ARM: cpuidle: heterogeneous systems extension

Some ARM systems (eg big.LITTLE ones) can be composed of CPUs having different hardware power management configurations and in the context of CPUidle, different idle states. The generic ARM CPUidle driver treats all possible CPUs as equal and initializes a common idle driver through DT idle states for all possible CPUs.

Current driver cannot manage systems where CPUs are heterogeneous and therefore can have different idle states.

This patch augments the generic ARM CPUidle driver, by adding code that at boot initializes CPUidle drivers by going through the cpu_possible_mask and through DT parsing detects the cpus sharing the same idle states, thus creating the CPUidle drivers cpumasks.

The drivers are then initialized through the DT idle states interface, that parses and initializes the DT idle states for the cpus set in the drivers cpumasks.

This patch instantiates a static array of idle drivers, some of which can turn out to be unused (eg platforms with uniform idle states on all possible CPUs), and relies on the config option CPU_IDLE_MULTIPLE_DRIVERS to be selected by default; this can cause a little memory overhead, but at the same time allows supporting most of the current and future ARM platforms through a single generic CPUidle driver.

Signed-off-by Lorenzo Pieralisi <[email protected]>

Cc Howard Chen <[email protected]>

Cc Rob Herring <[email protected]>

Cc Kevin Hilman <[email protected]>

Cc Sudeep Holla <[email protected]>

Cc Lina Iyer <[email protected]>

Cc Daniel Lezcano <[email protected]>

Cc Grant Likely <[email protected]>

Cc Mathieu Poirier <[email protected]>

Cc Mark Rutland <[email protected]>

---
  drivers/cpuidle/Kconfig.arm   |   1 +
  drivers/cpuidle/cpuidle-arm.c | 176 ++++++++++++++++++++++++++++++++++++------
  2 files changed, 152 insertions(+), 25 deletions(-)

drivers/cpuidle/Kconfig.arm +1 -0

index 21340e0..90c6553 100644
@@ -3,6 +3,7 @@
  #
  config ARM_CPUIDLE
          bool "Generic ARM/ARM64 CPU idle Driver"
+        select CPU_IDLE_MULTIPLE_DRIVERS
          select DT_IDLE_STATES
          help
            Select this to enable generic cpuidle driver for ARM.

drivers/cpuidle/cpuidle-arm.c +22 -17

index 545069d..251fa2a 100644
@@ -14,6 +14,7 @@
  #include <linux/cpuidle.h>
  #include <linux/cpumask.h>
  #include <linux/cpu_pm.h>
+#include <linux/of_device.h>
  #include <linux/kernel.h>
  #include <linux/module.h>
  #include <linux/of.h>

@@ -58,23 +59,27 @@ static int arm_enter_idle_state(struct cpuidle_device *dev,
  	return ret ? -1 : idx;
  }
 
-static struct cpuidle_driver arm_idle_driver = {
-	.name = "arm_idle",
-	.owner = THIS_MODULE,
-	/*
-	 * State at index 0 is standby wfi and considered standard
-	 * on all ARM platforms. If in some platforms simple wfi
-	 * can't be used as "state 0", DT bindings must be implemented
-	 * to work around this issue and allow installing a special
-	 * handler for idle state index 0.
-	 */
-	.states[0] = {
-		.enter                  = arm_enter_idle_state,
-		.exit_latency           = 1,
-		.target_residency       = 1,
-		.power_usage		= UINT_MAX,
-		.name                   = "WFI",
-		.desc                   = "ARM WFI",
+#define ARM_CPUIDLE_MAX_DRIVERS	2
+
+static struct cpuidle_driver arm_idle_drivers[ARM_CPUIDLE_MAX_DRIVERS] = {
+	[0 ... ARM_CPUIDLE_MAX_DRIVERS - 1] = {
+		.name = "arm_idle",
+		.owner = THIS_MODULE,
+		/*
+		 * State at index 0 is standby wfi and considered standard
+		 * on all ARM platforms. If in some platforms simple wfi
+		 * can't be used as "state 0", DT bindings must be implemented
+		 * to work around this issue and allow installing a special
+		 * handler for idle state index 0.
+		 */
+		.states[0] = {
+			.enter                  = arm_enter_idle_state,
+			.exit_latency           = 1,
+			.target_residency       = 1,
+			.power_usage		= UINT_MAX,
+			.name                   = "WFI",
+			.desc                   = "ARM WFI",
+		}
  	}
  };
 

