feat: add host PID broker server and device plugin integration

charts/hami/README.md

@@ -192,6 +192,8 @@ This document provides detailed descriptions of all configurable values paramete
 |-----------|-------------|---------------|
 | `devicePlugin.pluginPath` | Plugin path | `/var/lib/kubelet/device-plugins` |
 | `devicePlugin.libPath` | Library path | `/usr/local/vgpu` |
+| `devicePlugin.hostPID` | Use the host PID namespace for the device plugin | `true` |
+| `devicePlugin.hostPIDBroker.enabled` | Let HAMi core ask the device plugin for its host PID. This requires `devicePlugin.hostPID`. See [Host PID broker](../../docs/develop/hostpid-broker.md) | `false` |
 | `devicePlugin.nvidiaNodeSelector` | NVIDIA node selector | `{"gpu": "on"}` |
 | `devicePlugin.updateStrategy.type` | Update strategy type | `RollingUpdate` |
 | `devicePlugin.updateStrategy.rollingUpdate.maxUnavailable` | Maximum unavailable count | `1` |

charts/hami/templates/device-plugin/daemonsetnvidia.yaml

@@ -1,3 +1,6 @@
+{{- if and .Values.devicePlugin.enabled .Values.devicePlugin.hostPIDBroker.enabled (not .Values.devicePlugin.hostPID) }}
+{{- fail "devicePlugin.hostPIDBroker requires devicePlugin.hostPID" }}
+{{- end }}
 {{- if .Values.devicePlugin.enabled }}
 apiVersion: apps/v1
 kind: DaemonSet
@@ -94,6 +97,10 @@ spec:
               value: {{ .Values.devicePlugin.deviceListStrategy }}
             - name: HOOK_PATH
               value: {{ .Values.global.gpuHookPath }}
+            {{- if .Values.devicePlugin.hostPIDBroker.enabled }}
+            - name: LIBVGPU_HOSTPID_BROKER
+              value: "1"
+            {{- end }}
             {{- if typeIs "bool" .Values.devicePlugin.passDeviceSpecsEnabled }}
             - name: PASS_DEVICE_SPECS
               value: {{ .Values.devicePlugin.passDeviceSpecsEnabled | quote }}
@@ -169,6 +176,10 @@ spec:
               mountPath: /etc/hami/numa-refit-ca
               readOnly: true
             {{- end }}
+            {{- if .Values.devicePlugin.hostPIDBroker.enabled }}
+            - name: hostpid-broker
+              mountPath: /var/run/hami/hostpid
+            {{- end }}
             {{- if typeIs "string" .Values.devicePlugin.nvidiaDriverRoot }}
             # We always mount the driver root at /driver-root in the container.
             # This is required for CDI detection to work correctly.
@@ -266,6 +277,12 @@ spec:
           hostPath:
             path: /var/run/cdi
             type: DirectoryOrCreate
+        {{- if .Values.devicePlugin.hostPIDBroker.enabled }}
+        - name: hostpid-broker
+          hostPath:
+            path: /var/run/hami/hostpid
+            type: DirectoryOrCreate
+        {{- end }}
         - name: usrbin
           hostPath:
             path: /usr/bin

charts/hami/values.yaml

@@ -426,6 +426,10 @@ devicePlugin:
 
   podAnnotations: {}
   hostPID: true
+  # Let HAMi core ask the device plugin for its host PID.
+  # This requires hostPID to be true.
+  hostPIDBroker:
+    enabled: false
   hostNetwork: false
   securityContext:
     privileged: true

cmd/device-plugin/nvidia/hostpid_broker.go

@@ -0,0 +1,78 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Copyright (c) 2026 The HAMi Authors.
+ */
+
+package main
+
+import (
+	"errors"
+	"fmt"
+	"os"
+
+	"k8s.io/klog/v2"
+
+	"github.com/Project-HAMi/HAMi/pkg/device-plugin/nvidiadevice/nvinternal/hostpid"
+)
+
+type runningHostPIDBroker struct {
+	broker   hostPIDBroker
+	done     chan struct{}
+	serveErr error
+}
+
+type hostPIDBroker interface {
+	Serve() error
+	Close() error
+}
+
+type hostPIDBrokerListener func() (hostPIDBroker, error)
+
+func startHostPIDBroker() (*runningHostPIDBroker, error) {
+	return startHostPIDBrokerWithListener(func() (hostPIDBroker, error) {
+		return hostpid.ListenDefault()
+	})
+}
+
+func startHostPIDBrokerWithListener(
+	listen hostPIDBrokerListener) (*runningHostPIDBroker, error) {
+	if !hostpid.Enabled(os.Getenv(hostpid.EnvironmentVariable)) {
+		return nil, nil
+	}
+	broker, err := listen()
+	if err != nil {
+		return nil, err
+	}
+	running := &runningHostPIDBroker{
+		broker: broker,
+		done:   make(chan struct{}),
+	}
+	go func() {
+		running.serveErr = broker.Serve()
+		close(running.done)
+	}()
+	klog.Infof("Host PID broker is listening on %s", hostpid.ServerSocketPath)
+	return running, nil
+}
+
+func (running *runningHostPIDBroker) stop() error {
+	closeErr := running.broker.Close()
+	<-running.done
+	return closeErr
+}
+
+func (running *runningHostPIDBroker) failure() error {
+	if running == nil {
+		return nil
+	}
+	select {
+	case <-running.done:
+		if running.serveErr != nil {
+			return fmt.Errorf("host PID broker stopped: %w", running.serveErr)
+		}
+		return errors.New("host PID broker stopped unexpectedly")
+	default:
+		return nil
+	}
+}

