[PATCH v19 10/11] ARM: kprobes: check register usage for probed instruction.
Jon Medhurst (Tixy)
tixy at linaro.org
Tue Jan 13 07:01:08 PST 2015
On Mon, 2015-01-05 at 19:29 +0800, Wang Nan wrote:
> This patch utilizes previous introduced checker to check register usage
> for probed ARM instruction and saves it in a mask. Futher patch will
> use such information to avoid simuation or emulation.
There's a couple of spelling mistakes above and minor grammar nitpicks,
my suggestion to improve those would be to add the bits in [] in the
following...
This patch utilizes [the] previous[ly] introduced checker to check
register usage for probed ARM instruction and saves it in a mask. [A]
fu[r]ther patch will use such information to avoid simu[l]ation or
emulation.
My comments on the code below are mostly suggestions on how to simplify
things, though the code is already quite re is one bug.
> Signed-off-by: Wang Nan <wangnan0 at huawei.com>
> ---
> arch/arm/include/asm/probes.h | 10 +++
> arch/arm/probes/decode.c | 7 ++
> arch/arm/probes/kprobes/actions-arm.c | 2 +-
> arch/arm/probes/kprobes/checkers-arm.c | 124 +++++++++++++++++++++++++++++++++
> arch/arm/probes/kprobes/checkers.h | 1 +
> 5 files changed, 143 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm/include/asm/probes.h b/arch/arm/include/asm/probes.h
> index f0a1ee8..27b65b7 100644
> --- a/arch/arm/include/asm/probes.h
> +++ b/arch/arm/include/asm/probes.h
> @@ -41,6 +41,16 @@ struct arch_probes_insn {
> probes_insn_singlestep_t *insn_singlestep;
> probes_insn_fn_t *insn_fn;
> int stack_space;
> +
> + /* Use 1 bit for a register. */
> +#define REG_NO_USE (0)
> +#define REG_USE (1)
No need for () around the 0 and 1
> +#define __register_usage_flag(n, f) ((f) << (n))
> +#define __clean_register_flag(m, n) ((m) & (~(1 << (n))))
> +#define __set_register_flag(m, n, f) (__clean_register_flag(m, n) | __register_usage_flag(n, f))
> +#define set_register_nouse(m, n) do {(m) = __set_register_flag(m, n, REG_NO_USE);} while(0)
> +#define set_register_use(m, n) do {(m) = __set_register_flag(m, n, REG_USE);} while(0)
All these macros seem a bit convoluted to me, and I don't think it makes
things any clearer than open coding a bit set operation like
|= 1 << reg
especially as I don't thing those macros will end up getting much use.
> + unsigned long register_usage_flag;
To nit-pick, as this isn't a single flag I would add an 's' to make the
name plural, i.e. 'register_usage_flags'
> };
>
> #endif /* __ASSEMBLY__ */
> diff --git a/arch/arm/probes/decode.c b/arch/arm/probes/decode.c
> index f9d7c42..40f9402 100644
> --- a/arch/arm/probes/decode.c
> +++ b/arch/arm/probes/decode.c
> @@ -435,6 +435,13 @@ probes_decode_insn(probes_opcode_t insn, struct arch_probes_insn *asi,
> */
> asi->stack_space = 0;
>
> + /*
> + * Similay to stack_space, register_usage_flag is filled by
Typo, s/Similay/Similarly/
> + * checkers. Its default value is set to ~0, which is 'all
> + * registers are used', to prevent any potential optimization.
