Re: [PATCH V2 05/10] ACPI: cpufreq: Switch to QoS requests instead of cpufreq notifier

From: Rafael J. Wysocki
Date: Mon Aug 05 2019 - 05:43:03 EST


On Tuesday, July 23, 2019 8:14:05 AM CEST Viresh Kumar wrote:
> The cpufreq core now takes the min/max frequency constraints via QoS
> requests and the CPUFREQ_ADJUST notifier shall get removed later on.
>
> Switch over to using the QoS request for maximum frequency constraint
> for acpi driver.
>
> Signed-off-by: Viresh Kumar <viresh.kumar@xxxxxxxxxx>
> ---
> drivers/acpi/processor_driver.c | 44 +++++++++++--
> drivers/acpi/processor_perflib.c | 106 +++++++++++++------------------
> drivers/acpi/processor_thermal.c | 81 ++++++++++++-----------
> include/acpi/processor.h | 22 ++++---
> 4 files changed, 137 insertions(+), 116 deletions(-)
>
> diff --git a/drivers/acpi/processor_driver.c b/drivers/acpi/processor_driver.c
> index aea8d674a33d..e7a3f07e9879 100644
> --- a/drivers/acpi/processor_driver.c
> +++ b/drivers/acpi/processor_driver.c
> @@ -284,6 +284,35 @@ static int acpi_processor_stop(struct device *dev)
> return 0;
> }
>
> +bool acpi_processor_cpufreq_init;
> +
> +static int acpi_processor_notifier(struct notifier_block *nb,
> + unsigned long event, void *data)
> +{
> + struct cpufreq_policy *policy = data;
> + int cpu;
> +
> + if (event == CPUFREQ_CREATE_POLICY) {
> + for_each_cpu(cpu, policy->cpus)
> + per_cpu(processors, cpu)->policy = policy;
> +
> + acpi_thermal_cpufreq_init(policy);
> + acpi_processor_ppc_init(policy);
> + } else if (event == CPUFREQ_REMOVE_POLICY) {
> + acpi_processor_ppc_exit(policy);
> + acpi_thermal_cpufreq_exit(policy);
> +
> + for_each_cpu(cpu, policy->cpus)
> + per_cpu(processors, cpu)->policy = NULL;
> + }

It doesn't look like it is necessary to pass policy to the
functions here, just the CPU number.

Also I don't think it is necessary to squirrel the policy pointer.

> +
> + return 0;
> +}
> +
> +static struct notifier_block acpi_processor_notifier_block = {
> + .notifier_call = acpi_processor_notifier,
> +};
> +
> /*
> * We keep the driver loaded even when ACPI is not running.
> * This is needed for the powernow-k8 driver, that works even without
> @@ -310,8 +339,11 @@ static int __init acpi_processor_driver_init(void)
> cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD, "acpi/cpu-drv:dead",
> NULL, acpi_soft_cpu_dead);
>
> - acpi_thermal_cpufreq_init();
> - acpi_processor_ppc_init();
> + if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> + CPUFREQ_POLICY_NOTIFIER)) {
> + acpi_processor_cpufreq_init = true;

Can't that be set/cleared by acpi_processor_notifier() itself?

> + }

Redundant braces.

> +
> acpi_processor_throttling_init();
> return 0;
> err:
> @@ -324,8 +356,12 @@ static void __exit acpi_processor_driver_exit(void)
> if (acpi_disabled)
> return;
>
> - acpi_processor_ppc_exit();
> - acpi_thermal_cpufreq_exit();
> + if (acpi_processor_cpufreq_init) {
> + cpufreq_unregister_notifier(&acpi_processor_notifier_block,
> + CPUFREQ_POLICY_NOTIFIER);
> + acpi_processor_cpufreq_init = false;
> + }
> +
> cpuhp_remove_state_nocalls(hp_online);
> cpuhp_remove_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD);
> driver_unregister(&acpi_processor_driver);
> diff --git a/drivers/acpi/processor_perflib.c b/drivers/acpi/processor_perflib.c
> index ee87cb6f6e59..1a22b2415a8b 100644
> --- a/drivers/acpi/processor_perflib.c
> +++ b/drivers/acpi/processor_perflib.c
> @@ -50,57 +50,13 @@ module_param(ignore_ppc, int, 0644);
> MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
> "limited by BIOS, this should help");
>
> -#define PPC_REGISTERED 1
> -#define PPC_IN_USE 2
> -
> -static int acpi_processor_ppc_status;
> -
> -static int acpi_processor_ppc_notifier(struct notifier_block *nb,
> - unsigned long event, void *data)
> -{
> - struct cpufreq_policy *policy = data;
> - struct acpi_processor *pr;
> - unsigned int ppc = 0;
> -
> - if (ignore_ppc < 0)
> - ignore_ppc = 0;
> -
> - if (ignore_ppc)
> - return 0;
> -
> - if (event != CPUFREQ_ADJUST)
> - return 0;
> -
> - mutex_lock(&performance_mutex);
> -
> - pr = per_cpu(processors, policy->cpu);
> - if (!pr || !pr->performance)
> - goto out;
> -
> - ppc = (unsigned int)pr->performance_platform_limit;
> -
> - if (ppc >= pr->performance->state_count)
> - goto out;
> -
> - cpufreq_verify_within_limits(policy, 0,
> - pr->performance->states[ppc].
> - core_frequency * 1000);
> -
> - out:
> - mutex_unlock(&performance_mutex);
> -
> - return 0;
> -}
> -
> -static struct notifier_block acpi_ppc_notifier_block = {
> - .notifier_call = acpi_processor_ppc_notifier,
> -};
> +static int acpi_processor_ppc_in_use;

