Skip to content

Secure, Fast, and Extensible Sandbox runtime for AI agents.

33.0 KB
Raw Blame
// Copyright 2025 The OpenSandbox Authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

package controller

import (
	"context"
	"fmt"
	"time"

	corev1 "k8s.io/api/core/v1"
	"k8s.io/apimachinery/pkg/api/errors"
	"k8s.io/apimachinery/pkg/api/resource"
	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
	"k8s.io/apimachinery/pkg/types"
	"k8s.io/client-go/util/retry"
	"k8s.io/utils/ptr"
	ctrl "sigs.k8s.io/controller-runtime"
	"sigs.k8s.io/controller-runtime/pkg/client"
	"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
	logf "sigs.k8s.io/controller-runtime/pkg/log"

	sandboxv1alpha1 "github.com/alibaba/OpenSandbox/sandbox-k8s/apis/sandbox/v1alpha1"
	"github.com/alibaba/OpenSandbox/sandbox-k8s/internal/controller/strategy"
	snapshotcontract "github.com/alibaba/OpenSandbox/sandbox-k8s/internal/snapshot"
	"github.com/alibaba/OpenSandbox/sandbox-k8s/internal/utils"
	controllerutils "github.com/alibaba/OpenSandbox/sandbox-k8s/internal/utils/controller"
)

const internalPauseSnapshotSuffix = "-pause"
const supportedPauseReplicas int32 = 1

const (
	vmStateRestoreVolumeName = "opensandbox-vmstate-restore"
	vmStateRestoreInitName   = "opensandbox-vmstate-restore"
	vmStateRestoreMountPath  = "/run/opensandbox/vmstate"
	// vmStateRestoreHeadroomBytes covers the manifest file and the copied
	// vmstate-loader binary stored next to the VM state payload.
	vmStateRestoreHeadroomBytes int64 = 64 << 20
)

func internalPauseSnapshotName(batchSandboxName string) string {
	// TODO: handle Kubernetes resource name length limits for long BatchSandbox names.
	return batchSandboxName + internalPauseSnapshotSuffix
}

func unsupportedPauseReplicasMessage(replicas *int32) string {
	if replicas == nil {
		return "pause/resume currently supports only BatchSandbox spec.replicas=1; spec.replicas is unset"
	}
	return fmt.Sprintf("pause/resume currently supports only BatchSandbox spec.replicas=1; got spec.replicas=%d", *replicas)
}

func ensureImagePullSecret(template *corev1.PodTemplateSpec, secretName string) {
	if template == nil || secretName == "" {
		return
	}
	for _, secret := range template.Spec.ImagePullSecrets {
		if secret.Name == secretName {
			return
		}
	}
	template.Spec.ImagePullSecrets = append(template.Spec.ImagePullSecrets, corev1.LocalObjectReference{Name: secretName})
}

func snapshotFailureMessage(snapshot *sandboxv1alpha1.SandboxSnapshot) string {
	for _, cond := range snapshot.Status.Conditions {
		if cond.Type == sandboxv1alpha1.SandboxSnapshotConditionFailed && cond.Status == sandboxv1alpha1.ConditionTrue && cond.Message != "" {
			return cond.Message
		}
	}
	return "snapshot failed"
}

func copyPodTemplateMap(values map[string]string) map[string]string {
	if len(values) == 0 {
		return nil
	}
	copied := make(map[string]string, len(values))
	for key, value := range values {
		copied[key] = value
	}
	if len(copied) == 0 {
		return nil
	}
	return copied
}

func sourcePodTemplateForPause(pod *corev1.Pod) *corev1.PodTemplateSpec {
	if pod == nil {
		return nil
	}
	labels := copyPodTemplateMap(pod.Labels)
	delete(labels, labelPoolName)
	delete(labels, labelPoolRevision)
	delete(labels, labelBatchSandboxNameKey)
	delete(labels, labelBatchSandboxPodIndexKey)

	spec := *pod.Spec.DeepCopy()
	spec.NodeName = ""

	return &corev1.PodTemplateSpec{
		ObjectMeta: metav1.ObjectMeta{
			Labels:      labels,
			Annotations: copyPodTemplateMap(pod.Annotations),
		},
		Spec: spec,
	}
}

func (r *BatchSandboxReconciler) deleteInternalPauseSnapshot(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) error {
	log := logf.FromContext(ctx)
	snapshot := &sandboxv1alpha1.SandboxSnapshot{}
	if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: internalPauseSnapshotName(bs.Name)}, snapshot); err != nil {
		if errors.IsNotFound(err) {
			return nil
		}
		return err
	}
	if snapshot.Status.Phase != sandboxv1alpha1.SandboxSnapshotPhaseSucceed {
		return nil
	}
	if err := r.Delete(ctx, snapshot); err != nil && !errors.IsNotFound(err) {
		return err
	}
	log.Info("Deleted SandboxSnapshot after successful resume", "snapshot", snapshot.Name)
	return nil
}

