RDMA allocatable resources changed to 0 after kubelet restart

[WulixuanS]

Version: v1.3.2

RDMA device plugin log:

As can be seen from the log, when kubelet restart, it triggers context canceled and restart will block because channel size is 0, context listener added in this issue: https://github.com/Mellanox/k8s-rdma-shared-dev-plugin/pull/51.

When the kubelet restarts, ListAndWatch will receive the event from the stop channel, there is no need to watch context, so I fixed the bug by removing the context listener. If necessary, i can submit a PR.

func (rs *resourceServer) ListAndWatch(e *pluginapi.Empty, s pluginapi.DevicePlugin_ListAndWatchServer) error {
        resp := new(pluginapi.ListAndWatchResponse)


        // Send initial list of devices
        if err := rs.sendDevices(resp, s); err != nil {
                return err
        }


        for {
                select {
                case <-s.Context().Done():
                        log.Printf("ListAndWatch stream close: %v", s.Context().Err())
                        return nil
                case <-rs.stop:
                        return nil
                case d := <-rs.health:
                        // FIXME: there is no way to recover from the Unhealthy state.
                        d.Health = pluginapi.Unhealthy
                        _ = s.Send(&pluginapi.ListAndWatchResponse{Devices: rs.devs})
                case <-rs.updateResource:
                        if err := rs.sendDevices(resp, s); err != nil {
                                // The old stream may not be closed properly, return to close it
                                // and pass the update event to the new stream for processing
                                rs.updateResource <- true
                                return err
                        }
                }
        }
}



func (rs *resourceServer) Restart() error {
        log.Printf("restarting %s device plugin server...", rs.resourceName)
        if rs.rsConnector == nil || rs.rsConnector.GetServer() == nil {
                return fmt.Errorf("grpc server instance not found for %s", rs.resourceName)
        }


        rs.rsConnector.Stop()
        rs.rsConnector.DeleteServer()


        // Send terminate signal to ListAndWatch()
        rs.stop <- true


        return rs.Start()
}

[WulixuanS]

cc @adrianchiris

[adrianchiris]

What is the K8s version you are using ?

i see in the logs:

Using Deprecated Device Plugin Registry path

does the following path exist in your system: /var/lib/kubelet/plugins_registry ?

[adrianchiris]

please check #82 it should solve the issue.

[adrianchiris]

v1.4.0 is out please check :)

[hvp4]

@adrianchiris v1.4.0 release seems to be broken, cannot find the release: docker pull nvcr.io/nvidia/cloud-native/k8s-rdma-shared-dev-plugin:v1.4.0 Error response from daemon: manifest for nvcr.io/nvidia/cloud-native/k8s-rdma-shared-dev-plugin:v1.4.0 not found: manifest unknown: manifest unknown

Nor can be seen here: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/cloud-native/containers/k8s-rdma-shared-dev-plugin/tags

[adrianchiris]

image is here: https://github.com/Mellanox/k8s-rdma-shared-dev-plugin/pkgs/container/k8s-rdma-shared-dev-plugin