Environment

The Toolstack runs in an environment on a server (host) that has:

• Physical hardware.
• The Xen hypervisor.
• The control domain (domain 0): the priviledged domain that the Toolstack runs in.
• Other, mostly unpriviledged domains, usually for guests (VMs).

The Toolstack relies on various bits of software inside the control domain, and directly communicates with most of these:

• Linux kernel including drivers for hardware and Xen paravirtualised devices (e.g. netback and blkback).
  • Interacts through /sys and /proc, udev scripts, xenstore, …
• CentOS distibution including userspace tools and libraries.
  • systemd, networking tools, …
• Xen-specific libraries, especially libxenctrl (a.k.a. libxc)
• xenstored: a key-value pair configuration database
  • Accessible from all domains on a host, which makes it useful for inter-domain communication.
  • The control domain has access to the entire xenstore database, while other domains only see sub-trees that are specific to that domain.
  • Used for connecting VM disks and network interfaces, and other VM configuration options.
  • Used for VM status reporting, e.g. the capabilities of the PV drivers (if installed), the IP address, etc.
• SM: Storage Manager plugins which connect xapi’s internal storage interfaces to the control APIs of external storage systems.
• stunnel: a daemon which decodes TLS and forwards traffic to xapi (and the other way around).
• Open vSwitch (OVS): a virtual network switch, used to connect VMs to network interfaces. The OVS offers several networking features that xapi takes advantage of.
• QEMU: emulation of various bits of hardware
• DEMU: emulation of Nvidia vGPUs
• xenguest
• emu-manager
• pvsproxy
• xenconsoled: allows access to guest consoles. This is common to all Xen hosts.

The Toolstack also interacts with software that runs inside the guests:

• PV drivers
• The guest agent

Daemons

The Toolstack consists of a set of co-operating daemons:

xapi

manages clusters of hosts, co-ordinating access to shared storage and networking.

xenopsd

a low-level “domain manager” which takes care of creating, suspending, resuming, migrating, rebooting domains by interacting with Xen via libxc and libxl.

xcp-rrdd

a performance counter monitoring daemon which aggregates “datasources” defined via a plugin API and records history for each. There are various rrdd-plugin daemons:

• xcp-rrdd-gpumon
• xcp-rrdd-iostat
• xcp-rrdd-squeezed
• xcp-rrdd-xenpm

xcp-networkd

a host network manager which takes care of configuring interfaces, bridges and OpenVSwitch instances

squeezed

a daemon in charge of VM memory management

xapi-storage-script

for storage manipulation over SMAPIv3

message-switch

exchanges messages between the daemons on a host

xapi-guard

forwards uefi and vtpm persistence calls from domains to xapi

v6d

controls which features are enabled.

forkexecd

a helper daemon that assists the above daemons with executing binaries and scripts

xhad

The High-Availability daemon

perfmon

a daemon which monitors performance counters and sends “alerts” if values exceed some pre-defined threshold

mpathalert

a daemon which monitors “storage paths” and sends “alerts” if paths fail and need repair

wsproxy

handles access to VM consoles

Interfaces

Communication between the Toolstack daemon is built upon libraries from a component called xapi-idl.

• Abstracts communication between daemons over the message-switch using JSON/RPC.
• Contains the definition of the interfaces exposed by the daemons (except xapi).

Disaster Recovery

The HA feature will restart VMs after hosts have failed, but what happens if a whole site (e.g. datacenter) is lost? A disaster recovery configuration is shown in the following diagram:

We rely on the storage array’s built-in mirroring to replicate (synchronously or asynchronously: the admin’s choice) between the primary and the secondary site. When DR is enabled the VM disk data and VM metadata are written to the storage server and mirrored. The secondary site contains the other side of the data mirror and a set of hosts, which may be powered off.

In normal operation, the DR feature allows a “dry-run” recovery where a host on the secondary site checks that it can indeed see all the VM disk data and metadata. This should be done regularly, so that admins are familiar with the process.

After a disaster, the admin breaks the mirror on the secondary site and triggers a remote power-on of the offline hosts (either using an out-of-band tool or the built-in host power-on feature of xapi). The pool master on the secondary site can connect to the storage and extract all the VM metadata. Finally the VMs can all be restarted.

When the primary site is fully recovered, the mirror can be re-synchronised and the VMs can be moved back.

Event handling in the Control Plane - Xapi, Xenopsd and Xenstore

Introduction

Xapi, xenopsd and xenstore use a number of different events to obtain indications that some state changed in dom0 or in the guests. The events are used as an efficient alternative to polling all these states periodically.

• xenstore provides a very configurable approach in which each and any key can be watched individually by a xenstore client. Once the value of a watched key changes, xenstore will indicate to the client that the value for that key has changed. An ocaml xenstore client library provides a way for ocaml programs such as xenopsd, message-cli and rrdd to provide high-level ocaml callback functions to watch specific key. It’s very common, for instance, for xenopsd to watch specific keys in the xenstore keyspace of a guest and then after receiving events for some or all of them, read other keys or subkeys in xenstored to update its internal state mirroring the state of guests and its devices (for instance, if the guest has pv drivers and specific frontend devices have established connections with the backend devices in dom0).
• xapi also provides a very configurable event mechanism in which the xenapi can be used to provide events whenever a xapi object (for instance, a VM, a VBD etc) changes state. This event mechanism is very reliable and is extensively used by XenCenter to provide real-time update on the XenCenter GUI.
• xenopsd provides a somewhat less configurable event mechanism, where it always provides signals for all objects (VBDs, VMs etc) whose state changed (so it’s not possible to select a subset of objects to watch for as in xenstore or in xapi). It’s up to the xenopsd client (eg. xapi) to receive these events and then filter out or act on each received signal by calling back xenopsd and asking it information for the specific signalled object.  The main use in xapi for the xenopsd signals is to update xapi’s database of the current state of each object controlled by xenopsd (VBDs, VMs etc).

Given a choice between polling states and receiving events when the state change, we should in general opt for receiving events in the code in order to avoid adding bottlenecks in dom0 that will prevent the scalability of XenServer to many VMs and virtual devices.

Xapi

Sending events from the xenapi

A xenapi user client, such as XenCenter, the xe-cli or a python script, can register to receive events from XAPI for specific objects in the XAPI DB. XAPI will generate events for those registered clients whenever the corresponding XAPI DB object changes.

This small python scripts shows how to register a simple event watch loop for XAPI:

import XenAPI
session = XenAPI.Session("http://xshost")
session.login_with_password("username","password")
session.xenapi.event.register(["VM","pool"]) # register for events in the pool and VM objects                                                
while True:
  try:
    events = session.xenapi.event.next() # block until a xapi event on a xapi DB object is available
    for event in events:
      print "received event op=%s class=%s ref=%s" % (event['operation'], event['class'], event['ref'])                                      
      if event['class'] == 'vm' and event['operatoin'] == 'mod':
        vm = event['snapshot']
        print "xapi-event on vm: vm_uuid=%s, power_state=%s, current_operation=%s" % (vm['uuid'],vm['name_label'],vm['power_state'],vm['current_operations'].values())
  except XenAPI.Failure, e:
    if len(e.details) > 0 and e.details[0] == 'EVENTS_LOST':
      session.xenapi.event.unregister(["VM","pool"])
      session.xenapi.event.register(["VM","pool"])

Receiving events from xenopsd

Xapi receives all events from xenopsd via the function xapi_xenops.events_watch() in its own independent thread. This is a single-threaded function that is responsible for handling all of the signals sent by xenopsd. In some situations with lots of VMs and virtual devices such as VBDs, this loop may saturate a single dom0 vcpu, which will slow down handling all of the xenopsd events and may cause the xenopsd signals to accumulate unboundedly in the worst case in the updates queue in xenopsd (see Figure 1).

The function xapi_xenops.events_watch() calls xenops_client.UPDATES.get() to obtain a list of (barrier,  barrier_events), and then it process each one of the barrier_event, which can be one of the following events:

• Vm id: something changed in this VM, run xapi_xenops.update_vm() to query xenopsd about its state. The function update_vm() will update power_state, allowed_operations, console and guest_agent state in the xapi DB.
• Vbd id: something changed in this VM, run xapi_xenops.update_vbd() to query xenopsd about its state. The function update_vbd() will update currently_attached and connected in the xapi DB.
• Vif id: something changed in this VM, run xapi_xenops.update_vif() to query xenopsd about its state. The function update_vif() will update activate and plugged state of in the xapi DB.
• Pci id: something changed in this VM, run xapi_xenops.update_pci() to query xenopsd about its state.
• Vgpu id: something changed in this VM, run xapi_xenops.update_vgpu() to query xenopsd about its state.
• Task id: something changed in this VM, run xapi_xenops.update_task() to query xenopsd about its state. The function update_task() will update the progress of the task in the xapi DB using the information of the task in xenopsd.

All the xapi_xenops.update_X() functions above will call Xenopsd_client.X.stat() functions to obtain the current state of X from xenopsd:

There are a couple of optimisations while processing the events in xapi_xenops.events_watch():

• if an event X=(vm_id,dev_id) (eg. Vbd dev_id) has already been processed in a barrier_events, it’s not processed again. A typical value for X is eg. “<vm_uuid>.xvda” for a VBD.
• if Events_from_xenopsd.are_supressed X, then this event is ignored. Events are supressed if VM X.vm_id is migrating away from the host

Barriers

When xapi needs to execute (and to wait for events indicating completion of) a xapi operation (such as VM.start and VM.shutdown) containing many xenopsd sub-operations (such as VM.start – to force xenopsd to change the VM power_state, and VM.stat, VBD.stat, VIF.stat etc – to force the xapi DB to catch up with the xenopsd new state for these objects), xapi sends to the xenopsd input queue a barrier, indicating that xapi will then block and only continue execution of the barred operation when xenopsd returns the barrier. The barrier should only be returned when xenopsd has finished the execution of all the operations requested by xapi (such as VBD.stat and VM.stat in order to update the state of the VM in the xapi database after a VM.start has been issued to xenopsd). 

A recent problem has been detected in the xapi_xenops.events_watch()  function: when it needs to process many VM_check_state events, this may push for later the processing of barriers associated with a VM.start, delaying xapi in reporting (via a xapi event) that the VM state in the xapi DB has reached the running power_state. This needs further debugging, and is probably one of the reasons in CA-87377 why in some conditions a xapi event reporting that the VM power_state is running (causing it to go from yellow to green state in XenCenter) is taking so long to be returned, way after the VM is already running.

Xenopsd

Xenopsd has a few queues that are used by xapi to store commands to be executed (eg. VBD.stat) and update events to be picked up by xapi. The main ones, easily seen at runtime by running the following command in dom0, are:

# xenops-cli diagnostics --queue=org.xen.xapi.xenops.classic
{
   queues: [  # XENOPSD INPUT QUEUE
            ... stuff that still needs to be processed by xenopsd
            VM.stat
            VBD.stat
            VM.start
            VM.shutdown
            VIF.plug
            etc
           ]
   workers: [ # XENOPSD WORKER THREADS
            ... which stuff each worker thread is processing
   ]
   updates: {
     updates: [ # XENOPSD OUTPUT QUEUE
            ... signals from xenopsd that need to be picked up by xapi
               VM_check_state
               VBD_check_state
               etc
        ]
      } tasks: [ # XENOPSD TASKS
               ... state of each known task, before they are manually deleted after completion of the task
               ]
}

Sending events to xapi

Whenever xenopsd changes the state of a XenServer object such as a VBD or VM, or when it receives an event from xenstore indicating that the states of these objects have changed (perhaps because either a guest or the dom0 backend changed the state of a virtual device), it creates a signal for the corresponding object (VM_check_state, VBD_check_state etc) and send it up to xapi. Xapi will then process this event in its xapi_xenops.events_watch() function.

These signals may need to wait a long time to be processed if the single-threaded xapi_xenops.events_watch() function is having difficulties (ie taking a long time) to process previous signals in the UPDATES queue from xenopsd.  

Receiving events from xenstore

Xenopsd watches a number of keys in xenstore, both in dom0 and in each guest. Xenstore is responsible to send watch events to xenopsd whenever the watched keys change state. Xenopsd uses a xenstore client library to make it easier to create a callback function that is called whenever xenstore sends these events.

Xenopsd also needs to complement sometimes these watch events with polling of some values. An example is the @introduceDomain event in xenstore (handled in xenopsd/xc/xenstore_watch.ml), which indicates that a new VM has been created. This event unfortunately does not indicate the domid of the VM, and xenopsd needs to query Xen (via libxc) which domains are now available in the host and compare with the previous list of known domains, in order to figure out the domid of the newly introduced domain.

 It is not good practice to poll xenstore for changes of values. This will add a large overhead to both xenstore and xenopsd, and decrease the scalability of XenServer in terms of number of VMs/host and virtual devices per VM. A much better approach is to rely on the watch events of xenstore to indicate when a specific value has changed in xenstore.

Xenstore

Sending events to xenstore clients

If a xenstore client has created watch events for a key, then xenstore will send events to this client whenever this key changes state.

Receiving events from xenstore clients

Xenstore clients indicate to xenstore that something state changed by writing to some xenstore key. This may or may not cause xenstore to create watch events for the corresponding key, depending on if other xenstore clients have watches on this key.

High-Availability

High-Availability (HA) tries to keep VMs running, even when there are hardware failures in the resource pool, when the admin is not present. Without HA the following may happen:

• during the night someone spills a cup of coffee over an FC switch; then
• VMs running on the affected hosts will lose access to their storage; then
• business-critical services will go down; then
• monitoring software will send a text message to an off-duty admin; then
• the admin will travel to the office and fix the problem by restarting the VMs elsewhere.

With HA the following will happen:

• during the night someone spills a cup of coffee over an FC switch; then
• VMs running on the affected hosts will lose access to their storage; then
• business-critical services will go down; then
• the HA software will determine which hosts are affected and shut them down; then
• the HA software will restart the VMs on unaffected hosts; then
• services are restored; then on the next working day
• the admin can arrange for the faulty switch to be replaced.

HA is designed to handle an emergency and allow the admin time to fix failures properly.

Example

The following diagram shows an HA-enabled pool, before and after a network link between two hosts fails.

When HA is enabled, all hosts in the pool

• exchange periodic heartbeat messages over the network
• send heartbeats to a shared storage device.
• attempt to acquire a “master lock” on the shared storage.

HA is designed to recover as much as possible of the pool after a single failure i.e. it removes single points of failure. When some subset of the pool suffers a failure then the remaining pool members

• figure out whether they are in the largest fully-connected set (the “liveset”);
  • if they are not in the largest set then they “fence” themselves (i.e. force reboot via the hypervisor watchdog)
• elect a master using the “master lock”
• restart all lost VMs.

After HA has recovered a pool, it is important that the original failure is addressed because the remaining pool members may not be able to cope with any more failures.

Design

HA must never violate the following safety rules:

1. there must be at most one master at all times. This is because the master holds the VM and disk locks.
2. there must be at most one instance of a particular VM at all times. This is because starting the same VM twice will result in severe filesystem corruption.

However to be useful HA must:

• detect failures quickly;
• minimise the number of false-positives in the failure detector; and
• make the failure handling logic as robust as possible.

The implementation difficulty arises when trying to be both useful and safe at the same time.

Terminology

We use the following terminology:

• fencing: also known as I/O fencing, refers to the act of isolating a host from network and storage. Once a host has been fenced, any VMs running there cannot generate side-effects observable to a third party. This means it is safe to restart the running VMs on another node without violating the safety-rule and running the same VM simultaneously in two locations.
• heartbeating: exchanging status updates with other hosts at regular pre-arranged intervals. Heartbeat messages reveal that hosts are alive and that I/O paths are working.
• statefile: a shared disk (also known as a “quorum disk”) on the “Heartbeat” SR which is mapped as a block device into every host’s domain 0. The shared disk acts both as a channel for heartbeat messages and also as a building block of a Pool master lock, to prevent multiple hosts becoming masters in violation of the safety-rule (a dangerous situation also known as “split-brain”).
• management network: the network over which the XenAPI XML/RPC requests flow and also used to send heartbeat messages.
• liveset: a per-Host view containing a subset of the Hosts in the Pool which are considered by that Host to be alive i.e. responding to XenAPI commands and running the VMs marked as resident_on there. When a Host b leaves the liveset as seen by Host a it is safe for Host a to assume that Host b has been fenced and to take recovery actions (e.g. restarting VMs), without violating either of the safety-rules.
• properly shared SR: an SR which has field shared=true; and which has a PBD connecting it to every enabled Host in the Pool; and where each of these PBDs has field currently_attached set to true. A VM whose disks are in a properly shared SR could be restarted on any enabled Host, memory and network permitting.
• properly shared Network: a Network which has a PIF connecting it to every enabled Host in the Pool; and where each of these PIFs has field currently_attached set to true. A VM whose VIFs connect to properly shared Networks could be restarted on any enabled Host, memory and storage permitting.
• agile: a VM is said to be agile if all disks are in properly shared SRs and all network interfaces connect to properly shared Networks.
• unprotected: an unprotected VM has field ha_always_run set to false and will never be restarted automatically on failure or have reconfiguration actions blocked by the HA overcommit protection.
• best-effort: a best-effort VM has fields ha_always_run set to true and ha_restart_priority set to best-effort. A best-effort VM will only be restarted if (i) the failure is directly observed; and (ii) capacity exists for an immediate restart. No more than one restart attempt will ever be made.
• protected: a VM is said to be protected if it will be restarted by HA i.e. has field ha_always_run set to true and field ha_restart_priority not set to `best-effort.
• survival rule 1: describes the situation where hosts survive because they are in the largest network partition with statefile access. This is the normal state of the xhad daemon.
• survival rule 2: describes the situation where all hosts have lost access to the statefile but remain alive while they can all see each-other on the network. In this state any further failure will cause all nodes to self-fence. This state is intended to cope with the system-wide temporary loss of the storage service underlying the statefile.