Why not bool?

>
> static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
> {
> acpi_status status = 0;
> unsigned long long ppc = 0;
> -
> + int ret;
>
> if (!pr)
> return -EINVAL;
> @@ -112,7 +68,7 @@ static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
> status = acpi_evaluate_integer(pr->handle, "_PPC", NULL, &ppc);
>
> if (status != AE_NOT_FOUND)
> - acpi_processor_ppc_status |= PPC_IN_USE;
> + acpi_processor_ppc_in_use = 1;
>
> if (ACPI_FAILURE(status) && status != AE_NOT_FOUND) {
> ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PPC"));
> @@ -124,6 +80,16 @@ static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
>
> pr->performance_platform_limit = (int)ppc;
>
> + if (ignore_ppc || ppc >= pr->performance->state_count)
> + return 0;
> +
> + ret = dev_pm_qos_update_request(pr->perflib_req,
> + pr->performance->states[ppc].core_frequency * 1000);
> + if (ret) {
> + pr_warn("Failed to update perflib freq constraint: cpu%d (%d)\n",
> + pr->id, ret);
> + }
> +
> return 0;
> }
>
> @@ -184,23 +150,39 @@ int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
> }
> EXPORT_SYMBOL(acpi_processor_get_bios_limit);
>
> -void acpi_processor_ppc_init(void)
> +void acpi_processor_ppc_init(struct cpufreq_policy *policy)
> {
> - if (!cpufreq_register_notifier
> - (&acpi_ppc_notifier_block, CPUFREQ_POLICY_NOTIFIER))
> - acpi_processor_ppc_status |= PPC_REGISTERED;
> - else
> - printk(KERN_DEBUG
> - "Warning: Processor Platform Limit not supported.\n");
> + struct acpi_processor *pr = per_cpu(processors, policy->cpu);
> + struct dev_pm_qos_request *req;
> + int ret;
> +
> + req = kzalloc(sizeof(*req), GFP_KERNEL);
> + if (!req)
> + return;

The above wouldn't be necessary if the request struct was part of struct acpi_processor.

> +
> + ret = dev_pm_qos_add_request(get_cpu_device(policy->cpu),
> + req, DEV_PM_QOS_MAX_FREQUENCY,
> + policy->cpuinfo.max_freq);

The last argument doesn't need to be policy->cpuinfo.max_freq AFAICS. It might as well
be a large constant ("infinity"), so the function doesn't seem to need a policy pointer,
just the CPU number.

