Re: [PATCH 01/02] ARM: shmobile: Add r8a73a4 SMP support using APMU code

[Olof Johansson <olof@lixom.net>]

On Wed, Aug 28, 2013 at 03:26:56PM +0900, Magnus Damm wrote:
> Hi Olof,
> 
> On Thu, Aug 22, 2013 at 2:48 PM, Olof Johansson <olof@lixom.net> wrote:
> > [+khilman]
> >
> > Hi,
> >
> > On Wed, Aug 21, 2013 at 05:57:15PM +0900, Simon Horman wrote:
> >> On Thu, Aug 08, 2013 at 02:52:32PM +0900, Magnus Damm wrote:
> >> > Hi Arnd and Olof,
> >> >
> >> > On Thu, Aug 8, 2013 at 7:55 AM, Magnus Damm <magnus.damm@gmail.com> wrote:
> >> > > From: Magnus Damm <damm@opensource.se>
> >> > >
> >> > > Add r8a73a4 SMP support using the shared APMU code. To enable
> >> > > SMP the r8a73a4 specific DTS needs to be updated to include
> >> > > CPU cores, and this is happening in a separate patch.
> >> > >
> >> > > Signed-off-by: Magnus Damm <damm@opensource.se>
> >> > > ---
> >> > >
> >> > >  arch/arm/mach-shmobile/Makefile               |    1
> >> > >  arch/arm/mach-shmobile/include/mach/r8a73a4.h |    1
> >> > >  arch/arm/mach-shmobile/setup-r8a73a4.c        |    3 +
> >> > >  arch/arm/mach-shmobile/smp-r8a73a4.c          |   75 +++++++++++++++++++++++++
> >> > >  4 files changed, 80 insertions(+)
> >> > >
> >> > > --- 0001/arch/arm/mach-shmobile/Makefile
> >> > > +++ work/arch/arm/mach-shmobile/Makefile        2013-08-07 20:07:31.000000000 +0900
> >> > > @@ -33,6 +33,7 @@ endif
> >> > >  # SMP objects
> >> > >  smp-y                          := platsmp.o headsmp.o
> >> > >  smp-$(CONFIG_ARCH_SH73A0)      += smp-sh73a0.o headsmp-scu.o platsmp-scu.o
> >> > > +smp-$(CONFIG_ARCH_R8A73A4)     += smp-r8a73a4.o platsmp-apmu.o
> >> > >  smp-$(CONFIG_ARCH_R8A7779)     += smp-r8a7779.o headsmp-scu.o platsmp-scu.o
> >> > >  smp-$(CONFIG_ARCH_EMEV2)       += smp-emev2.o headsmp-scu.o platsmp-scu.o
> >> > >
> >> > > --- 0008/arch/arm/mach-shmobile/include/mach/r8a73a4.h
> >> > > +++ work/arch/arm/mach-shmobile/include/mach/r8a73a4.h  2013-08-07 20:08:02.000000000 +0900
> >> > > @@ -6,5 +6,6 @@ void r8a73a4_add_dt_devices(void);
> >> > >  void r8a73a4_clock_init(void);
> >> > >  void r8a73a4_pinmux_init(void);
> >> > >  void r8a73a4_init_early(void);
> >> > > +extern struct smp_operations r8a73a4_smp_ops;
> >> > >
> >> > >  #endif /* __ASM_R8A73A4_H__ */
> >> > > --- 0008/arch/arm/mach-shmobile/setup-r8a73a4.c
> >> > > +++ work/arch/arm/mach-shmobile/setup-r8a73a4.c 2013-08-07 20:07:48.000000000 +0900
> >> > > @@ -212,6 +212,9 @@ void __init r8a73a4_init_early(void)
> >> > >  #ifndef CONFIG_ARM_ARCH_TIMER
> >> > >         shmobile_setup_delay(1500, 2, 4); /* Cortex-A15 @ 1500MHz */
> >> > >  #endif
> >> > > +#ifdef CONFIG_SMP
> >> > > +       smp_set_ops(&r8a73a4_smp_ops);
> >> > > +#endif
> >> > >  }
> >> >
> >> > Arnd or Olof, may I ask for your advice here?
> >> >
> >> > I think it's quite nice to use smp_set_ops() in ->init_early() to
> >> > install the per-SoC SMP operations. I prefer that over using the
> >> > smp_ops directly in the per-board DT_MACHINE_START. Which way do you
> >> > prefer?
> >>
> >> Olof, Arnd, Ping.
> >
> >
> > Sorry, missed the original question when it was posted.
> >
> > I guess I don't see the benefit of doing it in code vs smp_init? You
> > already have a DT_MACHINE section for r8a73a4 in this case.
> 
> The benefit would be to reduce the number of callbacks used in
> DT_MACHINE. Today on mach-shmobile we are already using ->init_early()
> for delay and timer setup, and if we also can invoke smp_set_ops()
> there then we would have per-SoC SMP support hooked in without having
> to use either ->smp or ->smp_init() for every board.
> 
> As you may have seen we have quite a few DT_MACHINE entires in
> mach-shmobile, and I'd like to keep them as small and consistent as
> ever possible. I can't really see how a board specific DT_MACHINE has
> anything to do with state of SMP on a particular SoC, so for any given
> board I'd prefer to keep the board code without any SMP dependencies
> and simply use ->init_early() to setup delay, timers and SMP.
> 
> > If you use .smp_init = smp_init_ops(r8a73a4_smp_ops) then there's no
> > ifdef needed down in the DT_MACHINE section either.
> 
> It's nice to not use the #ifdefs, I agree. I'd like to have a dummy
> stub for smp_set_ops() too in the case of CONFIG_SMP=n, that would
> make the above code a bit prettier.
> 
> I hope this clarifies the reasons behind my question!
> 
> If I understand you correctly then I guess you prefer keeping the code
> as is with DT_MACHINE ->smp = smp_ops(foo) over using smp_set_ops() in
> ->init_early()?

Yes, that was my initial preference. However, that was based on the assumption
that you would keep the current setup with several DT_MACHINE entries.

If you think you will shortly start to distill down and only keep a couple of
DT_MACHINE structures (and start sharing them between SoCs), then doing this in
code makes sense. However, if that is further out than a release or two then
I think doing it in DT_MACHINE makes more sense.


-Olof