Merge 修复 runner 工作目录权限

.gitea/workflows/pod-smoke.yml

@@ -1,5 +1,7 @@
+---
 name: dynamic Pod smoke test
 
+# yamllint disable-line rule:truthy
 on:
   workflow_dispatch:
 
@@ -17,3 +19,24 @@ jobs:
             -socketPath /run/spire/agent-sockets/spire-agent.sock \
             >/dev/null
           test "$(id -u)" = 2000
+
+      - name: Build and run image
+        shell: bash
+        run: |
+          set -euo pipefail
+          findmnt /var/lib/docker
+          context=$(mktemp -d)
+          cleanup() {
+            docker image rm --force pod-docker-smoke:test \
+              >/dev/null 2>&1 || true
+            rm -rf -- "$context"
+          }
+          trap cleanup EXIT
+          printf '%s\n' \
+            'FROM alpine:3.22' \
+            'RUN printf pod-docker-ok >/result' \
+            >"$context/Dockerfile"
+          docker build --tag pod-docker-smoke:test "$context"
+          output=$(docker run --rm pod-docker-smoke:test cat /result)
+          test "$output" = pod-docker-ok
+          test "$(docker info --format '{{.Driver}}')" = overlay2

.gitea/workflows/publish-images.yml

@@ -1,26 +1,22 @@
 ---
-name: publish images
+name: publish controller image
 
 on:
   push:
     branches: [main]
     paths:
       - '.gitea/workflows/publish-images.yml'
-      - 'config/**'
+      - 'container/controller.Dockerfile'
-      - 'container/**'
       - 'cmd/**'
       - 'internal/**'
-      - 'scripts/**'
+      - 'scripts/publish-image'
-      - 'src/**'
       - 'go.mod'
       - 'go.sum'
-      - 'pyproject.toml'
-      - 'README.md'
   workflow_dispatch:
 
 jobs:
   publish-images:
-    name: publish-images
+    name: publish-controller
     runs-on: [self-hosted, vm]
     timeout-minutes: 45
     permissions:
@@ -28,101 +24,13 @@ jobs:
     env:
       PUSH_REGISTRY: zot-push.ad.ddupan.top
       PULL_REGISTRY: zot.ad.ddupan.top
-      CONTROLLER_REPOSITORY: panxiao81/gitea-dynamic-runner-controller
+      IMAGE_NAME: controller
-      RUNNER_REPOSITORY: panxiao81/gitea-dynamic-runner-runner
+      IMAGE_REPOSITORY: panxiao81/gitea-dynamic-runner-controller
+      IMAGE_DOCKERFILE: container/controller.Dockerfile
       SPIRE_AGENT_SOCKET: /run/spire/agent-sockets/spire-agent.sock
     steps:
       - uses: actions/checkout@v4
 
-      - name: Test source
-        shell: bash
-        run: |
-          set -euo pipefail
-          go test ./...
-          go vet ./...
-          python3 -m pip install --break-system-packages -e '.[test]'
-          pytest -q
-          python3 -m compileall -q src tests
-          apt-get update
-          apt-get install --yes --no-install-recommends shellcheck
-          shellcheck scripts/*
-
       - name: Build and publish
         shell: bash
-        run: |
+        run: scripts/publish-image
-          set -euo pipefail
-          set +x
-
-          : "${GITHUB_SHA:?GITHUB_SHA is required}"
-          image_tag="sha-${GITHUB_SHA}"
-          docker_config=$(mktemp -d)
-          jwt_file=$(mktemp)
-          cleanup() {
-            docker buildx rm ci-builder >/dev/null 2>&1 || true
-            rm -rf -- "$docker_config" "$jwt_file"
-          }
-          trap cleanup EXIT
-          export DOCKER_CONFIG="$docker_config"
-
-          /opt/spire/bin/spire-agent api fetch jwt \
-            -audience zot \
-            -socketPath "$SPIRE_AGENT_SOCKET" \
-            -output json >"$jwt_file"
-          # shellcheck disable=SC2016
-          jq -er '.[0].svids[0].svid' "$jwt_file" | \
-            docker login "$PUSH_REGISTRY" --username zot --password-stdin
-
-          docker buildx create \
-            --name ci-builder \
-            --driver docker-container \
-            --use
-
-          publish() {
-            local repository=$1
-            local dockerfile=$2
-            local metadata=$3
-            docker buildx build \
-              --builder ci-builder \
-              --platform linux/amd64 \
-              --file "$dockerfile" \
-              --tag "${PUSH_REGISTRY}/${repository}:${image_tag}" \
-              --tag "${PUSH_REGISTRY}/${repository}:main" \
-              --provenance=mode=max \
-              --sbom=true \
-              --metadata-file "$metadata" \
-              --push \
-              .
-          }
-
-          publish \
-            "$CONTROLLER_REPOSITORY" \
-            container/controller.Dockerfile \
-            controller-metadata.json
-          publish \
-            "$RUNNER_REPOSITORY" \
-            container/runner.Dockerfile \
-            runner-metadata.json
-
-          controller_digest=$(
-            # shellcheck disable=SC2016
-            jq -er '."containerimage.digest"' controller-metadata.json
-          )
-          runner_digest=$(
-            # shellcheck disable=SC2016
-            jq -er '."containerimage.digest"' runner-metadata.json
-          )
-          controller_ref="${PULL_REGISTRY}/${CONTROLLER_REPOSITORY}@${controller_digest}"
-          runner_ref="${PULL_REGISTRY}/${RUNNER_REPOSITORY}@${runner_digest}"
-
-          printf 'controller=%s\nrunner=%s\n' "$controller_ref" "$runner_ref"
-          if [[ -n "${GITHUB_STEP_SUMMARY:-}" ]]; then
-            {
-              printf '## Published images\n\n'
-              # shellcheck disable=SC2016
-              printf -- '- Controller: `%s`\n' "$controller_ref"
-              # shellcheck disable=SC2016
-              printf -- '- Runner: `%s`\n' "$runner_ref"
-              # shellcheck disable=SC2016
-              printf -- '- Source: `%s`\n' "$GITHUB_SHA"
-            } >>"$GITHUB_STEP_SUMMARY"
-          fi

