct bpf_insn *pc)
    276 {
    277 
    278 	switch (BPF_SIZE(pc->code)) {
    279 	case BPF_W:
    280 		return 4;
    281 	case BPF_H:
    282 		return 2;
    283 	case BPF_B:
    284 		return 1;
    285 	default:
    286 		BJ_ASSERT(false);
    287 		return 0;
    288 	}
    289 }
    290 
    291 /*
    292  * Copy buf and buflen members of bpf_args from BJ_ARGS
    293  * pointer to BJ_BUF and BJ_BUFLEN registers.
    294  */
    295 static int
    296 load_buf_buflen(struct sljit_compiler *compiler)
    297 {
    298 	int status;
    299 
    300 	status = sljit_emit_op1(compiler,
    301 	    SLJIT_MOV_P,
    302 	    BJ_BUF, 0,
    303 	    SLJIT_MEM1(BJ_ARGS),
    304 	    offsetof(struct bpf_args, pkt));
    305 	if (status != SLJIT_SUCCESS)
    306 		return status;
    307 
    308 	status = sljit_emit_op1(compiler,
    309 	    SLJIT_MOV, /* size_t source */
    310 	    BJ_BUFLEN, 0,
    311 	    SLJIT_MEM1(BJ_ARGS),
    312 	    offsetof(struct bpf_args, buflen));
    313 
    314 	return status;
    315 }
    316 
    317 static bool
    318 grow_jumps(struct sljit_jump ***jumps, size_t *size)
    319 {
    320 	struct sljit_jump **newptr;
    321 	const size_t elemsz = sizeof(struct sljit_jump *);
    322 	size_t old_size = *size;
    323 	size_t new_size = 2 * old_size;
    324 
    325 	if (new_size < old_size || new_size > SIZE_MAX / elemsz)
    326 		return false;
    327 
    328 	newptr = BJ_ALLOC(new_size * elemsz);
    329 	if (newptr == NULL)
    330 		return false;
    331 
    332 	memcpy(newptr, *jumps, old_size * elemsz);
    333 	BJ_FREE(*jumps, old_size * elemsz);
    334 
    335 	*jumps = newptr;
    336 	*size = new_size;
    337 	return true;
    338 }
    339 
    340 static bool
    341 append_jump(struct sljit_jump *jump, struct sljit_jump ***jumps,
    342     size_t *size, size_t *max_size)
    343 {
    344 	if (*size == *max_size && !grow_jumps(jumps, max_size))
    345 		return false;
    346 
    347 	(*jumps)[(*size)++] = jump;
    348 	return true;
    349 }
    350 
    351 /*
    352  * Emit code for BPF_LD+BPF_B+BPF_ABS    A <- P[k:1].
    353  */
    354 static int
    355 emit_read8(struct sljit_compiler *compiler, sljit_si src, uint32_t k)
    356 {
    357 
    358 	return sljit_emit_op1(compiler,
    359 	    SLJIT_MOV_UB,
    360 	    BJ_AREG, 0,
    361 	    SLJIT_MEM1(src), k);
    362 }
    363 
    364 /*
    365  * Emit code for BPF_LD+BPF_H+BPF_ABS    A <- P[k:2].
    366  */
    367 static int
    368 emit_read16(struct sljit_compiler *compiler, sljit_si src, uint32_t k)
    369 {
    370 	int status;
    371 
    372 	BJ_ASSERT(k <= UINT32_MAX - 1);
    373 
    374 	/* A = buf[k]; */
    375 	status = sljit_emit_op1(compiler,
    376 	    SLJIT_MOV_UB,
    377 	    BJ_AREG, 0,
    378 	    SLJIT_MEM1(src), k);
    379 	if (status != SLJIT_SUCCESS)
    380 		return status;
    381 
    382 	/* tmp1 = buf[k+1]; */
    383 	status = sljit_emit_op1(compiler,
    384 	    SLJIT_MOV_UB,
    385 	    BJ_TMP1REG, 0,
    386 	    SLJIT_MEM1(src), k+1);
    387 	if (status != SLJIT_SUCCESS)
    388 		return status;
    389 
    390 	/* A = A << 8; */
    391 	status = sljit_emit_op2(compiler,
    392 	    SLJIT_SHL,
    393 	    BJ_AREG, 0,
    394 	    BJ_AREG, 0,
    395 	    SLJIT_IMM, 8);
    396 	if (status != SLJIT_SUCCESS)
    397 		return status;
    398 
    399 	/* A = A + tmp1; */
    400 	status = sljit_emit_op2(compiler,
    401 	    SLJIT_ADD,
    402 	    BJ_AREG, 0,
    403 	    BJ_AREG, 0,
    404 	    BJ_TMP1REG, 0);
    405 	return status;
    406 }
    407 
    408 /*
    409  * Emit code for BPF_LD+BPF_W+BPF_ABS    A <- P[k:4].
    410  */
    411 static int
    412 emit_read32(struct sljit_compiler *compiler, sljit_si src, uint32_t k)
    413 {
    414 	int status;
    415 
    416 	BJ_ASSERT(k <= UINT32_MAX - 3);
    417 
    418 	/* A = buf[k]; */
    419 	status = sljit_emit_op1(compiler,
    420 	    SLJIT_MOV_UB,
    421 	    BJ_AREG, 0,
    422 	    SLJIT_MEM1(src), k);
    423 	if (status != SLJIT_SUCCESS)
    424 		return status;
    425 
    426 	/* tmp1 = buf[k+1]; */
    427 	status = sljit_emit_op1(compiler,
    428 	    SLJIT_MOV_UB,
    429 	    BJ_TMP1REG, 0,
    430 	    SLJIT_MEM1(src), k+1);
    431 	if (status != SLJIT_SUCCESS)
    432 		return status;
    433 
    434 	/* A = A << 8; */
    435 	status = sljit_emit_op2(compiler,
    436 	    SLJIT_SHL,
    437 	    BJ_AREG, 0,
    438 	    BJ_AREG, 0,
    439 	    SLJIT_IMM, 8);
    440 	if (status != SLJIT_SUCCESS)
    441 		return status;
    442 
    443 	/* A = A + tmp1; */
    444 	status = sljit_emit_op2(compiler,
    445 	    SLJIT_ADD,
    446 	    BJ_AREG, 0,
    447 	    BJ_AREG, 0,
    448 	    BJ_TMP1REG, 0);
    449 	if (status != SLJIT_SUCCESS)
    450 		return status;
    451 
    452 	/* tmp1 = buf[k+2]; */
    453 	status = sljit_emit_op1(compiler,
    454 	    SLJIT_MOV_UB,
    455 	    BJ_TMP1REG, 0,
    456 	    SLJIT_MEM1(src), k+2);
    457 	if (status != SLJIT_SUCCESS)
    458 		return status;
    459 
    460 	/* A = A << 8; */
    461 	status = sljit_emit_op2(compiler,
    462 	    SLJIT_SHL,
    463 	    BJ_AREG, 0,
    464 	    BJ_AREG, 0,
    465 	    SLJIT_IMM, 8);
    466 	if (status != SLJIT_SUCCESS)
    467 		return status;
    468 
    469 	/* A = A + tmp1; */
    470 	status = sljit_emit_op2(compiler,
    471 	    SLJIT_ADD,
    472 	    BJ_AREG, 0,
    473 	    BJ_AREG, 0,
    474 	    BJ_TMP1REG, 0);
    475 	if (status != SLJIT_SUCCESS)
    476 		return status;
    477 
    478 	/* tmp1 = buf[k+3]; */
    479 	status = sljit_emit_op1(compiler,
    480 	    SLJIT_MOV_UB,
    481 	    BJ_TMP1REG, 0,
    482 	    SLJIT_MEM1(src), k+3);
    483 	if (status != SLJIT_SUCCESS)
    484 		return status;
    485 
    486 	/* A = A << 8; */
    487 	status = sljit_emit_op2(compiler,
    488 	    SLJIT_SHL,
    489 	    BJ_AREG, 0,
    490 	    BJ_AREG, 0,
    491 	    SLJIT_IMM, 8);
    492 	if (status != SLJIT_SUCCESS)
    493 		return status;
    494 
    495 	/* A = A + tmp1; */
    496 	status = sljit_emit_op2(compiler,
    497 	    SLJIT_ADD,
    498 	    BJ_AREG, 0,
    499 	    BJ_AREG, 0,
    500 	    BJ_TMP1REG, 0);
    501 	return status;
    502 }
    503 
    504 #ifdef _KERNEL
    505 /*
    506  * Emit code for m_xword/m_xhalf/m_xbyte call.
    507  *
    508  * @pc BPF_LD+BPF_W+BPF_ABS    A <- P[k:4]
    509  *     BPF_LD+BPF_H+BPF_ABS    A <- P[k:2]
    510  *     BPF_LD+BPF_B+BPF_ABS    A <- P[k:1]
    511  *     BPF_LD+BPF_W+BPF_IND    A <- P[X+k:4]
    512  *     BPF_LD+BPF_H+BPF_IND    A <- P[X+k:2]
    513  *     BPF_LD+BPF_B+BPF_IND    A <- P[X+k:1]
    514  *     BPF_LDX+BPF_B+BPF_MSH   X <- 4*(P[k:1]&0xf)
    515  */
    516 static int
    517 emit_xcall(struct sljit_compiler *compiler, const struct bpf_insn *pc,
    518     int dst, struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize,
    519     uint32_t (*fn)(const struct mbuf *, uint32_t, int *))
    520 {
    521 #if BJ_XREG    == SLJIT_RETURN_REG   || \
    522     BJ_XREG    == SLJIT_SCRATCH_REG1 || \
    523     BJ_XREG    == SLJIT_SCRATCH_REG2 || \
    524     BJ_XREG    == SLJIT_SCRATCH_REG3 || \
    525     BJ_TMP3REG == SLJIT_RETURN_REG   || \
    526     BJ_TMP3REG == SLJIT_SCRATCH_REG1 || \
    527     BJ_TMP3REG == SLJIT_SCRATCH_REG2 || \
    528     BJ_TMP3REG == SLJIT_SCRATCH_REG3
    529 #error "Not supported assignment of registers."
