cloud-hypervisor/docs/live_migration.md

Live Migration

This document gives examples of how to use the live migration support in Cloud Hypervisor:

1. Local Migration: Migrating a VM from one Cloud Hypervisor instance to another on the same machine; also called UNIX socket migration.
2. Remote Migration (TCP Migration): migrating a VM between two TCP/IP hosts.

⚠️ These examples place sockets in /tmp. This is done for simplicity and should not be done in production.

Local Migration (Suitable for Live Upgrade of VMM)

Launch the source VM (on the host machine):

$ target/release/cloud-hypervisor
    --kernel ~/workloads/vmlinux \
    --disk path=~/workloads/focal.raw \
    --cpus boot=1 --memory size=1G,shared=on \
    --cmdline "root=/dev/vda1 console=ttyS0"  \
    --serial tty --console off --api-socket=/tmp/api1

Launch the destination VM from the same directory (on the host machine):

$ target/release/cloud-hypervisor --api-socket=/tmp/api2

Get ready for receiving migration for the destination VM (on the host machine):

$ target/release/ch-remote --api-socket=/tmp/api2 receive-migration receiver_url=unix:/tmp/sock

Start to send migration for the source VM (on the host machine):

$ target/release/ch-remote --api-socket=/tmp/api1 send-migration destination_url=unix:/tmp/sock,local=on

When the above commands completed, the source VM should be successfully migrated to the destination VM. Now the destination VM is running while the source VM is terminated gracefully.

Remote Migration (TCP Migration)

Hint: For developing purposes, same-host TCP migrations are also supported.

In this example, we will migrate a VM from one machine (src) to another (dst) across the network. To keep it simple, we will use a minimal VM setup without storage.

Preparation

Make sure that src and dst can reach each other via the network. You should be able to ping each machine. Also each machine should have an open TCP port.

You will need a kernel and initramfs for a minimal Linux system. For this example, we will use the Debian netboot image.

Place the kernel and initramfs into the same directory on both machines. This is important for the migration to succeed. We will use /var/images:

src $ export DEBIAN=https://ftp.debian.org/debian/dists/stable/main/installer-amd64/current/images/netboot/debian-installer/amd64
src $ mkdir -p /var/images
src $ curl $DEBIAN/linux > /var/images/linux
src $ curl $DEBIAN/initrd.gz > /var/images/initrd

Repeat the above steps on the destination host.

Unix Socket Migration

If Unix socket is selected for migration, we can tunnel traffic through "socat".

Starting the Receiver VM

On the receiver side, we prepare an empty VM:

dst $ cloud-hypervisor --api-socket /tmp/api

In a different terminal, configure the VM as a migration target:

dst $ ch-remote --api-socket=/tmp/api receive-migration receiver_url=unix:/tmp/sock

In yet another terminal, forward TCP connections to the Unix domain socket:

dst $ socat TCP-LISTEN:{port},reuseaddr UNIX-CLIENT:/tmp/sock

Starting the Sender VM

Let's start the VM on the source machine:

src $ cloud-hypervisor \
        --serial tty --console off \
        --cpus boot=2 --memory size=4G \
        --kernel /var/images/linux \
        --initramfs /var/images/initrd \
        --cmdline "console=ttyS0" \
        --api-socket /tmp/api

After a few seconds the VM should be up and you can interact with it.

Performing the Migration

First, we start socat:

src $ socat UNIX-LISTEN:/tmp/sock,reuseaddr TCP:{dst}:{port}

Replace {dst}:{port} with the actual IP address and port of your destination host.

Then we kick-off the migration itself:

src $ ch-remote --api-socket=/tmp/api send-migration destination_url=unix:/tmp/sock

When the above commands completed, the VM should be successfully migrated to the destination machine without interrupting the workload.

Network Announcements After Resume

After a VM resumes from migration, snapshot restore, or any other path that restores a previously paused VM, Cloud Hypervisor asks supported network devices to announce the VM from its new host. For virtio-net, the current implementation sets VIRTIO_NET_S_ANNOUNCE, raises a config interrupt, retries that request a few times in the background, and also sends host-side RARP announcements on the TAP interfaces. A guest re-announcement therefore only happens when the guest negotiated VIRTIO_NET_F_GUEST_ANNOUNCE. For vhost-user-net, the current implementation only uses the guest announcement path.

