github.com/GoogleCloudPlatform/guest-agent @20260529.00

Guest Agent for Google Compute Engine

This repository contains the source code and packaging artifacts for the Google guest agent and metadata script runner binaries. These components are installed on Windows and Linux GCE VMs in order to enable GCE platform features.

Table of Contents

• Overview
• Features
  • Account Management
  • Clock Skew
  • OS Login
  • Network
  • Windows Failover Cluster Support
  • Instance Setup
  • Telemetry
  • MTLS MDS
• Metadata Scripts
• Configuration
• Packaging

Overview

The repository contains these components:

• google-guest-agent daemon which handles all of the areas outlined below in "features"
• google-metadata-script-runner binary to run user-provided scripts at VM startup and shutdown.

Note: Starting in December 2024 guest agent installs a secondary systemd unit named google-guest-agent-manager, this unit manages on demand plugins lifecycle, such plugin architecture will support expansion of the guest agent in the future. In the upcoming releases both google-guest-agent and google-guest-agent-manager will be consolidated in a single unit. The source code for plugin manager can be found here.

Features

The guest agent functionality can be separated into various areas of responsibility. Historically, on Linux these were managed by separate independent processes, but today they are all managed by the guest agent.

The Daemons section of the instance configs file on Linux refers to these areas of responsibility. This allows a user to easily modify or disable functionality. Behaviors for each area of responsibility are detailed below.

Account management

On Windows, the agent handles creating user accounts and setting/resetting passwords.

Guest Agent automatically creates local user accounts for any SSH user defined in the Metadata SSH keys at the instance or project level (unless blocked) on Windows instances to support connecting to Windows VMs using SSH.

Active Directory Domain Controller does not use the local user account database except when it is booted into the recovery console or demoted, so any account created on the system would become an administrator of the Active Directory Domain. You can prevent unintended AD user provisioning by disabling the account manager on the AD controller VM. Refer deploy domain controllers for more information on setting up AD on GCE.

On Linux: If OS Login is not used, the guest agent will be responsible for provisioning and deprovisioning user accounts. The agent creates local user accounts and maintains the authorized SSH keys file for each. User account creation is based on adding and remove SSH Keys stored in metadata.

The guest agent has the following behaviors:

• Administrator permissions are managed with a google-sudoers Linux group. Members of this group are granted sudo permissions on the VM.
• All users provisioned by the account daemon are added to the google-sudoers group.
• The daemon stores a file in the guest to record which user accounts are managed by Google.
• User accounts not managed by the agent are not touched by the accounts daemon.
• The authorized keys file for a Google managed user is deleted when all SSH keys for the user are removed from metadata.
• Users accounts managed by the agent will be added to the groups config line in the Accounts section. If these groups do not exist, the agent will not create them.

OS Login

(Linux only)

If the user has configured OS Login via metadata, the guest agent will be responsible for configuring the OS to use OS Login, otherwise called 'enabling' OS Login. This consists of:

• Adding a Google config block to the SSHD configuration file and restarting SSHD.
• Adding OS Login entries to the nsswitch.conf file.
• Adding OS Login entries to the PAM configuration file for SSHD.

If the user disables OS login via metadata, the configuration changes will be removed.

Note that options under the Accounts section of the configuration do not apply to oslogin users.

Clock Skew

(Linux only)

The guest agent is responsible for syncing the software clock with the hypervisor clock after a stop/start event or after a migration. Preventing clock skew may result in system time has changed messages in VM logs.

Network

The guest agent uses network interface metadata to manage the network interfaces by performing the following tasks:

• Enable all associated network interfaces on boot.
  • Detect the current active network manager service that is managing the primary NIC.
  • Rollback and delete any guest agent-managed files/configurations left behind by all other supported network manager services.
  • Write and apply new configurations for the secondary NICs using the network manager service detected in the first step.
  • Create a route to the metadata server for the primary NIC only.
• Setup or remove IP routes in the guest for IP forwarding and IP aliases
  • Only IPv4 IP addresses are currently supported.
  • Routes are set on the primary ethernet interface.
  • Google routes are configured, by default, with the routing protocol ID 66. This ID is a namespace for daemon configured IP addresses. It can be changed with the config file, see below.

On Linux, supported network managers are as follows. These are listed by descending priority and include the location at which the configuration files are written.

• netplan
  • Config location: /run/netplan/
    • ex: /run/netplan/20-google-guest-agent-eth0.yaml
  • Dropin location: /etc/systemd/network/
    • ex: /etc/systemd/network/10-netplan-eth0.network.d/
• wicked
  • Config location: /etc/sysconfig/network/
    • ex: /etc/sysconfig/network/ifcfg-eth0
  • Notes:
    • Existing ifcfg files are not overwritten and are skipped instead.
• NetworkManager
  • Config location: /etc/NetworkManager/system-connections/
    • ex: /etc/NetworkManager/system-connections/google-guest-agent-eth0.nmconnection
• systemd-networkd
  • Config location: /usr/lib/systemd/network/
    • ex: /usr/lib/systemd/network/20-eth0-google-guest-agent.network
• dhclient
  • Config location: /run/
    • ex (pid): /run/dhclient.google-guest-agent.eth0.ipv4.pid
    • ex (lease): /run/dhclient.google-guest-agent.eth0.ipv4.lease
  • Notes:
    • The primary NIC setup, if enabled, is skipped if a dhclient process for the primary NIC is already running.

