Re: [PATCH 04/12] powerpc: Implement hw_breakpoint_arch_parse()

From: Michael Ellerman
Date: Tue Jun 05 2018 - 07:06:46 EST


Frederic Weisbecker <frederic@xxxxxxxxxx> writes:
> On Mon, May 28, 2018 at 09:31:07PM +1000, Michael Ellerman wrote:
>> Frederic Weisbecker <frederic@xxxxxxxxxx> writes:
>>
>> > On Thu, May 24, 2018 at 12:01:52PM +1000, Michael Ellerman wrote:
>> >> Frederic Weisbecker <frederic@xxxxxxxxxx> writes:
>> >>
>> >> > diff --git a/arch/powerpc/kernel/hw_breakpoint.c b/arch/powerpc/kernel/hw_breakpoint.c
>> >> > index 348cac9..fba6527 100644
>> >> > --- a/arch/powerpc/kernel/hw_breakpoint.c
>> >> > +++ b/arch/powerpc/kernel/hw_breakpoint.c
>> >> > @@ -139,30 +139,31 @@ int arch_bp_generic_fields(int type, int *gen_bp_type)
>> >> > /*
>> >> > * Validate the arch-specific HW Breakpoint register settings
>> >> > */
>> >> > -int arch_validate_hwbkpt_settings(struct perf_event *bp)
>> >> > +int hw_breakpoint_arch_parse(struct perf_event *bp,
>> >> > + struct perf_event_attr *attr,
>> >> > + struct arch_hw_breakpoint *hw)
>> >>
>> >> I think the semantics here are that we are reading from bp/attr and
>> >> writing to hw?
>> >>
>> >> If so would some sprinkling of const on the first two parameters help
>> >> make that clearer?
>> >
>> > I seem to remember there was an issue with that due to the various functions
>> > we call that need to be converted to take const as well. I thought I would
>> > do it in a seperate series but actually it should be no big deal to do it
>> > on this one.
>>
>> Yeah, that does sometimes snowball out of control.
>>
>> > Let me try that and respin.
>>
>> Cool. It would be nice to have, but obviously not crucial.
>
> So I managed to constify the perf_event_attr parameter but not the struct perf_event *bp
> because the task target is fetched from it on is_compat_bp() on ARM64. I could constify it
> all the way up to test_ti_thread_flag() but the thread info can only be retrieved through
> a call to task_thread_info() and that's where the qualifier control ends. The const can not
> be passed there and we can't afford to constify the function either, I fear, as it is used
> everywhere for any purpose, including thread_info modifications.

Thanks for trying. 1 out of 2 const parameters is better than none :)

cheers