func (r *BatchSandboxReconciler) hasReadyResumePod(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (bool, error) {
	poolStrategy := strategy.NewPoolStrategy(bs)
	if poolStrategy.IsPooledMode() {
		alloc, err := parseSandboxAllocation(bs)
		if err != nil {
			return false, err
		}
		for _, podName := range alloc.Pods {
			pod := &corev1.Pod{}
			if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: podName}, pod); err != nil {
				if errors.IsNotFound(err) {
					continue
				}
				return false, err
			}
			if utils.IsPodReady(pod) {
				return true, nil
			}
		}
	}

	podList := &corev1.PodList{}
	if err := r.List(ctx, podList,
		client.InNamespace(bs.Namespace),
		client.MatchingLabels{labelBatchSandboxNameKey: bs.Name},
	); err != nil {
		return false, err
	}
	for i := range podList.Items {
		if utils.IsPodReady(&podList.Items[i]) {
			return true, nil
		}
	}

	pod := &corev1.Pod{}
	if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: fmt.Sprintf("%s-%d", bs.Name, batchSandboxFirstPodIndex)}, pod); err != nil {
		if errors.IsNotFound(err) {
			return false, nil
		}
		return false, err
	}
	return utils.IsPodReady(pod), nil
}

// dispatchPauseResume implements the 5-case dispatch table from the design doc.
// Returns (result, handled, error). If handled=true, the caller should return immediately.
func (r *BatchSandboxReconciler) dispatchPauseResume(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (ctrl.Result, bool, error) {
	log := logf.FromContext(ctx)
	generation := bs.Generation
	pauseObservedGen := bs.Status.PauseObservedGeneration
	pause := bs.Spec.Pause

	// In-progress phases take priority over generation checks so that controller-owned
	// spec mutations (template solidification, image replacement, pool detachment) do not
	// get mistaken for new external pause/resume requests.
	log.Info("Dispatch: checking phase", "currentPhase", bs.Status.Phase, "generation", generation, "pauseObservedGen", pauseObservedGen, "pause", pause)
	if bs.Status.Phase == sandboxv1alpha1.BatchSandboxPhaseResuming {
		log.Info("Dispatch: phase is Resuming, continuing resume")
		result, err := r.continueResume(ctx, bs)
		if err != nil {
			return result, true, err
		}
		// Return handled=false to let normal flow update phase from Resuming to Succeed.
		return result, false, nil
	}
	if bs.Status.Phase == sandboxv1alpha1.BatchSandboxPhasePausing {
		log.Info("Dispatch: phase is Pausing, syncing pause state")
		result, err := r.syncPauseOrClear(ctx, bs)
		return result, true, err
	}

	if generation > pauseObservedGen {
		if pause != nil {
			if *pause {
				if bs.Status.Phase == sandboxv1alpha1.BatchSandboxPhasePaused {
					// Already paused: a generation bump (for example a renew-only spec
					// change) is not a new pause request.
					log.Info("Dispatch: already Paused, skipping pause", "generation", generation, "pauseObservedGeneration", pauseObservedGen)
					return ctrl.Result{}, false, nil
				}
				log.Info("Dispatch: handlePause", "generation", generation, "pauseObservedGeneration", pauseObservedGen)
				result, err := r.handlePause(ctx, bs)
				return result, true, err
			}
			log.Info("Dispatch: handleResume", "generation", generation, "pauseObservedGeneration", pauseObservedGen, "currentPhase", bs.Status.Phase)
			result, err := r.handleResume(ctx, bs)
			return result, true, err
		}
		// No pause intent — skip the dedicated ACK API call. The normal flow's
		// persistRuntimeView will update PauseObservedGeneration in its status patch.
		log.Info("Dispatch: no pause intent, deferring ACK to status patch", "generation", generation, "pauseObservedGeneration", pauseObservedGen)
		return ctrl.Result{}, false, nil
	}

	return ctrl.Result{}, false, nil
}

// handlePause implements the pause flow:
// 1. ACK (pauseObservedGeneration + phase=Pausing)
// 2. Stop task-executor tasks, if any, while keeping the source Pod allocated
// 3. Create SandboxSnapshot child resource
func (r *BatchSandboxReconciler) handlePause(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (ctrl.Result, error) {
	log := logf.FromContext(ctx)

