drm/amd/amdgpu: fix flr_work corner case

[Why]
In SRIOV multi-vf environment, the flr_work can be entered
even if the TDR thread has entered the recovery. This can
lead to GMC flush tlb with SDMA during full access while SDMA
is not initialized.

[How]
1. flr_work should take write_lock, otherwise there maybe hw access
during vf flr
2. (amdgpu_in_reset(adev) ||!down_write_trylock(&adev->reset_sem))
is the correct critera when the flr_work direct returns.

Acked-by: Christian König <[email protected]>
Signed-off-by: Jingwen Chen <[email protected]>

Author: Jingwen Chen

Diff for: drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c

@@ -259,7 +259,8 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
 	 * otherwise the mailbox msg will be ruined/reseted by
 	 * the VF FLR.
 	 */
-	if (atomic_cmpxchg(&adev->reset_domain->in_gpu_reset, 0, 1) != 0)
+	if (amdgpu_in_reset(adev) ||
+		atomic_cmpxchg(&adev->reset_domain->in_gpu_reset, 0, 1) != 0)
 		return;
 
 	down_write(&adev->reset_domain->sem);

Diff for: drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c

@@ -292,7 +292,8 @@ static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
 	 * otherwise the mailbox msg will be ruined/reseted by
 	 * the VF FLR.
 	 */
-	if (atomic_cmpxchg(&adev->reset_domain->in_gpu_reset, 0, 1) != 0)
+	if (amdgpu_in_reset(adev) ||
+		atomic_cmpxchg(&adev->reset_domain->in_gpu_reset, 0, 1) != 0)
 		return;
 
 	down_write(&adev->reset_domain->sem);