drivers/cpuidle/cpuidle-arm.c +103 -4

   */
-static int __init arm_idle_init(void)
+static bool __init idle_states_cmp(struct property *states1,
+				   struct property *states2)
+{
+	/*
+	 * NB: Implemented through code from drivers/of/unittest.c
+	 *     Function is generic and can be moved to generic OF code
+	 *     if needed
+	 */
+	return states1 && states2 &&
+	       (states1->length == states2->length) &&
+	       states1->value && states2->value &&
+	       !memcmp(states1->value, states2->value, states1->length);
+}
+
+static int __init arm_idle_init_driver(struct cpuidle_driver *drv)
  {
-	int cpu, ret;
-	struct cpuidle_driver *drv = &arm_idle_driver;
+	int ret, cpu;
  	struct cpuidle_device *dev;
+	struct property *curr_idle_states, *idle_states = NULL;
+	struct device_node *cpu_node;
+
+	for_each_cpu(cpu, drv->cpumask) {
+		cpu_node = of_cpu_device_node_get(cpu);
+		curr_idle_states = of_find_property(cpu_node,
+						    "cpu-idle-states", NULL);
+		of_node_put(cpu_node);
+
+		/*
+		 * Stash the first valid idle states phandle in the cpumask.
+		 * If curr_idle_states is NULL assigning it to idle_states
+		 * is harmless and it is managed by idle states comparison
+		 * code. Keep track of first valid phandle so that
+		 * subsequent cpus can compare against it.
+		 */
+		if (!idle_states)
+			idle_states = curr_idle_states;
+
+		/*
+		 * If idle states phandles are not equal, remove the
+		 * cpu from the driver mask since a CPUidle driver
+		 * is only capable of managing uniform idle states.
+		 *
+		 * Comparison works also when idle_states and
+		 * curr_idle_states are the same property, since
+		 * they can be == NULL so the cpu must be removed from
+		 * the driver mask in that case too (ie cpu has no idle
+		 * states).
+		 */
+		if (!idle_states_cmp(idle_states, curr_idle_states))
+			cpumask_clear_cpu(cpu, drv->cpumask);
+	}
+
+	/*
+	 *  If there are no valid states for this driver we rely on arch
+	 *  default idle behaviour, bail out
+	 */
+	if (!idle_states)
+		return -ENODEV;
 
  	/*
  	 * Initialize idle states data, starting at index 1.

@@ -117,7 +173,7 @@ static int __init arm_idle_init(void)
  	 * Call arch CPU operations in order to initialize
  	 * idle states suspend back-end specific data
  	 */
-	for_each_possible_cpu(cpu) {
+	for_each_cpu(cpu, drv->cpumask) {
  		ret = arm_cpuidle_init(cpu);
 
  		/*

@@ -157,7 +213,77 @@ out_fail:
  	}
 
  	cpuidle_unregister_driver(drv);
+	return ret;
+}
+
+/*
+ * arm_idle_init
+ *
+ * Registers the arm specific cpuidle driver(s) with the cpuidle
+ * framework. It relies on core code to parse the idle states
+ * and initialize them using driver data structures accordingly.
+ */
+static int __init arm_idle_init(void)
+{
+	int i, ret = -ENODEV;
+	struct cpuidle_driver *drv;
+	cpumask_var_t tmpmask;
+
+	/*
+	 * These drivers require DT idle states to be present.
+	 * If no idle states are detected there is no reason to
+	 * proceed any further hence we return early.
+	 */
+	if (!of_find_node_by_name(NULL, "idle-states"))
+		return -ENODEV;
+
+	if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
+		return -ENOMEM;
+
+	/*
+	 * We need to vet idle states to create CPUidle drivers
+	 * that share a common set of them. Create a tmp mask
+	 * that we use to keep track of initialized cpus.
+	 * Start off by initializing the mask with all possible
+	 * cpus, we clear it as we go, till either all cpus
+	 * have a CPUidle driver initialized or there are some
+	 * CPUs that have no idle states or a parsing error
+	 * occurs.
+	 */
+	cpumask_copy(tmpmask, cpu_possible_mask);
+
+	for (i = 0; !cpumask_empty(tmpmask); i++) {
+		if (i == ARM_CPUIDLE_MAX_DRIVERS) {
+			pr_warn("number of drivers exceeding static allocation\n");
+			break;
+		}
+
+		drv = &arm_idle_drivers[i];

drivers/cpuidle/cpuidle-arm.c +2 -0

+	}
 
+	free_cpumask_var(tmpmask);
  	return ret;
  }
  device_initcall(arm_idle_init);

On 04/16/2015 06:10 PM, Lorenzo Pieralisi wrote:

Quoted 8 lines

/*
- * arm_idle_init
+ * Compare idle states phandle properties
*
- * Registers the arm specific cpuidle driver with the cpuidle
- * framework. It relies on core code to parse the idle states
- * and initialize them using driver data structures accordingly.
+ * Return true if properties are valid and equal, false otherwise

Just a detail. Would be more consistent to return zero when valid and equal ?

