Develop for XenServer

Using the API

This chapter describes how to use the XenServer Management API to write real programs to manage XenServer hosts and VMs. The chapter begins with a walk-through of a typical client application and demonstrates how the API can be used to perform common tasks. Example code fragments are given in Python syntax but equivalent code in the other programming languages would look very similar. The chapter finishes with walk-throughs of two complete examples.

Anatomy of a typical application

This section describes the structure of a typical application using the XenServer Management API. Most client applications begin by connecting to a XenServer host and authenticating using a username and password. Assuming the authentication succeeds, the server will create a “session” object and return a reference to the client. This reference will be passed as an argument to all future API calls. Once authenticated, the client may search for references to other useful objects (XenServer hosts, VMs, SRs, and so on) and invoke operations on them. Operations may be invoked either synchronously or asynchronously; special task objects represent the state and progress of asynchronous operations. These application elements are all described in detail in the following sections.

Choosing a low-level transport

API calls can be issued over two transports:

  • SSL-encrypted TCP on port 443 (https) over an IP network

  • plaintext over a local Unix domain socket: /var/xapi/xapi

Switching from the XML-RPC to the JSON-RPC backend can be done by adding the suffix /jsonrpc to the host URL path.

The SSL-encrypted TCP transport is used for all off-host traffic, while the Unix domain socket can be used from services running directly on the XenServer host itself. In the SSL-encrypted TCP transport, all API calls must be directed at the pool coordinator. If directed at a pool supporter, the calls will fail with the error HOST_IS_SLAVE, which includes the IP address of the coordinator as an error parameter.

Because the coordinator host of a pool can change, especially if HA is enabled on a pool, clients must implement the following steps to detect a coordinator host change and connect to the new coordinator as required:

Handling pool coordinator changes

  1. Subscribe to updates in the list of hosts servers, and maintain a current list of hosts in the pool

  2. If the connection to the pool coordinator fails to respond, attempt to connect to all hosts in the list until one responds

  3. The first host to respond will return the HOST_IS_SLAVE error message, which contains the identity of the new pool coordinator (unless of course the host is the new coordinator)

  4. Connect to the new coordinator


As a special-case, all messages sent through the Unix domain socket are transparently forwarded to the correct node.

Authentication and session handling

The vast majority of API calls take a session reference as their first parameter; failure to supply a valid reference will result in a SESSION_INVALID error being returned. Acquire a session reference by supplying a user name and password to the login_with_password function.


As a special-case, if this call is run over the local Unix domain socket then the user name and password are ignored and the call always succeeds.

Every session has an associated “last active” timestamp which is updated on every API call. The server software currently has a built-in limit of 500 active sessions and will remove those with the oldest “last active” field if this limit is exceeded for a given username or originator. In addition all sessions whose “last active” field is older than 24 hours are also removed. Therefore it is important to:

  • Specify an appropriate originator when logging in; and

  • Remember to log out of active sessions to avoid leaking them; and

  • Be prepared to log in again to the server if a SESSION_INVALID error is caught.


A session reference obtained by a login request to the XML-RPC backend can be used in subsequent requests to the JSON-RPC backend, and vice-versa.

In the following Python fragment a connection is established over the Unix domain socket and a session is created:

import XenAPI

session = XenAPI.xapi_local()
    session.xenapi.login_with_password("root", "", "2.20", "My Widget v0.1")

Finding references to useful objects

Once an application has authenticated the next step is to acquire references to objects in order to query their state or invoke operations on them. All objects have a set of “implicit” messages which include the following:

  • get_by_name_label : return a list of all objects of a particular class with a particular label;

  • get_by_uuid : return a single object named by its UUID;

  • get_all : return a set of references to all objects of a particular class; and

  • get_all_records : return a map of reference to records for each object of a particular class.

For example, to list all hosts:

hosts =

To find all VMs with the name “my first VM”:

vms = session.xenapi.VM.get_by_name_label('my first VM')


Object name_label fields are not guaranteed to be unique and so the get_by_name_label API call returns a set of references rather than a single reference.

In addition to the methods of finding objects described above, most objects also contain references to other objects within fields. For example it is possible to find the set of VMs running on a particular host by calling:

vms =

Invoking synchronous operations on objects

Once object references have been acquired, operations may be invoked on them. For example to start a VM:

session.xenapi.VM.start(vm, False, False)

All API calls are by default synchronous and will not return until the operation has completed or failed. For example in the case of VM.start the call does not return until the VM has started booting.