    530 #endif
    531 	struct sljit_jump *jump;
    532 	int status;
    533 
    534 	BJ_ASSERT(dst != BJ_TMP2REG && dst != BJ_TMP3REG);
    535 
    536 	if (BPF_CLASS(pc->code) == BPF_LDX) {
    537 		/* save A */
    538 		status = sljit_emit_op1(compiler,
    539 		    SLJIT_MOV,
    540 		    BJ_TMP3REG, 0,
    541 		    BJ_AREG, 0);
    542 		if (status != SLJIT_SUCCESS)
    543 			return status;
    544 	}
    545 
    546 	/*
    547 	 * Prepare registers for fn(mbuf, k, &err) call.
    548 	 */
    549 	status = sljit_emit_op1(compiler,
    550 	    SLJIT_MOV,
    551 	    SLJIT_SCRATCH_REG1, 0,
    552 	    BJ_BUF, 0);
    553 	if (status != SLJIT_SUCCESS)
    554 		return status;
    555 
    556 	if (BPF_CLASS(pc->code) == BPF_LD && BPF_MODE(pc->code) == BPF_IND) {
    557 		/* if (X > UINT32_MAX - pc->k) return 0; */
    558 		jump = sljit_emit_cmp(compiler,
    559 		    SLJIT_C_GREATER,
    560 		    BJ_XREG, 0,
    561 		    SLJIT_IMM, UINT32_MAX - pc->k);
    562 		if (jump == NULL)
    563 			return SLJIT_ERR_ALLOC_FAILED;
    564 		if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
    565 			return SLJIT_ERR_ALLOC_FAILED;
    566 
    567 		/* k = X + pc->k; */
    568 		status = sljit_emit_op2(compiler,
    569 		    SLJIT_ADD,
    570 		    SLJIT_SCRATCH_REG2, 0,
    571 		    BJ_XREG, 0,
    572 		    SLJIT_IMM, (uint32_t)pc->k);
    573 		if (status != SLJIT_SUCCESS)
    574 			return status;
    575 	} else {
    576 		/* k = pc->k */
    577 		status = sljit_emit_op1(compiler,
    578 		    SLJIT_MOV,
    579 		    SLJIT_SCRATCH_REG2, 0,
    580 		    SLJIT_IMM, (uint32_t)pc->k);
    581 		if (status != SLJIT_SUCCESS)
    582 			return status;
    583 	}
    584 
    585 	/*
    586 	 * The third argument of fn is an address on stack.
    587 	 */
    588 	status = sljit_get_local_base(compiler,
    589 	    SLJIT_SCRATCH_REG3, 0,
    590 	    offsetof(struct bpfjit_stack, err));
    591 	if (status != SLJIT_SUCCESS)
    592 		return status;
    593 
    594 	/* fn(buf, k, &err); */
    595 	status = sljit_emit_ijump(compiler,
    596 	    SLJIT_CALL3,
    597 	    SLJIT_IMM, SLJIT_FUNC_OFFSET(fn));
    598 	if (status != SLJIT_SUCCESS)
    599 		return status;
    600 
    601 	if (dst != SLJIT_RETURN_REG) {
    602 		/* move return value to dst */
    603 		status = sljit_emit_op1(compiler,
    604 		    SLJIT_MOV,
    605 		    dst, 0,
    606 		    SLJIT_RETURN_REG, 0);
    607 		if (status != SLJIT_SUCCESS)
    608 			return status;
    609 	}
    610 
    611 	/* tmp2 = *err; */
    612 	status = sljit_emit_op1(compiler,
    613 	    SLJIT_MOV_UI,
    614 	    BJ_TMP2REG, 0,
    615 	    SLJIT_MEM1(SLJIT_LOCALS_REG),
    616 	    offsetof(struct bpfjit_stack, err));
    617 	if (status != SLJIT_SUCCESS)
    618 		return status;
    619 
    620 	/* if (tmp2 != 0) return 0; */
    621 	jump = sljit_emit_cmp(compiler,
    622 	    SLJIT_C_NOT_EQUAL,
    623 	    BJ_TMP2REG, 0,
    624 	    SLJIT_IMM, 0);
    625 	if (jump == NULL)
    626 		return SLJIT_ERR_ALLOC_FAILED;
    627 
    628 	if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
    629 		return SLJIT_ERR_ALLOC_FAILED;
    630 
    631 	if (BPF_CLASS(pc->code) == BPF_LDX) {
    632 		/* restore A */
    633 		status = sljit_emit_op1(compiler,
    634 		    SLJIT_MOV,
    635 		    BJ_AREG, 0,
    636 		    BJ_TMP3REG, 0);
    637 		if (status != SLJIT_SUCCESS)
    638 			return status;
    639 	}
    640 
    641 	return SLJIT_SUCCESS;
    642 }
    643 #endif
    644 
    645 /*
    646  * Emit code for BPF_COP and BPF_COPX instructions.
    647  */
    648 static int
    649 emit_cop(struct sljit_compiler *compiler,
    650     const bpf_ctx_t *bc, const struct bpf_insn *pc,
    651     struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize)
    652 {
    653 #if BJ_XREG    == SLJIT_RETURN_REG   || \
    654     BJ_XREG    == SLJIT_SCRATCH_REG1 || \
    655     BJ_XREG    == SLJIT_SCRATCH_REG2 || \
    656     BJ_XREG    == SLJIT_SCRATCH_REG3 || \
    657     BJ_TMP3REG == SLJIT_SCRATCH_REG1 || \
    658     BJ_TMP3REG == SLJIT_SCRATCH_REG2 || \
    659     BJ_TMP3REG == SLJIT_SCRATCH_REG3
    660 #error "Not supported assignment of registers."
    661 #endif
    662 
    663 	struct sljit_jump *jump;
    664 	sljit_si call_reg;
    665 	sljit_sw call_off;
    666 	int status;
    667 
    668 	BJ_ASSERT(bc != NULL && bc->copfuncs != NULL);
    669 
    670 	if (BPF_MISCOP(pc->code) == BPF_COP) {
    671 		call_reg = SLJIT_IMM;
    672 		call_off = SLJIT_FUNC_OFFSET(bc->copfuncs[pc->k]);
    673 	} else {
    674 		/* if (X >= bc->nfuncs) return 0; */
    675 		jump = sljit_emit_cmp(compiler,
    676 		    SLJIT_C_GREATER_EQUAL,
    677 		    BJ_XREG, 0,
    678 		    SLJIT_IMM, bc->nfuncs);
    679 		if (jump == NULL)
    680 			return SLJIT_ERR_ALLOC_FAILED;
    681 		if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
    682 			return SLJIT_ERR_ALLOC_FAILED;
    683 
    684 		/* tmp1 = ctx; */
    685 		status = sljit_emit_op1(compiler,
    686 		    SLJIT_MOV_P,
    687 		    BJ_TMP1REG, 0,
    688 		    SLJIT_MEM1(SLJIT_LOCALS_REG),
    689 		    offsetof(struct bpfjit_stack, ctx));
    690 		if (status != SLJIT_SUCCESS)
    691 			return status;
    692 
    693 		/* tmp1 = ctx->copfuncs; */
    694 		status = sljit_emit_op1(compiler,
    695 		    SLJIT_MOV_P,
    696 		    BJ_TMP1REG, 0,
    697 		    SLJIT_MEM1(BJ_TMP1REG),
    698 		    offsetof(struct bpf_ctx, copfuncs));
    699 		if (status != SLJIT_SUCCESS)
    700 			return status;
    701 
    702 		/* tmp2 = X; */
    703 		status = sljit_emit_op1(compiler,
    704 		    SLJIT_MOV,
    705 		    BJ_TMP2REG, 0,
    706 		    BJ_XREG, 0);
    707 		if (status != SLJIT_SUCCESS)
    708 			return status;
    709 
    710 		/* tmp3 = ctx->copfuncs[tmp2]; */
    711 		call_reg = BJ_TMP3REG;
    712 		call_off = 0;
    713 		status = sljit_emit_op1(compiler,
    714 		    SLJIT_MOV_P,
    715 		    call_reg, call_off,
    716 		    SLJIT_MEM2(BJ_TMP1REG, BJ_TMP2REG),
    717 		    SLJIT_WORD_SHIFT);
    718 		if (status != SLJIT_SUCCESS)
    719 			return status;
    720 	}
    721 
    722 	/*
    723 	 * Copy bpf_copfunc_t arguments to registers.
