Re: [PATCH v10 1/5] add infrastructure for tagging functions as error injectable

[Alexei Starovoitov <ast@fb.com>]

On 12/19/17 11:13 PM, Masami Hiramatsu wrote:
> On Tue, 19 Dec 2017 18:14:17 -0800
> Alexei Starovoitov <ast@fb.com> wrote:
>
>> On 12/18/17 10:29 PM, Masami Hiramatsu wrote:
>>>>
>>>> +#if defined(__KERNEL__) && !defined(__ASSEMBLY__)
>>>> +#ifdef CONFIG_BPF_KPROBE_OVERRIDE
>>>
>>> BTW, CONFIG_BPF_KPROBE_OVERRIDE is also confusable name.
>>> Since this feature override a function to just return with
>>> some return value (as far as I understand, or would you
>>> also plan to modify execution path inside a function?),
>>> I think it should be better CONFIG_BPF_FUNCTION_OVERRIDE or
>>> CONFIG_BPF_EXECUTION_OVERRIDE.
>>
>> I don't think such renaming makes sense.
>> The feature is overriding kprobe by changing how kprobe returns.
>> It doesn't override BPF_FUNCTION or BPF_EXECUTION.
>
> No, I meant this is BPF's feature which override FUNCTION, so
> BPF is a kind of namespace. (Is that only for a function entry
> because it can not tweak stackframe at this morment?)
>
>> The kernel enters and exists bpf program as normal.
>
> Yeah, but that bpf program modifies instruction pointer, am I correct?

no. bpf side is asking kprobe side to modify it.
bpf cannot do such things as modifying IP or any other register
directly.

>>
>>> Indeed, BPF is based on kprobes, but it seems you are limiting it
>>> with ftrace (function-call trace) (I'm not sure the reason why),
>>> so using "kprobes" for this feature seems strange for me.
>>
>> do you have an idea how kprobe override can happen when kprobe
>> placed in the middle of the function?
>
> For example, if you know a basic block in the function, maybe
> you can skip a block or something like that. But nowadays
> it is somewhat hard because optimizer mixed it up.

still missing how that can work...