	// The pause snapshot format records one source pod's container images, so this
	// controller only supports rootfs pause/resume for single-replica sandboxes.
	if bs.Spec.Replicas == nil || *bs.Spec.Replicas != supportedPauseReplicas {
		msg := unsupportedPauseReplicasMessage(bs.Spec.Replicas)
		log.Info("Rejecting pause for unsupported replica count", "message", msg)
		phase := bs.Status.Phase
		if phase == "" {
			phase = sandboxv1alpha1.BatchSandboxPhaseSucceed
		}
		if err := r.ackPauseWithPhase(ctx, bs, phase, ""); err != nil {
			return ctrl.Result{}, err
		}
		if err := r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionPauseFailed, sandboxv1alpha1.ConditionTrue, "UnsupportedReplicas", msg); err != nil {
			return ctrl.Result{}, err
		}
		return ctrl.Result{}, nil
	}

	snapshot := &sandboxv1alpha1.SandboxSnapshot{}
	snapshotName := internalPauseSnapshotName(bs.Name)
	err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: snapshotName}, snapshot)
	if err == nil && snapshot.DeletionTimestamp != nil {
		log.Info("Waiting for stale SandboxSnapshot deletion before retrying pause", "snapshot", snapshotName)
		return ctrl.Result{RequeueAfter: time.Second}, nil
	}
	if err == nil && (snapshot.Status.Phase == sandboxv1alpha1.SandboxSnapshotPhaseFailed || snapshot.Status.Phase == sandboxv1alpha1.SandboxSnapshotPhaseSucceed) {
		log.Info("Deleting terminal SandboxSnapshot for retry", "snapshot", snapshotName, "phase", snapshot.Status.Phase)
		if err := r.Delete(ctx, snapshot); err != nil && !errors.IsNotFound(err) {
			return ctrl.Result{}, err
		}
		return ctrl.Result{RequeueAfter: time.Second}, nil
	}
	if err != nil && !errors.IsNotFound(err) {
		return ctrl.Result{}, err
	}

	_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionPauseFailed, sandboxv1alpha1.ConditionFalse, "", "")

	if err := r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhasePausing, ""); err != nil {
		return ctrl.Result{}, err
	}

	if errors.IsNotFound(err) {
		if created, err := r.ensureInternalPauseSnapshot(ctx, bs, snapshotName); err != nil {
			return ctrl.Result{}, err
		} else if !created {
			log.Info("Waiting for task cleanup before creating SandboxSnapshot", "snapshot", snapshotName)
		}
	}

	return ctrl.Result{RequeueAfter: time.Second}, nil
}

func (r *BatchSandboxReconciler) ensureInternalPauseSnapshot(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox, snapshotName string) (bool, error) {
	log := logf.FromContext(ctx)

	tasksStopped, err := r.stopTasksBeforePause(ctx, bs)
	if err != nil {
		return false, err
	}
	if !tasksStopped {
		return false, nil
	}

	snapshot := &sandboxv1alpha1.SandboxSnapshot{
		ObjectMeta: metav1.ObjectMeta{
			Name:      snapshotName,
			Namespace: bs.Namespace,
		},
		Spec: sandboxv1alpha1.SandboxSnapshotSpec{
			SandboxName: bs.Name,
		},
	}
	if err := controllerutil.SetControllerReference(bs, snapshot, r.Scheme); err != nil {
		return false, err
	}
	if err := r.Create(ctx, snapshot); err != nil {
		if errors.IsAlreadyExists(err) {
			return true, nil
		}
		return false, err
	}
	log.Info("Created SandboxSnapshot", "snapshot", snapshotName)
	return true, nil
}

func (r *BatchSandboxReconciler) stopTasksBeforePause(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (bool, error) {
	log := logf.FromContext(ctx)
	taskStrategy := strategy.NewTaskSchedulingStrategy(bs)
	if !taskStrategy.NeedTaskScheduling() {
		return true, nil
	}

	poolStrategy := strategy.NewPoolStrategy(bs)
	pods, err := r.listPods(ctx, poolStrategy, bs)
	if err != nil {
		return false, err
	}
	sch, err := r.getTaskScheduler(ctx, bs, pods)
	if err != nil {
		return false, err
	}

	stoppingTasks := sch.StopTask()
	if len(stoppingTasks) > 0 {
		log.Info("Stopping tasks before pause", "count", len(stoppingTasks))
	}

	if err := sch.Schedule(); err != nil {
		return false, fmt.Errorf("failed to stop tasks before pause: %w", err)
	}
	unfinishedTasks := r.getTasksCleanupUnfinished(bs, sch)
	if len(unfinishedTasks) > 0 {
		log.Info("Task cleanup before pause is unfinished", "unfinishedCount", len(unfinishedTasks))
		return false, nil
	}
	log.Info("Task cleanup before pause is finished")
	return true, nil
}

