Skip to contentSkip to conversation

cpufreq: skip loading acpi_cpufreq after intel_pstate

PATCH In review Yinghai Lu 2 participants 3 messages Sep 20, 2013

Subsystems (2)
lore
Share & tools
lore

Yinghai Lu

author
#0 lore raw

PATCH 1/1 cpufreq: skip loading acpi_cpufreq after intel_pstate

If the hw support intel_pstate and acpi_cpufreq, intel_pstate will get loaded at first.

acpi_cpufreq_init will call acpi_cpufreq_early_init()
and it allocate perf data and init those perf data in ACPI core, (that will go over all cpus).
but late it will free them as cpufreq_register_driver(acpi_cpufreq) will return fail as init_pstate already take over before.

Use cpufreq_get_current_driver() to check if we can skip the acpi_cpufreq loading.

Also there is racing in

	__acpi_processor_start
		 ==> acpi_processor_load_module
			 ==> request_module_nowait/requested = 1

before first pr path to have requested set, second cpu would request again. that will cause acpi_cpufreq_early_init to be called in parallel. that is cause for intermittent crashes.

So add mutex to protect it.

---
 drivers/acpi/processor_perflib.c |    8 +++++++-
 drivers/cpufreq/acpi-cpufreq.c   |    4 ++++
 2 files changed, 11 insertions(+), 1 deletion(-)
drivers/cpufreq/acpi-cpufreq.c +4 -0
Index: linux-2.6/drivers/cpufreq/acpi-cpufreq.c
===================================================================
@@ -986,6 +986,10 @@ static int __init acpi_cpufreq_init(void
 {
 	int ret;
 
+	/* don't keep reloading if cpufreq_driver exists */
+	if (cpufreq_get_current_driver())
+		return 0;
+
 	if (acpi_disabled)
 		return 0;
 
drivers/acpi/processor_perflib.c +7 -1
Index: linux-2.6/drivers/acpi/processor_perflib.c
===================================================================
@@ -235,6 +235,7 @@ void acpi_processor_ppc_exit(void)
 	acpi_processor_ppc_status &= ~PPC_REGISTERED;
 }
 
+static DEFINE_MUTEX(acpi_cpufreq_load_lock);
 /*
  * Do a quick check if the systems looks like it should use ACPI
  * cpufreq. We look at a _PCT method being available, but don't
@@ -246,8 +247,12 @@ void acpi_processor_load_module(struct a
 	acpi_status status = 0;
 	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 
-	if (!arch_has_acpi_pdc() || requested)
+	mutex_lock(&acpi_cpufreq_load_lock);
+	if (!arch_has_acpi_pdc() || requested ||cpufreq_get_current_driver()) {
+		mutex_unlock(&acpi_cpufreq_load_lock);
 		return;
+	}
+
 	status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
 	if (!ACPI_FAILURE(status)) {
 		printk(KERN_INFO PREFIX "Requesting acpi_cpufreq\n");
@@ -255,6 +260,7 @@ void acpi_processor_load_module(struct a
 		requested = 1;
 	}
 	kfree(buffer.pointer);
+	mutex_unlock(&acpi_cpufreq_load_lock);
 }
 
 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
maintainer
#1 lore raw
Quoted 2 lines · Yinghai Lu

If the hw support intel_pstate and acpi_cpufreq, intel_pstate will get loaded at first.

s/at //

Quoted 3 lines

acpi_cpufreq_init will call acpi_cpufreq_early_init()
and it allocate perf data and init those perf data in ACPI core, (that

s/it /that will /

Quoted 1 line

will go over all cpus).

s/go over/cover/

Quoted 1 line

but late it will free them as cpufreq_register_driver(acpi_cpufreq) will

s/late/later

Quoted 1 line

return fail as init_pstate already take over before.

write it as: fail as intel_pstate is already registered.

Quoted 2 lines

Use cpufreq_get_current_driver() to check if we can skip the acpi_cpufreq loading.

The below material looks to be a separate issue and so we probably break this patch into two?

Quoted 10 lines

Also there is racing in

        __acpi_processor_start
                 ==> acpi_processor_load_module
                         ==> request_module_nowait/requested = 1

Index: linux-2.6/drivers/cpufreq/acpi-cpufreq.c
===================================================================

drivers/cpufreq/acpi-cpufreq.c static int __init acpi_cpufreq_init(void · line 994

Quoted 9 lines
drivers/cpufreq/acpi-cpufreq.c +4 -0
@@ -986,6 +986,10 @@ static int __init acpi_cpufreq_init(void
 {
        int ret;
 
+       /* don't keep reloading if cpufreq_driver exists */
+       if (cpufreq_get_current_driver())
+               return 0;
+
        if (acpi_disabled)
                return 0;

This looks fine to me..

drivers/acpi/processor_perflib.c

Quoted 4 lines

Index: linux-2.6/drivers/acpi/processor_perflib.c
===================================================================

drivers/acpi/processor_perflib.c +0 -0

Can't really review this one.. Rafael will do it..

author
#2 lore raw
Quoted 22 lines · Viresh Kumar
Quoted 2 lines · Yinghai Lu

If the hw support intel_pstate and acpi_cpufreq, intel_pstate will get loaded at first.

s/at //

Quoted 3 lines

acpi_cpufreq_init will call acpi_cpufreq_early_init()
and it allocate perf data and init those perf data in ACPI core, (that

s/it /that will /

Quoted 1 line

will go over all cpus).

s/go over/cover/

Quoted 1 line

but late it will free them as cpufreq_register_driver(acpi_cpufreq) will

s/late/later

Quoted 1 line

return fail as init_pstate already take over before.

Quoted 7 lines

write it as: fail as intel_pstate is already registered.

Quoted 2 lines

Use cpufreq_get_current_driver() to check if we can skip the acpi_cpufreq loading.

The below material looks to be a separate issue and so we probably break this patch into two?

ok, check that.

Thanks

Yinghai