Assumptions

We assume:

• All I/O used for monitoring the health of hosts (i.e. both storage and network-based heartbeating) is along redundant paths, so that it survives a single hardware failure (e.g. a broken switch or an accidentally-unplugged cable). It is up to the admin to ensure their environment is setup correctly.
• The hypervisor watchdog mechanism will be able to guarantee the isolation of nodes, once communication has been lost, within a pre-arranged time period. Therefore no active power fencing equipment is required.
• VMs may only be marked as protected if they are fully agile i.e. able to run on any host, memory permitting. No additional constraints of any kind may be specified e.g. it is not possible to make “CPU reservations”.
• Pools are assumed to be homogenous with respect to CPU type and presence of VT/SVM support (also known as “HVM”). If a Pool is created with non-homogenous hosts using the --force flag then the additional constraints will not be noticed by the VM failover planner resulting in runtime failures while trying to execute the failover plans.
• No attempt will ever be made to shutdown or suspend “lower” priority VMs to guarantee the survival of “higher” priority VMs.
• Once HA is enabled it is not possible to reconfigure the management network or the SR used for storage heartbeating.
• VMs marked as protected are considered to have failed if they are offline i.e. the VM failure handling code is level-sensitive rather than edge-sensitive.
• VMs marked as best-effort are considered to have failed only when the host where they are resident is declared offline i.e. the best-effort VM failure handling code is edge-sensitive rather than level-sensitive. A single restart attempt is attempted and if this fails no further start is attempted.
• HA can only be enabled if all Pool hosts are online and actively responding to requests.
• when HA is enabled the database is configured to write all updates to the “Heartbeat” SR, guaranteeing that VM configuration changes are not lost when a host fails.

Components

The implementation is split across the following components:

• xhad: the cluster membership daemon maintains a quorum of hosts through network and storage heartbeats
• xapi: used to configure the HA policy i.e. which network and storage to use for heartbeating and which VMs to restart after a failure.
• xen: the Xen watchdog is used to reliably fence the host when the host has been (partially or totally) isolated from the cluster

To avoid a “split-brain”, the cluster membership daemon must “fence” (i.e. isolate) nodes when they are not part of the cluster. In general there are 2 approaches:

• cut the power of remote hosts which you can’t talk to on the network any more. This is the approach taken by most open-source clustering software since it is simpler. However it has the downside of requiring the customer buy more hardware and set it up correctly.
• rely on the remote hosts using a watchdog to cut their own power (i.e. halt or reboot) after a timeout. This relies on the watchdog being reliable. Most other people don’t trust the Linux watchdog; after all the Linux kernel is highly threaded, performs a lot of (useful) functions and kernel bugs which result in deadlocks do happen. We use the Xen watchdog because we believe that the Xen hypervisor is simple enough to reliably fence the host (via triggering a reboot of domain 0 which then triggers a host reboot).

xhad

xhad is the cluster membership daemon: it exchanges heartbeats with the other nodes to determine which nodes are still in the cluster (the “live set”) and which nodes have definitely failed (through watchdog fencing). When a host has definitely failed, xapi will unlock all the disks and restart the VMs according to the HA policy.

Since Xapi is a critical part of the system, the xhad also acts as a Xapi watchdog. It polls Xapi every few seconds and checks if Xapi can respond. If Xapi seems to have failed then xhad will restart it. If restarts continue to fail then xhad will consider the host to have failed and self-fence.

xhad is configured via a simple config file written on each host in /etc/xensource/xhad.conf. The file must be identical on each host in the cluster. To make changes to the file, HA must be disabled and then re-enabled afterwards. Note it may not be possible to re-enable HA depending on the configuration change (e.g. if a host has been added but that host has a broken network configuration then this will block HA enable).

The xhad.conf file is written in XML and contains

• pool-wide configuration: this includes a list of all hosts which should be in the liveset and global timeout information
• local host configuration: this identifies the local host and described which local network interface and block device to use for heartbeating.

The following is an example xhad.conf file:

<?xml version="1.0" encoding="utf-8"?>
<xhad-config version="1.0">

  <!--pool-wide configuration-->
  <common-config>
    <GenerationUUID>xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx</GenerationUUID>
    <UDPport>694</UDPport>

    <!--for each host, specify host UUID, and IP address-->
    <host>
      <HostID>xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx</HostID>
      <IPaddress>xxx.xxx.xxx.xx1</IPaddress>
    </host>

    <host>
      <HostID>xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx</HostID>
      <IPaddress>xxx.xxx.xxx.xx2</IPaddress>
    </host>

    <host>
      <HostID>xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx</HostID>
      <IPaddress>xxx.xxx.xxx.xx3</IPaddress>
    </host>

    <!--optional parameters [sec] -->
    <parameters>
      <HeartbeatInterval>4</HeartbeatInterval>
      <HeartbeatTimeout>30</HeartbeatTimeout>
      <StateFileInterval>4</StateFileInterval>
      <StateFileTimeout>30</StateFileTimeout>
      <HeartbeatWatchdogTimeout>30</HeartbeatWatchdogTimeout>
      <StateFileWatchdogTimeout>45</StateFileWatchdogTimeout>
      <BootJoinTimeout>90</BootJoinTimeout>
      <EnableJoinTimeout>90</EnableJoinTimeout>
      <XapiHealthCheckInterval>60</XapiHealthCheckInterval>
      <XapiHealthCheckTimeout>10</XapiHealthCheckTimeout>
      <XapiRestartAttempts>1</XapiRestartAttempts>
      <XapiRestartTimeout>30</XapiRestartTimeout>
      <XapiLicenseCheckTimeout>30</XapiLicenseCheckTimeout>
    </parameters>
  </common-config>

  <!--local host configuration-->
  <local-config>
    <localhost>
      <HostID>xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxx2</HostID>
      <HeartbeatInterface> xapi1</HeartbeatInterface>
      <HeartbeatPhysicalInterface>bond0</HeartbeatPhysicalInterface>
      <StateFile>/dev/statefiledevicename</StateFile>
    </localhost>
  </local-config>

</xhad-config>

The fields have the following meaning:

• GenerationUUID: a UUID generated each time HA is reconfigured. This allows xhad to tell an old host which failed; had been removed from the configuration; repaired and then restarted that the world has changed while it was away.
• UDPport: the port number to use for network heartbeats. It’s important to allow this traffic through the firewall and to make sure the same port number is free on all hosts (beware of portmap services occasionally binding to it).
• HostID: a UUID identifying a host in the pool. We would normally use xapi’s notion of a host uuid.
• IPaddress: any IP address on the remote host. We would normally use xapi’s notion of a management network.
• HeartbeatTimeout: if a heartbeat packet is not received for this many seconds, then xhad considers the heartbeat to have failed. This is the user-supplied “HA timeout” value, represented below as T. T must be bigger than 10; we would normally use 60s.
• StateFileTimeout: if a storage update is not seen for a host for this many seconds, then xhad considers the storage heartbeat to have failed. We would normally use the same value as the HeartbeatTimeout T.
• HeartbeatInterval: interval between heartbeat packets sent. We would normally use a value 2 <= t <= 6, derived from the user-supplied HA timeout via t = (T + 10) / 10
• StateFileInterval: interval betwen storage updates (also known as “statefile updates”). This would normally be set to the same value as HeartbeatInterval.
• HeartbeatWatchdogTimeout: If the host does not send a heartbeat for this amount of time then the host self-fences via the Xen watchdog. We normally set this to T.
• StateFileWatchdogTimeout: If the host does not update the statefile for this amount of time then the host self-fences via the Xen watchdog. We normally set this to T+15.
• BootJoinTimeout: When the host is booting and joining the liveset (i.e. the cluster), consider the join a failure if it takes longer than this amount of time. We would normally set this to T+60.
• EnableJoinTimeout: When the host is enabling HA for the first time, consider the enable a failure if it takes longer than this amount of time. We would normally set this to T+60.
• XapiHealthCheckInterval: Interval between “health checks” where we run a script to check whether Xapi is responding or not.
• XapiHealthCheckTimeout: Number of seconds to wait before assuming that Xapi has deadlocked during a “health check”.
• XapiRestartAttempts: Number of Xapi restarts to attempt before concluding Xapi has permanently failed.
• XapiRestartTimeout: Number of seconds to wait for a Xapi restart to complete before concluding it has failed.
• XapiLicenseCheckTimeout: Number of seconds to wait for a Xapi license check to complete before concluding that xhad should terminate.

In addition to the config file, Xhad exposes a simple control API which is exposed as scripts:

• ha_set_pool_state (Init | Invalid): sets the global pool state to “Init” (before starting HA) or “Invalid” (causing all other daemons who can see the statefile to shutdown)
• ha_start_daemon: if the pool state is “Init” then the daemon will attempt to contact other daemons and enable HA. If the pool state is “Active” then the host will attempt to join the existing liveset.
• ha_query_liveset: returns the current state of the cluster.
• ha_propose_master: returns whether the current node has been elected pool master.
• ha_stop_daemon: shuts down the xhad on the local host. Note this will not disarm the Xen watchdog by itself.
• ha_disarm_fencing: disables fencing on the local host.
• ha_set_excluded: when a host is being shutdown cleanly, record the fact that the VMs have all been shutdown so that this host can be ignored in future cluster membership calculations.

Fencing

Xhad continuously monitors whether the host should remain alive, or if it should self-fence. There are two “survival rules” which will keep a host alive; if neither rule applies (or if xhad crashes or deadlocks) then the host will fence. The rules are:

1. Xapi is running; the storage heartbeats are visible; this host is a member of the “best” partition (as seen through the storage heartbeats)
2. Xapi is running; the storage is inaccessible; all hosts which should be running (i.e. not those “excluded” by being cleanly shutdown) are online and have also lost storage access (as seen through the network heartbeats).

where the “best” partition is the largest one if that is unique, or if there are multiple partitions of the same size then the one containing the lowest host uuid is considered best.

The first survival rule is the “normal” case. The second rule exists only to prevent the storage from becoming a single point of failure: all hosts can remain alive until the storage is repaired. Note that if a host has failed and has not yet been repaired, then the storage becomes a single point of failure for the degraded pool. HA removes single point of failures, but multiple failures can still cause problems. It is important to fix failures properly after HA has worked around them.

xapi

Xapi is responsible for

• exposing an interface for setting HA policy
• creating VDIs (disks) on shared storage for heartbeating and storing the pool database
• arranging for these disks to be attached on host boot, before the “SRmaster” is online
• configuring and managing the xhad heartbeating daemon

The HA policy APIs include

• methods to determine whether a VM is agile i.e. can be restarted in principle on any host after a failure
• planning for a user-specified number of host failures and enforcing access control
• restarting failed protected VMs in policy order

The HA policy settings are stored in the Pool database which is written (synchronously) to a VDI in the same SR that’s being used for heartbeating. This ensures that the database can be recovered after a host fails and the VMs are recovered.

Xapi stores 2 settings in its local database:

• ha_disable_failover_actions: this is set to false when we want nodes to be able to recover VMs – this is the normal case. It is set to true during the HA disable process to prevent a split-brain forming while HA is only partially enabled.
• ha_armed: this is set to true to tell Xapi to start Xhad during host startup and wait to join the liveset.

Disks on shared storage

The regular disk APIs for creating, destroying, attaching, detaching (etc) disks need the SRmaster (usually but not always the Pool master) to be online to allow the disks to be locked. The SRmaster cannot be brought online until the host has joined the liveset. Therefore we have a cyclic dependency: joining the liveset needs the statefile disk to be attached but attaching a disk requires being a member of the liveset already.

The dependency is broken by adding an explicit “unlocked” attach storage API called VDI_ATTACH_FROM_CONFIG. Xapi uses the VDI_GENERATE_CONFIG API during the HA enable operation and stores away the result. When the system boots the VDI_ATTACH_FROM_CONFIG is able to attach the disk without the SRmaster.

The role of Host.enabled

The Host.enabled flag is used to mean, “this host is ready to start VMs and should be included in failure planning”. The VM restart planner assumes for simplicity that all protected VMs can be started anywhere; therefore all involved networks and storage must be properly shared. If a host with an unplugged PBD were to become enabled then the corresponding SR would cease to be properly shared, all the VMs would cease to be agile and the VM restart logic would fail.

To ensure the VM restart logic always works, great care is taken to make sure that Hosts may only become enabled when their networks and storage are properly configured. This is achieved by:

• when the master boots and initialises its database it sets all Hosts to dead and disabled and then signals the HA background thread (signal_database_state_valid) to wake up from sleep and start processing liveset information (and potentially setting hosts to live)
• when a slave calls Pool.hello (i.e. after the slave has rebooted), the master sets it to disabled, allowing it a grace period to plug in its storage;
• when a host (master or slave) successfully plugs in its networking and storage it calls consider_enabling_host which checks that the preconditions are met and then sets the host to enabled; and
• when a slave notices its database connection to the master restart (i.e. after the master xapi has just restarted) it calls consider_enabling_host}

The steady-state

When HA is enabled and all hosts are running normally then each calls ha_query_liveset every 10s.

Slaves check to see if the host they believe is the master is alive and has the master lock. If another node has become master then the slave will rewrite its pool.conf and restart. If no node is the master then the slave will call on_master_failure, proposing itself and, if it is rejected, checking the liveset to see which node acquired the lock.

The master monitors the liveset and updates the Host_metrics.live flag of every host to reflect the liveset value. For every host which is not in the liveset (i.e. has fenced) it enumerates all resident VMs and marks them as Halted. For each protected VM which is not running, the master computes a VM restart plan and attempts to execute it. If the plan fails then a best-effort VM.start call is attempted. Finally an alert is generated if the VM could not be restarted.

Note that XenAPI heartbeats are still sent when HA is enabled, even though they are not used to drive the values of the Host_metrics.live field. Note further that, when a host is being shutdown, the host is immediately marked as dead and its host reference is added to a list used to prevent the Host_metrics.live being accidentally reset back to live again by the asynchronous liveset query. The Host reference is removed from the list when the host restarts and calls Pool.hello.

Planning and overcommit

The VM failover planning code is sub-divided into two pieces, stored in separate files:

• binpack.ml: contains two algorithms for packing items of different sizes (i.e. VMs) into bins of different sizes (i.e. Hosts); and
• xapi_ha_vm_failover.ml: interfaces between the Pool database and the binpacker; also performs counterfactual reasoning for overcommit protection.

The input to the binpacking algorithms are configuration values which represent an abstract view of the Pool:

type ('a, 'b) configuration = {
  hosts:        ('a * int64) list; (** a list of live hosts and free memory *)
  vms:          ('b * int64) list; (** a list of VMs and their memory requirements *)
  placement:    ('b * 'a) list;    (** current VM locations *)
  total_hosts:  int;               (** total number of hosts in the pool 'n' *)
  num_failures: int;               (** number of failures to tolerate 'r' *)
}

Note that:

• the memory required by the VMs listed in placement has already been substracted from the free memory of the hosts; it doesn’t need to be subtracted again.
• the free memory of each host has already had per-host miscellaneous overheads subtracted from it, including that used by unprotected VMs, which do not appear in the VM list.
• the total number of hosts in the pool (total_hosts) is a constant for any particular invocation of HA.
• the number of failures to tolerate (num_failures) is the user-settable value from the XenAPI Pool.ha_host_failures_to_tolerate.

There are two algorithms which satisfy the interface:

sig
  plan_always_possible: ('a, 'b) configuration -> bool;
  get_specific_plan: ('a, 'b) configuration -> 'b list -> ('b * 'a) list
end

The function get_specific_plan takes a configuration and a list of VMs( the host where they are resident on have failed). It returns a VM restart plan represented as a VM to Host association list. This is the function called by the background HA VM restart thread on the master.

The function plan_always_possible returns true if every sequence of Host failures of length num_failures (irrespective of whether all hosts failed at once, or in multiple separate episodes) would result in calls to get_specific_plan which would allow all protected VMs to be restarted. This function is heavily used by the overcommit protection logic as well as code in XenCenter which aims to maximise failover capacity using the counterfactual reasoning APIs:

Pool.ha_compute_max_host_failures_to_tolerate
Pool.ha_compute_hypothetical_max_host_failures_to_tolerate

There are two binpacking algorithms: the more detailed but expensive algorithmm is used for smaller/less complicated pool configurations while the less detailed, cheaper algorithm is used for the rest. The choice between algorithms is based only on total_hosts (n) and num_failures (r). Note that the choice of algorithm will only change if the number of Pool hosts is varied (requiring HA to be disabled and then enabled) or if the user requests a new num_failures target to plan for.

The expensive algorithm uses an exchaustive search with a “biggest-fit-decreasing” strategy that takes the biggest VMs first and allocates them to the biggest remaining Host. The implementation keeps the VMs and Hosts as sorted lists throughout. There are a number of transformations to the input configuration which are guaranteed to preserve the existence of a VM to host allocation (even if the actual allocation is different). These transformations which are safe are:

• VMs may be removed from the list
• VMs may have their memory requirements reduced
• Hosts may be added
• Hosts may have additional memory added.

