Re: possible regression caused by "clocksource: sh_tmu: Set cpu_possible_mask to fix SMP broadcast"

[Simon Horman <horms@verge.net.au>]

On Thu, Jun 18, 2015 at 08:50:11AM +0900, Magnus Damm wrote:
> Hi Simon,
> 
> On Wed, Jun 17, 2015 at 5:42 PM, Simon Horman <horms@verge.net.au> wrote:
> > On Wed, Jun 17, 2015 at 04:51:07PM +0900, Magnus Damm wrote:
> >> Hi Geert,
> >>
> >> On Wed, Jun 17, 2015 at 4:45 PM, Geert Uytterhoeven
> >> <geert@linux-m68k.org> wrote:
> >> > On Wed, Jun 17, 2015 at 8:52 AM, Magnus Damm <magnus.damm@gmail.com> wrote:
> >> >> On Wed, Jun 17, 2015 at 11:48 AM, Simon Horman <horms@verge.net.au> wrote:
> >> >>> I have observed what appears to be a regression caused by
> >> >>> f2a5473861cf ("clocksource: sh_tmu: Set cpu_possible_mask to fix SMP
> >> >>> broadcast"), which was included in v3.19.
> >> >>>
> >> >>> The problem I see is that with a kernel compiled using marzen_defconfig on
> >> >>> the r8a7779/marzen board: i.e. a "legacy" kernel.
> >> >>>
> >> >>> The problem does not manifest when booting the same board using
> >> >>> a kernel built with shmobile_defconfig: i.e. a multiplatform kernel.
> >> >>>
> >> >>> It does not appear to affect other boards I have access to:
> >> >>> emev2/kzm9d, r7s72100/genmai, r8a73a4/ape6evm, r8a7740/armadillo900eva,
> >> >>> r8a77798/bockw r8a7790/lager r8a7791/koelsch, r8a7794/alt, sh73a0/kzm9g.
> >> >>>
> >> >>> I have observed this problem on the renesas-devel-20150617-v4.1-rc8
> >> >>> tag of my renesas tree. But I do not believe it is exclusive to that tag.
> >> >>>
> >> >>> The problem appears to be that there is no clocksource available during
> >> >>> early boot. The boot log at the end of the email was obtained by
> >> >>> enabling DEBUG_LL and earlyprintk.
> >> >>
> >> >> Thanks for reporting this. I suspect that the reason why this triggers
> >> >> on legacy is that we i such case most likely don't have TWD together
> >> >> with CCF. Since this is just old legacy code and the multiplatform
> >> >
> >> > We do have other A9 SMP SoCs (sh73a0, r8a7778).
> >> >
> >> > r8a7779.dtsi has an "arm,cortex-a9-twd-timer" node. r8a7778 hasn't.
> >>
> >> Right, r8a7778 is not SMP so I don't think there is any TWD there. Or
> >> at least the Linux driver did not use to support running in UP mode.
> >>
> >> > sh73a0 also has it, but see http://www.spinics.net/lists/linux-sh/msg39128.html
> >> > As bockw and marzen still have -reference variants, it may be more
> >> > difficult to fix.
> >>
> >> Your TWD DT node is nicely pointing out the clock via CCF. On legacy
> >> this is most likely not possible, so we end up with trying to
> >> automatically determining the clock by assuming another clock is
> >> running. At least it used to be like that.
> >>
> >> >> version should be equivalent I suggest that the correct approach is to
> >> >> use multiplatform support instead of legacy. And legacy r8a7779
> >> >> support should be removed right away if it hasn't already.
> >> >
> >> > Please wait, so I don't have to rebase the sh73a0/r8a7740 patches I'm
> >> > about to send out ;-)
> >>
> >> Sure, please push out so we can clean up. =)
> >
> > I'm happy to proceed with an orderly removal of marzen legacy
> > as a resolution to this problem.
> 
> I agree about removing the marzen legacy code. For anyone wanting to
> use the legacy code then the following mangled patch can be used to
> work around the "calibrating delay" issue:

As it is small I wonder if we we should consider applying the patch below
as a fix and then treating the removal of r8a7779 legacy as a separate
issue.

> From: Magnus Damm <damm+renesas@opensource.se>
> 
> Rely on CPU frequency information provided by the r8a7779 DTB for
> delay calculation.
> 
> Signed-off-by: Magnus Damm <damm+renesas@opensource.se>
> 
> ---
> 
> 
>  arch/arm/mach-shmobile/setup-r8a7779.c |    2 ++
> 
>  1 file changed, 2 insertions(+)
> 
> 
> --- 0001/arch/arm/mach-shmobile/setup-r8a7779.c
> 
> +++ work/arch/arm/mach-shmobile/setup-r8a7779.c 2015-06-18
> 08:46:21.062366518 +0900
> 
> @@ -675,6 +675,8 @@ void __init r8a7779_add_standard_devices
> 
> 
> 
>  void __init r8a7779_add_early_devices(void)
> 
>  {
> 
> + shmobile_init_delay();
> 
> +
> 
>   early_platform_add_devices(r8a7779_early_devices,
> 
>     ARRAY_SIZE(r8a7779_early_devices));
>