linux/arch/mips/net/bpf_jit_comp32.c

1907 lines
54 KiB
C
Raw Normal View History

mips, bpf: Add eBPF JIT for 32-bit MIPS This is an implementation of an eBPF JIT for 32-bit MIPS I-V and MIPS32. The implementation supports all 32-bit and 64-bit ALU and JMP operations, including the recently-added atomics. 64-bit div/mod and 64-bit atomics are implemented using function calls to math64 and atomic64 functions, respectively. All 32-bit operations are implemented natively by the JIT, except if the CPU lacks ll/sc instructions. Register mapping ================ All 64-bit eBPF registers are mapped to native 32-bit MIPS register pairs, and does not use any stack scratch space for register swapping. This means that all eBPF register data is kept in CPU registers all the time, and this simplifies the register management a lot. It also reduces the JIT's pressure on temporary registers since we do not have to move data around. Native register pairs are ordered according to CPU endiannes, following the O32 calling convention for passing 64-bit arguments and return values. The eBPF return value, arguments and callee-saved registers are mapped to their native MIPS equivalents. Since the 32 highest bits in the eBPF FP (frame pointer) register are always zero, only one general-purpose register is actually needed for the mapping. The MIPS fp register is used for this purpose. The high bits are mapped to MIPS register r0. This saves us one CPU register, which is much needed for temporaries, while still allowing us to treat the R10 (FP) register just like any other eBPF register in the JIT. The MIPS gp (global pointer) and at (assembler temporary) registers are used as internal temporary registers for constant blinding. CPU registers t6-t9 are used internally by the JIT when constructing more complex 64-bit operations. This is precisely what is needed - two registers to store an operand value, and two more as scratch registers when performing the operation. The register mapping is shown below. R0 - $v1, $v0 return value R1 - $a1, $a0 argument 1, passed in registers R2 - $a3, $a2 argument 2, passed in registers R3 - $t1, $t0 argument 3, passed on stack R4 - $t3, $t2 argument 4, passed on stack R5 - $t4, $t3 argument 5, passed on stack R6 - $s1, $s0 callee-saved R7 - $s3, $s2 callee-saved R8 - $s5, $s4 callee-saved R9 - $s7, $s6 callee-saved FP - $r0, $fp 32-bit frame pointer AX - $gp, $at constant-blinding $t6 - $t9 unallocated, JIT temporaries Jump offsets ============ The JIT tries to map all conditional JMP operations to MIPS conditional PC-relative branches. The MIPS branch offset field is 18 bits, in bytes, which is equivalent to the eBPF 16-bit instruction offset. However, since the JIT may emit more than one CPU instruction per eBPF instruction, the field width may overflow. If that happens, the JIT converts the long conditional jump to a short PC-relative branch with the condition inverted, jumping over a long unconditional absolute jmp (j). This conversion will change the instruction offset mapping used for jumps, and may in turn result in more branch offset overflows. The JIT therefore dry-runs the translation until no more branches are converted and the offsets do not change anymore. There is an upper bound on this of course, and if the JIT hits that limit, the last two iterations are run with all branches being converted. Tail call count =============== The current tail call count is stored in the 16-byte area of the caller's stack frame that is reserved for the callee in the o32 ABI. The value is initialized in the prologue, and propagated to the tail-callee by skipping the initialization instructions when emitting the tail call. Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20211005165408.2305108-4-johan.almbladh@anyfinetworks.com
2021-10-06 00:54:04 +08:00
// SPDX-License-Identifier: GPL-2.0-only
/*
* Just-In-Time compiler for eBPF bytecode on MIPS.
* Implementation of JIT functions for 32-bit CPUs.
*
* Copyright (c) 2021 Anyfi Networks AB.
* Author: Johan Almbladh <johan.almbladh@gmail.com>
*
* Based on code and ideas from
* Copyright (c) 2017 Cavium, Inc.
* Copyright (c) 2017 Shubham Bansal <illusionist.neo@gmail.com>
* Copyright (c) 2011 Mircea Gherzan <mgherzan@gmail.com>
*/
#include <linux/math64.h>
#include <linux/errno.h>
#include <linux/filter.h>
#include <linux/bpf.h>
#include <asm/cpu-features.h>
#include <asm/isa-rev.h>
#include <asm/uasm.h>
#include "bpf_jit_comp.h"
/* MIPS a4-a7 are not available in the o32 ABI */
#undef MIPS_R_A4
#undef MIPS_R_A5
#undef MIPS_R_A6
#undef MIPS_R_A7
/* Stack is 8-byte aligned in o32 ABI */
#define MIPS_STACK_ALIGNMENT 8
/*
* The top 16 bytes of a stack frame is reserved for the callee in O32 ABI.
* This corresponds to stack space for register arguments a0-a3.
*/
#define JIT_RESERVED_STACK 16
/* Temporary 64-bit register used by JIT */
#define JIT_REG_TMP MAX_BPF_JIT_REG
/*
* Number of prologue bytes to skip when doing a tail call.
* Tail call count (TCC) initialization (8 bytes) always, plus
* R0-to-v0 assignment (4 bytes) if big endian.
*/
#ifdef __BIG_ENDIAN
#define JIT_TCALL_SKIP 12
#else
#define JIT_TCALL_SKIP 8
#endif
/* CPU registers holding the callee return value */
#define JIT_RETURN_REGS \
(BIT(MIPS_R_V0) | \
BIT(MIPS_R_V1))
/* CPU registers arguments passed to callee directly */
#define JIT_ARG_REGS \
(BIT(MIPS_R_A0) | \
BIT(MIPS_R_A1) | \
BIT(MIPS_R_A2) | \
BIT(MIPS_R_A3))
/* CPU register arguments passed to callee on stack */
#define JIT_STACK_REGS \
(BIT(MIPS_R_T0) | \
BIT(MIPS_R_T1) | \
BIT(MIPS_R_T2) | \
BIT(MIPS_R_T3) | \
BIT(MIPS_R_T4) | \
BIT(MIPS_R_T5))
/* Caller-saved CPU registers */
#define JIT_CALLER_REGS \
(JIT_RETURN_REGS | \
JIT_ARG_REGS | \
JIT_STACK_REGS)
/* Callee-saved CPU registers */
#define JIT_CALLEE_REGS \
(BIT(MIPS_R_S0) | \
BIT(MIPS_R_S1) | \
BIT(MIPS_R_S2) | \
BIT(MIPS_R_S3) | \
BIT(MIPS_R_S4) | \
BIT(MIPS_R_S5) | \
BIT(MIPS_R_S6) | \
BIT(MIPS_R_S7) | \
BIT(MIPS_R_GP) | \
BIT(MIPS_R_FP) | \
BIT(MIPS_R_RA))
/*
* Mapping of 64-bit eBPF registers to 32-bit native MIPS registers.
*
* 1) Native register pairs are ordered according to CPU endiannes, following
* the MIPS convention for passing 64-bit arguments and return values.
* 2) The eBPF return value, arguments and callee-saved registers are mapped
* to their native MIPS equivalents.
* 3) Since the 32 highest bits in the eBPF FP register are always zero,
* only one general-purpose register is actually needed for the mapping.
* We use the fp register for this purpose, and map the highest bits to
* the MIPS register r0 (zero).
* 4) We use the MIPS gp and at registers as internal temporary registers
* for constant blinding. The gp register is callee-saved.
* 5) One 64-bit temporary register is mapped for use when sign-extending
* immediate operands. MIPS registers t6-t9 are available to the JIT
* for as temporaries when implementing complex 64-bit operations.
*
* With this scheme all eBPF registers are being mapped to native MIPS
* registers without having to use any stack scratch space. The direct
* register mapping (2) simplifies the handling of function calls.
*/
static const u8 bpf2mips32[][2] = {
/* Return value from in-kernel function, and exit value from eBPF */
[BPF_REG_0] = {MIPS_R_V1, MIPS_R_V0},
/* Arguments from eBPF program to in-kernel function */
[BPF_REG_1] = {MIPS_R_A1, MIPS_R_A0},
[BPF_REG_2] = {MIPS_R_A3, MIPS_R_A2},
/* Remaining arguments, to be passed on the stack per O32 ABI */
[BPF_REG_3] = {MIPS_R_T1, MIPS_R_T0},
[BPF_REG_4] = {MIPS_R_T3, MIPS_R_T2},
[BPF_REG_5] = {MIPS_R_T5, MIPS_R_T4},
/* Callee-saved registers that in-kernel function will preserve */
[BPF_REG_6] = {MIPS_R_S1, MIPS_R_S0},
[BPF_REG_7] = {MIPS_R_S3, MIPS_R_S2},
[BPF_REG_8] = {MIPS_R_S5, MIPS_R_S4},
[BPF_REG_9] = {MIPS_R_S7, MIPS_R_S6},
/* Read-only frame pointer to access the eBPF stack */
#ifdef __BIG_ENDIAN
[BPF_REG_FP] = {MIPS_R_FP, MIPS_R_ZERO},
#else
[BPF_REG_FP] = {MIPS_R_ZERO, MIPS_R_FP},
#endif
/* Temporary register for blinding constants */
[BPF_REG_AX] = {MIPS_R_GP, MIPS_R_AT},
/* Temporary register for internal JIT use */
[JIT_REG_TMP] = {MIPS_R_T7, MIPS_R_T6},
};
/* Get low CPU register for a 64-bit eBPF register mapping */
static inline u8 lo(const u8 reg[])
{
#ifdef __BIG_ENDIAN
return reg[0];
#else
return reg[1];
#endif
}
/* Get high CPU register for a 64-bit eBPF register mapping */
static inline u8 hi(const u8 reg[])
{
#ifdef __BIG_ENDIAN
return reg[1];
#else
return reg[0];
#endif
}
/*
* Mark a 64-bit CPU register pair as clobbered, it needs to be
* saved/restored by the program if callee-saved.
*/
static void clobber_reg64(struct jit_context *ctx, const u8 reg[])
{
clobber_reg(ctx, reg[0]);
clobber_reg(ctx, reg[1]);
}
/* dst = imm (sign-extended) */
static void emit_mov_se_i64(struct jit_context *ctx, const u8 dst[], s32 imm)
{
emit_mov_i(ctx, lo(dst), imm);
if (imm < 0)
emit(ctx, addiu, hi(dst), MIPS_R_ZERO, -1);
else
emit(ctx, move, hi(dst), MIPS_R_ZERO);
clobber_reg64(ctx, dst);
}
/* Zero extension, if verifier does not do it for us */
static void emit_zext_ver(struct jit_context *ctx, const u8 dst[])
{
if (!ctx->program->aux->verifier_zext) {
emit(ctx, move, hi(dst), MIPS_R_ZERO);
clobber_reg(ctx, hi(dst));
}
}
/* Load delay slot, if ISA mandates it */
static void emit_load_delay(struct jit_context *ctx)
{
if (!cpu_has_mips_2_3_4_5_r)
emit(ctx, nop);
}
/* ALU immediate operation (64-bit) */
static void emit_alu_i64(struct jit_context *ctx,
const u8 dst[], s32 imm, u8 op)
{
u8 src = MIPS_R_T6;
/*
* ADD/SUB with all but the max negative imm can be handled by
* inverting the operation and the imm value, saving one insn.