TCP Socket Migration

If TCP socket is selected for migration, we need to consider migrating in a trusted network.

Starting the Receiver VM

On the receiver side, we prepare an empty VM:

dst $ cloud-hypervisor --api-socket /tmp/api

In a different terminal, prepare to receive the migration:

dst $ ch-remote --api-socket=/tmp/api receive-migration receiver_url=tcp:0.0.0.0:{port}

Starting the Sender VM

Let's start the VM on the source machine:

src $ cloud-hypervisor \
        --serial tty --console off \
        --cpus boot=2 --memory size=4G \
        --kernel /var/images/linux \
        --initramfs /var/images/initrd \
        --cmdline "console=ttyS0" \
        --api-socket /tmp/api

After a few seconds the VM should be up and you can interact with it.

Performing the Migration

Initiate the Migration over TCP:

src $ ch-remote --api-socket=/tmp/api send-migration destination_url=tcp:{dst}:{port}

With migration parameters:

src $ ch-remote --api-socket=/tmp/api send-migration destination_url=tcp:{dst}:{port},downtime_ms=200,timeout_s=3600,timeout_strategy=cancel

Replace {dst}:{port} with the actual IP address and port of your destination host.

After completing the above commands, the source VM will be migrated to the destination host and continue running there. The source VM instance will terminate normally. All ongoing processes and connections within the VM should remain intact after the migration. See Network Announcements After Resume for the announcement behavior after a VM resumes.

Encryption

TCP migration can be protected with TLS by passing tls_dir=<path> to both receive-migration and send-migration.

The destination host needs a directory containing:

• ca-cert.pem: the CA certificate used to verify source certificates
• server-cert.pem: the certificate presented by the destination
• server-key.pem: the private key for server-cert.pem

The source host needs a directory containing:

• ca-cert.pem: the CA certificate used to verify the destination certificate
• client-cert.pem: the certificate presented by the source
• client-key.pem: the private key for client-cert.pem

Protect the private key files with file mode 0600 to reduce the risk of accidental disclosure:

$ chmod 600 server-key.pem client-key.pem

Current TCP migration uses mutual TLS (mTLS) authentication. The source verifies the destination certificate against ca-cert.pem and presents client-cert.pem and client-key.pem. The destination presents server-cert.pem and server-key.pem, and only accepts client certificates that chain to ca-cert.pem.

Example receiver command:

dst $ ch-remote --api-socket=/tmp/api receive-migration receiver_url=tcp:0.0.0.0:{port},tls_dir=/path/to/dst-tls

Example sender command:

src $ ch-remote --api-socket=/tmp/api send-migration destination_url=tcp:{dst}:{port},tls_dir=/path/to/src-tls

TLS encryption is only supported with tcp:<host>:<port> migration URLs, not with local UNIX-socket migration.

The commands below describe how to create the encryption material necessary to perform a same-host TCP migration with TLS.

Setup the Certificate Authority

First, set up the CA with a private key and a self-signed certificate.

$ certtool --generate-privkey \
           --outfile ca-key.pem

To generate the certificate, provide at least your organization name in a template file, for example ca.info:

cn = Name of your organization
ca
cert_signing_key

Then you can create the certificate:

$ certtool --generate-self-signed \
           --load-privkey ca-key.pem \
           --template ca.info \
           --outfile ca-cert.pem

Issuing host certificates

Generate a private key and a certificate for each side. Cloud Hypervisor enforces mTLS, thus you also need a key for the migration sender.

The certificates also require template data. You can use separate templates for sender and receiver, but this example uses a single template, hosts.info:

organization = Name of your organization
cn = localhost
tls_www_server
tls_www_client
signing_key
dns_name = localhost
ip_address = 127.0.0.1

Then you can create the necessary files for the client (migration sender):

$ certtool --generate-privkey \
           --outfile client-key.pem

$ certtool --generate-certificate \
           --load-ca-certificate ca-cert.pem \
           --load-ca-privkey ca-key.pem \
           --load-privkey client-key.pem \
           --template hosts.info \
           --outfile client-cert.pem

And the necessary files for the server (migration receiver):

$ certtool --generate-privkey \
           --outfile server-key.pem

$ certtool --generate-certificate \
           --load-ca-certificate ca-cert.pem \
           --load-ca-privkey ca-key.pem \
           --load-privkey server-key.pem \
           --template hosts.info \
           --outfile server-cert.pem

For a same-host test, place these files in a single directory and use that directory as tls_dir for both commands.

