kgsl: convert some workqueues to use kthreads

adreno_dispatch_work and _kgsl_event_worker are both low-latency
low-runtime functions that are in the critical path of GPU
rendering. Moving them out of workqueues and into a dedicated FIFO
kthread avoids significant jitter.

bug 30342017

Change-Id: I5feb4e829064d422b4b9af2acb449afd1f981899

5 files changed, 26 insertions, 9 deletions

diff --git a/drivers/gpu/msm/adreno_dispatch.c b/drivers/gpu/msm/adreno_dispatch.c
index 9e526fb329a..0c3134dcd86 100644
--- a/drivers/gpu/msm/adreno_dispatch.c
+++ b/drivers/gpu/msm/adreno_dispatch.c
@@ -1879,7 +1879,7 @@ static int adreno_dispatch_process_cmdqueue(struct adreno_device *adreno_dev,
  *
  * Process expired commands and send new ones.
  */
-static void adreno_dispatcher_work(struct work_struct *work)
+static void adreno_dispatcher_work(struct kthread_work *work)
 {
 	struct adreno_dispatcher *dispatcher =
 		container_of(work, struct adreno_dispatcher, work);
@@ -1954,7 +1954,7 @@ void adreno_dispatcher_schedule(struct kgsl_device *device)
 	struct adreno_device *adreno_dev = ADRENO_DEVICE(device);
 	struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher;
 
-	queue_work(device->work_queue, &dispatcher->work);
+	queue_kthread_work(&kgsl_driver.worker, &dispatcher->work);
 }
 
 /**
@@ -2249,7 +2249,7 @@ int adreno_dispatcher_init(struct adreno_device *adreno_dev)
 	if (adreno_is_a304(adreno_dev))
 		_fault_timer_interval = 400;
 
-	INIT_WORK(&dispatcher->work, adreno_dispatcher_work);
+	init_kthread_work(&dispatcher->work, adreno_dispatcher_work);
 
 	init_completion(&dispatcher->idle_gate);
 	complete_all(&dispatcher->idle_gate);
diff --git a/drivers/gpu/msm/adreno_dispatch.h b/drivers/gpu/msm/adreno_dispatch.h
index c2721ad0c88..d7637146fe9 100644
--- a/drivers/gpu/msm/adreno_dispatch.h
+++ b/drivers/gpu/msm/adreno_dispatch.h
@@ -72,7 +72,7 @@ struct adreno_dispatcher {
 	atomic_t fault;
 	struct plist_head pending;
 	spinlock_t plist_lock;
-	struct work_struct work;
+	struct kthread_work work;
 	struct kobject kobj;
 	struct completion idle_gate;
 };
diff --git a/drivers/gpu/msm/kgsl.c b/drivers/gpu/msm/kgsl.c
index 2a92bf5a951..52b246687eb 100644
--- a/drivers/gpu/msm/kgsl.c
+++ b/drivers/gpu/msm/kgsl.c
@@ -4816,6 +4816,8 @@ static void kgsl_core_exit(void)
 static int __init kgsl_core_init(void)
 {
 	int result = 0;
+	struct sched_param param = { .sched_priority = 2 };
+
 	/* alloc major and minor device numbers */
 	result = alloc_chrdev_region(&kgsl_driver.major, 0, KGSL_DEVICE_MAX,
 		"kgsl");
@@ -4877,6 +4879,18 @@ static int __init kgsl_core_init(void)
 
 	kgsl_mmu_set_mmutype(ksgl_mmu_type);
 
+	init_kthread_worker(&kgsl_driver.worker);
+
+	kgsl_driver.worker_thread = kthread_run(kthread_worker_fn,
+		&kgsl_driver.worker, "kgsl_worker_thread");
+
+	if (IS_ERR(kgsl_driver.worker_thread)) {
+		pr_err("unable to start kgsl thread\n");
+		goto err;
+	}
+
+	sched_setscheduler(kgsl_driver.worker_thread, SCHED_FIFO, &param);
+
 	kgsl_events_init();
 
 	/* create the memobjs kmem cache */
diff --git a/drivers/gpu/msm/kgsl.h b/drivers/gpu/msm/kgsl.h
index 661119b8c1c..f0b66b7f496 100644
--- a/drivers/gpu/msm/kgsl.h
+++ b/drivers/gpu/msm/kgsl.h
@@ -24,6 +24,7 @@
 #include <linux/cdev.h>
 #include <linux/regulator/consumer.h>
 #include <linux/mm.h>
+#include <linux/kthread.h>
 
 /* The number of memstore arrays limits the number of contexts allowed.
  * If more contexts are needed, update multiple for MEMSTORE_SIZE
@@ -94,6 +95,8 @@ struct kgsl_driver {
 		unsigned int mapped_max;
 	} stats;
 	unsigned int full_cache_threshold;
+	struct kthread_worker worker;
+	struct task_struct *worker_thread;
 };
 
 extern struct kgsl_driver kgsl_driver;
@@ -220,7 +223,7 @@ struct kgsl_event {
 	void *priv;
 	struct list_head node;
 	unsigned int created;
-	struct work_struct work;
+	struct kthread_work work;
 	int result;
 	struct kgsl_event_group *group;
 };
diff --git a/drivers/gpu/msm/kgsl_events.c b/drivers/gpu/msm/kgsl_events.c
index 7f2f21e0d62..90ff200184d 100644
--- a/drivers/gpu/msm/kgsl_events.c
+++ b/drivers/gpu/msm/kgsl_events.c
@@ -32,7 +32,7 @@ static inline void signal_event(struct kgsl_device *device,
 {
 	list_del(&event->node);
 	event->result = result;
-	queue_work(device->events_wq, &event->work);
+	queue_kthread_work(&kgsl_driver.worker, &event->work);
 }
 
 /**
@@ -42,7 +42,7 @@ static inline void signal_event(struct kgsl_device *device,
  * Each event callback has its own work struct and is run on a event specific
  * workqeuue.  This is the worker that queues up the event callback function.
  */
-static void _kgsl_event_worker(struct work_struct *work)
+static void _kgsl_event_worker(struct kthread_work *work)
 {
 	struct kgsl_event *event = container_of(work, struct kgsl_event, work);
 	int id = KGSL_CONTEXT_ID(event->context);
@@ -238,7 +238,7 @@ int kgsl_add_event(struct kgsl_device *device, struct kgsl_event_group *group,
 	event->created = jiffies;
 	event->group = group;
 
-	INIT_WORK(&event->work, _kgsl_event_worker);
+	init_kthread_work(&event->work, _kgsl_event_worker);
 
 	trace_kgsl_register_event(KGSL_CONTEXT_ID(context), timestamp, func);
 
@@ -253,7 +253,7 @@ int kgsl_add_event(struct kgsl_device *device, struct kgsl_event_group *group,
 
 	if (timestamp_cmp(retired, timestamp) >= 0) {
 		event->result = KGSL_EVENT_RETIRED;
-		queue_work(device->events_wq, &event->work);
+		queue_kthread_work(&kgsl_driver.worker, &event->work);
 		spin_unlock(&group->lock);
 		return 0;
 	}