The cheaper algorithm is used for larger Pools where the state space to search is too large. It uses the same “biggest-fit-decreasing” strategy with the following simplifying approximations:

• every VM that fails is as big as the biggest
• the number of VMs which fail due to a single Host failure is always the maximum possible (even if these are all very small VMs)
• the largest and most capable Hosts fail

An informal argument that these approximations are safe is as follows: if the maximum number of VMs fail, each of which is size of the largest and we can find a restart plan using only the smaller hosts then any real failure:

• can never result in the failure of more VMs;
• can never result in the failure of bigger VMs; and
• can never result in less host capacity remaining.

Therefore we can take this almost-certainly-worse-than-worst-case failure plan and:

• replace the remaining hosts in the worst case plan with the real remaining hosts, which will be the same size or larger; and
• replace the failed VMs in the worst case plan with the real failed VMs, which will be fewer or the same in number and smaller or the same in size.

Note that this strategy will perform best when each host has the same number of VMs on it and when all VMs are approximately the same size. If one very big VM exists and a lot of smaller VMs then it will probably fail to find a plan. It is more tolerant of differing amounts of free host memory.

Overcommit protection

Overcommit protection blocks operations which would prevent the Pool being able to restart protected VMs after host failure. The Pool may become unable to restart protected VMs in two general ways: (i) by running out of resource i.e. host memory; and (ii) by altering host configuration in such a way that VMs cannot be started (or the planner thinks that VMs cannot be started).

API calls which would change the amount of host memory currently in use (VM.start, VM.resume, VM.migrate etc) have been modified to call the planning functions supplying special “configuration change” parameters. Configuration change values represent the proposed operation and have type

type configuration_change = {
  (** existing VMs which are leaving *)
  old_vms_leaving: (API.ref_host * (API.ref_VM * API.vM_t)) list;
  (** existing VMs which are arriving *)
  old_vms_arriving: (API.ref_host * (API.ref_VM * API.vM_t)) list;  
  (** hosts to pretend to disable *)
  hosts_to_disable: API.ref_host list;
  (** new number of failures to consider *)
  num_failures: int option;
  (** new VMs to restart *)  
  new_vms_to_protect: API.ref_VM list;
}

A VM migration will be represented by saying the VM is “leaving” one host and “arriving” at another. A VM start or resume will be represented by saying the VM is “arriving” on a host.

Note that no attempt is made to integrate the overcommit protection with the general VM.start host chooser as this would be quite expensive.

Note that the overcommit protection calls are written as asserts called within the message forwarder in the master, holding the main forwarding lock.

API calls which would change the system configuration in such a way as to prevent the HA restart planner being able to guarantee to restart protected VMs are also blocked. These calls include:

• VBD.create: where the disk is not in a properly shared SR
• VBD.insert: where the CDROM is local to a host
• VIF.create: where the network is not properly shared
• PIF.unplug: when the network would cease to be properly shared
• PBD.unplug: when the storage would cease to be properly shared
• Host.enable: when some network or storage would cease to be properly shared (e.g. if this host had a broken storage configuration)

xen

The Xen hypervisor has per-domain watchdog counters which, when enabled, decrement as time passes and can be reset from a hypercall from the domain. If the domain fails to make the hypercall and the timer reaches zero then the domain is immediately shutdown with reason reboot. We configure Xen to reboot the host when domain 0 enters this state.

High-level operations

Enabling HA

Before HA can be enabled the admin must take care to configure the environment properly. In particular:

• NIC bonds should be available for network heartbeats;
• multipath should be configured for the storage heartbeats;
• all hosts should be online and fully-booted.

The XenAPI client can request a specific shared SR to be used for storage heartbeats, otherwise Xapi will use the Pool’s default SR. Xapi will use VDI_GENERATE_CONFIG to ensure the disk will be attached automatically on system boot before the liveset has been joined.

Note that extra effort is made to re-use any existing heartbeat VDIS so that

• if HA is disabled with some hosts offline, when they are rebooted they stand a higher chance of seeing a well-formed statefile with an explicit invalid state. If the VDIs were destroyed on HA disable then hosts which boot up later would fail to attach the disk and it would be harder to distinguish between a temporary storage failure and a permanent HA disable.
• the heartbeat SR can be created on expensive low-latency high-reliability storage and made as small as possible (to minimise infrastructure cost), safe in the knowledge that if HA enables successfully once, it won’t run out of space and fail to enable in the future.

The Xapi-to-Xapi communication looks as follows:

The Xapi Pool master calls Host.ha_join_liveset on all hosts in the pool simultaneously. Each host runs the ha_start_daemon script which starts Xhad. Each Xhad starts exchanging heartbeats over the network and storage defined in the xhad.conf.

Joining a liveset

The Xhad instances exchange heartbeats and decide which hosts are in the “liveset” and which have been fenced.

After joining the liveset, each host clears the “excluded” flag which would have been set if the host had been shutdown cleanly before – this is only needed when a host is shutdown cleanly and then restarted.

Xapi periodically queries the state of xhad via the ha_query_liveset command. The state will be Starting until the liveset is fully formed at which point the state will be Online.

When the ha_start_daemon script returns then Xapi will decide whether to stand for master election or not. Initially when HA is being enabled and there is a master already, this node will be expected to stand unopposed. Later when HA notices that the master host has been fenced, all remaining hosts will stand for election and one of them will be chosen.

Shutting down a host

When a host is to be shutdown cleanly, it can be safely “excluded” from the pool such that a future failure of the storage heartbeat will not cause all pool hosts to self-fence (see survival rule 2 above). When a host is “excluded” all other hosts know that the host does not consider itself a master and has no resources locked i.e. no VMs are running on it. An excluded host will never allow itself to form part of a “split brain”.

Once a host has given up its master role and shutdown any VMs, it is safe to disable fencing with ha_disarm_fencing and stop xhad with ha_stop_daemon. Once the daemon has been stopped the “excluded” bit can be set in the statefile via ha_set_excluded and the host safely rebooted.

Restarting a host

When a host restarts after a failure Xapi notices that ha_armed is set in the local database. Xapi

• runs the attach-static-vdis script to attach the statefile and database VDIs. This can fail if the storage is inaccessible; Xapi will retry until it succeeds.
• runs the ha_start_daemon to join the liveset, or determine that HA has been cleanly disabled (via setting the state to Invalid).

In the special case where Xhad fails to access the statefile and the host used to be a slave then Xapi will try to contact the previous master and find out

• who the new master is;
• whether HA is enabled on the Pool or not.

If Xapi can confirm that HA was disabled then it will disarm itself and join the new master. Otherwise it will keep waiting for the statefile to recover.

In the special case where the statefile has been destroyed and cannot be recovered, there is an emergency HA disable API the admin can use to assert that HA really has been disabled, and it’s not simply a connectivity problem. Obviously this API should only be used if the admin is totally sure that HA has been disabled.

Disabling HA

There are 2 methods of disabling HA: one for the “normal” case when the statefile is available; and the other for the “emergency” case when the statefile has failed and can’t be recovered.

Disabling HA cleanly

HA can be shutdown cleanly when the statefile is working i.e. when hosts are alive because of survival rule 1. First the master Xapi tells the local Xhad to mark the pool state as “invalid” using ha_set_pool_state. Every xhad instance will notice this state change the next time it performs a storage heartbeat. The Xhad instances will shutdown and Xapi will notice that HA has been disabled the next time it attempts to query the liveset.

If a host loses access to the statefile (or if none of the hosts have access to the statefile) then HA can be disabled uncleanly.

Disabling HA uncleanly

The Xapi master first calls Host.ha_disable_failover_actions on each host which sets ha_disable_failover_decisions in the lcoal database. This prevents the node rebooting, gaining statefile access, acquiring the master lock and restarting VMs when other hosts have disabled their fencing (i.e. a “split brain”).

Once the master is sure that no host will suddenly start recovering VMs it is safe to call Host.ha_disarm_fencing which runs the script ha_disarm_fencing and then shuts down the Xhad with ha_stop_daemon.

Add a host to the pool

We assume that adding a host to the pool is an operation the admin will perform manually, so it is acceptable to disable HA for the duration and to re-enable it afterwards. If a failure happens during this operation then the admin will take care of it by hand.

NUMA

NUMA in a nutshell

Systems that contain more than one CPU socket are typically built on a Non-Uniform Memory Architecture (NUMA) ¹². In a NUMA system each node has fast, lower latency access to local memory.

In the diagram ³ above we have 4 NUMA nodes:

• 2 of those are due to 2 separate physical packages (sockets)
• a further 2 is due to Sub-NUMA-Clustering (aka Nodes Per Socket for AMD) where the L3 cache is split

The L3 cache is shared among multiple cores, but cores 0-5 have lower latency access to one part of it, than cores 6-11, and this is also reflected by splitting memory addresses into 4 31GiB ranges in total.

In the diagram the closer the memory is to the core, the lower the access latency:

• per-core caches: L1, L2
• per-package shared cache: L3 (local part), L3 (remote part)
• local NUMA node (to a group of cores, e.g. L#0 P#0), node 0
• remote NUMA node in same package (L#1 P#2), node 1
• remote NUMA node in other packages (L#2 P#1 and ‘L#3P#3’), node 2 and 3

The NUMA distance matrix

Accessing remote NUMA node in the other package has to go through a shared interconnect, which has lower bandwidth than the direct connections, and also a bottleneck if both cores have to access remote memory: the bandwidth for a single core is effectively at most half.

This is reflected in the NUMA distance/latency matrix. The units are arbitrary, and by convention access latency to the local NUMA node is given distance ‘10’.

Relative latency matrix by logical indexes:

This follows the latencies described previously:

• fast access to local NUMA node memory (by definition), node 0, cost 10
• slightly slower access latency to the other NUMA node in same package, node 1, cost 11
• twice as slow access latency to remote NUMA memory in the other physical package (socket): nodes 2 and 3, cost 21

There is also I/O NUMA where a cost is similarly associated to where a PCIe is plugged in, but exploring that is future work (it requires exposing NUMA topology to the Dom0 kernel to benefit from it), and for simplicity the diagram above does not show it.

Advantages of NUMA

NUMA does have advantages though: if each node accesses only its local memory, then each node can independently achieve maximum throughput.

For best performance we should:

• minimize the amount of interconnect bandwidth we are using
• run code that accesses memory allocated on the closest NUMA node
• maximize the number of NUMA nodes that we use in the system as a whole

If a VM’s memory and vCPUs can entirely fit within a single NUMA node then we should tell Xen to prefer to allocate memory from and run the vCPUs on a single NUMA node.

Xen vCPU soft-affinity

The Xen scheduler supports 2 kinds of constraints:

• hard pinning: a vCPU may only run on the specified set of pCPUs and nowhere else
• soft pinning: a vCPU is preferably run on the specified set of pCPUs, but if they are all busy then it may run elsewhere

The former is useful if you want strict separation, but it can potentially leave part of the system idle while another part is bottlenecked with lots of vCPUs all competing for the same limited set of pCPUs.

Xen does not migrate workloads between NUMA nodes on its own (the Linux kernel does), although it is possible to achieve a similar effect with explicit migration. However migration introduces additional delays and is best avoided for entire VMs.

The latter (soft pinning) is preferred: running a workload now, even on a potentially suboptimal pCPU (higher NUMA latency) is still better than not running it at all and waiting until a pCPU is freed up.

Xen will also allocate memory for the VM according to the vCPU (soft) pinning: if the vCPUs are pinned only to NUMA nodes A and B, then it will allocate the VM’s memory from NUMA nodes A and B (in a round-robin way, resulting in interleaving).

By default (no pinning) it will interleave memory from all NUMA nodes, which provides average performance, but individual tasks’ performance may be significantly higher or lower depending on which NUMA node the application may have “landed” on. Furthermore restarting processes will speed them up or slow them down as address space randomization picks different memory regions inside a VM.

Note that this is not the worst case: the worst case would be for memory to be allocated on one NUMA node, but the vCPU always running on the furthest away NUMA node.

Best effort NUMA-aware memory allocation for VMs

By default Xen stripes the VM’s memory accross all NUMA nodes of the host, which means that every VM has to go through all the interconnects. The goal here is to find a better allocation than the default, not necessarily an optimal allocation. An optimal allocation would require knowing what VMs you would start/create in the future, and planning across hosts too.

Overall we want to balance the VMs across NUMA nodes, such that we use all NUMA nodes to take advantage of the maximum memory bandwidth available on the system. For now this proposed balancing will be done only by balancing memory usage: always heuristically allocating VMs on the NUMA node that has the most available memory. Note that this allocation has a race condition for now when multiple VMs are booted in parallel, because we don’t wait until Xen has constructed the domain for each one (that’d serialize domain construction, which is currently parallel). This may be improved in the future by having an API to query Xen where it has allocated the memory, and to explicitly ask it to place memory on a given NUMA node (instead of best_effort).

If a VM doesn’t fit into a single node then it is not so clear what the best approach is. One criteria to consider is minimizing the NUMA distance between the nodes chosen for the VM. Large NUMA systems may not be fully connected in a mesh requiring multiple hops to each a node, or even have assymetric links, or links with different bitwidth. These tradeoff should be approximatively reflected in the ACPI SLIT tables, as a matrix of distances between nodes. It is possible that 3 NUMA nodes have a smaller average/maximum distance than 2, so we need to consider all possibilities.

For N nodes there would be 2^N possibilities, so [Topology.NUMA.candidates] limits the number of choices to 65520+N (full set of 2^N possibilities for 16 NUMA nodes, and a reduced set of choices for larger systems).

[Topology.NUMA.candidates] is a sorted sequence of node sets, in ascending order of maximum/average distances. Once we’ve eliminated the candidates not suitable for this VM (that do not have enough total memory/pCPUs) we are left with a monotonically increasing sequence of nodes. There are still multiple possibilities with same average distance. This is where we consider our second criteria - balancing - and pick the node with most available free memory.

Once a suitable set of NUMA nodes are picked we compute the CPU soft affinity as the union of the CPUs from all these NUMA nodes. If we didn’t find a solution then we let Xen use its default allocation.

The “distances” between NUMA nodes may not all be equal, e.g. some nodes may have shorter links to some remote NUMA nodes, while others may have to go through multiple hops to reach it. See page 13 in ⁴ for a diagram of an AMD Opteron 6272 system.

Limitations and tradeoffs

• Booting multiple VMs in parallel will result in potentially allocating both on the same NUMA node (race condition)
• When we’re about to run out of host memory we’ll fall back to striping memory again, but the soft affinity mask won’t reflect that (this needs an API to query Xen on where it has actually placed the VM, so we can fix up the mask accordingly)
• XAPI is not aware of NUMA balancing across a pool, and choses hosts purely based on total amount of free memory, even if a better NUMA placement could be found on another host
• Very large (>16 NUMA nodes) systems may only explore a limited number of choices (fit into a single node vs fallback to full interleaving)
• The exact VM placement is not yet controllable
• Microbenchmarks with a single VM on a host show both performance improvements and regressions on memory bandwidth usage: previously a single VM may have been able to take advantage of the bandwidth of both NUMA nodes if it happened to allocate memory from the right places, whereas now it’ll be forced to use just a single node. As soon as you have more than 1 VM that is busy on a system enabling NUMA balancing should almost always be an improvement though.
• it is not supported to combine hard vCPU masks with soft affinity: if hard affinities are used then no NUMA scheduling is done by the toolstack and we obey exactly what the user has asked for with hard affinities. This shouldn’t affect other VMs since the memory used by hard-pinned VMs will still be reflected in overall less memory available on individual NUMA nodes.
• Corner case: the ACPI standard allows certain NUMA nodes to be unreachable (distance 0xFF = -1 in the Xen bindings). This is not supported and will cause an exception to be raised. If this is an issue in practice the NUMA matrix could be pre-filtered to contain only reachable nodes. NUMA nodes with 0 CPUs are accepted (it can result from hard affinity pinnings)
• NUMA balancing is not considered during HA planning
• Dom0 is a single VM that needs to communicate with all other VMs, so NUMA balancing is not applied to it (we’d need to expose NUMA topology to the Dom0 kernel so it can better allocate processes)
• IO NUMA is out of scope for now

XAPI datamodel design

• New API field: Host.numa_affinity_policy.
• Choices: default_policy, any, best_effort.
• On upgrade the field is set to default_policy
• Changes in the field only affect newly (re)booted VMs, for changes to take effect on existing VMs a host evacuation or reboot is needed

There may be more choices in the future (e.g. strict, which requires both Xen and toolstack changes).

Meaning of the policy:

• any: the Xen default where it allocated memory by striping across NUMA nodes

• best_effort: the algorithm described in this document, where soft pinning is used to achieve better balancing and lower latency

• default_policy: when the admin hasn’t expressed a preference

• Currently default_policy is treated as any, but the admin can change it, and then the system will remember that change across upgrades. If we didn’t have a default_policy then changing the “default” policy on an upgrade would be tricky: we either risk overriding an explicit choice of the admin, or existing installs cannot take advantage of the improved performance from best_effort

• Future XAPI versions may change default_policy to mean best_effort. Admins can still override it to any if they wish on a host by host basis.

It is not expected that users would have to change best_effort, unless they run very specific workloads, so a pool level control is not provided at this moment.

There is also no separate feature flag: this host flag acts as a feature flag that can be set through the API without restarting the toolstack. Although obviously only new VMs will benefit.

Debugging the allocator is done by running xl vcpu-list and investigating the soft pinning masks, and by analyzing xensource.log.