*/
if (imm > S32_MIN && imm < 0)
switch (op) {
case BPF_ADD:
op = BPF_SUB;
imm = -imm;
break;
case BPF_SUB:
op = BPF_ADD;
imm = -imm;
break;
}
/* Move immediate to temporary register */
emit_mov_i(ctx, src, imm);
switch (op) {
/* dst = dst + imm */
case BPF_ADD:
emit(ctx, addu, lo(dst), lo(dst), src);
emit(ctx, sltu, MIPS_R_T9, lo(dst), src);
emit(ctx, addu, hi(dst), hi(dst), MIPS_R_T9);
if (imm < 0)
emit(ctx, addiu, hi(dst), hi(dst), -1);
break;
/* dst = dst - imm */
case BPF_SUB:
emit(ctx, sltu, MIPS_R_T9, lo(dst), src);
emit(ctx, subu, lo(dst), lo(dst), src);
emit(ctx, subu, hi(dst), hi(dst), MIPS_R_T9);
if (imm < 0)
emit(ctx, addiu, hi(dst), hi(dst), 1);
break;
/* dst = dst | imm */
case BPF_OR:
emit(ctx, or, lo(dst), lo(dst), src);
if (imm < 0)
emit(ctx, addiu, hi(dst), MIPS_R_ZERO, -1);
break;
/* dst = dst & imm */
case BPF_AND:
emit(ctx, and, lo(dst), lo(dst), src);
if (imm >= 0)
emit(ctx, move, hi(dst), MIPS_R_ZERO);
break;
/* dst = dst ^ imm */
case BPF_XOR:
emit(ctx, xor, lo(dst), lo(dst), src);
if (imm < 0) {
emit(ctx, subu, hi(dst), MIPS_R_ZERO, hi(dst));
emit(ctx, addiu, hi(dst), hi(dst), -1);
}
break;
}
clobber_reg64(ctx, dst);
}
/* ALU register operation (64-bit) */
static void emit_alu_r64(struct jit_context *ctx,
const u8 dst[], const u8 src[], u8 op)
{
switch (BPF_OP(op)) {
/* dst = dst + src */
case BPF_ADD:
if (src == dst) {
emit(ctx, srl, MIPS_R_T9, lo(dst), 31);
emit(ctx, addu, lo(dst), lo(dst), lo(dst));
} else {
emit(ctx, addu, lo(dst), lo(dst), lo(src));
emit(ctx, sltu, MIPS_R_T9, lo(dst), lo(src));
}
emit(ctx, addu, hi(dst), hi(dst), hi(src));
emit(ctx, addu, hi(dst), hi(dst), MIPS_R_T9);
break;
/* dst = dst - src */
case BPF_SUB:
emit(ctx, sltu, MIPS_R_T9, lo(dst), lo(src));
emit(ctx, subu, lo(dst), lo(dst), lo(src));
emit(ctx, subu, hi(dst), hi(dst), hi(src));
emit(ctx, subu, hi(dst), hi(dst), MIPS_R_T9);
break;
/* dst = dst | src */
case BPF_OR:
emit(ctx, or, lo(dst), lo(dst), lo(src));
emit(ctx, or, hi(dst), hi(dst), hi(src));
break;
/* dst = dst & src */
case BPF_AND:
emit(ctx, and, lo(dst), lo(dst), lo(src));
emit(ctx, and, hi(dst), hi(dst), hi(src));
break;
/* dst = dst ^ src */
case BPF_XOR:
emit(ctx, xor, lo(dst), lo(dst), lo(src));
emit(ctx, xor, hi(dst), hi(dst), hi(src));
break;
}
clobber_reg64(ctx, dst);
}
/* ALU invert (64-bit) */
static void emit_neg_i64(struct jit_context *ctx, const u8 dst[])
{
emit(ctx, sltu, MIPS_R_T9, MIPS_R_ZERO, lo(dst));
emit(ctx, subu, lo(dst), MIPS_R_ZERO, lo(dst));
emit(ctx, subu, hi(dst), MIPS_R_ZERO, hi(dst));
emit(ctx, subu, hi(dst), hi(dst), MIPS_R_T9);
clobber_reg64(ctx, dst);
}
/* ALU shift immediate (64-bit) */
static void emit_shift_i64(struct jit_context *ctx,
const u8 dst[], u32 imm, u8 op)
{
switch (BPF_OP(op)) {
/* dst = dst << imm */
case BPF_LSH:
if (imm < 32) {
emit(ctx, srl, MIPS_R_T9, lo(dst), 32 - imm);
emit(ctx, sll, lo(dst), lo(dst), imm);
emit(ctx, sll, hi(dst), hi(dst), imm);
emit(ctx, or, hi(dst), hi(dst), MIPS_R_T9);
} else {
emit(ctx, sll, hi(dst), lo(dst), imm - 32);
emit(ctx, move, lo(dst), MIPS_R_ZERO);
}
break;
/* dst = dst >> imm */
case BPF_RSH:
if (imm < 32) {
emit(ctx, sll, MIPS_R_T9, hi(dst), 32 - imm);
emit(ctx, srl, lo(dst), lo(dst), imm);
emit(ctx, srl, hi(dst), hi(dst), imm);
emit(ctx, or, lo(dst), lo(dst), MIPS_R_T9);
} else {
emit(ctx, srl, lo(dst), hi(dst), imm - 32);
emit(ctx, move, hi(dst), MIPS_R_ZERO);
}
break;
/* dst = dst >> imm (arithmetic) */
case BPF_ARSH:
if (imm < 32) {
emit(ctx, sll, MIPS_R_T9, hi(dst), 32 - imm);
emit(ctx, srl, lo(dst), lo(dst), imm);
emit(ctx, sra, hi(dst), hi(dst), imm);
emit(ctx, or, lo(dst), lo(dst), MIPS_R_T9);
} else {
emit(ctx, sra, lo(dst), hi(dst), imm - 32);
emit(ctx, sra, hi(dst), hi(dst), 31);
}
break;
}
clobber_reg64(ctx, dst);
}
/* ALU shift register (64-bit) */
static void emit_shift_r64(struct jit_context *ctx,
const u8 dst[], u8 src, u8 op)
{
u8 t1 = MIPS_R_T8;
u8 t2 = MIPS_R_T9;
emit(ctx, andi, t1, src, 32); /* t1 = src & 32 */
emit(ctx, beqz, t1, 16); /* PC += 16 if t1 == 0 */
emit(ctx, nor, t2, src, MIPS_R_ZERO); /* t2 = ~src (delay slot) */
switch (BPF_OP(op)) {
/* dst = dst << src */
case BPF_LSH:
/* Next: shift >= 32 */
emit(ctx, sllv, hi(dst), lo(dst), src); /* dh = dl << src */
emit(ctx, move, lo(dst), MIPS_R_ZERO); /* dl = 0 */
emit(ctx, b, 20); /* PC += 20 */
/* +16: shift < 32 */
emit(ctx, srl, t1, lo(dst), 1); /* t1 = dl >> 1 */
emit(ctx, srlv, t1, t1, t2); /* t1 = t1 >> t2 */
emit(ctx, sllv, lo(dst), lo(dst), src); /* dl = dl << src */
emit(ctx, sllv, hi(dst), hi(dst), src); /* dh = dh << src */
emit(ctx, or, hi(dst), hi(dst), t1); /* dh = dh | t1 */
break;
/* dst = dst >> src */
case BPF_RSH:
/* Next: shift >= 32 */
emit(ctx, srlv, lo(dst), hi(dst), src); /* dl = dh >> src */
emit(ctx, move, hi(dst), MIPS_R_ZERO); /* dh = 0 */
emit(ctx, b, 20); /* PC += 20 */
/* +16: shift < 32 */
emit(ctx, sll, t1, hi(dst), 1); /* t1 = dl << 1 */
emit(ctx, sllv, t1, t1, t2); /* t1 = t1 << t2 */
emit(ctx, srlv, lo(dst), lo(dst), src); /* dl = dl >> src */
emit(ctx, srlv, hi(dst), hi(dst), src); /* dh = dh >> src */
emit(ctx, or, lo(dst), lo(dst), t1); /* dl = dl | t1 */
break;
/* dst = dst >> src (arithmetic) */
case BPF_ARSH:
/* Next: shift >= 32 */
emit(ctx, srav, lo(dst), hi(dst), src); /* dl = dh >>a src */
emit(ctx, sra, hi(dst), hi(dst), 31); /* dh = dh >>a 31 */
emit(ctx, b, 20); /* PC += 20 */
/* +16: shift < 32 */
emit(ctx, sll, t1, hi(dst), 1); /* t1 = dl << 1 */
emit(ctx, sllv, t1, t1, t2); /* t1 = t1 << t2 */
emit(ctx, srlv, lo(dst), lo(dst), src); /* dl = dl >>a src */
emit(ctx, srav, hi(dst), hi(dst), src); /* dh = dh >> src */
emit(ctx, or, lo(dst), lo(dst), t1); /* dl = dl | t1 */
break;
}
/* +20: Done */
clobber_reg64(ctx, dst);
}
/* ALU mul immediate (64x32-bit) */
static void emit_mul_i64(struct jit_context *ctx, const u8 dst[], s32 imm)
{
u8 src = MIPS_R_T6;
u8 tmp = MIPS_R_T9;
switch (imm) {
/* dst = dst * 1 is a no-op */
case 1:
break;
/* dst = dst * -1 */
case -1:
emit_neg_i64(ctx, dst);
break;
case 0:
emit_mov_r(ctx, lo(dst), MIPS_R_ZERO);
emit_mov_r(ctx, hi(dst), MIPS_R_ZERO);
break;
/* Full 64x32 multiply */
default:
/* hi(dst) = hi(dst) * src(imm) */
emit_mov_i(ctx, src, imm);
if (cpu_has_mips32r1 || cpu_has_mips32r6) {
emit(ctx, mul, hi(dst), hi(dst), src);
} else {
emit(ctx, multu, hi(dst), src);
emit(ctx, mflo, hi(dst));
}
/* hi(dst) = hi(dst) - lo(dst) */
if (imm < 0)
emit(ctx, subu, hi(dst), hi(dst), lo(dst));
/* tmp = lo(dst) * src(imm) >> 32 */
/* lo(dst) = lo(dst) * src(imm) */
if (cpu_has_mips32r6) {
emit(ctx, muhu, tmp, lo(dst), src);
emit(ctx, mulu, lo(dst), lo(dst), src);
} else {
emit(ctx, multu, lo(dst), src);
emit(ctx, mflo, lo(dst));
emit(ctx, mfhi, tmp);
}
/* hi(dst) += tmp */
emit(ctx, addu, hi(dst), hi(dst), tmp);
clobber_reg64(ctx, dst);
break;
}
}
/* ALU mul register (64x64-bit) */
static void emit_mul_r64(struct jit_context *ctx,
const u8 dst[], const u8 src[])
{
u8 acc = MIPS_R_T8;
u8 tmp = MIPS_R_T9;
/* acc = hi(dst) * lo(src) */
if (cpu_has_mips32r1 || cpu_has_mips32r6) {
emit(ctx, mul, acc, hi(dst), lo(src));
} else {
emit(ctx, multu, hi(dst), lo(src));
emit(ctx, mflo, acc);
}
/* tmp = lo(dst) * hi(src) */
if (cpu_has_mips32r1 || cpu_has_mips32r6) {
emit(ctx, mul, tmp, lo(dst), hi(src));
} else {
emit(ctx, multu, lo(dst), hi(src));
emit(ctx, mflo, tmp);
}
/* acc += tmp */
emit(ctx, addu, acc, acc, tmp);
/* tmp = lo(dst) * lo(src) >> 32 */
/* lo(dst) = lo(dst) * lo(src) */
if (cpu_has_mips32r6) {
emit(ctx, muhu, tmp, lo(dst), lo(src));
emit(ctx, mulu, lo(dst), lo(dst), lo(src));
} else {
emit(ctx, multu, lo(dst), lo(src));
emit(ctx, mflo, lo(dst));
emit(ctx, mfhi, tmp);
}
/* hi(dst) = acc + tmp */
emit(ctx, addu, hi(dst), acc, tmp);
clobber_reg64(ctx, dst);
}
/* Helper function for 64-bit modulo */