cmd/device-plugin/nvidia/hostpid_broker_lifecycle_test.go

@@ -0,0 +1,130 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Copyright (c) 2026 The HAMi Authors.
+ */
+
+package main
+
+import (
+	"errors"
+	"testing"
+
+	"github.com/Project-HAMi/HAMi/pkg/device-plugin/nvidiadevice/nvinternal/plugin"
+	"github.com/Project-HAMi/HAMi/pkg/device-plugin/nvidiadevice/nvinternal/rm"
+)
+
+type fakeDevicePlugin struct {
+	devices    rm.Devices
+	start      func(string) error
+	stopErr    error
+	startCalls int
+	stopCalls  int
+}
+
+func (p *fakeDevicePlugin) Devices() rm.Devices {
+	return p.devices
+}
+
+func (p *fakeDevicePlugin) Start(socket string) error {
+	p.startCalls++
+	if p.start != nil {
+		return p.start(socket)
+	}
+	return nil
+}
+
+func (p *fakeDevicePlugin) Stop() error {
+	p.stopCalls++
+	return p.stopErr
+}
+
+func devicePluginWithDevice() *fakeDevicePlugin {
+	return &fakeDevicePlugin{
+		devices: rm.Devices{"GPU-0": &rm.Device{}},
+	}
+}
+
+func TestStartPluginServersDetectsBrokerFailureBeforeStart(t *testing.T) {
+	wantErr := errors.New("broker failed")
+	done := make(chan struct{})
+	close(done)
+	running := &runningHostPIDBroker{
+		done:     done,
+		serveErr: wantErr,
+	}
+	p := devicePluginWithDevice()
+
+	started, restart, err := startPluginServers(
+		[]plugin.Interface{p}, "/tmp/kubelet.sock", running)
+	if started != 0 || restart || !errors.Is(err, wantErr) {
+		t.Fatalf("started=%d restart=%v err=%v, want broker failure",
+			started, restart, err)
+	}
+	if p.startCalls != 0 || p.stopCalls != 0 {
+		t.Fatalf("start calls=%d stop calls=%d, want 0 and 0",
+			p.startCalls, p.stopCalls)
+	}
+}
+
+func TestStartPluginServersCleansUpAfterBrokerFailure(t *testing.T) {
+	wantServeErr := errors.New("broker failed")
+	wantStopErr := errors.New("plugin stop failed")
+	done := make(chan struct{})
+	running := &runningHostPIDBroker{done: done}
+	p := devicePluginWithDevice()
+	p.stopErr = wantStopErr
+	p.start = func(string) error {
+		running.serveErr = wantServeErr
+		close(done)
+		return nil
+	}
+
+	started, restart, err := startPluginServers(
+		[]plugin.Interface{p}, "/tmp/kubelet.sock", running)
+	if started != 0 || restart || !errors.Is(err, wantServeErr) ||
+		!errors.Is(err, wantStopErr) {
+		t.Fatalf("started=%d restart=%v err=%v, want joined failures",
+			started, restart, err)
+	}
+	if p.startCalls != 1 || p.stopCalls != 1 {
+		t.Fatalf("start calls=%d stop calls=%d, want 1 and 1",
+			p.startCalls, p.stopCalls)
+	}
+}
+
+func TestStartPluginServersRequestsRestartAfterStartFailure(t *testing.T) {
+	wantErr := errors.New("plugin start failed")
+	p := devicePluginWithDevice()
+	p.start = func(socket string) error {
+		if socket != "/tmp/kubelet.sock" {
+			t.Fatalf("socket=%q", socket)
+		}
+		return wantErr
+	}
+
+	started, restart, err := startPluginServers(
+		[]plugin.Interface{p}, "/tmp/kubelet.sock", nil)
+	if started != 0 || !restart || err != nil {
+		t.Fatalf("started=%d restart=%v err=%v", started, restart, err)
+	}
+	if p.startCalls != 1 || p.stopCalls != 0 {
+		t.Fatalf("start calls=%d stop calls=%d, want 1 and 0",
+			p.startCalls, p.stopCalls)
+	}
+}
+
+func TestStartPluginServersSkipsEmptyPlugins(t *testing.T) {
+	empty := &fakeDevicePlugin{}
+	ready := devicePluginWithDevice()
+
+	started, restart, err := startPluginServers(
+		[]plugin.Interface{empty, ready}, "/tmp/kubelet.sock", nil)
+	if started != 1 || restart || err != nil {
+		t.Fatalf("started=%d restart=%v err=%v", started, restart, err)
+	}
+	if empty.startCalls != 0 || ready.startCalls != 1 {
+		t.Fatalf("empty starts=%d ready starts=%d, want 0 and 1",
+			empty.startCalls, ready.startCalls)
+	}
+}