.gitea/workflows/publish-runner-image.yml

@@ -0,0 +1,26 @@
+---
+name: publish runner image
+
+on:
+  workflow_dispatch:
+
+jobs:
+  publish-images:
+    name: publish-runner
+    runs-on: [self-hosted, vm]
+    timeout-minutes: 45
+    permissions:
+      contents: read
+    env:
+      PUSH_REGISTRY: zot-push.ad.ddupan.top
+      PULL_REGISTRY: zot.ad.ddupan.top
+      IMAGE_NAME: runner
+      IMAGE_REPOSITORY: panxiao81/gitea-dynamic-runner-runner
+      IMAGE_DOCKERFILE: container/runner.Dockerfile
+      SPIRE_AGENT_SOCKET: /run/spire/agent-sockets/spire-agent.sock
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Build and publish
+        shell: bash
+        run: scripts/publish-image

.gitea/workflows/test.yml

@@ -17,7 +17,7 @@ jobs:
       - run: go vet ./...
 
   python:
-    runs-on: self-hosted
+    runs-on: [self-hosted, pod]
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
@@ -28,7 +28,7 @@ jobs:
       - run: python -m compileall -q src
 
   shell:
-    runs-on: self-hosted
+    runs-on: [self-hosted, pod]
     steps:
       - uses: actions/checkout@v4
       - run: |

.gitea/workflows/vm-kind-smoke.yml

@@ -1,5 +1,7 @@
+---
 name: VM kind smoke
 
+# yamllint disable-line rule:truthy
 on:
   workflow_dispatch:
 
@@ -7,25 +9,123 @@ jobs:
   kind:
     runs-on: [self-hosted, vm]
     steps:
+      - name: Prepare nested kubelet device
+        shell: bash
+        run: |
+          set -euo pipefail
+          if [[ ! -e /dev/kmsg ]]; then
+            sudo mknod /dev/kmsg c 1 11
+          fi
+
       - name: Verify Docker
-        run: docker info
+        shell: bash
+        run: |
+          set -euo pipefail
+          findmnt /var/lib/docker
+          docker info
+          echo '### runner cgroup'
+          cat /proc/self/cgroup
+          cat /sys/fs/cgroup/cgroup.type
+          cat /sys/fs/cgroup/cgroup.controllers
+          cat /sys/fs/cgroup/cgroup.subtree_control
+          echo '### nested private cgroup namespace'
+          docker run --rm --privileged --cgroupns=private alpine:3.22 \
+            sh -c 'cat /proc/self/cgroup; cat /sys/fs/cgroup/cgroup.type'
+          echo '### nested host cgroup namespace'
+          docker run --rm --privileged --cgroupns=host alpine:3.22 \
+            sh -c 'cat /proc/self/cgroup; cat /sys/fs/cgroup/cgroup.type'
 
       - name: Install kind
         shell: bash
         run: |
           set -euo pipefail
           version=v0.33.0
+          base_url="https://kind.sigs.k8s.io/dl/${version}"
           curl --fail --location --silent --show-error \
-            --output /tmp/kind "https://kind.sigs.k8s.io/dl/${version}/kind-linux-amd64"
+            --output /tmp/kind "${base_url}/kind-linux-amd64"
           curl --fail --location --silent --show-error \
-            --output /tmp/kind.sha256sum "https://kind.sigs.k8s.io/dl/${version}/kind-linux-amd64.sha256sum"
+            --output /tmp/kind.sha256sum \
-          printf '%s  %s\n' "$(cut -d ' ' -f1 /tmp/kind.sha256sum)" /tmp/kind | sha256sum --check
+            "${base_url}/kind-linux-amd64.sha256sum"
+          checksum=$(cut -d ' ' -f1 /tmp/kind.sha256sum)
+          printf '%s  %s\n' "$checksum" /tmp/kind | sha256sum --check
           chmod 0755 /tmp/kind
 
       - name: Create and delete kind cluster
         shell: bash
         run: |
           set -euo pipefail
-          trap '/tmp/kind delete cluster --name smoke' EXIT
+          diagnose_and_cleanup() {
-          /tmp/kind create cluster --name smoke --wait 180s
+            status=$?
+            node=smoke-control-plane
+            if (( status != 0 )) && docker inspect "$node" >/dev/null 2>&1; then
+              echo '::group::kind node inspect'
+              docker inspect "$node"
+              echo '::endgroup::'
+              echo '::group::kind node logs'
+              docker logs "$node" 2>&1 || true
+              echo '::endgroup::'
+              echo '::group::kind node guest state'
+              docker exec "$node" bash -c '
+                set +e
+                echo "### pid 1"
+                ps -p 1 -o pid,ppid,user,stat,comm,args
+                cat /proc/1/status
+                echo "### cgroup"
+                cat /proc/1/cgroup
+                findmnt -R /sys/fs/cgroup
+                stat -fc "%T %a" /sys/fs/cgroup
+                echo "### systemd"
+                systemctl --no-pager --failed
+                systemctl --no-pager status \
+                  multi-user.target containerd.service kubelet.service
+                echo "### CRI containers"
+                endpoint=unix:///run/containerd/containerd.sock
+                crictl --runtime-endpoint "$endpoint" ps --all
+                for id in $(
+                  crictl --runtime-endpoint "$endpoint" ps --all --quiet
+                ); do
+                  echo "### CRI container $id"
+                  crictl --runtime-endpoint "$endpoint" inspect "$id"
+                  crictl --runtime-endpoint "$endpoint" logs "$id"
+                done
+                echo "### containerd metadata"
+                timeout 10 ctr --namespace k8s.io containers list
+                timeout 10 ctr --namespace k8s.io snapshots list
+                echo "### runtime process stacks"
+                ps -e -o pid,ppid,stat,wchan:32,comm,args
+                for pid in $(pidof containerd containerd-shim-runc-v2); do
+                  echo "### kernel stack $pid"
+                  cat "/proc/$pid/stack"
+                done
+                kill -USR1 "$(pidof containerd)"
+                sleep 2
+                journalctl --no-pager -b -n 500
+              ' 2>&1 || true
+              echo '::endgroup::'
+            fi
+            /tmp/kind delete cluster --name smoke || true
+            exit "$status"
+          }
+          trap diagnose_and_cleanup EXIT
+          # Nested Kata + Docker + kind cold starts can take longer than
+          # kubeadm's one-minute API-call default even after the static pods
+          # have been accepted. Give the API server enough time to become
+          # responsive before kubeadm creates its initial RBAC objects.
+          cat >/tmp/kind-config.yml <<'EOF'
+          kind: Cluster
+          apiVersion: kind.x-k8s.io/v1alpha4
+          nodes:
+            - role: control-plane
+              kubeadmConfigPatches:
+                - |
+                  apiVersion: kubeadm.k8s.io/v1beta4
+                  kind: InitConfiguration
+                  timeouts:
+                    kubernetesAPICall: 5m0s
+          EOF
+          /tmp/kind create cluster \
+            --name smoke \
+            --config /tmp/kind-config.yml \
+            --wait 300s \
+            --retain
           /tmp/kind get clusters | grep -Fx smoke

