Re: [PATCH v2 6/6] arm64: use activity monitors for frequency invariance

[Valentin Schneider <valentin.schneider@arm.com>]

Only commenting on the bits that should be there regardless of using the
workqueues or not;

On 18/12/2019 18:26, Ionela Voinescu wrote:
> +static void cpu_amu_fie_init_workfn(struct work_struct *work)
> +{
> +	u64 core_cnt, const_cnt, ratio;
> +	struct cpu_amu_work *amu_work;
> +	int cpu = smp_processor_id();
> +
> +	if (!cpu_has_amu_feat()) {
> +		pr_debug("CPU%d: counters are not supported.\n", cpu);
> +		return;
> +	}
> +
> +	core_cnt = read_sysreg_s(SYS_AMEVCNTR0_CORE_EL0);
> +	const_cnt = read_sysreg_s(SYS_AMEVCNTR0_CONST_EL0);
> +
> +	if (unlikely(!core_cnt || !const_cnt)) {
> +		pr_err("CPU%d: cycle counters are not enabled.\n", cpu);
> +		return;
> +	}
> +
> +	amu_work = container_of(work, struct cpu_amu_work, cpu_work);
> +	if (unlikely(!(amu_work->cpuinfo_max_freq))) {
> +		pr_err("CPU%d: invalid maximum frequency.\n", cpu);
> +		return;
> +	}
> +
> +	/*
> +	 * Pre-compute the fixed ratio between the frequency of the
> +	 * constant counter and the maximum frequency of the CPU (hz).

I can't resist: s/hz/Hz/

> +	 */
> +	ratio = (u64)arch_timer_get_rate() << (2 * SCHED_CAPACITY_SHIFT);
> +	ratio = div64_u64(ratio, amu_work->cpuinfo_max_freq * 1000);

Nit: we're missing a comment somewhere that the unit of this is in kHz
(which explains the * 1000).

> +	this_cpu_write(arch_max_freq_scale, (unsigned long)ratio);
> +

Okay so what we get in the tick is:

  /\ core
  --------
  /\ const

And we want that to be SCHED_CAPACITY_SCALE when running at max freq. IOW we
want to turn

  max_freq
  ----------
  const_freq

into SCHED_CAPACITY_SCALE, so we can just multiply that by:

  const_freq
  ---------- * SCHED_CAPACITY_SCALE
  max_freq

But what the ratio you are storing here is 

                          const_freq
  arch_max_freq_scale =   ---------- * SCHED_CAPACITY_SCALE²
                           max_freq

(because x << 2 * SCHED_CAPACITY_SHIFT == x << 20)


In topology_freq_scale_tick() you end up doing

  /\ core   arch_max_freq_scale
  ------- * --------------------
  /\ const  SCHED_CAPACITY_SCALE

which gives us what we want (SCHED_CAPACITY_SCALE at max freq).


Now, the reason why we multiply our ratio by the square of
SCHED_CAPACITY_SCALE was not obvious to me, but you pointed me out that the
frequency of the arch timer can be *really* low compared to the max CPU freq.

For instance on h960:

  [    0.000000] arch_timer: cp15 timer(s) running at 1.92MHz (phys)

  $ root@valsch-h960:~# cat /sys/devices/system/cpu/cpufreq/policy4/cpuinfo_max_freq 
  2362000

So our ratio would be

  1'920'000 * 1024
  ----------------
    2'362'000'000

Which is ~0.83, so that becomes simply 0...


I had a brief look at the Arm ARM, for the arch timer it says it is
"typically in the range 1-50MHz", but then also gives an example with 20KHz
in a low-power mode.

If we take say 5GHz max CPU frequency, our lower bound for the arch timer
(with that SCHED_CAPACITY_SCALE² trick) is about ~4.768KHz. It's not *too*
far from that 20KHz, but I'm not sure we would actually be executing stuff
in that low-power mode.

Long story short, we're probably fine, but it would nice to shove some of
the above into comments (especially the SCHED_CAPACITY_SCALE² trick)