Xenopsd implementation

See the documentation in [softaffinity.mli] and [topology.mli].

• [Softaffinity.plan] returns a [CPUSet] given a host’s NUMA allocation state and a VM’s NUMA allocation request.
• [Topology.CPUSet] provides helpers for operating on a set of CPU indexes.
• [Topology.NUMAResource] is a [CPUSet] and the free memory available on a NUMA node.
• [Topology.NUMARequest] is a request for a given number of vCPUs and memory in bytes.
• [Topology.NUMA] represents a host’s NUMA allocation state.
• [Topology.NUMA.candidates] are groups of nodes orderd by minimum average distance. The sequence is limited to [N+65520], where [N] is the number of NUMA nodes. This avoids exponential state space explosion on very large systems (>16 NUMA nodes).
• [Topology.NUMA.choose] will choose one NUMA node deterministically, while trying to keep overall NUMA node usage balanced.
• [Domain.numa_placement] builds a [NUMARequest] and uses the above [Topology] and [Softaffinity] functions to compute and apply a plan.

We used to have a xenopsd.conf configuration option to enable numa placement, for backwards compatibility this is still supported, but only if the admin hasn’t set an explicit policy on the Host. It is best to remove the experimental xenopsd.conf entry though, a future version may completely drop it.

Tests are in [test_topology.ml] which checks balancing properties and whether the plan has improved best/worst/average-case access times in a simulated test based on 2 predefined NUMA distance matrixes (one from Intel and one from an AMD system).

Future work

• enable ‘best_effort’ mode by default once more testing has been done
• an API to query Xen where it has actually allocated the VM’s memory. Currently only an xl debug-keys interface exists which is not supported in production as it can result in killing the host via the watchdog, and is not a proper API, but a textual debug output with no stability guarantees.
• more host policies (e.g. strict). Requires the XAPI pool scheduler to be NUMA aware and consider it as part of chosing hosts.
• VM level policy that can set a NUMA affinity index, mapped to a NUMA node modulo NUMA nodes available on the system (this is needed so that after migration we don’t end up trying to allocate vCPUs to a non-existent NUMA node)
• VM level anti-affinity rules for NUMA placement (can be achieved by setting unique NUMA affinity indexes)

Snapshots

Snapshots represent the state of a VM, or a disk (VDI) at a point in time. They can be used for:

• backups (hourly, daily, weekly etc)
• experiments (take snapshot, try something, revert back again)
• golden images (install OS, get it just right, clone it 1000s of times)

Read more about the Snapshot APIs.

Disk snapshots

Disks are represented in the XenAPI as VDI objects. Disk snapshots are represented as VDI objects with the flag is_a_snapshot set to true. Snapshots are always considered read-only, and should only be used for backup or cloning into new disks. Disk snapshots have a lifetime independent of the disk they are a snapshot of i.e. if someone deletes the original disk, the snapshots remain. This contrasts with some storage arrays in which snapshots are “second class” objects which are automatically deleted when the original disk is deleted.

Disks are implemented in Xapi via “Storage Manager” (SM) plugins. The SM plugins conform to an api (the SMAPI) which has operations including

• vdi_create: make a fresh disk, full of zeroes
• vdi_snapshot: create a snapshot of a disk

File-based vhd implementation

The existing “EXT” and “NFS” file-based Xapi SM plugins store disk data in trees of .vhd files as in the following diagram:

From the XenAPI point of view, we have one current VDI and a set of snapshots, each taken at a different point in time. These VDIs correspond to leaf vhds in a tree stored on disk, where the non-leaf nodes contain all the shared blocks.

The vhd files are always thinly-provisioned which means they only allocate new blocks on an as-needed basis. The snapshot leaf vhd files only contain vhd metadata and therefore are very small (a few KiB). The parent nodes containing the shared blocks only contain the shared blocks. The current leaf initially contains only the vhd metadata and therefore is very small (a few KiB) and will only grow when the VM writes blocks.

File-based vhd implementations are a good choice if a “gold image” snapshot is going to be cloned lots of times.

Block-based vhd implementation

The existing “LVM”, “LVMoISCSI” and “LVMoHBA” block-based Xapi SM plugins store disk data in trees of .vhd files contained within LVM logical volumes:

Non-snapshot VDIs are always stored full size (a.k.a. thickly-provisioned). When parent nodes are created they are automatically shrunk to the minimum size needed to store the shared blocks. The LVs corresponding with snapshot VDIs only contain vhd metadata and by default consume 8MiB. Note: this is different to VDI.clones which are stored full size.

Block-based vhd implementations are not a good choice if a “gold image” snapshot is going to be cloned lots of times, since each clone will be stored full size.

Hypothetical LUN implementation

A hypothetical Xapi SM plugin could use LUNs on an iSCSI storage array as VDIs, and the array’s custom control interface to implement the “snapshot” operation:

From the XenAPI point of view, we have one current VDI and a set of snapshots, each taken at a different point in time. These VDIs correspond to LUNs on the same iSCSI target, and internally within the target these LUNs are comprised of blocks from a large shared copy-on-write pool with support for dedup.

Reverting disk snapshots

There is no current way to revert in-place a disk to a snapshot, but it is possible to create a writable disk by “cloning” a snapshot.

VM snapshots

Let’s say we have a VM, “VM1” that has 2 disks. Concentrating only on the VM, VBDs and VDIs, we have the following structure:

When we take a snapshot, we first ask the storage backends to snapshot all of the VDIs associated with the VM, producing new VDI objects. Then we copy all of the metadata, producing a new ‘snapshot’ VM object, complete with its own VBDs copied from the original, but now pointing at the snapshot VDIs. We also copy the VIFs and VGPUs but for now we will ignore those.

This process leads to a set of objects that look like this:

We have fields that help navigate the new objects: VM.snapshot_of, and VDI.snapshot_of. These, like you would expect, point to the relevant other objects.

Deleting VM snapshots

When a snapshot is deleted Xapi calls the SM API vdi_delete. The Xapi SM plugins which use vhd format data do not reclaim space immediately; instead they mark the corresponding vhd leaf node as “hidden” and, at some point later, run a garbage collector process.

The garbage collector will first determine whether a “coalesce” should happen i.e. whether any parent nodes have only one child i.e. the “shared” blocks are only “shared” with one other node. In the following example the snapshot delete leaves such a parent node and the coalesce process copies blocks from the redundant parent’s only child into the parent:

Note that if the vhd data is being stored in LVM, then the parent node will have had to be expanded to full size to accommodate the writes. Unfortunately this means the act of reclaiming space actually consumes space itself, which means it is important to never completely run out of space in such an SR.

Once the blocks have been copied, we can now cut one of the parents out of the tree by relinking its children into their grandparent:

Finally the garbage collector can remove unused vhd files / LVM LVs:

Reverting VM snapshots

The XenAPI call VM.revert overwrites the VM metadata with the snapshot VM metadata, deletes the current VDIs and replaces them with clones of the snapshot VDIs. Note there is no “vdi_revert” in the SMAPI.

Revert implementation details

This is the process by which we revert a VM to a snapshot. The first thing to notice is that there is some logic that is called from message_forwarding.ml, which uses some low-level database magic to turn the current VM record into one that looks like the snapshot object. We then go to the rest of the implementation in xapi_vm_snapshot.ml. First, we shut down the VM if it is currently running. Then, we revert all of the VBDs, VIFs and VGPUs. To revert the VBDs, we need to deal with the VDIs underneath them. In order to create space, the first thing we do is delete all of the VDIs currently attached via VBDs to the VM. We then clone the disks from the snapshot. Note that there is no SMAPI operation ‘revert’ currently - we simply clone from the snapshot VDI. It’s important to note that cloning creates a new VDI object: this is not the one we started with gone.

vGPU

XenServer has supported passthrough for GPU devices since XenServer 6.0. Since the advent of NVIDIA’s vGPU-capable GRID K1/K2 cards it has been possible to carve up a GPU into smaller pieces yielding a more scalable solution to boosting graphics performance within virtual machines.

The K1 has four GK104 GPUs and the K2 two GK107 GPUs. Each of these will be exposed through Xapi so a host with a single K1 card will have access to four independent PGPUs.

Each of the GPUs can then be subdivided into vGPUs. For each type of PGPU, there are a few options of vGPU type which consume different amounts of the PGPU. For example, K1 and K2 cards can currently be configured in the following ways:

Note, this diagram is not to scale, the PGPU resource required by each vGPU type is as follows:

Currently each physical GPU (PGPU) only supports homogeneous vGPU configurations but different configurations are supported on different PGPUs across a single K1/K2 card. This means that, for example, a host with a K1 card can run 64 VMs with k100 vGPUs (8 per PGPU).

XenServer’s vGPU architecture

A new display type has been added to the device model:

@@ -4519,6 +4522,7 @@ static const QEMUOption qemu_options[] =

     /* Xen tree options: */
     { "std-vga", 0, QEMU_OPTION_std_vga },
+    { "vgpu", 0, QEMU_OPTION_vgpu },
     { "videoram", HAS_ARG, QEMU_OPTION_videoram },
     { "d", HAS_ARG, QEMU_OPTION_domid }, /* deprecated; for xend compatibility */
     { "domid", HAS_ARG, QEMU_OPTION_domid },

With this in place, qemu can now be started using a new option that will enable it to communicate with a new display emulator, vgpu to expose the graphics device to the guest. The vgpu binary is responsible for handling the VGX-capable GPU and, once it has been successfully passed through, the in-guest drivers can be installed in the same way as when it detects new hardware.

The diagram below shows the relevant parts of the architecture for this project.

Relevant code

• In Xenopsd: Xenops_server_xen is where Xenopsd gets the vGPU information from the values passed from Xapi;
• In Xenopsd: Device.__start is where the vgpu process is started, if necessary, before Qemu.

Xapi’s API and data model

A lot of work has gone into the toolstack to handle the creation and management of VMs with vGPUs. We revised our data model, introducing a semantic link between VGPU and PGPU objects to help with utilisation tracking; we maintained the GPU_group concept as a pool-wide abstraction of PGPUs available for VMs; and we added VGPU_types which are configurations for VGPU objects.

Aside: The VGPU type in Xapi’s data model predates this feature and was synonymous with GPU-passthrough. A VGPU is simply a display device assigned to a VM which may be a vGPU (this feature) or a whole GPU (a VGPU of type passthrough).

VGPU_types can be enabled/disabled on a per-PGPU basis allowing for reservation of particular PGPUs for certain workloads. VGPUs are allocated on PGPUs within their GPU group in either a depth-first or breadth-first manner, which is configurable on a per-group basis.

VGPU_types are created by xapi at startup depending on the available hardware and config files present in dom0. They exist in the pool database, and a primary key is used to avoid duplication. In XenServer 6.x the tuple of (vendor_name, model_name) was used as the primary key, however this was not ideal as these values are subject to change. XenServer 7.0 switched to a new primary key generated from static metadata, falling back to the old method for backwards compatibility.

A VGPU_type will be garbage collected when there is no VGPU of that type and there is no hardware which supports that type. On VM import, all VGPUs and VGPU_types will be created if necessary - if this results in the creation of a new VGPU_type then the VM will not be usable until the required hardware and drivers are installed.

Relevant code

• In Xapi: Xapi_vgpu_type contains the type definitions and parsing logic for vGPUs;
• In Xapi: Xapi_pgpu_helpers defines the functions used to allocate vGPUs on PGPUs.

Xapi <-> Xenopsd interface

In XenServer 6.x, all VGPU config was added to the VM’s platform field at startup, and this information was used by xenopsd to start the display emulator. See the relevant code here.

In XenServer 7.0, to facilitate support of VGPU on Intel hardware in parallel with the existing NVIDIA support, VGPUs were made first-class objects in the xapi-xenopsd interface. The interface is described here.

VM startup

On the pool master:

• Assuming no WLB, all VM.start tasks pass through Xapi_vm_helpers.choose_host_for_vm_no_wlb. If the VM has a vGPU, the list of all hosts in the pool is split into a list of lists, where the first list is the most optimal in terms of the GPU group’s allocation mode and the PGPU availability on each host.
• Each list of hosts in turn is passed to Xapi_vm_placement.select_host, which checks storage, network and memory availability, until a suitable host is found.
• Once a host has been chosen, allocate_vm_to_host will set the VM.scheduled_to_be_resident_on and VGPU.scheduled_to_be_resident_on fields.

The task is then ready to be forwarded to the host on which the VM will start:

• If the VM has a VGPU, the startup task is wrapped in Xapi_gpumon.with_gpumon_stopped. This makes sure that the NVIDIA driver is not in use so can be loaded or unloaded from physical GPUs as required.
• The VM metadata, including VGPU metadata, is passed to xenopsd. The creation of the VGPU metadata is done by vgpus_of_vm. Note that at this point passthrough VGPUs are represented by the PCI device type, and metadata is generated by pcis_of_vm.
• As part of starting up the VM, xenopsd should report a VGPU event or a PCI event, which xapi will use to indicate that the xapi VGPU object can be marked as currently_attached.

Usage

To create a VGPU of a given type you can use vgpu-create:

$ xe vgpu-create vm-uuid=... gpu-group-uuid=... vgpu-type-uuid=...

To see a list of VGPU types available for use on your XenServer, run the following command. Note: these will only be populated if you have installed the relevant NVIDIA RPMs and if there is hardware installed on that host supported each type. Using params=all will display more information such as the maximum number of heads supported by that VGPU type and which PGPUs have this type enabled and supported.

$ xe vgpu-type-list [params=all]

To access the new and relevant parameters on a PGPU (i.e. supported_VGPU_types, enabled_VGPU_types, resident_VGPUs) you can use pgpu-param-get with param-name=supported-vgpu-types param-name=enabled-vgpu-types and param-name=resident-vgpus respectively. Or, alternatively, you can use the following command to list all the parameters for the PGPU. You can get the types supported or enabled for a given PGPU:

$ xe pgpu-list uuid=... params=all

Xapi Storage Migration

The Xapi Storage Migration (XSM) also known as “Storage Motion” allows

• a running VM to be migrated within a pool, between different hosts and different storage simultaneously;
• a running VM to be migrated to another pool;
• a disk attached to a running VM to be moved to another SR.

The following diagram shows how XSM works at a high level:

The slowest part of a storage migration is migrating the storage, since virtual disks can be very large. Xapi starts by taking a snapshot and copying that to the destination as a background task. Before the datapath connecting the VM to the disk is re-established, xapi tells tapdisk to start mirroring all writes to a remote tapdisk over NBD. From this point on all VM disk writes are written to both the old and the new disk. When the background snapshot copy is complete, xapi can migrate the VM memory across. Once the VM memory image has been received, the destination VM is complete and the original can be safely destroyed.

How to add....

Metadata-on-LUN

In the present version of XenServer, metadata changes resulting in writes to the database are not persisted in non-volatile storage. Hence, in case of failure, up to five minutes’ worth of metadata changes could be lost. The Metadata-on-LUN feature addresses the issue by ensuring that all database writes are retained. This will be used to improve recovery from failure by storing incremental deltas which can be re-applied to an old version of the database to bring it more up-to-date. An implication of this is that clients will no longer be required to perform a ‘pool-sync-database’ to protect critical writes, because all writes will be implicitly protected.

This is implemented by saving descriptions of all persistent database writes to a LUN when HA is active. Upon xapi restart after failure, such as on master fail-over, these descriptions are read and parsed to restore the latest version of the database.

Layout on block device

It is useful to store the database on the block device as well as the deltas, so that it is unambiguous on recovery which version of the database the deltas apply to.

The content of the block device will be structured as shown in the table below. It consists of a header; the rest of the device is split into two halves.

After the header, one or both halves may be devoid of content. In a half which contains a database, there may be zero or more deltas (repetitions of the last three entries in each half).

The structure of the device is split into two halves to provide double-buffering. In case of failure during write to one half, the other half remains intact.

The magic identifier at the start of the file protect against attempting to treat a different device as a redo log.

The validity byte is a single `ascii character indicating the state of the two halves. It can take the following values:

The use of lengths preceding data sections permit convenient reading. The constant repetitions of the UUIDs act as nonces to protect against reading in invalid data in the case of an incomplete or corrupt write.

Architecture

The I/O to and from the block device may involve long delays. For example, if there is a network problem, or the iSCSI device disappears, the I/O calls may block indefinitely. It is important to isolate this from xapi. Hence, I/O with the block device will occur in a separate process.

Xapi will communicate with the I/O process via a UNIX domain socket using a simple text-based protocol described below. The I/O process will use to ensure that it can always accept xapi’s requests with a guaranteed upper limit on the delay. Xapi can therefore communicate with the process using blocking I/O.

Xapi will interact with the I/O process in a best-effort fashion. If it cannot communicate with the process, or the process indicates that it has not carried out the requested command, xapi will continue execution regardless. Redo-log entries are idempotent (modulo the raising of exceptions in some cases) so it is of little consequence if a particular entry cannot be written but others can. If xapi notices that the process has died, it will attempt to restart it.

The I/O process keeps track of a pointer for each half indicating the position at which the next delta will be written in that half.

Protocol

Upon connection to the control socket, the I/O process will attempt to connect to the block device. Depending on whether this is successful or unsuccessful, one of two responses will be sent to the client.

• connect|ack_ if it is successful; or

• connect|nack|<length>|<message> if it is unsuccessful, perhaps because the block device does not exist or cannot be read from. The <message> is a description of the error; the <length> of the message is expressed using 16 digits of decimal ascii.