.gitea/workflows/vm-runtime-smoke.yml

@@ -0,0 +1,36 @@
+---
+name: VM runtime smoke
+
+# yamllint disable-line rule:truthy
+on:
+  workflow_dispatch:
+
+jobs:
+  runtime:
+    runs-on: [self-hosted, vm]
+    steps:
+      - name: Verify workload identity socket
+        shell: bash
+        run: |
+          set -euo pipefail
+          : "${SPIFFE_ENDPOINT_SOCKET:?SPIFFE_ENDPOINT_SOCKET is required}"
+          socket_path=${SPIFFE_ENDPOINT_SOCKET#unix://}
+          test -S "$socket_path"
+
+      - name: Verify Docker daemon
+        shell: bash
+        run: |
+          set -euo pipefail
+          format='{{json .ServerVersion}} {{json .Driver}}'
+          format="$format {{json .CgroupVersion}}"
+          docker info --format "$format"
+
+      - name: Run and clean nested container
+        shell: bash
+        run: |
+          set -euo pipefail
+          name=vm-runtime-smoke
+          trap 'docker rm --force "$name" >/dev/null 2>&1 || true' EXIT
+          output=$(docker run --name "$name" alpine:3.22 /bin/sh -c \
+            'test "$(uname -m)" = x86_64 && printf vm-runtime-ok')
+          test "$output" = vm-runtime-ok

README.md

@@ -15,6 +15,10 @@ runs-on: [self-hosted, vm]
 只执行一个 job，并在 job 结束后连同本地状态一起销毁。完整的设计约束见
 [`docs/design-principles.md`](docs/design-principles.md)。
 
+集成期间可将 `VM_RUNNER_LABEL=vm-dev`，只接取显式使用
+`runs-on: [self-hosted, vm-dev]` 的测试任务；生产 `vm` job 将保持在 Gitea pending，
+不会在 backend 修复过程中继续涌入。
+
 目标 Go controller 组件：
 
 - `scheduler`：以常驻 Gitea RunnerService 身份直接领取 task，并把完整 assignment
@@ -32,8 +36,9 @@ runs-on: [self-hosted, vm]
   entry；不持有 OpenSandbox API key、Gitea token 或 Bao 凭据。身份与 Pool 契约见
   [`docs/opensandbox-runner.md`](docs/opensandbox-runner.md)。
 - Pod executor：在 Kubernetes 中创建一次性 privileged Pod；Pod 内的 workflow 使用
-  host executor，Docker、BuildKit 和 kind 等工具由 pipeline 按需 setup。Runner 固定在
+  host executor。Runner 固定在支持原生 job hooks 的 3.x 版本，在 workflow 第一步前
-  支持原生 job hooks 的 3.x 版本，在 workflow 第一步前等待实际任务对应的 SVID。
+  等待实际任务对应的 SVID，并启动 job-local Docker daemon；workflow 可直接使用与
+  GitHub-hosted runner 相同的 Docker/BuildKit action。
 - `jwt-broker`：早期共享 Kubernetes runner 的过渡实验；目标架构不部署它，每个
   动态 Pod 或 VM 直接取得自己的 SPIFFE 身份。
 

cmd/gitea-dynamic-runner/controller.go

@@ -4,7 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"log"
+	"log/slog"
 	"net/http"
 	"os"
 	"slices"
@@ -49,7 +49,7 @@ type controllerConfig struct {
 	PodNamespace, PodImage, PodServiceAccount, SPIRECluster, SPIREClass string
 	SPIREAgentID                                                        string
 	PodExecutorUID, PodCapacity                                         int
-	OpenSandboxURL, OpenSandboxAPIKey, OpenSandboxPool                  string
+	OpenSandboxURL, OpenSandboxAPIKey, OpenSandboxPool, VMRunnerLabel   string
 	VMTimeout, VMCapacity                                               int
 }
 
