PM / Domains: Add preliminary support for cpuidle, v2
Prashant Gaikwad [Sun, 14 Oct 2012 17:59:57 +0000 (22:59 +0530)]
On some systems there are CPU cores located in the same power
domains as I/O devices.  Then, power can only be removed from the
domain if all I/O devices in it are not in use and the CPU core
is idle.  Add preliminary support for that to the generic PM domains
framework.

First, the platform is expected to provide a cpuidle driver with one
extra state designated for use with the generic PM domains code.
This state should be initially disabled and its exit_latency value
should be set to whatever time is needed to bring up the CPU core
itself after restoring power to it, not including the domain's
power on latency.  Its .enter() callback should point to a procedure
that will remove power from the domain containing the CPU core at
the end of the CPU power transition.

The remaining characteristics of the extra cpuidle state, referred to
as the "domain" cpuidle state below, (e.g. power usage, target
residency) should be populated in accordance with the properties of
the hardware.

Next, the platform should execute genpd_attach_cpuidle() on the PM
domain containing the CPU core.  That will cause the generic PM
domains framework to treat that domain in a special way such that:

 * When all devices in the domain have been suspended and it is about
   to be turned off, the states of the devices will be saved, but
   power will not be removed from the domain.  Instead, the "domain"
   cpuidle state will be enabled so that power can be removed from
   the domain when the CPU core is idle and the state has been chosen
   as the target by the cpuidle governor.

 * When the first I/O device in the domain is resumed and
   __pm_genpd_poweron(() is called for the first time after
   power has been removed from the domain, the "domain" cpuidle
   state will be disabled to avoid subsequent surprise power removals
   via cpuidle.

The effective exit_latency value of the "domain" cpuidle state
depends on the time needed to bring up the CPU core itself after
restoring power to it as well as on the power on latency of the
domain containing the CPU core.  Thus the "domain" cpuidle state's
exit_latency has to be recomputed every time the domain's power on
latency is updated, which may happen every time power is restored
to the domain, if the measured power on latency is greater than
the latency stored in the corresponding generic_pm_domain structure.

[Prashant: Resolved conflicts]

Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Reviewed-by: Kevin Hilman <khilman@ti.com>
(cherry picked from commit cbc9ef0287ab764d3da0129efa673808df641fe3)

Change-Id: I1ec4dda167b2a687986bb5da491d43941b8f6ae8
Signed-off-by: Prashant Gaikwad <pgaikwad@nvidia.com>
Reviewed-on: http://git-master/r/145220
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Bharat Nihalani <bnihalani@nvidia.com>

drivers/base/power/domain.c
drivers/cpuidle/cpuidle.c
drivers/cpuidle/governors/menu.c
drivers/cpuidle/sysfs.c
include/linux/cpuidle.h
include/linux/pm_domain.h

index da1d525..9046c65 100644 (file)
@@ -139,6 +139,19 @@ static void genpd_set_active(struct generic_pm_domain *genpd)
                genpd->status = GPD_STATE_ACTIVE;
 }
 
+static void genpd_recalc_cpu_exit_latency(struct generic_pm_domain *genpd)
+{
+       s64 usecs64;
+
+       if (!genpd->cpu_data)
+               return;
+
+       usecs64 = genpd->power_on_latency_ns;
+       do_div(usecs64, NSEC_PER_USEC);
+       usecs64 += genpd->cpu_data->saved_exit_latency;
+       genpd->cpu_data->idle_state->exit_latency = usecs64;
+}
+
 /**
  * __pm_genpd_poweron - Restore power to a given PM domain and its masters.
  * @genpd: PM domain to power up.
@@ -176,6 +189,13 @@ int __pm_genpd_poweron(struct generic_pm_domain *genpd)
                return 0;
        }
 
+       if (genpd->cpu_data) {
+               cpuidle_pause_and_lock();
+               genpd->cpu_data->idle_state->disabled = true;
+               cpuidle_resume_and_unlock();
+               goto out;
+       }
+
        /*
         * The list is guaranteed not to change while the loop below is being
         * executed, unless one of the masters' .power_on() callbacks fiddles
@@ -215,6 +235,7 @@ int __pm_genpd_poweron(struct generic_pm_domain *genpd)
                if (elapsed_ns > genpd->power_on_latency_ns) {
                        genpd->power_on_latency_ns = elapsed_ns;
                        genpd->max_off_time_changed = true;
+                       genpd_recalc_cpu_exit_latency(genpd);
                        if (genpd->name)
                                pr_warning("%s: Power-on latency exceeded, "
                                        "new value %lld ns\n", genpd->name,
@@ -222,6 +243,7 @@ int __pm_genpd_poweron(struct generic_pm_domain *genpd)
                }
        }
 
