Hello,

both Rafael and Viresh make a similar remark: why adding a new "max_boost"
variable, since "max_freq" is already available and could be used instead.

Replying here to both.

On Tue, 2021-02-02 at 20:26 +0100, Rafael J. Wysocki wrote:

On Tue, Feb 2, 2021 at 7:59 PM Rafael J. Wysocki [off-list ref] wrote:

quoted

On Fri, Jan 22, 2021 at 9:47 PM Giovanni Gherdovich [off-list ref] wrote:

[cut]

quoted

@@ -779,15 +829,25 @@ static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
                    freq_table[valid_states-1].frequency / 1000)
                        continue;

+               freq = perf->states[i].core_frequency * 1000;
                freq_table[valid_states].driver_data = i;
-               freq_table[valid_states].frequency =
-                   perf->states[i].core_frequency * 1000;
+               freq_table[valid_states].frequency = freq;
+
+               if (freq > max_freq)
+                       max_freq = freq;
+
                valid_states++;
        }
        freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
        policy->freq_table = freq_table;
        perf->state = 0;

+       if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
+           amd_max_boost(max_freq, &max_boost)) {
+               policy->cpuinfo.max_boost = max_boost;

Why not to set max_freq to max_boost instead?

I mean, would setting the frequency in the last table entry to max_boost work?

Alternatively, one more (artificial) entry with the frequency equal to
max_boost could be added.

On Wed, 2021-02-03 at 11:34 +0530, Viresh Kumar wrote:

[cut]

On 22-01-21, 21:40, Giovanni Gherdovich wrote:

quoted

diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
index 6931f0cdeb80..541f3db3f576 100644
--- a/kernel/sched/cpufreq_schedutil.c
+++ b/kernel/sched/cpufreq_schedutil.c

@@ -159,8 +159,12 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 				  unsigned long util, unsigned long max)
 {
 	struct cpufreq_policy *policy = sg_policy->policy;
-	unsigned int freq = arch_scale_freq_invariant() ?
-				policy->cpuinfo.max_freq : policy->cur;
+	unsigned int freq, max_freq;
+
+	max_freq = cpufreq_driver_has_max_boost() ?
+			policy->cpuinfo.max_boost : policy->cpuinfo.max_freq;

Also, can't we update max_freq itself from the cpufreq driver? What
troubles will it cost ?

I could add the max boost frequency to the frequency table (and
policy->cpuinfo.max_freq would follow), yes, but that would trigger the
following warning from acpi-cpufreq.c:

static void acpi_cpufreq_cpu_ready(struct cpufreq_policy *policy)
{
        struct acpi_processor_performance *perf = per_cpu_ptr(acpi_perf_data,
                                                              policy->cpu);

        if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
                pr_warn(FW_WARN "P-state 0 is not max freq\n");
}

so I thought that to stay out of troubles I'd supply a different variable,
max_boost, to be used only in the schedutil formula. After schedutil
figures out a desired next_freq then the usual comparison with the
firmware-supplied frequency table could take place.

Altering the frequency table seemed more invasive because once a freq value is
in there, it's going to be actually requested by the driver to the platform. I
only want my max_boost to stretch the range of schedutil's next_freq.

On Tue, 2021-02-02 at 19:59 +0100, Rafael J. Wysocki wrote:

[cut]
Also notice that the static branch is global and the max_boost value
for different CPUs may be different, at least in theory.

In theory yes, but I'm guarding the code with two conditions:

* vendor is X86_VENDOR_AMD
* cppc_get_perf_caps() returns success

this identifies AMD EPYC cpus with model 7xx2 and later, where max_boost is
the same on all cores. I may have added synchronization so that only one cpu
sets the value, but I didn't in the interest of simplicity for an -rc patch
(I'd have to consider hotplug, the maxcpus= command line param, ecc).


Giovanni