@@ -131,7 +131,7 @@ func runController(ctx context.Context) error {
 		labels = append(labels, string(taskassignment.BackendPod))
 	}
 	if slices.Contains(config.Components, controller.VMWorker) {
-		labels = append(labels, string(taskassignment.BackendVM))
+		labels = append(labels, config.VMRunnerLabel)
 	}
 	poller := taskscheduler.Poller{
 		Client: giteaClient,
@@ -139,7 +139,7 @@ func runController(ctx context.Context) error {
 			JetStream: producerJS, SubjectBase: config.SubjectBase,
 		}},
 		Config:  taskscheduler.PollerConfig{Version: "gitea-dynamic-runner/0.4", Labels: labels, Capacity: config.PodCapacity + config.VMCapacity},
-		OnError: func(err error) { log.Printf("scheduler: %v", err) },
+		OnError: func(err error) { slog.Error("scheduler error", "component", "scheduler", "error", err) },
 	}
 	kubernetesConfig, err := rest.InClusterConfig()
 	if err != nil {
@@ -190,11 +190,13 @@ func runController(ctx context.Context) error {
 				return fmt.Errorf("recover Pod facade claim %s: %w", assignment.ID, err)
 			}
 		}
-		component, err := workerComponent(ctx, workerJS, config, taskassignment.BackendPod, config.PodCapacity, taskworker.Worker{Backend: backend, Bootstrap: bootstrap}, registry, podPool)
+		component, err := workerComponent(ctx, workerJS, config, taskassignment.BackendPod, config.PodCapacity, taskworker.Worker{Backend: backend, Bootstrap: bootstrap, OnEvent: workerEventLogger(taskassignment.BackendPod)}, registry, podPool)
 		if err != nil {
 			return err
 		}
-		lifecycle := podbackend.Lifecycle{Backend: backend, OnError: func(err error) { log.Printf("pod lifecycle: %v", err) }}
+		lifecycle := podbackend.Lifecycle{Backend: backend, OnError: func(err error) {
+			slog.Error("backend lifecycle error", "component", "lifecycle", "backend", taskassignment.BackendPod, "error", err)
+		}}
 		components[controller.PodWorker] = runComponent(func(ctx context.Context) error {
 			group, groupContext := errgroup.WithContext(ctx)
 			group.Go(func() error { return component.Run(groupContext) })
@@ -220,11 +222,13 @@ func runController(ctx context.Context) error {
 				return fmt.Errorf("recover VM facade claim %s: %w", assignment.ID, err)
 			}
 		}
-		component, err := workerComponent(ctx, workerJS, config, taskassignment.BackendVM, config.VMCapacity, taskworker.Worker{Backend: backend, Bootstrap: bootstrap}, registry, vmPool)
+		component, err := workerComponent(ctx, workerJS, config, taskassignment.BackendVM, config.VMCapacity, taskworker.Worker{Backend: backend, Bootstrap: bootstrap, OnEvent: workerEventLogger(taskassignment.BackendVM)}, registry, vmPool)
 		if err != nil {
 			return err
 		}
-		lifecycleReconciler := opensandboxbackend.LifecycleReconciler{Backend: backend, OnError: func(err error) { log.Printf("VM lifecycle: %v", err) }}
+		lifecycleReconciler := opensandboxbackend.LifecycleReconciler{Backend: backend, OnError: func(err error) {
+			slog.Error("backend lifecycle error", "component", "lifecycle", "backend", taskassignment.BackendVM, "error", err)
+		}}
 		components[controller.VMWorker] = runComponent(func(ctx context.Context) error {
 			group, groupContext := errgroup.WithContext(ctx)
 			group.Go(func() error { return component.Run(groupContext) })
@@ -294,11 +298,21 @@ func workerComponent(ctx context.Context, js jetstream.JetStream, config control
 	}
 	return assignmentqueue.ConsumerComponent{
 		Consumer: consumer, Capacity: capacity,
-		Processor: assignmentqueue.Processor{TrustDomain: config.TrustDomain, Accepter: accepter, Claims: claims, Admission: admission},
+		Processor: assignmentqueue.Processor{TrustDomain: config.TrustDomain, Accepter: accepter, Claims: claims, Admission: admission, OnEvent: func(event assignmentqueue.Event) {
-		OnError:   func(err error) { log.Printf("%s worker: %v", backend, err) },
+			slog.Info("assignment transition", "component", "worker", "event", event.Name, "backend", event.Backend, "assignment", event.AssignmentID, "retry_delay", event.RetryDelay)
+		}},
+		OnError: func(err error) {
+			slog.Error("assignment processing error", "component", "worker", "backend", backend, "error", err)
+		},
 	}, nil
 }
 
+func workerEventLogger(backend taskassignment.Backend) func(taskworker.Event) {
+	return func(event taskworker.Event) {
+		slog.Info("executor transition", "component", "worker", "event", event.Name, "backend", backend, "assignment", event.AssignmentID, "executor", event.Executor, "phase", event.Phase)
+	}
+}
+
 func loadControllerConfig() (controllerConfig, error) {
 	selection, err := controller.ParseSelection(os.Getenv("COMPONENTS"))
 	if err != nil {
@@ -345,7 +359,7 @@ func loadControllerConfig() (controllerConfig, error) {
 		FacadeListen: env("RUNNER_FACADE_LISTEN", ":8443"), FacadeURL: os.Getenv("RUNNER_FACADE_URL"), FacadeSPIFFEID: os.Getenv("RUNNER_FACADE_SPIFFE_ID"), CapabilityKey: []byte(capabilityKey),
 		PodNamespace: env("POD_NAMESPACE", "gitea-actions"), PodImage: os.Getenv("POD_EXECUTOR_IMAGE"), PodServiceAccount: env("POD_SERVICE_ACCOUNT", "gitea-task-executor"),
 		SPIRECluster: env("SPIRE_CLUSTER", "homelab"), SPIREClass: env("SPIRE_CLASS", "spire-mgmt-spire"), SPIREAgentID: os.Getenv("SPIRE_AGENT_ID"), PodExecutorUID: envInt("POD_EXECUTOR_UID", 2000), PodCapacity: envInt("POD_CAPACITY", 4),
-		OpenSandboxURL: os.Getenv("OPENSANDBOX_API"), OpenSandboxPool: env("OPENSANDBOX_POOL", "ci-vm"), VMTimeout: envInt("VM_TIMEOUT_SECONDS", 14400), VMCapacity: envInt("VM_CAPACITY", 1),
+		OpenSandboxURL: os.Getenv("OPENSANDBOX_API"), OpenSandboxPool: env("OPENSANDBOX_POOL", "ci-vm"), VMRunnerLabel: env("VM_RUNNER_LABEL", "vm"), VMTimeout: envInt("VM_TIMEOUT_SECONDS", 14400), VMCapacity: envInt("VM_CAPACITY", 1),
 	}
 	if config.WorkloadAPIAddr == "" || config.FacadeURL == "" || config.FacadeSPIFFEID == "" {
 		return controllerConfig{}, errors.New("SPIFFE_ENDPOINT_SOCKET, RUNNER_FACADE_URL, and RUNNER_FACADE_SPIFFE_ID are required")
@@ -357,6 +371,9 @@ func loadControllerConfig() (controllerConfig, error) {
 		return controllerConfig{}, errors.New("SPIRE_AGENT_ID is required for pod-worker")
 	}
 	if slices.Contains(selection, controller.VMWorker) {
+		if config.VMRunnerLabel != "vm" && config.VMRunnerLabel != "vm-dev" {
+			return controllerConfig{}, errors.New("VM_RUNNER_LABEL must be vm or vm-dev")
+		}
 		if config.OpenSandboxURL == "" {
 			return controllerConfig{}, errors.New("OPENSANDBOX_API is required for vm-worker")
 		}

cmd/gitea-dynamic-runner/controller_test.go

@@ -56,6 +56,7 @@ func TestLoadControllerConfigRequiresOpenSandboxSecretOnlyForVM(t *testing.T) {
 	t.Setenv("RUNNER_FACADE_URL", "https://facade:8443")
 	t.Setenv("RUNNER_FACADE_SPIFFE_ID", "spiffe://ddupan.top/controller")
 	t.Setenv("OPENSANDBOX_API", "http://opensandbox.internal")
+	t.Setenv("VM_RUNNER_LABEL", "vm-dev")
 	if _, err := loadControllerConfig(); err == nil {
 		t.Fatal("expected missing OpenSandbox API key file error")
 	}

