Pci device plugin fix orphan goroutine and remove unused channel

[WebberHuang1118]

Problem:
After a PCI device plugin is disabled, there is Go routine (invoked as the device plugin started) blocked by chan stop, however, chan stop is never touched, this brings an orphan Go routine once a PCI device is enabled and disabled.

Solution:
PCI device plugin fix the orphan goroutine and removes unused channel

Related Issue:
harvester/harvester#5179

Additional Context:
The CodeFactor Complex Method error could be fixed once PR #66 merged

[Yu-Jack]

I traced the original stop usage in kubevirt.

The heartBeat will trigger kubevirt to refresh the permitted devices.
-> Then it will start and stop all devices.
-> Then the DeviceController uses controlledDevice to close stopChan

In device_plugin part, it uses errChan to wait which waits for grpc server and health check error in Start function.
-> and there is a healthCheck function to detect stopChan to return nil if stopChan is closed.

The structure is one DeviceController uses multiple controlledDevice, and one controlledDevice is mapped to one devicePlugin. But, in our structure, we just used devicePlugin, and our device plugin Start function doesn't wait for health check, so I think that's why stop chan is weird in our structure.

[WebberHuang1118]

I traced the original stop usage in kubevirt.

The heartBeat will trigger kubevirt to refresh the permitted devices. -> Then it will start and stop all devices. -> Then the DeviceController uses controlledDevice to close stopChan

In device_plugin part, it uses errChan to wait which waits for grpc server and health check error in Start function. -> and there is a healthCheck function to detect stopChan to return nil if stopChan is closed.

The structure is one DeviceController uses multiple controlledDevice, and one controlledDevice is mapped to one devicePlugin. But, in our structure, we just used devicePlugin, and our device plugin Start function doesn't wait for health check, so I think that's why stop chan is weird in our structure.

@Yu-Jack, Thanks for your explanation,

IIUC, in the current implementation stop chan doesn't take effect, its existence is just because the PCI controller references some Kubevirt code snippet from the beginning.

IMHO, since this PR doesn't introduce any structural or control flow change, I'd like to remove the unused code path. WDYT? cc @bk201 @ibrokethecloud , thanks.

[Yu-Jack]

Just leave link to original comment #66 (comment) if we don't want to remove existing codes.

[Yu-Jack]

I prefer not to delete stop channel. This stop is a control flag to control device plugin. For example, if we remove this stop, it can't return in health check function. Instead, we can remove done channel in this case.

[ibrokethecloud]

Looks like one channel is enough. we could either use done or stop to trigger cleanup. that channel should be used in the healthcheck.

this is legacy stuff which we copied over when we tried to setup the device plugins.

Thanks for the PR @WebberHuang1118 to help clean this up.

[Copilot]

Pull request overview

This pull request fixes an orphan goroutine issue in the PCI device plugin lifecycle management. Previously, when a PCI device plugin was disabled, a goroutine would remain blocked on an unused stop channel that was never closed, causing a goroutine leak.

Changes:

• Removed the unused stop channel parameter from device plugin lifecycle methods
• Removed blocking on the stop channel in the device plugin startup goroutine
• Cleaned up import ordering in vgpu_controller.go

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
pkg/deviceplugins/device_manager.go	Removed stop channel field and updated Start() and SetStarted() method signatures to not accept stop channel parameter; removed dp.stop case from ListAndWatch select statement
pkg/controller/pcideviceclaim/pcideviceclaim_controller.go	Removed creation of unused stop channel and blocking <-stop statement that caused orphan goroutine
pkg/controller/gpudevice/vgpu_controller.go	Reordered imports to follow standard grouping convention

Comments suppressed due to low confidence (1)

pkg/deviceplugins/device_manager.go:372

• The healthCheck() method is missing a case for the dp.done channel in its select statement. When stopDevicePlugin() closes dp.done, the healthCheck goroutine will not exit properly, which can lead to goroutine leaks. This is critical because healthCheck is spawned in ListAndWatch() and needs to terminate when the device plugin stops.

The select statement should include a case for <-dp.done that returns nil, similar to how VGPUDevicePlugin.performCheck() handles both dp.stop and dp.done in vgpu_device_manager.go:324-325.

	for {
		select {
		case err := <-watcher.Errors:
			logger.Reason(err).Errorf("error watching devices and device plugin directory")
		case event := <-watcher.Events:
			logger.V(4).Infof("health Event: %v", event)
			if monDevID, exist := monitoredDevices[event.Name]; exist {
				// Health in this case is if the device path actually exists
				if event.Op == fsnotify.Create {
					logger.Infof("monitored device %s appeared", dp.resourceName)
					dp.health <- deviceHealth{
						DevID:  monDevID,
						Health: pluginapi.Healthy,
					}
				} else if (event.Op == fsnotify.Remove) || (event.Op == fsnotify.Rename) {
					logger.Infof("monitored device %s disappeared", dp.resourceName)
					dp.health <- deviceHealth{
						DevID:  monDevID,
						Health: pluginapi.Unhealthy,
					}
				}
			} else if event.Name == dp.socketPath && event.Op == fsnotify.Remove {
				logger.Infof("device socket file for device %s was removed, kubelet probably restarted.", dp.resourceName)
				return nil
			}
		}
	}

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.