If none of the first 4 network manager services are detected on the system, then the agent will default to using dhclient for managing network interfaces.

Note: Ubuntu 18.04, while having netplan installed, ships a outdated and unsupported version of networkctl. This older version lacks essential commands like networkctl reload, causing compatibility issues. Guest agent is designed to fallback to dhclient on Ubuntu 18.04, even when netplan is present, to ensure proper network configuration.

The following configuration flags can control the behavior:

• manage_primary_nic: When enabled, the agent will start managing the primary NIC in addition to the secondary NICs.

For more information about the instance configuration, see the Configuration section.

The guest agent will also setup VLANs if VLAN is enabled. The setup and configuration for this work similarly to the normal NIC configuration.

If the VLANs' parent interface is the primary NIC, it will apply the VLAN configurations regardless of whether manage_primary_nic is set.

Windows Failover Cluster Support

(Windows only)

The agent can monitor the active node in the Windows Failover Cluster and coordinate with GCP Internal Load Balancer to forward all cluster traffic to the expected node.

The following fields on instance metadata or instance_configs.cfg can control the behavior:

• enable-wsfc: If set to true, all IP forwarding info will be ignored and agent will start responding to the health check port. Default false.
• wsfc-agent-port: The port which the agent will respond to health checks. Default 59998.
• wsfc-addrs: A comma separated list of IP address. This is an advanced setting to enable user have both normal forwarding IPs and cluster IPs on the same instance. If set, agent will only skip-auto configuring IPs in the list. Default empty.

Instance Setup

(Linux only)

The guest agent will perform some actions once each time on startup:

• Optimize for local SSD.
• Enable multi-queue on all the virtionet devices.

The guest agent will perform some actions one time only, on the first VM boot:

• Generate SSH host keys.
• Create the boto config for using Google Cloud Storage.

Telemetry

The guest agent will record some basic system telemetry information at start and then once every 24 hours.

• Guest agent version and architecture
• Operating system name and version
• Operating system kernel release and version

Telemetry can be disabled by setting the metadata key disable-guest-telemetry to true.

MTLS MDS

GCE Shielded VMs now support HTTPS endpoint https://metadata.google.internal/computeMetadata/v1 for Metadata Server. To enable communication with secure HTTPS endpoint, Guest Agent retrieves and stores credentials on the VM's disk in a standard location, making them accessible to any client application running on the VM. Both the root certificate and client credentials are updated each time the guest-agent process starts. For enhanced security, client credentials are automatically refreshed every 48 hours. The agent generates and saves new credentials, while the old ones remain valid. This overlap period ensures that clients have sufficient time to transition to the new credentials before the old ones expire, and it allows the agent to retry in case of failure and obtain valid credentials before the existing ones become invalid. Client credentials are basically EC private key and the client certificate concatenated. These credentials are unique to an instance and would not work elsewhere.

Refer this for more information on HTTPS metadata server endpoint and credential details including their lifespan.

Credentials can be stored at these supported locations -

• Linux:

  • Client credentials: /run/google-mds-mtls/client.key
  • Root certificate: /run/google-mds-mtls/root.crt and local trust store based on target OS. Refer this for local trust store location for each target OS.

• Windows:

  • Client credentials: C:\ProgramData\Google\ComputeEngine\mds-mtls-client.key and Cert:\LocalMachine\My
  • Root certificate: C:\ProgramData\Google\ComputeEngine\mds-mtls-root.crt and Cert:\LocalMachine\Root
  • PFX: C:\ProgramData\Google\Compute Engine\mds-mtls-client.key.pfx

  Credentials can be stored on disk as well as in Certificate Store on Windows

Note that this is disabled by default, if HTTPS endpoint is supported on a VM, the feature can be enabled by setting disable-https-mds-setup = false under [MDS] section in instance_configs.cfg file.

If enabled, agent will write certificates only on disk by default and users can opt-in to have certificates in OS Native stores. This means in case of Linux based VMs MDS Root certificate will be added to trust store like /etc/pki/ca-trust/extracted/pem/tls-ca-bundle.pem on RHEL based systems and /etc/ssl/certs/ca-certificates.crt on Debian based. Local root trust store is updated by running update-ca-certificates or update-ca-trust tool based on the OS. On Windows, Client credentials will be added in Cert:\LocalMachine\My and Root certificate in Cert:\LocalMachine\Root. This can be enabled by setting enable-https-mds-native-cert-store = true under same [MDS] section.

As documented by Microsoft here there could be issues with LDAPS process when multiple certificates are added in personal store. Avoid enabling OS Native stores on Domain Controllers. Credentials can still be used from disk if required.

Metadata Scripts

Metadata scripts implement support for running user provided startup scripts and shutdown scripts. The guest support for metadata scripts consider the following details:

• Metadata scripts a