The former message indicates that the I/O process is ready to receive commands. The latter message indicates that commands can not be sent to the I/O process.

There are three commands which xapi can send to the I/O process. These are described below, with a high level description of the operational semantics of the I/O process’ actions, and the corresponding responses. For ease of parsing, each command is ten bytes in length.

Write database

Xapi requests that a new database is written to the block device, and sends its content using the data socket.

Command:

: writedb___|<uuid>|<generation-count>|<length>

The UUID is expressed as 36 ASCII characters. The length of the data and the generation-count are expressed using 16 digits of decimal ASCII.

Semantics:

1. Read the validity byte.
2. If one half is valid, we will use the other half. If no halves are valid, we will use the first half.
3. Read the data from the data socket and write it into the chosen half.
4. Set the pointer for the chosen half to point to the position after the data.
5. Set the validity byte to indicate the chosen half is valid.

Response:

: writedb|ack_ in case of successful write; or

writedb|nack|<length>|<message> otherwise.

For error messages, the length of the message is expressed using 16 digits of decimal ascii. In particular, the error message for timeouts is the string Timeout.

Write database delta

Xapi sends a description of a database delta to append to the block device.

Command:

: writedelta|<uuid>|<generation-count>|<length>|<data>

The UUID is expressed as 36 ASCII characters. The length of the data and the generation-count are expressed using 16 digits of decimal ASCII.

Semantics:

1. Read the validity byte to establish which half is valid. If neither half is valid, return with a nack.
2. If the half’s pointer is set, seek to that position. Otherwise, scan through the half and stop at the position after the last write.
3. Write the entry.
4. Update the half’s pointer to point to the position after the entry.

Response:

: writedelta|ack_ in case of successful append; or

writedelta|nack|<length>|<message> otherwise.

For error messages, the length of the message is expressed using 16 digits of decimal ASCII. In particular, the error message for timeouts is the string Timeout.

Read log

Xapi requests the contents of the log.

Command:

: read______

Semantics:

1. Read the validity byte to establish which half is valid. If neither half is valid, return with an end.
2. Attempt to read the database from the current half.
3. If this is successful, continue in that half reading entries up to the position of the half’s pointer. If the pointer is not set, read until a record of length zero is found or the end of the half is reached. Otherwise—if the attempt to the read the database was not successful—switch to using the other half and try again from step 2.
4. Finally output an end.

Response:

: read|nack_|<length>|<message> in case of error; or

read|db___|<generation-count>|<length>|<data> for a database record, then a sequence of zero or more

read|delta|<generation-count>|<length>|<data> for each delta record, then

read|end__

For each record, and for error messages, the length of the data or message is expressed using 16 digits of decimal ascii. In particular, the error message for timeouts is the string Timeout.

Re-initialise log

Xapi requests that the block device is re-initialised with a fresh redo-log.

Command:

: empty_____\

Semantics:

: 1. Set the validity byte to indicate that neither half is valid.

Response:

: empty|ack_ in case of successful re-initialisation; or

empty|nack|<length>|<message> otherwise.

For error messages, the length of the message is expressed using 16 digits of decimal ASCII. In particular, the error message for timeouts is the string Timeout.

Impact on xapi performance

The implementation of the feature causes a slow-down in xapi of around 6% in the general case. However, if the LUN becomes inaccessible this can cause a slow-down of up to 25% in the worst case.

The figure below shows the result of testing four configurations, counting the number of database writes effected through a command-line ‘xe pool-param-set’ call.

• The first and second configurations are xapi without the Metadata-on-LUN feature, with HA disabled and enabled respectively.

• The third configuration shows xapi with the Metadata-on-LUN feature using a healthy LUN to which all database writes can be successfully flushed.

• The fourth configuration shows xapi with the Metadata-on-LUN feature using an inaccessible LUN for which all database writes fail.

Testing strategy

The section above shows how xapi performance is affected by this feature. The sections below describe the dev-testing which has already been undertaken, and propose how this feature will impact on regression testing.

Dev-testing performed

A variety of informal tests have been performed as part of the development process:

Enable HA.

Confirm LUN starts being used to persist database writes.

Enable HA, disable HA.

Confirm LUN stops being used.

Enable HA, kill xapi on master, restart xapi on master.

Confirm that last database write before kill is successfully restored on restart.

Repeatedly enable and disable HA.

Confirm that no file descriptors are leaked (verified by counting the number of descriptors in /proc/pid/fd/).

Enable HA, reboot the master.

Due to HA, a slave becomes the master (or this can be forced using ‘xe pool-emergency-transition-to-master’). Confirm that the new master starts is able to restore the database from the LUN from the point the old master left off, and begins to write new changes to the LUN.

Enable HA, disable the iSCSI volume.

Confirm that xapi continues to make progress, although database writes are not persisted.

Enable HA, disable and enable the iSCSI volume.

Confirm that xapi begins to use the LUN when the iSCSI volume is re-enabled and subsequent writes are persisted.

These tests have been undertaken using an iSCSI target VM and a real iSCSI volume on lannik. In these scenarios, disabling the iSCSI volume consists of stopping the VM and unmapping the LUN, respectively.

Proposed new regression test

A new regression test is proposed to confirm that all database writes are persisted across failure.

There are three types of database modification to test: row creation, field-write and row deletion. Although these three kinds of write could be tested in separate tests, the means of setting up the pre-conditions for a field-write and a row deletion require a row creation, so it is convenient to test them all in a single test.

1. Start a pool containing three hosts.

2. Issue a CLI command on the master to create a row in the database, e.g.

   xe network-create name-label=a.

3. Forcefully power-cycle the master.

4. On fail-over, issue a CLI command on the new master to check that the row creation persisted:

   xe network-list name-label=a,

   confirming that the returned string is non-empty.

5. Issue a CLI command on the master to modify a field in the new row in the database:

   xe network-param-set uuid=<uuid> name-description=abcd,

   where <uuid> is the UUID returned from step 2.

6. Forcefully power-cycle the master.

7. On fail-over, issue a CLI command on the new master to check that the field-write persisted:

   xe network-param-get uuid=<uuid> param-name=name-description,

   where <uuid> is the UUID returned from step 2. The returned string should contain

   abcd.

8. Issue a CLI command on the master to delete the row from the database:

   xe network-destroy uuid=<uuid>,

   where <uuid> is the UUID returned from step 2.

9. Forcefully power-cycle the master.

10. On fail-over, issue a CLI command on the new master to check that the row does not exist:

    xe network-list name-label=a,

    confirming that the returned string is empty.

Impact on existing regression tests

The Metadata-on-LUN feature should mean that there is no need to perform an ‘xe pool-sync-database’ operation in existing HA regression tests to ensure that database state persists on xapi failure.

From RPC migration request to xapi internals

Overview

In this document we will use the VM.pool_migrate request to illustrate the interaction between various components within the XAPI toolstack during migration. However this schema can be applied to other requests as well.

Not all parts of the Xapi toolstack are shown here as not all are involved in the migration process. For instance you won’t see the squeezed nor mpathalert two daemons that belong to the toolstack but don’t participate in the migration of a VM.

Anatomy of a VM migration

• Migration is initiated by a Xapi client that sends VM.pool_migrate, an RPC XML request.
• The Xen API server handles this request and dispatches it to the server.
• The server is generated using XAPI IDL and requests are wrapped whithin a context, either to be forwarded to a host or executed locally. Broadly, the context follows RBAC rules. The executed function is related to the message of the request (refer to XenAPI Reference).
• In the case of the migration you can refer to ocaml/idl/datamodel_vm.ml.
• The server will dispatch the operation to server helpers, executing the operation synchronously or asynchronously and returning the RPC answer.
• Message forwarding decides if operation must be executed by another host of the pool and then forward the call or if is executed locally.
• When executed locally the high-level migration operation is send to the Xenopsd daemon by posting a message on a known queue on the message switch.
• Xenopsd will get the command and will split it into several atomic operations that will be run by the xenopsd backend.
• Xenopsd with its backend can then access xenstore or execute hypercall to interact with xen a server the micro operation.

A diagram is worth a thousand words

Storage migration

Overview

Receiving SXM

These are the remote calls in the above diagram sent from the remote host to the receiving end of storage motion:

• Remote SMAPIv2 -> local SMAPIv2 RPC calls:
  • SR.list
  • SR.scan
  • SR.update_snapshot_info_dest
  • VDI.attach2
  • VDI.activate
  • VDI.snapshot
  • VDI.destroy
  • For copying:
    • For copying from base:
      • VDI.clone
      • VDI.resize
    • For copying without base:
      • VDI.create
  • For mirroring:
    • DATA.MIRROR.receive_start
    • VDI.compose
    • VDI.remove_from_sm_config
    • DATA.MIRROR.receive_finalize
• HTTP requests to xapi:
  • Connecting to NBD URI via xapi’s HTTP handler

This is how xapi coordinates storage migration. We’ll do it as a code walkthrough through the two layers: xapi and storage-in-xapi (SMAPIv2).

Xapi code

The entry point is in xapi_vm_migration.ml

The function takes several arguments:

• a vm reference (vm)
• a dictionary of (string * string) key-value pairs about the destination (dest). This is the result of a previous call to the destination pool, Host.migrate_receive
• live, a boolean of whether we should live-migrate or suspend-resume,
• vdi_map, a mapping of VDI references to destination SR references,
• vif_map, a mapping of VIF references to destination network references,
• vgpu_map, similar for VGPUs
• options, another dictionary of options

let migrate_send'  ~__context ~vm ~dest ~live ~vdi_map ~vif_map ~vgpu_map ~options =
  SMPERF.debug "vm.migrate_send called vm:%s" (Db.VM.get_uuid ~__context ~self:vm);

  let open Xapi_xenops in

  let localhost = Helpers.get_localhost ~__context in
  let remote = remote_of_dest dest in

  (* Copy mode means we don't destroy the VM on the source host. We also don't
     	   copy over the RRDs/messages *)
  let copy = try bool_of_string (List.assoc "copy" options) with _ -> false in

It begins by getting the local host reference, deciding whether we’re copying or moving, and converting the input dest parameter from an untyped string association list to a typed record, remote, which is declared further up the file:

type remote = {
  rpc : Rpc.call -> Rpc.response;
  session : API.ref_session;
  sm_url : string;
  xenops_url : string;
  master_url : string;
  remote_ip : string; (* IP address *)
  remote_master_ip : string; (* IP address *)
  dest_host : API.ref_host;
}

this contains:

• A function, rpc, for calling XenAPI RPCs on the destination
• A session valid on the destination
• A sm_url on which SMAPIv2 APIs can be called on the destination
• A master_url on which XenAPI commands can be called (not currently used)
• The IP address, remote_ip, of the destination host
• The IP address, remote_master_ip, of the master of the destination pool

Next, we determine which VDIs to copy:

  (* The first thing to do is to create mirrors of all the disks on the remote.
     We look through the VM's VBDs and all of those of the snapshots. We then
     compile a list of all of the associated VDIs, whether we mirror them or not
     (mirroring means we believe the VDI to be active and new writes should be
     mirrored to the destination - otherwise we just copy it)
     We look at the VDIs of the VM, the VDIs of all of the snapshots, and any
     suspend-image VDIs. *)

  let vm_uuid = Db.VM.get_uuid ~__context ~self:vm in
  let vbds = Db.VM.get_VBDs ~__context ~self:vm in
  let vifs = Db.VM.get_VIFs ~__context ~self:vm in
  let snapshots = Db.VM.get_snapshots ~__context ~self:vm in
  let vm_and_snapshots = vm :: snapshots in
  let snapshots_vbds = List.flatten (List.map (fun self -> Db.VM.get_VBDs ~__context ~self) snapshots) in
  let snapshot_vifs = List.flatten (List.map (fun self -> Db.VM.get_VIFs ~__context ~self) snapshots) in

we now decide whether we’re intra-pool or not, and if we’re intra-pool whether we’re migrating onto the same host (localhost migrate). Intra-pool is decided by trying to do a lookup of our current host uuid on the destination pool.

  let is_intra_pool = try ignore(Db.Host.get_uuid ~__context ~self:remote.dest_host); true with _ -> false in
  let is_same_host = is_intra_pool && remote.dest_host == localhost in

  if copy && is_intra_pool then raise (Api_errors.Server_error(Api_errors.operation_not_allowed, [ "Copy mode is disallowed on intra pool storage migration, try efficient alternatives e.g. VM.copy/clone."]));

Having got all of the VBDs of the VM, we now need to find the associated VDIs, filtering out empty CDs, and decide whether we’re going to copy them or mirror them - read-only VDIs can be copied but RW VDIs must be mirrored.

  let vms_vdis = List.filter_map (vdi_filter __context true) vbds in

where vdi_filter is defined earler:

(* We ignore empty or CD VBDs - nothing to do there. Possible redundancy here:
   I don't think any VBDs other than CD VBDs can be 'empty' *)
let vdi_filter __context allow_mirror vbd =
  if Db.VBD.get_empty ~__context ~self:vbd || Db.VBD.get_type ~__context ~self:vbd = `CD
  then None
  else
    let do_mirror = allow_mirror && (Db.VBD.get_mode ~__context ~self:vbd = `RW) in
    let vm = Db.VBD.get_VM ~__context ~self:vbd in
    let vdi = Db.VBD.get_VDI ~__context ~self:vbd in
    Some (get_vdi_mirror __context vm vdi do_mirror)

This in turn calls get_vdi_mirror which gathers together some important info:

let get_vdi_mirror __context vm vdi do_mirror =
  let snapshot_of = Db.VDI.get_snapshot_of ~__context ~self:vdi in
  let size = Db.VDI.get_virtual_size ~__context ~self:vdi in
  let xenops_locator = Xapi_xenops.xenops_vdi_locator ~__context ~self:vdi in
  let location = Db.VDI.get_location ~__context ~self:vdi in
  let dp = Storage_access.presentative_datapath_of_vbd ~__context ~vm ~vdi in
  let sr = Db.SR.get_uuid ~__context ~self:(Db.VDI.get_SR ~__context ~self:vdi) in
  {vdi; dp; location; sr; xenops_locator; size; snapshot_of; do_mirror}

The record is helpfully commented above:

type vdi_mirror = {
  vdi : [ `VDI ] API.Ref.t;           (* The API reference of the local VDI *)
  dp : string;                        (* The datapath the VDI will be using if the VM is running *)
  location : string;                  (* The location of the VDI in the current SR *)
  sr : string;                        (* The VDI's current SR uuid *)
  xenops_locator : string;            (* The 'locator' xenops uses to refer to the VDI on the current host *)
  size : Int64.t;                     (* Size of the VDI *)
  snapshot_of : [ `VDI ] API.Ref.t;   (* API's snapshot_of reference *)
  do_mirror : bool;                   (* Whether we should mirror or just copy the VDI *)
}

xenops_locator is <sr uuid>/<vdi uuid>, and dp is vbd/<domid>/<device> if the VM is running and vbd/<vm_uuid>/<vdi_uuid> if not.

