linux-kernel - Re: [PATCH 1/3] arm64: topology: Improve AMU-based frequency calculation

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <aRIbBBBl8mxd2rtC@arm.com>
Date: Mon, 10 Nov 2025 18:04:04 +0100
From: Beata Michalska <beata.michalska@....com>
To: Bowen Yu <yubowen8@...wei.com>
Cc: linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
	catalin.marinas@....com, will@...nel.org, ptsm@...ux.microsoft.com,
	linuxarm@...wei.com, jonathan.cameron@...wei.com,
	zhanjie9@...ilicon.com, prime.zeng@...ilicon.com,
	wanghuiqiang@...wei.com, xuwei5@...wei.com, zhenglifeng1@...wei.com,
	zhangpengjie2@...wei.com
Subject: Re: [PATCH 1/3] arm64: topology: Improve AMU-based frequency
 calculation

On Tue, Nov 04, 2025 at 03:55:42PM +0800, Bowen Yu wrote:
> The current approach of reverse-calculating CPU frequency from capacity
> values introduces quantization errors due to intermediate scaling of
> arch_scale_freq_capacity, which results in the calculated frequency having
> only 1/1024 resolution.
> 
> This patch:
> 1. Directly computes frequency using AMU counters in amu_scale_freq_tick():
> freq = (core_cycles_delta * timer_freq) / (const_cycles_delta * 1000)
>  - core_cycles_delta: Measured CPU cycles
>  - timer_freq: Architectural timer frequency
>  - const_cycles_delta: Reference cycles from fixed-frequency timer
> 2. Returns pre-computed avgfreq in arch_freq_get_on_cpu()
> 
> examples:
> Before change
> [root@...alhost ~]# cat /sys/devices/system/cpu/cpufreq/policy*/cpuinfo_avg_freq
> 2297851
> 2297851
> 2295312
> 2297851
> 2297851
> 2295312
> 2297851
> 2295312
> 2297851
> 2297851
> 2297851
> 2295312
> 2295312
> 2297851
> 2297851
> 2297851
> 2297851
> 2300390
> 2297851
> 2297851
> 2297851
> 
> After change
> [root@...alhost ~]# cat /sys/devices/system/cpu/cpufreq/policy*/cpuinfo_avg_freq
> 2299177
> 2298117
> 2299188
> 2297330
> 2296530
> 2298817
> 2298434
> 2298986
> 2298596
> 2299395
> 2299560
> 2298446
> 2299108
> 2299294
> 2298707
> 2298453
> 2298632
> 2299218
> 2297962
Based on your numbers the shift is on average ~0.055–0.057%.
I'm not entirely convinced it is worth it, especially that this is an average
frequency. What is the use case here if < 0,2% makes a difference ?

---
BR
Beata
> 
> Signed-off-by: Bowen Yu <yubowen8@...wei.com>
> ---
>  arch/arm64/kernel/topology.c | 20 +++++++++++---------
>  1 file changed, 11 insertions(+), 9 deletions(-)
> 
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 5d07ee85bdae..c0dbc27289ea 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -20,6 +20,7 @@
>  #include <linux/percpu.h>
>  #include <linux/sched/isolation.h>
>  #include <linux/xarray.h>
> +#include <linux/units.h>
>  
>  #include <asm/cpu.h>
>  #include <asm/cputype.h>
> @@ -144,6 +145,8 @@ int __init parse_acpi_topology(void)
>   */
>  static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale) =  1UL << (2 * SCHED_CAPACITY_SHIFT);
>  static cpumask_var_t amu_fie_cpus;
> +static DEFINE_PER_CPU(unsigned long, core_delta);
> +static DEFINE_PER_CPU(unsigned long, const_delta);
>  
>  struct amu_cntr_sample {
>  	u64		arch_const_cycles_prev;
> @@ -246,6 +249,7 @@ static void amu_scale_freq_tick(void)
>  	 * arch_max_freq_scale and the use of SCHED_CAPACITY_SHIFT.
>  	 */
>  	scale = core_cnt - prev_core_cnt;
> +	this_cpu_write(core_delta, scale);
>  	scale *= this_cpu_read(arch_max_freq_scale);
>  	scale = div64_u64(scale >> SCHED_CAPACITY_SHIFT,
>  			  const_cnt - prev_const_cnt);
> @@ -253,6 +257,7 @@ static void amu_scale_freq_tick(void)
>  	scale = min_t(unsigned long, scale, SCHED_CAPACITY_SCALE);
>  	this_cpu_write(arch_freq_scale, (unsigned long)scale);
>  
> +	this_cpu_write(const_delta, const_cnt - prev_const_cnt);
>  	amu_sample->last_scale_update = jiffies;
>  }
>  
> @@ -288,7 +293,7 @@ int arch_freq_get_on_cpu(int cpu)
>  	unsigned int start_cpu = cpu;
>  	unsigned long last_update;
>  	unsigned int freq = 0;
> -	u64 scale;
> +	u64 delta_core_kHz;
>  
>  	if (!amu_fie_cpu_supported(cpu) || !arch_scale_freq_ref(cpu))
>  		return -EOPNOTSUPP;
> @@ -340,14 +345,11 @@ int arch_freq_get_on_cpu(int cpu)
>  			break;
>  		}
>  	}
> -	/*
> -	 * Reversed computation to the one used to determine
> -	 * the arch_freq_scale value
> -	 * (see amu_scale_freq_tick for details)
> -	 */
> -	scale = arch_scale_freq_capacity(cpu);
> -	freq = scale * arch_scale_freq_ref(cpu);
> -	freq >>= SCHED_CAPACITY_SHIFT;
> +
> +	if (check_mul_overflow(per_cpu(core_delta, cpu), arch_timer_get_cntfrq(), &delta_core_kHz))
> +		return -EINVAL;
> +
> +	freq = div_u64(delta_core_kHz, per_cpu(const_delta, cpu) * HZ_PER_KHZ);
>  	return freq;
>  }
>  
> -- 
> 2.33.0
>