static u64 jit_mod64(u64 a, u64 b)
{
u64 rem;
div64_u64_rem(a, b, &rem);
return rem;
}
/* ALU div/mod register (64-bit) */
static void emit_divmod_r64(struct jit_context *ctx,
const u8 dst[], const u8 src[], u8 op)
{
const u8 *r0 = bpf2mips32[BPF_REG_0]; /* Mapped to v0-v1 */
const u8 *r1 = bpf2mips32[BPF_REG_1]; /* Mapped to a0-a1 */
const u8 *r2 = bpf2mips32[BPF_REG_2]; /* Mapped to a2-a3 */
int exclude, k;
u32 addr = 0;
/* Push caller-saved registers on stack */
push_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
0, JIT_RESERVED_STACK);
/* Put 64-bit arguments 1 and 2 in registers a0-a3 */
for (k = 0; k < 2; k++) {
emit(ctx, move, MIPS_R_T9, src[k]);
emit(ctx, move, r1[k], dst[k]);
emit(ctx, move, r2[k], MIPS_R_T9);
}
/* Emit function call */
switch (BPF_OP(op)) {
/* dst = dst / src */
case BPF_DIV:
addr = (u32)&div64_u64;
break;
/* dst = dst % src */
case BPF_MOD:
addr = (u32)&jit_mod64;
break;
}
emit_mov_i(ctx, MIPS_R_T9, addr);
emit(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
emit(ctx, nop); /* Delay slot */
/* Store the 64-bit result in dst */
emit(ctx, move, dst[0], r0[0]);
emit(ctx, move, dst[1], r0[1]);
/* Restore caller-saved registers, excluding the computed result */
exclude = BIT(lo(dst)) | BIT(hi(dst));
pop_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
exclude, JIT_RESERVED_STACK);
emit_load_delay(ctx);
clobber_reg64(ctx, dst);
clobber_reg(ctx, MIPS_R_V0);
clobber_reg(ctx, MIPS_R_V1);
clobber_reg(ctx, MIPS_R_RA);
}
/* Swap bytes in a register word */
static void emit_swap8_r(struct jit_context *ctx, u8 dst, u8 src, u8 mask)
{
u8 tmp = MIPS_R_T9;
emit(ctx, and, tmp, src, mask); /* tmp = src & 0x00ff00ff */
emit(ctx, sll, tmp, tmp, 8); /* tmp = tmp << 8 */
emit(ctx, srl, dst, src, 8); /* dst = src >> 8 */
emit(ctx, and, dst, dst, mask); /* dst = dst & 0x00ff00ff */
emit(ctx, or, dst, dst, tmp); /* dst = dst | tmp */
}
/* Swap half words in a register word */
static void emit_swap16_r(struct jit_context *ctx, u8 dst, u8 src)
{
u8 tmp = MIPS_R_T9;
emit(ctx, sll, tmp, src, 16); /* tmp = src << 16 */
emit(ctx, srl, dst, src, 16); /* dst = src >> 16 */
emit(ctx, or, dst, dst, tmp); /* dst = dst | tmp */
}
/* Swap bytes and truncate a register double word, word or half word */
static void emit_bswap_r64(struct jit_context *ctx, const u8 dst[], u32 width)
{
u8 tmp = MIPS_R_T8;
switch (width) {
/* Swap bytes in a double word */
case 64:
if (cpu_has_mips32r2 || cpu_has_mips32r6) {
emit(ctx, rotr, tmp, hi(dst), 16);
emit(ctx, rotr, hi(dst), lo(dst), 16);
emit(ctx, wsbh, lo(dst), tmp);
emit(ctx, wsbh, hi(dst), hi(dst));
} else {
emit_swap16_r(ctx, tmp, lo(dst));
emit_swap16_r(ctx, lo(dst), hi(dst));
emit(ctx, move, hi(dst), tmp);
emit(ctx, lui, tmp, 0xff); /* tmp = 0x00ff0000 */
emit(ctx, ori, tmp, tmp, 0xff); /* tmp = 0x00ff00ff */
emit_swap8_r(ctx, lo(dst), lo(dst), tmp);
emit_swap8_r(ctx, hi(dst), hi(dst), tmp);
}
break;
/* Swap bytes in a word */
/* Swap bytes in a half word */
case 32:
case 16:
emit_bswap_r(ctx, lo(dst), width);
emit(ctx, move, hi(dst), MIPS_R_ZERO);
break;
}
clobber_reg64(ctx, dst);
}
/* Truncate a register double word, word or half word */
static void emit_trunc_r64(struct jit_context *ctx, const u8 dst[], u32 width)
{
switch (width) {
case 64:
break;
/* Zero-extend a word */
case 32:
emit(ctx, move, hi(dst), MIPS_R_ZERO);
clobber_reg(ctx, hi(dst));
break;
/* Zero-extend a half word */
case 16:
emit(ctx, move, hi(dst), MIPS_R_ZERO);
emit(ctx, andi, lo(dst), lo(dst), 0xffff);
clobber_reg64(ctx, dst);
break;
}
}
/* Load operation: dst = *(size*)(src + off) */
static void emit_ldx(struct jit_context *ctx,
const u8 dst[], u8 src, s16 off, u8 size)
{
switch (size) {
/* Load a byte */
case BPF_B:
emit(ctx, lbu, lo(dst), off, src);
emit(ctx, move, hi(dst), MIPS_R_ZERO);
break;
/* Load a half word */
case BPF_H:
emit(ctx, lhu, lo(dst), off, src);
emit(ctx, move, hi(dst), MIPS_R_ZERO);
break;
/* Load a word */
case BPF_W:
emit(ctx, lw, lo(dst), off, src);
emit(ctx, move, hi(dst), MIPS_R_ZERO);
break;
/* Load a double word */
case BPF_DW:
if (dst[1] == src) {
emit(ctx, lw, dst[0], off + 4, src);
emit(ctx, lw, dst[1], off, src);
} else {
emit(ctx, lw, dst[1], off, src);
emit(ctx, lw, dst[0], off + 4, src);
}
emit_load_delay(ctx);
break;
}
clobber_reg64(ctx, dst);
}
/* Store operation: *(size *)(dst + off) = src */
static void emit_stx(struct jit_context *ctx,
const u8 dst, const u8 src[], s16 off, u8 size)
{
switch (size) {
/* Store a byte */
case BPF_B:
emit(ctx, sb, lo(src), off, dst);
break;
/* Store a half word */
case BPF_H:
emit(ctx, sh, lo(src), off, dst);
break;
/* Store a word */
case BPF_W:
emit(ctx, sw, lo(src), off, dst);
break;
/* Store a double word */
case BPF_DW:
emit(ctx, sw, src[1], off, dst);
emit(ctx, sw, src[0], off + 4, dst);
break;
}
}
/* Atomic read-modify-write (32-bit, non-ll/sc fallback) */
static void emit_atomic_r32(struct jit_context *ctx,
u8 dst, u8 src, s16 off, u8 code)
{
u32 exclude = 0;
u32 addr = 0;
/* Push caller-saved registers on stack */
push_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
0, JIT_RESERVED_STACK);
/*
* Argument 1: dst+off if xchg, otherwise src, passed in register a0
* Argument 2: src if xchg, otherwise dst+off, passed in register a1
mips, bpf: Add eBPF JIT for 32-bit MIPS This is an implementation of an eBPF JIT for 32-bit MIPS I-V and MIPS32. The implementation supports all 32-bit and 64-bit ALU and JMP operations, including the recently-added atomics. 64-bit div/mod and 64-bit atomics are implemented using function calls to math64 and atomic64 functions, respectively. All 32-bit operations are implemented natively by the JIT, except if the CPU lacks ll/sc instructions. Register mapping ================ All 64-bit eBPF registers are mapped to native 32-bit MIPS register pairs, and does not use any stack scratch space for register swapping. This means that all eBPF register data is kept in CPU registers all the time, and this simplifies the register management a lot. It also reduces the JIT's pressure on temporary registers since we do not have to move data around. Native register pairs are ordered according to CPU endiannes, following the O32 calling convention for passing 64-bit arguments and return values. The eBPF return value, arguments and callee-saved registers are mapped to their native MIPS equivalents. Since the 32 highest bits in the eBPF FP (frame pointer) register are always zero, only one general-purpose register is actually needed for the mapping. The MIPS fp register is used for this purpose. The high bits are mapped to MIPS register r0. This saves us one CPU register, which is much needed for temporaries, while still allowing us to treat the R10 (FP) register just like any other eBPF register in the JIT. The MIPS gp (global pointer) and at (assembler temporary) registers are used as internal temporary registers for constant blinding. CPU registers t6-t9 are used internally by the JIT when constructing more complex 64-bit operations. This is precisely what is needed - two registers to store an operand value, and two more as scratch registers when performing the operation. The register mapping is shown below. R0 - $v1, $v0 return value R1 - $a1, $a0 argument 1, passed in registers R2 - $a3, $a2 argument 2, passed in registers R3 - $t1, $t0 argument 3, passed on stack R4 - $t3, $t2 argument 4, passed on stack R5 - $t4, $t3 argument 5, passed on stack R6 - $s1, $s0 callee-saved R7 - $s3, $s2 callee-saved R8 - $s5, $s4 callee-saved R9 - $s7, $s6 callee-saved FP - $r0, $fp 32-bit frame pointer AX - $gp, $at constant-blinding $t6 - $t9 unallocated, JIT temporaries Jump offsets ============ The JIT tries to map all conditional JMP operations to MIPS conditional PC-relative branches. The MIPS branch offset field is 18 bits, in bytes, which is equivalent to the eBPF 16-bit instruction offset. However, since the JIT may emit more than one CPU instruction per eBPF instruction, the field width may overflow. If that happens, the JIT converts the long conditional jump to a short PC-relative branch with the condition inverted, jumping over a long unconditional absolute jmp (j). This conversion will change the instruction offset mapping used for jumps, and may in turn result in more branch offset overflows. The JIT therefore dry-runs the translation until no more branches are converted and the offsets do not change anymore. There is an upper bound on this of course, and if the JIT hits that limit, the last two iterations are run with all branches being converted. Tail call count =============== The current tail call count is stored in the 16-byte area of the caller's stack frame that is reserved for the callee in the o32 ABI. The value is initialized in the prologue, and propagated to the tail-callee by skipping the initialization instructions when emitting the tail call. Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20211005165408.2305108-4-johan.almbladh@anyfinetworks.com
2021-10-06 00:54:04 +08:00
*/
emit(ctx, move, MIPS_R_T9, dst);
if (code == BPF_XCHG) {
emit(ctx, move, MIPS_R_A1, src);
emit(ctx, addiu, MIPS_R_A0, MIPS_R_T9, off);
} else {
emit(ctx, move, MIPS_R_A0, src);
emit(ctx, addiu, MIPS_R_A1, MIPS_R_T9, off);
}
/* Emit function call */
switch (code) {
case BPF_ADD:
addr = (u32)&atomic_add;
break;
case BPF_ADD | BPF_FETCH:
addr = (u32)&atomic_fetch_add;
break;
case BPF_SUB:
addr = (u32)&atomic_sub;
break;
case BPF_SUB | BPF_FETCH:
addr = (u32)&atomic_fetch_sub;
break;
case BPF_OR:
addr = (u32)&atomic_or;
break;
case BPF_OR | BPF_FETCH:
addr = (u32)&atomic_fetch_or;
break;
case BPF_AND:
addr = (u32)&atomic_and;
break;
case BPF_AND | BPF_FETCH:
addr = (u32)&atomic_fetch_and;
break;
case BPF_XOR:
addr = (u32)&atomic_xor;
break;
case BPF_XOR | BPF_FETCH:
addr = (u32)&atomic_fetch_xor;
break;
case BPF_XCHG:
addr = (u32)&atomic_xchg;
break;
}
emit_mov_i(ctx, MIPS_R_T9, addr);
emit(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
emit(ctx, nop); /* Delay slot */
/* Update src register with old value, if specified */
if (code & BPF_FETCH) {
emit(ctx, move, src, MIPS_R_V0);
exclude = BIT(src);
clobber_reg(ctx, src);
}
/* Restore caller-saved registers, except any fetched value */
pop_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
exclude, JIT_RESERVED_STACK);
emit_load_delay(ctx);
clobber_reg(ctx, MIPS_R_RA);
}
/* Helper function for 64-bit atomic exchange */
static s64 jit_xchg64(s64 a, atomic64_t *v)
{
return atomic64_xchg(v, a);
}
/* Atomic read-modify-write (64-bit) */
static void emit_atomic_r64(struct jit_context *ctx,
u8 dst, const u8 src[], s16 off, u8 code)
{
const u8 *r0 = bpf2mips32[BPF_REG_0]; /* Mapped to v0-v1 */
const u8 *r1 = bpf2mips32[BPF_REG_1]; /* Mapped to a0-a1 */
u32 exclude = 0;
u32 addr = 0;
/* Push caller-saved registers on stack */
push_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
0, JIT_RESERVED_STACK);
/*
* Argument 1: 64-bit src, passed in registers a0-a1
* Argument 2: 32-bit dst+off, passed in register a2
*/
emit(ctx, move, MIPS_R_T9, dst);
emit(ctx, move, r1[0], src[0]);
emit(ctx, move, r1[1], src[1]);
emit(ctx, addiu, MIPS_R_A2, MIPS_R_T9, off);
/* Emit function call */
switch (code) {
case BPF_ADD:
addr = (u32)&atomic64_add;
break;
case BPF_ADD | BPF_FETCH:
addr = (u32)&atomic64_fetch_add;
break;
case BPF_SUB:
addr = (u32)&atomic64_sub;
break;
case BPF_SUB | BPF_FETCH:
addr = (u32)&atomic64_fetch_sub;
break;
case BPF_OR:
addr = (u32)&atomic64_or;
break;
case BPF_OR | BPF_FETCH:
addr = (u32)&atomic64_fetch_or;
break;
case BPF_AND:
addr = (u32)&atomic64_and;
break;
case BPF_AND | BPF_FETCH:
addr = (u32)&atomic64_fetch_and;
break;
case BPF_XOR:
addr = (u32)&atomic64_xor;
break;
case BPF_XOR | BPF_FETCH:
addr = (u32)&atomic64_fetch_xor;
break;
case BPF_XCHG:
addr = (u32)&jit_xchg64;
break;
}
emit_mov_i(ctx, MIPS_R_T9, addr);
emit(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
emit(ctx, nop); /* Delay slot */
/* Update src register with old value, if specified */
if (code & BPF_FETCH) {
emit(ctx, move, lo(src), lo(r0));
emit(ctx, move, hi(src), hi(r0));
exclude = BIT(src[0]) | BIT(src[1]);
clobber_reg64(ctx, src);
}
/* Restore caller-saved registers, except any fetched value */
pop_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
exclude, JIT_RESERVED_STACK);
emit_load_delay(ctx);
clobber_reg(ctx, MIPS_R_RA);
}
/* Atomic compare-and-exchange (32-bit, non-ll/sc fallback) */
static void emit_cmpxchg_r32(struct jit_context *ctx, u8 dst, u8 src, s16 off)
{
const u8 *r0 = bpf2mips32[BPF_REG_0];
/* Push caller-saved registers on stack */
push_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
JIT_RETURN_REGS, JIT_RESERVED_STACK + 2 * sizeof(u32));
/*
* Argument 1: 32-bit dst+off, passed in register a0
* Argument 2: 32-bit r0, passed in register a1
* Argument 3: 32-bit src, passed in register a2
*/
emit(ctx, addiu, MIPS_R_T9, dst, off);
emit(ctx, move, MIPS_R_T8, src);
emit(ctx, move, MIPS_R_A1, lo(r0));
emit(ctx, move, MIPS_R_A0, MIPS_R_T9);
emit(ctx, move, MIPS_R_A2, MIPS_R_T8);
/* Emit function call */
emit_mov_i(ctx, MIPS_R_T9, (u32)&atomic_cmpxchg);
emit(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
emit(ctx, nop); /* Delay slot */
#ifdef __BIG_ENDIAN
emit(ctx, move, lo(r0), MIPS_R_V0);
#endif
/* Restore caller-saved registers, except the return value */
pop_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
JIT_RETURN_REGS, JIT_RESERVED_STACK + 2 * sizeof(u32));
emit_load_delay(ctx);
clobber_reg(ctx, MIPS_R_V0);
clobber_reg(ctx, MIPS_R_V1);
clobber_reg(ctx, MIPS_R_RA);
}
/* Atomic compare-and-exchange (64-bit) */
static void emit_cmpxchg_r64(struct jit_context *ctx,
u8 dst, const u8 src[], s16 off)
{
const u8 *r0 = bpf2mips32[BPF_REG_0];
const u8 *r2 = bpf2mips32[BPF_REG_2];
/* Push caller-saved registers on stack */
push_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
JIT_RETURN_REGS, JIT_RESERVED_STACK + 2 * sizeof(u32));
/*
* Argument 1: 32-bit dst+off, passed in register a0 (a1 unused)
* Argument 2: 64-bit r0, passed in registers a2-a3
* Argument 3: 64-bit src, passed on stack
*/
push_regs(ctx, BIT(src[0]) | BIT(src[1]), 0, JIT_RESERVED_STACK);
emit(ctx, addiu, MIPS_R_T9, dst, off);
emit(ctx, move, r2[0], r0[0]);
emit(ctx, move, r2[1], r0[1]);
emit(ctx, move, MIPS_R_A0, MIPS_R_T9);
/* Emit function call */
emit_mov_i(ctx, MIPS_R_T9, (u32)&atomic64_cmpxchg);
emit(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
emit(ctx, nop); /* Delay slot */
/* Restore caller-saved registers, except the return value */
pop_regs(ctx, ctx->clobbered & JIT_CALLER_REGS,
JIT_RETURN_REGS, JIT_RESERVED_STACK + 2 * sizeof(u32));
emit_load_delay(ctx);
clobber_reg(ctx, MIPS_R_V0);
clobber_reg(ctx, MIPS_R_V1);
clobber_reg(ctx, MIPS_R_RA);
}
/*
* Conditional movz or an emulated equivalent.
* Note that the rs register may be modified.
*/
static void emit_movz_r(struct jit_context *ctx, u8 rd, u8 rs, u8 rt)
{
if (cpu_has_mips_2) {
emit(ctx, movz, rd, rs, rt); /* rd = rt ? rd : rs */
} else if (cpu_has_mips32r6) {
if (rs != MIPS_R_ZERO)
emit(ctx, seleqz, rs, rs, rt); /* rs = 0 if rt == 0 */
emit(ctx, selnez, rd, rd, rt); /* rd = 0 if rt != 0 */
if (rs != MIPS_R_ZERO)
emit(ctx, or, rd, rd, rs); /* rd = rd | rs */
} else {
emit(ctx, bnez, rt, 8); /* PC += 8 if rd != 0 */
emit(ctx, nop); /* +0: delay slot */
emit(ctx, or, rd, rs, MIPS_R_ZERO); /* +4: rd = rs */
}
clobber_reg(ctx, rd);
clobber_reg(ctx, rs);
}
/*
* Conditional movn or an emulated equivalent.
* Note that the rs register may be modified.
