Add interlocks to ensure operations are not interrupted

.drone.yml

@@ -160,96 +160,96 @@ volumes:
   host:
     path: /var/run/docker.sock
 
----
-kind: pipeline
-name: s390x
-
-platform:
-  os: linux
-  arch: amd64
-
-# Hack needed for s390x: https://gist.github.com/colstrom/c2f359f72658aaabb44150ac20b16d7c#gistcomment-3858388
-node:
-  arch: s390x
-
-steps:
-- name: build
-  image: rancher/dapper:v0.6.0
-  commands:
-  - dapper ci
-  volumes:
-  - name: docker
-    path: /var/run/docker.sock
-
-- name: github_binary_release
-  image: rancher/drone-images:github-release-s390x
-  settings:
-    api_key:
-      from_secret: github_token
-    prerelease: true
-    checksum:
-    - sha256
-    checksum_file: CHECKSUMsum-s390x.txt
-    checksum_flatten: true
-    files:
-    - "dist/artifacts/*"
-  when:
-    instance:
-    - drone-publish.rancher.io
-    ref:
-    - refs/head/master
-    - refs/tags/*
-    event:
-    - tag
-
-- name: docker-publish
-  image: rancher/drone-images:docker-s390x
-  volumes:
-    - name: docker
-      path: /var/run/docker.sock
-  settings:
-    dockerfile: package/Dockerfile
-    password:
-      from_secret: docker_password
-    repo: "rancher/system-agent"
-    tag: "${DRONE_TAG}-s390x"
-    username:
-      from_secret: docker_username
-  when:
-    instance:
-    - drone-publish.rancher.io
-    ref:
-    - refs/head/master
-    - refs/tags/*
-    event:
-    - tag
-
-- name: docker-publish-suc
-  image: rancher/drone-images:docker-s390x
-  volumes:
-    - name: docker
-      path: /var/run/docker.sock
-  settings:
-    dockerfile: package/Dockerfile.suc
-    password:
-      from_secret: docker_password
-    repo: "rancher/system-agent"
-    tag: "${DRONE_TAG}-suc-s390x"
-    username:
-      from_secret: docker_username
-  when:
-    instance:
-    - drone-publish.rancher.io
-    ref:
-    - refs/head/master
-    - refs/tags/*
-    event:
-    - tag
-
-volumes:
-- name: docker
-  host:
-    path: /var/run/docker.sock
+#---
+#kind: pipeline
+#name: s390x
+#
+#platform:
+#  os: linux
+#  arch: amd64
+#
+## Hack needed for s390x: https://gist.github.com/colstrom/c2f359f72658aaabb44150ac20b16d7c#gistcomment-3858388
+#node:
+#  arch: s390x
+#
+#steps:
+#- name: build
+#  image: rancher/dapper:v0.6.0
+#  commands:
+#  - dapper ci
+#  volumes:
+#  - name: docker
+#    path: /var/run/docker.sock
+#
+#- name: github_binary_release
+#  image: rancher/drone-images:github-release-s390x
+#  settings:
+#    api_key:
+#      from_secret: github_token
+#    prerelease: true
+#    checksum:
+#    - sha256
+#    checksum_file: CHECKSUMsum-s390x.txt
+#    checksum_flatten: true
+#    files:
+#    - "dist/artifacts/*"
+#  when:
+#    instance:
+#    - drone-publish.rancher.io
+#    ref:
+#    - refs/head/master
+#    - refs/tags/*
+#    event:
+#    - tag
+#
+#- name: docker-publish
+#  image: rancher/drone-images:docker-s390x
+#  volumes:
+#    - name: docker
+#      path: /var/run/docker.sock
+#  settings:
+#    dockerfile: package/Dockerfile
+#    password:
+#      from_secret: docker_password
+#    repo: "rancher/system-agent"
+#    tag: "${DRONE_TAG}-s390x"
+#    username:
+#      from_secret: docker_username
+#  when:
+#    instance:
+#    - drone-publish.rancher.io
+#    ref:
+#    - refs/head/master
+#    - refs/tags/*
+#    event:
+#    - tag
+#
+#- name: docker-publish-suc
+#  image: rancher/drone-images:docker-s390x
+#  volumes:
+#    - name: docker
+#      path: /var/run/docker.sock
+#  settings:
+#    dockerfile: package/Dockerfile.suc
+#    password:
+#      from_secret: docker_password
+#    repo: "rancher/system-agent"
+#    tag: "${DRONE_TAG}-suc-s390x"
+#    username:
+#      from_secret: docker_username
+#  when:
+#    instance:
+#    - drone-publish.rancher.io
+#    ref:
+#    - refs/head/master
+#    - refs/tags/*
+#    event:
+#    - tag
+#
+#volumes:
+#- name: docker
+#  host:
+#    path: /var/run/docker.sock
 
 ---
 kind: pipeline
@@ -270,7 +270,7 @@ steps:
     platforms:
       - linux/amd64
       - linux/arm64
-      - linux/s390x
+#      - linux/s390x
     target: "rancher/system-agent:${DRONE_TAG}"
     template: "rancher/system-agent:${DRONE_TAG}-ARCH"
   when:
@@ -285,7 +285,7 @@ steps:
 depends_on:
 - amd64
 - arm64
-- s390x
+#- s390x
 
 ---
 kind: pipeline
@@ -306,7 +306,7 @@ steps:
     platforms:
       - linux/amd64
       - linux/arm64
-      - linux/s390x
+#      - linux/s390x
     target: "rancher/system-agent:${DRONE_TAG}-suc"
     template: "rancher/system-agent:${DRONE_TAG}-suc-ARCH"
   when:
@@ -321,4 +321,4 @@ steps:
 depends_on:
 - amd64
 - arm64
-- s390x
+#- s390x

install.sh

@@ -47,6 +47,7 @@ fi
 FALLBACK=v0.2.9
 CACERTS_PATH=cacerts
 RETRYCOUNT=4500
+APPLYINATOR_ACTIVE_WAIT_COUNT=60 # If the system-agent is unhealthy but had created an interlock file to indicate it was actively applying a plan, after 5 minutes, ignore the interlock.
 
