QoS: Enhance framework to support cpu/irq specific QoS requests

QoS request for CPU_DMA_LATENCY can be better optimized if the request
can be set only for the required cpus and not all cpus. This helps save
power on other cores, while still gauranteeing the quality of service.

Enhance the QoS constraints data structures to support target value for
each core. Requests specify if the QoS is applicable to all cores
(default) or to a selective subset of the cores or to a core(s), that the
IRQ is affine to.

QoS requests that need to track an IRQ can be set to apply only on the
cpus to which the IRQ's smp_affinity attribute is set to. The QoS
framework will automatically track IRQ migration between the cores. The
QoS is updated to be applied only to the core(s) that the IRQ has been
migrated to.

Idle and interested drivers can request a PM QoS value for a constraint
across all cpus, or a specific cpu or a set of cpus. Separate APIs have
been added to request for individual cpu or a cpumask.  The default
behaviour of PM QoS is maintained i.e, requests that do not specify a
type of the request will continue to be effected on all cores.  Requests
that want to specify an affinity of cpu(s) or an irq, can modify the PM
QoS request data structures by specifying the type of the request and
either the mask of the cpus or the IRQ number depending on the type.
Updating the request does not reset the type of the request.

The userspace sysfs interface does not support CPU/IRQ affinity.

Change-Id: I09ae85a1e8585d44440e86d63504ad734e8e3e36
Signed-off-by: Lina Iyer <ilina@codeaurora.org>

Documentation/power/pm_qos_interface.txt

@@ -43,6 +43,17 @@ registered notifiers are called only if the target value is now different.
 Clients of pm_qos need to save the returned handle for future use in other
 pm_qos API functions.
 
+The handle is a pm_qos_request object. By default the request object sets the
+request type to PM_QOS_REQ_ALL_CORES, in which case, the PM QoS request
+applies to all cores. However, the driver can also specify a request type to
+be either of
+        PM_QOS_REQ_ALL_CORES,
+        PM_QOS_REQ_AFFINE_CORES,
+        PM_QOS_REQ_AFFINE_IRQ,
+
+Specify the cpumask when type is set to PM_QOS_REQ_AFFINE_CORES and specify
+the IRQ number with PM_QOS_REQ_AFFINE_IRQ.
+
 void pm_qos_update_request(handle, new_target_value):
 Will update the list element pointed to by the handle with the new target value
 and recompute the new aggregated target, calling the notification tree if the
@@ -56,6 +67,13 @@ the request.
 int pm_qos_request(param_class):
 Returns the aggregated value for a given PM QoS class.
 
+int pm_qos_request_for_cpu(param_class, cpu):
+Returns the aggregated value for a given PM QoS class for the specified cpu.
+
+int pm_qos_request_for_cpumask(param_class, cpumask):
+Returns the aggregated value for a given PM QoS class for the specified
+cpumask.
+
 int pm_qos_request_active(handle):
 Returns if the request is still active, i.e. it has not been removed from a
 PM QoS class constraints list.

include/linux/pm_qos.h

@@ -9,6 +9,8 @@
 #include <linux/notifier.h>
 #include <linux/device.h>
 #include <linux/workqueue.h>
+#include <linux/cpumask.h>
+#include <linux/interrupt.h>
 
 enum {
 	PM_QOS_RESERVED = 0,
@@ -44,7 +46,22 @@ enum pm_qos_flags_status {
 
 #define PM_QOS_FLAG_NO_POWER_OFF	(1 << 0)
 
+enum pm_qos_req_type {
+	PM_QOS_REQ_ALL_CORES = 0,
+	PM_QOS_REQ_AFFINE_CORES,
+#ifdef CONFIG_SMP
+	PM_QOS_REQ_AFFINE_IRQ,
+#endif
+};
+
 struct pm_qos_request {
+	enum pm_qos_req_type type;
+	struct cpumask cpus_affine;
+#ifdef CONFIG_SMP
+	uint32_t irq;
+	/* Internal structure members */
+	struct irq_affinity_notify irq_notify;
+#endif
 	struct plist_node node;
 	int pm_qos_class;
 	struct delayed_work work; /* for pm_qos_update_request_timeout */
@@ -85,6 +102,7 @@ enum pm_qos_type {
 struct pm_qos_constraints {
 	struct plist_head list;
 	s32 target_value;	/* Do not change to 64 bit */
+	s32 target_per_cpu[NR_CPUS];
 	s32 default_value;
 	s32 no_constraint_value;
 	enum pm_qos_type type;
@@ -131,6 +149,8 @@ void pm_qos_update_request_timeout(struct pm_qos_request *req,
 void pm_qos_remove_request(struct pm_qos_request *req);
 
 int pm_qos_request(int pm_qos_class);
+int pm_qos_request_for_cpu(int pm_qos_class, int cpu);
+int pm_qos_request_for_cpumask(int pm_qos_class, struct cpumask *mask);
 int pm_qos_add_notifier(int pm_qos_class, struct notifier_block *notifier);
 int pm_qos_remove_notifier(int pm_qos_class, struct notifier_block *notifier);
 int pm_qos_request_active(struct pm_qos_request *req);

kernel/power/qos.c

@@ -43,6 +43,8 @@
 #include <linux/kernel.h>
 #include <linux/debugfs.h>
 #include <linux/seq_file.h>
+#include <linux/irq.h>
+#include <linux/irqdesc.h>
 
