On 22-02-21, 16:57, Rafael J. Wysocki wrote:

On Mon, Feb 22, 2021 at 12:20 PM Viresh Kumar [off-list ref] wrote:
Even though the driver is located in drivers/cpufreq/ CPPC is part of
ACPI and so a CC to linux-acpi is missing.

I just used get-maintainers, perhaps we should add an entry for this
in MAINTAINERS, will be orphan though..

quoted

diff --git a/drivers/cpufreq/Kconfig.arm b/drivers/cpufreq/Kconfig.arm
index e65e0a43be64..a3e2d6dfea70 100644
--- a/drivers/cpufreq/Kconfig.arm
+++ b/drivers/cpufreq/Kconfig.arm

@@ -19,6 +19,15 @@ config ACPI_CPPC_CPUFREQ

          If in doubt, say N.

+config ACPI_CPPC_CPUFREQ_FIE
+       bool "Frequency Invariance support for CPPC cpufreq driver"
+       depends on ACPI_CPPC_CPUFREQ

In theory, the CPPC cpufreq driver can be used on systems with
nontrivial arch_freq_scale_tick() in which case the latter should be
used I suppose.

Would that actually happen if this option is enabled?

IIUC, you are saying that if this driver runs on x86 then we want
arch_freq_scale_tick() from arch/x86/kernel/smpboot.c to run instead
of this ? Yes that will happen because x86 doesn't enable
CONFIG_GENERIC_ARCH_TOPOLOGY and so this code will never trigger.

For other cases, like ARM AMU counters, the arch specific
implementation takes precedence to this.

quoted

+static void __init cppc_freq_invariance_init(void)
+{
+       struct cppc_perf_fb_ctrs fb_ctrs = {0};
+       struct cppc_freq_invariance *cppc_fi;
+       struct sched_attr attr = {
+               .size           = sizeof(struct sched_attr),
+               .sched_policy   = SCHED_DEADLINE,
+               .sched_nice     = 0,
+               .sched_priority = 0,
+               /*
+                * Fake (unused) bandwidth; workaround to "fix"
+                * priority inheritance.
+                */
+               .sched_runtime  = 1000000,
+               .sched_deadline = 10000000,
+               .sched_period   = 10000000,
+       };
+       int i, ret;
+
+       if (cppc_cpufreq_driver.get == hisi_cppc_cpufreq_get_rate)
+               return;
+
+       kworker_fie = kthread_create_worker(0, "cppc_fie");
+       if (IS_ERR(kworker_fie))
+               return;
+
+       for_each_possible_cpu(i) {
+               cppc_fi = &per_cpu(cppc_freq_inv, i);
+
+               /* A policy failed to initialize, abort */
+               if (unlikely(!cppc_fi->cpu_data))
+                       return cppc_freq_invariance_exit();
+
+               kthread_init_work(&cppc_fi->work, cppc_scale_freq_workfn);
+               init_irq_work(&cppc_fi->irq_work, cppc_irq_work);

What would be wrong with doing the above in
cppc_freq_invariance_policy_init()?  It looks like a better place to
me.

Can move it there as well, I just kept policy specific stuff there as
ideally I wanted to do everything here.

quoted

+               ret = sched_setattr_nocheck(kworker_fie->task, &attr);

And this needs to be done only once if I'm not mistaken.

Yes, I failed to fix this when I went to a single kworker.

-- 
viresh