> + if (ret < 0) {
> + pr_err("Failed to add freq constraint for cpu%d (%d)\n",
> + policy->cpu, ret);
> + kfree(req);
> + return;
> + }
> +
> + pr->perflib_req = req;
> +
> + if (ignore_ppc < 0)
> + ignore_ppc = 0;
> }
>
> -void acpi_processor_ppc_exit(void)
> +void acpi_processor_ppc_exit(struct cpufreq_policy *policy)
> {
> - if (acpi_processor_ppc_status & PPC_REGISTERED)
> - cpufreq_unregister_notifier(&acpi_ppc_notifier_block,
> - CPUFREQ_POLICY_NOTIFIER);
> + struct acpi_processor *pr = per_cpu(processors, policy->cpu);
>
> - acpi_processor_ppc_status &= ~PPC_REGISTERED;
> + dev_pm_qos_remove_request(pr->perflib_req);
> + kfree(pr->perflib_req);
> + pr->perflib_req = NULL;
> }
>
> static int acpi_processor_get_performance_control(struct acpi_processor *pr)
> @@ -477,7 +459,7 @@ int acpi_processor_notify_smm(struct module *calling_module)
> static int is_done = 0;
> int result;
>
> - if (!(acpi_processor_ppc_status & PPC_REGISTERED))
> + if (!acpi_processor_cpufreq_init)
> return -EBUSY;
>
> if (!try_module_get(calling_module))
> @@ -513,7 +495,7 @@ int acpi_processor_notify_smm(struct module *calling_module)
> * we can allow the cpufreq driver to be rmmod'ed. */
> is_done = 1;
>
> - if (!(acpi_processor_ppc_status & PPC_IN_USE))
> + if (!acpi_processor_ppc_in_use)
> module_put(calling_module);
>
> return 0;
> @@ -742,7 +724,7 @@ acpi_processor_register_performance(struct acpi_processor_performance
> {
> struct acpi_processor *pr;
>
> - if (!(acpi_processor_ppc_status & PPC_REGISTERED))
> + if (!acpi_processor_cpufreq_init)
> return -EINVAL;
>
> mutex_lock(&performance_mutex);
> diff --git a/drivers/acpi/processor_thermal.c b/drivers/acpi/processor_thermal.c
> index 50fb0107375e..02407b33b874 100644
> --- a/drivers/acpi/processor_thermal.c
> +++ b/drivers/acpi/processor_thermal.c
> @@ -35,7 +35,6 @@ ACPI_MODULE_NAME("processor_thermal");
> #define CPUFREQ_THERMAL_MAX_STEP 3
>
> static DEFINE_PER_CPU(unsigned int, cpufreq_thermal_reduction_pctg);
> -static unsigned int acpi_thermal_cpufreq_is_init = 0;
>
> #define reduction_pctg(cpu) \
> per_cpu(cpufreq_thermal_reduction_pctg, phys_package_first_cpu(cpu))
> @@ -61,35 +60,11 @@ static int phys_package_first_cpu(int cpu)
> static int cpu_has_cpufreq(unsigned int cpu)
> {
> struct cpufreq_policy policy;
> - if (!acpi_thermal_cpufreq_is_init || cpufreq_get_policy(&policy, cpu))
> + if (!acpi_processor_cpufreq_init || cpufreq_get_policy(&policy, cpu))
> return 0;
> return 1;
> }
>
> -static int acpi_thermal_cpufreq_notifier(struct notifier_block *nb,
> - unsigned long event, void *data)
> -{
> - struct cpufreq_policy *policy = data;
> - unsigned long max_freq = 0;
> -
> - if (event != CPUFREQ_ADJUST)
> - goto out;
> -
> - max_freq = (
> - policy->cpuinfo.max_freq *
> - (100 - reduction_pctg(policy->cpu) * 20)
> - ) / 100;
> -
> - cpufreq_verify_within_limits(policy, 0, max_freq);
> -
> - out:
> - return 0;
> -}
> -
> -static struct notifier_block acpi_thermal_cpufreq_notifier_block = {
> - .notifier_call = acpi_thermal_cpufreq_notifier,
> -};
> -
> static int cpufreq_get_max_state(unsigned int cpu)
> {
> if (!cpu_has_cpufreq(cpu))
> @@ -108,7 +83,9 @@ static int cpufreq_get_cur_state(unsigned int cpu)
>
> static int cpufreq_set_cur_state(unsigned int cpu, int state)
> {
> - int i;
> + struct acpi_processor *pr;
> + unsigned long max_freq;
> + int i, ret;
>
> if (!cpu_has_cpufreq(cpu))
> return 0;
> @@ -121,33 +98,53 @@ static int cpufreq_set_cur_state(unsigned int cpu, int state)
> * frequency.
> */
> for_each_online_cpu(i) {
> - if (topology_physical_package_id(i) ==
> + if (topology_physical_package_id(i) !=
> topology_physical_package_id(cpu))
> - cpufreq_update_policy(i);
> + continue;
> +
> + pr = per_cpu(processors, i);

This could use cpufreq_cpu_get() to get to the policy (given the CPU number).

> + max_freq = (pr->policy->cpuinfo.max_freq * (100 - reduction_pctg(i) * 20)) / 100;
> +
> + ret = dev_pm_qos_update_request(pr->thermal_req, max_freq);
> + if (ret) {
> + pr_warn("Failed to update thermal freq constraint: cpu%d (%d)\n",
> + pr->id, ret);

Please spell CPU in capitals in messages.

> + }
> }
> return 0;
> }
>
> -void acpi_thermal_cpufreq_init(void)
> +void acpi_thermal_cpufreq_init(struct cpufreq_policy *policy)
> {
> - int i;
> + struct acpi_processor *pr = per_cpu(processors, policy->cpu);
> + struct dev_pm_qos_request *req;
> + int ret;
> +
> + req = kzalloc(sizeof(*req), GFP_KERNEL);
> + if (!req)
> + return;

Again, the request struct could be part of struct acpi_processor and the code
would be simpler then.