> + */
> + asi->register_usage_flag = ~(0UL);
No need for () around 0UL
> +
> if (emulate)
> insn = prepare_emulated_insn(insn, asi, thumb);
>
> diff --git a/arch/arm/probes/kprobes/actions-arm.c b/arch/arm/probes/kprobes/actions-arm.c
> index 4fedd4c..26e435b 100644
> --- a/arch/arm/probes/kprobes/actions-arm.c
> +++ b/arch/arm/probes/kprobes/actions-arm.c
> @@ -340,4 +340,4 @@ const union decode_action kprobes_arm_actions[NUM_PROBES_ARM_ACTIONS] = {
> [PROBES_LDMSTM] = {.decoder = kprobe_decode_ldmstm}
> };
>
> -const struct decode_checker *kprobes_arm_checkers[] = {arm_stack_checker, NULL};
> +const struct decode_checker *kprobes_arm_checkers[] = {arm_stack_checker, arm_regs_checker, NULL};
> diff --git a/arch/arm/probes/kprobes/checkers-arm.c b/arch/arm/probes/kprobes/checkers-arm.c
> index f817663..1929225 100644
> --- a/arch/arm/probes/kprobes/checkers-arm.c
> +++ b/arch/arm/probes/kprobes/checkers-arm.c
> @@ -97,3 +97,127 @@ const struct decode_checker arm_stack_checker[NUM_PROBES_ARM_ACTIONS] = {
> [PROBES_STORE] = {.checker = arm_check_stack},
> [PROBES_LDMSTM] = {.checker = arm_check_stack},
> };
> +
> +static enum probes_insn __kprobes arm_check_regs_nouse(probes_opcode_t insn,
> + struct arch_probes_insn *asi,
> + const struct decode_header *h)
> +{
> + asi->register_usage_flag = 0;
> + return INSN_GOOD;
> +}
> +
> +static void __arm_check_regs(probes_opcode_t insn,
> + const struct decode_header *h,
> + int *quintuple)
> +{
> + int i;
> + u32 regs = h->type_regs.bits >> DECODE_TYPE_BITS;
> + probes_opcode_t mask, shifted;
> +
> + memset(quintuple, 0xff, sizeof(int) * 5);
Not sure that filling with 0xff is a C standard compliant way of
initialising int's to a negative value. Anyway, memset isn't really
needed, see next comment.
> + for (i = 0, shifted = insn, mask = 0xf; regs != 0;
> + regs >>= 4, mask <<= 4, shifted >>= 4, i++)
> + if (regs & 0xf)
> + quintuple[i] = shifted & 0xf;
Looks a bit complicated, the last 6 lines could be more simply expressed
as:
for (i = 0; i < 5; regs >>= 4, insn >>= 4, i++)
quintuple[i] = regs & 0xf ? insn & 0xf : -1;
Though I think that this function may be a bit overkill anyway, it is
only really needed in one place, so may as well be inlined there, that
is the function below...
> +}
> +
> +static enum probes_insn arm_check_regs_normal(probes_opcode_t insn,
> + struct arch_probes_insn *asi,
> + const struct decode_header *h)
> +{
> + int quintuple[5], i;
> + asi->register_usage_flag = 0;
> + __arm_check_regs(insn, h, quintuple);
> + for (i = 0; i < 5; i++) {
> + int r = quintuple[i];
> + if (r < 0)
> + continue;
> + set_register_use(asi->register_usage_flag, r);
> + }
> +
The above can be as simply written without using _arm_check_regs like...
u32 regs = h->type_regs.bits >> DECODE_TYPE_BITS;
int i;
asi->register_usage_flag = 0;
for (i = 0; i < 5; regs >>= 4, insn >>= 4, i++)
if (regs & 0xf)
asi->register_usage_flag |= 1 << (insn & 0xf);
> + return INSN_GOOD;
> +}
> +
> +static enum probes_insn arm_check_regs_ldmstm(probes_opcode_t insn,
> + struct arch_probes_insn *asi,
> + const struct decode_header *h)
> +{
> + unsigned int reglist = insn & 0xffff;
> + unsigned int rn = (insn >> 16) & 0xf;
> + int i;
> +
Think you forgot to clear asi->register_usage_flag here, it will still
contain 0xffffffff set by from probes_decode_insn(). However, we don't
need this if we follow my next suggestion.
> + set_register_use(asi->register_usage_flag, rn);
> + for (i = 0; reglist > 0; i++, reglist >>= 1)
> + if (reglist & 1)
> + set_register_use(asi->register_usage_flag, i);
The preceding 5 lines are equivalent to
asi->register_usage_flag = reglist | (1 << rn);
so how about we just use that?
> + return INSN_GOOD;
> +}
> +
> +static enum probes_insn arm_check_regs_mov_ip_sp(probes_opcode_t insn,
> + struct arch_probes_insn *asi,
> + const struct decode_header *h)
> +{
> + /* should be 'mov ip, sp' */
> + set_register_use(asi->register_usage_flag, 12);
> + set_register_use(asi->register_usage_flag, 13);
register_usage_flag needs setting to zero before setting the flags. Or
we could just write this function as
/* Instruction is 'mov ip, sp' i.e. 'mov r12, r13' */
asi->register_usage_flag = (1 << 12) | (1<< 13);
> + return INSN_GOOD;
> +}
> +
> +/*
> + * | Rn |Rt/d| | Rm |
> + * LDRD (register) cccc 000x x0x0 xxxx xxxx xxxx 1101 xxxx
> + * STRD (register) cccc 000x x0x0 xxxx xxxx xxxx 1111 xxxx
> + * | Rn |Rt/d| |imm4L|
> + * LDRD (immediate) cccc 000x x1x0 xxxx xxxx xxxx 1101 xxxx
> + * STRD (immediate) cccc 000x x1x0 xxxx xxxx xxxx 1111 xxxx
> + *
> + * Such instructions access Rt/d and its next register, so different
> + * from others, a specific checker is required for Rt/d and Rt2/d2.