So now we have a list of these records for all VDIs attached to the VM. For these we check explicitly that they’re all defined in the vdi_map, the mapping of VDI references to their destination SR references.

  check_vdi_map ~__context vms_vdis vdi_map;

We then figure out the VIF map:

 let vif_map =
    if is_intra_pool then vif_map
    else infer_vif_map ~__context (vifs @ snapshot_vifs) vif_map
  in

More sanity checks: We can’t do a storage migration if any of the VDIs is a reset-on-boot one - since the state will be lost on the destination when it’s attached:

(* Block SXM when VM has a VDI with on_boot=reset *)
  List.(iter (fun vconf ->
      let vdi = vconf.vdi in
      if (Db.VDI.get_on_boot ~__context ~self:vdi ==`reset) then
        raise (Api_errors.Server_error(Api_errors.vdi_on_boot_mode_incompatible_with_operation, [Ref.string_of vdi]))) vms_vdis) ;

We now consider all of the VDIs associated with the snapshots. As for the VM’s VBDs above, we end up with a vdi_mirror list. Note we pass false to the allow_mirror parameter of the get_vdi_mirror function as none of these snapshot VDIs will ever require mirrorring.

let snapshots_vdis = List.filter_map (vdi_filter __context false)

Finally we get all of the suspend-image VDIs from all snapshots as well as the actual VM, since it might be suspended itself:

snapshots_vbds in
  let suspends_vdis =
    List.fold_left
      (fun acc vm ->
         if Db.VM.get_power_state ~__context ~self:vm = `Suspended
         then
           let vdi = Db.VM.get_suspend_VDI ~__context ~self:vm in
           let sr = Db.VDI.get_SR ~__context ~self:vdi in
           if is_intra_pool && Helpers.host_has_pbd_for_sr ~__context ~host:remote.dest_host ~sr
           then acc
           else (get_vdi_mirror __context vm vdi false):: acc
         else acc)
      [] vm_and_snapshots in

Sanity check that we can see all of the suspend-image VDIs on this host:

 (* Double check that all of the suspend VDIs are all visible on the source *)
  List.iter (fun vdi_mirror ->
      let sr = Db.VDI.get_SR ~__context ~self:vdi_mirror.vdi in
      if not (Helpers.host_has_pbd_for_sr ~__context ~host:localhost ~sr)
      then raise (Api_errors.Server_error (Api_errors.suspend_image_not_accessible, [ Ref.string_of vdi_mirror.vdi ]))) suspends_vdis;

Next is a fairly complex piece that determines the destination SR for all of these VDIs. We don’t require API uses to decide destinations for all of the VDIs on snapshots and hence we have to make some decisions here:

  let dest_pool = List.hd (XenAPI.Pool.get_all remote.rpc remote.session) in
  let default_sr_ref =
    XenAPI.Pool.get_default_SR remote.rpc remote.session dest_pool in
  let suspend_sr_ref =
    let pool_suspend_SR = XenAPI.Pool.get_suspend_image_SR remote.rpc remote.session dest_pool
    and host_suspend_SR = XenAPI.Host.get_suspend_image_sr remote.rpc remote.session remote.dest_host in
    if pool_suspend_SR <> Ref.null then pool_suspend_SR else host_suspend_SR in

  (* Resolve placement of unspecified VDIs here - unspecified VDIs that
            are 'snapshot_of' a specified VDI go to the same place. suspend VDIs
            that are unspecified go to the suspend_sr_ref defined above *)

  let extra_vdis = suspends_vdis @ snapshots_vdis in

  let extra_vdi_map =
    List.map
      (fun vconf ->
         let dest_sr_ref =
           let is_mapped = List.mem_assoc vconf.vdi vdi_map
           and snapshot_of_is_mapped = List.mem_assoc vconf.snapshot_of vdi_map
           and is_suspend_vdi = List.mem vconf suspends_vdis
           and remote_has_suspend_sr = suspend_sr_ref <> Ref.null
           and remote_has_default_sr = default_sr_ref <> Ref.null in
           let log_prefix =
             Printf.sprintf "Resolving VDI->SR map for VDI %s:" (Db.VDI.get_uuid ~__context ~self:vconf.vdi) in
           if is_mapped then begin
             debug "%s VDI has been specified in the map" log_prefix;
             List.assoc vconf.vdi vdi_map
           end else if snapshot_of_is_mapped then begin
             debug "%s Snapshot VDI has entry in map for it's snapshot_of link" log_prefix;
             List.assoc vconf.snapshot_of vdi_map
           end else if is_suspend_vdi && remote_has_suspend_sr then begin
             debug "%s Mapping suspend VDI to remote suspend SR" log_prefix;
             suspend_sr_ref
           end else if is_suspend_vdi && remote_has_default_sr then begin
             debug "%s Remote suspend SR not set, mapping suspend VDI to remote default SR" log_prefix;
             default_sr_ref
           end else if remote_has_default_sr then begin
             debug "%s Mapping unspecified VDI to remote default SR" log_prefix;
             default_sr_ref
           end else begin
             error "%s VDI not in VDI->SR map and no remote default SR is set" log_prefix;
             raise (Api_errors.Server_error(Api_errors.vdi_not_in_map, [ Ref.string_of vconf.vdi ]))
           end in
         (vconf.vdi, dest_sr_ref))
      extra_vdis in

At the end of this we’ve got all of the VDIs that need to be copied and destinations for all of them:

  let vdi_map = vdi_map @ extra_vdi_map in
  let all_vdis = vms_vdis @ extra_vdis in

  (* The vdi_map should be complete at this point - it should include all the
     VDIs in the all_vdis list. *)

Now we gather some final information together:

  assert_no_cbt_enabled_vdi_migrated ~__context ~vdi_map;

  let dbg = Context.string_of_task __context in
  let open Xapi_xenops_queue in
  let queue_name = queue_of_vm ~__context ~self:vm in
  let module XenopsAPI = (val make_client queue_name : XENOPS) in

  let remote_vdis = ref [] in

  let ha_always_run_reset = not is_intra_pool && Db.VM.get_ha_always_run ~__context ~self:vm in

  let cd_vbds = find_cds_to_eject __context vdi_map vbds in
  eject_cds __context cd_vbds;

check there’s no CBT (we can’t currently migrate the CBT metadata), make our client to talk to Xenopsd, make a mutable list of remote VDIs (which I think is redundant right now), decide whether we need to do anything for HA (we disable HA protection for this VM on the destination until it’s fully migrated) and eject any CDs from the VM.

Up until now this has mostly been gathering info (aside from the ejecting CDs bit), but now we’ll start to do some actions, so we begin a try-catch block:

try

but we’ve still got a bit of thinking to do: we sort the VDIs to copy based on age/size:

    (* Sort VDIs by size in principle and then age secondly. This gives better
       chances that similar but smaller VDIs would arrive comparatively
       earlier, which can serve as base for incremental copying the larger
       ones. *)
    let compare_fun v1 v2 =
      let r = Int64.compare v1.size v2.size in
      if r = 0 then
        let t1 = Date.to_unix_time (Db.VDI.get_snapshot_time ~__context ~self:v1.vdi) in
        let t2 = Date.to_unix_time (Db.VDI.get_snapshot_time ~__context ~self:v2.vdi) in
        compare t1 t2
      else r in
    let all_vdis = all_vdis |> List.sort compare_fun in

    let total_size = List.fold_left (fun acc vconf -> Int64.add acc vconf.size) 0L all_vdis in
    let so_far = ref 0L in

OK, let’s copy/mirror:

    with_many (vdi_copy_fun __context dbg vdi_map remote is_intra_pool remote_vdis so_far total_size copy) all_vdis @@ fun all_map ->

The copy functions are written such that they take continuations. This it to make the error handling simpler - each individual component function can perform its setup and execute the continuation. In the event of an exception coming from the continuation it can then unroll its bit of state and rethrow the exception for the next layer to handle.

with_many is a simple helper function for nesting invocations of functions that take continuations. It has the delightful type:

('a -> ('b -> 'c) -> 'c) -> 'a list -> ('b list -> 'c) -> 'c

(* Helper function to apply a 'with_x' function to a list *)
let rec with_many withfn many fn =
  let rec inner l acc =
    match l with
    | [] -> fn acc
    | x::xs -> withfn x (fun y -> inner xs (y::acc))
  in inner many []

As an example of its operation, imagine our withfn is as follows:

let withfn x c =
  Printf.printf "Starting withfn: x=%d\n" x;
  try
    c (string_of_int x)
  with e ->
    Printf.printf "Handling exception for x=%d\n" x;
    raise e;;

applying this gives the output:

utop # with_many withfn [1;2;3;4] (String.concat ",");;
Starting with fn: x=1
Starting with fn: x=2
Starting with fn: x=3
Starting with fn: x=4
- : string = "4,3,2,1"

whereas raising an exception in the continutation results in the following:

utop # with_many with_fn [1;2;3;4] (fun _ -> failwith "error");;
Starting with fn: x=1
Starting with fn: x=2
Starting with fn: x=3
Starting with fn: x=4
Handling exception for x=4
Handling exception for x=3
Handling exception for x=2
Handling exception for x=1
Exception: Failure "error".

All the real action is in vdi_copy_fun, which copies or mirrors a single VDI:

let vdi_copy_fun __context dbg vdi_map remote is_intra_pool remote_vdis so_far total_size copy vconf continuation =
  TaskHelper.exn_if_cancelling ~__context;
  let open Storage_access in
  let dest_sr_ref = List.assoc vconf.vdi vdi_map in
  let dest_sr_uuid = XenAPI.SR.get_uuid remote.rpc remote.session dest_sr_ref in

  (* Plug the destination shared SR into destination host and pool master if unplugged.
     Plug the local SR into destination host only if unplugged *)
  let dest_pool = List.hd (XenAPI.Pool.get_all remote.rpc remote.session) in
  let master_host = XenAPI.Pool.get_master remote.rpc remote.session dest_pool in
  let pbds = XenAPI.SR.get_PBDs remote.rpc remote.session dest_sr_ref in
  let pbd_host_pair = List.map (fun pbd -> (pbd, XenAPI.PBD.get_host remote.rpc remote.session pbd)) pbds in
  let hosts_to_be_attached = [master_host; remote.dest_host] in
  let pbds_to_be_plugged = List.filter (fun (_, host) ->
      (List.mem host hosts_to_be_attached) && (XenAPI.Host.get_enabled remote.rpc remote.session host)) pbd_host_pair in
  List.iter (fun (pbd, _) ->
      if not (XenAPI.PBD.get_currently_attached remote.rpc remote.session pbd) then
        XenAPI.PBD.plug remote.rpc remote.session pbd) pbds_to_be_plugged;

It begins by attempting to ensure the SRs we require are definitely attached on the destination host and on the destination pool master.

There’s now a little logic to support the case where we have cross-pool SRs and the VDI is already visible to the destination pool. Since this is outside our normal support envelope there is a key in xapi_globs that has to be set (via xapi.conf) to enable this:

  let rec dest_vdi_exists_on_sr vdi_uuid sr_ref retry =
    try
      let dest_vdi_ref = XenAPI.VDI.get_by_uuid remote.rpc remote.session vdi_uuid in
      let dest_vdi_sr_ref = XenAPI.VDI.get_SR remote.rpc remote.session dest_vdi_ref in
      if dest_vdi_sr_ref = sr_ref then
        true
      else
        false
    with _ ->
      if retry then
        begin
          XenAPI.SR.scan remote.rpc remote.session sr_ref;
          dest_vdi_exists_on_sr vdi_uuid sr_ref false
        end
      else
        false
  in

  (* CP-4498 added an unsupported mode to use cross-pool shared SRs - the initial
     use case is for a shared raw iSCSI SR (same uuid, same VDI uuid) *)
  let vdi_uuid = Db.VDI.get_uuid ~__context ~self:vconf.vdi in
  let mirror = if !Xapi_globs.relax_xsm_sr_check then
      if (dest_sr_uuid = vconf.sr) then
        begin
          (* Check if the VDI uuid already exists in the target SR *)
          if (dest_vdi_exists_on_sr vdi_uuid dest_sr_ref true) then
            false
          else
            failwith ("SR UUID matches on destination but VDI does not exist")
        end
      else
        true
    else
      (not is_intra_pool) || (dest_sr_uuid <> vconf.sr)
  in

The check also covers the case where we’re doing an intra-pool migration and not copying all of the disks, in which case we don’t need to do anything for that disk.

We now have a wrapper function that creates a new datapath and passes it to a continuation function. On error it handles the destruction of the datapath:

let with_new_dp cont =
    let dp = Printf.sprintf (if vconf.do_mirror then "mirror_%s" else "copy_%s") vconf.dp in
    try cont dp
    with e ->
      (try SMAPI.DP.destroy ~dbg ~dp ~allow_leak:false with _ -> info "Failed to cleanup datapath: %s" dp);
      raise e in

and now a helper that, given a remote VDI uuid, looks up the reference on the remote host and gives it to a continuation function. On failure of the continuation it will destroy the remote VDI:

  let with_remote_vdi remote_vdi cont =
    debug "Executing remote scan to ensure VDI is known to xapi";
    XenAPI.SR.scan remote.rpc remote.session dest_sr_ref;
    let query = Printf.sprintf "(field \"location\"=\"%s\") and (field \"SR\"=\"%s\")" remote_vdi (Ref.string_of dest_sr_ref) in
    let vdis = XenAPI.VDI.get_all_records_where remote.rpc remote.session query in
    let remote_vdi_ref = match vdis with
      | [] -> raise (Api_errors.Server_error(Api_errors.vdi_location_missing, [Ref.string_of dest_sr_ref; remote_vdi]))
      | h :: [] -> debug "Found remote vdi reference: %s" (Ref.string_of (fst h)); fst h
      | _ -> raise (Api_errors.Server_error(Api_errors.location_not_unique, [Ref.string_of dest_sr_ref; remote_vdi])) in
    try cont remote_vdi_ref
    with e ->
      (try XenAPI.VDI.destroy remote.rpc remote.session remote_vdi_ref with _ -> error "Failed to destroy remote VDI");
      raise e in

another helper to gather together info about a mirrored VDI:

let get_mirror_record ?new_dp remote_vdi remote_vdi_reference =
    { mr_dp = new_dp;
      mr_mirrored = mirror;
      mr_local_sr = vconf.sr;
      mr_local_vdi = vconf.location;
      mr_remote_sr = dest_sr_uuid;
      mr_remote_vdi = remote_vdi;
      mr_local_xenops_locator = vconf.xenops_locator;
      mr_remote_xenops_locator = Xapi_xenops.xenops_vdi_locator_of_strings dest_sr_uuid remote_vdi;
      mr_local_vdi_reference = vconf.vdi;
      mr_remote_vdi_reference = remote_vdi_reference } in

and finally the really important function:

let mirror_to_remote new_dp =
    let task =
      if not vconf.do_mirror then
        SMAPI.DATA.copy ~dbg ~sr:vconf.sr ~vdi:vconf.location ~dp:new_dp ~url:remote.sm_url ~dest:dest_sr_uuid
      else begin
        (* Though we have no intention of "write", here we use the same mode as the
           associated VBD on a mirrored VDIs (i.e. always RW). This avoids problem
           when we need to start/stop the VM along the migration. *)
        let read_write = true in
        (* DP set up is only essential for MIRROR.start/stop due to their open ended pattern.
           It's not necessary for copy which will take care of that itself. *)
        ignore(SMAPI.VDI.attach ~dbg ~dp:new_dp ~sr:vconf.sr ~vdi:vconf.location ~read_write);
        SMAPI.VDI.activate ~dbg ~dp:new_dp ~sr:vconf.sr ~vdi:vconf.location;
        ignore(Storage_access.register_mirror __context vconf.location);
        SMAPI.DATA.MIRROR.start ~dbg ~sr:vconf.sr ~vdi:vconf.location ~dp:new_dp ~url:remote.sm_url ~dest:dest_sr_uuid
      end in

    let mapfn x =
      let total = Int64.to_float total_size in
      let done_ = Int64.to_float !so_far /. total in
      let remaining = Int64.to_float vconf.size /. total in
      done_ +. x *. remaining in

    let open Storage_access in

    let task_result =
      task |> register_task __context
      |> add_to_progress_map mapfn
      |> wait_for_task dbg
      |> remove_from_progress_map
      |> unregister_task __context
      |> success_task dbg in

    let mirror_id, remote_vdi =
      if not vconf.do_mirror then
        let vdi = task_result |> vdi_of_task dbg in
        remote_vdis := vdi.vdi :: !remote_vdis;
        None, vdi.vdi
      else
        let mirrorid = task_result |> mirror_of_task dbg in
        let m = SMAPI.DATA.MIRROR.stat ~dbg ~id:mirrorid in
        Some mirrorid, m.Mirror.dest_vdi in

    so_far := Int64.add !so_far vconf.size;
    debug "Local VDI %s %s to %s" vconf.location (if vconf.do_mirror then "mirrored" else "copied") remote_vdi;
    mirror_id, remote_vdi in

This is the bit that actually starts the mirroring or copying. Before the call to mirror we call VDI.attach and VDI.activate locally to ensure that if the VM is shutdown then the detach/deactivate there doesn’t kill the mirroring process.

Note the parameters to the SMAPI call are sr and vdi, locating the local VDI and SM backend, new_dp, the datapath we’re using for the mirroring, url, which is the remote url on which SMAPI calls work, and dest, the destination SR uuid. These are also the arguments to copy above too.

There’s a little function to calculate the overall progress of the task, and the function waits until the completion of the task before it continues. The function success_task will raise an exception if the task failed. For DATA.mirror, completion implies both that the disk data has been copied to the destination and that all local writes are being mirrored to the destination. Hence more cleanup must be done on cancellation. In contrast, if the DATA.copy path had been taken then the operation at this point has completely finished.

The result of this function is an optional mirror id and the remote VDI uuid.

Next, there is a post_mirror function:

  let post_mirror mirror_id mirror_record =
    try
      let result = continuation mirror_record in
      (match mirror_id with
       | Some mid -> ignore(Storage_access.unregister_mirror mid);
       | None -> ());
      if mirror && not (Xapi_fist.storage_motion_keep_vdi () || copy) then
        Helpers.call_api_functions ~__context (fun rpc session_id ->
            XenAPI.VDI.destroy rpc session_id vconf.vdi);
      result
    with e ->
      let mirror_failed =
        match mirror_id with
        | Some mid ->
          ignore(Storage_access.unregister_mirror mid);
          let m = SMAPI.DATA.MIRROR.stat ~dbg ~id:mid in
          (try SMAPI.DATA.MIRROR.stop ~dbg ~id:mid with _ -> ());
          m.Mirror.failed
        | None -> false in
      if mirror_failed then raise (Api_errors.Server_error(Api_errors.mirror_failed,[Ref.string_of vconf.vdi]))
      else raise e in

This is poorly named - it is post mirror and copy. The aim of this function is to destroy the source VDIs on successful completion of the continuation function, which will have migrated the VM to the destination. In its exception handler it will stop the mirroring, but before doing so it will check to see if the mirroring process it was looking after has itself failed, and raise mirror_failed if so. This is because a failed mirror can result in a range of actual errors, and we decide here that the failed mirror was probably the root cause.

These functions are assembled together at the end of the vdi_copy_fun function:

   if mirror then
    with_new_dp (fun new_dp ->
        let mirror_id, remote_vdi = mirror_to_remote new_dp in
        with_remote_vdi remote_vdi (fun remote_vdi_ref ->
            let mirror_record = get_mirror_record ~new_dp remote_vdi remote_vdi_ref in
            post_mirror mirror_id mirror_record))
  else
    let mirror_record = get_mirror_record vconf.location (XenAPI.VDI.get_by_uuid remote.rpc remote.session vdi_uuid) in
    continuation mirror_record

again, mirror here is poorly named, and means mirror or copy.