Quoted 130 lines

   */

drivers/cpuidle/cpuidle-arm.c +109 -4

-static int __init arm_idle_init(void)
+static bool __init idle_states_cmp(struct property *states1,
+                                struct property *states2)
+{
+     /*
+      * NB: Implemented through code from drivers/of/unittest.c
+      *     Function is generic and can be moved to generic OF code
+      *     if needed
+      */
+     return states1 && states2 &&
+            (states1->length == states2->length) &&
+            states1->value && states2->value &&
+            !memcmp(states1->value, states2->value, states1->length);
+}
+
+static int __init arm_idle_init_driver(struct cpuidle_driver *drv)
  {
-     int cpu, ret;
-     struct cpuidle_driver *drv = &arm_idle_driver;
+     int ret, cpu;
      struct cpuidle_device *dev;
+     struct property *curr_idle_states, *idle_states = NULL;
+     struct device_node *cpu_node;
+
+     for_each_cpu(cpu, drv->cpumask) {
+             cpu_node = of_cpu_device_node_get(cpu);
+             curr_idle_states = of_find_property(cpu_node,
+                                                 "cpu-idle-states", NULL);
+             of_node_put(cpu_node);
+
+             /*
+              * Stash the first valid idle states phandle in the cpumask.
+              * If curr_idle_states is NULL assigning it to idle_states
+              * is harmless and it is managed by idle states comparison
+              * code. Keep track of first valid phandle so that
+              * subsequent cpus can compare against it.
+              */
+             if (!idle_states)
+                     idle_states = curr_idle_states;
+
+             /*
+              * If idle states phandles are not equal, remove the
+              * cpu from the driver mask since a CPUidle driver
+              * is only capable of managing uniform idle states.
+              *
+              * Comparison works also when idle_states and
+              * curr_idle_states are the same property, since
+              * they can be == NULL so the cpu must be removed from
+              * the driver mask in that case too (ie cpu has no idle
+              * states).
+              */
+             if (!idle_states_cmp(idle_states, curr_idle_states))
+                     cpumask_clear_cpu(cpu, drv->cpumask);
+     }
+
+     /*
+      *  If there are no valid states for this driver we rely on arch
+      *  default idle behaviour, bail out
+      */
+     if (!idle_states)
+             return -ENODEV;
 
      /*
       * Initialize idle states data, starting at index 1.

@@ -117,7 +173,7 @@ static int __init arm_idle_init(void)
       * Call arch CPU operations in order to initialize
       * idle states suspend back-end specific data
       */
-     for_each_possible_cpu(cpu) {
+     for_each_cpu(cpu, drv->cpumask) {
              ret = arm_cpuidle_init(cpu);
 
              /*

@@ -157,7 +213,77 @@ out_fail:
      }
 
      cpuidle_unregister_driver(drv);
+     return ret;
+}
+
+/*
+ * arm_idle_init
+ *
+ * Registers the arm specific cpuidle driver(s) with the cpuidle
+ * framework. It relies on core code to parse the idle states
+ * and initialize them using driver data structures accordingly.
+ */
+static int __init arm_idle_init(void)
+{
+     int i, ret = -ENODEV;
+     struct cpuidle_driver *drv;
+     cpumask_var_t tmpmask;
+
+     /*
+      * These drivers require DT idle states to be present.
+      * If no idle states are detected there is no reason to
+      * proceed any further hence we return early.
+      */
+     if (!of_find_node_by_name(NULL, "idle-states"))
+             return -ENODEV;
+
+     if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
+             return -ENOMEM;
+
+     /*
+      * We need to vet idle states to create CPUidle drivers
+      * that share a common set of them. Create a tmp mask
+      * that we use to keep track of initialized cpus.
+      * Start off by initializing the mask with all possible
+      * cpus, we clear it as we go, till either all cpus
+      * have a CPUidle driver initialized or there are some
+      * CPUs that have no idle states or a parsing error
+      * occurs.
+      */
+     cpumask_copy(tmpmask, cpu_possible_mask);
+
+     for (i = 0; !cpumask_empty(tmpmask); i++) {
+             if (i == ARM_CPUIDLE_MAX_DRIVERS) {
+                     pr_warn("number of drivers exceeding static allocation\n");
+                     break;
+             }
+
+             drv = &arm_idle_drivers[i];
+             drv->cpumask = kzalloc(cpumask_size(), GFP_KERNEL);
+             if (!drv->cpumask) {
+                     ret = -ENOMEM;
+                     break;
+             }
+             /*

drivers/cpuidle/cpuidle-arm.c out_fail: · line 271

Quoted 4 lines

drivers/cpuidle/cpuidle-arm.c +4 -0

+              * Force driver mask, arm_idle_init_driver()
+              * will tweak it by vetting idle states.
+              */
+             cpumask_copy(drv->cpumask, tmpmask);

Why do you need to copy tmpmask ? Isn't simpler to have a zero-ed cpumask and let the arm_idle_init_driver function to set a bit instead of clearing it ?

drivers/cpuidle/cpuidle-arm.c out_fail: · line 283

Quoted 12 lines

drivers/cpuidle/cpuidle-arm.c +11 -0

+             ret = arm_idle_init_driver(drv);
+             if (ret) {
+                     kfree(drv->cpumask);
+                     break;
+             }
+             /*
+              * Remove the cpus that were part of the registered
+              * driver from the mask of cpus to be initialized
+              * and restart.
+              */
+             cpumask_andnot(tmpmask, tmpmask, drv->cpumask);