Performing the Migration

On the receiver side, we prepare an empty VM:

dst $ cloud-hypervisor --api-socket /tmp/api-rcv

In a different terminal, prepare to receive the migration:

dst $ ch-remote --api-socket=/tmp/api-rcv receive-migration receiver_url=tcp:0.0.0.0:9000,tls_dir=/path/to/certificates

On the sender side, we start a VM:

src $ cloud-hypervisor \
        --serial tty --console off \
        --cpus boot=2 --memory size=4G \
        --kernel /var/images/linux \
        --initramfs /var/images/initrd \
        --cmdline "console=ttyS0" \
        --api-socket /tmp/api-src

After a few seconds the VM should be up and you can interact with it. Then trigger the migration:

src $ ch-remote --api-socket=/tmp/api-src send-migration destination_url=tcp:localhost:9000,tls_dir=/path/to/certificates

Migration Parameters

Cloud Hypervisor supports additional parameters to control the migration process. Via the API or ch-remote, you may specify:

• downtime_ms <milliseconds>:
  The maximum downtime the migration aims for, in milliseconds. Defaults to 300ms.
• timeout_s <seconds>:
  The timeout for the migration (maximum total duration), in seconds. Defaults to 3600s (one hour).
• timeout_strategy <strategy> ([cancel, ignore]):
  The strategy to apply when the migration timeout is reached. Cancel will abort the migration and keep the VM running on the source. Ignore will proceed with the migration regardless of the downtime requirement. Defaults to cancel.
• connections <amount>:
  The number of parallel TCP connections to use for migration. Must be between 1 and 128. Defaults to 1. Multiple connections are not supported with local UNIX-socket migration.
• memory_mode <precopy|postcopy>:
  Memory transfer mode. postcopy resumes the destination first and faults guest pages in on demand over a dedicated connection. Defaults to precopy.
• zone_updates <list of zone updates>:
  A list of updates to apply to memory zones on the receiver side. For example, this can be used to remap memory zones to another NUMA node. Each zone update has the form zone_id@host_numa_node.

Migrating VMs with VFIO Devices

VMs with migratable VFIO devices, for example NIC VFs bound to mlx5_vfio_pci, can be live migrated. The opaque device state is captured during the downtime window and the destination receives new device file descriptors through the vfio_fds and iommufd_fd options of receive-migration.

See Live Migration of VFIO Devices for the requirements, the dirty tracking behavior, and an example.

Version Compatibility

Cloud Hypervisor live migration compatibility has two dimensions: the Cloud Hypervisor version and the migration protocol version.

Cloud Hypervisor uses a versioned migration protocol for the messages exchanged between source and destination. Each Cloud Hypervisor release sends its current migration protocol version and accepts that version plus the immediately previous protocol version.

This means migration is supported from an older protocol version to the same or next protocol version. If the source and destination are more than one migration protocol version apart, the VM must be migrated through an intermediate Cloud Hypervisor version first.

Technical Details

The source sends its migration protocol version in the initial Start request. The destination accepts the migration only if that protocol version is supported. A zeroed Start command header is handled as protocol v0, so older deployments that do not explicitly send a protocol version remain compatible.

The migration protocol version covers the protocol spoken after a migration connection has been established. Examples include adding a mandatory migration command, changing the order of migration protocol messages, or changing the framing or encoding of protocol command payloads.

The migration protocol version does not cover migration transport setup. For example, choosing TCP vs. UNIX sockets or opening the initial connection needs separate compatibility handling.

Migration protocol versioning is separate from snapshot state compatibility. Device and VM state changes still need to be handled by the respective snapshot serialization/deserialization code. That compatibility is Cloud Hypervisor version dependent, but it is not the responsibility of migration protocol versioning.