Let’s trace through interesting operations to see how the whole system works.
A XenAPI client wishes to migrate a VM from one host to another within the same pool.
The client will issue a command to migrate the VM and it will be dispatched
by the autogenerated dispatch_call function from xapi/server.ml. For
more information about the generated functions you can have a look to
XAPI IDL model.
The command will trigger the operation VM_migrate that has low level operations performed by the backend. These atomics operations that we will describe in the documentation are:
The command have serveral parameters such as: should it be ran asynchronously, should it be forwared to another host, how arguments should be marshalled and so on. A new thread is created by xapi/server_helpers.ml to handle the command asynchronously. At this point the helper also check if the command should be passed to the message forwarding layer in order to be executed on another host (the destination) or locally if we are already at the right place.
It will finally reach xapi/api_server.ml that will take the action of posted a command to the message broker message switch. It is a JSON-RPC HTTP request sends on a Unix socket to communicate between some XAPI daemons. In the case of the migration this message sends by XAPI will be consumed by the xenopsd daemon that will do the job of migrating the VM.
The migration is an asynchronous task and a thread is created to handle this task. The tasks’s reference is returned to the client, which can then check its status until completion.
As we see in the introduction the xenopsd daemon will pop the operation VM_migrate from the message broker.
Only one backend is know available that interacts with libxc, libxenguest and xenstore. It is the xc backend.
The entities that need to be migrated are: VDI, VIF, VGPU and PCI components.
During the migration process the destination domain will be built with the same
uuid than the original VM but the last part of the UUID will be
XXXXXXXX-XXXX-XXXX-XXXX-000000000001. The original domain will be removed using
XXXXXXXX-XXXX-XXXX-XXXX-000000000000.
There are some points called hooks at which xenopsd can execute some script.
Before starting a migration a command is send to the original domain to execute
a pre migrate script if it exists.
Before starting the migration a command is sent to Qemu using the Qemu Machine Protocol (QMP) to check that the domain can be suspended (see xenopsd/xc/device_common.ml). After checking with Qemu that the VM is suspendable we can start the migration.
As for hooks, commands to source domain are sent using stunnel a daemon which is used as a wrapper to manage SSL encryption communication between two hosts on the same pool. To import metada an XML RPC command is sent to the original domain.
Once imported it will give us a reference id and will allow to build the new domain
on the destination using the temporary VM uuid XXXXXXXX-XXXX-XXXX-XXXX-000000000001
where XXX... is the reference id of the original VM.
One of the first thing to do is to setup the memory. The backend will check that there is no ballooning operation in progress. At this point the migration can fail if a ballooning operation is in progress and takes too much time.
Once memory checked the daemon will get the state of the VM (running, halted, …) and information about the VM are retrieve by the backend like the maximum memory the domain can consume but also information about quotas for example. Information are retrieve by the backend from xenstore.
Once this is complete, we can restore VIF and create the domain.
The synchronisation of the memory is the first point of synchronisation and everythin is ready for VM migration.
After receiving memory we can set up the destination domain. If we have a vGPU we need to kick off its migration process. We will need to wait the acknowledge that indicates that the entry for the GPU has been well initialized. before starting the main VM migration.
Their is a mechanism of handshake for synchronizing between the source and the destination. Using the handshake protocol the receiver inform the sender of the request that everything has been setup and ready to save/restore.
VM restore is a low level atomic operation VM.restore. This operation is represented by a function call to backend. It uses Xenguest, a low-level utility from XAPI toolstack, to interact with the Xen hypervisor and libxc for sending a request of migration to the emu-manager.
After sending the request results coming from emu-manager are collected by the main thread. It blocks until results are received.
During the live migration, emu-manager helps in ensuring the correct state transitions for the devices and handling the message passing for the VM as it’s moved between hosts. This includes making sure that the state of the VM’s virtual devices, like disks or network interfaces, is correctly moved over.
Once all operations are done we can rename the VM on the target from its temporary name to its real UUID. This operation is another low level atomic one VM.rename that will take care of updating the xenstore on the destination.
The next step is the restauration of devices and unpause the domain.
Restoring devices starts by activating VBD using the low level atomic operation VBD.set_active. It is an update of Xenstore. VBDs that are read-write must be plugged before read-only ones. Once activated the low level atomic operation VBD.plug is called. VDI are attached and activate.
Next devices are VIFs that are set as active VIF.set_active and plug VIF.plug. If there are VGPUs we will set them as active now using the atomic VGPU.set_active.
We are almost done. The next step is to create the device model
Create device model is done by using the atomic operation VM.create_device_model. This will configure qemu-dm and started. This allow to manage PCI devices.