cmd/gitea-dynamic-runner/main.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"log/slog"
 	"os"
 	"os/signal"
 	"syscall"
@@ -12,8 +13,9 @@ import (
 )
 
 func main() {
+	slog.SetDefault(slog.New(slog.NewJSONHandler(os.Stderr, nil)))
 	if err := run(); err != nil {
-		fmt.Fprintln(os.Stderr, err)
+		slog.Error("runner stopped", "error", err)
 		os.Exit(1)
 	}
 }

container/runner.Dockerfile

@@ -21,7 +21,7 @@ RUN groupadd --gid 2000 runner \
     && useradd --uid 2000 --gid 2000 --groups docker --create-home --shell /bin/bash runner \
     && printf 'runner ALL=(ALL) NOPASSWD:ALL\n' >/etc/sudoers.d/runner \
     && chmod 0440 /etc/sudoers.d/runner \
-    && install -d -o 2000 -g 2000 /data
+    && install -d -o 2000 -g 2000 /data /workspace
 
 COPY --from=runner /usr/local/bin/gitea-runner /usr/local/bin/gitea-runner
 COPY --from=controller /out/gitea-dynamic-runner /usr/local/bin/gitea-dynamic-runner
@@ -29,6 +29,7 @@ COPY --from=spire /opt/spire/bin/spire-agent /opt/spire/bin/spire-agent
 COPY config/runner.yaml /etc/gitea-runner/config.yaml
 COPY --chmod=0755 scripts/gitea-job-started /usr/local/libexec/gitea-job-started
 COPY --chmod=0755 scripts/gitea-opensandbox-runner /usr/local/libexec/gitea-opensandbox-runner
+COPY --chmod=0755 scripts/setup-job-docker /usr/local/libexec/setup-job-docker
 
 VOLUME ["/data"]
 ENV HOME=/home/runner

docs/opensandbox-runner.md

@@ -40,9 +40,11 @@ UID attestation 的临时 Agent 失去父级。
 
 `ci-vm` 使用 `kata-clh-runtime-rs`；`ci-pod` 使用默认 runc。两者都要求：
 
-- runner 镜像包含 Gitea Runner、Node.js action userspace、SPIRE CLI 和 identity gate；
+- runner 镜像包含 Gitea Runner、Node.js action userspace、SPIRE CLI、Docker 工具和
-- runner UID 2000，SPIRE Agent 与 privileged dockerd 使用不同 UID；
+  identity gate；Pod 与 VM backend 使用同一个镜像；
-- Docker socket 通过 group 2000 共享，Docker 数据仅存在于 sandbox emptyDir；
+- runner UID 2000；SPIRE Agent 独立运行；job-started hook 在第一步 workflow 之前
+  启动 job-local Docker daemon，业务 workflow 不负责 runner 基础设施初始化；
+- Kata VM 中 Docker 数据使用 guest 内的 loop-backed ext4，并随 sandbox 一起删除；
 - `self-hosted` 必须是所有 runner labels 的前缀；
 - ephemeral/once runner 完成一项任务后退出。
 

internal/assignmentqueue/jetstream.go

@@ -78,6 +78,22 @@ type Processor struct {
 	Admission    Admission
 	RetryDelay   time.Duration
 	ClaimTimeout time.Duration
+	OnEvent      func(Event)
+}
+
+// Event describes a non-sensitive assignment handoff transition. It never
+// contains task payloads, credentials, capabilities, or workload identities.
+type Event struct {
+	Name         string
+	AssignmentID string
+	Backend      taskassignment.Backend
+	RetryDelay   time.Duration
+}
+
+func (p Processor) event(name string, assignment taskassignment.Assignment, retryDelay time.Duration) {
+	if p.OnEvent != nil {
+		p.OnEvent(Event{Name: name, AssignmentID: assignment.ID, Backend: assignment.Backend, RetryDelay: retryDelay})
+	}
 }
 
 func (p Processor) Process(ctx context.Context, message Message) error {
@@ -88,6 +104,7 @@ func (p Processor) Process(ctx context.Context, message Message) error {
 	if err != nil {
 		return errors.Join(err, message.TermWithReason("invalid assignment"))
 	}
+	p.event("received", assignment, 0)
 	if _, err := p.Claims.Offer(assignment); err != nil {
 		return errors.Join(err, message.TermWithReason("conflicting assignment"))
 	}
@@ -96,8 +113,10 @@ func (p Processor) Process(ctx context.Context, message Message) error {
 		if delay <= 0 {
 			delay = 2 * time.Second
 		}
+		p.event("capacity_wait", assignment, delay)
 		return message.NakWithDelay(delay)
 	}
+	p.event("capacity_acquired", assignment, 0)
 	accepted, err := p.Accepter.Accept(ctx, assignment)
 	if err != nil {
 		p.Admission.Release(assignment.ID)
@@ -105,9 +124,11 @@ func (p Processor) Process(ctx context.Context, message Message) error {
 		if delay <= 0 {
 			delay = 15 * time.Second
 		}
+		p.event("backend_retry", assignment, delay)
 		return errors.Join(err, message.NakWithDelay(delay))
 	}
 	if accepted {
+		p.event("backend_ready", assignment, 0)
 		timeout := p.ClaimTimeout
 		if timeout <= 0 {
 			timeout = 4 * time.Minute
@@ -120,17 +141,21 @@ func (p Processor) Process(ctx context.Context, message Message) error {
 			if delay <= 0 {
 				delay = 2 * time.Second
 			}
+			p.event("claim_timeout", assignment, delay)
 			return errors.Join(err, message.NakWithDelay(delay))
 		}
+		p.event("runner_claimed", assignment, 0)
 		if err := message.DoubleAck(ctx); err != nil {
 			return fmt.Errorf("ack assignment %s: %w", assignment.ID, err)
 		}
+		p.event("acked", assignment, 0)
 		return nil
 	}
 	delay := p.RetryDelay
 	if delay <= 0 {
 		delay = 2 * time.Second
 	}
+	p.event("backend_pending", assignment, delay)
 	return message.NakWithDelay(delay)
 }
 

