drm/amdgpu: rework scheduler submission handling.
authorChristian König <christian.koenig@amd.com>
Wed, 19 Aug 2015 14:12:15 +0000 (16:12 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 25 Aug 2015 14:39:31 +0000 (10:39 -0400)
Remove active_hw_rq and it's protecting queue_lock, they are unused.

User 32bit atomic for hw_rq_count, 64bits for counting to three is a bit
overkill.

Cleanup the function name and remove incorrect comments.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
drivers/gpu/drm/amd/scheduler/gpu_scheduler.h

index f8d46b0b4e3b8e4563998c571945e3ee6db8fdc1..981420ea1cb5fec7460688bf021a36347b69b0d6 100644 (file)
@@ -94,25 +94,12 @@ amd_sched_rq_select_entity(struct amd_sched_rq *rq)
 }
 
 /**
- * Note: This function should only been called inside scheduler main
- * function for thread safety, there is no other protection here.
- * return ture if scheduler has something ready to run.
- *
- * For active_hw_rq, there is only one producer(scheduler thread) and
- * one consumer(ISR). It should be safe to use this function in scheduler
- * main thread to decide whether to continue emit more IBs.
-*/
-static bool is_scheduler_ready(struct amd_gpu_scheduler *sched)
+ * Return ture if we can push more jobs to the hw.
+ */
+static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
 {
-       unsigned long flags;
-       bool full;
-
-       spin_lock_irqsave(&sched->queue_lock, flags);
-       full = atomic64_read(&sched->hw_rq_count) <
-               sched->hw_submission_limit ? true : false;
-       spin_unlock_irqrestore(&sched->queue_lock, flags);
-
-       return full;
+       return atomic_read(&sched->hw_rq_count) <
+               sched->hw_submission_limit;
 }
 
 /**
@@ -124,7 +111,7 @@ select_context(struct amd_gpu_scheduler *sched)
        struct amd_sched_entity *wake_entity = NULL;
        struct amd_sched_entity *tmp;
 
-       if (!is_scheduler_ready(sched))
+       if (!amd_sched_ready(sched))
                return NULL;
 
        /* Kernel run queue has higher priority than normal run queue*/
@@ -293,14 +280,10 @@ static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
        struct amd_sched_job *sched_job =
                container_of(cb, struct amd_sched_job, cb);
        struct amd_gpu_scheduler *sched;
-       unsigned long flags;
 
        sched = sched_job->sched;
        amd_sched_fence_signal(sched_job->s_fence);
-       spin_lock_irqsave(&sched->queue_lock, flags);
-       list_del(&sched_job->list);
-       atomic64_dec(&sched->hw_rq_count);
-       spin_unlock_irqrestore(&sched->queue_lock, flags);
+       atomic_dec(&sched->hw_rq_count);
        fence_put(&sched_job->s_fence->base);
        sched->ops->process_job(sched, sched_job);
        wake_up_interruptible(&sched->wait_queue);
@@ -320,7 +303,7 @@ static int amd_sched_main(void *param)
                struct fence *fence;
 
                wait_event_interruptible(sched->wait_queue,
-                                        is_scheduler_ready(sched) &&
+                                        amd_sched_ready(sched) &&
                                         (c_entity = select_context(sched)));
                r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
                if (r != sizeof(void *))
@@ -329,11 +312,7 @@ static int amd_sched_main(void *param)
                if (sched->ops->prepare_job)
                        r = sched->ops->prepare_job(sched, c_entity, job);
                if (!r) {
-                       unsigned long flags;
-                       spin_lock_irqsave(&sched->queue_lock, flags);
-                       list_add_tail(&job->list, &sched->active_hw_rq);
-                       atomic64_inc(&sched->hw_rq_count);
-                       spin_unlock_irqrestore(&sched->queue_lock, flags);
+                       atomic_inc(&sched->hw_rq_count);
                }
                mutex_lock(&sched->sched_lock);
                fence = sched->ops->run_job(sched, c_entity, job);
@@ -384,13 +363,11 @@ struct amd_gpu_scheduler *amd_sched_create(void *device,
        sched->hw_submission_limit = hw_submission;
        snprintf(name, sizeof(name), "gpu_sched[%d]", ring);
        mutex_init(&sched->sched_lock);
-       spin_lock_init(&sched->queue_lock);
        amd_sched_rq_init(&sched->sched_rq);
        amd_sched_rq_init(&sched->kernel_rq);
 
        init_waitqueue_head(&sched->wait_queue);
-       INIT_LIST_HEAD(&sched->active_hw_rq);
-       atomic64_set(&sched->hw_rq_count, 0);
+       atomic_set(&sched->hw_rq_count, 0);
        /* Each scheduler will run on a seperate kernel thread */
        sched->thread = kthread_create(amd_sched_main, sched, name);
        if (sched->thread) {
index d328e968beb374e5de5b88a0bf53aa6baaf79748..81c00daa23fb26ae78ab83d870f80848315e19a4 100644 (file)
@@ -74,7 +74,6 @@ struct amd_sched_fence {
 };
 
 struct amd_sched_job {
-       struct list_head                list;
        struct fence_cb                 cb;
        struct amd_gpu_scheduler        *sched;
        struct amd_sched_entity         *s_entity;
@@ -115,8 +114,7 @@ struct amd_gpu_scheduler {
        struct task_struct              *thread;
        struct amd_sched_rq             sched_rq;
        struct amd_sched_rq             kernel_rq;
-       struct list_head                active_hw_rq;
-       atomic64_t                      hw_rq_count;
+       atomic_t                        hw_rq_count;
        struct amd_sched_backend_ops    *ops;
        uint32_t                        ring_id;
        uint32_t                        granularity; /* in ms unit */
@@ -124,7 +122,6 @@ struct amd_gpu_scheduler {
        wait_queue_head_t               wait_queue;
        struct amd_sched_entity *current_entity;
        struct mutex                    sched_lock;
-       spinlock_t                      queue_lock;
        uint32_t                        hw_submission_limit;
 };