PCI.plug is executed by the backend. It plugs a PCI device and advertise it to QEMU if this option is set. It is the case for NVIDIA SR-IOV vGPUS.
At this point devices have been restored. The new domain is considered survivable. We can unpause the domain and performs last actions
Unpause is done by managing the state of the domain using bindings to xenctrl. Once hypervisor has unpaused the domain some actions can be requested using VM.set_domain_action_request. It is a path in xenstore. By default no action is done but a reboot can be for example initiated.
Previously we spoke about some points called hooks at which xenopsd can execute some script. There
is also a hook to run a post migrate script. After the execution of the script if there is one
the migration is almost done. The last step is a handskake to seal the success of the migration
and the old VM can now be cleaned.
Some links are old but even if many changes occured they are relevant for a global understanding of the XAPI toolstack.
A Xenopsd client wishes to start a VM. They must first tell Xenopsd the VM configuration to use. A VM configuration is broken down into objects:
Treating devices as first-class objects is convenient because we wish to expose operations on the devices such as hotplug, unplug, eject (for removable media), carrier manipulation (for network interfaces) etc.
The “add” functions in the Xenopsd interface cause Xenopsd to create the objects:
In the case of xapi, there are a set of functions which convert between the XenAPI objects and the Xenopsd objects. The two interfaces are slightly different because they have different expected users:
The Xenopsd “VM.add” function has code like this:
let add' x =
debug "VM.add %s" (Jsonrpc.to_string (rpc_of_t x));
DB.write x.id x;
let module B = (val get_backend () : S) in
B.VM.add x;
x.idThis function does 2 things:
The Xenopsd database is really a set of config files in the filesystem. All objects belonging to a VM (recall we only have VMs, VBDs, VIFs, PCIs and not stand-alone entities like disks) and are placed into a subdirectory named after the VM e.g.:
# ls /run/nonpersistent/xenopsd/xenlight/VM/7b719ce6-0b17-9733-e8ee-dbc1e6e7b701
config vbd.xvda vbd.xvdb
# cat /run/nonpersistent/xenopsd/xenlight/VM/7b719ce6-0b17-9733-e8ee-dbc1e6e7b701/config
{"id": "7b719ce6-0b17-9733-e8ee-dbc1e6e7b701", "name": "fedora",
...
}Xenopsd doesn’t have as persistent a notion of a VM as xapi, it is expected that all objects are deleted when the host is rebooted. However the objects should be persisted over a simple Xenopsd restart, which is why the objects are stored in the filesystem.
Aside: it would probably be more appropriate to store the metadata in Xenstore since this has the exact object lifetime we need. This will require a more performant Xenstore to realise.
Every running Xenopsd process is linked with a single backend. Currently backends exist for:
From here we shall assume the use of the “Xen via libxc, libxenguest and xenstore” (a.k.a. “Xenopsd classic”) backend.
The backend VM.add function checks whether the VM we have to manage already exists – and if it does then it ensures the Xenstore configuration is intact. This Xenstore configuration is important because at any time a client can query the state of a VM with VM.stat and this relies on certain Xenstore keys being present.
Once the VM metadata has been registered with Xenopsd, the client can call VM.start. Like all potentially-blocking Xenopsd APIs, this function returns a Task id. Please refer to the Task handling design for a general overview of how tasks are handled.
Clients can poll the state of a task by calling TASK.stat but most clients will prefer to use the event system instead. Please refer to the Event handling design for a general overview of how events are handled.
The event model is similar to the XenAPI: clients call a blocking UPDATES.get passing in a token which represents the point in time when the last UPDATES.get returned. The call blocks until some objects have changed state, and these object ids are returned (NB in the XenAPI the current object states are returned) The client must then call the relevant “stat” function, in this case TASK.stat
The client will be able to see the task make progress and use this to – for example – populate a progress bar in a UI. If the client needs to cancel the task then it can call the TASK.cancel; again see the Task handling design to understand how this is implemented.
When the Task has completed successfully, then calls to *.stat will show:
Note: before a Task completes, calls to *.stat will show partial updates e.g. the power state may be Paused but none of the disks may have become plugged. UI clients must choose whether they are happy displaying this in-between state or whether they wish to hide it and pretend the whole operation has happened transactionally. If a particular client wishes to perform side-effects in response to Xenopsd state changes – for example to clean up an external resource when a VIF becomes unplugged – then it must be very careful to avoid responding to these in-between states. Generally it is safest to passively report these values without driving things directly from them. Think of them as status lights on the front panel of a PC: fine to look at but it’s not a good idea to wire them up to actuators which actually do things.
Note: the Xenopsd implementation guarantees that, if it is restarted at any point during the start operation, on restart the VM state shall be “fixed” by either (i) shutting down the VM; or (ii) ensuring the VM is intact and running.