// handleResume implements the resume flow:
// 1. ACK (pauseObservedGeneration + phase=Resuming)
// 2. Requeue so a subsequent pass can consume the snapshot result
func (r *BatchSandboxReconciler) handleResume(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (ctrl.Result, error) {
	log := logf.FromContext(ctx)

	_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionFalse, "", "")

	log.Info("ACK Resuming phase")
	if err := r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhaseResuming, ""); err != nil {
		return ctrl.Result{}, err
	}

	return ctrl.Result{RequeueAfter: time.Second}, nil
}

// syncPauseOrClear waits for the internal pause snapshot to finish and transitions the
// BatchSandbox into Paused or a retryable/terminal failure state.
func (r *BatchSandboxReconciler) syncPauseOrClear(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (ctrl.Result, error) {
	log := logf.FromContext(ctx)

	snapshot := &sandboxv1alpha1.SandboxSnapshot{}
	err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: internalPauseSnapshotName(bs.Name)}, snapshot)
	if errors.IsNotFound(err) {
		snapshotName := internalPauseSnapshotName(bs.Name)
		if created, createErr := r.ensureInternalPauseSnapshot(ctx, bs, snapshotName); createErr != nil {
			return ctrl.Result{}, createErr
		} else if !created {
			log.Info("SandboxSnapshot not created yet; waiting for task cleanup", "snapshot", snapshotName)
		}
		return ctrl.Result{RequeueAfter: time.Second}, nil
	}
	if err != nil {
		return ctrl.Result{}, err
	}

	switch snapshot.Status.Phase {
	case sandboxv1alpha1.SandboxSnapshotPhaseSucceed:
		log.Info("SandboxSnapshot Succeed, completing pause")
		if err := r.completePause(ctx, bs); err != nil {
			return ctrl.Result{}, err
		}
		return ctrl.Result{RequeueAfter: time.Second}, nil
	case sandboxv1alpha1.SandboxSnapshotPhaseFailed:
		msg := snapshotFailureMessage(snapshot)
		log.Info("SandboxSnapshot Failed", "message", msg)

		phase := sandboxv1alpha1.BatchSandboxPhaseSucceed
		reason := "SnapshotFailed"
		if _, podErr := r.findPodForSandbox(ctx, bs); podErr != nil {
			phase = sandboxv1alpha1.BatchSandboxPhaseFailed
			reason = "PodNotFound"
		}

		if err := r.ackPauseWithPhase(ctx, bs, phase, ""); err != nil {
			return ctrl.Result{}, err
		}
		if err := r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionPauseFailed, sandboxv1alpha1.ConditionTrue, reason, msg); err != nil {
			return ctrl.Result{}, err
		}
		return ctrl.Result{}, nil
	case sandboxv1alpha1.SandboxSnapshotPhasePending, sandboxv1alpha1.SandboxSnapshotPhaseCommitting:
		log.Info("SandboxSnapshot in progress", "phase", snapshot.Status.Phase)
		return ctrl.Result{RequeueAfter: time.Second}, nil
	default:
		return ctrl.Result{RequeueAfter: time.Second}, nil
	}
}

// completePause finalizes the pause operation:
//  1. Normal mode: delete all Pods (cascade via OwnerRef)
//  2. Pool mode: solidify the source Pod template and clear poolRef. Pool Controller GC
//     then releases and deletes the old allocated pool Pod.
//  3. Set phase=Paused only after pod deletion/release is accepted.
//  4. spec.pause remains unchanged; the next external request (or server retry bridge)
//     is responsible for changing it.
func (r *BatchSandboxReconciler) completePause(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) error {
	log := logf.FromContext(ctx)
	poolStrategy := strategy.NewPoolStrategy(bs)
	wasPooled := poolStrategy.IsPooledMode()

