Re: [Qemu-devel] [PATCH v2 02/20] target-i386: move back cpu_exec_init() to init

[Eduardo Habkost]

On Fri, Oct 14, 2016 at 04:12:20PM +0200, Laurent Vivier wrote:
> 
> 
> On 14/10/2016 15:48, Eduardo Habkost wrote:
> > On Fri, Oct 14, 2016 at 03:34:17PM +0200, Laurent Vivier wrote:
> >> On 14/10/2016 15:33, Eduardo Habkost wrote:
> >>> On Thu, Oct 13, 2016 at 06:24:44PM +0200, Laurent Vivier wrote:
> >>>> We have now the cpu_exec_realize() in realize,
> >>>> so the init part must be in init.
> >>>>
> >>>> I've removed the cannot_destroy_with_object_finalize_yet field as
> >>>> unsafe references have been moved to cpu_exec_realize().
> >>>> (tested with QOM command provided by commit 4c315c27 with
> >>>> "athlon-x86_64-cpu")
> >>>>
> >>>> CC: Eduardo Habkost <address@hidden>
> >>>> Signed-off-by: Laurent Vivier <address@hidden>
> >>>
> >>> Resending the question I asked in my reply to v1:
> >>>
> >>> Instead of creating requiring each subclass to manually call
> >>> cpu_exec_init()) on instance_init, why don't we move parts of
> >>> cpu_exec_init()/cpu_exec_realize() code to cpu_common_initfn()?
> >>> (TYPE_CPU's instance_init)
> >>>
> >>> (And if there's any code that needs to be run after the
> >>> subclasses instance_init functions, we can just add a
> >>> instance_post_init function to TYPE_CPU).
> >>>
> >>
> >> It's done in PATCH 20/20.
> >>
> >> Is that what you want?
> > 
> > Yes (except that I would have inlined the the cpu_exec_init()
> > code inside cpu_common_init()).
> > 
> > I think I expected this to be done in a single step, that
> > wouldn't require touching code for all architectures three times.
> > Something like:
> 
> Well, I've added several steps to help to review and break nothing.

To me, it made review harder. Architecture maintainers have to
review 3 different patches that touch their code, instead of a
single one.

Anyway, I would still give my Reviewed-by line for the i386 part,
if you believe it's better to do it this way.

> 
> > 1) Move cpu->as, cpu->num_ases, cpu->thread_id, cpu->memory
> >    initialization, and "memory" property registration from
> >    cpu_exec_init() to cpu_common_init() (no architecture code
> >    touched).
> 
> system_memory (for the "memory" property) is declared as static in
> exec.c, so we can't move it to cpu_common_init().

In this case we can't inline it, that's true. But it still
doesn't require touching each architecture 3 times (we could just
create a cpu_exec_instance_init() function in exec.c and call it
from cpu_common_init()).

(Or we could move system_memory to MachineState, but that could
be done in a follow-up patch).

-- 
Eduardo