Re: [PATCH v3 12/15] xen/x86: implement EPP support for the amd-cppc driver in active mode

[Jan Beulich <jbeulich@xxxxxxxx>]

On 06.03.2025 09:39, Penny Zheng wrote:
> amd-cppc has 2 operation modes: autonomous (active) mode,
> non-autonomous (passive) mode.
> In active mode, platform ignores the requestd done in the Desired
> Performance Target register and takes into account only the values
> set to the minimum, maximum and energy performance preference(EPP)
> registers.
> The EPP is used in the CCLK DPM controller to drive the frequency
> that a core is going to operate during short periods of activity.
> The SOC EPP targets are configured on a scale from 0 to 255 where 0
> represents maximum performance and 255 represents maximum efficiency.

So this is the other way around from "perf" values, where aiui 0xff is
"highest"?

> @@ -537,6 +537,12 @@ choice of `dom0-kernel` is deprecated and not supported 
> by all Dom0 kernels.
>  * `amd-cppc` selects ACPI Collaborative Performance and Power Control (CPPC)
>    on supported AMD hardware to provide finer grained frequency control
>    mechanism. The default is disabled.
> +* `active` is to enable amd-cppc driver in active(autonomous) mode. In this
> +  mode, users could write to energy performance preference register to tell
> +  hardware if they want to bias toward performance or energy efficiency. Then
> +  built-in CPPC power algorithm will calculate the runtime workload and 
> adjust
> +  the realtime cores frequency automatically according to the power supply 
> and

What are "the realtime cores"?

> +  thermal, core voltage and some other hardware conditions.

I think there better would be only one "and" in the enumeration of conditions.

> @@ -261,7 +276,20 @@ static int cf_check amd_cppc_cpufreq_target(struct 
> cpufreq_policy *policy,
>          return res;
>  
>      return amd_cppc_write_request(policy->cpu, 
> data->caps.lowest_nonlinear_perf,
> -                                  des_perf, data->caps.highest_perf);
> +                                  des_perf, data->caps.highest_perf,
> +                                  /* Pre-defined BIOS value for passive mode 
> */
> +                                  per_cpu(epp_init, policy->cpu));
> +}
> +
> +static int read_epp_init(void)
> +{
> +    uint64_t val;
> +
> +    if ( rdmsr_safe(MSR_AMD_CPPC_REQ, val) )
> +        return -EINVAL;

I'm unconvinced of using rdmsr_safe() everywhere (i.e. this also goes for 
earlier
patches). Unless you can give a halfway reasonable scenario under which by the
time we get here there's still a chance that the MSR isn't implemented in the
next lower layer (hardware or another hypervisor, just to explain what's meant,
without me assuming that the driver should come into play in the first place 
when
we run virtualized ourselves).

Furthermore you call this function unconditionally, i.e. if there was a chance
for the MSR read to fail, CPU init would needlessly fail when in passive mode.

> +    this_cpu(epp_init) = (val >> 24) & 0xFF;

Please can you #define a suitable mask constant in msr-index.h, such that you 
can
use MASK_EXTR() here?

> @@ -411,12 +441,78 @@ static int cf_check amd_cppc_cpufreq_cpu_init(struct 
> cpufreq_policy *policy)
>  
>      amd_cppc_boost_init(policy, data);
>  
> +    return 0;
> +}
> +
> +static int cf_check amd_cppc_cpufreq_cpu_init(struct cpufreq_policy *policy)
> +{
> +    int ret;
> +
> +    ret = amd_cppc_cpufreq_init_perf(policy);
> +    if ( ret )
> +        return ret;
> +
>      amd_cppc_verbose("CPU %u initialized with amd-cppc passive mode\n",
>                       policy->cpu);
>  
>      return 0;
>  }
>  
> +static int cf_check amd_cppc_epp_cpu_init(struct cpufreq_policy *policy)
> +{
> +    int ret;
> +
> +    ret = amd_cppc_cpufreq_init_perf(policy);
> +    if ( ret )
> +        return ret;
> +
> +    policy->policy = cpufreq_parse_policy(policy->governor);
> +
> +    amd_cppc_verbose("CPU %u initialized with amd-cppc active mode\n", 
> policy->cpu);
> +
> +    return 0;
> +}
> +
> +static int amd_cppc_epp_update_limit(const struct cpufreq_policy *policy)
> +{
> +    const struct amd_cppc_drv_data *data = per_cpu(amd_cppc_drv_data,
> +                                                    policy->cpu);

Nit: Indentation is off by one here.