	pods, err := r.listPods(ctx, poolStrategy, bs)
	if err != nil {
		return err
	}

	var pooledTemplate *corev1.PodTemplateSpec
	if wasPooled {
		if len(pods) == 0 {
			return fmt.Errorf("no allocated pods found for pooled BatchSandbox %s/%s", bs.Namespace, bs.Name)
		}
		pooledTemplate = sourcePodTemplateForPause(pods[0])
		if err := retry.RetryOnConflict(retry.DefaultBackoff, func() error {
			latest := &sandboxv1alpha1.BatchSandbox{}
			if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: bs.Name}, latest); err != nil {
				return err
			}
			patch := client.MergeFrom(latest.DeepCopy())
			latest.Spec.Template = pooledTemplate.DeepCopy()
			latest.Spec.PoolRef = ""
			controllerutil.RemoveFinalizer(latest, finalizerPoolAllocation)
			if latest.Annotations != nil {
				delete(latest.Annotations, annoAllocReleaseKey)
			}
			return r.Patch(ctx, latest, patch)
		}); err != nil {
			return err
		}
		bs.Spec.Template = pooledTemplate.DeepCopy()
		bs.Spec.PoolRef = ""
		controllerutil.RemoveFinalizer(bs, finalizerPoolAllocation)
		if bs.Annotations != nil {
			delete(bs.Annotations, annoAllocReleaseKey)
		}
		log.Info("Detached pooled BatchSandbox after pause", "sourcePod", pods[0].Name)
	}

	controllerKey := controllerutils.GetControllerKey(bs)
	batchSandboxScaleExpectations.DeleteExpectations(controllerKey)
	log.Info("Cleared scale expectations before pod deletion", "controllerKey", controllerKey)

	if !wasPooled {
		for _, pod := range pods {
			if err := r.Delete(ctx, pod); err != nil && !errors.IsNotFound(err) {
				log.Error(err, "Failed to delete pod during pause", "pod", pod.Name)
				return err
			}
			log.Info("Deleted pod during pause", "pod", pod.Name)
		}
	}

	r.deleteTaskScheduler(ctx, bs)

	var latest *sandboxv1alpha1.BatchSandbox
	if err := retry.RetryOnConflict(retry.DefaultBackoff, func() error {
		latest = &sandboxv1alpha1.BatchSandbox{}
		if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: bs.Name}, latest); err != nil {
			return err
		}
		latest.Status.Phase = sandboxv1alpha1.BatchSandboxPhasePaused
		// Pool pause rewrites spec.template and clears spec.poolRef, which advances generation.
		// Acknowledge that controller-owned spec change only while pause is still requested,
		// so spec.pause=true is not replayed and a queued resume request is still observed.
		if latest.Spec.Pause != nil && *latest.Spec.Pause {
			latest.Status.PauseObservedGeneration = latest.Generation
		}
		applyBatchSandboxPhaseConditions(&latest.Status)
		return r.Status().Update(ctx, latest)
	}); err != nil {
		return err
	}
	r.StatusRVExpectation.Expect(latest)

	return nil
}

// continueResume continues the resume flow:
//  1. Read SandboxSnapshot status for image URIs
//  2. Replace template container images
//  3. Pool mode: clear poolRef
//  4. Leave spec.pause and spec.replicas untouched; normal reconciliation recreates
//     the single supported runtime replica from the rewritten template.
func (r *BatchSandboxReconciler) continueResume(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (ctrl.Result, error) {
	log := logf.FromContext(ctx)

	snapshot := &sandboxv1alpha1.SandboxSnapshot{}
	if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: internalPauseSnapshotName(bs.Name)}, snapshot); err != nil {
		if errors.IsNotFound(err) {
			readyPodExists, readyErr := r.hasReadyResumePod(ctx, bs)
			if readyErr != nil {
				return ctrl.Result{}, readyErr
			}
			if readyPodExists {
				log.Info("SandboxSnapshot missing for resume, but a ready pod already exists; treating resume as complete")
				_ = r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhaseSucceed, "")
				_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionFalse, "", "")
				_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionPodFailed, sandboxv1alpha1.ConditionFalse, "", "")
				return ctrl.Result{}, nil
			}
			log.Info("SandboxSnapshot not found for resume, rolling back to Paused")
			_ = r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhasePaused, "")
			_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionTrue, "SnapshotNotFound", "SandboxSnapshot not found")
			return ctrl.Result{}, nil
		}
		return ctrl.Result{}, err
	}

	if snapshot.Status.Phase != sandboxv1alpha1.SandboxSnapshotPhaseSucceed {
		msg := fmt.Sprintf("snapshot not ready: phase=%s", snapshot.Status.Phase)
		log.Error(nil, msg)
		_ = r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhasePaused, "")
		_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionTrue, "SnapshotNotReady", msg)
		return ctrl.Result{}, nil
	}