internal/assignmentqueue/jetstream_test.go

@@ -126,13 +126,23 @@ func encodedAssignment(t *testing.T) []byte {
 
 func TestProcessorAcknowledgesPersistedHandoff(t *testing.T) {
 	message := &fakeMessage{data: encodedAssignment(t)}
-	processor := Processor{TrustDomain: "ddupan.top", Accepter: &fakeAccepter{accepted: true}, Claims: &fakeClaims{claimed: true}, Admission: &fakeAdmission{allowed: true}}
+	var events []Event
+	processor := Processor{TrustDomain: "ddupan.top", Accepter: &fakeAccepter{accepted: true}, Claims: &fakeClaims{claimed: true}, Admission: &fakeAdmission{allowed: true}, OnEvent: func(event Event) { events = append(events, event) }}
 	if err := processor.Process(context.Background(), message); err != nil {
 		t.Fatal(err)
 	}
 	if message.acked != 1 || message.nacked != 0 {
 		t.Fatalf("message = %#v", message)
 	}
+	want := []string{"received", "capacity_acquired", "backend_ready", "runner_claimed", "acked"}
+	if len(events) != len(want) {
+		t.Fatalf("events = %#v", events)
+	}
+	for index := range want {
+		if events[index].Name != want[index] || events[index].AssignmentID != "gitea-task-42" || events[index].Backend != taskassignment.BackendPod {
+			t.Fatalf("event[%d] = %#v", index, events[index])
+		}
+	}
 }
 
 func TestProcessorRetriesUntilBackendHandoffIsDurable(t *testing.T) {

internal/podbackend/client.go

@@ -8,6 +8,7 @@ import (
 
 	corev1 "k8s.io/api/core/v1"
 	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
 	"k8s.io/apimachinery/pkg/runtime/schema"
@@ -76,17 +77,26 @@ func (c *Client) CreatePod(ctx context.Context, manifest PodManifest) (Pod, erro
 			Containers: []corev1.Container{{
 				Name: "executor", Image: manifest.Image, Args: manifest.Args, Env: environment,
 				SecurityContext: &corev1.SecurityContext{Privileged: boolPointer(true)},
-				VolumeMounts: []corev1.VolumeMount{{
+				VolumeMounts: []corev1.VolumeMount{
-					Name: "spire-agent-socket", MountPath: "/run/spire/agent-sockets", ReadOnly: true,
+					{Name: "spire-agent-socket", MountPath: "/run/spire/agent-sockets", ReadOnly: true},
+					{Name: "docker-data", MountPath: "/var/lib/docker"},
+				},
 			}},
-			}},
+			Volumes: []corev1.Volume{
-			Volumes: []corev1.Volume{{
+				{
 					Name: "spire-agent-socket",
 					VolumeSource: corev1.VolumeSource{CSI: &corev1.CSIVolumeSource{
 						Driver: "csi.spiffe.io", ReadOnly: boolPointer(true),
 					}},
+				},
+				{
+					Name: "docker-data",
+					VolumeSource: corev1.VolumeSource{EmptyDir: &corev1.EmptyDirVolumeSource{
+						SizeLimit: resourceQuantity("20Gi"),
 					}},
 				},
+			},
+		},
 	}
 	created, err := c.Kubernetes.CoreV1().Pods(manifest.Namespace).Create(ctx, document, metav1.CreateOptions{})
 	if err != nil {
@@ -169,6 +179,11 @@ func podFromKubernetes(pod corev1.Pod) Pod {
 
 func boolPointer(value bool) *bool { return &value }
 
+func resourceQuantity(value string) *resource.Quantity {
+	quantity := resource.MustParse(value)
+	return &quantity
+}
+
 func stringMap(values map[string]string) map[string]any {
 	result := make(map[string]any, len(values))
 	for key, value := range values {

internal/podbackend/client_test.go

@@ -37,6 +37,12 @@ func TestClientPodLifecycleUsesTypedClient(t *testing.T) {
 	if got := pod.Spec.Containers[0].Env; len(got) != 1 || got[0].Name != "CI_RUNNER_CAPABILITY" || got[0].Value != "capability" {
 		t.Fatalf("environment = %#v", got)
 	}
+	if got := pod.Spec.Containers[0].VolumeMounts; len(got) != 2 || got[1].Name != "docker-data" || got[1].MountPath != "/var/lib/docker" {
+		t.Fatalf("volume mounts = %#v", got)
+	}
+	if got := pod.Spec.Volumes; len(got) != 2 || got[1].EmptyDir == nil || got[1].EmptyDir.SizeLimit == nil || got[1].EmptyDir.SizeLimit.String() != "20Gi" {
+		t.Fatalf("volumes = %#v", got)
+	}
 	if err := client.LabelPod(context.Background(), "gitea-actions", created.Name, map[string]string{terminalLabel: "true"}); err != nil {
 		t.Fatal(err)
 	}

internal/runnerbootstrap/executor.go

@@ -22,6 +22,7 @@ type ExecutorConfig struct {
 	FacadeSPIFFEID  string
 	WorkloadAPIAddr string
 	RunnerBinary    string
+	RunnerConfig    string
 	ListenAddress   string
 	WorkDir         string
 	Stdout          *os.File
@@ -35,6 +36,9 @@ func RunExecutor(ctx context.Context, config ExecutorConfig) error {
 	if config.RunnerBinary == "" {
 		config.RunnerBinary = "gitea-runner"
 	}
+	if config.RunnerConfig == "" {
+		config.RunnerConfig = "/etc/gitea-runner/config.yaml"
+	}
 	if config.ListenAddress == "" {
 		config.ListenAddress = "127.0.0.1:0"
 	}
@@ -93,7 +97,7 @@ func RunExecutor(ctx context.Context, config ExecutorConfig) error {
 		return errors.Join(readyErr, shutdownErr, serverErr)
 	}
 
-	command := exec.CommandContext(ctx, config.RunnerBinary, "daemon", "--once")
+	command := exec.CommandContext(ctx, config.RunnerBinary, runnerArguments(config.RunnerConfig)...)
 	command.Dir = workDir
 	command.Stdout = config.Stdout
 	command.Stderr = config.Stderr
@@ -108,6 +112,10 @@ func RunExecutor(ctx context.Context, config ExecutorConfig) error {
 	return errors.Join(runnerErr, shutdownErr, serverErr)
 }
 
