Re: [PATCH] power/amd_pstate: fix frequency matching for continuous scaling

[Thomas Monjalon <thomas@monjalon.net>]

28/03/2026 20:34, Stephen Hemminger:
> The power_init_for_setting_freq() function fails on systems using the
> amd-pstate-epp driver because the current CPU frequency read from
> scaling_setspeed does not exactly match any of the synthesized
> frequency buckets. Unlike acpi_cpufreq which provides a discrete list
> of frequencies, amd-pstate operates with continuously variable
> frequencies, so an exact match will rarely succeed.
> 
> For example, on a Ryzen 9 7945HX the sysfs file reports 2797172
> which rounds to 2797000, but this value does not appear in the
> generated frequency table.
> 
> Replace the exact match lookup with a nearest-frequency search.
> 
[...]
> -	freq = strtoul(buf, NULL, POWER_CONVERT_TO_DECIMAL);
> +	errno = 0;
> +	freq = strtoul(buf, &endptr, POWER_CONVERT_TO_DECIMAL);
> +	if (errno != 0 || endptr == buf || freq == 0) {
> +		POWER_LOG(ERR, "Failed to parse frequency '%s' for lcore %u",
> +				buf, pi->lcore_id);
> +		goto err;
> +	}
>  
>  	/* convert the frequency to nearest 1000 value
>  	 * Ex: if freq=1396789 then freq_conv=1397000
>  	 * Ex: if freq=800030 then freq_conv=800000
>  	 */
> -	unsigned int freq_conv = 0;
> -	freq_conv = (freq + FREQ_ROUNDING_DELTA)
> -				/ ROUND_FREQ_TO_N_1000;
> +	freq_conv = (freq + FREQ_ROUNDING_DELTA) / ROUND_FREQ_TO_N_1000;
>  	freq_conv = freq_conv * ROUND_FREQ_TO_N_1000;
>  
> -	for (i = 0; i < pi->nb_freqs; i++) {
> -		if (freq_conv == pi->freqs[i]) {
> -			pi->curr_idx = i;
> -			pi->f = f;
> -			return 0;
> +	/* Find the nearest frequency in the table.
> +	 * With amd-pstate the CPU runs at continuously variable
> +	 * frequencies so the current frequency will not exactly
> +	 * match one of the synthesized frequency buckets.
> +	 */
> +	best_idx = 0;
> +	best_diff = abs_diff(freq_conv, pi->freqs[0]);
> +
> +	for (i = 1; i < pi->nb_freqs; i++) {
> +		diff = abs_diff(freq_conv, pi->freqs[i]);
> +		if (diff < best_diff) {
> +			best_diff = diff;
> +			best_idx = i;
>  		}
>  	}

GPT found this problem:

power_init_for_setting_freq() now assigns pi->curr_idx = best_idx
after finding the nearest synthesized frequency bucket.
However, set_freq_internal() skips the sysfs write
whenever idx == pi->curr_idx.

This means that if the current scaling_setspeed value is merely close
to a bucket but not equal to it, a later request to set that bucket
will return success without actually writing the requested frequency.
This can happen during init too: power_amd_pstate_cpufreq_init()
calls freq_max() after initialization, but if the current frequency
is nearest to the max bucket, freq_max() will be skipped even when
the actual sysfs value is not the synthesized max.
The nearest-bucket match should not be treated as an exact programmed
frequency, or the next explicit set to that bucket should be forced.