*/
static void emit_movn_r(struct jit_context *ctx, u8 rd, u8 rs, u8 rt)
{
if (cpu_has_mips_2) {
emit(ctx, movn, rd, rs, rt); /* rd = rt ? rs : rd */
} else if (cpu_has_mips32r6) {
if (rs != MIPS_R_ZERO)
emit(ctx, selnez, rs, rs, rt); /* rs = 0 if rt == 0 */
emit(ctx, seleqz, rd, rd, rt); /* rd = 0 if rt != 0 */
if (rs != MIPS_R_ZERO)
emit(ctx, or, rd, rd, rs); /* rd = rd | rs */
} else {
emit(ctx, beqz, rt, 8); /* PC += 8 if rd == 0 */
emit(ctx, nop); /* +0: delay slot */
emit(ctx, or, rd, rs, MIPS_R_ZERO); /* +4: rd = rs */
}
clobber_reg(ctx, rd);
clobber_reg(ctx, rs);
}
/* Emulation of 64-bit sltiu rd, rs, imm, where imm may be S32_MAX + 1 */
static void emit_sltiu_r64(struct jit_context *ctx, u8 rd,
const u8 rs[], s64 imm)
{
u8 tmp = MIPS_R_T9;
if (imm < 0) {
emit_mov_i(ctx, rd, imm); /* rd = imm */
emit(ctx, sltu, rd, lo(rs), rd); /* rd = rsl < rd */
emit(ctx, sltiu, tmp, hi(rs), -1); /* tmp = rsh < ~0U */
emit(ctx, or, rd, rd, tmp); /* rd = rd | tmp */
} else { /* imm >= 0 */
if (imm > 0x7fff) {
emit_mov_i(ctx, rd, (s32)imm); /* rd = imm */
emit(ctx, sltu, rd, lo(rs), rd); /* rd = rsl < rd */
} else {
emit(ctx, sltiu, rd, lo(rs), imm); /* rd = rsl < imm */
}
emit_movn_r(ctx, rd, MIPS_R_ZERO, hi(rs)); /* rd = 0 if rsh */
}
}
/* Emulation of 64-bit sltu rd, rs, rt */
static void emit_sltu_r64(struct jit_context *ctx, u8 rd,
const u8 rs[], const u8 rt[])
{
u8 tmp = MIPS_R_T9;
emit(ctx, sltu, rd, lo(rs), lo(rt)); /* rd = rsl < rtl */
emit(ctx, subu, tmp, hi(rs), hi(rt)); /* tmp = rsh - rth */
emit_movn_r(ctx, rd, MIPS_R_ZERO, tmp); /* rd = 0 if tmp != 0 */
emit(ctx, sltu, tmp, hi(rs), hi(rt)); /* tmp = rsh < rth */
emit(ctx, or, rd, rd, tmp); /* rd = rd | tmp */
}
/* Emulation of 64-bit slti rd, rs, imm, where imm may be S32_MAX + 1 */
static void emit_slti_r64(struct jit_context *ctx, u8 rd,
const u8 rs[], s64 imm)
{
u8 t1 = MIPS_R_T8;
u8 t2 = MIPS_R_T9;
u8 cmp;
/*
* if ((rs < 0) ^ (imm < 0)) t1 = imm >u rsl
* else t1 = rsl <u imm
*/
emit_mov_i(ctx, rd, (s32)imm);
emit(ctx, sltu, t1, lo(rs), rd); /* t1 = rsl <u imm */
emit(ctx, sltu, t2, rd, lo(rs)); /* t2 = imm <u rsl */
emit(ctx, srl, rd, hi(rs), 31); /* rd = rsh >> 31 */
if (imm < 0)
emit_movz_r(ctx, t1, t2, rd); /* t1 = rd ? t1 : t2 */
else
emit_movn_r(ctx, t1, t2, rd); /* t1 = rd ? t2 : t1 */
/*
* if ((imm < 0 && rsh != 0xffffffff) ||
* (imm >= 0 && rsh != 0))
* t1 = 0
*/
if (imm < 0) {
emit(ctx, addiu, rd, hi(rs), 1); /* rd = rsh + 1 */
cmp = rd;
} else { /* imm >= 0 */
cmp = hi(rs);
}
emit_movn_r(ctx, t1, MIPS_R_ZERO, cmp); /* t1 = 0 if cmp != 0 */
/*
* if (imm < 0) rd = rsh < -1
* else rd = rsh != 0
* rd = rd | t1
*/
emit(ctx, slti, rd, hi(rs), imm < 0 ? -1 : 0); /* rd = rsh < hi(imm) */
emit(ctx, or, rd, rd, t1); /* rd = rd | t1 */
}
/* Emulation of 64-bit(slt rd, rs, rt) */
static void emit_slt_r64(struct jit_context *ctx, u8 rd,
const u8 rs[], const u8 rt[])
{
u8 t1 = MIPS_R_T7;
u8 t2 = MIPS_R_T8;
u8 t3 = MIPS_R_T9;
/*
* if ((rs < 0) ^ (rt < 0)) t1 = rtl <u rsl
* else t1 = rsl <u rtl
* if (rsh == rth) t1 = 0
*/
emit(ctx, sltu, t1, lo(rs), lo(rt)); /* t1 = rsl <u rtl */
emit(ctx, sltu, t2, lo(rt), lo(rs)); /* t2 = rtl <u rsl */
emit(ctx, xor, t3, hi(rs), hi(rt)); /* t3 = rlh ^ rth */
emit(ctx, srl, rd, t3, 31); /* rd = t3 >> 31 */
emit_movn_r(ctx, t1, t2, rd); /* t1 = rd ? t2 : t1 */
emit_movn_r(ctx, t1, MIPS_R_ZERO, t3); /* t1 = 0 if t3 != 0 */
/* rd = (rsh < rth) | t1 */
emit(ctx, slt, rd, hi(rs), hi(rt)); /* rd = rsh <s rth */
emit(ctx, or, rd, rd, t1); /* rd = rd | t1 */
}
/* Jump immediate (64-bit) */
static void emit_jmp_i64(struct jit_context *ctx,
const u8 dst[], s32 imm, s32 off, u8 op)
{
u8 tmp = MIPS_R_T6;
switch (op) {
/* No-op, used internally for branch optimization */
case JIT_JNOP:
break;
/* PC += off if dst == imm */
/* PC += off if dst != imm */
case BPF_JEQ:
case BPF_JNE:
if (imm >= -0x7fff && imm <= 0x8000) {
emit(ctx, addiu, tmp, lo(dst), -imm);
} else if ((u32)imm <= 0xffff) {
emit(ctx, xori, tmp, lo(dst), imm);
} else { /* Register fallback */
emit_mov_i(ctx, tmp, imm);
emit(ctx, xor, tmp, lo(dst), tmp);
}
if (imm < 0) { /* Compare sign extension */
emit(ctx, addu, MIPS_R_T9, hi(dst), 1);
emit(ctx, or, tmp, tmp, MIPS_R_T9);
} else { /* Compare zero extension */
emit(ctx, or, tmp, tmp, hi(dst));
}
if (op == BPF_JEQ)
emit(ctx, beqz, tmp, off);
else /* BPF_JNE */
emit(ctx, bnez, tmp, off);
break;
/* PC += off if dst & imm */
/* PC += off if (dst & imm) == 0 (not in BPF, used for long jumps) */
case BPF_JSET:
case JIT_JNSET:
if ((u32)imm <= 0xffff) {
emit(ctx, andi, tmp, lo(dst), imm);
} else { /* Register fallback */
emit_mov_i(ctx, tmp, imm);
emit(ctx, and, tmp, lo(dst), tmp);
}
if (imm < 0) /* Sign-extension pulls in high word */
emit(ctx, or, tmp, tmp, hi(dst));
if (op == BPF_JSET)
emit(ctx, bnez, tmp, off);
else /* JIT_JNSET */
emit(ctx, beqz, tmp, off);
break;
/* PC += off if dst > imm */
case BPF_JGT:
emit_sltiu_r64(ctx, tmp, dst, (s64)imm + 1);
emit(ctx, beqz, tmp, off);
break;
/* PC += off if dst >= imm */
case BPF_JGE:
emit_sltiu_r64(ctx, tmp, dst, imm);
emit(ctx, beqz, tmp, off);
break;
/* PC += off if dst < imm */
case BPF_JLT:
emit_sltiu_r64(ctx, tmp, dst, imm);
emit(ctx, bnez, tmp, off);
break;
/* PC += off if dst <= imm */
case BPF_JLE:
emit_sltiu_r64(ctx, tmp, dst, (s64)imm + 1);
emit(ctx, bnez, tmp, off);
break;
/* PC += off if dst > imm (signed) */
case BPF_JSGT:
emit_slti_r64(ctx, tmp, dst, (s64)imm + 1);
emit(ctx, beqz, tmp, off);
break;
/* PC += off if dst >= imm (signed) */
case BPF_JSGE:
emit_slti_r64(ctx, tmp, dst, imm);
emit(ctx, beqz, tmp, off);
break;
/* PC += off if dst < imm (signed) */
case BPF_JSLT:
emit_slti_r64(ctx, tmp, dst, imm);
emit(ctx, bnez, tmp, off);
break;
/* PC += off if dst <= imm (signed) */
case BPF_JSLE:
emit_slti_r64(ctx, tmp, dst, (s64)imm + 1);
emit(ctx, bnez, tmp, off);
break;
}
}
/* Jump register (64-bit) */
static void emit_jmp_r64(struct jit_context *ctx,
const u8 dst[], const u8 src[], s32 off, u8 op)
{
u8 t1 = MIPS_R_T6;
u8 t2 = MIPS_R_T7;
switch (op) {
/* No-op, used internally for branch optimization */
case JIT_JNOP:
break;
/* PC += off if dst == src */
/* PC += off if dst != src */
case BPF_JEQ:
case BPF_JNE:
emit(ctx, subu, t1, lo(dst), lo(src));
emit(ctx, subu, t2, hi(dst), hi(src));
emit(ctx, or, t1, t1, t2);
if (op == BPF_JEQ)
emit(ctx, beqz, t1, off);
else /* BPF_JNE */
emit(ctx, bnez, t1, off);
break;
/* PC += off if dst & src */
/* PC += off if (dst & imm) == 0 (not in BPF, used for long jumps) */
case BPF_JSET:
case JIT_JNSET:
emit(ctx, and, t1, lo(dst), lo(src));
emit(ctx, and, t2, hi(dst), hi(src));
emit(ctx, or, t1, t1, t2);
if (op == BPF_JSET)
emit(ctx, bnez, t1, off);
else /* JIT_JNSET */
emit(ctx, beqz, t1, off);
break;
/* PC += off if dst > src */
case BPF_JGT:
emit_sltu_r64(ctx, t1, src, dst);
emit(ctx, bnez, t1, off);
break;
/* PC += off if dst >= src */
case BPF_JGE:
emit_sltu_r64(ctx, t1, dst, src);
emit(ctx, beqz, t1, off);
break;
/* PC += off if dst < src */
case BPF_JLT:
emit_sltu_r64(ctx, t1, dst, src);
emit(ctx, bnez, t1, off);
break;
/* PC += off if dst <= src */
case BPF_JLE:
emit_sltu_r64(ctx, t1, src, dst);
emit(ctx, beqz, t1, off);
break;
/* PC += off if dst > src (signed) */
case BPF_JSGT:
emit_slt_r64(ctx, t1, src, dst);
emit(ctx, bnez, t1, off);
break;
/* PC += off if dst >= src (signed) */
case BPF_JSGE:
emit_slt_r64(ctx, t1, dst, src);
emit(ctx, beqz, t1, off);
break;
/* PC += off if dst < src (signed) */
case BPF_JSLT:
emit_slt_r64(ctx, t1, dst, src);
emit(ctx, bnez, t1, off);
break;
/* PC += off if dst <= src (signed) */
case BPF_JSLE:
emit_slt_r64(ctx, t1, src, dst);
emit(ctx, beqz, t1, off);
break;
}
}
/* Function call */
static int emit_call(struct jit_context *ctx, const struct bpf_insn *insn)
{
bool fixed;
u64 addr;
/* Decode the call address */
if (bpf_jit_get_func_addr(ctx->program, insn, false,
&addr, &fixed) < 0)
return -1;
if (!fixed)
return -1;
/* Push stack arguments */
push_regs(ctx, JIT_STACK_REGS, 0, JIT_RESERVED_STACK);
/* Emit function call */
emit_mov_i(ctx, MIPS_R_T9, addr);
emit(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
emit(ctx, nop); /* Delay slot */
clobber_reg(ctx, MIPS_R_RA);
clobber_reg(ctx, MIPS_R_V0);
clobber_reg(ctx, MIPS_R_V1);
return 0;
}
/* Function tail call */
static int emit_tail_call(struct jit_context *ctx)
{
u8 ary = lo(bpf2mips32[BPF_REG_2]);
u8 ind = lo(bpf2mips32[BPF_REG_3]);
u8 t1 = MIPS_R_T8;
u8 t2 = MIPS_R_T9;
int off;
/*
* Tail call:
* eBPF R1 - function argument (context ptr), passed in a0-a1
* eBPF R2 - ptr to object with array of function entry points
* eBPF R3 - array index of function to be called
* stack[sz] - remaining tail call count, initialized in prologue
*/
/* if (ind >= ary->map.max_entries) goto out */
off = offsetof(struct bpf_array, map.max_entries);
if (off > 0x7fff)
return -1;
emit(ctx, lw, t1, off, ary); /* t1 = ary->map.max_entries*/
emit_load_delay(ctx); /* Load delay slot */
emit(ctx, sltu, t1, ind, t1); /* t1 = ind < t1 */
emit(ctx, beqz, t1, get_offset(ctx, 1)); /* PC += off(1) if t1 == 0 */
/* (next insn delay slot) */
/* if (TCC-- <= 0) goto out */
emit(ctx, lw, t2, ctx->stack_size, MIPS_R_SP); /* t2 = *(SP + size) */
emit_load_delay(ctx); /* Load delay slot */
emit(ctx, blez, t2, get_offset(ctx, 1)); /* PC += off(1) if t2 <= 0 */