    724 	 */
    725 #if BJ_AREG != SLJIT_SCRATCH_REG3
    726 	status = sljit_emit_op1(compiler,
    727 	    SLJIT_MOV_UI,
    728 	    SLJIT_SCRATCH_REG3, 0,
    729 	    BJ_AREG, 0);
    730 	if (status != SLJIT_SUCCESS)
    731 		return status;
    732 #endif
    733 
    734 	status = sljit_emit_op1(compiler,
    735 	    SLJIT_MOV_P,
    736 	    SLJIT_SCRATCH_REG1, 0,
    737 	    SLJIT_MEM1(SLJIT_LOCALS_REG),
    738 	    offsetof(struct bpfjit_stack, ctx));
    739 	if (status != SLJIT_SUCCESS)
    740 		return status;
    741 
    742 	status = sljit_emit_op1(compiler,
    743 	    SLJIT_MOV_P,
    744 	    SLJIT_SCRATCH_REG2, 0,
    745 	    BJ_ARGS, 0);
    746 	if (status != SLJIT_SUCCESS)
    747 		return status;
    748 
    749 	status = sljit_emit_ijump(compiler,
    750 	    SLJIT_CALL3, call_reg, call_off);
    751 	if (status != SLJIT_SUCCESS)
    752 		return status;
    753 
    754 #if BJ_AREG != SLJIT_RETURN_REG
    755 	status = sljit_emit_op1(compiler,
    756 	    SLJIT_MOV,
    757 	    BJ_AREG, 0,
    758 	    SLJIT_RETURN_REG, 0);
    759 	if (status != SLJIT_SUCCESS)
    760 		return status;
    761 #endif
    762 
    763 	return SLJIT_SUCCESS;
    764 }
    765 
    766 /*
    767  * Generate code for
    768  * BPF_LD+BPF_W+BPF_ABS    A <- P[k:4]
    769  * BPF_LD+BPF_H+BPF_ABS    A <- P[k:2]
    770  * BPF_LD+BPF_B+BPF_ABS    A <- P[k:1]
    771  * BPF_LD+BPF_W+BPF_IND    A <- P[X+k:4]
    772  * BPF_LD+BPF_H+BPF_IND    A <- P[X+k:2]
    773  * BPF_LD+BPF_B+BPF_IND    A <- P[X+k:1]
    774  */
    775 static int
    776 emit_pkt_read(struct sljit_compiler *compiler,
    777     const struct bpf_insn *pc, struct sljit_jump *to_mchain_jump,
    778     struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize)
    779 {
    780 	int status = SLJIT_ERR_ALLOC_FAILED;
    781 	uint32_t width;
    782 	sljit_si ld_reg;
    783 	struct sljit_jump *jump;
    784 #ifdef _KERNEL
    785 	struct sljit_label *label;
    786 	struct sljit_jump *over_mchain_jump;
    787 	const bool check_zero_buflen = (to_mchain_jump != NULL);
    788 #endif
    789 	const uint32_t k = pc->k;
    790 
    791 #ifdef _KERNEL
    792 	if (to_mchain_jump == NULL) {
    793 		to_mchain_jump = sljit_emit_cmp(compiler,
    794 		    SLJIT_C_EQUAL,
    795 		    BJ_BUFLEN, 0,
    796 		    SLJIT_IMM, 0);
    797 		if (to_mchain_jump == NULL)
    798 			return SLJIT_ERR_ALLOC_FAILED;
    799 	}
    800 #endif
    801 
    802 	ld_reg = BJ_BUF;
    803 	width = read_width(pc);
    804 
    805 	if (BPF_MODE(pc->code) == BPF_IND) {
    806 		/* tmp1 = buflen - (pc->k + width); */
    807 		status = sljit_emit_op2(compiler,
    808 		    SLJIT_SUB,
    809 		    BJ_TMP1REG, 0,
    810 		    BJ_BUFLEN, 0,
    811 		    SLJIT_IMM, k + width);
    812 		if (status != SLJIT_SUCCESS)
    813 			return status;
    814 
    815 		/* ld_reg = buf + X; */
    816 		ld_reg = BJ_TMP2REG;
    817 		status = sljit_emit_op2(compiler,
    818 		    SLJIT_ADD,
    819 		    ld_reg, 0,
    820 		    BJ_BUF, 0,
    821 		    BJ_XREG, 0);
    822 		if (status != SLJIT_SUCCESS)
    823 			return status;
    824 
    825 		/* if (tmp1 < X) return 0; */
    826 		jump = sljit_emit_cmp(compiler,
    827 		    SLJIT_C_LESS,
    828 		    BJ_TMP1REG, 0,
    829 		    BJ_XREG, 0);
    830 		if (jump == NULL)
    831 			return SLJIT_ERR_ALLOC_FAILED;
    832 		if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
    833 			return SLJIT_ERR_ALLOC_FAILED;
    834 	}
    835 
    836 	switch (width) {
    837 	case 4:
    838 		status = emit_read32(compiler, ld_reg, k);
    839 		break;
    840 	case 2:
    841 		status = emit_read16(compiler, ld_reg, k);
    842 		break;
    843 	case 1:
    844 		status = emit_read8(compiler, ld_reg, k);
    845 		break;
    846 	}
    847 
    848 	if (status != SLJIT_SUCCESS)
    849 		return status;
    850 
    851 #ifdef _KERNEL
    852 	over_mchain_jump = sljit_emit_jump(compiler, SLJIT_JUMP);
    853 	if (over_mchain_jump == NULL)
    854 		return SLJIT_ERR_ALLOC_FAILED;
    855 
    856 	/* entry point to mchain handler */
    857 	label = sljit_emit_label(compiler);
    858 	if (label == NULL)
    859 		return SLJIT_ERR_ALLOC_FAILED;
    860 	sljit_set_label(to_mchain_jump, label);
    861 
    862 	if (check_zero_buflen) {
    863 		/* if (buflen != 0) return 0; */
    864 		jump = sljit_emit_cmp(compiler,
    865 		    SLJIT_C_NOT_EQUAL,
    866 		    BJ_BUFLEN, 0,
    867 		    SLJIT_IMM, 0);
    868 		if (jump == NULL)
    869 			return SLJIT_ERR_ALLOC_FAILED;
    870 		if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
    871 			return SLJIT_ERR_ALLOC_FAILED;
    872 	}
    873 
    874 	switch (width) {
    875 	case 4:
    876 		status = emit_xcall(compiler, pc, BJ_AREG,
    877 		    ret0, ret0_size, ret0_maxsize, &m_xword);
    878 		break;
    879 	case 2:
    880 		status = emit_xcall(compiler, pc, BJ_AREG,
    881 		    ret0, ret0_size, ret0_maxsize, &m_xhalf);
    882 		break;
    883 	case 1:
    884 		status = emit_xcall(compiler, pc, BJ_AREG,
    885 		    ret0, ret0_size, ret0_maxsize, &m_xbyte);
    886 		break;
    887 	}
    888 
    889 	if (status != SLJIT_SUCCESS)
    890 		return status;
    891 
    892 	label = sljit_emit_label(compiler);
    893 	if (label == NULL)
    894 		return SLJIT_ERR_ALLOC_FAILED;
    895 	sljit_set_label(over_mchain_jump, label);
    896 #endif
    897 
    898 	return SLJIT_SUCCESS;
    899 }
    900 
    901 static int
    902 emit_memload(struct sljit_compiler *compiler,
    903     sljit_si dst, uint32_t k, size_t extwords)
    904 {
    905 	int status;
    906 	sljit_si src;
    907 	sljit_sw srcw;
    908 
    909 	srcw = k * sizeof(uint32_t);
    910 
    911 	if (extwords == 0) {
    912 		src = SLJIT_MEM1(SLJIT_LOCALS_REG);
    913 		srcw += offsetof(struct bpfjit_stack, mem);
    914 	} else {
    915 		/* copy extmem pointer to the tmp1 register */
    916 		status = sljit_emit_op1(compiler,
    917 		    SLJIT_MOV_P,
    918 		    BJ_TMP1REG, 0,
    919 		    SLJIT_MEM1(SLJIT_LOCALS_REG),
    920 		    offsetof(struct bpfjit_stack, extmem));
    921 		if (status != SLJIT_SUCCESS)
    922 			return status;
    923 		src = SLJIT_MEM1(BJ_TMP1REG);
    924 	}
    925 
    926 	return sljit_emit_op1(compiler, SLJIT_MOV_UI, dst, 0, src, srcw);
    927 }
    928 
    929 static int
    930 emit_memstore(struct sljit_compiler *compiler,
    931     sljit_si src, uint32_t k, size_t extwords)
    932 {
    933 	int status;
    934 	sljit_si dst;
    935 	sljit_sw dstw;
    936 
    937 	dstw = k * sizeof(uint32_t);
    938 
    939 	if (extwords == 0) {
    940 		dst = SLJIT_MEM1(SLJIT_LOCALS_REG);
    941 		dstw += offsetof(struct bpfjit_stack, mem);
    942 	} else {
    943 		/* copy extmem pointer to the tmp1 register */
    944 		status = sljit_emit_op1(compiler,
    945 		    SLJIT_MOV_P,
    946 		    BJ_TMP1REG, 0,
    947 		    SLJIT_MEM1(SLJIT_LOCALS_REG),
    948 		    offsetof(struct bpfjit_stack, extmem));
    949 		if (status != SLJIT_SUCCESS)
    950 			return status;
    951 		dst = SLJIT_MEM1(BJ_TMP1REG);
    952 	}
    953 
    954 	return sljit_emit_op1(compiler, SLJIT_MOV_UI, dst, dstw, src, 0);
    955 }
    956 
    957 /*
    958  * Emit code for BPF_LDX+BPF_B+BPF_MSH    X <- 4*(P[k:1]&0xf).