+func runnerArguments(configFile string) []string {
+	return []string{"daemon", "--config", configFile, "--once"}
+}
+
 func waitForFacade(ctx context.Context, endpoint string) error {
 	client := &http.Client{Timeout: 2 * time.Second}
 	ticker := time.NewTicker(250 * time.Millisecond)
@@ -143,7 +151,8 @@ func ExecutorConfigFromEnvironment() (ExecutorConfig, error) {
 	config := ExecutorConfig{
 		AssignmentID: os.Getenv(EnvAssignmentID), Capability: os.Getenv(EnvCapability),
 		Backend: backend, FacadeURL: os.Getenv(EnvFacadeURL), FacadeSPIFFEID: os.Getenv(EnvFacadeID),
-		RunnerBinary: os.Getenv("GITEA_RUNNER_BINARY"), ListenAddress: "127.0.0.1:0",
+		RunnerBinary: os.Getenv("GITEA_RUNNER_BINARY"), RunnerConfig: os.Getenv("GITEA_RUNNER_CONFIG_FILE"),
+		ListenAddress: "127.0.0.1:0",
 		Stdout:        os.Stdout, Stderr: os.Stderr,
 	}
 	if config.AssignmentID == "" || config.Capability == "" || config.FacadeURL == "" || config.FacadeSPIFFEID == "" {

internal/runnerbootstrap/executor_test.go

@@ -4,11 +4,19 @@ import (
 	"context"
 	"net/http"
 	"net/http/httptest"
+	"reflect"
 	"sync/atomic"
 	"testing"
 	"time"
 )
 
+func TestRunnerArgumentsLoadJobHooksConfig(t *testing.T) {
+	want := []string{"daemon", "--config", "/etc/gitea-runner/config.yaml", "--once"}
+	if got := runnerArguments("/etc/gitea-runner/config.yaml"); !reflect.DeepEqual(got, want) {
+		t.Fatalf("runner arguments = %q, want %q", got, want)
+	}
+}
+
 func TestWaitForFacadeRetriesTransientGatewayFailure(t *testing.T) {
 	var requests atomic.Int32
 	server := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, _ *http.Request) {

internal/taskassignment/assignment.go

@@ -102,7 +102,7 @@ func backendFromTask(task *runnerv1.Task) (Backend, error) {
 		return "", fmt.Errorf("task runs-on labels must include self-hosted: %v", labels)
 	}
 	hasPod := slices.Contains(labels, string(BackendPod))
-	hasVM := slices.Contains(labels, string(BackendVM))
+	hasVM := slices.Contains(labels, string(BackendVM)) || slices.Contains(labels, "vm-dev")
 	if hasPod == hasVM {
 		return "", fmt.Errorf("task runs-on labels must select exactly one of pod or vm: %v", labels)
 	}

internal/taskassignment/assignment_test.go

@@ -28,6 +28,7 @@ func TestNewSelectsBackendFromRunsOn(t *testing.T) {
 	}{
 		{"[self-hosted, pod]", BackendPod},
 		{"[self-hosted, vm]", BackendVM},
+		{"[self-hosted, vm-dev]", BackendVM},
 	} {
 		assignment, err := New(task(t, test.labels), "ddupan.top")
 		if err != nil {

internal/taskworker/worker.go

@@ -71,6 +71,28 @@ type Worker struct {
 	Backend   Backend
 	Tasks     TaskState
 	Bootstrap Bootstrap
+	OnEvent   func(Event)
+}
+
+// Event describes a backend lifecycle transition without exposing launch
+// environment values or other credentials.
+type Event struct {
+	Name         string
+	AssignmentID string
+	Executor     string
+	Phase        Phase
+}
+
+func (w Worker) event(name string, assignment taskassignment.Assignment, executor *Executor) {
+	if w.OnEvent == nil {
+		return
+	}
+	event := Event{Name: name, AssignmentID: assignment.ID}
+	if executor != nil {
+		event.Executor = executor.Name
+		event.Phase = executor.Phase
+	}
+	w.OnEvent(event)
 }
 
 // Accept completes the durable handoff from JetStream to the backend. Once it
@@ -88,6 +110,7 @@ func (w Worker) Accept(ctx context.Context, assignment taskassignment.Assignment
 		return false, err
 	}
 	if executor == nil {
+		w.event("executor_absent", assignment, nil)
 		launch, launchErr := w.launchSpec(assignment)
 		if launchErr != nil {
 			return false, launchErr
@@ -96,13 +119,18 @@ func (w Worker) Accept(ctx context.Context, assignment taskassignment.Assignment
 		if err != nil {
 			return false, err
 		}
+		w.event("executor_created", assignment, executor)
+	} else {
+		w.event("executor_found", assignment, executor)
 	}
 	if executor.IdentityTarget == "" {
+		w.event("identity_target_pending", assignment, executor)
 		return false, nil
 	}
 	if err := w.Backend.BindIdentity(ctx, executor, assignment.Identity); err != nil {
 		return false, err
 	}
+	w.event("identity_bound", assignment, executor)
 	return true, nil
 }
 

internal/taskworker/worker_test.go

@@ -79,7 +79,8 @@ func TestHandleRecoversExistingExecutorWithoutCreatingAnother(t *testing.T) {
 
 func TestAcceptAcknowledgesAfterBackendAndIdentityAreDurable(t *testing.T) {
 	backend := &fakeBackend{}
-	worker := Worker{Backend: backend, Bootstrap: fakeBootstrap{}}
+	var events []Event
+	worker := Worker{Backend: backend, Bootstrap: fakeBootstrap{}, OnEvent: func(event Event) { events = append(events, event) }}
 
 	accepted, err := worker.Accept(context.Background(), assignment())
 	if err != nil || !accepted {
@@ -88,6 +89,18 @@ func TestAcceptAcknowledgesAfterBackendAndIdentityAreDurable(t *testing.T) {
 	if backend.created != 1 || backend.bound != 1 || backend.deleted != 0 {
 		t.Fatalf("created=%d bound=%d deleted=%d", backend.created, backend.bound, backend.deleted)
 	}
+	want := []string{"executor_absent", "executor_created", "identity_bound"}
+	if len(events) != len(want) {
+		t.Fatalf("events = %#v", events)
+	}
+	for index := range want {
+		if events[index].Name != want[index] || events[index].AssignmentID != "gitea-task-42" {
+			t.Fatalf("event[%d] = %#v", index, events[index])
+		}
+	}
+	if events[1].Executor != "executor" || events[1].Phase != PhaseRunning {
+		t.Fatalf("created event = %#v", events[1])
+	}
 }
 