mips, bpf: Add eBPF JIT for 32-bit MIPS This is an implementation of an eBPF JIT for 32-bit MIPS I-V and MIPS32. The implementation supports all 32-bit and 64-bit ALU and JMP operations, including the recently-added atomics. 64-bit div/mod and 64-bit atomics are implemented using function calls to math64 and atomic64 functions, respectively. All 32-bit operations are implemented natively by the JIT, except if the CPU lacks ll/sc instructions. Register mapping ================ All 64-bit eBPF registers are mapped to native 32-bit MIPS register pairs, and does not use any stack scratch space for register swapping. This means that all eBPF register data is kept in CPU registers all the time, and this simplifies the register management a lot. It also reduces the JIT's pressure on temporary registers since we do not have to move data around. Native register pairs are ordered according to CPU endiannes, following the O32 calling convention for passing 64-bit arguments and return values. The eBPF return value, arguments and callee-saved registers are mapped to their native MIPS equivalents. Since the 32 highest bits in the eBPF FP (frame pointer) register are always zero, only one general-purpose register is actually needed for the mapping. The MIPS fp register is used for this purpose. The high bits are mapped to MIPS register r0. This saves us one CPU register, which is much needed for temporaries, while still allowing us to treat the R10 (FP) register just like any other eBPF register in the JIT. The MIPS gp (global pointer) and at (assembler temporary) registers are used as internal temporary registers for constant blinding. CPU registers t6-t9 are used internally by the JIT when constructing more complex 64-bit operations. This is precisely what is needed - two registers to store an operand value, and two more as scratch registers when performing the operation. The register mapping is shown below. R0 - $v1, $v0 return value R1 - $a1, $a0 argument 1, passed in registers R2 - $a3, $a2 argument 2, passed in registers R3 - $t1, $t0 argument 3, passed on stack R4 - $t3, $t2 argument 4, passed on stack R5 - $t4, $t3 argument 5, passed on stack R6 - $s1, $s0 callee-saved R7 - $s3, $s2 callee-saved R8 - $s5, $s4 callee-saved R9 - $s7, $s6 callee-saved FP - $r0, $fp 32-bit frame pointer AX - $gp, $at constant-blinding $t6 - $t9 unallocated, JIT temporaries Jump offsets ============ The JIT tries to map all conditional JMP operations to MIPS conditional PC-relative branches. The MIPS branch offset field is 18 bits, in bytes, which is equivalent to the eBPF 16-bit instruction offset. However, since the JIT may emit more than one CPU instruction per eBPF instruction, the field width may overflow. If that happens, the JIT converts the long conditional jump to a short PC-relative branch with the condition inverted, jumping over a long unconditional absolute jmp (j). This conversion will change the instruction offset mapping used for jumps, and may in turn result in more branch offset overflows. The JIT therefore dry-runs the translation until no more branches are converted and the offsets do not change anymore. There is an upper bound on this of course, and if the JIT hits that limit, the last two iterations are run with all branches being converted. Tail call count =============== The current tail call count is stored in the 16-byte area of the caller's stack frame that is reserved for the callee in the o32 ABI. The value is initialized in the prologue, and propagated to the tail-callee by skipping the initialization instructions when emitting the tail call. Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20211005165408.2305108-4-johan.almbladh@anyfinetworks.com
2021-10-06 00:54:04 +08:00
emit(ctx, addiu, t2, t2, -1); /* t2-- (delay slot) */
emit(ctx, sw, t2, ctx->stack_size, MIPS_R_SP); /* *(SP + size) = t2 */
/* prog = ary->ptrs[ind] */
off = offsetof(struct bpf_array, ptrs);
if (off > 0x7fff)
return -1;
emit(ctx, sll, t1, ind, 2); /* t1 = ind << 2 */
emit(ctx, addu, t1, t1, ary); /* t1 += ary */
emit(ctx, lw, t2, off, t1); /* t2 = *(t1 + off) */
emit_load_delay(ctx); /* Load delay slot */
/* if (prog == 0) goto out */
emit(ctx, beqz, t2, get_offset(ctx, 1)); /* PC += off(1) if t2 == 0 */
emit(ctx, nop); /* Delay slot */
/* func = prog->bpf_func + 8 (prologue skip offset) */
off = offsetof(struct bpf_prog, bpf_func);
if (off > 0x7fff)
return -1;
emit(ctx, lw, t1, off, t2); /* t1 = *(t2 + off) */
emit_load_delay(ctx); /* Load delay slot */
emit(ctx, addiu, t1, t1, JIT_TCALL_SKIP); /* t1 += skip (8 or 12) */
/* goto func */
build_epilogue(ctx, t1);
return 0;
}
/*
* Stack frame layout for a JITed program (stack grows down).
*
* Higher address : Caller's stack frame :
* :----------------------------:
* : 64-bit eBPF args r3-r5 :
* :----------------------------:
* : Reserved / tail call count :
* +============================+ <--- MIPS sp before call
* | Callee-saved registers, |
* | including RA and FP |
* +----------------------------+ <--- eBPF FP (MIPS zero,fp)
* | Local eBPF variables |
* | allocated by program |
* +----------------------------+
* | Reserved for caller-saved |
* | registers |
* +----------------------------+
* | Reserved for 64-bit eBPF |
* | args r3-r5 & args passed |
* | on stack in kernel calls |
* Lower address +============================+ <--- MIPS sp
*/
/* Build program prologue to set up the stack and registers */
void build_prologue(struct jit_context *ctx)
{
const u8 *r1 = bpf2mips32[BPF_REG_1];
const u8 *fp = bpf2mips32[BPF_REG_FP];
int stack, saved, locals, reserved;
/*
* In the unlikely event that the TCC limit is raised to more
* than 16 bits, it is clamped to the maximum value allowed for
* the generated code (0xffff). It is better fail to compile
* instead of degrading gracefully.
*/
BUILD_BUG_ON(MAX_TAIL_CALL_CNT > 0xffff);
mips, bpf: Add eBPF JIT for 32-bit MIPS This is an implementation of an eBPF JIT for 32-bit MIPS I-V and MIPS32. The implementation supports all 32-bit and 64-bit ALU and JMP operations, including the recently-added atomics. 64-bit div/mod and 64-bit atomics are implemented using function calls to math64 and atomic64 functions, respectively. All 32-bit operations are implemented natively by the JIT, except if the CPU lacks ll/sc instructions. Register mapping ================ All 64-bit eBPF registers are mapped to native 32-bit MIPS register pairs, and does not use any stack scratch space for register swapping. This means that all eBPF register data is kept in CPU registers all the time, and this simplifies the register management a lot. It also reduces the JIT's pressure on temporary registers since we do not have to move data around. Native register pairs are ordered according to CPU endiannes, following the O32 calling convention for passing 64-bit arguments and return values. The eBPF return value, arguments and callee-saved registers are mapped to their native MIPS equivalents. Since the 32 highest bits in the eBPF FP (frame pointer) register are always zero, only one general-purpose register is actually needed for the mapping. The MIPS fp register is used for this purpose. The high bits are mapped to MIPS register r0. This saves us one CPU register, which is much needed for temporaries, while still allowing us to treat the R10 (FP) register just like any other eBPF register in the JIT. The MIPS gp (global pointer) and at (assembler temporary) registers are used as internal temporary registers for constant blinding. CPU registers t6-t9 are used internally by the JIT when constructing more complex 64-bit operations. This is precisely what is needed - two registers to store an operand value, and two more as scratch registers when performing the operation. The register mapping is shown below. R0 - $v1, $v0 return value R1 - $a1, $a0 argument 1, passed in registers R2 - $a3, $a2 argument 2, passed in registers R3 - $t1, $t0 argument 3, passed on stack R4 - $t3, $t2 argument 4, passed on stack R5 - $t4, $t3 argument 5, passed on stack R6 - $s1, $s0 callee-saved R7 - $s3, $s2 callee-saved R8 - $s5, $s4 callee-saved R9 - $s7, $s6 callee-saved FP - $r0, $fp 32-bit frame pointer AX - $gp, $at constant-blinding $t6 - $t9 unallocated, JIT temporaries Jump offsets ============ The JIT tries to map all conditional JMP operations to MIPS conditional PC-relative branches. The MIPS branch offset field is 18 bits, in bytes, which is equivalent to the eBPF 16-bit instruction offset. However, since the JIT may emit more than one CPU instruction per eBPF instruction, the field width may overflow. If that happens, the JIT converts the long conditional jump to a short PC-relative branch with the condition inverted, jumping over a long unconditional absolute jmp (j). This conversion will change the instruction offset mapping used for jumps, and may in turn result in more branch offset overflows. The JIT therefore dry-runs the translation until no more branches are converted and the offsets do not change anymore. There is an upper bound on this of course, and if the JIT hits that limit, the last two iterations are run with all branches being converted. Tail call count =============== The current tail call count is stored in the 16-byte area of the caller's stack frame that is reserved for the callee in the o32 ABI. The value is initialized in the prologue, and propagated to the tail-callee by skipping the initialization instructions when emitting the tail call. Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20211005165408.2305108-4-johan.almbladh@anyfinetworks.com
2021-10-06 00:54:04 +08:00
/*
* The first two instructions initialize TCC in the reserved (for us)
* 16-byte area in the parent's stack frame. On a tail call, the
* calling function jumps into the prologue after these instructions.
