drm/amdgpu: protect RAS sysfs during GPU reset

MMHub EDC becomes dirty after BACO reset

EDC registers should be cleared early on in reset phase

Reviewed-by: Hawking Zhang <[email protected]>
Signed-off-by: John Clements <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

Author: John Clements

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

@@ -2742,6 +2742,9 @@ static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
 
 		if (adev->asic_reset_res)
 			goto fail;
+
+		if (adev->mmhub.funcs && adev->mmhub.funcs->reset_ras_error_count)
+			adev->mmhub.funcs->reset_ras_error_count(adev);
 	} else {
 
 		task_barrier_full(&hive->tb);
@@ -3910,8 +3913,15 @@ static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
 		}
 	}
 
-	if (!r && amdgpu_ras_intr_triggered())
+	if (!r && amdgpu_ras_intr_triggered()) {
+		list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
+			if (tmp_adev->mmhub.funcs &&
+			    tmp_adev->mmhub.funcs->reset_ras_error_count)
+				tmp_adev->mmhub.funcs->reset_ras_error_count(tmp_adev);
+		}
+
 		amdgpu_ras_intr_cleared();
+	}
 
 	list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
 		if (need_full_reset) {

drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c

@@ -281,6 +281,11 @@ static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f, const char __user *
 	struct ras_debug_if data;
 	int ret = 0;
 
+	if (amdgpu_ras_intr_triggered()) {
+		DRM_WARN("RAS WARN: error injection currently inaccessible\n");
+		return size;
+	}
+
 	ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
 	if (ret)
 		return -EINVAL;
@@ -394,6 +399,10 @@ static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
 		.head = obj->head,
 	};
 
+	if (amdgpu_ras_intr_triggered())
+		return snprintf(buf, PAGE_SIZE,
+				"Query currently inaccessible\n");
+
 	if (amdgpu_ras_error_query(obj->adev, &info))
 		return -EINVAL;