    959  */
    960 static int
    961 emit_msh(struct sljit_compiler *compiler,
    962     const struct bpf_insn *pc, struct sljit_jump *to_mchain_jump,
    963     struct sljit_jump ***ret0, size_t *ret0_size, size_t *ret0_maxsize)
    964 {
    965 	int status;
    966 #ifdef _KERNEL
    967 	struct sljit_label *label;
    968 	struct sljit_jump *jump, *over_mchain_jump;
    969 	const bool check_zero_buflen = (to_mchain_jump != NULL);
    970 #endif
    971 	const uint32_t k = pc->k;
    972 
    973 #ifdef _KERNEL
    974 	if (to_mchain_jump == NULL) {
    975 		to_mchain_jump = sljit_emit_cmp(compiler,
    976 		    SLJIT_C_EQUAL,
    977 		    BJ_BUFLEN, 0,
    978 		    SLJIT_IMM, 0);
    979 		if (to_mchain_jump == NULL)
    980 			return SLJIT_ERR_ALLOC_FAILED;
    981 	}
    982 #endif
    983 
    984 	/* tmp1 = buf[k] */
    985 	status = sljit_emit_op1(compiler,
    986 	    SLJIT_MOV_UB,
    987 	    BJ_TMP1REG, 0,
    988 	    SLJIT_MEM1(BJ_BUF), k);
    989 	if (status != SLJIT_SUCCESS)
    990 		return status;
    991 
    992 	/* tmp1 &= 0xf */
    993 	status = sljit_emit_op2(compiler,
    994 	    SLJIT_AND,
    995 	    BJ_TMP1REG, 0,
    996 	    BJ_TMP1REG, 0,
    997 	    SLJIT_IMM, 0xf);
    998 	if (status != SLJIT_SUCCESS)
    999 		return status;
   1000 
   1001 	/* tmp1 = tmp1 << 2 */
   1002 	status = sljit_emit_op2(compiler,
   1003 	    SLJIT_SHL,
   1004 	    BJ_XREG, 0,
   1005 	    BJ_TMP1REG, 0,
   1006 	    SLJIT_IMM, 2);
   1007 	if (status != SLJIT_SUCCESS)
   1008 		return status;
   1009 
   1010 #ifdef _KERNEL
   1011 	over_mchain_jump = sljit_emit_jump(compiler, SLJIT_JUMP);
   1012 	if (over_mchain_jump == NULL)
   1013 		return SLJIT_ERR_ALLOC_FAILED;
   1014 
   1015 	/* entry point to mchain handler */
   1016 	label = sljit_emit_label(compiler);
   1017 	if (label == NULL)
   1018 		return SLJIT_ERR_ALLOC_FAILED;
   1019 	sljit_set_label(to_mchain_jump, label);
   1020 
   1021 	if (check_zero_buflen) {
   1022 		/* if (buflen != 0) return 0; */
   1023 		jump = sljit_emit_cmp(compiler,
   1024 		    SLJIT_C_NOT_EQUAL,
   1025 		    BJ_BUFLEN, 0,
   1026 		    SLJIT_IMM, 0);
   1027 		if (jump == NULL)
   1028 			return SLJIT_ERR_ALLOC_FAILED;
   1029 		if (!append_jump(jump, ret0, ret0_size, ret0_maxsize))
   1030 			return SLJIT_ERR_ALLOC_FAILED;
   1031 	}
   1032 
   1033 	status = emit_xcall(compiler, pc, BJ_TMP1REG,
   1034 	    ret0, ret0_size, ret0_maxsize, &m_xbyte);
   1035 	if (status != SLJIT_SUCCESS)
   1036 		return status;
   1037 
   1038 	/* tmp1 &= 0xf */
   1039 	status = sljit_emit_op2(compiler,
   1040 	    SLJIT_AND,
   1041 	    BJ_TMP1REG, 0,
   1042 	    BJ_TMP1REG, 0,
   1043 	    SLJIT_IMM, 0xf);
   1044 	if (status != SLJIT_SUCCESS)
   1045 		return status;
   1046 
   1047 	/* tmp1 = tmp1 << 2 */
   1048 	status = sljit_emit_op2(compiler,
   1049 	    SLJIT_SHL,
   1050 	    BJ_XREG, 0,
   1051 	    BJ_TMP1REG, 0,
   1052 	    SLJIT_IMM, 2);
   1053 	if (status != SLJIT_SUCCESS)
   1054 		return status;
   1055 
   1056 
   1057 	label = sljit_emit_label(compiler);
   1058 	if (label == NULL)
   1059 		return SLJIT_ERR_ALLOC_FAILED;
   1060 	sljit_set_label(over_mchain_jump, label);
   1061 #endif
   1062 
   1063 	return SLJIT_SUCCESS;
   1064 }
   1065 
   1066 static int
   1067 emit_pow2_division(struct sljit_compiler *compiler, uint32_t k)
   1068 {
   1069 	int shift = 0;
   1070 	int status = SLJIT_SUCCESS;
   1071 
   1072 	while (k > 1) {
   1073 		k >>= 1;
   1074 		shift++;
   1075 	}
   1076 
   1077 	BJ_ASSERT(k == 1 && shift < 32);
   1078 
   1079 	if (shift != 0) {
   1080 		status = sljit_emit_op2(compiler,
   1081 		    SLJIT_LSHR|SLJIT_INT_OP,
   1082 		    BJ_AREG, 0,
   1083 		    BJ_AREG, 0,
   1084 		    SLJIT_IMM, shift);
   1085 	}
   1086 
   1087 	return status;
   1088 }
   1089 
   1090 #if !defined(BPFJIT_USE_UDIV)
   1091 static sljit_uw
   1092 divide(sljit_uw x, sljit_uw y)
   1093 {
   1094 
   1095 	return (uint32_t)x / (uint32_t)y;
   1096 }
   1097 #endif
   1098 
   1099 /*
   1100  * Emit code for A = A / div.
   1101  * divt,divw are either SLJIT_IMM,pc->k or BJ_XREG,0.
   1102  */
   1103 static int
   1104 emit_division(struct sljit_compiler *compiler, int divt, sljit_sw divw)
   1105 {
   1106 	int status;
   1107 
   1108 #if BJ_XREG == SLJIT_RETURN_REG   || \
   1109     BJ_XREG == SLJIT_SCRATCH_REG1 || \
   1110     BJ_XREG == SLJIT_SCRATCH_REG2 || \
   1111     BJ_AREG == SLJIT_SCRATCH_REG2
   1112 #error "Not supported assignment of registers."
   1113 #endif
   1114 
   1115 #if BJ_AREG != SLJIT_SCRATCH_REG1
   1116 	status = sljit_emit_op1(compiler,
   1117 	    SLJIT_MOV,
   1118 	    SLJIT_SCRATCH_REG1, 0,
   1119 	    BJ_AREG, 0);
   1120 	if (status != SLJIT_SUCCESS)
   1121 		return status;
   1122 #endif
   1123 
   1124 	status = sljit_emit_op1(compiler,
   1125 	    SLJIT_MOV,
   1126 	    SLJIT_SCRATCH_REG2, 0,
   1127 	    divt, divw);
   1128 	if (status != SLJIT_SUCCESS)
   1129 		return status;
   1130 
   1131 #if defined(BPFJIT_USE_UDIV)
   1132 	status = sljit_emit_op0(compiler, SLJIT_UDIV|SLJIT_INT_OP);
   1133 
   1134 #if BJ_AREG != SLJIT_SCRATCH_REG1
   1135 	status = sljit_emit_op1(compiler,
   1136 	    SLJIT_MOV,
   1137 	    BJ_AREG, 0,
   1138 	    SLJIT_SCRATCH_REG1, 0);
   1139 	if (status != SLJIT_SUCCESS)
   1140 		return status;
   1141 #endif
   1142 #else
   1143 	status = sljit_emit_ijump(compiler,
   1144 	    SLJIT_CALL2,
   1145 	    SLJIT_IMM, SLJIT_FUNC_OFFSET(divide));
   1146 
   1147 #if BJ_AREG != SLJIT_RETURN_REG
   1148 	status = sljit_emit_op1(compiler,
   1149 	    SLJIT_MOV,
   1150 	    BJ_AREG, 0,
   1151 	    SLJIT_RETURN_REG, 0);
   1152 	if (status != SLJIT_SUCCESS)
   1153 		return status;
   1154 #endif
   1155 #endif
   1156 
   1157 	return status;
   1158 }
   1159 
   1160 /*
   1161  * Return true if pc is a "read from packet" instruction.
   1162  * If length is not NULL and return value is true, *length will
   1163  * be set to a safe length required to read a packet.
   1164  */
   1165 static bool
   1166 read_pkt_insn(const struct bpf_insn *pc, bpfjit_abc_length_t *length)
   1167 {
   1168 	bool rv;
   1169 	bpfjit_abc_length_t width;
   1170 
   1171 	switch (BPF_CLASS(pc->code)) {
   1172 	default:
   1173 		rv = false;
   1174 		break;
   1175 
   1176 	case BPF_LD:
   1177 		rv = BPF_MODE(pc->code) == BPF_ABS ||
   1178 		     BPF_MODE(pc->code) == BPF_IND;
   1179 		if (rv)
   1180 			width = read_width(pc);
   1181 		break;
   1182 
   1183 	case BPF_LDX:
   1184 		rv = pc->code == (BPF_LDX|BPF_B|BPF_MSH);
   1185 		width = 1;
   1186 		break;
   1187 	}
   1188 
   1189 	if (rv && length != NULL) {
   1190 		/*
   1191 		 * Values greater than UINT32_MAX will generate
   1192 		 * unconditional "return 0".
   1193 		 */
   1194 		*length = (uint32_t)pc->k + width;
   1195 	}
   1196 
   1197 	return rv;
   1198 }
   1199 
   1200 static void
   1201 optimize_init(struct bpfjit_insn_data *insn_dat, size_t insn_count)
   1202 {
   1203 	size_t i;
   1204 
   1205 	for (i = 0; i < insn_count; i++) {
   1206 		SLIST_INIT(&insn_dat[i].bjumps);
   1207 		insn_dat[i].invalid = BJ_INIT_NOBITS;
   1208 	}
   1209 }
   1210 
   1211 /*
   1212  * The function divides instructions into blocks. Destination of a jump
   1213  * instruction starts a new block. BPF_RET and BPF_JMP instructions
   1214  * terminate a block. Blocks are linear, that is, there are no jumps out
   1215  * from the middle of a block and there are no jumps in to the middle of
   1216  * a block.
   1217  *
   1218  * The function also sets bits in *initmask for memwords that
   1219  * need to be initialized to zero. Note that this set should be empty
   1220  * for any valid kernel filter program.