*/
emit(ctx, ori, MIPS_R_T9, MIPS_R_ZERO, MAX_TAIL_CALL_CNT);
mips, bpf: Add eBPF JIT for 32-bit MIPS This is an implementation of an eBPF JIT for 32-bit MIPS I-V and MIPS32. The implementation supports all 32-bit and 64-bit ALU and JMP operations, including the recently-added atomics. 64-bit div/mod and 64-bit atomics are implemented using function calls to math64 and atomic64 functions, respectively. All 32-bit operations are implemented natively by the JIT, except if the CPU lacks ll/sc instructions. Register mapping ================ All 64-bit eBPF registers are mapped to native 32-bit MIPS register pairs, and does not use any stack scratch space for register swapping. This means that all eBPF register data is kept in CPU registers all the time, and this simplifies the register management a lot. It also reduces the JIT's pressure on temporary registers since we do not have to move data around. Native register pairs are ordered according to CPU endiannes, following the O32 calling convention for passing 64-bit arguments and return values. The eBPF return value, arguments and callee-saved registers are mapped to their native MIPS equivalents. Since the 32 highest bits in the eBPF FP (frame pointer) register are always zero, only one general-purpose register is actually needed for the mapping. The MIPS fp register is used for this purpose. The high bits are mapped to MIPS register r0. This saves us one CPU register, which is much needed for temporaries, while still allowing us to treat the R10 (FP) register just like any other eBPF register in the JIT. The MIPS gp (global pointer) and at (assembler temporary) registers are used as internal temporary registers for constant blinding. CPU registers t6-t9 are used internally by the JIT when constructing more complex 64-bit operations. This is precisely what is needed - two registers to store an operand value, and two more as scratch registers when performing the operation. The register mapping is shown below. R0 - $v1, $v0 return value R1 - $a1, $a0 argument 1, passed in registers R2 - $a3, $a2 argument 2, passed in registers R3 - $t1, $t0 argument 3, passed on stack R4 - $t3, $t2 argument 4, passed on stack R5 - $t4, $t3 argument 5, passed on stack R6 - $s1, $s0 callee-saved R7 - $s3, $s2 callee-saved R8 - $s5, $s4 callee-saved R9 - $s7, $s6 callee-saved FP - $r0, $fp 32-bit frame pointer AX - $gp, $at constant-blinding $t6 - $t9 unallocated, JIT temporaries Jump offsets ============ The JIT tries to map all conditional JMP operations to MIPS conditional PC-relative branches. The MIPS branch offset field is 18 bits, in bytes, which is equivalent to the eBPF 16-bit instruction offset. However, since the JIT may emit more than one CPU instruction per eBPF instruction, the field width may overflow. If that happens, the JIT converts the long conditional jump to a short PC-relative branch with the condition inverted, jumping over a long unconditional absolute jmp (j). This conversion will change the instruction offset mapping used for jumps, and may in turn result in more branch offset overflows. The JIT therefore dry-runs the translation until no more branches are converted and the offsets do not change anymore. There is an upper bound on this of course, and if the JIT hits that limit, the last two iterations are run with all branches being converted. Tail call count =============== The current tail call count is stored in the 16-byte area of the caller's stack frame that is reserved for the callee in the o32 ABI. The value is initialized in the prologue, and propagated to the tail-callee by skipping the initialization instructions when emitting the tail call. Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Link: https://lore.kernel.org/bpf/20211005165408.2305108-4-johan.almbladh@anyfinetworks.com
2021-10-06 00:54:04 +08:00
emit(ctx, sw, MIPS_R_T9, 0, MIPS_R_SP);
/*
* Register eBPF R1 contains the 32-bit context pointer argument.
* A 32-bit argument is always passed in MIPS register a0, regardless
* of CPU endianness. Initialize R1 accordingly and zero-extend.
*/
#ifdef __BIG_ENDIAN
emit(ctx, move, lo(r1), MIPS_R_A0);
#endif
/* === Entry-point for tail calls === */
/* Zero-extend the 32-bit argument */
emit(ctx, move, hi(r1), MIPS_R_ZERO);
/* If the eBPF frame pointer was accessed it must be saved */
if (ctx->accessed & BIT(BPF_REG_FP))
clobber_reg64(ctx, fp);
/* Compute the stack space needed for callee-saved registers */
saved = hweight32(ctx->clobbered & JIT_CALLEE_REGS) * sizeof(u32);
saved = ALIGN(saved, MIPS_STACK_ALIGNMENT);
/* Stack space used by eBPF program local data */
locals = ALIGN(ctx->program->aux->stack_depth, MIPS_STACK_ALIGNMENT);
/*
* If we are emitting function calls, reserve extra stack space for
* caller-saved registers and function arguments passed on the stack.
* The required space is computed automatically during resource
* usage discovery (pass 1).
*/
reserved = ctx->stack_used;
/* Allocate the stack frame */
stack = ALIGN(saved + locals + reserved, MIPS_STACK_ALIGNMENT);
emit(ctx, addiu, MIPS_R_SP, MIPS_R_SP, -stack);
/* Store callee-saved registers on stack */
push_regs(ctx, ctx->clobbered & JIT_CALLEE_REGS, 0, stack - saved);
/* Initialize the eBPF frame pointer if accessed */
if (ctx->accessed & BIT(BPF_REG_FP))
emit(ctx, addiu, lo(fp), MIPS_R_SP, stack - saved);
ctx->saved_size = saved;
ctx->stack_size = stack;
}
/* Build the program epilogue to restore the stack and registers */
void build_epilogue(struct jit_context *ctx, int dest_reg)
{
/* Restore callee-saved registers from stack */
pop_regs(ctx, ctx->clobbered & JIT_CALLEE_REGS, 0,
ctx->stack_size - ctx->saved_size);
/*
* A 32-bit return value is always passed in MIPS register v0,
* but on big-endian targets the low part of R0 is mapped to v1.
*/
#ifdef __BIG_ENDIAN
emit(ctx, move, MIPS_R_V0, MIPS_R_V1);
#endif
/* Jump to the return address and adjust the stack pointer */
emit(ctx, jr, dest_reg);
emit(ctx, addiu, MIPS_R_SP, MIPS_R_SP, ctx->stack_size);
}
/* Build one eBPF instruction */
int build_insn(const struct bpf_insn *insn, struct jit_context *ctx)
{
const u8 *dst = bpf2mips32[insn->dst_reg];
const u8 *src = bpf2mips32[insn->src_reg];
const u8 *res = bpf2mips32[BPF_REG_0];
const u8 *tmp = bpf2mips32[JIT_REG_TMP];
u8 code = insn->code;
s16 off = insn->off;
s32 imm = insn->imm;
s32 val, rel;
u8 alu, jmp;
switch (code) {
/* ALU operations */
/* dst = imm */
case BPF_ALU | BPF_MOV | BPF_K:
emit_mov_i(ctx, lo(dst), imm);
emit_zext_ver(ctx, dst);
break;
/* dst = src */
case BPF_ALU | BPF_MOV | BPF_X:
if (imm == 1) {
/* Special mov32 for zext */
emit_mov_i(ctx, hi(dst), 0);
} else {
emit_mov_r(ctx, lo(dst), lo(src));
emit_zext_ver(ctx, dst);
}
break;
/* dst = -dst */
case BPF_ALU | BPF_NEG:
emit_alu_i(ctx, lo(dst), 0, BPF_NEG);
emit_zext_ver(ctx, dst);
break;
/* dst = dst & imm */
/* dst = dst | imm */
/* dst = dst ^ imm */
/* dst = dst << imm */
/* dst = dst >> imm */
/* dst = dst >> imm (arithmetic) */
/* dst = dst + imm */
/* dst = dst - imm */
/* dst = dst * imm */
/* dst = dst / imm */
/* dst = dst % imm */
case BPF_ALU | BPF_OR | BPF_K:
case BPF_ALU | BPF_AND | BPF_K:
case BPF_ALU | BPF_XOR | BPF_K:
case BPF_ALU | BPF_LSH | BPF_K:
case BPF_ALU | BPF_RSH | BPF_K:
case BPF_ALU | BPF_ARSH | BPF_K:
case BPF_ALU | BPF_ADD | BPF_K:
case BPF_ALU | BPF_SUB | BPF_K:
case BPF_ALU | BPF_MUL | BPF_K:
case BPF_ALU | BPF_DIV | BPF_K:
case BPF_ALU | BPF_MOD | BPF_K:
if (!valid_alu_i(BPF_OP(code), imm)) {
emit_mov_i(ctx, MIPS_R_T6, imm);
emit_alu_r(ctx, lo(dst), MIPS_R_T6, BPF_OP(code));
} else if (rewrite_alu_i(BPF_OP(code), imm, &alu, &val)) {
emit_alu_i(ctx, lo(dst), val, alu);
}
emit_zext_ver(ctx, dst);
break;
/* dst = dst & src */
/* dst = dst | src */
/* dst = dst ^ src */
/* dst = dst << src */
/* dst = dst >> src */
/* dst = dst >> src (arithmetic) */
/* dst = dst + src */
/* dst = dst - src */
/* dst = dst * src */
/* dst = dst / src */
/* dst = dst % src */
case BPF_ALU | BPF_AND | BPF_X:
case BPF_ALU | BPF_OR | BPF_X:
case BPF_ALU | BPF_XOR | BPF_X:
case BPF_ALU | BPF_LSH | BPF_X:
case BPF_ALU | BPF_RSH | BPF_X:
case BPF_ALU | BPF_ARSH | BPF_X:
case BPF_ALU | BPF_ADD | BPF_X:
case BPF_ALU | BPF_SUB | BPF_X:
case BPF_ALU | BPF_MUL | BPF_X:
case BPF_ALU | BPF_DIV | BPF_X:
case BPF_ALU | BPF_MOD | BPF_X:
emit_alu_r(ctx, lo(dst), lo(src), BPF_OP(code));
emit_zext_ver(ctx, dst);
break;
/* dst = imm (64-bit) */
case BPF_ALU64 | BPF_MOV | BPF_K:
emit_mov_se_i64(ctx, dst, imm);
break;
/* dst = src (64-bit) */
case BPF_ALU64 | BPF_MOV | BPF_X:
emit_mov_r(ctx, lo(dst), lo(src));
emit_mov_r(ctx, hi(dst), hi(src));
break;
/* dst = -dst (64-bit) */
case BPF_ALU64 | BPF_NEG:
emit_neg_i64(ctx, dst);
break;
/* dst = dst & imm (64-bit) */
case BPF_ALU64 | BPF_AND | BPF_K:
emit_alu_i64(ctx, dst, imm, BPF_OP(code));
break;
/* dst = dst | imm (64-bit) */
/* dst = dst ^ imm (64-bit) */
/* dst = dst + imm (64-bit) */
/* dst = dst - imm (64-bit) */
case BPF_ALU64 | BPF_OR | BPF_K:
case BPF_ALU64 | BPF_XOR | BPF_K:
case BPF_ALU64 | BPF_ADD | BPF_K:
case BPF_ALU64 | BPF_SUB | BPF_K:
if (imm)
emit_alu_i64(ctx, dst, imm, BPF_OP(code));
break;
/* dst = dst << imm (64-bit) */
/* dst = dst >> imm (64-bit) */
/* dst = dst >> imm (64-bit, arithmetic) */
case BPF_ALU64 | BPF_LSH | BPF_K:
case BPF_ALU64 | BPF_RSH | BPF_K:
case BPF_ALU64 | BPF_ARSH | BPF_K:
if (imm)
emit_shift_i64(ctx, dst, imm, BPF_OP(code));
break;
/* dst = dst * imm (64-bit) */
case BPF_ALU64 | BPF_MUL | BPF_K:
emit_mul_i64(ctx, dst, imm);
break;
/* dst = dst / imm (64-bit) */
/* dst = dst % imm (64-bit) */
case BPF_ALU64 | BPF_DIV | BPF_K:
case BPF_ALU64 | BPF_MOD | BPF_K:
/*
* Sign-extend the immediate value into a temporary register,
* and then do the operation on this register.
