On Fri, 2017-02-17 at 21:26 +1100, Michael Ellerman wrote:
> Russell Currey <rus...@russell.cc> writes:
>
> > diff --git a/arch/powerpc/kernel/cputable.c b/arch/powerpc/kernel/cputable.c
> > index 6a82ef039c50..d23a54b09436 100644
> > --- a/arch/powerpc/kernel/cputable.c
> > +++ b/arch/powerpc/kernel/cputable.c
> > @@ -386,6 +386,25 @@ static struct cpu_spec __initdata cpu_specs[] = {
> > .machine_check_early =
> > __machine_check_early_realmode_p8,
> > .platform = "power8",
> > },
> > + { /* 3.00-compliant processor, i.e. Power9 "architected"
> > mode */
> > + .pvr_mask = 0xffffffff,
> > + .pvr_value = 0x0f000005,
> > + .cpu_name = "POWER9 (architected)",
> > + .cpu_features = CPU_FTRS_POWER9,
> > + .cpu_user_features = COMMON_USER_POWER9,
> > + .cpu_user_features2 = COMMON_USER2_POWER9,
> > + .mmu_features = MMU_FTRS_POWER9,
> > + .icache_bsize = 128,
> > + .dcache_bsize = 128,
> > + .num_pmcs = 6,
>
> It's important *not* to set num_pmcs for the architected PVRs.
>
> See the comment in setup_cpu_spec():
>
> /*
> * If we are overriding a previous value derived from the real
> * PVR with a new value obtained using a logical PVR value,
> * don't modify the performance monitor fields.
> */
> if (old.num_pmcs && !s->num_pmcs) {
> t->num_pmcs = old.num_pmcs;
> t->pmc_type = old.pmc_type;
>
> I realise that having that requirement in the code is serious foot gun
> material on our part, but c'est la vie.
>
> The reason we do that is there's no "compat mode" for the PMU. So if you
> boot on a Power9, and then the logical PVR says "actually pretend you're
> on a Power8", we flip most of the cpu_spec to have the Power8 values,
> but *not* the PMU fields. That way the Power9 PMU code will still detect
> that it's on a Power9 and work correctly.
>
> Possibly now that oprofile is more or less dead we can rip all that crap
> out, and have perf just look at the PVR directly.
Thanks a lot for explaining, that's interesting. I thought it might just have
been an accidental omission in the architected entries but I should've dug
deeper.
>
> cheers
