Re: [PATCH v1 1/3] arm64: topology: Add arch_freq_get_on_cpu() support

[lihuisong (C)]

Hi Vanshidhar,

在 2024/3/1 0:25, Vanshidhar Konda 写道:
> AMU counters are used by the Frequency Invariance Engine (FIE) to
> estimate the CPU utilization during each tick. The delta of the AMU
> counters between two ticks can also be used to estimate the average CPU
> frequency of each core over the tick duration. Measure the AMU counters
> during tick, compute the delta and store it. When the frequency of the
> core is queried, use the stored delta to determine the frequency.
>
> arch_freq_get_on_cpu() is used on x86 systems to estimate the frequency
> of each CPU. It can be wired up on arm64 for the same functionality.
>
> Signed-off-by: Vanshidhar Konda <vanshikonda@xxxxxxxxxxxxxxxxxxxxxx>
> ---
>   arch/arm64/kernel/topology.c | 114 +++++++++++++++++++++++++++++------
>   1 file changed, 96 insertions(+), 18 deletions(-)
>
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 1a2c72f3e7f8..db8d14525cf4 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -17,6 +17,8 @@
>   #include <linux/cpufreq.h>
>   #include <linux/init.h>
>   #include <linux/percpu.h>
> +#include <linux/sched/isolation.h>
> +#include <linux/seqlock_types.h>
>   
>   #include <asm/cpu.h>
>   #include <asm/cputype.h>
> @@ -82,20 +84,54 @@ int __init parse_acpi_topology(void)
>   #undef pr_fmt
>   #define pr_fmt(fmt) "AMU: " fmt
>   
> +struct amu_counters {
> +	seqcount_t	seq;
> +	unsigned long	last_update;
> +	u64		core_cnt;
> +	u64		const_cnt;
> +	u64		delta_core_cnt;
> +	u64		delta_const_cnt;
> +};
> +
>   /*
>    * 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);
> -static DEFINE_PER_CPU(u64, arch_const_cycles_prev);
> -static DEFINE_PER_CPU(u64, arch_core_cycles_prev);
> +static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale) =
> +	1UL << (2 * SCHED_CAPACITY_SHIFT);
> +static DEFINE_PER_CPU_SHARED_ALIGNED(struct amu_counters, cpu_samples) = {
> +	.seq = SEQCNT_ZERO(cpu_samples.seq)
> +};
>   static cpumask_var_t amu_fie_cpus;
>   
>   void update_freq_counters_refs(void)
>   {
> -	this_cpu_write(arch_core_cycles_prev, read_corecnt());
> -	this_cpu_write(arch_const_cycles_prev, read_constcnt());
> +	struct amu_counters *cpu_sample = this_cpu_ptr(&cpu_samples);
> +	u64 core_cnt, const_cnt, delta_core_cnt, delta_const_cnt;
> +
> +	const_cnt = read_constcnt();
> +	core_cnt = read_corecnt();
> +
> +	if (unlikely(core_cnt < cpu_sample->core_cnt) ||
> +	    unlikely(const_cnt < cpu_sample->const_cnt)) {

The two counter register might be wrap around. So here is not fully 
true, right?

If we don't consider this case, below warning should be removed.

> +		WARN(1, "AMU counter values should be monotonic.\n");
> +		cpu_sample->delta_const_cnt = 0;
> +		cpu_sample->delta_core_cnt = 0;
> +		return;
> +	}
> +
> +	delta_core_cnt = core_cnt - cpu_sample->core_cnt;
> +	delta_const_cnt = const_cnt - cpu_sample->const_cnt;
> +
> +	cpu_sample->core_cnt = core_cnt;
> +	cpu_sample->const_cnt = const_cnt;
> +
> +	raw_write_seqcount_begin(&cpu_sample->seq);
> +	cpu_sample->last_update = jiffies;
> +	cpu_sample->delta_const_cnt = delta_const_cnt;
> +	cpu_sample->delta_core_cnt = delta_core_cnt;
> +	raw_write_seqcount_end(&cpu_sample->seq);
>   }
<...>