If there is a DT definition with just a cluster with the cpuidle driver set and another one without any idle state, we will have always a pr_warn because i == ARM_CPUIDLE_MAX_DRIVERS due to tmpmask being never equal to a zero mask. Is it the purpose to detect such cases ?

drivers/cpuidle/cpuidle-arm.c out_fail: · line 289

Quoted 8 lines

drivers/cpuidle/cpuidle-arm.c +2 -0

+     }
 
+     free_cpumask_var(tmpmask);
      return ret;
  }
  device_initcall(arm_idle_init);

--
<http://www.linaro.org/> Linaro.org ??? Open source software for ARM SoCs

Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog

Quoted 1 line · Daniel Lezcano

On 04/16/2015 06:10 PM, Lorenzo Pieralisi wrote:

<snip>

Quoted 11 lines

Quoted 8 lines

/*
- * arm_idle_init
+ * Compare idle states phandle properties
*
- * Registers the arm specific cpuidle driver with the cpuidle
- * framework. It relies on core code to parse the idle states
- * and initialize them using driver data structures accordingly.
+ * Return true if properties are valid and equal, false otherwise

Just a detail. Would be more consistent to return zero when valid and equal ?

Consistent with memcmp you mean ? Yes, I can change it.

Quoted 130 lines

Quoted 130 lines

   */

drivers/cpuidle/cpuidle-arm.c +109 -4

-static int __init arm_idle_init(void)
+static bool __init idle_states_cmp(struct property *states1,
+                                struct property *states2)
+{
+     /*
+      * NB: Implemented through code from drivers/of/unittest.c
+      *     Function is generic and can be moved to generic OF code
+      *     if needed
+      */
+     return states1 && states2 &&
+            (states1->length == states2->length) &&
+            states1->value && states2->value &&
+            !memcmp(states1->value, states2->value, states1->length);
+}
+
+static int __init arm_idle_init_driver(struct cpuidle_driver *drv)
  {
-     int cpu, ret;
-     struct cpuidle_driver *drv = &arm_idle_driver;
+     int ret, cpu;
      struct cpuidle_device *dev;
+     struct property *curr_idle_states, *idle_states = NULL;
+     struct device_node *cpu_node;
+
+     for_each_cpu(cpu, drv->cpumask) {
+             cpu_node = of_cpu_device_node_get(cpu);
+             curr_idle_states = of_find_property(cpu_node,
+                                                 "cpu-idle-states", NULL);
+             of_node_put(cpu_node);
+
+             /*
+              * Stash the first valid idle states phandle in the cpumask.
+              * If curr_idle_states is NULL assigning it to idle_states
+              * is harmless and it is managed by idle states comparison
+              * code. Keep track of first valid phandle so that
+              * subsequent cpus can compare against it.
+              */
+             if (!idle_states)
+                     idle_states = curr_idle_states;
+
+             /*
+              * If idle states phandles are not equal, remove the
+              * cpu from the driver mask since a CPUidle driver
+              * is only capable of managing uniform idle states.
+              *
+              * Comparison works also when idle_states and
+              * curr_idle_states are the same property, since
+              * they can be == NULL so the cpu must be removed from
+              * the driver mask in that case too (ie cpu has no idle
+              * states).
+              */
+             if (!idle_states_cmp(idle_states, curr_idle_states))
+                     cpumask_clear_cpu(cpu, drv->cpumask);
+     }
+
+     /*
+      *  If there are no valid states for this driver we rely on arch
+      *  default idle behaviour, bail out
+      */
+     if (!idle_states)
+             return -ENODEV;
 
      /*
       * Initialize idle states data, starting at index 1.

@@ -117,7 +173,7 @@ static int __init arm_idle_init(void)
       * Call arch CPU operations in order to initialize
       * idle states suspend back-end specific data
       */
-     for_each_possible_cpu(cpu) {
+     for_each_cpu(cpu, drv->cpumask) {
              ret = arm_cpuidle_init(cpu);
 
              /*

@@ -157,7 +213,77 @@ out_fail:
      }
 
      cpuidle_unregister_driver(drv);
+     return ret;
+}
+
+/*
+ * arm_idle_init
+ *
+ * Registers the arm specific cpuidle driver(s) with the cpuidle
+ * framework. It relies on core code to parse the idle states
+ * and initialize them using driver data structures accordingly.
+ */
+static int __init arm_idle_init(void)
+{
+     int i, ret = -ENODEV;
+     struct cpuidle_driver *drv;
+     cpumask_var_t tmpmask;
+
+     /*
+      * These drivers require DT idle states to be present.
+      * If no idle states are detected there is no reason to
+      * proceed any further hence we return early.
+      */
+     if (!of_find_node_by_name(NULL, "idle-states"))
+             return -ENODEV;
+
+     if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
+             return -ENOMEM;
+
+     /*
+      * We need to vet idle states to create CPUidle drivers
+      * that share a common set of them. Create a tmp mask
+      * that we use to keep track of initialized cpus.
+      * Start off by initializing the mask with all possible
+      * cpus, we clear it as we go, till either all cpus
+      * have a CPUidle driver initialized or there are some
+      * CPUs that have no idle states or a parsing error
+      * occurs.
+      */
+     cpumask_copy(tmpmask, cpu_possible_mask);
+
+     for (i = 0; !cpumask_empty(tmpmask); i++) {
+             if (i == ARM_CPUIDLE_MAX_DRIVERS) {
+                     pr_warn("number of drivers exceeding static allocation\n");
+                     break;
+             }
+
+             drv = &arm_idle_drivers[i];
+             drv->cpumask = kzalloc(cpumask_size(), GFP_KERNEL);
+             if (!drv->cpumask) {
+                     ret = -ENOMEM;
+                     break;
+             }
+             /*

