Re: Porting DTrace to ARM

[Ryota Ozaki <ozaki-r%NetBSD.org@localhost>]

On Thu, Mar 6, 2014 at 10:46 PM, Christos Zoulas 
<christos%zoulas.com@localhost> wrote:
> On Mar 6,  1:55pm, ozaki-r%NetBSD.org@localhost (Ryota Ozaki) wrote:
> -- Subject: Re: Porting DTrace to ARM
>
> | On Thu, Mar 6, 2014 at 1:26 PM, Matt Thomas 
> <matt%3am-software.com@localhost> wrote:
> | >
> | > On Mar 5, 2014, at 8:25 PM, Ryota Ozaki <ozaki-r%netbsd.org@localhost> 
> wrote:
> | >
> | >> On Thu, Mar 6, 2014 at 12:48 PM, Matt Thomas 
> <matt%3am-software.com@localhost> wrote:
> | >>>
> | >>> On Mar 5, 2014, at 7:33 PM, Ryota Ozaki <ozaki-r%NetBSD.org@localhost> 
> wrote:
> | >>>
> | >>>> On Thu, Mar 6, 2014 at 1:39 AM, Matt Thomas 
> <matt%3am-software.com@localhost> wrote:
> | >>>>>
> | >>>>> On Mar 5, 2014, at 3:12 AM, Ryota Ozaki 
> <ozaki-r%netbsd.org@localhost> wrote:
> | >>>>>
> | >>>>>> - Replace cpu_id with cpuid in sys/arch/arm
> | >>>>>> - Can I commit the change?
> | >>>>>
> | >>>>> Why?  It's just churn for no reason I can see.
> | >>>>
> | >>>> The background is that cpu_id in sys/arch/arm conflicts with
> | >>>> the code in cddl and we have to change either one. I and christos
> | >>>> (he already replied in another mail) decided to change
> | >>>> sys/arch/arm, which seems less pain.
> | >>>
> | >>> The problem is that all the functions in cpufunc.h are cpu_xxx
> | >>> cpuid would be an outlier.
> | >>>
> | >>> I think a better solution might be to put a field for dtrace
> | >>> into cpu_data and just curcpu()->ci_dtraceinfo->foo
> | >>> to get to it instead have a parallel structure.
> | >>>
> | >>> You want to put a dtrace in mi_attach_cpu to initialize/allocate it.
> | >>
> | >> Sounds reasonable (except that it needs to modify cddl much though).
> | >> Can we do attach_cpu in a module?
> | >
> | > Too late for the most part.
> |
> | # oh, mi_attach_cpu is correct. I found it now :)
> |
> | Okay, so we need to have some code in src/sys. Hmm, big change
> | (for me). I can do it, but I don't know if it's ok or not.
> |
> | Christos, how do you think?
>
> It is probably easier to change cddl at this point or make cpu_id() an inline.

Well...I found that -current now doesn't require cpu_id tweaks for both
arch/arm and dtrace anymore. I guess changes in arch/arm headers
during my development solved the problem without realizing.

So I think we can postpone per-cpu data improvement after works
for arm finish. I'm listing the per-cpu work in my todo.

Regards,
  ozaki-r