[PATCH 3/3] early kprobes: x86: don't try to recover ftraced instruction before ftrace get ready.

[Masami Hiramatsu]

(2015/03/04 13:39), Wang Nan wrote:
> On 2015/3/4 11:36, Masami Hiramatsu wrote:
>> (2015/03/04 11:24), Wang Nan wrote:
>>> On 2015/3/4 1:06, Petr Mladek wrote:
>>>> On Tue 2015-03-03 13:09:05, Wang Nan wrote:
>>>>> Before ftrace convertin instruction to nop, if an early kprobe is
>>>>> registered then unregistered, without this patch its first bytes will
>>>>> be replaced by head of NOP, which may confuse ftrace.
>>>>>
>>>>> Actually, since we have a patch which convert ftrace entry to nop
>>>>> when probing, this problem should never be triggered. Provide it for
>>>>> safety.
>>>>>
>>>>> Signed-off-by: Wang Nan <wangnan0 at huawei.com>
>>>>> ---
>>>>>  arch/x86/kernel/kprobes/core.c | 3 +++
>>>>>  1 file changed, 3 insertions(+)
>>>>>
>>>>> diff --git a/arch/x86/kernel/kprobes/core.c b/arch/x86/kernel/kprobes/core.c
>>>>> index 87beb64..c7d304d 100644
>>>>> --- a/arch/x86/kernel/kprobes/core.c
>>>>> +++ b/arch/x86/kernel/kprobes/core.c
>>>>> @@ -225,6 +225,9 @@ __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
>>>>>  	struct kprobe *kp;
>>>>>  	unsigned long faddr;
>>>>>  
>>>>> +	if (!kprobes_on_ftrace_initialized)
>>>>> +		return addr;
>>>>
>>>> This is not correct. The function has to return a buffer with the original
>>>> code also when it is modified by normal kprobes. If it is a normal
>>>> Kprobe, it reads the current code and replaces the first byte (INT3
>>>> instruction) with the saved kp->opcode.
>>>>
>>>>> +
>>>>>  	kp = get_kprobe((void *)addr);
>>>>>  	faddr = ftrace_location(addr);
>>>>
>>>> IMHO, the proper fix might be to replace the above line with
>>>>
>>>> 	if (kprobes_on_ftrace_initialized)
>>>> 		faddr = ftrace_location(addr);
>>>> 	else
>>>> 		faddr = 0UL;
>>>>
>>>> By other words, it might pretend that it is not a ftrace location
>>>> when the ftrace is not ready yet.
>>>>
>>>
>>> Thanks for your reply. I'll follow your suggection in my next version. I change
>>> it as follow to enable the checking.
>>>
>>> diff --git a/arch/x86/kernel/kprobes/core.c b/arch/x86/kernel/kprobes/core.c
>>> index 4e3d5a9..3241677 100644
>>> --- a/arch/x86/kernel/kprobes/core.c
>>> +++ b/arch/x86/kernel/kprobes/core.c
>>> @@ -234,6 +234,20 @@ __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
>>> 	 */
>>> 	if (WARN_ON(faddr && faddr != addr))
>>> 		return 0UL;
>>> +
>>> +	/*
>>> +	 * If ftrace is not ready yet, pretend this is not an ftrace
>>> +	 * location, because currently the target instruction has not
>>> +	 * been replaced by a NOP yet. When ftrace trying to convert
>>> +	 * it to NOP, kprobe should be notified and the kprobe data
>>> +	 * should be fixed at that time.
>>> +	 *
>>> +	 * Since it is possible that an early kprobe already on that
>>> +	 * place, don't return addr directly.
>>> +	 */
>>> +	if (likely(kprobes_on_ftrace_initialized))
>>> +		faddr = 0UL;
>>> +
>>> 	/*
>>> 	 * Use the current code if it is not modified by Kprobe
>>> 	 * and it cannot be modified by ftrace
>>>
>>
>> This is better, but I don't think we need bool XXX_initialized flags
>> for each subfunctions. Those should be serialized.
>>
>> Thank you,
>>
> 
> For this specific case, calling __recover_probed_insn() is mandatory for
> can_boost(). However, we can disallow early kprobes to be unregistered before
> ftrace is ready, and let ftrace fix all inconsistency by calling
> kprobe_on_ftrace_get_old_insn(). Which will make things simpler, and constrain
> the using scope of kprobes_on_ftrace_initialized to kernel/kprobes.c.

What I meant was consolidating those XXX_initialized flag to
kprobes_init_stage flag and enum kprobes_stage which has
{ KP_STG_NONE, KP_STG_EARLY, KP_STG_NORMAL } etc.
This can serialize staging phases and do not cause unexpected
initialized-flags combination.

> The
> cost is unable to do smoke test for early ftrace because it will remove all
> kprobe before returning. I think it should be acceptable. What do you think?

Ah, I see. We should change it to just remove only the kprobes which smoke test
inserted. Or, sort the test order to move it after the ftrace is initialized.

I guess latter is better, since at the point of smoke test is executed,
all the kprobe-events feature should be available.

Thank you,

> 
> Thank you.
> 
> 


-- 
Masami HIRAMATSU
Software Platform Research Dept. Linux Technology Research Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt at hitachi.com