Quoted 8 lines

drivers/cpuidle/cpuidle-arm.c out_fail: · line 271

Quoted 4 lines

drivers/cpuidle/cpuidle-arm.c +4 -0

+              * Force driver mask, arm_idle_init_driver()
+              * will tweak it by vetting idle states.
+              */
+             cpumask_copy(drv->cpumask, tmpmask);

Why do you need to copy tmpmask ? Isn't simpler to have a zero-ed cpumask and let the arm_idle_init_driver function to set a bit instead of clearing it ?

No, because we need to keep track of logical cpus that have been already parsed, we need a tmpmask to keep track of that, how could we do it otherwise ? We can have more than two cpumask sets.

Quoted 13 lines

drivers/cpuidle/cpuidle-arm.c out_fail: · line 283

Quoted 12 lines

drivers/cpuidle/cpuidle-arm.c +11 -0

+             ret = arm_idle_init_driver(drv);
+             if (ret) {
+                     kfree(drv->cpumask);
+                     break;
+             }
+             /*
+              * Remove the cpus that were part of the registered
+              * driver from the mask of cpus to be initialized
+              * and restart.
+              */
+             cpumask_andnot(tmpmask, tmpmask, drv->cpumask);

Quoted 4 lines

If there is a DT definition with just a cluster with the cpuidle driver set and another one without any idle state, we will have always a pr_warn because i == ARM_CPUIDLE_MAX_DRIVERS due to tmpmask being never equal to a zero mask. Is it the purpose to detect such cases ?

Not really, because arm_idle_init_driver() would return -ENODEV when it detects cpus with no idle states and the loop will break before warning.

If we had two cluster of cpus with an idle-states set per cluster plus some cpus with no idle states the warning would be hit, because in actual facts we have more cpuidle sets than drivers (I know, a cpu set with no idle states falls back to default idle, but I think that's a detail that is easy to sort out).

I can create idle drivers dynamically so that the ARM_CPUIDLE_MAX_DRIVERS check can be removed or relaxed (but the code becomes slightly more complex).

Thanks for having a look, apart from these comments do we think it is a reasonable approach ?

Quoted 189 lines · Lorenzo Pieralisi

Quoted 1 line · Daniel Lezcano

On 04/16/2015 06:10 PM, Lorenzo Pieralisi wrote:

<snip>

Quoted 11 lines

Quoted 8 lines

/*
- * arm_idle_init
+ * Compare idle states phandle properties
*
- * Registers the arm specific cpuidle driver with the cpuidle
- * framework. It relies on core code to parse the idle states
- * and initialize them using driver data structures accordingly.
+ * Return true if properties are valid and equal, false otherwise

Just a detail. Would be more consistent to return zero when valid and equal ?

Consistent with memcmp you mean ? Yes, I can change it.

Quoted 130 lines

Quoted 130 lines

   */

drivers/cpuidle/cpuidle-arm.c +109 -4

-static int __init arm_idle_init(void)
+static bool __init idle_states_cmp(struct property *states1,
+                                struct property *states2)
+{
+     /*
+      * NB: Implemented through code from drivers/of/unittest.c
+      *     Function is generic and can be moved to generic OF code
+      *     if needed
+      */
+     return states1 && states2 &&
+            (states1->length == states2->length) &&
+            states1->value && states2->value &&
+            !memcmp(states1->value, states2->value, states1->length);
+}
+
+static int __init arm_idle_init_driver(struct cpuidle_driver *drv)
  {
-     int cpu, ret;
-     struct cpuidle_driver *drv = &arm_idle_driver;
+     int ret, cpu;
      struct cpuidle_device *dev;
+     struct property *curr_idle_states, *idle_states = NULL;
+     struct device_node *cpu_node;
+
+     for_each_cpu(cpu, drv->cpumask) {
+             cpu_node = of_cpu_device_node_get(cpu);
+             curr_idle_states = of_find_property(cpu_node,
+                                                 "cpu-idle-states", NULL);
+             of_node_put(cpu_node);
+
+             /*
+              * Stash the first valid idle states phandle in the cpumask.
+              * If curr_idle_states is NULL assigning it to idle_states
+              * is harmless and it is managed by idle states comparison
+              * code. Keep track of first valid phandle so that
+              * subsequent cpus can compare against it.
+              */
+             if (!idle_states)
+                     idle_states = curr_idle_states;
+
+             /*
+              * If idle states phandles are not equal, remove the
+              * cpu from the driver mask since a CPUidle driver
+              * is only capable of managing uniform idle states.
+              *
+              * Comparison works also when idle_states and
+              * curr_idle_states are the same property, since
+              * they can be == NULL so the cpu must be removed from
+              * the driver mask in that case too (ie cpu has no idle
+              * states).
+              */
+             if (!idle_states_cmp(idle_states, curr_idle_states))
+                     cpumask_clear_cpu(cpu, drv->cpumask);
+     }
+
+     /*
+      *  If there are no valid states for this driver we rely on arch
+      *  default idle behaviour, bail out
+      */
+     if (!idle_states)
+             return -ENODEV;
 
