Message ID | 1409116109-67330-1-git-send-email-wangnan0@huawei.com |
---|---|
State | New |
Headers | show |
On Wed, 2014-08-27 at 13:08 +0800, Wang Nan wrote: > When kprobe decoding instruction, original code calls instruction > specific decoder if emulate is set to false. However, instructions with > DECODE_TYPE_EMULATE are in fact don't have their decoder. What in the > action table are in fact handlers. For example: > > /* LDRD (immediate) cccc 000x x1x0 xxxx xxxx xxxx 1101 xxxx */ > /* STRD (immediate) cccc 000x x1x0 xxxx xxxx xxxx 1111 xxxx */ > DECODE_EMULATEX (0x0e5000d0, 0x004000d0, PROBES_LDRSTRD, > REGS(NOPCWB, NOPCX, 0, 0, 0)), > > and > > const union decode_action kprobes_arm_actions[NUM_PROBES_ARM_ACTIONS] = { > ... > [PROBES_LDRSTRD] = {.handler = emulate_ldrdstrd}, > ... > > In this situation, original code calls 'emulate_ldrdstrd' as a decoder, > which is obviously incorrect. Except that situation can't occur because when arm_probes_decode_insn() is called with kprobes_arm_actions then emulate is set to true (see arch_prepare_kprobe). For the situation where emulate is false, i.e. when called from arch_uprobe_analyze_insn(), then this provides these actions... const union decode_action uprobes_probes_actions[] = { ... [PROBES_LDRSTRD] = {.decoder = decode_pc_ro}, ... and we do want that decode function called. Basically, the code is behaving as it was designed and when passed emulate=false it turns the behaviour of emulated instruction forms into into the 'custom' action. > > This patch makes it returns INSN_GOOD directly when 'emulate' is not > true. > > Signed-off-by: Wang Nan <wangnan0@huawei.com> > Cc: "David A. Long" <dave.long@linaro.org> > Cc: Russell King <linux@arm.linux.org.uk> > Cc: Jon Medhurst <tixy@linaro.org> > Cc: Taras Kondratiuk <taras.kondratiuk@linaro.org> > Cc: Ben Dooks <ben.dooks@codethink.co.uk> > --- > arch/arm/kernel/probes.c | 3 +-- > 1 file changed, 1 insertion(+), 2 deletions(-) > > diff --git a/arch/arm/kernel/probes.c b/arch/arm/kernel/probes.c > index a8ab540..1c77b8d 100644 > --- a/arch/arm/kernel/probes.c > +++ b/arch/arm/kernel/probes.c > @@ -436,8 +436,7 @@ probes_decode_insn(probes_opcode_t insn, struct arch_probes_insn *asi, > struct decode_emulate *d = (struct decode_emulate *)h; > > if (!emulate) > - return actions[d->handler.action].decoder(insn, > - asi, h); > + return INSN_GOOD; > > asi->insn_handler = actions[d->handler.action].handler; > set_emulated_insn(insn, asi, thumb);
diff --git a/arch/arm/kernel/probes.c b/arch/arm/kernel/probes.c index a8ab540..1c77b8d 100644 --- a/arch/arm/kernel/probes.c +++ b/arch/arm/kernel/probes.c @@ -436,8 +436,7 @@ probes_decode_insn(probes_opcode_t insn, struct arch_probes_insn *asi, struct decode_emulate *d = (struct decode_emulate *)h; if (!emulate) - return actions[d->handler.action].decoder(insn, - asi, h); + return INSN_GOOD; asi->insn_handler = actions[d->handler.action].handler; set_emulated_insn(insn, asi, thumb);
When kprobe decoding instruction, original code calls instruction specific decoder if emulate is set to false. However, instructions with DECODE_TYPE_EMULATE are in fact don't have their decoder. What in the action table are in fact handlers. For example: /* LDRD (immediate) cccc 000x x1x0 xxxx xxxx xxxx 1101 xxxx */ /* STRD (immediate) cccc 000x x1x0 xxxx xxxx xxxx 1111 xxxx */ DECODE_EMULATEX (0x0e5000d0, 0x004000d0, PROBES_LDRSTRD, REGS(NOPCWB, NOPCX, 0, 0, 0)), and const union decode_action kprobes_arm_actions[NUM_PROBES_ARM_ACTIONS] = { ... [PROBES_LDRSTRD] = {.handler = emulate_ldrdstrd}, ... In this situation, original code calls 'emulate_ldrdstrd' as a decoder, which is obviously incorrect. This patch makes it returns INSN_GOOD directly when 'emulate' is not true. Signed-off-by: Wang Nan <wangnan0@huawei.com> Cc: "David A. Long" <dave.long@linaro.org> Cc: Russell King <linux@arm.linux.org.uk> Cc: Jon Medhurst <tixy@linaro.org> Cc: Taras Kondratiuk <taras.kondratiuk@linaro.org> Cc: Ben Dooks <ben.dooks@codethink.co.uk> --- arch/arm/kernel/probes.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)