Walkthrough: Starting a VM

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:

  • VM: A device-less Virtual Machine
  • VBD: A virtual block device for a VM
  • VIF: A virtual network interface for a VM
  • PCI: A virtual PCI device for a VM

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 XenAPI has many clients which are updated on long release cycles. The main property needed is backwards compatibility, so that new release of xapi remain compatible with these older clients. Quite often we will chose to “grandfather in” some poorly designed interface simply because we wish to avoid imposing churn on 3rd parties.
  • the Xenopsd API clients are all open-source and are part of the xapi-project. These clients can be updated as the API is changed. The main property needed is to keep the interface clean, so that it properly hides the complexity of dealing with Xen from other components.

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.id

This function does 2 things:

  • it stores the VM configuration in the “database”
  • it tells the “backend” that the VM exists

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:

  • Xen via libxc, libxenguest and xenstore
  • Xen via libxl, libxc and xenstore
  • Xen via libvirt
  • KVM by direct invocation of qemu
  • Simulation for testing

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:

  • the power state is Paused
  • exactly one valid Xen domain id
  • all VBDs have active = plugged = true
  • all VIFs have active = plugged = true
  • all PCI devices have plugged = true
  • at least one active console
  • a valid start time
  • valid “targets” for memory and vCPU

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);
	task

The “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:

  • pushing operations to the right queue
  • ensuring at most one worker thread is working on a VM’s operations
  • reducing the queue size by coalescing items together
  • providing a diagnostics interface

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 id

Each “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:

1. run the “VM_pre_start” scripts

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.

2. create a Xen domain

The VM_create micro-op calls the VM.create function in the backend. In the classic Xenopsd backend the VM.create_exn function must

  1. check if we’re creating a domain for a fresh VM or resuming an existing one: if it’s a resume then the domain configuration stored in the VmExtra database table must be used
  2. ask squeezed to create a memory “reservation” big enough to hold the VM memory. Unfortunately the domain cannot be created until the memory is free because domain create often fails in low-memory conditions. This means the “reservation” is associated with our “session” with squeezed; if Xenopsd crashes and restarts the reservation will be freed automatically.
  3. create the Domain via the libxc hypercall
  4. “transfer” the squeezed reservation to the domain such that squeezed will free the memory if the domain is destroyed later
  5. compute and set an initial balloon target depending on the amount of memory reserved (recall we ask for a range between dynamic_min and dynamic_max)
  6. apply the “suppress spurious page faults” workaround if requested
  7. set the “machine address size”
  8. “hotplug” the vCPUs. This operates a lot like memory ballooning – Xen creates lots of vCPUs and then the guest is asked to only use some of them. Every VM therefore starts with the “VCPUs_max” setting and co-operative hotplug is used to reduce the number. Note there is no enforcement mechanism: a VM which cheats and uses too many vCPUs would have to be caught by looking at the performance statistics.

3. build the domain

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

  1. run pygrub (or eliloader) to extract the kernel and initrd, if necessary
  2. invoke the xenguest binary to interact with libxenguest.
  3. apply the cpuid configuration
  4. store the current domain configuration on disk – it’s important to know the difference between the configuration you started with and the configuration you would use after a reboot because some properties (such as maximum memory and vCPUs) as fixed on create.

The 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

  • the vCPU affinity
  • the vCPU credit2 weight/cap parameters
  • whether the NX bit is exposed
  • whether the viridian CPUID leaf is exposed
  • whether the system has PAE or not
  • whether the system has ACPI or not
  • whether the system has nested HVM or not
  • whether the system has an HPET or not

4. mark each VBD as “active”

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.

5. handle non-persistent disks

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.

6. plug VBDs

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

  • call VDI.attach and VDI.activate in the storage API to make the devices ready (start the tapdisk processes etc)
  • add the Xenstore frontend/backend directories containing the block device info
  • add the extra xenstore keys returned by the VDI.attach call that are needed for SCSIid passthrough which is needed to support VSS
  • write the VBD information to the Xenopsd database so that future calls to VBD.stat can be told about the associated disk (this is needed so clients like xapi can cope with CD insert/eject etc)
  • if the qemu is going to be in a different domain to the storage, a frontend device in the qemu domain is created.

The 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.

7. mark each VIF as “active”

This is for the same reason as VBDs are marked “active”.

8. plug VIFs

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

  • compute the port locking configuration required and write this to a well-known location in the filesystem where it can be read from the udev scripts. This really should be written to Xenstore instead, since this scheme doesn’t work with driver domains.
  • add the Xenstore frontend/backend directories containing the network device info
  • write the VIF information to the Xenopsd database so that future calls to VIF.stat can be told about the associated network
  • if the qemu is going to be in a different domain to the storage, a frontend device in the qemu domain is created.

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.

9. create the device model

The VM_create_device_model micro-op will create a qemu device model if

  • the VM is HVM; or
  • the VM uses a PV keyboard or mouse (since only qemu currently has backend support for these devices).

The function VM.create_device_model_exn will

  • (if using a qemu stubdom) it will create and build the qemu domain
  • compute the necessary qemu arguments and launch it.

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).

10. plug PCI devices

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.

11. mark the domain as alive

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.

12. unpause the domain

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)