Re: [PATCH v3 6/7] arm64: use activity monitors for frequency invariance

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

Hi Ionela,

Overall I think this approach is much better, and apart from a few nits below
this is looking pretty good.

On 2/11/20 6:45 PM, Ionela Voinescu wrote:
> @@ -120,4 +121,188 @@ int __init parse_acpi_topology(void)
>  }
>  #endif
>  
> +#ifdef CONFIG_ARM64_AMU_EXTN
>  
> +#undef pr_fmt
> +#define pr_fmt(fmt) "AMU: " fmt
> +
> +static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale);
> +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;
> +
> +/* Obtain max frequency (in KHz) as reported by hardware */
> +__weak unsigned int cpu_get_max_freq(unsigned int cpu)
> +{
> +	return 0;
> +}
> +
> +#ifdef CONFIG_CPU_FREQ
> +/* Replace max frequency getter with cpufreq based function */
> +#define cpu_get_max_freq cpufreq_get_hw_max_freq
> +#endif
> +
> +/* Initialize counter reference per-cpu variables for the current CPU */
> +void init_cpu_freq_invariance_counters(void)
> +{
> +	this_cpu_write(arch_core_cycles_prev,
> +		       read_sysreg_s(SYS_AMEVCNTR0_CORE_EL0));
> +	this_cpu_write(arch_const_cycles_prev,
> +		       read_sysreg_s(SYS_AMEVCNTR0_CONST_EL0));
> +}
> +
> +static int validate_cpu_freq_invariance_counters(int cpu)
> +{
> +	u64 max_freq_hz, ratio;
> +
> +	if (!cpu_has_amu_feat(cpu)) {
> +		pr_debug("CPU%d: counters are not supported.\n", cpu);
> +		return -EINVAL;
> +	}
> +
> +	if (unlikely(!per_cpu(arch_const_cycles_prev, cpu) ||
> +		     !per_cpu(arch_core_cycles_prev, cpu))) {
> +		pr_debug("CPU%d: cycle counters are not enabled.\n", cpu);
> +		return -EINVAL;
> +	}
> +
> +	/* Convert maximum frequency from KHz to Hz and validate */
> +	max_freq_hz = cpu_get_max_freq(cpu) * 1000;
> +	if (unlikely(!max_freq_hz)) {
> +		pr_debug("CPU%d: invalid maximum frequency.\n", cpu);
> +		return -EINVAL;
> +	}
> +
> +	/*
> +	 * Pre-compute the fixed ratio between the frequency of the constant
> +	 * counter and the maximum frequency of the CPU.
> +	 *
> +	 *			      const_freq
> +	 * arch_max_freq_scale =   ---------------- * SCHED_CAPACITY_SCALE²
> +	 *			   cpuinfo_max_freq
> +	 *
> +	 * 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 arch timer frequencies (up to the KHz range which should be
                                            ^^^^^
<pedantic hat on>: s/up to/down to/

> +	 * unlikely).
> +	 */
> +	ratio = (u64)arch_timer_get_rate() << (2 * SCHED_CAPACITY_SHIFT);
> +	ratio = div64_u64(ratio, max_freq_hz);
> +	if (!ratio) {
> +		pr_err("System timer frequency too low.\n");

Should that be a WARN_ONCE() instead? If the arch timer freq is too low,
we'll end up spamming this message, since we go through this for all CPUs.


> +		return -EINVAL;
> +	}
> +
> +	per_cpu(arch_max_freq_scale, cpu) = (unsigned long)ratio;
> +

It occurred to me that this isn't strictly speaking a per-CPU information as
it only depends on the max possible frequency. Not really worth bothering
about though, I think.

> +	return 0;
> +}
> +
> +static inline int
> +enable_policy_freq_counters(int cpu, cpumask_var_t valid_cpus)
> +{
> +	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
> +
> +	if (!policy) {
> +		pr_debug("CPU%d: No cpufreq policy found.\n", cpu);
> +		return false;
> +	}
> +
> +	if (cpumask_subset(policy->related_cpus, valid_cpus)) {
> +		cpumask_or(amu_fie_cpus, policy->related_cpus,
> +			   amu_fie_cpus);
> +		pr_info("CPUs[%*pbl]: counters will be used for FIE.",
> +			cpumask_pr_args(amu_fie_cpus));

Could we have a single print of all CPUs in one go? AIUI this will generate a
line per cpufreq policy. Maybe just something at the tail of init_amu_fie():

if (!cpumask_empty(amu_fie_cpus))
	pr_info(<blah>);

> +	}
> +
> +	cpufreq_cpu_put(policy);
> +
> +	return true;
> +}
> +
> +static int __init init_amu_fie(void)
> +{
> +	cpumask_var_t valid_cpus;
> +	bool have_policy = false;
> +	int cpu;
> +
> +	if (!zalloc_cpumask_var(&valid_cpus, GFP_KERNEL) ||
> +	    !zalloc_cpumask_var(&amu_fie_cpus, GFP_KERNEL))
> +		return -ENOMEM;
> +
> +	for_each_possible_cpu(cpu) {
> +		if (validate_cpu_freq_invariance_counters(cpu))
> +			continue;
> +		cpumask_set_cpu(cpu, valid_cpus);
> +		have_policy = enable_policy_freq_counters(cpu, valid_cpus) ||
> +			      have_policy;

What about:
		have_policy |= enable_policy_freq_counters(cpu, valid_cpus);

> +	}
> +
> +	if (!have_policy) {
> +		/*
> +		 * If we are not restricted by cpufreq policies, we only enable
> +		 * the use of the AMU feature for FIE if all CPUs support AMU.
> +		 * Otherwise, enable_policy_freq_counters has already enabled
> +		 * policy cpus.
> +		 */
> +		if (cpumask_equal(valid_cpus, cpu_possible_mask)) {

Mmm so I'm thinking what we want here is the cpu_present_mask rather than
the possible one. This is very corner-casy, but I think that if we fail to
boot a secondary, we'll have it possible but not present.