> + */
> +static enum probes_insn arm_check_regs_ldrdstrd(probes_opcode_t insn,
> + struct arch_probes_insn *asi,
> + const struct decode_header *h)
> +{
> + int quintuple[5], rn, rdt, rm;
> + asi->register_usage_flag = 0;
> + __arm_check_regs(insn, h, quintuple);
> +
> + rn = quintuple[4];
> + rdt = quintuple[3];
> + rm = quintuple[0];
> + set_register_use(asi->register_usage_flag, rn);
> + set_register_use(asi->register_usage_flag, rdt);
> + set_register_use(asi->register_usage_flag, rdt + 1);
> + if (rm >= 0)
> + set_register_use(asi->register_usage_flag, rm);
> +
We can avoid most of this code if we reuse arm_check_regs_normal and
then add in the use of rdt+1, i.e. this function can be simply...
int rdt = (insn >> 12) & 0xf;
arm_check_regs_normal(insn, asi, h);
asi->register_usage_flag |= 1 << (rdt + 1);
> + return INSN_GOOD;
> +}
> +
> +
> +const struct decode_checker arm_regs_checker[NUM_PROBES_ARM_ACTIONS] = {
> + [PROBES_EMULATE_NONE] = {.checker = arm_check_regs_nouse},
> + [PROBES_SIMULATE_NOP] = {.checker = arm_check_regs_nouse},
> + [PROBES_MRS] = {.checker = arm_check_regs_normal},
> + [PROBES_SATURATING_ARITHMETIC] = {.checker = arm_check_regs_normal},
> + [PROBES_MUL1] = {.checker = arm_check_regs_normal},
> + [PROBES_MUL2] = {.checker = arm_check_regs_normal},
> + [PROBES_MUL_ADD_LONG] = {.checker = arm_check_regs_normal},
> + [PROBES_MUL_ADD] = {.checker = arm_check_regs_normal},
> + [PROBES_LOAD] = {.checker = arm_check_regs_normal},
> + [PROBES_LOAD_EXTRA] = {.checker = arm_check_regs_normal},
> + [PROBES_STORE] = {.checker = arm_check_regs_normal},
> + [PROBES_STORE_EXTRA] = {.checker = arm_check_regs_normal},
> + [PROBES_DATA_PROCESSING_REG] = {.checker = arm_check_regs_normal},
> + [PROBES_DATA_PROCESSING_IMM] = {.checker = arm_check_regs_normal},
> + [PROBES_SATURATE] = {.checker = arm_check_regs_normal},
> + [PROBES_REV] = {.checker = arm_check_regs_normal},
> + [PROBES_MMI] = {.checker = arm_check_regs_normal},
> + [PROBES_PACK] = {.checker = arm_check_regs_normal},
> + [PROBES_EXTEND] = {.checker = arm_check_regs_normal},
> + [PROBES_EXTEND_ADD] = {.checker = arm_check_regs_normal},
> + [PROBES_BITFIELD] = {.checker = arm_check_regs_normal},
> + [PROBES_LDMSTM] = {.checker = arm_check_regs_ldmstm},
> + [PROBES_MOV_IP_SP] = {.checker = arm_check_regs_mov_ip_sp},
> + [PROBES_LDRSTRD] = {.checker = arm_check_regs_ldrdstrd},
> +};
> diff --git a/arch/arm/probes/kprobes/checkers.h b/arch/arm/probes/kprobes/checkers.h
> index bddfa0e..cf6c9e7 100644
> --- a/arch/arm/probes/kprobes/checkers.h
> +++ b/arch/arm/probes/kprobes/checkers.h
> @@ -47,6 +47,7 @@ extern const union decode_action stack_check_actions[];
>
> #ifndef CONFIG_THUMB2_KERNEL
> extern const struct decode_checker arm_stack_checker[];
> +extern const struct decode_checker arm_regs_checker[];
> #else
> #endif
> extern const struct decode_checker t32_stack_checker[];
I've made a lot, of suggestions for code changes, so I think it's best
if I follow up to this email with a new version of this patch which
applies them all so we can see what the bigger picture is...
--
Tixy
More information about the linux-arm-kernel
mailing list