   1221  */
   1222 static bool
   1223 optimize_pass1(const bpf_ctx_t *bc, const struct bpf_insn *insns,
   1224     struct bpfjit_insn_data *insn_dat, size_t insn_count,
   1225     bpf_memword_init_t *initmask, bpfjit_hint_t *hints)
   1226 {
   1227 	struct bpfjit_jump *jtf;
   1228 	size_t i;
   1229 	uint32_t jt, jf;
   1230 	bpfjit_abc_length_t length;
   1231 	bpf_memword_init_t invalid; /* borrowed from bpf_filter() */
   1232 	bool unreachable;
   1233 
   1234 	const size_t memwords = GET_MEMWORDS(bc);
   1235 
   1236 	*hints = 0;
   1237 	*initmask = BJ_INIT_NOBITS;
   1238 
   1239 	unreachable = false;
   1240 	invalid = ~BJ_INIT_NOBITS;
   1241 
   1242 	for (i = 0; i < insn_count; i++) {
   1243 		if (!SLIST_EMPTY(&insn_dat[i].bjumps))
   1244 			unreachable = false;
   1245 		insn_dat[i].unreachable = unreachable;
   1246 
   1247 		if (unreachable)
   1248 			continue;
   1249 
   1250 		invalid |= insn_dat[i].invalid;
   1251 
   1252 		if (read_pkt_insn(&insns[i], &length) && length > UINT32_MAX)
   1253 			unreachable = true;
   1254 
   1255 		switch (BPF_CLASS(insns[i].code)) {
   1256 		case BPF_RET:
   1257 			if (BPF_RVAL(insns[i].code) == BPF_A)
   1258 				*initmask |= invalid & BJ_INIT_ABIT;
   1259 
   1260 			unreachable = true;
   1261 			continue;
   1262 
   1263 		case BPF_LD:
   1264 			if (BPF_MODE(insns[i].code) == BPF_ABS)
   1265 				*hints |= BJ_HINT_ABS;
   1266 
   1267 			if (BPF_MODE(insns[i].code) == BPF_IND) {
   1268 				*hints |= BJ_HINT_IND | BJ_HINT_XREG;
   1269 				*initmask |= invalid & BJ_INIT_XBIT;
   1270 			}
   1271 
   1272 			if (BPF_MODE(insns[i].code) == BPF_MEM &&
   1273 			    (uint32_t)insns[i].k < memwords) {
   1274 				*initmask |= invalid & BJ_INIT_MBIT(insns[i].k);
   1275 			}
   1276 
   1277 			invalid &= ~BJ_INIT_ABIT;
   1278 			continue;
   1279 
   1280 		case BPF_LDX:
   1281 			*hints |= BJ_HINT_XREG | BJ_HINT_LDX;
   1282 
   1283 			if (BPF_MODE(insns[i].code) == BPF_MEM &&
   1284 			    (uint32_t)insns[i].k < memwords) {
   1285 				*initmask |= invalid & BJ_INIT_MBIT(insns[i].k);
   1286 			}
   1287 
   1288 			invalid &= ~BJ_INIT_XBIT;
   1289 			continue;
   1290 
   1291 		case BPF_ST:
   1292 			*initmask |= invalid & BJ_INIT_ABIT;
   1293 
   1294 			if ((uint32_t)insns[i].k < memwords)
   1295 				invalid &= ~BJ_INIT_MBIT(insns[i].k);
   1296 
   1297 			continue;
   1298 
   1299 		case BPF_STX:
   1300 			*hints |= BJ_HINT_XREG;
   1301 			*initmask |= invalid & BJ_INIT_XBIT;
   1302 
   1303 			if ((uint32_t)insns[i].k < memwords)
   1304 				invalid &= ~BJ_INIT_MBIT(insns[i].k);
   1305 
   1306 			continue;
   1307 
   1308 		case BPF_ALU:
   1309 			*initmask |= invalid & BJ_INIT_ABIT;
   1310 
   1311 			if (insns[i].code != (BPF_ALU|BPF_NEG) &&
   1312 			    BPF_SRC(insns[i].code) == BPF_X) {
   1313 				*hints |= BJ_HINT_XREG;
   1314 				*initmask |= invalid & BJ_INIT_XBIT;
   1315 			}
   1316 
   1317 			invalid &= ~BJ_INIT_ABIT;
   1318 			continue;
   1319 
   1320 		case BPF_MISC:
   1321 			switch (BPF_MISCOP(insns[i].code)) {
   1322 			case BPF_TAX: // X <- A
   1323 				*hints |= BJ_HINT_XREG;
   1324 				*initmask |= invalid & BJ_INIT_ABIT;
   1325 				invalid &= ~BJ_INIT_XBIT;
   1326 				continue;
   1327 
   1328 			case BPF_TXA: // A <- X
   1329 				*hints |= BJ_HINT_XREG;
   1330 				*initmask |= invalid & BJ_INIT_XBIT;
   1331 				invalid &= ~BJ_INIT_ABIT;
   1332 				continue;
   1333 
   1334 			case BPF_COPX:
   1335 				*hints |= BJ_HINT_XREG | BJ_HINT_COPX;
   1336 				/* FALLTHROUGH */
   1337 
   1338 			case BPF_COP:
   1339 				*hints |= BJ_HINT_COP;
   1340 				*initmask |= invalid & BJ_INIT_ABIT;
   1341 				invalid &= ~BJ_INIT_ABIT;
   1342 				continue;
   1343 			}
   1344 
   1345 			continue;
   1346 
   1347 		case BPF_JMP:
   1348 			/* Initialize abc_length for ABC pass. */
   1349 			insn_dat[i].u.jdata.abc_length = MAX_ABC_LENGTH;
   1350 
   1351 			if (BPF_OP(insns[i].code) == BPF_JA) {
   1352 				jt = jf = insns[i].k;
   1353 			} else {
   1354 				jt = insns[i].jt;
   1355 				jf = insns[i].jf;
   1356 			}
   1357 
   1358 			if (jt >= insn_count - (i + 1) ||
   1359 			    jf >= insn_count - (i + 1)) {
   1360 				return false;
   1361 			}
   1362 
   1363 			if (jt > 0 && jf > 0)
   1364 				unreachable = true;
   1365 
   1366 			jt += i + 1;
   1367 			jf += i + 1;
   1368 
   1369 			jtf = insn_dat[i].u.jdata.jtf;
   1370 
   1371 			jtf[0].jdata = &insn_dat[i].u.jdata;
   1372 			SLIST_INSERT_HEAD(&insn_dat[jt].bjumps,
   1373 			    &jtf[0], entries);
   1374 
   1375 			if (jf != jt) {
   1376 				jtf[1].jdata = &insn_dat[i].u.jdata;
   1377 				SLIST_INSERT_HEAD(&insn_dat[jf].bjumps,
   1378 				    &jtf[1], entries);
   1379 			}
   1380 
   1381 			insn_dat[jf].invalid |= invalid;
   1382 			insn_dat[jt].invalid |= invalid;
   1383 			invalid = 0;
   1384 
   1385 			continue;
   1386 		}
   1387 	}
   1388 
   1389 	return true;
   1390 }
   1391 
   1392 /*
   1393  * Array Bounds Check Elimination (ABC) pass.
   1394  */
   1395 static void
   1396 optimize_pass2(const bpf_ctx_t *bc, const struct bpf_insn *insns,
   1397     struct bpfjit_insn_data *insn_dat, size_t insn_count)
   1398 {
   1399 	struct bpfjit_jump *jmp;
   1400 	const struct bpf_insn *pc;
   1401 	struct bpfjit_insn_data *pd;
   1402 	size_t i;
   1403 	bpfjit_abc_length_t length, abc_length = 0;
   1404 
   1405 	const size_t extwords = GET_EXTWORDS(bc);
   1406 
   1407 	for (i = insn_count; i != 0; i--) {
   1408 		pc = &insns[i-1];
   1409 		pd = &insn_dat[i-1];
   1410 
   1411 		if (pd->unreachable)
   1412 			continue;
   1413 
   1414 		switch (BPF_CLASS(pc->code)) {
   1415 		case BPF_RET:
   1416 			/*
   1417 			 * It's quite common for bpf programs to
   1418 			 * check packet bytes in increasing order
   1419 			 * and return zero if bytes don't match
   1420 			 * specified critetion. Such programs disable
   1421 			 * ABC optimization completely because for
   1422 			 * every jump there is a branch with no read
   1423 			 * instruction.
   1424 			 * With no side effects, BPF_STMT(BPF_RET+BPF_K, 0)
   1425 			 * is indistinguishable from out-of-bound load.
   1426 			 * Therefore, abc_length can be set to
   1427 			 * MAX_ABC_LENGTH and enable ABC for many
   1428 			 * bpf programs.
   1429 			 * If this optimization encounters any
   1430 			 * instruction with a side effect, it will
   1431 			 * reset abc_length.
   1432 			 */
   1433 			if (BPF_RVAL(pc->code) == BPF_K && pc->k == 0)
   1434 				abc_length = MAX_ABC_LENGTH;
   1435 			else
   1436 				abc_length = 0;
   1437 			break;
   1438 
   1439 		case BPF_MISC:
   1440 			if (BPF_MISCOP(pc->code) == BPF_COP ||
   1441 			    BPF_MISCOP(pc->code) == BPF_COPX) {
   1442 				/* COP instructions can have side effects. */
   1443 				abc_length = 0;
   1444 			}
   1445 			break;
   1446 
   1447 		case BPF_ST:
   1448 		case BPF_STX:
   1449 			if (extwords != 0) {
   1450 				/* Write to memory is visible after a call. */
   1451 				abc_length = 0;
   1452 			}
   1453 			break;
   1454 
   1455 		case BPF_JMP:
   1456 			abc_length = pd->u.jdata.abc_length;
   1457 			break;
   1458 
   1459 		default:
   1460 			if (read_pkt_insn(pc, &length)) {
   1461 				if (abc_length < length)
   1462 					abc_length = length;
   1463 				pd->u.rdata.abc_length = abc_length;
   1464 			}
   1465 			break;
   1466 		}
   1467 
   1468 		SLIST_FOREACH(jmp, &pd->bjumps, entries) {
   1469 			if (jmp->jdata->abc_length > abc_length)
   1470 				jmp->jdata->abc_length = abc_length;
   1471 		}
   1472 	}
   1473 }
   1474 
   1475 static void
   1476 optimize_pass3(const struct bpf_insn *insns,
   1477     struct bpfjit_insn_data *insn_dat, size_t insn_count)
   1478 {
   1479 	struct bpfjit_jump *jmp;
   1480 	size_t i;
   1481 	bpfjit_abc_length_t checked_length = 0;
   1482 
   1483 	for (i = 0; i < insn_count; i++) {
   1484 		if (insn_dat[i].unreachable)
   1485 			continue;
   1486 
   1487 		SLIST_FOREACH(jmp, &insn_dat[i].bjumps, entries) {
   1488 			if (jmp->jdata->checked_length < checked_length)
   1489 				checked_length = jmp->jdata->checked_length;
   1490 		}
   1491 
   1492 		if (BPF_CLASS(insns[i].code) == BPF_JMP) {
   1493 			insn_dat[i].u.jdata.checked_length = checked_length;
   1494 		} else if (read_pkt_insn(&insns[i], NULL)) {
   1495 			struct bpfjit_read_pkt_data *rdata =
   1496 			    &insn_dat[i].u.rdata;
   1497 			rdata->check_length = 0;
   1498 			if (checked_length < rdata->abc_length) {
   1499 				checked_length = rdata->abc_length;
   1500 				rdata->check_length = checked_length;
   1501 			}
   1502 		}
   1503 	}
   1504 }
   1505 
   1506 static bool
   1507 optimize(const bpf_ctx_t *bc, const struct bpf_insn *insns,
   1508     struct bpfjit_insn_data *insn_dat, size_t insn_count,
   1509     bpf_memword_init_t *initmask, bpfjit_hint_t *hints)
   1510 {
   1511 
   1512 	optimize_init(insn_dat, insn_count);
   1513 
   1514 	if (!optimize_pass1(bc, insns, insn_dat, insn_count, initmask, hints))
   1515 		return false;
   1516 
   1517 	optimize_pass2(bc, insns, insn_dat, insn_count);
   1518 	optimize_pass3(insns, insn_dat, insn_count);
   1519 
   1520 	return true;
   1521 }
   1522 
   1523 /*
   1524  * Convert BPF_ALU operations except BPF_NEG and BPF_DIV to sljit operation.