When the VM.start call returns, the VM will be booting. To determine when the booting has finished, wait for the in-guest agent to report internal metrics through the VM_guest_metrics object.

Using Tasks to manage asynchronous operations

To simplify managing operations which take quite a long time (for example, VM.clone and VM.copy) functions are available in two forms: synchronous (the default) and asynchronous. Each asynchronous function returns a reference to a task object which contains information about the in-progress operation including:

  • whether it is pending

  • whether it is has succeeded or failed

  • progress (in the range 0-1)

  • the result or error code returned by the operation

An application which wanted to track the progress of a VM.clone operation and display a progress bar would have code like the following:

vm = session.xenapi.VM.get_by_name_label('my vm')
task = session.xenapi.Async.VM.clone(vm)
while session.xenapi.task.get_status(task) == "pending":
    progress = session.xenapi.task.get_progress(task)


A well-behaved client must delete tasks created by asynchronous operations when it has finished reading the result or error. If the number of tasks exceeds a built-in threshold then the server will delete the oldest of the completed tasks.

Subscribing to and listening for events

With the exception of the task and metrics classes, whenever an object is modified the server generates an event. Clients can subscribe to this event stream on a per-class basis and receive updates rather than resorting to frequent polling. Events come in three types:

  • add - generated when an object has been created;

  • del - generated immediately before an object is destroyed; and

  • mod - generated when an object’s field has changed.

Events also contain a monotonically increasing ID, the name of the class of object, and a snapshot of the object state equivalent to the result of a get_record().

Clients can receive events by calling event.from() with a list of class names or the special string * (to receive events for all classes). The output of event.from() includes a token, which can be passed into a subsequent event.from() call to receive only the events that have occurred since the last call. If an empty string is passed, event.from() will return all events.

The following Python code fragment demonstrates how to print a summary of events for all classes generated by a system:

token = ''
fmt = "%8s %s %20s  %5s  %s %s"

while True:
        output = session.xenapi.event_from("*", token, 30.0)
        events = output['events']
        token = output['token']

        for event in events:
            name = "n/a"
            snapshot = ''
            if "snapshot" in event.keys():
                snapshot = event['snapshot']
                if "name_label" in snapshot.keys():
                    name = snapshot['name_label']
            print fmt % (event['id'], event["ref"], event['class'], event['operation'], name, repr(snapshot))


The full script can be found at on Github.

Complete application examples

This section describes two complete examples of real programs using the API.

Simultaneously migrating VMs using live migration

This python example (the full script can be found at on Github) demonstrates how to use live migration to move VMs simultaneously between hosts in a Resource Pool. The example makes use of asynchronous API calls and shows how to wait for a set of tasks to complete.

The program begins with some standard boilerplate and imports the API module

import sys, time
import XenAPI

Next the commandline arguments containing a server URL, user name, password and a number of iterations are parsed. The user name and password are used to establish a session which is passed to the function main, which is called multiple times in a loop. Note the use of try: finally: to make sure the program logs out of its session at the end.

if __name__ == "__main__":
    if len(sys.argv) <> 5:
        print "Usage:"
        print sys.argv[0], " <url> <username> <password> <iterations>"
    url = sys.argv[1]
    username = sys.argv[2]
    password = sys.argv[3]
    iterations = int(sys.argv[4])
    # First acquire a valid session by logging in:
    session = XenAPI.Session(url)
    session.xenapi.login_with_password(username, password, "2.3",
                                        "Example migration-demo v0.1")
        for i in range(iterations):
            main(session, i)

The main function examines each running VM in the system, taking care to filter out control domains (which are part of the system and not controllable by the user). A list of running VMs and their current hosts is constructed.

def main(session, iteration):
    # Find a non-template VM object
    all = session.xenapi.VM.get_all()
    vms = []
    hosts = []
    for vm in all:
        record = session.xenapi.VM.get_record(vm)
        if not(record["is_a_template"]) and \
            not(record["is_control_domain"]) and \
            record["power_state"] == "Running":
    print "%d: Found %d suitable running VMs" % (iteration, len(vms))

