Re: [PATCH v5 7/7] arm64/amu: Use capacity_ref_freq to set AMU ratio

[Vincent Guittot]

On Tue, 7 Nov 2023 at 13:01, Will Deacon <will@xxxxxxxxxx> wrote:
>
> On Tue, Nov 07, 2023 at 12:18:20PM +0100, Vincent Guittot wrote:
> > On Tue, 7 Nov 2023 at 11:38, Will Deacon <will@xxxxxxxxxx> wrote:
> > >
> > > On Sat, Nov 04, 2023 at 11:59:07AM +0100, Vincent Guittot wrote:
> > > > Use the new capacity_ref_freq to set the ratio that is used by AMU for
> > > > computing the arch_scale_freq_capacity().
> > > > This helps to keep everything aligned using the same reference for
> > > > computing CPUs capacity.
> > > >
> > > > The default value of the ratio (stored in per_cpu(arch_max_freq_scale))
> > > > ensures that arch_scale_freq_capacity() returns max capacity until it is
> > > > set to its correct value with the cpu capacity and capacity_ref_freq.
> > > >
> > > > Signed-off-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > > > ---
> > > >  arch/arm64/kernel/topology.c  | 26 ++++++++++++++------------
> > > >  drivers/base/arch_topology.c  | 12 +++++++++++-
> > > >  include/linux/arch_topology.h |  1 +
> > > >  3 files changed, 26 insertions(+), 13 deletions(-)
> > > >
> > > > diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> > > > index 817d788cd866..615c1a20129f 100644
> > > > --- a/arch/arm64/kernel/topology.c
> > > > +++ b/arch/arm64/kernel/topology.c
> > > > @@ -82,7 +82,12 @@ int __init parse_acpi_topology(void)
> > > >  #undef pr_fmt
> > > >  #define pr_fmt(fmt) "AMU: " fmt
> > > >
> > > > -static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale);
> > > > +/*
> > > > + * Ensure that amu_scale_freq_tick() will return SCHED_CAPACITY_SCALE until
> > > > + * the CPU capacity and its associated frequency have been correctly
> > > > + * initialized.
> > > > + */
> > > > +static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale) =  1UL << (2 * SCHED_CAPACITY_SHIFT);
> > >
> > > This doesn't seem to match the comment? SCHED_CAPACITY_SCALE doesn't have
> > > the '2 *' multiplier.
> >
> > The comment in freq_inv_set_max_ratio() says:
> >
> > * We use a factor of 2 * SCHED_CAPACITY_SHIFT -> SCHED_CAPACITY_SCALE²
> > * in order to ensure a good resolution for arch_max_freq_scale for
> > * very low reference frequencies (down to the KHz range which should
> > * be unlikely).
> >
> > Then there is a  "  * arch_max_freq_scale) >> SCHED_CAPACITY_SHIFT"
> > when computing the scale which brings the result back to
> > SCHED_CAPACITY_SHIFT
>
> Ah, I see; I'd not spotted that amu_scale_freq_tick() is doing some
> arithmetic on the value (it doesn't return anything because it's 'void').
> It's slightly confusing because the comment talks about SCHED_CAPACITY_SCALE
> whereas all the code works on the shift, but I get it now.
>
> > > >  static DEFINE_PER_CPU(u64, arch_const_cycles_prev);
> > > >  static DEFINE_PER_CPU(u64, arch_core_cycles_prev);
> > > >  static cpumask_var_t amu_fie_cpus;
> > > > @@ -112,14 +117,14 @@ static inline bool freq_counters_valid(int cpu)
> > > >       return true;
> > > >  }
> > > >
> > > > -static int freq_inv_set_max_ratio(int cpu, u64 max_rate, u64 ref_rate)
> > > > +void freq_inv_set_max_ratio(int cpu, u64 max_rate)
> > > >  {
> > > > -     u64 ratio;
> > > > +     u64 ratio, ref_rate = arch_timer_get_rate();
> > > >
> > > >       if (unlikely(!max_rate || !ref_rate)) {
> > > > -             pr_debug("CPU%d: invalid maximum or reference frequency.\n",
> > > > +             WARN_ONCE(1, "CPU%d: invalid maximum or reference frequency.\n",
> > > >                        cpu);
> > > > -             return -EINVAL;
> > > > +             return;
> > > >       }
> > > >
> > > >       /*
> > > > @@ -139,12 +144,12 @@ static int freq_inv_set_max_ratio(int cpu, u64 max_rate, u64 ref_rate)
> > > >       ratio = div64_u64(ratio, max_rate);
> > > >       if (!ratio) {
> > > >               WARN_ONCE(1, "Reference frequency too low.\n");
> > > > -             return -EINVAL;
> > > > +             return;
> > > >       }
> > > >
> > > > -     per_cpu(arch_max_freq_scale, cpu) = (unsigned long)ratio;
> > > > +     WRITE_ONCE(per_cpu(arch_max_freq_scale, cpu), (unsigned long)ratio);
> > >
> > > Why is WRITE_ONCE() now needed?
> >
> > the tick can already use it. We want to make sure to use either the
> > old or the new one but not an intermediate value
>
> Isn't that already the case without this patch? In other words, this should
> be a separate change.

Currently the ratio is set in the same notifier callback before
registering the AMU function in topology_scale_freq_tick whereas there
are now done in 2 separate notifier callbacks (still for the same
event) but we don't know the order

>
> Will