+ out:
        genpd_set_active(genpd);
 
        return 0;
@@ -455,6 +477,21 @@ static int pm_genpd_poweroff(struct generic_pm_domain *genpd)
                }
        }
 
+       if (genpd->cpu_data) {
+               /*
+                * If cpu_data is set, cpuidle should turn the domain off when
+                * the CPU in it is idle.  In that case we don't decrement the
+                * subdomain counts of the master domains, so that power is not
+                * removed from the current domain prematurely as a result of
+                * cutting off the masters' power.
+                */
+               genpd->status = GPD_STATE_POWER_OFF;
+               cpuidle_pause_and_lock();
+               genpd->cpu_data->idle_state->disabled = false;
+               cpuidle_resume_and_unlock();
+               goto out;
+       }
+
        if (genpd->power_off) {
                ktime_t time_start;
                s64 elapsed_ns;
@@ -1600,6 +1637,92 @@ int __pm_genpd_remove_callbacks(struct device *dev, bool clear_td)
 }
 EXPORT_SYMBOL_GPL(__pm_genpd_remove_callbacks);
 
+int genpd_attach_cpuidle(struct generic_pm_domain *genpd, int state,
+                               struct cpuidle_device *cpuidle_dev)
+{
+       struct cpuidle_driver *cpuidle_drv;
+       struct gpd_cpu_data *cpu_data;
+       struct cpuidle_state *idle_state;
+       int ret = 0;
+
+       if (IS_ERR_OR_NULL(genpd) || state < 0)
+               return -EINVAL;
+
+       genpd_acquire_lock(genpd);
+
+       if (!cpuidle_dev) {
+               return -ENODEV;
+               goto out;
+       }
+
+       if (genpd->cpu_data) {
+               ret = -EEXIST;
+               goto out;
+       }
+       cpu_data = kzalloc(sizeof(*cpu_data), GFP_KERNEL);
+       if (!cpu_data) {
+               ret = -ENOMEM;
+               goto out;
+       }
+       cpuidle_drv = cpuidle_driver_ref();
+       if (!cpuidle_drv) {
+               ret = -ENODEV;
+               goto out;
+       }
+       if (cpuidle_dev->state_count <= state) {
+               ret = -EINVAL;
+               goto err;
+       }
+       idle_state = &cpuidle_dev->states[state];
+       if (!idle_state->disabled) {
+               ret = -EAGAIN;
+               goto err;
+       }
+       cpu_data->idle_state = idle_state;
+       cpu_data->saved_exit_latency = idle_state->exit_latency;
+       genpd->cpu_data = cpu_data;
+       genpd_recalc_cpu_exit_latency(genpd);
+
+ out:
+       genpd_release_lock(genpd);
+       return ret;
+
+ err:
+       cpuidle_driver_unref();
+       goto out;
+}
+
+int genpd_detach_cpuidle(struct generic_pm_domain *genpd)
+{
+       struct gpd_cpu_data *cpu_data;
+       struct cpuidle_state *idle_state;
+       int ret = 0;
+
+       if (IS_ERR_OR_NULL(genpd))
+               return -EINVAL;
+
+       genpd_acquire_lock(genpd);
+
+       cpu_data = genpd->cpu_data;
+       if (!cpu_data) {
+               ret = -ENODEV;
+               goto out;
+       }
+       idle_state = cpu_data->idle_state;
+       if (!idle_state->disabled) {
+               ret = -EAGAIN;
+               goto out;
+       }
+       idle_state->exit_latency = cpu_data->saved_exit_latency;
+       cpuidle_driver_unref();
+       genpd->cpu_data = NULL;
+       kfree(cpu_data);
+
+ out:
+       genpd_release_lock(genpd);
+       return ret;
+}
+
 /* Default device callbacks for generic PM domains. */
 
 /**
index 90f48c0..4210dc5 100644 (file)
@@ -280,7 +280,7 @@ static void poll_idle_init(struct cpuidle_device *dev)
        state->power_usage = -1;
        state->flags = 0;
        state->enter = poll_idle;
-       state->disable = 0;
+       state->disabled = false;
 }
 #else
 static void poll_idle_init(struct cpuidle_device *dev) {}
index a43a524..36c7fa9 100644 (file)
@@ -285,7 +285,7 @@ static int menu_select(struct cpuidle_device *dev)
         * unless the timer is happening really really soon.
         */
        if (data->expected_us > 5 &&
-               dev->states[CPUIDLE_DRIVER_STATE_START].disable == 0)
+           !dev->states[CPUIDLE_DRIVER_STATE_START].disabled)
                data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
 