In the case of xapi every Xenopsd Task id bound one-to-one with a XenAPI task by the function sync_with_task. The function update_task is called when xapi receives a notification that a Xenopsd Task has changed state, and updates the corresponding XenAPI task. Xapi launches exactly one thread per Xenopsd instance (“queue”) to monitor for background events via the function events_watch while each thread performing a XenAPI call waits for its specific Task to complete via the function event_wait.
It is the responsibility of the client to call TASK.destroy when the Task is nolonger needed. Xenopsd won’t destroy the task because it contains the success/failure result of the operation which is needed by the client.
What happens when a Xenopsd receives a VM.start request?
When Xenopsd receives the request it adds it to the appropriate per-VM queue via the function queue_operation. To understand this and other internal details of Xenopsd, consult the architecture description. The queue_operation_int function looks like this:
let queue_operation_int dbg id op =
let task = Xenops_task.add tasks dbg (fun t -> perform op t; None) in
Redirector.push id (op, task);
taskThe “task” is a record containing Task metadata plus a “do it now” function which will be executed by a thread from the thread pool. The module Redirector takes care of:
Once a thread from the worker pool becomes free, it will execute the “do it now”
function. In the example above this is perform op t where op is
VM_start vm and t is the Task. The function
perform
has fragments like this:
| VM_start id ->
debug "VM.start %s" id;
perform_atomics (atomics_of_operation op) t;
VM_DB.signal idEach “operation” (e.g. VM_start vm) is decomposed into “micro-ops” by the
function
atomics_of_operation
where the micro-ops are small building-block actions common to the higher-level
operations. Each operation corresponds to a list of “micro-ops”, where there is
no if/then/else. Some of the “micro-ops” may be a no-op depending on the VM
configuration (for example a PV domain may not need a qemu). In the case of
VM_start vm this decomposes into the sequence:
The VM_hook_script micro-op runs the corresponding “hook” scripts. The
code is all in the
Xenops_hooks
module and looks for scripts in the hardcoded path /etc/xapi.d.
The VM_create micro-op calls the VM.create function in the backend.
In the classic Xenopsd backend the
VM.create_exn
function must
On a Xen system a domain is created empty, and memory is actually allocated from the host in the “build” phase via functions in libxenguest. The VM.build_domain_exn function must
cpuid configurationThe xenguest binary was originally a separate binary for two reasons: (i) the libxenguest functions weren’t threadsafe since they used lots of global variables; and (ii) the libxenguest functions used to have a different, incompatible license, which prevent us linking. Both these problems have been resolved but we still shell out to the xenguest binary.
The xenguest binary has also evolved to configure more of the initial domain state. It also reads Xenstore and configures
VBDs and VIFs are said to be “active” when they are intended to be used by a
particular VM, even if the backend/frontend connection hasn’t been established,
or has been closed. If someone calls VBD.stat or VIF.stat then
the result includes both “active” and “plugged”, where “plugged” is true if
the frontend/backend connection is established.
For example xapi will
set VBD.currently_attached
to “active || plugged”. The “active” flag is conceptually very similar to the
traditional “online” flag (which is not documented in the upstream Xen tree
as of Oct/2014 but really should be) except that on unplug, one would set
the “online” key to “0” (false) first before initiating the hotunplug. By
contrast the “active” flag is set to false after the unplug i.e. “set_active”
calls bracket plug/unplug. If the “active” flag was set before the unplug
attempt then as soon as the frontend/backend connection is removed clients
would see the VBD as completely dissociated from the VM – this would be misleading
because Xenopsd will not have had time to use the storage API to release locks
on the disks. By doing all the cleanup before setting “active” to false, clients
can be assured that the disks are now free to be reassigned.
A non-persistent disk is one which is reset to a known-good state on every
VM start. The VBD_epoch_begin is the signal to perform any necessary reset.
The VBD_plug micro-op will plug the VBD into the VM. Every VBD is plugged
in a carefully-chosen order.
Generally, plug order is important for all types of devices. For VBDs, we must
work around the deficiency in the storage interface where a VDI, once attached
read/only, cannot be attached read/write. Since it is legal to attach the same
VDI with multiple VBDs, we must plug them in such that the read/write VBDs
come first. From the guest’s point of view the order we plug them doesn’t
matter because they are indexed by the Xenstore device id (e.g. 51712 = xvda).
