Jiri Slaby cf2094
From: YuBiao Wang <YuBiao.Wang@amd.com>
Jiri Slaby cf2094
Date: Thu, 16 Mar 2023 11:30:32 +0800
Jiri Slaby cf2094
Subject: [PATCH] drm/amdgpu: Force signal hw_fences that are embedded in
Jiri Slaby cf2094
 non-sched jobs
Jiri Slaby cf2094
References: bsc#1012628
Jiri Slaby cf2094
Patch-mainline: 6.2.12
Jiri Slaby cf2094
Git-commit: 033c56474acf567a450f8bafca50e0b610f2b716
Jiri Slaby cf2094
Jiri Slaby cf2094
[ Upstream commit 033c56474acf567a450f8bafca50e0b610f2b716 ]
Jiri Slaby cf2094
Jiri Slaby cf2094
[Why]
Jiri Slaby cf2094
For engines not supporting soft reset, i.e. VCN, there will be a failed
Jiri Slaby cf2094
ib test before mode 1 reset during asic reset. The fences in this case
Jiri Slaby cf2094
are never signaled and next time when we try to free the sa_bo, kernel
Jiri Slaby cf2094
will hang.
Jiri Slaby cf2094
Jiri Slaby cf2094
[How]
Jiri Slaby cf2094
During pre_asic_reset, driver will clear job fences and afterwards the
Jiri Slaby cf2094
fences' refcount will be reduced to 1. For drm_sched_jobs it will be
Jiri Slaby cf2094
released in job_free_cb, and for non-sched jobs like ib_test, it's meant
Jiri Slaby cf2094
to be released in sa_bo_free but only when the fences are signaled. So
Jiri Slaby cf2094
we have to force signal the non_sched bad job's fence during
Jiri Slaby cf2094
pre_asic_reset or the clear is not complete.
Jiri Slaby cf2094
Jiri Slaby cf2094
Signed-off-by: YuBiao Wang <YuBiao.Wang@amd.com>
Jiri Slaby cf2094
Acked-by: Luben Tuikov <luben.tuikov@amd.com>
Jiri Slaby cf2094
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Jiri Slaby cf2094
Signed-off-by: Sasha Levin <sashal@kernel.org>
Jiri Slaby cf2094
Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Jiri Slaby cf2094
---
Jiri Slaby cf2094
 drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 9 +++++++++
Jiri Slaby cf2094
 1 file changed, 9 insertions(+)
Jiri Slaby cf2094
Jiri Slaby cf2094
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
Jiri Slaby cf2094
index faff4a3f..f52d0ba9 100644
Jiri Slaby cf2094
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
Jiri Slaby cf2094
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
Jiri Slaby cf2094
@@ -678,6 +678,15 @@ void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
Jiri Slaby cf2094
 		ptr = &ring->fence_drv.fences[i];
Jiri Slaby cf2094
 		old = rcu_dereference_protected(*ptr, 1);
Jiri Slaby cf2094
 		if (old && old->ops == &amdgpu_job_fence_ops) {
Jiri Slaby cf2094
+			struct amdgpu_job *job;
Jiri Slaby cf2094
+
Jiri Slaby cf2094
+			/* For non-scheduler bad job, i.e. failed ib test, we need to signal
Jiri Slaby cf2094
+			 * it right here or we won't be able to track them in fence_drv
Jiri Slaby cf2094
+			 * and they will remain unsignaled during sa_bo free.
Jiri Slaby cf2094
+			 */
Jiri Slaby cf2094
+			job = container_of(old, struct amdgpu_job, hw_fence);
Jiri Slaby cf2094
+			if (!job->base.s_fence && !dma_fence_is_signaled(old))
Jiri Slaby cf2094
+				dma_fence_signal(old);
Jiri Slaby cf2094
 			RCU_INIT_POINTER(*ptr, NULL);
Jiri Slaby cf2094
 			dma_fence_put(old);
Jiri Slaby cf2094
 		}
Jiri Slaby cf2094
-- 
Jiri Slaby cf2094
2.35.3
Jiri Slaby cf2094