Once all of the disks have been mirrored or copied, we jump back to the body of migrate_send. We split apart the mirror records according to the source of the VDI:

      let was_from vmap = List.exists (fun vconf -> vconf.vdi = vmap.mr_local_vdi_reference) in

      let suspends_map, snapshots_map, vdi_map = List.fold_left (fun (suspends, snapshots, vdis) vmap ->
          if was_from vmap suspends_vdis then  vmap :: suspends, snapshots, vdis
          else if was_from vmap snapshots_vdis then suspends, vmap :: snapshots, vdis
          else suspends, snapshots, vmap :: vdis
        ) ([],[],[]) all_map in

then we reassemble all_map from this, for some reason:

    let all_map = List.concat [suspends_map; snapshots_map; vdi_map] in

Now we need to update the snapshot-of links:

     (* All the disks and snapshots have been created in the remote SR(s),
       * so update the snapshot links if there are any snapshots. *)
      if snapshots_map <> [] then
        update_snapshot_info ~__context ~dbg ~url:remote.sm_url ~vdi_map ~snapshots_map;

I’m not entirely sure why this is done in this layer as opposed to in the storage layer.

A little housekeeping:

     let xenops_vdi_map = List.map (fun mirror_record -> (mirror_record.mr_local_xenops_locator, mirror_record.mr_remote_xenops_locator)) all_map in

      (* Wait for delay fist to disappear *)
      wait_for_fist __context Xapi_fist.pause_storage_migrate "pause_storage_migrate";

      TaskHelper.exn_if_cancelling ~__context;

the fist thing here simply allows tests to put in a delay at this specific point.

We also check the task to see if we’ve been cancelled and raise an exception if so.

The VM metadata is now imported into the remote pool, with all the XenAPI level objects remapped:

let new_vm =
        if is_intra_pool
        then vm
        else
          (* Make sure HA replaning cycle won't occur right during the import process or immediately after *)
          let () = if ha_always_run_reset then XenAPI.Pool.ha_prevent_restarts_for ~rpc:remote.rpc ~session_id:remote.session ~seconds:(Int64.of_float !Xapi_globs.ha_monitor_interval) in
          (* Move the xapi VM metadata to the remote pool. *)
          let vms =
            let vdi_map =
              List.map (fun mirror_record -> {
                    local_vdi_reference = mirror_record.mr_local_vdi_reference;
                    remote_vdi_reference = Some mirror_record.mr_remote_vdi_reference;
                  })
                all_map in
            let vif_map =
              List.map (fun (vif, network) -> {
                    local_vif_reference = vif;
                    remote_network_reference = network;
                  })
                vif_map in
            let vgpu_map =
              List.map (fun (vgpu, gpu_group) -> {
                    local_vgpu_reference = vgpu;
                    remote_gpu_group_reference = gpu_group;
                  })
                vgpu_map
            in
            inter_pool_metadata_transfer ~__context ~remote ~vm ~vdi_map
              ~vif_map ~vgpu_map ~dry_run:false ~live:true ~copy
          in
          let vm = List.hd vms in
          let () = if ha_always_run_reset then XenAPI.VM.set_ha_always_run ~rpc:remote.rpc ~session_id:remote.session ~self:vm ~value:false in
          (* Reserve resources for the new VM on the destination pool's host *)
          let () = XenAPI.Host.allocate_resources_for_vm remote.rpc remote.session remote.dest_host vm true in
          vm in

More waiting for fist points:

     wait_for_fist __context Xapi_fist.pause_storage_migrate2 "pause_storage_migrate2";

      (* Attach networks on remote *)
      XenAPI.Network.attach_for_vm ~rpc:remote.rpc ~session_id:remote.session ~host:remote.dest_host ~vm:new_vm;

also make sure all the networks are plugged for the VM on the destination. Next we create the xenopsd-level vif map, equivalent to the vdi_map above:

  (* Create the vif-map for xenops, linking VIF devices to bridge names on the remote *)
      let xenops_vif_map =
        let vifs = XenAPI.VM.get_VIFs ~rpc:remote.rpc ~session_id:remote.session ~self:new_vm in
        List.map (fun vif ->
            let vifr = XenAPI.VIF.get_record ~rpc:remote.rpc ~session_id:remote.session ~self:vif in
            let bridge = Xenops_interface.Network.Local
                (XenAPI.Network.get_bridge ~rpc:remote.rpc ~session_id:remote.session ~self:vifr.API.vIF_network) in
            vifr.API.vIF_device, bridge
          ) vifs
      in

Now we destroy any extra mirror datapaths we set up previously:

     (* Destroy the local datapaths - this allows the VDIs to properly detach, invoking the migrate_finalize calls *)
      List.iter (fun mirror_record ->
          if mirror_record.mr_mirrored
          then match mirror_record.mr_dp with | Some dp ->  SMAPI.DP.destroy ~dbg ~dp ~allow_leak:false | None -> ()) all_map;

More housekeeping:

    SMPERF.debug "vm.migrate_send: migration initiated vm:%s" vm_uuid;

      (* In case when we do SXM on the same host (mostly likely a VDI
         migration), the VM's metadata in xenopsd will be in-place updated
         as soon as the domain migration starts. For these case, there
         will be no (clean) way back from this point. So we disable task
         cancellation for them here.
       *)
      if is_same_host then (TaskHelper.exn_if_cancelling ~__context; TaskHelper.set_not_cancellable ~__context);

Finally we get to the memory-image part of the migration:

      (* It's acceptable for the VM not to exist at this point; shutdown commutes with storage migrate *)
      begin
        try
          Xapi_xenops.Events_from_xenopsd.with_suppressed queue_name dbg vm_uuid
            (fun () ->
               let xenops_vgpu_map = (* can raise VGPU_mapping *)
                 infer_vgpu_map ~__context ~remote new_vm in
               migrate_with_retry
                 ~__context queue_name dbg vm_uuid xenops_vdi_map
                 xenops_vif_map xenops_vgpu_map remote.xenops_url;
               Xapi_xenops.Xenopsd_metadata.delete ~__context vm_uuid)
        with
        | Xenops_interface.Does_not_exist ("VM",_)
        | Xenops_interface.Does_not_exist ("extra",_) ->
          info "%s: VM %s stopped being live during migration"
            "vm_migrate_send" vm_uuid
        | VGPU_mapping(msg) ->
          info "%s: VM %s - can't infer vGPU map: %s"
            "vm_migrate_send" vm_uuid msg;
          raise Api_errors.
                  (Server_error
                     (vm_migrate_failed,
                      ([ vm_uuid
                       ; Helpers.get_localhost_uuid ()
                       ; Db.Host.get_uuid ~__context ~self:remote.dest_host
                       ; "The VM changed its power state during migration"
                       ])))
      end;

      debug "Migration complete";
      SMPERF.debug "vm.migrate_send: migration complete vm:%s" vm_uuid;

Now we tidy up after ourselves:

      (* So far the main body of migration is completed, and the rests are
         updates, config or cleanup on the source and destination. There will
         be no (clean) way back from this point, due to these destructive
         changes, so we don't want user intervention e.g. task cancellation.
       *)
      TaskHelper.exn_if_cancelling ~__context;
      TaskHelper.set_not_cancellable ~__context;
      XenAPI.VM.pool_migrate_complete remote.rpc remote.session new_vm remote.dest_host;

      detach_local_network_for_vm ~__context ~vm ~destination:remote.dest_host;
      Xapi_xenops.refresh_vm ~__context ~self:vm;

the function pool_migrate_complete is called on the destination host, and consists of a few things that ordinarily would be set up during VM.start or the like:

let pool_migrate_complete ~__context ~vm ~host =
  let id = Db.VM.get_uuid ~__context ~self:vm in
  debug "VM.pool_migrate_complete %s" id;
  let dbg = Context.string_of_task __context in
  let queue_name = Xapi_xenops_queue.queue_of_vm ~__context ~self:vm in
  if Xapi_xenops.vm_exists_in_xenopsd queue_name dbg id then begin
    Cpuid_helpers.update_cpu_flags ~__context ~vm ~host;
    Xapi_xenops.set_resident_on ~__context ~self:vm;
    Xapi_xenops.add_caches id;
    Xapi_xenops.refresh_vm ~__context ~self:vm;
    Monitor_dbcalls_cache.clear_cache_for_vm ~vm_uuid:id
  end

More tidying up, remapping some remaining VBDs and clearing state on the sender:

      (* Those disks that were attached at the point the migration happened will have been
         remapped by the Events_from_xenopsd logic. We need to remap any other disks at
         this point here *)

      if is_intra_pool
      then
        List.iter
          (fun vm' ->
             intra_pool_vdi_remap ~__context vm' all_map;
             intra_pool_fix_suspend_sr ~__context remote.dest_host vm')
          vm_and_snapshots;

      (* If it's an inter-pool migrate, the VBDs will still be 'currently-attached=true'
         because we supressed the events coming from xenopsd. Destroy them, so that the
         VDIs can be destroyed *)
      if not is_intra_pool && not copy
      then List.iter (fun vbd -> Db.VBD.destroy ~__context ~self:vbd) (vbds @ snapshots_vbds);

      new_vm
    in

The remark about the Events_from_xenopsd is that we have a thread watching for events that are emitted by xenopsd, and we resynchronise xapi’s state according to xenopsd’s state for several fields for which xenopsd is considered the canonical source of truth. One of these is the exact VDI the VBD is associated with.

The suspend_SR field of the VM is set to the source’s value, so we reset that.

Now we move the RRDs:

  if not copy then begin
      Rrdd_proxy.migrate_rrd ~__context ~remote_address:remote.remote_ip ~session_id:(Ref.string_of remote.session)
        ~vm_uuid:vm_uuid ~host_uuid:(Ref.string_of remote.dest_host) ()
    end;

This can be done for intra- and inter- pool migrates in the same way, simplifying the logic.

However, for messages and blobs we have to only migrate them for inter-pool migrations:

   if not is_intra_pool && not copy then begin
      (* Replicate HA runtime flag if necessary *)
      if ha_always_run_reset then XenAPI.VM.set_ha_always_run ~rpc:remote.rpc ~session_id:remote.session ~self:new_vm ~value:true;
      (* Send non-database metadata *)
      Xapi_message.send_messages ~__context ~cls:`VM ~obj_uuid:vm_uuid
        ~session_id:remote.session ~remote_address:remote.remote_master_ip;
      Xapi_blob.migrate_push ~__context ~rpc:remote.rpc
        ~remote_address:remote.remote_master_ip ~session_id:remote.session ~old_vm:vm ~new_vm ;
      (* Signal the remote pool that we're done *)
    end;

Lastly, we destroy the VM record on the source:

    Helpers.call_api_functions ~__context (fun rpc session_id ->
        if not is_intra_pool && not copy then begin
          info "Destroying VM ref=%s uuid=%s" (Ref.string_of vm) vm_uuid;
          Xapi_vm_lifecycle.force_state_reset ~__context ~self:vm ~value:`Halted;
          List.iter (fun self -> Db.VM.destroy ~__context ~self) vm_and_snapshots
        end);
    SMPERF.debug "vm.migrate_send exiting vm:%s" vm_uuid;
    new_vm

The exception handler still has to clean some state, but mostly things are handled in the CPS functions declared above:

with e ->
    error "Caught %s: cleaning up" (Printexc.to_string e);

    (* We do our best to tidy up the state left behind *)
    Events_from_xenopsd.with_suppressed queue_name dbg vm_uuid (fun () ->
        try
          let _, state = XenopsAPI.VM.stat dbg vm_uuid in
          if Xenops_interface.(state.Vm.power_state = Suspended) then begin
            debug "xenops: %s: shutting down suspended VM" vm_uuid;
            Xapi_xenops.shutdown ~__context ~self:vm None;
          end;
        with _ -> ());

    if not is_intra_pool && Db.is_valid_ref __context vm then begin
      List.map (fun self -> Db.VM.get_uuid ~__context ~self) vm_and_snapshots
      |> List.iter (fun self ->
          try
            let vm_ref = XenAPI.VM.get_by_uuid remote.rpc remote.session self in
            info "Destroying stale VM uuid=%s on destination host" self;
            XenAPI.VM.destroy remote.rpc remote.session vm_ref
          with e -> error "Caught %s while destroying VM uuid=%s on destination host" (Printexc.to_string e) self)
    end;

    let task = Context.get_task_id __context in
    let oc = Db.Task.get_other_config ~__context ~self:task in
    if List.mem_assoc "mirror_failed" oc then begin
      let failed_vdi = List.assoc "mirror_failed" oc in
      let vconf = List.find (fun vconf -> vconf.location=failed_vdi) vms_vdis in
      debug "Mirror failed for VDI: %s" failed_vdi;
      raise (Api_errors.Server_error(Api_errors.mirror_failed,[Ref.string_of vconf.vdi]))
    end;
    TaskHelper.exn_if_cancelling ~__context;
    begin match e with
      | Storage_interface.Backend_error(code, params) -> raise (Api_errors.Server_error(code, params))
      | Storage_interface.Unimplemented(code) -> raise (Api_errors.Server_error(Api_errors.unimplemented_in_sm_backend, [code]))
      | Xenops_interface.Cancelled _ -> TaskHelper.raise_cancelled ~__context
      | _ -> raise e
    end

Failures during the migration can result in the VM being in a suspended state. There’s no point leaving it like this since there’s nothing that can be done to resume it, so we force shut it down.

We also try to remove the VM record from the destination if we managed to send it there.

Finally we check for mirror failure in the task - this is set by the events thread watching for events from the storage layer, in storage_access.ml

Storage code

The part of the code that is conceptually in the storage layer, but physically in xapi, is located in storage_migrate.ml. There are logically a few separate parts to this file:

• A stateful module for persisting state across xapi restarts.
• Some general helper functions
• Some quite specific helper functions related to actions to be taken on deactivate/detach
• An NBD handler
• The implementations of the SMAPIv2 mirroring APIs

Let’s start by considering the way the storage APIs are intended to be used.

Copying a VDI

DATA.copy takes several parameters:

• dbg - a debug string
• sr - the source SR (a uuid)
• vdi - the source VDI (a uuid)
• dp - unused
• url - a URL on which SMAPIv2 API calls can be made
• sr - the destination SR in which the VDI should be copied

and returns a parameter of type Task.id. The API call is intended to be called in an asynchronous fashion - ie., the caller makes the call, receives the task ID back and polls or uses the event mechanism to wait until the task has completed. The task may be cancelled via the Task.cancel API call. The result of the operation is obtained by calling TASK.stat, which returns a record:

	type t = {
		id: id;
		dbg: string;
		ctime: float;
		state: state;
		subtasks: (string * state) list;
		debug_info: (string * string) list;
		backtrace: string;
	}

Where the state field contains the result once the task has completed:

type async_result_t =
	| Vdi_info of vdi_info
	| Mirror_id of Mirror.id

type completion_t = {
	duration : float;
	result : async_result_t option
}

type state =
	| Pending of float
	| Completed of completion_t
	| Failed of Rpc.t

Once the result has been obtained from the task, the task should be destroyed via the TASK.destroy API call.

The implementation uses the url parameter to make SMAPIv2 calls to the destination SR. This is used, for example, to invoke a VDI.create call if necessary. The URL contains an authentication token within it (valid for the duration of the XenAPI call that caused this DATA.copy API call).

The implementation tries to minimize the amount of data copied by looking for related VDIs on the destination SR. See below for more details.

Mirroring a VDI

DATA.MIRROR.start takes a similar set of parameters to that of copy:

• dbg - a debug string
• sr - the source SR (a uuid)
• vdi - the source VDI (a uuid)
• dp - the datapath on which the VDI has been attached
• url - a URL on which SMAPIv2 API calls can be made
• sr - the destination SR in which the VDI should be copied

Similar to copy above, this returns a task id. The task ‘completes’ once the mirror has been set up - that is, at any point afterwards we can detach the disk and the destination disk will be identical to the source. Unlike for copy the operation is ongoing after the API call completes, since new writes need to be mirrored to the destination. Therefore the completion type of the mirror operation is Mirror_id which contains a handle on which further API calls related to the mirror call can be made. For example MIRROR.stat whose signature is:

MIRROR.stat: dbg:debug_info -> id:Mirror.id -> Mirror.t

The return type of this call is a record containing information about the mirror:

type state =
	| Receiving
	| Sending
	| Copying

type t = {
	source_vdi : vdi;
	dest_vdi : vdi;
	state : state list;
	failed : bool;
}

Note that state is a list since the initial phase of the operation requires both copying and mirroring.

Additionally the mirror can be cancelled using the MIRROR.stop API call.

Code walkthrough

let’s go through the implementation of copy:

DATA.copy

let copy ~task ~dbg ~sr ~vdi ~dp ~url ~dest =
  debug "copy sr:%s vdi:%s url:%s dest:%s" sr vdi url dest;
  let remote_url = Http.Url.of_string url in
  let module Remote = Client(struct let rpc = rpc ~srcstr:"smapiv2" ~dststr:"dst_smapiv2" remote_url end) in

Here we are constructing a module Remote on which we can do SMAPIv2 calls directly on the destination.

  try

Wrap the whole function in an exception handler.