 #include <linux/uaccess.h>
 #include <linux/export.h>
@@ -67,6 +69,8 @@ static BLOCKING_NOTIFIER_HEAD(cpu_dma_lat_notifier);
 static struct pm_qos_constraints cpu_dma_constraints = {
 	.list = PLIST_HEAD_INIT(cpu_dma_constraints.list),
 	.target_value = PM_QOS_CPU_DMA_LAT_DEFAULT_VALUE,
+	.target_per_cpu = { [0 ... (NR_CPUS - 1)] =
+				PM_QOS_CPU_DMA_LAT_DEFAULT_VALUE },
 	.default_value = PM_QOS_CPU_DMA_LAT_DEFAULT_VALUE,
 	.no_constraint_value = PM_QOS_CPU_DMA_LAT_DEFAULT_VALUE,
 	.type = PM_QOS_MIN,
@@ -81,6 +85,8 @@ static BLOCKING_NOTIFIER_HEAD(network_lat_notifier);
 static struct pm_qos_constraints network_lat_constraints = {
 	.list = PLIST_HEAD_INIT(network_lat_constraints.list),
 	.target_value = PM_QOS_NETWORK_LAT_DEFAULT_VALUE,
+	.target_per_cpu = { [0 ... (NR_CPUS - 1)] =
+				PM_QOS_NETWORK_LAT_DEFAULT_VALUE },
 	.default_value = PM_QOS_NETWORK_LAT_DEFAULT_VALUE,
 	.no_constraint_value = PM_QOS_NETWORK_LAT_DEFAULT_VALUE,
 	.type = PM_QOS_MIN,
@@ -91,11 +97,12 @@ static struct pm_qos_object network_lat_pm_qos = {
 	.name = "network_latency",
 };
 
-
 static BLOCKING_NOTIFIER_HEAD(network_throughput_notifier);
 static struct pm_qos_constraints network_tput_constraints = {
 	.list = PLIST_HEAD_INIT(network_tput_constraints.list),
 	.target_value = PM_QOS_NETWORK_THROUGHPUT_DEFAULT_VALUE,
+	.target_per_cpu = { [0 ... (NR_CPUS - 1)] =
+				PM_QOS_NETWORK_THROUGHPUT_DEFAULT_VALUE },
 	.default_value = PM_QOS_NETWORK_THROUGHPUT_DEFAULT_VALUE,
 	.no_constraint_value = PM_QOS_NETWORK_THROUGHPUT_DEFAULT_VALUE,
 	.type = PM_QOS_MAX,
@@ -258,6 +265,33 @@ static const struct file_operations pm_qos_debug_fops = {
 	.release        = single_release,
 };
 
+static inline void pm_qos_set_value_for_cpus(struct pm_qos_constraints *c)
+{
+	struct pm_qos_request *req = NULL;
+	int cpu;
+	s32 qos_val[NR_CPUS] = { [0 ... (NR_CPUS - 1)] = c->default_value };
+
+	plist_for_each_entry(req, &c->list, node) {
+		for_each_cpu(cpu, &req->cpus_affine) {
+			switch (c->type) {
+			case PM_QOS_MIN:
+				if (qos_val[cpu] > req->node.prio)
+					qos_val[cpu] = req->node.prio;
+				break;
+			case PM_QOS_MAX:
+				if (req->node.prio > qos_val[cpu])
+					qos_val[cpu] = req->node.prio;
+				break;
+			default:
+				break;
+			}
+		}
+	}
+
+	for_each_possible_cpu(cpu)
+		c->target_per_cpu[cpu] = qos_val[cpu];
+}
+
 /**
  * pm_qos_update_target - manages the constraints list and calls the notifiers
  *  if needed
@@ -306,6 +340,7 @@ int pm_qos_update_target(struct pm_qos_constraints *c, struct plist_node *node,
 
 	curr_value = pm_qos_get_value(c);
 	pm_qos_set_value(c, curr_value);
+	pm_qos_set_value_for_cpus(c);
 
 	spin_unlock_irqrestore(&pm_qos_lock, flags);
 
@@ -399,12 +434,49 @@ int pm_qos_request(int pm_qos_class)
 }
 EXPORT_SYMBOL_GPL(pm_qos_request);
 
+int pm_qos_request_for_cpu(int pm_qos_class, int cpu)
+{
+	return pm_qos_array[pm_qos_class]->constraints->target_per_cpu[cpu];
+}
+EXPORT_SYMBOL(pm_qos_request_for_cpu);
+
 int pm_qos_request_active(struct pm_qos_request *req)
 {
 	return req->pm_qos_class != 0;
 }
 EXPORT_SYMBOL_GPL(pm_qos_request_active);
 