> +    uint8_t max_perf, min_perf, epp;
> +
> +    /* Initial min/max values for CPPC Performance Controls Register */
> +    /*
> +     * Continuous CPPC performance scale in active mode is [lowest_perf,
> +     * highest_perf]
> +     */
> +    max_perf = data->caps.highest_perf;
> +    min_perf = data->caps.lowest_perf;
> +
> +    epp = per_cpu(epp_init, policy->cpu);
> +    if ( policy->policy == CPUFREQ_POLICY_PERFORMANCE )

This may want to be switch() instead.

> +    {
> +        /* Force the epp value to be zero for performance policy */
> +        epp = CPPC_ENERGY_PERF_MAX_PERFORMANCE;
> +        min_perf = max_perf;
> +    }
> +    else if ( policy->policy == CPUFREQ_POLICY_POWERSAVE )
> +        /* Force the epp value to be 0xff for powersave policy */
> +        /*
> +         * If set max_perf = min_perf = lowest_perf, we are putting
> +         * cpu cores in idle.
> +         */

Nit: Such two successive comments want combining. (Same near the top of the
function, as I notice only now.)

Furthermore I'm in trouble with interpreting this comment: To me "lowest"
doesn't mean "doing nothing" but "doing things as efficiently in terms of
power use as possible". IOW that's not idle. Yet the comment reads as if it
was meant to be an explanation of why we can't set max_perf from min_perf
here. That is, not matter what's meant to be said, I think this needs re-
wording (and possibly using subjunctive mood).

> +        epp = CPPC_ENERGY_PERF_MAX_POWERSAVE;
> +
> +    return amd_cppc_write_request(policy->cpu, min_perf,
> +                                  /* des_perf = 0 for epp mode */
> +                                  0,

The comment could do with putting on the same line as the 0, e.g.
(slightly adjusted)

    return amd_cppc_write_request(policy->cpu, min_perf,
                                  0 /* no des_perf for epp mode */,
                                  max_perf, epp);

> +static int cf_check amd_cppc_epp_set_policy(struct cpufreq_policy *policy)
> +{
> +    return amd_cppc_epp_update_limit(policy);
> +}

So the purpose of this wrapper is solely to have the actual function's
parameter be pointer-to-const? I don't think that's worth it; I also don't
think we do such elsewhere.

> --- a/xen/drivers/cpufreq/utility.c
> +++ b/xen/drivers/cpufreq/utility.c
> @@ -491,3 +491,14 @@ int __cpufreq_set_policy(struct cpufreq_policy *data,
>  
>      return __cpufreq_governor(data, CPUFREQ_GOV_LIMITS);
>  }
> +
> +unsigned int cpufreq_parse_policy(const struct cpufreq_governor *gov)
> +{
> +    if ( !strncasecmp(gov->name, "performance", CPUFREQ_NAME_LEN) )
> +        return CPUFREQ_POLICY_PERFORMANCE;
> +
> +    if ( !strncasecmp(gov->name, "powersave", CPUFREQ_NAME_LEN) )
> +        return CPUFREQ_POLICY_POWERSAVE;
> +
> +    return CPUFREQ_POLICY_UNKNOWN;
> +}

Hmm, this isn't really parsing (in the sense of dealing with e.g. command
line elements). Maybe cpufreq_get_policy() or, more explicitly,
cpufreq_policy_from_governor()? Or something along these lines?

I also don't see why the more expensive case-insensitive comparison
routine needs using here.

Jan