      /*
       * Initialize idle states data, starting at index 1.

@@ -117,7 +173,7 @@ static int __init arm_idle_init(void)
       * Call arch CPU operations in order to initialize
       * idle states suspend back-end specific data
       */
-     for_each_possible_cpu(cpu) {
+     for_each_cpu(cpu, drv->cpumask) {
              ret = arm_cpuidle_init(cpu);
 
              /*

@@ -157,7 +213,77 @@ out_fail:
      }
 
      cpuidle_unregister_driver(drv);
+     return ret;
+}
+
+/*
+ * arm_idle_init
+ *
+ * Registers the arm specific cpuidle driver(s) with the cpuidle
+ * framework. It relies on core code to parse the idle states
+ * and initialize them using driver data structures accordingly.
+ */
+static int __init arm_idle_init(void)
+{
+     int i, ret = -ENODEV;
+     struct cpuidle_driver *drv;
+     cpumask_var_t tmpmask;
+
+     /*
+      * These drivers require DT idle states to be present.
+      * If no idle states are detected there is no reason to
+      * proceed any further hence we return early.
+      */
+     if (!of_find_node_by_name(NULL, "idle-states"))
+             return -ENODEV;
+
+     if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
+             return -ENOMEM;
+
+     /*
+      * We need to vet idle states to create CPUidle drivers
+      * that share a common set of them. Create a tmp mask
+      * that we use to keep track of initialized cpus.
+      * Start off by initializing the mask with all possible
+      * cpus, we clear it as we go, till either all cpus
+      * have a CPUidle driver initialized or there are some
+      * CPUs that have no idle states or a parsing error
+      * occurs.
+      */
+     cpumask_copy(tmpmask, cpu_possible_mask);
+
+     for (i = 0; !cpumask_empty(tmpmask); i++) {
+             if (i == ARM_CPUIDLE_MAX_DRIVERS) {
+                     pr_warn("number of drivers exceeding static allocation\n");
+                     break;
+             }
+
+             drv = &arm_idle_drivers[i];
+             drv->cpumask = kzalloc(cpumask_size(), GFP_KERNEL);
+             if (!drv->cpumask) {
+                     ret = -ENOMEM;
+                     break;
+             }
+             /*

Quoted 8 lines

drivers/cpuidle/cpuidle-arm.c out_fail: · line 271

Quoted 4 lines

drivers/cpuidle/cpuidle-arm.c +4 -0

+              * Force driver mask, arm_idle_init_driver()
+              * will tweak it by vetting idle states.
+              */
+             cpumask_copy(drv->cpumask, tmpmask);

Why do you need to copy tmpmask ? Isn't simpler to have a zero-ed cpumask and let the arm_idle_init_driver function to set a bit instead of clearing it ?

No, because we need to keep track of logical cpus that have been already parsed, we need a tmpmask to keep track of that, how could we do it otherwise ? We can have more than two cpumask sets.

Quoted 13 lines

drivers/cpuidle/cpuidle-arm.c out_fail: · line 283

Quoted 12 lines

drivers/cpuidle/cpuidle-arm.c +11 -0

+             ret = arm_idle_init_driver(drv);
+             if (ret) {
+                     kfree(drv->cpumask);
+                     break;
+             }
+             /*
+              * Remove the cpus that were part of the registered
+              * driver from the mask of cpus to be initialized
+              * and restart.
+              */
+             cpumask_andnot(tmpmask, tmpmask, drv->cpumask);

Quoted 4 lines

If there is a DT definition with just a cluster with the cpuidle driver set and another one without any idle state, we will have always a pr_warn because i == ARM_CPUIDLE_MAX_DRIVERS due to tmpmask being never equal to a zero mask. Is it the purpose to detect such cases ?

Not really, because arm_idle_init_driver() would return -ENODEV when it detects cpus with no idle states and the loop will break before warning.

If we had two cluster of cpus with an idle-states set per cluster plus some cpus with no idle states the warning would be hit, because in actual facts we have more cpuidle sets than drivers (I know, a cpu set with no idle states falls back to default idle, but I think that's a detail that is easy to sort out).

Quoted 5 lines

I can create idle drivers dynamically so that the ARM_CPUIDLE_MAX_DRIVERS check can be removed or relaxed (but the code becomes slightly more complex).

