| 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
| ||
|
Message-ID: <CAJZ5v0hqzsoANFVPJXqYM6ZL4Ad6rDh=+G0qKmnmKLtyvADwmg@mail.gmail.com>
Date: Fri, 4 Feb 2022 19:07:57 +0100
From: "Rafael J. Wysocki" <rafael@...nel.org>
To: Pierre Gondois <Pierre.Gondois@....com>
Cc: Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
Ionela Voinescu <Ionela.Voinescu@....com>,
Lukasz Luba <lukasz.luba@....com>,
Morten Rasmussen <morten.rasmussen@....com>,
"Rafael J. Wysocki" <rafael@...nel.org>,
Viresh Kumar <viresh.kumar@...aro.org>,
Linux PM <linux-pm@...r.kernel.org>
Subject: Re: [PATCH v2] cpufreq: CPPC: Fix performance/frequency conversion
On Tue, Jan 11, 2022 at 4:54 PM Pierre Gondois <Pierre.Gondois@....com> wrote:
>
> CPUfreq governors request CPU frequencies using information
> on current CPU usage. The CPPC driver converts them to
> performance requests. Frequency targets are computed as:
> target_freq = (util / cpu_capacity) * max_freq
> target_freq is then clamped between [policy->min, policy->max].
>
> The CPPC driver converts performance values to frequencies
> (and vice-versa) using cppc_cpufreq_perf_to_khz() and
> cppc_cpufreq_khz_to_perf(). These functions both use two different
> factors depending on the range of the input value. For
> cppc_cpufreq_khz_to_perf():
> - (NOMINAL_PERF / NOMINAL_FREQ) or
> - (LOWEST_PERF / LOWEST_FREQ)
> and for cppc_cpufreq_perf_to_khz():
> - (NOMINAL_FREQ / NOMINAL_PERF) or
> - ((NOMINAL_PERF - LOWEST_FREQ) / (NOMINAL_PERF - LOWEST_PERF))
>
> This means:
> 1- the functions are not inverse for some values:
> (perf_to_khz(khz_to_perf(x)) != x)
> 2- cppc_cpufreq_perf_to_khz(LOWEST_PERF) can sometimes give
> a different value from LOWEST_FREQ due to integer approximation
> 3- it is implied that performance and frequency are proportional
> (NOMINAL_FREQ / NOMINAL_PERF) == (LOWEST_PERF / LOWEST_FREQ)
>
> This patch changes the conversion functions to an affine function.
> This fixes the 3 points above.
>
> Suggested-by: Lukasz Luba <lukasz.luba@....com>
> Suggested-by: Morten Rasmussen <morten.rasmussen@....com>
> Signed-off-by: Pierre Gondois <Pierre.Gondois@....com>
> ---
> drivers/cpufreq/cppc_cpufreq.c | 43 +++++++++++++++++-----------------
> 1 file changed, 21 insertions(+), 22 deletions(-)
>
> diff --git a/drivers/cpufreq/cppc_cpufreq.c b/drivers/cpufreq/cppc_cpufreq.c
> index db17196266e4..5024d9af2e6e 100644
> --- a/drivers/cpufreq/cppc_cpufreq.c
> +++ b/drivers/cpufreq/cppc_cpufreq.c
> @@ -303,52 +303,48 @@ static u64 cppc_get_dmi_max_khz(void)
>
> /*
> * If CPPC lowest_freq and nominal_freq registers are exposed then we can
> - * use them to convert perf to freq and vice versa
> - *
> - * If the perf/freq point lies between Nominal and Lowest, we can treat
> - * (Low perf, Low freq) and (Nom Perf, Nom freq) as 2D co-ordinates of a line
> - * and extrapolate the rest
> - * For perf/freq > Nominal, we use the ratio perf:freq at Nominal for conversion
> + * use them to convert perf to freq and vice versa. The conversion is
> + * extrapolated as an affine function passing by the 2 points:
> + * - (Low perf, Low freq)
> + * - (Nominal perf, Nominal perf)
> */
> static unsigned int cppc_cpufreq_perf_to_khz(struct cppc_cpudata *cpu_data,
> unsigned int perf)
> {
> struct cppc_perf_caps *caps = &cpu_data->perf_caps;
> + s64 retval, offset = 0;
> static u64 max_khz;
> u64 mul, div;
>
> if (caps->lowest_freq && caps->nominal_freq) {
> - if (perf >= caps->nominal_perf) {
> - mul = caps->nominal_freq;
> - div = caps->nominal_perf;
> - } else {
> - mul = caps->nominal_freq - caps->lowest_freq;
> - div = caps->nominal_perf - caps->lowest_perf;
> - }
> + mul = caps->nominal_freq - caps->lowest_freq;
> + div = caps->nominal_perf - caps->lowest_perf;
> + offset = caps->nominal_freq - (u64)caps->nominal_perf * mul / div;
Since mult is a u64, the other operands need not be cast to u64
explicitly AFAICS.
Moreover, it might be better to use div64_u64() instead of the plain
integer division.
> } else {
> if (!max_khz)
> max_khz = cppc_get_dmi_max_khz();
> mul = max_khz;
> div = caps->highest_perf;
> }
> - return (u64)perf * mul / div;
> +
> + retval = offset + (u64)perf * mul / div;
> + if (retval >= 0)
> + return retval;
> + return 0;
> }
>
> static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
> unsigned int freq)
> {
> struct cppc_perf_caps *caps = &cpu_data->perf_caps;
> + s64 retval, offset = 0;
> static u64 max_khz;
> u64 mul, div;
>
> if (caps->lowest_freq && caps->nominal_freq) {
> - if (freq >= caps->nominal_freq) {
> - mul = caps->nominal_perf;
> - div = caps->nominal_freq;
> - } else {
> - mul = caps->lowest_perf;
> - div = caps->lowest_freq;
> - }
> + mul = caps->nominal_perf - caps->lowest_perf;
> + div = caps->nominal_freq - caps->lowest_freq;
> + offset = caps->nominal_perf - (u64)caps->nominal_freq * mul / div;
> } else {
> if (!max_khz)
> max_khz = cppc_get_dmi_max_khz();
> @@ -356,7 +352,10 @@ static unsigned int cppc_cpufreq_khz_to_perf(struct cppc_cpudata *cpu_data,
> div = max_khz;
> }
>
> - return (u64)freq * mul / div;
> + retval = offset + (u64)freq * mul / div;
> + if (retval >= 0)
> + return retval;
> + return 0;
> }
>
> static int cppc_cpufreq_set_target(struct cpufreq_policy *policy,
> --
> 2.25.1
>
Powered by blists - more mailing lists