    (* Find the local VDI *)
    let vdis = Local.SR.scan ~dbg ~sr in
    let local_vdi =
      try List.find (fun x -> x.vdi = vdi) vdis
      with Not_found -> failwith (Printf.sprintf "Local VDI %s not found" vdi) in

We first find the metadata for our source VDI by doing a local SMAPIv2 call SR.scan. This returns a list of VDI metadata, out of which we extract the VDI we’re interested in.

    try

Another exception handler. This looks redundant to me right now.

      let similar_vdis = Local.VDI.similar_content ~dbg ~sr ~vdi in
      let similars = List.map (fun vdi -> vdi.content_id) similar_vdis in
      debug "Similar VDIs to %s = [ %s ]" vdi (String.concat "; " (List.map (fun x -> Printf.sprintf "(vdi=%s,content_id=%s)" x.vdi x.content_id) similar_vdis));

Here we look for related VDIs locally using the VDI.similar_content SMAPIv2 API call. This searches for related VDIs and returns an ordered list where the most similar is first in the list. It returns both clones and snapshots, and hence is more general than simply following snapshot_of links.

      let remote_vdis = Remote.SR.scan ~dbg ~sr:dest in
      (** We drop cbt_metadata VDIs that do not have any actual data *)
      let remote_vdis = List.filter (fun vdi -> vdi.ty <> "cbt_metadata") remote_vdis in

      let nearest = List.fold_left
          (fun acc content_id -> match acc with
             | Some x -> acc
             | None ->
               try Some (List.find (fun vdi -> vdi.content_id = content_id && vdi.virtual_size <= local_vdi.virtual_size) remote_vdis)
               with Not_found -> None) None similars in

      debug "Nearest VDI: content_id=%s vdi=%s"
        (Opt.default "None" (Opt.map (fun x -> x.content_id) nearest))
        (Opt.default "None" (Opt.map (fun x -> x.vdi) nearest));

Here we look for VDIs on the destination with the same content_id as one of the locally similar VDIs. We will use this as a base image and only copy deltas to the destination. This is done by cloning the VDI on the destination and then using sparse_dd to find the deltas from our local disk to our local copy of the content_id disk and streaming these to the destination. Note that we need to ensure the VDI is smaller than the one we want to copy since we can’t resize disks downwards in size.

      let remote_base = match nearest with
        | Some vdi ->
          debug "Cloning VDI %s" vdi.vdi;
          let vdi_clone = Remote.VDI.clone ~dbg ~sr:dest ~vdi_info:vdi in
          if vdi_clone.virtual_size <> local_vdi.virtual_size then begin
            let new_size = Remote.VDI.resize ~dbg ~sr:dest ~vdi:vdi_clone.vdi ~new_size:local_vdi.virtual_size in
            debug "Resize remote VDI %s to %Ld: result %Ld" vdi_clone.vdi local_vdi.virtual_size new_size;
          end;
          vdi_clone
        | None ->
          debug "Creating a blank remote VDI";
          Remote.VDI.create ~dbg ~sr:dest ~vdi_info:{ local_vdi with sm_config = [] }  in

If we’ve found a base VDI we clone it and resize it immediately. If there’s nothing on the destination already we can use, we just create a new VDI. Note that the calls to create and clone may well fail if the destination host is not the SRmaster. This is handled purely in the rpc function:

let rec rpc ~srcstr ~dststr url call =
  let result = XMLRPC_protocol.rpc ~transport:(transport_of_url url)
      ~srcstr ~dststr ~http:(xmlrpc ~version:"1.0" ?auth:(Http.Url.auth_of url) ~query:(Http.Url.get_query_params url) (Http.Url.get_uri url)) call
  in
  if not result.Rpc.success then begin
    debug "Got failure: checking for redirect";
    debug "Call was: %s" (Rpc.string_of_call call);
    debug "result.contents: %s" (Jsonrpc.to_string result.Rpc.contents);
    match Storage_interface.Exception.exnty_of_rpc result.Rpc.contents with
    | Storage_interface.Exception.Redirect (Some ip) ->
      let open Http.Url in
      let newurl =
        match url with
        | (Http h, d) ->
          (Http {h with host=ip}, d)
        | _ ->
          remote_url ip in
      debug "Redirecting to ip: %s" ip;
      let r = rpc ~srcstr ~dststr newurl call in
      debug "Successfully redirected. Returning";
      r
    | _ ->
      debug "Not a redirect";
      result
  end
  else result

Back to the copy function:

      let remote_copy = copy' ~task ~dbg ~sr ~vdi ~url ~dest ~dest_vdi:remote_base.vdi |> vdi_info in

This calls the actual data copy part. See below for more on that.

      let snapshot = Remote.VDI.snapshot ~dbg ~sr:dest ~vdi_info:remote_copy in
      Remote.VDI.destroy ~dbg ~sr:dest ~vdi:remote_copy.vdi;
      Some (Vdi_info snapshot)

Finally we snapshot the remote VDI to ensure we’ve got a VDI of type ‘snapshot’ on the destination, and we delete the non-snapshot VDI.

    with e ->
      error "Caught %s: copying snapshots vdi" (Printexc.to_string e);
      raise (Internal_error (Printexc.to_string e))
  with
  | Backend_error(code, params)
  | Api_errors.Server_error(code, params) ->
    raise (Backend_error(code, params))
  | e ->
    raise (Internal_error(Printexc.to_string e))

The exception handler does nothing - so we leak remote VDIs if the exception happens after we’ve done our cloning :-(

DATA.copy_into

Let’s now look at the data-copying part. This is common code shared between VDI.copy, VDI.copy_into and MIRROR.start and hence has some duplication of the calls made above.

let copy_into ~task ~dbg ~sr ~vdi ~url ~dest ~dest_vdi =
  copy' ~task ~dbg ~sr ~vdi ~url ~dest ~dest_vdi

copy_into is a stub and just calls copy'

let copy' ~task ~dbg ~sr ~vdi ~url ~dest ~dest_vdi =
  let remote_url = Http.Url.of_string url in
  let module Remote = Client(struct let rpc = rpc ~srcstr:"smapiv2" ~dststr:"dst_smapiv2" remote_url end) in
  debug "copy local=%s/%s url=%s remote=%s/%s" sr vdi url dest dest_vdi;

This call takes roughly the same parameters as the ``DATA.copy` call above, except it specifies the destination VDI. Once again we construct a module to do remote SMAPIv2 calls

  (* Check the remote SR exists *)
  let srs = Remote.SR.list ~dbg in
  if not(List.mem dest srs)
  then failwith (Printf.sprintf "Remote SR %s not found" dest);

Sanity check.

  let vdis = Remote.SR.scan ~dbg ~sr:dest in
  let remote_vdi =
    try List.find (fun x -> x.vdi = dest_vdi) vdis
    with Not_found -> failwith (Printf.sprintf "Remote VDI %s not found" dest_vdi)
  in

Find the metadata of the destination VDI

  let dest_content_id = remote_vdi.content_id in

If we’ve got a local VDI with the same content_id as the destination, we only need copy the deltas, so we make a note of the destination content ID here.

  (* Find the local VDI *)
  let vdis = Local.SR.scan ~dbg ~sr in
  let local_vdi =
    try List.find (fun x -> x.vdi = vdi) vdis
    with Not_found -> failwith (Printf.sprintf "Local VDI %s not found" vdi) in

  debug "copy local=%s/%s content_id=%s" sr vdi local_vdi.content_id;
  debug "copy remote=%s/%s content_id=%s" dest dest_vdi remote_vdi.content_id;

Find the source VDI metadata.

  if local_vdi.virtual_size > remote_vdi.virtual_size then begin
    (* This should never happen provided the higher-level logic is working properly *)
    error "copy local=%s/%s virtual_size=%Ld > remote=%s/%s virtual_size = %Ld" sr vdi local_vdi.virtual_size dest dest_vdi remote_vdi.virtual_size;
    failwith "local VDI is larger than the remote VDI";
  end;

Sanity check - the remote VDI can’t be smaller than the source.

  let on_fail : (unit -> unit) list ref = ref [] in

We do some ugly error handling here by keeping a mutable list of operations to perform in the event of a failure.

  let base_vdi =
    try
      let x = (List.find (fun x -> x.content_id = dest_content_id) vdis).vdi in
      debug "local VDI %s has content_id = %s; we will perform an incremental copy" x dest_content_id;
      Some x
    with _ ->
      debug "no local VDI has content_id = %s; we will perform a full copy" dest_content_id;
      None
  in

See if we can identify a local VDI with the same content_id as the destination. If not, no problem.

  try
    let remote_dp = Uuid.string_of_uuid (Uuid.make_uuid ()) in
    let base_dp = Uuid.string_of_uuid (Uuid.make_uuid ()) in
    let leaf_dp = Uuid.string_of_uuid (Uuid.make_uuid ()) in

Construct some datapaths - named reasons why the VDI is attached - that we will pass to VDI.attach/activate.

    let dest_vdi_url = Http.Url.set_uri remote_url (Printf.sprintf "%s/nbd/%s/%s/%s" (Http.Url.get_uri remote_url) dest dest_vdi remote_dp) |> Http.Url.to_string in

    debug "copy remote=%s/%s NBD URL = %s" dest dest_vdi dest_vdi_url;

Here we are constructing a URI that we use to connect to the destination xapi. The handler for this particular path will verify the credentials and then pass the connection on to tapdisk which will behave as a NBD server. The VDI has to be attached and activated for this to work, unlike the new NBD handler in xapi-nbd that is smarter. The handler for this URI is declared in this file

    let id=State.copy_id_of (sr,vdi) in
    debug "Persisting state for copy (id=%s)" id;
    State.add id State.(Copy_op Copy_state.({
        base_dp; leaf_dp; remote_dp; dest_sr=dest; copy_vdi=remote_vdi.vdi; remote_url=url}));

Since we’re about to perform a long-running operation that is stateful, we persist the state here so that if xapi is restarted we can cancel the operation and not leak VDI attaches. Normally in xapi code we would be doing VBD.plug operations to persist the state in the xapi db, but this is storage code so we have to use a different mechanism.

    SMPERF.debug "mirror.copy: copy initiated local_vdi:%s dest_vdi:%s" vdi dest_vdi;

    Pervasiveext.finally (fun () ->
        debug "activating RW datapath %s on remote=%s/%s" remote_dp dest dest_vdi;
        ignore(Remote.VDI.attach ~dbg ~sr:dest ~vdi:dest_vdi ~dp:remote_dp ~read_write:true);
        Remote.VDI.activate ~dbg ~dp:remote_dp ~sr:dest ~vdi:dest_vdi;

        with_activated_disk ~dbg ~sr ~vdi:base_vdi ~dp:base_dp
          (fun base_path ->
             with_activated_disk ~dbg ~sr ~vdi:(Some vdi) ~dp:leaf_dp
               (fun src ->
                  let dd = Sparse_dd_wrapper.start ~progress_cb:(progress_callback 0.05 0.9 task) ?base:base_path true (Opt.unbox src)
                      dest_vdi_url remote_vdi.virtual_size in
                  Storage_task.with_cancel task
                    (fun () -> Sparse_dd_wrapper.cancel dd)
                    (fun () ->
                       try Sparse_dd_wrapper.wait dd
                       with Sparse_dd_wrapper.Cancelled -> Storage_task.raise_cancelled task)
               )
          );
      )
      (fun () ->
         Remote.DP.destroy ~dbg ~dp:remote_dp ~allow_leak:false;
         State.remove_copy id
      );

In this chunk of code we attach and activate the disk on the remote SR via the SMAPI, then locally attach and activate both the VDI we’re copying and the base image we’re copying deltas from (if we’ve got one). We then call sparse_dd to copy the data to the remote NBD URL. There is some logic to update progress indicators and to cancel the operation if the SMAPIv2 call TASK.cancel is called.

Once the operation has terminated (either on success, error or cancellation), we remove the local attach and activations in the with_activated_disk function and the remote attach and activation by destroying the datapath on the remote SR. We then remove the persistent state relating to the copy.

    SMPERF.debug "mirror.copy: copy complete local_vdi:%s dest_vdi:%s" vdi dest_vdi;

    debug "setting remote=%s/%s content_id <- %s" dest dest_vdi local_vdi.content_id;
    Remote.VDI.set_content_id ~dbg ~sr:dest ~vdi:dest_vdi ~content_id:local_vdi.content_id;
    (* PR-1255: XXX: this is useful because we don't have content_ids by default *)
    debug "setting local=%s/%s content_id <- %s" sr local_vdi.vdi local_vdi.content_id;
    Local.VDI.set_content_id ~dbg ~sr ~vdi:local_vdi.vdi ~content_id:local_vdi.content_id;
    Some (Vdi_info remote_vdi)

The last thing we do is to set the local and remote content_id. The local set_content_id is there because the content_id of the VDI is constructed from the location if it is unset in the storage_access.ml module of xapi (still part of the storage layer)

  with e ->
    error "Caught %s: performing cleanup actions" (Printexc.to_string e);
    perform_cleanup_actions !on_fail;
    raise e

Here we perform the list of cleanup operations. Theoretically. It seems we don’t ever actually set this to anything, so this is dead code.

DATA.MIRROR.start

let start' ~task ~dbg ~sr ~vdi ~dp ~url ~dest =
  debug "Mirror.start sr:%s vdi:%s url:%s dest:%s" sr vdi url dest;
  SMPERF.debug "mirror.start called sr:%s vdi:%s url:%s dest:%s" sr vdi url dest;
  let remote_url = Http.Url.of_string url in
  let module Remote = Client(struct let rpc = rpc ~srcstr:"smapiv2" ~dststr:"dst_smapiv2" remote_url end) in

  (* Find the local VDI *)
  let vdis = Local.SR.scan ~dbg ~sr in
  let local_vdi =
    try List.find (fun x -> x.vdi = vdi) vdis
    with Not_found -> failwith (Printf.sprintf "Local VDI %s not found" vdi) in

As with the previous calls, we make a remote module for SMAPIv2 calls on the destination, and we find local VDI metadata via SR.scan

  let id = State.mirror_id_of (sr,local_vdi.vdi) in

Mirror ids are deterministically constructed.

  (* A list of cleanup actions to perform if the operation should fail. *)
  let on_fail : (unit -> unit) list ref = ref [] in

This on_fail list is actually used.

  try
    let similar_vdis = Local.VDI.similar_content ~dbg ~sr ~vdi in
    let similars = List.filter (fun x -> x <> "") (List.map (fun vdi -> vdi.content_id) similar_vdis) in
    debug "Similar VDIs to %s = [ %s ]" vdi (String.concat "; " (List.map (fun x -> Printf.sprintf "(vdi=%s,content_id=%s)" x.vdi x.content_id) similar_vdis));

As with copy we look locally for similar VDIs. However, rather than use that here we actually pass this information on to the destination SR via the receive_start internal SMAPIv2 call:

    let result_ty = Remote.DATA.MIRROR.receive_start ~dbg ~sr:dest ~vdi_info:local_vdi ~id ~similar:similars in
    let result = match result_ty with
        Mirror.Vhd_mirror x -> x
    in

This gives the destination SR a chance to say what sort of migration it can support. We only support Vhd_mirror style migrations which require the destination to support the compose SMAPIv2 operation. The type of x is a record:

type mirror_receive_result_vhd_t = {
	mirror_vdi : vdi_info;
	mirror_datapath : dp;
	copy_diffs_from : content_id option;
	copy_diffs_to : vdi;
	dummy_vdi : vdi;
}

Field descriptions:

• mirror_vdi is the VDI to which new writes should be mirrored.
• mirror_datapath is the remote datapath on which the VDI has been attached and activated. This is required to construct the remote NBD url
• copy_diffs_from represents the source base VDI to be used for the non-mirrored data copy.
• copy_diffs_to is the remote VDI to copy those diffs to
• dummy_vdi exists to prevent leaf-coalesce on the mirror_vdi

    (* Enable mirroring on the local machine *)
    let mirror_dp = result.Mirror.mirror_datapath in

    let uri = (Printf.sprintf "/services/SM/nbd/%s/%s/%s" dest result.Mirror.mirror_vdi.vdi mirror_dp) in
    let dest_url = Http.Url.set_uri remote_url uri in
    let request = Http.Request.make ~query:(Http.Url.get_query_params dest_url) ~version:"1.0" ~user_agent:"smapiv2" Http.Put uri in
    let transport = Xmlrpc_client.transport_of_url dest_url in

This is where we connect to the NBD server on the destination.

    debug "Searching for data path: %s" dp;
    let attach_info = Local.DP.attach_info ~dbg:"nbd" ~sr ~vdi ~dp in
    debug "Got it!";

we need the local attach_info to find the local tapdisk so we can send it the connected NBD socket.

    on_fail := (fun () -> Remote.DATA.MIRROR.receive_cancel ~dbg ~id) :: !on_fail;

This should probably be set directly after the call to receive_start

    let tapdev = match tapdisk_of_attach_info attach_info with
      | Some tapdev ->
        debug "Got tapdev";
        let pid = Tapctl.get_tapdisk_pid tapdev in
        let path = Printf.sprintf "/var/run/blktap-control/nbdclient%d" pid in
        with_transport transport (with_http request (fun (response, s) ->
            debug "Here inside the with_transport";
            let control_fd = Unix.socket Unix.PF_UNIX Unix.SOCK_STREAM 0 in
            finally
              (fun () ->
                 debug "Connecting to path: %s" path;
                 Unix.connect control_fd (Unix.ADDR_UNIX path);
                 let msg = dp in
                 let len = String.length msg in
                 let written = Unixext.send_fd control_fd msg 0 len [] s in
                 debug "Sent fd";
                 if written <> len then begin
                   error "Failed to transfer fd to %s" path;
                   failwith "foo"
                 end)
              (fun () ->
                 Unix.close control_fd)));
        tapdev
      | None ->
        failwith "Not attached"
    in