While at it you could make the loop only target present CPUs, but I think the
one bit that matters is this check right here (!present should fail at
validate_cpu_freq_invariance_counters()).

> +			cpumask_or(amu_fie_cpus, amu_fie_cpus, valid_cpus);
> +			pr_info("CPUs[%*pbl]: counters will be used for FIE.",
> +				cpumask_pr_args(amu_fie_cpus));
> +		}
> +	}
> +
> +	free_cpumask_var(valid_cpus);
> +
> +	return 0;
> +}
> +late_initcall_sync(init_amu_fie);
> +
> +bool topology_cpu_freq_counters(struct cpumask *cpus)
> +{
> +	return cpumask_available(amu_fie_cpus) &&
> +	       cpumask_subset(cpus, amu_fie_cpus);
> +}
> +
> +void topology_scale_freq_tick(void)
> +{
> +	u64 prev_core_cnt, prev_const_cnt;
> +	u64 core_cnt, const_cnt, scale;
> +	int cpu = smp_processor_id();
> +
> +	if (!cpumask_available(amu_fie_cpus) ||
> +	    !cpumask_test_cpu(cpu, amu_fie_cpus))
> +		return;

It might be a good idea to have a static key to gate our entry into this
function - that way we can lessen our impact on older platforms (without AMUs)
running a recent kernel with CONFIG_ARM64_AMU_EXTN=y.

x86 does just that, if you look at their arch_scale_freq_tick()
implementation. FWIW I don't think we should bother with playing with the
key counter to count AMU-enabled CPUs, just enable it at startup if we have
> 1 such CPU and let the cpumask drive the rest.

In your check here, the static key check could replace the cpumask_available()
check. The static key could also be used for topology_cpu_freq_counters().

> +
> +	const_cnt = read_sysreg_s(SYS_AMEVCNTR0_CONST_EL0);
> +	core_cnt = read_sysreg_s(SYS_AMEVCNTR0_CORE_EL0);
> +	prev_const_cnt = this_cpu_read(arch_const_cycles_prev);
> +	prev_core_cnt = this_cpu_read(arch_core_cycles_prev);
> +
> +	if (unlikely(core_cnt <= prev_core_cnt ||
> +		     const_cnt <= prev_const_cnt))
> +		goto store_and_exit;
> +
> +	/*
> +	 *	    /\core    arch_max_freq_scale
> +	 * scale =  ------- * --------------------
> +	 *	    /\const   SCHED_CAPACITY_SCALE
> +	 *
> +	 * We shift by SCHED_CAPACITY_SHIFT (divide by SCHED_CAPACITY_SCALE)
> +	 * in order to compensate for the SCHED_CAPACITY_SCALE² factor in
> +	 * arch_max_freq_scale (used to ensure its resolution) while keeping
> +	 * the scale value in the 0-SCHED_CAPACITY_SCALE capacity range.
> +	 */

A simple "See validate_cpu_freq_invariance_counters() for details on the
scale factor" would suffice wrt the shifting details.

> +	scale = core_cnt - prev_core_cnt;
> +	scale *= this_cpu_read(arch_max_freq_scale);
> +	scale = div64_u64(scale >> SCHED_CAPACITY_SHIFT,
> +			  const_cnt - prev_const_cnt);
> +
> +	scale = min_t(unsigned long, scale, SCHED_CAPACITY_SCALE);
> +	this_cpu_write(freq_scale, (unsigned long)scale);
> +
> +store_and_exit:
> +	this_cpu_write(arch_core_cycles_prev, core_cnt);
> +	this_cpu_write(arch_const_cycles_prev, const_cnt);
> +}
> +#endif /* CONFIG_ARM64_AMU_EXTN */
> diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
> index 1eb81f113786..1ab2b7503d63 100644
> --- a/drivers/base/arch_topology.c
> +++ b/drivers/base/arch_topology.c
> @@ -29,6 +29,14 @@ void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq,
>  	unsigned long scale;
>  	int i;
>  
> +	/*
> +	 * If the use of counters for FIE is enabled, just return as we don't
> +	 * want to update the scale factor with information from CPUFREQ.
> +	 * Instead the scale factor will be updated from arch_scale_freq_tick.
> +	 */
> +	if (arch_cpu_freq_counters(cpus))
> +		return;
> +
>  	scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
>  
>  	for_each_cpu(i, cpus)
> diff --git a/include/linux/topology.h b/include/linux/topology.h
> index eb2fe6edd73c..397aad6ae163 100644
> --- a/include/linux/topology.h
> +++ b/include/linux/topology.h
> @@ -227,5 +227,12 @@ static inline const struct cpumask *cpu_cpu_mask(int cpu)
>  	return cpumask_of_node(cpu_to_node(cpu));
>  }
>  
> +#ifndef arch_cpu_freq_counters
> +static __always_inline
> +bool arch_cpu_freq_counters(struct cpumask *cpus)
> +{
> +	return false;
> +}
> +#endif
>  
>  #endif /* _LINUX_TOPOLOGY_H */
>