*/
emit_mov_se_i64(ctx, tmp, imm);
emit_divmod_r64(ctx, dst, tmp, BPF_OP(code));
break;
/* dst = dst & src (64-bit) */
/* dst = dst | src (64-bit) */
/* dst = dst ^ src (64-bit) */
/* dst = dst + src (64-bit) */
/* dst = dst - src (64-bit) */
case BPF_ALU64 | BPF_AND | BPF_X:
case BPF_ALU64 | BPF_OR | BPF_X:
case BPF_ALU64 | BPF_XOR | BPF_X:
case BPF_ALU64 | BPF_ADD | BPF_X:
case BPF_ALU64 | BPF_SUB | BPF_X:
emit_alu_r64(ctx, dst, src, BPF_OP(code));
break;
/* dst = dst << src (64-bit) */
/* dst = dst >> src (64-bit) */
/* dst = dst >> src (64-bit, arithmetic) */
case BPF_ALU64 | BPF_LSH | BPF_X:
case BPF_ALU64 | BPF_RSH | BPF_X:
case BPF_ALU64 | BPF_ARSH | BPF_X:
emit_shift_r64(ctx, dst, lo(src), BPF_OP(code));
break;
/* dst = dst * src (64-bit) */
case BPF_ALU64 | BPF_MUL | BPF_X:
emit_mul_r64(ctx, dst, src);
break;
/* dst = dst / src (64-bit) */
/* dst = dst % src (64-bit) */
case BPF_ALU64 | BPF_DIV | BPF_X:
case BPF_ALU64 | BPF_MOD | BPF_X:
emit_divmod_r64(ctx, dst, src, BPF_OP(code));
break;
/* dst = htole(dst) */
/* dst = htobe(dst) */
case BPF_ALU | BPF_END | BPF_FROM_LE:
case BPF_ALU | BPF_END | BPF_FROM_BE:
if (BPF_SRC(code) ==
#ifdef __BIG_ENDIAN
BPF_FROM_LE
#else
BPF_FROM_BE
#endif
)
emit_bswap_r64(ctx, dst, imm);
else
emit_trunc_r64(ctx, dst, imm);
break;
/* dst = imm64 */
case BPF_LD | BPF_IMM | BPF_DW:
emit_mov_i(ctx, lo(dst), imm);
emit_mov_i(ctx, hi(dst), insn[1].imm);
return 1;
/* LDX: dst = *(size *)(src + off) */
case BPF_LDX | BPF_MEM | BPF_W:
case BPF_LDX | BPF_MEM | BPF_H:
case BPF_LDX | BPF_MEM | BPF_B:
case BPF_LDX | BPF_MEM | BPF_DW:
emit_ldx(ctx, dst, lo(src), off, BPF_SIZE(code));
break;
/* ST: *(size *)(dst + off) = imm */
case BPF_ST | BPF_MEM | BPF_W:
case BPF_ST | BPF_MEM | BPF_H:
case BPF_ST | BPF_MEM | BPF_B:
case BPF_ST | BPF_MEM | BPF_DW:
switch (BPF_SIZE(code)) {
case BPF_DW:
/* Sign-extend immediate value into temporary reg */
emit_mov_se_i64(ctx, tmp, imm);
break;
case BPF_W:
case BPF_H:
case BPF_B:
emit_mov_i(ctx, lo(tmp), imm);
break;
}
emit_stx(ctx, lo(dst), tmp, off, BPF_SIZE(code));
break;
/* STX: *(size *)(dst + off) = src */
case BPF_STX | BPF_MEM | BPF_W:
case BPF_STX | BPF_MEM | BPF_H:
case BPF_STX | BPF_MEM | BPF_B:
case BPF_STX | BPF_MEM | BPF_DW:
emit_stx(ctx, lo(dst), src, off, BPF_SIZE(code));
break;
/* Speculation barrier */
case BPF_ST | BPF_NOSPEC:
break;
/* Atomics */
case BPF_STX | BPF_ATOMIC | BPF_W:
switch (imm) {
case BPF_ADD:
case BPF_ADD | BPF_FETCH:
case BPF_AND:
case BPF_AND | BPF_FETCH:
case BPF_OR:
case BPF_OR | BPF_FETCH:
case BPF_XOR:
case BPF_XOR | BPF_FETCH:
case BPF_XCHG:
if (cpu_has_llsc)
emit_atomic_r(ctx, lo(dst), lo(src), off, imm);
else /* Non-ll/sc fallback */
emit_atomic_r32(ctx, lo(dst), lo(src),
off, imm);
if (imm & BPF_FETCH)
emit_zext_ver(ctx, src);
break;
case BPF_CMPXCHG:
if (cpu_has_llsc)
emit_cmpxchg_r(ctx, lo(dst), lo(src),
lo(res), off);
else /* Non-ll/sc fallback */
emit_cmpxchg_r32(ctx, lo(dst), lo(src), off);
/* Result zero-extension inserted by verifier */
break;
default:
goto notyet;
}
break;
/* Atomics (64-bit) */
case BPF_STX | BPF_ATOMIC | BPF_DW:
switch (imm) {
case BPF_ADD:
case BPF_ADD | BPF_FETCH:
case BPF_AND:
case BPF_AND | BPF_FETCH:
case BPF_OR:
case BPF_OR | BPF_FETCH:
case BPF_XOR:
case BPF_XOR | BPF_FETCH:
case BPF_XCHG:
emit_atomic_r64(ctx, lo(dst), src, off, imm);
break;
case BPF_CMPXCHG:
emit_cmpxchg_r64(ctx, lo(dst), src, off);
break;
default:
goto notyet;
}
break;
/* PC += off if dst == src */
/* PC += off if dst != src */
/* PC += off if dst & src */
/* PC += off if dst > src */
/* PC += off if dst >= src */
/* PC += off if dst < src */
/* PC += off if dst <= src */
/* PC += off if dst > src (signed) */
/* PC += off if dst >= src (signed) */
/* PC += off if dst < src (signed) */
/* PC += off if dst <= src (signed) */
case BPF_JMP32 | BPF_JEQ | BPF_X:
case BPF_JMP32 | BPF_JNE | BPF_X:
case BPF_JMP32 | BPF_JSET | BPF_X:
case BPF_JMP32 | BPF_JGT | BPF_X:
case BPF_JMP32 | BPF_JGE | BPF_X:
case BPF_JMP32 | BPF_JLT | BPF_X:
case BPF_JMP32 | BPF_JLE | BPF_X:
case BPF_JMP32 | BPF_JSGT | BPF_X:
case BPF_JMP32 | BPF_JSGE | BPF_X:
case BPF_JMP32 | BPF_JSLT | BPF_X:
case BPF_JMP32 | BPF_JSLE | BPF_X:
if (off == 0)
break;
setup_jmp_r(ctx, dst == src, BPF_OP(code), off, &jmp, &rel);
emit_jmp_r(ctx, lo(dst), lo(src), rel, jmp);
if (finish_jmp(ctx, jmp, off) < 0)
goto toofar;
break;
/* PC += off if dst == imm */
/* PC += off if dst != imm */
/* PC += off if dst & imm */
/* PC += off if dst > imm */
/* PC += off if dst >= imm */
/* PC += off if dst < imm */
/* PC += off if dst <= imm */
/* PC += off if dst > imm (signed) */
/* PC += off if dst >= imm (signed) */
/* PC += off if dst < imm (signed) */
/* PC += off if dst <= imm (signed) */
case BPF_JMP32 | BPF_JEQ | BPF_K:
case BPF_JMP32 | BPF_JNE | BPF_K:
case BPF_JMP32 | BPF_JSET | BPF_K:
case BPF_JMP32 | BPF_JGT | BPF_K:
case BPF_JMP32 | BPF_JGE | BPF_K:
case BPF_JMP32 | BPF_JLT | BPF_K:
case BPF_JMP32 | BPF_JLE | BPF_K:
case BPF_JMP32 | BPF_JSGT | BPF_K:
case BPF_JMP32 | BPF_JSGE | BPF_K:
case BPF_JMP32 | BPF_JSLT | BPF_K:
case BPF_JMP32 | BPF_JSLE | BPF_K:
if (off == 0)
break;
setup_jmp_i(ctx, imm, 32, BPF_OP(code), off, &jmp, &rel);
if (valid_jmp_i(jmp, imm)) {
emit_jmp_i(ctx, lo(dst), imm, rel, jmp);
} else {
/* Move large immediate to register */
emit_mov_i(ctx, MIPS_R_T6, imm);
emit_jmp_r(ctx, lo(dst), MIPS_R_T6, rel, jmp);
}
if (finish_jmp(ctx, jmp, off) < 0)
goto toofar;
break;
/* PC += off if dst == src */
/* PC += off if dst != src */
/* PC += off if dst & src */
/* PC += off if dst > src */
/* PC += off if dst >= src */
/* PC += off if dst < src */
/* PC += off if dst <= src */
/* PC += off if dst > src (signed) */
/* PC += off if dst >= src (signed) */
/* PC += off if dst < src (signed) */
/* PC += off if dst <= src (signed) */
case BPF_JMP | BPF_JEQ | BPF_X:
case BPF_JMP | BPF_JNE | BPF_X:
case BPF_JMP | BPF_JSET | BPF_X:
case BPF_JMP | BPF_JGT | BPF_X:
case BPF_JMP | BPF_JGE | BPF_X:
case BPF_JMP | BPF_JLT | BPF_X:
case BPF_JMP | BPF_JLE | BPF_X:
case BPF_JMP | BPF_JSGT | BPF_X:
case BPF_JMP | BPF_JSGE | BPF_X:
case BPF_JMP | BPF_JSLT | BPF_X:
case BPF_JMP | BPF_JSLE | BPF_X:
if (off == 0)
break;
setup_jmp_r(ctx, dst == src, BPF_OP(code), off, &jmp, &rel);
emit_jmp_r64(ctx, dst, src, rel, jmp);
if (finish_jmp(ctx, jmp, off) < 0)
goto toofar;
break;
/* PC += off if dst == imm */
/* PC += off if dst != imm */
/* PC += off if dst & imm */
/* PC += off if dst > imm */
/* PC += off if dst >= imm */
/* PC += off if dst < imm */
/* PC += off if dst <= imm */
/* PC += off if dst > imm (signed) */
/* PC += off if dst >= imm (signed) */
/* PC += off if dst < imm (signed) */
/* PC += off if dst <= imm (signed) */
case BPF_JMP | BPF_JEQ | BPF_K:
case BPF_JMP | BPF_JNE | BPF_K:
case BPF_JMP | BPF_JSET | BPF_K:
case BPF_JMP | BPF_JGT | BPF_K:
case BPF_JMP | BPF_JGE | BPF_K:
case BPF_JMP | BPF_JLT | BPF_K:
case BPF_JMP | BPF_JLE | BPF_K:
case BPF_JMP | BPF_JSGT | BPF_K:
case BPF_JMP | BPF_JSGE | BPF_K:
case BPF_JMP | BPF_JSLT | BPF_K:
case BPF_JMP | BPF_JSLE | BPF_K:
if (off == 0)
break;
setup_jmp_i(ctx, imm, 64, BPF_OP(code), off, &jmp, &rel);
emit_jmp_i64(ctx, dst, imm, rel, jmp);
if (finish_jmp(ctx, jmp, off) < 0)
goto toofar;
break;
/* PC += off */
case BPF_JMP | BPF_JA:
if (off == 0)
break;
if (emit_ja(ctx, off) < 0)
goto toofar;
break;
/* Tail call */
case BPF_JMP | BPF_TAIL_CALL:
if (emit_tail_call(ctx) < 0)
goto invalid;
break;
/* Function call */
case BPF_JMP | BPF_CALL:
if (emit_call(ctx, insn) < 0)
goto invalid;
break;
/* Function return */
case BPF_JMP | BPF_EXIT:
/*
* Optimization: when last instruction is EXIT
* simply continue to epilogue.
*/
if (ctx->bpf_index == ctx->program->len - 1)
break;
if (emit_exit(ctx) < 0)
goto toofar;
break;
default:
invalid:
pr_err_once("unknown opcode %02x\n", code);
return -EINVAL;
notyet:
pr_info_once("*** NOT YET: opcode %02x ***\n", code);
return -EFAULT;
toofar:
pr_info_once("*** TOO FAR: jump at %u opcode %02x ***\n",
ctx->bpf_index, code);
return -E2BIG;
}
return 0;
}