[PATCH 06/31] drm/amdgpu: rework queue reset scheduler interaction

[Alex Deucher]

From: Christian König <ckoenig.leichtzumerken at gmail.com>

Stopping the scheduler for queue reset is generally a good idea because
it prevents any worker from touching the ring buffer.

But using amdgpu_fence_driver_force_completion() before restarting it was
a really bad idea because it marked fences as failed while the work was
potentially still running.

Stop doing that and cleanup the comment a bit.

v2: keep amdgpu_fence_driver_force_completion() for non-gfx rings
v3: drop amdgpu_fence_driver_force_completion() for compute ring
v4: avoid a warning when setting an error on the fence

Reviewed-by: Alex Deucher <alexander.deucher at amd.com>
Signed-off-by: Christian König <christian.koenig at amd.com>
Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 37 +++++++++++++++----------
 1 file changed, 22 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index ddb9d3269357c..821f88b64f3f6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -91,8 +91,8 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
 	struct amdgpu_job *job = to_amdgpu_job(s_job);
 	struct amdgpu_task_info *ti;
 	struct amdgpu_device *adev = ring->adev;
-	int idx;
-	int r;
+	bool set_error = false;
+	int idx, r;
 
 	if (!drm_dev_enter(adev_to_drm(adev), &idx)) {
 		dev_info(adev->dev, "%s - device unplugged skipping recovery on scheduler:%s",
@@ -136,10 +136,12 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
 	} else if (amdgpu_gpu_recovery && ring->funcs->reset) {
 		bool is_guilty;
 
-		dev_err(adev->dev, "Starting %s ring reset\n", s_job->sched->name);
-		/* stop the scheduler, but don't mess with the
-		 * bad job yet because if ring reset fails
-		 * we'll fall back to full GPU reset.
+		dev_err(adev->dev, "Starting %s ring reset\n",
+			s_job->sched->name);
+
+		/*
+		 * Stop the scheduler to prevent anybody else from touching the
+		 * ring buffer.
 		 */
 		drm_sched_wqueue_stop(&ring->sched);
 
@@ -154,24 +156,29 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
 
 		if (is_guilty)
 			dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
+			set_error = true;
+		}
 
 		r = amdgpu_ring_reset(ring, job->vmid);
 		if (!r) {
-			if (amdgpu_ring_sched_ready(ring))
-				drm_sched_stop(&ring->sched, s_job);
 			if (is_guilty) {
 				atomic_inc(&ring->adev->gpu_reset_counter);
-				amdgpu_fence_driver_force_completion(ring);
+				if ((ring->funcs->type != AMDGPU_RING_TYPE_GFX) &&
+				    (ring->funcs->type != AMDGPU_RING_TYPE_COMPUTE))
+					amdgpu_fence_driver_force_completion(ring);
 			}
-			if (amdgpu_ring_sched_ready(ring))
-				drm_sched_start(&ring->sched, 0);
-			dev_err(adev->dev, "Ring %s reset succeeded\n", ring->sched.name);
-			drm_dev_wedged_event(adev_to_drm(adev), DRM_WEDGE_RECOVERY_NONE);
+			drm_sched_wqueue_start(&ring->sched);
+			dev_err(adev->dev, "Ring %s reset succeeded\n",
+				ring->sched.name);
+			drm_dev_wedged_event(adev_to_drm(adev),
+					     DRM_WEDGE_RECOVERY_NONE);
 			goto exit;
 		}
-		dev_err(adev->dev, "Ring %s reset failure\n", ring->sched.name);
+		dev_err(adev->dev, "Ring %s reset failed\n", ring->sched.name);
 	}
-	dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
+
+	if (!set_error)
+		dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
 
 	if (amdgpu_device_should_recover_gpu(ring->adev)) {
 		struct amdgpu_reset_context reset_context;
-- 
2.49.0