> +
> + ret = dev_pm_qos_add_request(get_cpu_device(policy->cpu),
> + req, DEV_PM_QOS_MAX_FREQUENCY,
> + policy->cpuinfo.max_freq);

And again, the last argument here could be a large ("inifinity") constant here, so
no need to take the policy pointer.

> + if (ret < 0) {
> + pr_err("Failed to add freq constraint for cpu%d (%d)\n",
> + policy->cpu, ret);
> + kfree(req);
> + return;
> + }
>
> - i = cpufreq_register_notifier(&acpi_thermal_cpufreq_notifier_block,
> - CPUFREQ_POLICY_NOTIFIER);
> - if (!i)
> - acpi_thermal_cpufreq_is_init = 1;
> + pr->thermal_req = req;
> }
>
> -void acpi_thermal_cpufreq_exit(void)
> +void acpi_thermal_cpufreq_exit(struct cpufreq_policy *policy)
> {
> - if (acpi_thermal_cpufreq_is_init)
> - cpufreq_unregister_notifier
> - (&acpi_thermal_cpufreq_notifier_block,
> - CPUFREQ_POLICY_NOTIFIER);
> + struct acpi_processor *pr = per_cpu(processors, policy->cpu);
>
> - acpi_thermal_cpufreq_is_init = 0;
> + dev_pm_qos_remove_request(pr->thermal_req);
> + kfree(pr->thermal_req);
> + pr->thermal_req = NULL;
> }
> -
> #else /* ! CONFIG_CPU_FREQ */
> static int cpufreq_get_max_state(unsigned int cpu)
> {
> diff --git a/include/acpi/processor.h b/include/acpi/processor.h
> index 1194a4c78d55..a1a7966bb755 100644
> --- a/include/acpi/processor.h
> +++ b/include/acpi/processor.h
> @@ -4,6 +4,8 @@
>
> #include <linux/kernel.h>
> #include <linux/cpu.h>
> +#include <linux/cpufreq.h>
> +#include <linux/pm_qos.h>
> #include <linux/thermal.h>
> #include <asm/acpi.h>
>
> @@ -230,6 +232,9 @@ struct acpi_processor {
> struct acpi_processor_limit limit;
> struct thermal_cooling_device *cdev;
> struct device *dev; /* Processor device. */
> + struct cpufreq_policy *policy;
> + struct dev_pm_qos_request *perflib_req;
> + struct dev_pm_qos_request *thermal_req;
> };
>
> struct acpi_processor_errata {
> @@ -296,16 +301,17 @@ static inline void acpi_processor_ffh_cstate_enter(struct acpi_processor_cx
> /* in processor_perflib.c */
>
> #ifdef CONFIG_CPU_FREQ
> -void acpi_processor_ppc_init(void);
> -void acpi_processor_ppc_exit(void);
> +extern bool acpi_processor_cpufreq_init;
> +void acpi_processor_ppc_init(struct cpufreq_policy *policy);
> +void acpi_processor_ppc_exit(struct cpufreq_policy *policy);
> void acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag);
> extern int acpi_processor_get_bios_limit(int cpu, unsigned int *limit);
> #else
> -static inline void acpi_processor_ppc_init(void)
> +static inline void acpi_processor_ppc_init(struct cpufreq_policy *policy)
> {
> return;
> }
> -static inline void acpi_processor_ppc_exit(void)
> +static inline void acpi_processor_ppc_exit(struct cpufreq_policy *policy)
> {
> return;
> }
> @@ -421,14 +427,14 @@ static inline int acpi_processor_hotplug(struct acpi_processor *pr)
> int acpi_processor_get_limit_info(struct acpi_processor *pr);
> extern const struct thermal_cooling_device_ops processor_cooling_ops;
> #if defined(CONFIG_ACPI_CPU_FREQ_PSS) & defined(CONFIG_CPU_FREQ)
> -void acpi_thermal_cpufreq_init(void);
> -void acpi_thermal_cpufreq_exit(void);
> +void acpi_thermal_cpufreq_init(struct cpufreq_policy *policy);
> +void acpi_thermal_cpufreq_exit(struct cpufreq_policy *policy);
> #else
> -static inline void acpi_thermal_cpufreq_init(void)
> +static inline void acpi_thermal_cpufreq_init(struct cpufreq_policy *policy)
> {
> return;
> }
> -static inline void acpi_thermal_cpufreq_exit(void)
> +static inline void acpi_thermal_cpufreq_exit(struct cpufreq_policy *policy)
> {
> return;
> }
>

Thanks!