   1525  */
   1526 static int
   1527 bpf_alu_to_sljit_op(const struct bpf_insn *pc)
   1528 {
   1529 
   1530 	/*
   1531 	 * Note: all supported 64bit arches have 32bit multiply
   1532 	 * instruction so SLJIT_INT_OP doesn't have any overhead.
   1533 	 */
   1534 	switch (BPF_OP(pc->code)) {
   1535 	case BPF_ADD: return SLJIT_ADD;
   1536 	case BPF_SUB: return SLJIT_SUB;
   1537 	case BPF_MUL: return SLJIT_MUL|SLJIT_INT_OP;
   1538 	case BPF_OR:  return SLJIT_OR;
   1539 	case BPF_AND: return SLJIT_AND;
   1540 	case BPF_LSH: return SLJIT_SHL;
   1541 	case BPF_RSH: return SLJIT_LSHR|SLJIT_INT_OP;
   1542 	default:
   1543 		BJ_ASSERT(false);
   1544 		return 0;
   1545 	}
   1546 }
   1547 
   1548 /*
   1549  * Convert BPF_JMP operations except BPF_JA to sljit condition.
   1550  */
   1551 static int
   1552 bpf_jmp_to_sljit_cond(const struct bpf_insn *pc, bool negate)
   1553 {
   1554 	/*
   1555 	 * Note: all supported 64bit arches have 32bit comparison
   1556 	 * instructions so SLJIT_INT_OP doesn't have any overhead.
   1557 	 */
   1558 	int rv = SLJIT_INT_OP;
   1559 
   1560 	switch (BPF_OP(pc->code)) {
   1561 	case BPF_JGT:
   1562 		rv |= negate ? SLJIT_C_LESS_EQUAL : SLJIT_C_GREATER;
   1563 		break;
   1564 	case BPF_JGE:
   1565 		rv |= negate ? SLJIT_C_LESS : SLJIT_C_GREATER_EQUAL;
   1566 		break;
   1567 	case BPF_JEQ:
   1568 		rv |= negate ? SLJIT_C_NOT_EQUAL : SLJIT_C_EQUAL;
   1569 		break;
   1570 	case BPF_JSET:
   1571 		rv |= negate ? SLJIT_C_EQUAL : SLJIT_C_NOT_EQUAL;
   1572 		break;
   1573 	default:
   1574 		BJ_ASSERT(false);
   1575 	}
   1576 
   1577 	return rv;
   1578 }
   1579 
   1580 /*
   1581  * Convert BPF_K and BPF_X to sljit register.
   1582  */
   1583 static int
   1584 kx_to_reg(const struct bpf_insn *pc)
   1585 {
   1586 
   1587 	switch (BPF_SRC(pc->code)) {
   1588 	case BPF_K: return SLJIT_IMM;
   1589 	case BPF_X: return BJ_XREG;
   1590 	default:
   1591 		BJ_ASSERT(false);
   1592 		return 0;
   1593 	}
   1594 }
   1595 
   1596 static sljit_sw
   1597 kx_to_reg_arg(const struct bpf_insn *pc)
   1598 {
   1599 
   1600 	switch (BPF_SRC(pc->code)) {
   1601 	case BPF_K: return (uint32_t)pc->k; /* SLJIT_IMM, pc->k, */
   1602 	case BPF_X: return 0;               /* BJ_XREG, 0,      */
   1603 	default:
   1604 		BJ_ASSERT(false);
   1605 		return 0;
   1606 	}
   1607 }
   1608 
   1609 static bool
   1610 generate_insn_code(struct sljit_compiler *compiler, const bpf_ctx_t *bc,
   1611     const struct bpf_insn *insns, struct bpfjit_insn_data *insn_dat,
   1612     size_t insn_count)
   1613 {
   1614 	/* a list of jumps to out-of-bound return from a generated function */
   1615 	struct sljit_jump **ret0;
   1616 	size_t ret0_size, ret0_maxsize;
   1617 
   1618 	struct sljit_jump *jump;
   1619 	struct sljit_label *label;
   1620 	const struct bpf_insn *pc;
   1621 	struct bpfjit_jump *bjump, *jtf;
   1622 	struct sljit_jump *to_mchain_jump;
   1623 
   1624 	size_t i;
   1625 	int status;
   1626 	int branching, negate;
   1627 	unsigned int rval, mode, src;
   1628 	uint32_t jt, jf;
   1629 
   1630 	bool unconditional_ret;
   1631 	bool rv;
   1632 
   1633 	const size_t extwords = GET_EXTWORDS(bc);
   1634 	const size_t memwords = GET_MEMWORDS(bc);
   1635 
   1636 	ret0 = NULL;
   1637 	rv = false;
   1638 
   1639 	ret0_size = 0;
   1640 	ret0_maxsize = 64;
   1641 	ret0 = BJ_ALLOC(ret0_maxsize * sizeof(ret0[0]));
   1642 	if (ret0 == NULL)
   1643 		goto fail;
   1644 
   1645 	/* reset sjump members of jdata */
   1646 	for (i = 0; i < insn_count; i++) {
   1647 		if (insn_dat[i].unreachable ||
   1648 		    BPF_CLASS(insns[i].code) != BPF_JMP) {
   1649 			continue;
   1650 		}
   1651 
   1652 		jtf = insn_dat[i].u.jdata.jtf;
   1653 		jtf[0].sjump = jtf[1].sjump = NULL;
   1654 	}
   1655 
   1656 	/* main loop */
   1657 	for (i = 0; i < insn_count; i++) {
   1658 		if (insn_dat[i].unreachable)
   1659 			continue;
   1660 
   1661 		/*
   1662 		 * Resolve jumps to the current insn.