Thanks for having a look, apart from these comments do we think it is a reasonable approach ?

Any further comments ? I can post a v2 with an updated idle_states_cmp() return value, if we think the multiple drivers approach is valid.

From the silence on the list, I guess there is no strong objection to

Thanks a lot,
Lorenzo

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Quoted 2 lines

Any further comments ? I can post a v2 with an updated idle_states_cmp() return value, if we think the multiple drivers approach is valid.

Can you post v2?
From the silence on the list, I guess there is no strong objection to your approach.

So, perhaps it is time to remove the "RFC" as well so this can get on track to be merged.

On 04/16/2015 06:10 PM, Lorenzo Pieralisi wrote:

Quoted 8 lines

/*
- * arm_idle_init
+ * Compare idle states phandle properties
*
- * Registers the arm specific cpuidle driver with the cpuidle
- * framework. It relies on core code to parse the idle states
- * and initialize them using driver data structures accordingly.
+ * Return true if properties are valid and equal, false otherwise

Just a detail. Would be more consistent to return zero when valid and equal ?

Quoted 124 lines

   */

drivers/cpuidle/cpuidle-arm.c +103 -4

-static int __init arm_idle_init(void)
+static bool __init idle_states_cmp(struct property *states1,
+                                struct property *states2)
+{
+     /*
+      * NB: Implemented through code from drivers/of/unittest.c
+      *     Function is generic and can be moved to generic OF code
+      *     if needed
+      */
+     return states1 && states2 &&
+            (states1->length == states2->length) &&
+            states1->value && states2->value &&
+            !memcmp(states1->value, states2->value, states1->length);
+}
+
+static int __init arm_idle_init_driver(struct cpuidle_driver *drv)
  {
-     int cpu, ret;
-     struct cpuidle_driver *drv = &arm_idle_driver;
+     int ret, cpu;
      struct cpuidle_device *dev;
+     struct property *curr_idle_states, *idle_states = NULL;
+     struct device_node *cpu_node;
+
+     for_each_cpu(cpu, drv->cpumask) {
+             cpu_node = of_cpu_device_node_get(cpu);
+             curr_idle_states = of_find_property(cpu_node,
+                                                 "cpu-idle-states", NULL);
+             of_node_put(cpu_node);
+
+             /*
+              * Stash the first valid idle states phandle in the cpumask.
+              * If curr_idle_states is NULL assigning it to idle_states
+              * is harmless and it is managed by idle states comparison
+              * code. Keep track of first valid phandle so that
+              * subsequent cpus can compare against it.
+              */
+             if (!idle_states)
+                     idle_states = curr_idle_states;
+
+             /*
+              * If idle states phandles are not equal, remove the
+              * cpu from the driver mask since a CPUidle driver
+              * is only capable of managing uniform idle states.
+              *
+              * Comparison works also when idle_states and
+              * curr_idle_states are the same property, since
+              * they can be == NULL so the cpu must be removed from
+              * the driver mask in that case too (ie cpu has no idle
+              * states).
+              */
+             if (!idle_states_cmp(idle_states, curr_idle_states))
+                     cpumask_clear_cpu(cpu, drv->cpumask);
+     }
+
+     /*
+      *  If there are no valid states for this driver we rely on arch
+      *  default idle behaviour, bail out
+      */
+     if (!idle_states)
+             return -ENODEV;
 
      /*
       * Initialize idle states data, starting at index 1.

@@ -117,7 +173,7 @@ static int __init arm_idle_init(void)
       * Call arch CPU operations in order to initialize
       * idle states suspend back-end specific data
       */
-     for_each_possible_cpu(cpu) {
+     for_each_cpu(cpu, drv->cpumask) {
              ret = arm_cpuidle_init(cpu);
 
              /*

@@ -157,7 +213,77 @@ out_fail:
      }
 
      cpuidle_unregister_driver(drv);
+     return ret;
+}
+
+/*
+ * arm_idle_init
+ *
+ * Registers the arm specific cpuidle driver(s) with the cpuidle
+ * framework. It relies on core code to parse the idle states
+ * and initialize them using driver data structures accordingly.
+ */
+static int __init arm_idle_init(void)
+{
+     int i, ret = -ENODEV;
+     struct cpuidle_driver *drv;
+     cpumask_var_t tmpmask;
+
+     /*
+      * These drivers require DT idle states to be present.
+      * If no idle states are detected there is no reason to
+      * proceed any further hence we return early.
+      */
+     if (!of_find_node_by_name(NULL, "idle-states"))
+             return -ENODEV;
+
+     if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
+             return -ENOMEM;
+
+     /*
+      * We need to vet idle states to create CPUidle drivers
+      * that share a common set of them. Create a tmp mask
+      * that we use to keep track of initialized cpus.
+      * Start off by initializing the mask with all possible
+      * cpus, we clear it as we go, till either all cpus
+      * have a CPUidle driver initialized or there are some
+      * CPUs that have no idle states or a parsing error
+      * occurs.
+      */
+     cpumask_copy(tmpmask, cpu_possible_mask);
+
+     for (i = 0; !cpumask_empty(tmpmask); i++) {
+             if (i == ARM_CPUIDLE_MAX_DRIVERS) {
+                     pr_warn("number of drivers exceeding static allocation\n");
+                     break;
+             }
+
+             drv = &arm_idle_drivers[i];