+int pm_qos_request_for_cpumask(int pm_qos_class, struct cpumask *mask)
+{
+	unsigned long irqflags;
+	int cpu;
+	struct pm_qos_constraints *c = NULL;
+	int val;
+
+	spin_lock_irqsave(&pm_qos_lock, irqflags);
+	c = pm_qos_array[pm_qos_class]->constraints;
+	val = c->default_value;
+
+	for_each_cpu(cpu, mask) {
+		switch (c->type) {
+		case PM_QOS_MIN:
+			if (c->target_per_cpu[cpu] < val)
+				val = c->target_per_cpu[cpu];
+			break;
+		case PM_QOS_MAX:
+			if (c->target_per_cpu[cpu] > val)
+				val = c->target_per_cpu[cpu];
+			break;
+		default:
+			break;
+		}
+	}
+	spin_unlock_irqrestore(&pm_qos_lock, irqflags);
+
+	return val;
+}
+EXPORT_SYMBOL(pm_qos_request_for_cpumask);
+
 static void __pm_qos_update_request(struct pm_qos_request *req,
 			   s32 new_value)
 {
@@ -431,6 +503,42 @@ static void pm_qos_work_fn(struct work_struct *work)
 	__pm_qos_update_request(req, PM_QOS_DEFAULT_VALUE);
 }
 
+#ifdef CONFIG_SMP
+static void pm_qos_irq_release(struct kref *ref)
+{
+	unsigned long flags;
+	struct irq_affinity_notify *notify = container_of(ref,
+					struct irq_affinity_notify, kref);
+	struct pm_qos_request *req = container_of(notify,
+					struct pm_qos_request, irq_notify);
+	struct pm_qos_constraints *c =
+				pm_qos_array[req->pm_qos_class]->constraints;
+
+	spin_lock_irqsave(&pm_qos_lock, flags);
+	cpumask_setall(&req->cpus_affine);
+	spin_unlock_irqrestore(&pm_qos_lock, flags);
+
+	pm_qos_update_target(c, &req->node, PM_QOS_UPDATE_REQ,
+			c->default_value);
+}
+
+static void pm_qos_irq_notify(struct irq_affinity_notify *notify,
+		const cpumask_t *mask)
+{
+	unsigned long flags;
+	struct pm_qos_request *req = container_of(notify,
+					struct pm_qos_request, irq_notify);
+	struct pm_qos_constraints *c =
+				pm_qos_array[req->pm_qos_class]->constraints;
+
+	spin_lock_irqsave(&pm_qos_lock, flags);
+	cpumask_copy(&req->cpus_affine, mask);
+	spin_unlock_irqrestore(&pm_qos_lock, flags);
+
+	pm_qos_update_target(c, &req->node, PM_QOS_UPDATE_REQ, req->node.prio);
+}
+#endif
+
 /**
  * pm_qos_add_request - inserts new qos request into the list
  * @req: pointer to a preallocated handle
@@ -454,6 +562,51 @@ void pm_qos_add_request(struct pm_qos_request *req,
 		WARN(1, KERN_ERR "pm_qos_add_request() called for already added request\n");
 		return;
 	}
+
+	switch (req->type) {
+	case PM_QOS_REQ_AFFINE_CORES:
+		if (cpumask_empty(&req->cpus_affine)) {
+			req->type = PM_QOS_REQ_ALL_CORES;
+			cpumask_setall(&req->cpus_affine);
+			WARN(1, "Affine cores not set for request with affinity flag\n");
+		}
+		break;
+#ifdef CONFIG_SMP
+	case PM_QOS_REQ_AFFINE_IRQ:
+		if (irq_can_set_affinity(req->irq)) {
+			int ret = 0;
+			struct irq_desc *desc = irq_to_desc(req->irq);
+			struct cpumask *mask = desc->irq_data.common->affinity;
+
+			/* Get the current affinity */
+			cpumask_copy(&req->cpus_affine, mask);
+			req->irq_notify.irq = req->irq;
+			req->irq_notify.notify = pm_qos_irq_notify;
+			req->irq_notify.release = pm_qos_irq_release;
+
+			ret = irq_set_affinity_notifier(req->irq,
+					&req->irq_notify);
+			if (ret) {
+				WARN(1, "IRQ affinity notify set failed\n");
+				req->type = PM_QOS_REQ_ALL_CORES;
+				cpumask_setall(&req->cpus_affine);
+			}
+		} else {
+			req->type = PM_QOS_REQ_ALL_CORES;
+			cpumask_setall(&req->cpus_affine);
+			WARN(1, "IRQ-%d not set for request with affinity flag\n",
+					req->irq);
+		}
+		break;
+#endif
+	default:
+		WARN(1, "Unknown request type %d\n", req->type);
+		/* fall through */
+	case PM_QOS_REQ_ALL_CORES:
+		cpumask_setall(&req->cpus_affine);
+		break;
+	}
+
 	req->pm_qos_class = pm_qos_class;
 	INIT_DELAYED_WORK(&req->work, pm_qos_work_fn);
 	trace_pm_qos_add_request(pm_qos_class, value);