	imageMap := make(map[string]string)
	for _, c := range snapshot.Status.Containers {
		if snapshot.Status.Format == "" && c.ImageDigest == "" {
			// Snapshots created before format/digest publication are kept
			// restorable for compatibility. New snapshots always use digests.
			imageMap[c.ContainerName] = c.ImageURI
			continue
		}
		immutableImage, err := snapshotcontract.ImmutableImageReference(c.ImageURI, c.ImageDigest)
		if err != nil {
			msg := fmt.Sprintf("invalid immutable snapshot image for container %s: %v", c.ContainerName, err)
			_ = r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhasePaused, "")
			_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionTrue, "InvalidSnapshotImage", msg)
			return ctrl.Result{}, nil
		}
		imageMap[c.ContainerName] = immutableImage
	}
	if snapshot.Status.Format == sandboxv1alpha1.SandboxSnapshotFormatQEMUV1 {
		if bs.Spec.Template == nil {
			msg := "qemu-v1 restore requires a concrete pod template"
			_ = r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhasePaused, "")
			_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionTrue, "InvalidQEMURestore", msg)
			return ctrl.Result{}, nil
		}
		if err := injectQEMURestore(bs.Spec.Template.DeepCopy(), snapshot); err != nil {
			msg := fmt.Sprintf("invalid QEMU restore plan: %v", err)
			_ = r.ackPauseWithPhase(ctx, bs, sandboxv1alpha1.BatchSandboxPhasePaused, "")
			_ = r.setCondition(ctx, bs, sandboxv1alpha1.BatchSandboxConditionResumeFailed, sandboxv1alpha1.ConditionTrue, "InvalidQEMURestore", msg)
			return ctrl.Result{}, nil
		}
	}

	var patched *sandboxv1alpha1.BatchSandbox
	if err := retry.RetryOnConflict(retry.DefaultBackoff, func() error {
		latest := &sandboxv1alpha1.BatchSandbox{}
		if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: bs.Name}, latest); err != nil {
			return err
		}
		patch := client.MergeFrom(latest.DeepCopy())

		if latest.Spec.Template != nil {
			for i := range latest.Spec.Template.Spec.Containers {
				if img, ok := imageMap[latest.Spec.Template.Spec.Containers[i].Name]; ok {
					latest.Spec.Template.Spec.Containers[i].Image = img
				}
			}
			if err := injectQEMURestore(latest.Spec.Template, snapshot); err != nil {
				return err
			}
			ensureImagePullSecret(latest.Spec.Template, r.ResumePullSecret)
		}

		if latest.Spec.PoolRef != "" {
			latest.Spec.PoolRef = ""
			controllerutil.RemoveFinalizer(latest, finalizerPoolAllocation)
		}

		if err := r.Patch(ctx, latest, patch); err != nil {
			return err
		}
		patched = latest.DeepCopy()
		return nil
	}); err != nil {
		return ctrl.Result{}, err
	}

	if patched != nil {
		bs.ObjectMeta = patched.ObjectMeta
		bs.Spec = patched.Spec
	}

	return ctrl.Result{RequeueAfter: time.Second}, nil
}

func injectQEMURestore(template *corev1.PodTemplateSpec, snapshotObject *sandboxv1alpha1.SandboxSnapshot) error {
	if snapshotObject.Status.Format == "" || snapshotObject.Status.Format == sandboxv1alpha1.SandboxSnapshotFormatRootfsV1 {
		return nil
	}
	if snapshotObject.Status.Format != sandboxv1alpha1.SandboxSnapshotFormatQEMUV1 {
		return fmt.Errorf("unsupported snapshot format %q", snapshotObject.Status.Format)
	}
	vm := snapshotObject.Status.VirtualMachine
	if vm == nil {
		return fmt.Errorf("qemu-v1 snapshot has no virtualMachine result")
	}
	immutableImage, err := snapshotcontract.ImmutableImageReference(vm.ImageURI, vm.ImageDigest)
	if err != nil {
		return fmt.Errorf("invalid immutable VM state image: %w", err)
	}
	manifest := snapshotcontract.VMStateManifest{
		FormatVersion: snapshotcontract.VMStateFormatVersion1,
		PayloadDigest: vm.PayloadDigest,
		PayloadSize:   vm.SizeBytes,
		Compression:   vm.Compression,
		Compatibility: snapshotcontract.QEMUCompatibility{
			Architecture:      vm.Compatibility.Architecture,
			QEMUVersion:       vm.Compatibility.QEMUVersion,
			MachineType:       vm.Compatibility.MachineType,
			CPUModel:          vm.Compatibility.CPUModel,
			VCPUs:             vm.Compatibility.VCPUs,
			MemoryBytes:       vm.Compatibility.MemoryBytes,
			QEMUConfigDigest:  vm.Compatibility.QEMUConfigDigest,
			RequiredNodeClass: vm.Compatibility.RequiredNodeClass,
		},
	}
	if err := manifest.Validate(); err != nil {
		return fmt.Errorf("invalid VM state manifest summary: %w", err)
	}
	if _, err := snapshotcontract.ParseSHA256Digest(vm.ManifestDigest); err != nil {
		return fmt.Errorf("invalid VM state manifest digest: %w", err)
	}
	qemuContainerName := template.Annotations[snapshotcontract.AnnotationQEMUContainer]
	if qemuContainerName == "" {
		return fmt.Errorf("QEMU restore template is missing annotation %q", snapshotcontract.AnnotationQEMUContainer)
	}
	containerIndex := -1
	for i := range template.Spec.Containers {
		if template.Spec.Containers[i].Name == qemuContainerName {
			containerIndex = i
			break
		}
	}
	if containerIndex < 0 {
		return fmt.Errorf("QEMU restore container %q does not exist", qemuContainerName)
	}