        /*
@@ -295,7 +295,7 @@ static int menu_select(struct cpuidle_device *dev)
        for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) {
                struct cpuidle_state *s = &dev->states[i];
 
-               if (s->disable)
+               if (s->disabled)
                        continue;
                if (s->target_residency > data->predicted_us)
                        continue;
index b50317c..6c4166a 100644 (file)
@@ -245,9 +245,9 @@ static ssize_t store_state_##_name(struct cpuidle_state *state, \
        if (err) \
                return err; \
        if (value) \
-               state->disable = 1; \
+               state->disabled = true; \
        else \
-               state->disable = 0; \
+               state->disabled = false; \
        return size; \
 }
 
@@ -273,8 +273,8 @@ define_show_state_ull_function(usage)
 define_show_state_ull_function(time)
 define_show_state_str_function(name)
 define_show_state_str_function(desc)
-define_show_state_function(disable)
-define_store_state_function(disable)
+define_show_state_function(disabled)
+define_store_state_function(disabled)
 
 define_one_state_ro(name, show_state_name);
 define_one_state_ro(desc, show_state_desc);
@@ -282,7 +282,7 @@ define_one_state_ro(latency, show_state_exit_latency);
 define_one_state_ro(power, show_state_power_usage);
 define_one_state_ro(usage, show_state_usage);
 define_one_state_ro(time, show_state_time);
-define_one_state_rw(disable, show_state_disable, store_state_disable);
+define_one_state_rw(disabled, show_state_disabled, store_state_disabled);
 
 static struct attribute *cpuidle_state_default_attrs[] = {
        &attr_name.attr,
@@ -291,7 +291,7 @@ static struct attribute *cpuidle_state_default_attrs[] = {
        &attr_power.attr,
        &attr_usage.attr,
        &attr_time.attr,
-       &attr_disable.attr,
+       &attr_disabled.attr,
        NULL
 };
 
index 085304c..0d14132 100644 (file)
@@ -45,7 +45,7 @@ struct cpuidle_state {
        unsigned int    exit_latency; /* in US */
        int             power_usage; /* in mW */
        unsigned int    target_residency; /* in US */
-       unsigned int    disable;
+       bool            disabled; /* disabled on all CPUs */
 
        int (*enter)    (struct cpuidle_device *dev,
                        int index);
index 30f794e..f99fb3c 100644 (file)
@@ -15,6 +15,7 @@
 #include <linux/err.h>
 #include <linux/of.h>
 #include <linux/notifier.h>
+#include <linux/cpuidle.h>
 
 enum gpd_status {
        GPD_STATE_ACTIVE = 0,   /* PM domain is active */
@@ -45,6 +46,11 @@ struct gpd_dev_ops {
        bool (*active_wakeup)(struct device *dev);
 };
 
+struct gpd_cpu_data {
+       unsigned int saved_exit_latency;
+       struct cpuidle_state *idle_state;
+};
+
 struct generic_pm_domain {
        struct dev_pm_domain domain;    /* PM domain operations */
        struct list_head gpd_list_node; /* Node in the global PM domains list */
@@ -75,6 +81,7 @@ struct generic_pm_domain {
        bool max_off_time_changed;
        bool cached_power_down_ok;
        struct device_node *of_node; /* Node in device tree */
+       struct gpd_cpu_data *cpu_data;
 };
 
 static inline struct generic_pm_domain *pd_to_genpd(struct dev_pm_domain *pd)
@@ -155,6 +162,9 @@ extern int pm_genpd_add_callbacks(struct device *dev,
                                  struct gpd_dev_ops *ops,
                                  struct gpd_timing_data *td);
 extern int __pm_genpd_remove_callbacks(struct device *dev, bool clear_td);
+extern int genpd_attach_cpuidle(struct generic_pm_domain *genpd, int state,
+                               struct cpuidle_device *cpuidle_dev);
+extern int genpd_detach_cpuidle(struct generic_pm_domain *genpd);
 extern void pm_genpd_init(struct generic_pm_domain *genpd,
                          struct dev_power_governor *gov, bool is_off);
 
@@ -211,6 +221,14 @@ static inline int __pm_genpd_remove_callbacks(struct device *dev, bool clear_td)
 {
        return -ENOSYS;
 }
+static inline int genpd_attach_cpuidle(struct generic_pm_domain *genpd, int st)
+{
+       return -ENOSYS;
+}
+static inline int genpd_detach_cpuidle(struct generic_pm_domain *genpd)
+{
+       return -ENOSYS;
+}
 static inline void pm_genpd_init(struct generic_pm_domain *genpd,
                                 struct dev_power_governor *gov, bool is_off)
 {