Next the list of hosts is rotated:

# use a rotation as a permutation
hosts = [hosts[-1]] + hosts[:(len(hosts)-1)]

Each VM is then moved using live migration to the new host under this rotation (that is, a VM running on host at position 2 in the list is moved to the host at position 1 in the list, and so on.) In order to execute each of the movements in parallel, the asynchronous version of the VM.pool_migrate is used and a list of task references constructed. Note the live flag passed to the VM.pool_migrate; this causes the VMs to be moved while they are still running.

tasks = []
    for i in range(0, len(vms)):
        vm = vms[i]
        host = hosts[i]
        task = session.xenapi.Async.VM.pool_migrate(vm, host, { "live": "true" })

The list of tasks is then polled for completion:

finished = False
    records = {}
    while not(finished):
        finished = True
        for task in tasks:
            record = session.xenapi.task.get_record(task)
            records[task] = record
            if record["status"] == "pending":
                finished = False

Once all tasks have left the pending state (i.e. they have successfully completed, failed or been cancelled) the tasks are polled once more to see if they all succeeded:

allok = True
    for task in tasks:
        record = records[task]
        if record["status"] <> "success":
            allok = False

If any one of the tasks failed then details are printed, an exception is raised and the task objects left around for further inspection. If all tasks succeeded then the task objects are destroyed and the function returns.

if not(allok):
        print "One of the tasks didn't succeed at", \
            time.strftime("%F:%HT%M:%SZ", time.gmtime())
        idx = 0
        for task in tasks:
            record = records[task]
            vm_name = session.xenapi.VM.get_name_label(vms[idx])
            host_name =[idx])
            print "%s : %12s %s -> %s [ status: %s; result = %s; error = %s ]" % \
                    (record["uuid"], record["name_label"], vm_name, host_name,      \
                    record["status"], record["result"], repr(record["error_info"]))
            idx = idx + 1
        raise "Task failed"
        for task in tasks:

Cloning a VM using the xe CLI

This example is a bash script which uses the xe CLI to clone a VM taking care to shut it down first if it is powered on.

The example begins with some boilerplate which first checks if the environment variable XE has been set: if it has it assumes that it points to the full path of the CLI, else it is assumed that the xe CLI is on the current path. Next the script prompts the user for a server name, user name and password:

# Allow the path to the 'xe' binary to be overridden by the XE environment variable
if [ -z "${XE}" ]; then

if [ ! -e "${HOME}/.xe" ]; then
    read -p "Server name: " SERVER
    read -p "Username: " USERNAME
    read -p "Password: " PASSWORD
    XE="${XE} -s ${SERVER} -u ${USERNAME} -pw ${PASSWORD}"

Next the script checks its commandline arguments. It requires exactly one: the UUID of the VM which is to be cloned:

# Check if there's a VM by the uuid specified
${XE} vm-list params=uuid | grep -q " ${vmuuid}$"
if [ $? -ne 0 ]; then
        echo "error: no vm uuid \"${vmuuid}\" found"
        exit 2

The script then checks the power state of the VM and if it is running, it attempts a clean shutdown. The event system is used to wait for the VM to enter state “Halted”.


The xe CLI supports a command-line argument --minimal which causes it to print its output without excess whitespace or formatting, ideal for use from scripts. If multiple values are returned they are comma-separated.

# Check the power state of the vm
name=$(${XE} vm-list uuid=${vmuuid} params=name-label --minimal)
state=$(${XE} vm-list uuid=${vmuuid} params=power-state --minimal)

# If the VM state is running, we shutdown the vm first
if [ "${state}" = "running" ]; then
        ${XE} vm-shutdown uuid=${vmuuid}
        ${XE} event-wait class=vm power-state=halted uuid=${vmuuid}

The VM is then cloned and the new VM has its name_label set to cloned_vm.

# Clone the VM
newuuid=$(${XE} vm-clone uuid=${vmuuid} new-name-label=cloned_vm)

Finally, if the original VM had been running and was shutdown, both it and the new VM are started.

# If the VM state was running before cloning, we start it again
# along with the new VM.
if [ "$wasrunning" -eq 1 ]; then
        ${XE} vm-start uuid=${vmuuid}
        ${XE} vm-start uuid=${newuuid}
Using the API