 func TestAcceptRetriesWhileBackendIdentityTargetIsUnavailable(t *testing.T) {

scripts/gitea-job-started

@@ -10,6 +10,7 @@ while [ "$(date +%s)" -lt "$deadline" ]; do
         -audience ci-job-ready \
         -socketPath "$socket" \
         >/dev/null 2>&1; then
+        /usr/local/libexec/setup-job-docker
         exit 0
     fi
     sleep 1

scripts/publish-image

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set +x
+
+: "${GITHUB_SHA:?GITHUB_SHA is required}"
+: "${IMAGE_NAME:?IMAGE_NAME is required}"
+: "${IMAGE_REPOSITORY:?IMAGE_REPOSITORY is required}"
+: "${IMAGE_DOCKERFILE:?IMAGE_DOCKERFILE is required}"
+: "${PUSH_REGISTRY:?PUSH_REGISTRY is required}"
+: "${PULL_REGISTRY:?PULL_REGISTRY is required}"
+: "${SPIRE_AGENT_SOCKET:?SPIRE_AGENT_SOCKET is required}"
+
+# Bootstrap the image that first introduces automatic Docker setup. Once that
+# runner is deployed, the job-started hook makes this an idempotent no-op.
+sudo scripts/setup-job-docker
+
+image_tag="sha-${GITHUB_SHA}"
+metadata="${IMAGE_NAME}-metadata.json"
+docker_config=$(mktemp -d)
+jwt_file=$(mktemp)
+cleanup() {
+  docker buildx rm ci-builder >/dev/null 2>&1 || true
+  rm -rf -- "$docker_config" "$jwt_file"
+}
+trap cleanup EXIT
+export DOCKER_CONFIG="$docker_config"
+
+/opt/spire/bin/spire-agent api fetch jwt \
+  -audience zot \
+  -socketPath "$SPIRE_AGENT_SOCKET" \
+  -output json >"$jwt_file"
+# shellcheck disable=SC2016
+jq -er '.[0].svids[0].svid' "$jwt_file" | \
+  docker login "$PUSH_REGISTRY" --username zot --password-stdin
+
+docker buildx create \
+  --name ci-builder \
+  --driver docker-container \
+  --use
+
+docker buildx build \
+  --builder ci-builder \
+  --platform linux/amd64 \
+  --file "$IMAGE_DOCKERFILE" \
+  --tag "${PUSH_REGISTRY}/${IMAGE_REPOSITORY}:${image_tag}" \
+  --tag "${PUSH_REGISTRY}/${IMAGE_REPOSITORY}:main" \
+  --provenance=mode=max \
+  --sbom=true \
+  --metadata-file "$metadata" \
+  --push \
+  .
+
+image_digest=$(
+  # shellcheck disable=SC2016
+  jq -er '."containerimage.digest"' "$metadata"
+)
+image_ref="${PULL_REGISTRY}/${IMAGE_REPOSITORY}@${image_digest}"
+
+printf '%s=%s\n' "$IMAGE_NAME" "$image_ref"
+if [[ -n "${GITHUB_STEP_SUMMARY:-}" ]]; then
+  {
+    printf '## Published image\n\n'
+    # shellcheck disable=SC2016
+    printf -- '- %s: `%s`\n' "$IMAGE_NAME" "$image_ref"
+    # shellcheck disable=SC2016
+    printf -- '- Source: `%s`\n' "$GITHUB_SHA"
+  } >>"$GITHUB_STEP_SUMMARY"
+fi

scripts/setup-job-docker

@@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+if docker info >/dev/null 2>&1; then
+  printf '%s\n' 'job Docker daemon is already ready'
+  exit 0
+fi
+
+storage_size=${DOCKER_DATA_SIZE:-20G}
+storage_driver=${DOCKER_STORAGE_DRIVER:-overlay2}
+wait_seconds=${DOCKER_START_WAIT_SECONDS:-60}
+
+sudo install -d /var/lib/docker
+if ! mountpoint --quiet /var/lib/docker; then
+  printf '%s\n' "preparing ${storage_size} loop-backed Docker storage"
+  if [[ ! -e /dev/loop-control ]]; then
+    sudo mknod /dev/loop-control c 10 237
+  fi
+  for minor in {0..7}; do
+    if [[ ! -e "/dev/loop${minor}" ]]; then
+      sudo mknod "/dev/loop${minor}" b 7 "$minor"
+    fi
+  done
+  sudo truncate -s "$storage_size" /tmp/docker-data.img
+  sudo mkfs.ext4 -F /tmp/docker-data.img
+  loop_device=$(sudo losetup --find --show /tmp/docker-data.img)
+  sudo mount "$loop_device" /var/lib/docker
+else
+  printf '%s\n' 'using mounted Docker storage at /var/lib/docker'
+fi
+
+if [[ -f /sys/fs/cgroup/cgroup.controllers ]]; then
+  sudo sh -c '
+    mkdir -p /sys/fs/cgroup/init
+    while read -r pid; do
+      printf "%s\n" "$pid" \
+        >/sys/fs/cgroup/init/cgroup.procs 2>/dev/null || true
+    done </sys/fs/cgroup/cgroup.procs
+    sed -e "s/ / +/g" -e "s/^/+/" \
+      /sys/fs/cgroup/cgroup.controllers \
+      >/sys/fs/cgroup/cgroup.subtree_control
+  '
+fi
+
+sudo sh -c 'nohup dockerd "$@" </dev/null >/tmp/dockerd.log 2>&1 &' sh \
+  --host=unix:///var/run/docker.sock \
+  --storage-driver="$storage_driver"
+printf '%s\n' 'waiting for job Docker daemon'
+for ((attempt = 0; attempt < wait_seconds; attempt++)); do
+  if docker info >/dev/null 2>&1; then
+    findmnt /var/lib/docker
+    docker info --format \
+      '{{json .ServerVersion}} {{json .Driver}} {{json .CgroupVersion}}'
+    exit 0
+  fi
+  sleep 1
+done
+
+cat /tmp/dockerd.log
+exit 1