The function VBD.plug will
VDI.attach and VDI.activate in the storage API to make the
devices ready (start the tapdisk processes etc)VDI.attach call that are
needed for SCSIid passthrough which is needed to support VSSThe Xenstore keys are written by the functions Device.Vbd.add_async and Device.Vbd.add_wait. In a Linux domain (such as dom0) when the backend directory is created, the kernel creates a “backend device”. Creating any device will cause a kernel UEVENT to fire which is picked up by udev. The udev rules run a script whose only job is to stat(2) the device (from the “params” key in the backend) and write the major and minor number to Xenstore for blkback to pick up. (Aside: FreeBSD doesn’t do any of this, instead the FreeBSD kernel module simply opens the device in the “params” key). The script also writes the backend key “hotplug-status=connected”. We currently wait for this key to be written so that later calls to VBD.stat will return with “plugged=true”. If the call returns before this key is written then sometimes we receive an event, call VBD.stat and conclude erroneously that a spontaneous VBD unplug occurred.
This is for the same reason as VBDs are marked “active”.
Again, the order matters. Unlike VBDs,
there is no read/write read/only constraint and the devices
have unique indices (0, 1, 2, …) but Linux kernels have often (always?)
ignored the actual index and instead relied on the order of results from the
xenstore-ls listing. The order that xenstored returns the items happens
to be the order the nodes were created so this means that (i) xenstored must
continue to store directories as ordered lists rather than maps (which would
be more efficient); and (ii) Xenopsd must make sure to plug the vifs in
the same order. Note that relying on ethX device numbering has always been a
bad idea but is still common. I bet if you change this lots of tests will
suddenly start to fail!
The function VIF.plug_exn will
Similarly to the VBD case, the function Device.Vif.add will write the Xenstore keys and wait for the “hotplug-status=connected” key. We do this because we cannot apply the port locking rules until the backend device has been created, and we cannot know the rules have been applied until after the udev script has written the key. If we didn’t wait for it then the VM might execute without all the port locking properly configured.
The VM_create_device_model micro-op will create a qemu device model if
The function VM.create_device_model_exn will
Note that qemu (aka the “device model”) is created after the VIFs and VBDs have been plugged but before the PCI devices have been plugged. Unfortunately qemu traditional infers the needed emulated hardware by inspecting the Xenstore VBD and VIF configuration and assuming that we want one emulated device per PV device, up to the natural limits of the emulated buses (i.e. there can be at most 4 IDE devices: {primary,secondary}{master,slave}). Not only does this create an ordering dependency that needn’t exist – and which impacts migration downtime – but it also completely ignores the plain fact that, on a Xen system, qemu can be in a different domain than the backend disk and network devices. This hack only works because we currently run everything in the same domain. There is an option (off by default) to list the emulated devices explicitly on the qemu command-line. If we switch to this by default then we ought to be able to start up qemu early, as soon as the domain has been created (qemu will need to know the domain id so it can map the I/O request ring).
PCI devices are treated differently to VBDs and VIFs. If we are attaching the device to an HVM guest then instead of relying on the traditional Xenstore frontend/backend state machine we instead send RPCs to qemu requesting they be hotplugged. Note the domain is paused at this point, but qemu still supports PCI hotplug/unplug. The reasons why this doesn’t follow the standard Xenstore model are known only to the people who contributed this support to qemu. Again the order matters because it determines the position of the virtual device in the VM.
Note that Xenopsd doesn’t know anything about the PCI devices; concepts such as “GPU groups” belong to higher layers, such as xapi.
A design principle of Xenopsd is that it should tolerate failures such as being
suddenly restarted. It guarantees to always leave the system in a valid state,
in particular there should never be any “half-created VMs”. We achieve this for
VM start by exploiting the mechanism which is necessary for reboot. When a VM
wishes to reboot it causes the domain to exit (via SCHEDOP_shutdown) with a
“reason code” of “reboot”. When Xenopsd sees this event VM_check_state
operation is queued. This operation calls
VM.get_domain_action_request
to ask the question, “what needs to be done to make this VM happy now?”. The
implementation checks the domain state for shutdown codes and also checks a
special Xenopsd Xenstore key. When Xenopsd creates a Xen domain it sets this
key to “reboot” (meaning “please reboot me if you see me”) and when Xenopsd
finishes starting the VM it clears this key. This means that if Xenopsd crashes
while starting a VM, the new Xenopsd will conclude that the VM needs to be rebooted
and will clean up the current domain and create a fresh one.
A Xenopsd VM.start will always leave the domain paused, so strictly speaking this is a separate “operation” queued by the client (such as xapi) after the VM.start has completed. The function VM.unpause is reassuringly simple:
if di.Xenctrl.total_memory_pages = 0n then raise (Domain_not_built);
Domain.unpause ~xc di.Xenctrl.domid;
Opt.iter
(fun stubdom_domid ->
Domain.unpause ~xc stubdom_domid
) (get_stubdom ~xs di.Xenctrl.domid)