	volumeSizeLimit := vmStateRestoreStorageSize(vm.SizeBytes)
	if err := ensureVMStateVolume(&template.Spec, volumeSizeLimit); err != nil {
		return err
	}
	target := &template.Spec.Containers[containerIndex]
	if err := ensureVMStateMount(target); err != nil {
		return err
	}
	setContainerEnv(target, "OPENSANDBOX_RESTORE_MODE", snapshotcontract.VMStateFormatVersion1)
	setContainerEnv(target, "OPENSANDBOX_VMSTATE_DIR", vmStateRestoreMountPath)

	restoreInit := corev1.Container{
		Name:            vmStateRestoreInitName,
		Image:           immutableImage,
		ImagePullPolicy: corev1.PullIfNotPresent,
		Command:         []string{"/usr/local/bin/vmstate-loader"},
		Args: []string{
			"restore",
			"--source", "/opensandbox/checkpoint",
			"--target", vmStateRestoreMountPath,
			"--manifest-digest", vm.ManifestDigest,
		},
		VolumeMounts: []corev1.VolumeMount{{Name: vmStateRestoreVolumeName, MountPath: vmStateRestoreMountPath}},
		// Reserve ephemeral storage for the copied VM state payload so the
		// scheduler only places resumed Pods on nodes with enough local disk.
		Resources: mergeResourceRequirements(corev1.ResourceRequirements{}, corev1.ResourceRequirements{
			Requests: corev1.ResourceList{corev1.ResourceEphemeralStorage: volumeSizeLimit.DeepCopy()},
			Limits:   corev1.ResourceList{corev1.ResourceEphemeralStorage: volumeSizeLimit.DeepCopy()},
		}),
		SecurityContext: &corev1.SecurityContext{
			RunAsUser: ptr.To(int64(0)),
			// The init container runs as root even when the source Pod template
			// sets pod-level runAsNonRoot, so override the inherited value.
			RunAsNonRoot:             ptr.To(false),
			AllowPrivilegeEscalation: ptr.To(false),
			Capabilities:             &corev1.Capabilities{Drop: []corev1.Capability{"ALL"}},
		},
	}
	initIndex := -1
	for i := range template.Spec.InitContainers {
		if template.Spec.InitContainers[i].Name == vmStateRestoreInitName {
			initIndex = i
			break
		}
	}
	if initIndex < 0 {
		// Kubernetes executes init containers in list order, so this preserves all
		// user init containers and runs VM state restore last.
		template.Spec.InitContainers = append(template.Spec.InitContainers, restoreInit)
	} else {
		template.Spec.InitContainers[initIndex] = restoreInit
	}

	if vm.Compatibility.RequiredNodeClass != "" {
		if template.Spec.NodeSelector == nil {
			template.Spec.NodeSelector = make(map[string]string)
		}
		if existing := template.Spec.NodeSelector[snapshotcontract.LabelQEMUNodeClass]; existing != "" && existing != vm.Compatibility.RequiredNodeClass {
			return fmt.Errorf("QEMU node class %q conflicts with template selector %q", vm.Compatibility.RequiredNodeClass, existing)
		}
		template.Spec.NodeSelector[snapshotcontract.LabelQEMUNodeClass] = vm.Compatibility.RequiredNodeClass
	}
	return nil
}

func ensureVMStateVolume(spec *corev1.PodSpec, sizeLimit resource.Quantity) error {
	for i := range spec.Volumes {
		if spec.Volumes[i].Name != vmStateRestoreVolumeName {
			continue
		}
		if spec.Volumes[i].EmptyDir == nil {
			return fmt.Errorf("reserved VM state restore volume %q is not an emptyDir", vmStateRestoreVolumeName)
		}
		return nil
	}
	spec.Volumes = append(spec.Volumes, corev1.Volume{
		Name: vmStateRestoreVolumeName,
		VolumeSource: corev1.VolumeSource{EmptyDir: &corev1.EmptyDirVolumeSource{
			SizeLimit: ptr.To(sizeLimit.DeepCopy()),
		}},
	})
	return nil
}