   1663 		 */
   1664 		label = NULL;
   1665 		SLIST_FOREACH(bjump, &insn_dat[i].bjumps, entries) {
   1666 			if (bjump->sjump != NULL) {
   1667 				if (label == NULL)
   1668 					label = sljit_emit_label(compiler);
   1669 				if (label == NULL)
   1670 					goto fail;
   1671 				sljit_set_label(bjump->sjump, label);
   1672 			}
   1673 		}
   1674 
   1675 		to_mchain_jump = NULL;
   1676 		unconditional_ret = false;
   1677 
   1678 		if (read_pkt_insn(&insns[i], NULL)) {
   1679 			if (insn_dat[i].u.rdata.check_length > UINT32_MAX) {
   1680 				/* Jump to "return 0" unconditionally. */
   1681 				unconditional_ret = true;
   1682 				jump = sljit_emit_jump(compiler, SLJIT_JUMP);
   1683 				if (jump == NULL)
   1684 					goto fail;
   1685 				if (!append_jump(jump, &ret0,
   1686 				    &ret0_size, &ret0_maxsize))
   1687 					goto fail;
   1688 			} else if (insn_dat[i].u.rdata.check_length > 0) {
   1689 				/* if (buflen < check_length) return 0; */
   1690 				jump = sljit_emit_cmp(compiler,
   1691 				    SLJIT_C_LESS,
   1692 				    BJ_BUFLEN, 0,
   1693 				    SLJIT_IMM,
   1694 				    insn_dat[i].u.rdata.check_length);
   1695 				if (jump == NULL)
   1696 					goto fail;
   1697 #ifdef _KERNEL
   1698 				to_mchain_jump = jump;
   1699 #else
   1700 				if (!append_jump(jump, &ret0,
   1701 				    &ret0_size, &ret0_maxsize))
   1702 					goto fail;
   1703 #endif
   1704 			}
   1705 		}
   1706 
   1707 		pc = &insns[i];
   1708 		switch (BPF_CLASS(pc->code)) {
   1709 
   1710 		default:
   1711 			goto fail;
   1712 
   1713 		case BPF_LD:
   1714 			/* BPF_LD+BPF_IMM          A <- k */
   1715 			if (pc->code == (BPF_LD|BPF_IMM)) {
   1716 				status = sljit_emit_op1(compiler,
   1717 				    SLJIT_MOV,
   1718 				    BJ_AREG, 0,
   1719 				    SLJIT_IMM, (uint32_t)pc->k);
   1720 				if (status != SLJIT_SUCCESS)
   1721 					goto fail;
   1722 
   1723 				continue;
   1724 			}
   1725 
   1726 			/* BPF_LD+BPF_MEM          A <- M[k] */
   1727 			if (pc->code == (BPF_LD|BPF_MEM)) {
   1728 				if ((uint32_t)pc->k >= memwords)
   1729 					goto fail;
   1730 				status = emit_memload(compiler,
   1731 				    BJ_AREG, pc->k, extwords);
   1732 				if (status != SLJIT_SUCCESS)
   1733 					goto fail;
   1734 
   1735 				continue;
   1736 			}
   1737 
   1738 			/* BPF_LD+BPF_W+BPF_LEN    A <- len */
   1739 			if (pc->code == (BPF_LD|BPF_W|BPF_LEN)) {
   1740 				status = sljit_emit_op1(compiler,
   1741 				    SLJIT_MOV, /* size_t source */
   1742 				    BJ_AREG, 0,
   1743 				    SLJIT_MEM1(BJ_ARGS),
   1744 				    offsetof(struct bpf_args, wirelen));
   1745 				if (status != SLJIT_SUCCESS)
   1746 					goto fail;
   1747 
   1748 				continue;
   1749 			}
   1750 
   1751 			mode = BPF_MODE(pc->code);
   1752 			if (mode != BPF_ABS && mode != BPF_IND)
   1753 				goto fail;
   1754 
   1755 			if (unconditional_ret)
   1756 				continue;
   1757 
   1758 			status = emit_pkt_read(compiler, pc,
   1759 			    to_mchain_jump, &ret0, &ret0_size, &ret0_maxsize);
   1760 			if (status != SLJIT_SUCCESS)
   1761 				goto fail;
   1762 
   1763 			continue;
   1764 
   1765 		case BPF_LDX:
   1766 			mode = BPF_MODE(pc->code);
   1767 
   1768 			/* BPF_LDX+BPF_W+BPF_IMM    X <- k */
   1769 			if (mode == BPF_IMM) {
   1770 				if (BPF_SIZE(pc->code) != BPF_W)
   1771 					goto fail;
   1772 				status = sljit_emit_op1(compiler,
   1773 				    SLJIT_MOV,
   1774 				    BJ_XREG, 0,
   1775 				    SLJIT_IMM, (uint32_t)pc->k);
   1776 				if (status != SLJIT_SUCCESS)
   1777 					goto fail;
   1778 
   1779 				continue;
   1780 			}
   1781 
   1782 			/* BPF_LDX+BPF_W+BPF_LEN    X <- len */
   1783 			if (mode == BPF_LEN) {
   1784 				if (BPF_SIZE(pc->code) != BPF_W)
   1785 					goto fail;
   1786 				status = sljit_emit_op1(compiler,
   1787 				    SLJIT_MOV, /* size_t source */
   1788 				    BJ_XREG, 0,
   1789 				    SLJIT_MEM1(BJ_ARGS),
   1790 				    offsetof(struct bpf_args, wirelen));
   1791 				if (status != SLJIT_SUCCESS)
   1792 					goto fail;
   1793 
   1794 				continue;
   1795 			}
   1796 
   1797 			/* BPF_LDX+BPF_W+BPF_MEM    X <- M[k] */
   1798 			if (mode == BPF_MEM) {
   1799 				if (BPF_SIZE(pc->code) != BPF_W)
   1800 					goto fail;
   1801 				if ((uint32_t)pc->k >= memwords)
   1802 					goto fail;
   1803 				status = emit_memload(compiler,
   1804 				    BJ_XREG, pc->k, extwords);
   1805 				if (status != SLJIT_SUCCESS)
   1806 					goto fail;
   1807 
   1808 				continue;
   1809 			}
   1810 
   1811 			/* BPF_LDX+BPF_B+BPF_MSH    X <- 4*(P[k:1]&0xf) */
   1812 			if (mode != BPF_MSH || BPF_SIZE(pc->code) != BPF_B)
   1813 				goto fail;
   1814 
   1815 			if (unconditional_ret)
   1816 				continue;
   1817 
   1818 			status = emit_msh(compiler, pc,
   1819 			    to_mchain_jump, &ret0, &ret0_size, &ret0_maxsize);
   1820 			if (status != SLJIT_SUCCESS)
   1821 				goto fail;
   1822 
   1823 			continue;
   1824 
   1825 		case BPF_ST:
   1826 			if (pc->code != BPF_ST ||
   1827 			    (uint32_t)pc->k >= memwords) {
   1828 				goto fail;
   1829 			}
   1830 
   1831 			status = emit_memstore(compiler,
   1832 			    BJ_AREG, pc->k, extwords);
   1833 			if (status != SLJIT_SUCCESS)
   1834 				goto fail;
   1835 
   1836 			continue;
   1837 
   1838 		case BPF_STX:
   1839 			if (pc->code != BPF_STX ||
   1840 			    (uint32_t)pc->k >= memwords) {
   1841 				goto fail;
   1842 			}
   1843 
   1844 			status = emit_memstore(compiler,
   1845 			    BJ_XREG, pc->k, extwords);
   1846 			if (status != SLJIT_SUCCESS)
   1847 				goto fail;
   1848 
   1849 			continue;
   1850 
   1851 		case BPF_ALU:
   1852 			if (pc->code == (BPF_ALU|BPF_NEG)) {
   1853 				status = sljit_emit_op1(compiler,
   1854 				    SLJIT_NEG,
   1855 				    BJ_AREG, 0,
   1856 				    BJ_AREG, 0);
   1857 				if (status != SLJIT_SUCCESS)
   1858 					goto fail;
   1859 
   1860 				continue;
   1861 			}
   1862 
   1863 			if (BPF_OP(pc->code) != BPF_DIV) {
   1864 				status = sljit_emit_op2(compiler,
   1865 				    bpf_alu_to_sljit_op(pc),
   1866 				    BJ_AREG, 0,
   1867 				    BJ_AREG, 0,
   1868 				    kx_to_reg(pc), kx_to_reg_arg(pc));
   1869 				if (status != SLJIT_SUCCESS)
   1870 					goto fail;
   1871 
   1872 				continue;
   1873 			}
   1874 
   1875 			/* BPF_DIV */
   1876 
   1877 			src = BPF_SRC(pc->code);
   1878 			if (src != BPF_X && src != BPF_K)
   1879 				goto fail;
   1880 
   1881 			/* division by zero? */
   1882 			if (src == BPF_X) {
   1883 				jump = sljit_emit_cmp(compiler,
   1884 				    SLJIT_C_EQUAL|SLJIT_INT_OP,
   1885 				    BJ_XREG, 0,
   1886 				    SLJIT_IMM, 0);
   1887 				if (jump == NULL)
   1888 					goto fail;
   1889 				if (!append_jump(jump, &ret0,
   1890 				    &ret0_size, &ret0_maxsize))
   1891 					goto fail;
   1892 			} else if (pc->k == 0) {
   1893 				jump = sljit_emit_jump(compiler, SLJIT_JUMP);
   1894 				if (jump == NULL)
   1895 					goto fail;
   1896 				if (!append_jump(jump, &ret0,
   1897 				    &ret0_size, &ret0_maxsize))
   1898 					goto fail;
   1899 			}
   1900 
   1901 			if (src == BPF_X) {
   1902 				status = emit_division(compiler, BJ_XREG, 0);
   1903 				if (status != SLJIT_SUCCESS)
   1904 					goto fail;
   1905 			} else if (pc->k != 0) {
   1906 				if (pc->k & (pc->k - 1)) {
   1907 				    status = emit_division(compiler,
   1908 				        SLJIT_IMM, (uint32_t)pc->k);
   1909 				} else {
   1910 				    status = emit_pow2_division(compiler,
   1911 				        (uint32_t)pc->k);
   1912 				}
   1913 				if (status != SLJIT_SUCCESS)
   1914 					goto fail;
   1915 			}
   1916 
   1917 			continue;
   1918 
   1919 		case BPF_JMP:
   1920 			if (BPF_OP(pc->code) == BPF_JA) {
   1921 				jt = jf = pc->k;
   1922 			} else {
   1923 				jt = pc->jt;
   1924 				jf = pc->jf;
   1925 			}
   1926 
   1927 			negate = (jt == 0) ? 1 : 0;
   1928 			branching = (jt == jf) ? 0 : 1;
   1929 			jtf = insn_dat[i].u.jdata.jtf;
   1930 
   1931 			if (branching) {
   1932 				if (BPF_OP(pc->code) != BPF_JSET) {
   1933 					jump = sljit_emit_cmp(compiler,
   1934 					    bpf_jmp_to_sljit_cond(pc, negate),
   1935 					    BJ_AREG, 0,
   1936 					    kx_to_reg(pc), kx_to_reg_arg(pc));
   1937 				} else {
