[PATCH v4 3/6] arm64: Kprobes with single stepping support

[dave.long@linaro.org (David Long)]

On 01/19/15 04:03, Pratyush Anand wrote:
>
>
> On Saturday 17 January 2015 12:58 AM, David Long wrote:
>>>> +static bool aarch64_insn_is_steppable(u32 insn)
>>>> +{
>>>> +       if (aarch64_get_insn_class(insn) == AARCH64_INSN_CLS_BR_SYS) {
>>>> +               if (aarch64_insn_is_branch(insn))
>>>> +                       return false;
>>>> +
>>>> +               /* modification of daif creates issues */
>>>> +               if (aarch64_insn_is_msr_daif(insn))
>>>> +                       return false;
>>>> +
>>>> +               if (aarch64_insn_is_hint(insn))
>>>> +                       return aarch64_insn_is_nop(insn);
>>>> +
>>>> +               return true;
>>>> +       }
>>>> +
>>>> +       if (aarch64_insn_uses_literal(insn))
>>>> +               return false;
>>>> +
>>>> +       if (aarch64_insn_is_exclusive(insn))
>>>> +               return false;
>>>> +
>>>> +       return true;
>>>
>>> Default true return may not be a good idea until we are sure that we
>>> are returning false for all possible
>>> simulation and rejection cases. In my opinion, its better to return
>>> true only for steppable and false for
>>> all remaining.
>>>
>>
>> I struggled a little with this when I did it but I decided if the
>> question was:  "should we have to recognize every instruction before
>> deciding it was single-steppable or should we only recognize
>> instructions that are *not* single-steppable", maybe it was OK to do the
>> latter while recognizing extensions to the instruction set *could* end
>> up (temporarly) allowing us to try and fail (badly) at single-stepping
>> any problematic new instructions.  Certainly opinions could differ.  If
>
> Lets see what others say, but I see that this approach will result in
> undesired behavior. For example: a probe has been tried to insert to svc
> instruction. SVC or any other exception generation instruction is
> expected to be rejected. But, current aarch64_insn_is_steppable will
> return true for it and then kprobe/uprobe code will allow to insert
> probe at that instruction, which will be wrong, no? I mean, I do not see
> a way to get into last else (INSN_REJECTED) of arm_kprobe_decode_insn.
>
> So, if we go with this approach we need to insure that we cover all
> simulation-able and reject-able cases in aarch64_insn_is_steppable.
>

yes, of course.  Any case that's missing in the current code needs to be 
fixed.  If the result starts to look less practical than the 
table-driven code then the new approach needs to be discarded.

> ~Pratyush
>
>
>
>> the consensus is that we can't allow this to ever happen (because old
>> kprobe code is running on new hardware) then I think the only choice is
>> to return to parsing binary tables.  Hopefully I could still find a way
>> to leverage insn.c in that case.