// vmStateRestoreStorageSize sizes the restore emptyDir and the init container's
// ephemeral storage. It covers the compressed payload recorded in the snapshot
// plus headroom for the manifest and the copied vmstate-loader binary.
func vmStateRestoreStorageSize(payloadSize int64) resource.Quantity {
	if payloadSize < 0 {
		payloadSize = 0
	}
	return *resource.NewQuantity(payloadSize+vmStateRestoreHeadroomBytes, resource.BinarySI)
}

func ensureVMStateMount(container *corev1.Container) error {
	for i := range container.VolumeMounts {
		mount := &container.VolumeMounts[i]
		if mount.Name == vmStateRestoreVolumeName {
			if mount.MountPath != vmStateRestoreMountPath {
				return fmt.Errorf("reserved VM state volume is already mounted at %q", mount.MountPath)
			}
			return nil
		}
		if mount.MountPath == vmStateRestoreMountPath {
			return fmt.Errorf("VM state restore path %q is already used by volume %q", vmStateRestoreMountPath, mount.Name)
		}
	}
	container.VolumeMounts = append(container.VolumeMounts, corev1.VolumeMount{Name: vmStateRestoreVolumeName, MountPath: vmStateRestoreMountPath})
	return nil
}

func setContainerEnv(container *corev1.Container, name, value string) {
	for i := range container.Env {
		if container.Env[i].Name == name {
			container.Env[i].Value = value
			container.Env[i].ValueFrom = nil
			return
		}
	}
	container.Env = append(container.Env, corev1.EnvVar{Name: name, Value: value})
}

func (r *BatchSandboxReconciler) ackPauseWithPhase(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox, phase sandboxv1alpha1.BatchSandboxPhase, _ string) error {
	var latest *sandboxv1alpha1.BatchSandbox
	if err := retry.RetryOnConflict(retry.DefaultBackoff, func() error {
		latest = &sandboxv1alpha1.BatchSandbox{}
		if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: bs.Name}, latest); err != nil {
			return err
		}
		latest.Status.PauseObservedGeneration = latest.Generation
		latest.Status.Phase = phase
		applyBatchSandboxPhaseConditions(&latest.Status)
		return r.Status().Update(ctx, latest)
	}); err != nil {
		return err
	}
	r.StatusRVExpectation.Expect(latest)
	bs.Status.PauseObservedGeneration = bs.Generation
	bs.Status.Phase = phase
	applyBatchSandboxPhaseConditions(&bs.Status)
	return nil
}

// findPodForSandbox finds the Pod associated with a BatchSandbox.
func (r *BatchSandboxReconciler) findPodForSandbox(ctx context.Context, bs *sandboxv1alpha1.BatchSandbox) (*corev1.Pod, error) {
	poolStrategy := strategy.NewPoolStrategy(bs)
	pods, err := r.listPods(ctx, poolStrategy, bs)
	if err != nil {
		return nil, err
	}
	if len(pods) == 0 {
		return nil, fmt.Errorf("no pods found for BatchSandbox %s/%s", bs.Namespace, bs.Name)
	}
	return pods[0], nil
}

func (r *BatchSandboxReconciler) setCondition(
	ctx context.Context,
	bs *sandboxv1alpha1.BatchSandbox,
	conditionType sandboxv1alpha1.BatchSandboxConditionType,
	status string,
	reason string,
	message string,
) error {
	var latest *sandboxv1alpha1.BatchSandbox
	if err := retry.RetryOnConflict(retry.DefaultBackoff, func() error {
		latest = &sandboxv1alpha1.BatchSandbox{}
		if err := r.Get(ctx, types.NamespacedName{Namespace: bs.Namespace, Name: bs.Name}, latest); err != nil {
			return err
		}

		var conditions []sandboxv1alpha1.BatchSandboxCondition
		found := false
		for _, c := range latest.Status.Conditions {
			if c.Type == conditionType {
				if status == sandboxv1alpha1.ConditionFalse {
					continue
				}
				c.Status = status
				c.Reason = reason
				c.Message = message
				c.LastTransitionTime = ptr.To(metav1.Now())
				found = true
			}
			conditions = append(conditions, c)
		}

		if !found && status == sandboxv1alpha1.ConditionTrue {
			conditions = append(conditions, sandboxv1alpha1.BatchSandboxCondition{
				Type:               conditionType,
				Status:             status,
				Reason:             reason,
				Message:            message,
				LastTransitionTime: ptr.To(metav1.Now()),
			})
		}

		latest.Status.Conditions = conditions
		return r.Status().Update(ctx, latest)
	}); err != nil {
		return err
	}
	r.StatusRVExpectation.Expect(latest)
	return nil
}