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
diff --git a/drivers/gpu/msm/adreno_dispatch.c b/drivers/gpu/msm/adreno_dispatch.c
index 5d3b2b8..67fd2f2 100644
--- a/drivers/gpu/msm/adreno_dispatch.c
+++ b/drivers/gpu/msm/adreno_dispatch.c
@@ -2096,7 +2096,7 @@
mutex_unlock(&device->mutex);
}
-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);
@@ -2156,7 +2156,7 @@
struct adreno_device *adreno_dev = ADRENO_DEVICE(device);
struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher;
- kgsl_schedule_work(&dispatcher->work);
+ queue_kthread_work(&kgsl_driver.worker, &dispatcher->work);
}
/**
@@ -2438,7 +2438,7 @@
setup_timer(&dispatcher->fault_timer, adreno_dispatcher_fault_timer,
(unsigned long) adreno_dev);
- 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 699c3e4..fee4317 100644
--- a/drivers/gpu/msm/adreno_dispatch.h
+++ b/drivers/gpu/msm/adreno_dispatch.h
@@ -91,7 +91,7 @@
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;
unsigned int disp_preempt_fair_sched;
diff --git a/drivers/gpu/msm/kgsl.c b/drivers/gpu/msm/kgsl.c
index 92ec8083..25f1f9d 100644
--- a/drivers/gpu/msm/kgsl.c
+++ b/drivers/gpu/msm/kgsl.c
@@ -4037,6 +4037,8 @@
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");
@@ -4101,6 +4103,18 @@
kgsl_driver.mem_workqueue = alloc_workqueue("kgsl-mementry",
WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
+ 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, ¶m);
+
kgsl_events_init();
result = kgsl_cmdbatch_init();
diff --git a/drivers/gpu/msm/kgsl.h b/drivers/gpu/msm/kgsl.h
index 90dc550..8460476 100644
--- a/drivers/gpu/msm/kgsl.h
+++ b/drivers/gpu/msm/kgsl.h
@@ -26,6 +26,7 @@
#include <linux/mm.h>
#include <linux/dma-attrs.h>
#include <linux/uaccess.h>
+#include <linux/kthread.h>
#include <asm/cacheflush.h>
/* The number of memstore arrays limits the number of contexts allowed.
@@ -132,6 +133,8 @@
unsigned int full_cache_threshold;
struct workqueue_struct *workqueue;
struct workqueue_struct *mem_workqueue;
+ struct kthread_worker worker;
+ struct task_struct *worker_thread;
};
extern struct kgsl_driver kgsl_driver;
@@ -273,7 +276,7 @@
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 6e8abf3..859511b 100644
--- a/drivers/gpu/msm/kgsl_events.c
+++ b/drivers/gpu/msm/kgsl_events.c
@@ -32,7 +32,7 @@
{
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 @@
* 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);
@@ -282,7 +282,7 @@
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);
@@ -297,7 +297,7 @@
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;
}