 # info logs the given argument at info log level.
 info() {
@@ -426,11 +427,13 @@ setup_env() {
 }
 
 ensure_directories() {
-    mkdir -p ${CATTLE_AGENT_VAR_DIR}
+    mkdir -p ${CATTLE_AGENT_VAR_DIR}/interlock
     mkdir -p ${CATTLE_AGENT_CONFIG_DIR}
     chmod 700 ${CATTLE_AGENT_VAR_DIR}
+    chmod 700 ${CATTLE_AGENT_VAR_DIR}/interlock
     chmod 700 ${CATTLE_AGENT_CONFIG_DIR}
     chown root:root ${CATTLE_AGENT_VAR_DIR}
+    chown root:root ${CATTLE_AGENT_VAR_DIR}/interlock
     chown root:root ${CATTLE_AGENT_CONFIG_DIR}
 }
 
@@ -783,6 +786,7 @@ appliedPlanDirectory: ${CATTLE_AGENT_VAR_DIR}/applied
 remoteEnabled: ${CATTLE_REMOTE_ENABLED}
 localEnabled: ${CATTLE_LOCAL_ENABLED}
 localPlanDirectory: ${CATTLE_AGENT_VAR_DIR}/plans
+interlockDirectory: ${CATTLE_AGENT_VAR_DIR}/interlock
 preserveWorkDirectory: ${CATTLE_PRESERVE_WORKDIR}
 EOF
     umask "${UMASK}"
@@ -796,6 +800,9 @@ generate_cattle_identifier() {
         info "Generating Cattle ID"
         if [ -f "${CATTLE_AGENT_CONFIG_DIR}/cattle-id" ]; then
             CATTLE_ID=$(cat ${CATTLE_AGENT_CONFIG_DIR}/cattle-id);
+            if [ -z "${CATTLE_ID}" ]; then
+              fatal "Cattle ID was empty, aborting installation"
+            fi
             info "Cattle ID was already detected as ${CATTLE_ID}. Not generating a new one."
             return
         fi
@@ -805,6 +812,9 @@ generate_cattle_identifier() {
         umask 0177
         echo "${CATTLE_ID}" > ${CATTLE_AGENT_CONFIG_DIR}/cattle-id
         umask "${UMASK}"
+        if [ ! -s ${CATTLE_AGENT_CONFIG_DIR}/cattle-id ]; then
+          fatal "Cattle ID could not be persisted. Aborting installation"
+        fi
         return
     fi
     info "Not generating Cattle ID"
@@ -832,6 +842,19 @@ create_env_file() {
     done
 }
 
+ensure_applyinator_not_active() {
+    i=1
+    while [ "${i}" -ne "${APPLYINATOR_ACTIVE_WAIT_COUNT}" ]; do
+      if [ -f "${CATTLE_AGENT_VAR_DIR}/interlock/applyinator-active" ]; then
+        i=$((i + 1))
+        info "Active plan reconciliation detected. Sleeping for 5 seconds and retrying check"
+        sleep 5
+        continue
+      fi
+      break
+    done
+}
+
 do_install() {
     if [ $(id -u) != 0 ]; then
       fatal "This script must be run as root."
@@ -843,6 +866,9 @@ do_install() {
     ensure_directories
     verify_downloader curl || fatal "can not find curl for downloading files"
 
+    touch ${CATTLE_AGENT_VAR_DIR}/interlock/restart-pending
+    ensure_applyinator_not_active
+
     if [ -n "${CATTLE_CA_CHECKSUM}" ]; then
         validate_ca_required
     fi
@@ -865,6 +891,7 @@ do_install() {
     systemctl enable rancher-system-agent
     info "Starting/restarting rancher-system-agent.service"
     systemctl restart rancher-system-agent
+    rm -f ${CATTLE_AGENT_VAR_DIR}/interlock/restart-pending
 }
 
 do_install "$@"

main.go

@@ -81,7 +81,7 @@ func run(c *cli.Context) error {
 	logrus.Infof("Using directory %s for work", cf.WorkDir)
 
 	imageUtil := image.NewUtility(cf.ImagesDir, cf.ImageCredentialProviderConfig, cf.ImageCredentialProviderBinDir, cf.AgentRegistriesFile)
-	applyinator := applyinator.NewApplyinator(cf.WorkDir, cf.PreserveWorkDir, cf.AppliedPlanDir, imageUtil)
+	applyinator := applyinator.NewApplyinator(cf.WorkDir, cf.PreserveWorkDir, cf.AppliedPlanDir, cf.InterlockDir, imageUtil)
 
 	if cf.RemoteEnabled {
 		logrus.Infof("Starting remote watch of plans")