drm/amdgpu: add timer to fence to detect scheduler lockup

Change-Id: I67e987db0efdca28faa80b332b75571192130d33
Signed-off-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: David Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

3 files changed, 61 insertions, 3 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index b3fc26c59787..fcad7e060938 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -628,8 +628,20 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
         init_waitqueue_head(&ring->fence_drv.fence_queue);
 
         if (amdgpu_enable_scheduler) {
+                long timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
+                if (timeout == 0) {
+                        /*
+                         * FIXME:
+                         * Delayed workqueue cannot use it directly,
+                         * so the scheduler will not use delayed workqueue if
+                         * MAX_SCHEDULE_TIMEOUT is set.
+                         * Currently keep it simple and silly.
+                         */
+                        timeout = MAX_SCHEDULE_TIMEOUT;
+                }
                 r = amd_sched_init(&ring->sched, &amdgpu_sched_ops,
-                                   amdgpu_sched_hw_submission, ring->name);
+                                   amdgpu_sched_hw_submission,
+                                   timeout, ring->name);
                 if (r) {
                         DRM_ERROR("Failed to create scheduler on ring %s.\n",
                                   ring->name);
diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
index 3697eeeecf82..a413dee7cd19 100644
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
@@ -327,19 +327,49 @@ static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
         struct amd_sched_fence *s_fence =
                 container_of(cb, struct amd_sched_fence, cb);
         struct amd_gpu_scheduler *sched = s_fence->sched;
+        unsigned long flags;
 
         atomic_dec(&sched->hw_rq_count);
         amd_sched_fence_signal(s_fence);
+        if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
+                cancel_delayed_work_sync(&s_fence->dwork);
+                spin_lock_irqsave(&sched->fence_list_lock, flags);
+                list_del_init(&s_fence->list);
+                spin_unlock_irqrestore(&sched->fence_list_lock, flags);
+        }
         fence_put(&s_fence->base);
         wake_up_interruptible(&sched->wake_up_worker);
 }
 
+static void amd_sched_fence_work_func(struct work_struct *work)
+{
+        struct amd_sched_fence *s_fence =
+                container_of(work, struct amd_sched_fence, dwork.work);
+        struct amd_gpu_scheduler *sched = s_fence->sched;
+        struct amd_sched_fence *entity, *tmp;
+        unsigned long flags;
+
+        DRM_ERROR("[%s] scheduler is timeout!\n", sched->name);
+
+        /* Clean all pending fences */
+        list_for_each_entry_safe(entity, tmp, &sched->fence_list, list) {
+                DRM_ERROR("  fence no %d\n", entity->base.seqno);
+                cancel_delayed_work_sync(&entity->dwork);
+                spin_lock_irqsave(&sched->fence_list_lock, flags);
+                list_del_init(&entity->list);
+                spin_unlock_irqrestore(&sched->fence_list_lock, flags);
+                fence_put(&entity->base);
+        }
+}
+
 static int amd_sched_main(void *param)
 {
         struct sched_param sparam = {.sched_priority = 1};
         struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
         int r, count;
 
+        spin_lock_init(&sched->fence_list_lock);
+        INIT_LIST_HEAD(&sched->fence_list);
         sched_setscheduler(current, SCHED_FIFO, &sparam);
 
         while (!kthread_should_stop()) {
@@ -347,6 +377,7 @@ static int amd_sched_main(void *param)
                 struct amd_sched_fence *s_fence;
                 struct amd_sched_job *sched_job;
                 struct fence *fence;
+                unsigned long flags;
 
                 wait_event_interruptible(sched->wake_up_worker,
                         kthread_should_stop() ||
@@ -357,6 +388,15 @@ static int amd_sched_main(void *param)
 
                 entity = sched_job->s_entity;
                 s_fence = sched_job->s_fence;
+
+                if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
+                        INIT_DELAYED_WORK(&s_fence->dwork, amd_sched_fence_work_func);
+                        schedule_delayed_work(&s_fence->dwork, sched->timeout);
+                        spin_lock_irqsave(&sched->fence_list_lock, flags);
+                        list_add_tail(&s_fence->list, &sched->fence_list);
+                        spin_unlock_irqrestore(&sched->fence_list_lock, flags);
+                }
+
                 atomic_inc(&sched->hw_rq_count);
                 fence = sched->ops->run_job(sched_job);
                 if (fence) {
@@ -392,11 +432,12 @@ static int amd_sched_main(void *param)
 */
 int amd_sched_init(struct amd_gpu_scheduler *sched,
                    struct amd_sched_backend_ops *ops,
-                   unsigned hw_submission, const char *name)
+                   unsigned hw_submission, long timeout, const char *name)
 {
         sched->ops = ops;
         sched->hw_submission_limit = hw_submission;
         sched->name = name;
+        sched->timeout = timeout;
         amd_sched_rq_init(&sched->sched_rq);
         amd_sched_rq_init(&sched->kernel_rq);
 
diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
index 80b64dc22214..929e9aced041 100644
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
@@ -68,6 +68,8 @@ struct amd_sched_fence {
         struct amd_gpu_scheduler        *sched;
         spinlock_t                      lock;
         void                            *owner;
+        struct delayed_work             dwork;
+        struct list_head                list;
 };
 
 struct amd_sched_job {
@@ -103,18 +105,21 @@ struct amd_sched_backend_ops {
 struct amd_gpu_scheduler {
         struct amd_sched_backend_ops    *ops;
         uint32_t                        hw_submission_limit;
+        long                            timeout;
         const char                      *name;
         struct amd_sched_rq             sched_rq;
         struct amd_sched_rq             kernel_rq;
         wait_queue_head_t               wake_up_worker;
         wait_queue_head_t               job_scheduled;
         atomic_t                        hw_rq_count;
+        struct list_head                fence_list;
+        spinlock_t                      fence_list_lock;
         struct task_struct              *thread;
 };
 
 int amd_sched_init(struct amd_gpu_scheduler *sched,
                    struct amd_sched_backend_ops *ops,
-                   uint32_t hw_submission, const char *name);
+                   uint32_t hw_submission, long timeout, const char *name);
 void amd_sched_fini(struct amd_gpu_scheduler *sched);
 
 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,