drivers/cpuidle/cpuidle-arm.c out_fail: · line 271

Quoted 10 lines

drivers/cpuidle/cpuidle-arm.c +10 -0

+             drv->cpumask = kzalloc(cpumask_size(), GFP_KERNEL);
+             if (!drv->cpumask) {
+                     ret = -ENOMEM;
+                     break;
+             }
+             /*
+              * Force driver mask, arm_idle_init_driver()
+              * will tweak it by vetting idle states.
+              */
+             cpumask_copy(drv->cpumask, tmpmask);

Why do you need to copy tmpmask ? Isn't simpler to have a zero-ed cpumask and let the arm_idle_init_driver function to set a bit instead of clearing it ?

drivers/cpuidle/cpuidle-arm.c out_fail: · line 283

Quoted 11 lines

drivers/cpuidle/cpuidle-arm.c +11 -0

+             ret = arm_idle_init_driver(drv);
+             if (ret) {
+                     kfree(drv->cpumask);
+                     break;
+             }
+             /*
+              * Remove the cpus that were part of the registered
+              * driver from the mask of cpus to be initialized
+              * and restart.
+              */
+             cpumask_andnot(tmpmask, tmpmask, drv->cpumask);

If there is a DT definition with just a cluster with the cpuidle driver set and another one without any idle state, we will have always a pr_warn because i == ARM_CPUIDLE_MAX_DRIVERS due to tmpmask being never equal to a zero mask. Is it the purpose to detect such cases ?

drivers/cpuidle/cpuidle-arm.c out_fail: · line 289

Quoted 7 lines

drivers/cpuidle/cpuidle-arm.c +2 -0

+     }
 
+     free_cpumask_var(tmpmask);
      return ret;
  }
  device_initcall(arm_idle_init);

--
<http://www.linaro.org/> Linaro.org ??? Open source software for ARM SoCs

Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog

This patch sets the maximum number of static CPUidle drivers allowed to two, since it is hard to foresee systems with more than two sets of CPUs having different idle states

On Thu, Apr 16, 2015 at 9:10 AM, Lorenzo Pieralisi <[email protected]> wrote:

[...]

Quoted 3 lines

This patch sets the maximum number of static CPUidle drivers allowed to two, since it is hard to foresee systems with more than two sets of CPUs having different idle states

It's not hard to forsee anymore:

   http://www.cnx-software.com/2015/04/21/mediatek-helio-x20-deca-core-processor/
   https://liliputing.com/2015/04/closer-look-at-mediateks-upcoming-10-core-helio-x20-processor.html

Quoted 11 lines · Kevin Hilman

On Thu, Apr 16, 2015 at 9:10 AM, Lorenzo Pieralisi <[email protected]> wrote:

[...]

Quoted 3 lines

This patch sets the maximum number of static CPUidle drivers allowed to two, since it is hard to foresee systems with more than two sets of CPUs having different idle states

It's not hard to forsee anymore:

Quoted 2 lines

   http://www.cnx-software.com/2015/04/21/mediatek-helio-x20-deca-core-processor/
   https://liliputing.com/2015/04/closer-look-at-mediateks-upcoming-10-core-helio-x20-processor.html

Eheh I stand corrected :). Not a big deal, drivers can be allocated dynamically; it seems that the designs above leave us no choice, the idle states are the only differentiating factor when it comes to CPUidle driver logical cpumask creation, happy to hear more opinions on this, but I do not really see what we can use instead.

This patch is a v2 of a previous posting:

http://lists.infradead.org/pipermail/linux-arm-kernel/2015-March/328192.html

v1 -> v2:

- Rebased against 4.0
- Removed DT affinity probe interface
- Improved and optimized idle states comparison and cpumask creation
- Merged DT idle states vetting code in the ARM CPUidle driver

This patch aims at enabling the generic CPUidle driver on ARM systems with heterogeneous idle states. In the CPUidle context the kernel considers cpus equal (and group them together through a CPUidle driver) if they share the same set of idle states.
Since the idle states represent the differentiation point when it comes to creating CPUidle drivers (their states and related cpu logical mask), the idle states can be used as a means to create logical masks of cpus for which the CPUidle driver applies.

This patch sets the maximum number of static CPUidle drivers allowed to two, since it is hard to foresee systems with more than two sets of CPUs having different idle states; it is meant to keep allocation simple while keeping functionality.

The code carrying out idle states comparison is completely generic and borrowed from OF unit tests; it overrides the OF API entirely though to carry out properties (ie phandles list) comparison in the fastest possible way. If deemed worthwhile it can be moved to generic OF layer.

Tested through several dts idle states permutations on Juno board with CPUidle and related idle states enabled.