   1938 					status = sljit_emit_op2(compiler,
   1939 					    SLJIT_AND,
   1940 					    BJ_TMP1REG, 0,
   1941 					    BJ_AREG, 0,
   1942 					    kx_to_reg(pc), kx_to_reg_arg(pc));
   1943 					if (status != SLJIT_SUCCESS)
   1944 						goto fail;
   1945 
   1946 					jump = sljit_emit_cmp(compiler,
   1947 					    bpf_jmp_to_sljit_cond(pc, negate),
   1948 					    BJ_TMP1REG, 0,
   1949 					    SLJIT_IMM, 0);
   1950 				}
   1951 
   1952 				if (jump == NULL)
   1953 					goto fail;
   1954 
   1955 				BJ_ASSERT(jtf[negate].sjump == NULL);
   1956 				jtf[negate].sjump = jump;
   1957 			}
   1958 
   1959 			if (!branching || (jt != 0 && jf != 0)) {
   1960 				jump = sljit_emit_jump(compiler, SLJIT_JUMP);
   1961 				if (jump == NULL)
   1962 					goto fail;
   1963 
   1964 				BJ_ASSERT(jtf[branching].sjump == NULL);
   1965 				jtf[branching].sjump = jump;
   1966 			}
   1967 
   1968 			continue;
   1969 
   1970 		case BPF_RET:
   1971 			rval = BPF_RVAL(pc->code);
   1972 			if (rval == BPF_X)
   1973 				goto fail;
   1974 
   1975 			/* BPF_RET+BPF_K    accept k bytes */
   1976 			if (rval == BPF_K) {
   1977 				status = sljit_emit_return(compiler,
   1978 				    SLJIT_MOV_UI,
   1979 				    SLJIT_IMM, (uint32_t)pc->k);
   1980 				if (status != SLJIT_SUCCESS)
   1981 					goto fail;
   1982 			}
   1983 
   1984 			/* BPF_RET+BPF_A    accept A bytes */
   1985 			if (rval == BPF_A) {
   1986 				status = sljit_emit_return(compiler,
   1987 				    SLJIT_MOV_UI,
   1988 				    BJ_AREG, 0);
   1989 				if (status != SLJIT_SUCCESS)
   1990 					goto fail;
   1991 			}
   1992 
   1993 			continue;
   1994 
   1995 		case BPF_MISC:
   1996 			switch (BPF_MISCOP(pc->code)) {
   1997 			case BPF_TAX:
   1998 				status = sljit_emit_op1(compiler,
   1999 				    SLJIT_MOV_UI,
   2000 				    BJ_XREG, 0,
   2001 				    BJ_AREG, 0);
   2002 				if (status != SLJIT_SUCCESS)
   2003 					goto fail;
   2004 
   2005 				continue;
   2006 
   2007 			case BPF_TXA:
   2008 				status = sljit_emit_op1(compiler,
   2009 				    SLJIT_MOV,
   2010 				    BJ_AREG, 0,
   2011 				    BJ_XREG, 0);
   2012 				if (status != SLJIT_SUCCESS)
   2013 					goto fail;
   2014 
   2015 				continue;
   2016 
   2017 			case BPF_COP:
   2018 			case BPF_COPX:
   2019 				if (bc == NULL || bc->copfuncs == NULL)
   2020 					goto fail;
   2021 				if (BPF_MISCOP(pc->code) == BPF_COP &&
   2022 				    (uint32_t)pc->k >= bc->nfuncs) {
   2023 					goto fail;
   2024 				}
   2025 
   2026 				status = emit_cop(compiler, bc, pc,
   2027 				    &ret0, &ret0_size, &ret0_maxsize);
   2028 				if (status != SLJIT_SUCCESS)
   2029 					goto fail;
   2030 
   2031 				continue;
   2032 			}
   2033 
   2034 			goto fail;
   2035 		} /* switch */
   2036 	} /* main loop */
   2037 
   2038 	BJ_ASSERT(ret0_size <= ret0_maxsize);
   2039 
   2040 	if (ret0_size > 0) {
   2041 		label = sljit_emit_label(compiler);
   2042 		if (label == NULL)
   2043 			goto fail;
   2044 		for (i = 0; i < ret0_size; i++)
   2045 			sljit_set_label(ret0[i], label);
   2046 	}
   2047 
   2048 	status = sljit_emit_return(compiler,
   2049 	    SLJIT_MOV_UI,
   2050 	    SLJIT_IMM, 0);
   2051 	if (status != SLJIT_SUCCESS)
   2052 		goto fail;
   2053 
   2054 	rv = true;
   2055 
   2056 fail:
   2057 	if (ret0 != NULL)
   2058 		BJ_FREE(ret0, ret0_maxsize * sizeof(ret0[0]));
   2059 
   2060 	return rv;
   2061 }
   2062 
   2063 bpfjit_func_t
   2064 bpfjit_generate_code(const bpf_ctx_t *bc,
   2065     const struct bpf_insn *insns, size_t insn_count)
   2066 {
   2067 	void *rv;
   2068 	struct sljit_compiler *compiler;
   2069 
   2070 	size_t i;
   2071 	int status;
   2072 
   2073 	/* optimization related */
   2074 	bpf_memword_init_t initmask;
   2075 	bpfjit_hint_t hints;
   2076 
   2077 	/* memory store location for initial zero initialization */
   2078 	sljit_si mem_reg;
   2079 	sljit_sw mem_off;
   2080 
   2081 	struct bpfjit_insn_data *insn_dat;
   2082 
   2083 	const size_t extwords = GET_EXTWORDS(bc);
   2084 	const size_t memwords = GET_MEMWORDS(bc);
   2085 	const bpf_memword_init_t preinited = extwords ? bc->preinited : 0;
   2086 
   2087 	rv = NULL;
   2088 	compiler = NULL;
   2089 	insn_dat = NULL;
   2090 
   2091 	if (memwords > MAX_MEMWORDS)
   2092 		goto fail;
   2093 
   2094 	if (insn_count == 0 || insn_count > SIZE_MAX / sizeof(insn_dat[0]))
   2095 		goto fail;
   2096 
   2097 	insn_dat = BJ_ALLOC(insn_count * sizeof(insn_dat[0]));
   2098 	if (insn_dat == NULL)
   2099 		goto fail;
   2100 
   2101 	if (!optimize(bc, insns, insn_dat, insn_count, &initmask, &hints))
   2102 		goto fail;
   2103 
   2104 	compiler = sljit_create_compiler();
   2105 	if (compiler == NULL)
   2106 		goto fail;
   2107 
   2108 #if !defined(_KERNEL) && defined(SLJIT_VERBOSE) && SLJIT_VERBOSE
   2109 	sljit_compiler_verbose(compiler, stderr);
   2110 #endif
   2111 
   2112 	status = sljit_emit_enter(compiler,
   2113 	    2, nscratches(hints), 3, sizeof(struct bpfjit_stack));
   2114 	if (status != SLJIT_SUCCESS)
   2115 		goto fail;
   2116 
   2117 	if (hints & BJ_HINT_COP) {
   2118 		/* save ctx argument */
   2119 		status = sljit_emit_op1(compiler,
   2120 		    SLJIT_MOV_P,
   2121 		    SLJIT_MEM1(SLJIT_LOCALS_REG),
   2122 		    offsetof(struct bpfjit_stack, ctx),
   2123 		    BJ_CTX_ARG, 0);
   2124 		if (status != SLJIT_SUCCESS)
   2125 			goto fail;
   2126 	}
   2127 
   2128 	if (extwords == 0) {
   2129 		mem_reg = SLJIT_MEM1(SLJIT_LOCALS_REG);
   2130 		mem_off = offsetof(struct bpfjit_stack, mem);
   2131 	} else {
   2132 		/* copy "mem" argument from bpf_args to bpfjit_stack */
   2133 		status = sljit_emit_op1(compiler,
   2134 		    SLJIT_MOV_P,
   2135 		    BJ_TMP1REG, 0,
   2136 		    SLJIT_MEM1(BJ_ARGS), offsetof(struct bpf_args, mem));
   2137 		if (status != SLJIT_SUCCESS)
   2138 			goto fail;
   2139 
   2140 		status = sljit_emit_op1(compiler,
   2141 		    SLJIT_MOV_P,
   2142 		    SLJIT_MEM1(SLJIT_LOCALS_REG),
   2143 		    offsetof(struct bpfjit_stack, extmem),
   2144 		    BJ_TMP1REG, 0);
   2145 		if (status != SLJIT_SUCCESS)
   2146 			goto fail;
   2147 
   2148 		mem_reg = SLJIT_MEM1(BJ_TMP1REG);
   2149 		mem_off = 0;
   2150 	}
   2151 
   2152 	/*
   2153 	 * Exclude pre-initialised external memory words but keep
   2154 	 * initialization statuses of A and X registers in case
   2155 	 * bc->preinited wrongly sets those two bits.
   2156 	 */
   2157 	initmask &= ~preinited | BJ_INIT_ABIT | BJ_INIT_XBIT;
   2158 
   2159 #if defined(_KERNEL)
   2160 	/* bpf_filter() checks initialization of memwords. */
   2161 	BJ_ASSERT((initmask & (BJ_INIT_MBIT(memwords) - 1)) == 0);
   2162 #endif
   2163 	for (i = 0; i < memwords; i++) {
   2164 		if (initmask & BJ_INIT_MBIT(i)) {
   2165 			/* M[i] = 0; */
   2166 			status = sljit_emit_op1(compiler,
   2167 			    SLJIT_MOV_UI,
   2168 			    mem_reg, mem_off + i * sizeof(uint32_t),
   2169 			    SLJIT_IMM, 0);
   2170 			if (status != SLJIT_SUCCESS)
   2171 				goto fail;
   2172 		}
   2173 	}
   2174 
   2175 	if (initmask & BJ_INIT_ABIT) {
   2176 		/* A = 0; */
   2177 		status = sljit_emit_op1(compiler,
   2178 		    SLJIT_MOV,
   2179 		    BJ_AREG, 0,
   2180 		    SLJIT_IMM, 0);
   2181 		if (status != SLJIT_SUCCESS)
   2182 			goto fail;
   2183 	}
   2184 
   2185 	if (initmask & BJ_INIT_XBIT) {
   2186 		/* X = 0; */
   2187 		status = sljit_emit_op1(compiler,
   2188 		    SLJIT_MOV,
   2189 		    BJ_XREG, 0,
   2190 		    SLJIT_IMM, 0);
   2191 		if (status != SLJIT_SUCCESS)
   2192 			goto fail;
   2193 	}
   2194 
   2195 	status = load_buf_buflen(compiler);
   2196 	if (status != SLJIT_SUCCESS)
   2197 		goto fail;
   2198 
   2199 	if (!generate_insn_code(compiler, bc, insns, insn_dat, insn_count))
   2200 		goto fail;
   2201 
   2202 	rv = sljit_generate_code(compiler);
   2203 
   2204 fail:
   2205 	if (compiler != NULL)
   2206 		sljit_free_compiler(compiler);
   2207 
   2208 	if (insn_dat != NULL)
   2209 		BJ_FREE(insn_dat, insn_count * sizeof(insn_dat[0]));
   2210 
   2211 	return (bpfjit_func_t)rv;
   2212 }
   2213 
   2214 void
   2215 bpfjit_free_code(bpfjit_func_t code)
   2216 {
   2217 
   2218 	sljit_free_code((void *)code);
   2219 }
   2220