AToL: LibVirt API

27.6.2012 09:45 | Přečteno: 3930× | Linux

Autor: Daniel Vrátil
Nasledovný príspevok je študentskou prácou, ktorá vznikla v rámci predmetu Advanced Topics of Linux Administration. Predmet je vypisovaný na Fakulte informatiky MU v spolupráci so spoločnosťou Red Hat Czech. Vyučovacím jazykom je angličtina a preto je v nej aj nasledovný príspevok, študent nemá právo výberu iného jazyka. Komentáre k obsahu príspevku sú vítané v ľubovoľnom zrozumiteľnom jazyku.

This article aims to provide introduction to programmatic use of Libvirt through it's C API.

Libvirt is a toolkit that allows programmers to control virtual machines running under different hypervisors through common API. At this moment Libvirt supports QEMU, Xen, LXC, OpenVZ, UML, VMWare and Microsoft Hyper-V.

The design is that application connects through libvirt to a libvirtd deamon (local or remote), which controls it's local (from daemon's point of view) virtual machines through libvirt. Libvirt itself consists of the user-side API and backends, which essentially translate the libvirt API calls into hypervisor-specific commands.

A lot has been written about libvirt(d) and how to install, configure it and how to use it's management tools, but less about how programmers can utilize it. Simple management tasks can be done through console applications like virsh (which supports both interactive and non-interactive mode). But sometimes programmer or administrator might want to watch a libvirtd server and be notified about events. This cannot be achieved through the console application and one must resort to use of Libvirt API.

When it comes to programming, hands-on approach is always better then cheap talks. I will demonstrate use of Libvirt's API on an application I have written for this article. Libvirt officially offers native C API and Python bindings, but 3^rd party bindings for Ruby, PHP, Perl, Java, C# and OCaml are available too. I have decided to use the native C API and write a KDE Plasma Applet and DataEngine in C++/Qt to demonstrate use of Libvirt in event-driven environment.

I will not go into much detail of the Plasma-related code, there are numerous articles about it, including this serial (in Czech) here on ABCLinuxu. The Plasma applet is written completely in QML and is there only to visualize data, the libvirt interaaction is implemented in a DataEgine.

The DataEngine is completely written in C++. It is a sort of a backend where each virtual machine is provided as a Source. Visualizers (like the applet) can connect to one or more sources and poll them for changes or let themselves be notified about changes. Each Source is essentially a hash map property => value. Every Source can also have so called services which visualizers can invoke to let the DataEngine perform an action or change a property of a Source.

Libvirt offers two ways how to be notified about changes. Either by polling or by event notifications. I have decided to use event notifications, because polling causes unnecessary load on server and network. Libvirt comes with it's own implementation of an event loop, so that it can easily be used in applications without any other event-based framework (Qt, Glib, ...). The elementary use is same as with any other event-based application:

int main(int argc, char **argv)
{
  /* initialize libvirt */
  virInitialize(); 

  /* connect to server */
  virConnectPtr conn = virConnectOpen(....);

  /* initialize event loop */
  virEventRegisterDefaultImpl();

  /* connect callbacks to events */
  ....

  /* process events in a loop */
  quit = false;
  while (!quit) {
    virEventRunDefaultImpl();
  }

  /* disconnect callbacks */
  ....

  /* disconnect from server */
  virConnectClose(conn);

  return 0;
}

Problem arises if you try to use Libvirt's events implementation together with an another event-based framework. For this situation Libvirt offers virEventRegisterImpl() to which user can pass callbacks for registering, modifying and removing handlers and timeouts. User then have to implement handling of the handlers and timeouts in the other event loop. In Qt, it would be possible to reimplement QAbstractEventDispatcher to do this. Unfortunately DataEngines are just a plugins run by Plasma desktop and we can't change the event dispatcher on runtime. An alternative approach if you want to save some work, is to use the original virEventRegisterDefaultImpl() in separate thread.

(libvirteventloop.cpp)

void LibvirtEventLoop::run()
{
    /* Initialize libvirt event loop */
    virEventRegisterDefaultImpl();

    /* Setup separate connection */
    virConnectPtr connection = virConnectOpen("qemu:///system");

    /* Make sure the connection stays alive in case there were no events for some time */
    if (virConnectSetKeepAlive(connection, 5, 3) < 0) {
        virConnectClose(connection);
        return;
    }

    /* Connect event handlers here */
    .....

    m_stopMutex.lock();
    while (!m_stop) {
        m_stopMutex.unlock();


        /* Die another day */
        if (virConnectIsAlive(connection) != 1)
            break;

        /* Process all queued libvirt events */
        if (virEventRunDefaultImpl() < 0) {
            m_stop = true;
        }

        m_stopMutex.lock();
    }
    m_stopMutex.unlock();

    /* Disconnect event handlers here */
    ....

    /* Disconnect from server */
    virConnectClose(connection);
}

The thread has it's own connection to server and will notify the main thread (which has it's own connection too) through standard Qt's slot/signal mechanism. The main thread then forces respective Source to update and notifies visualizers (see LibvirtEngine::eventLoopEvent()).

To connect to event handlers, there are virConnectDomainEventRegister() and virConnectDomainEventRegisterAny() functions. The first one notifies only about changes in domain (Libvirt term for a virtual machine) lifecycle. The latter allows users to specify which type of events they want to listen to and to specify domain for which they want to receive events (otherwise events for all domains are delivered).

Now when we are notified about events, we need to actually get a list of the domains that are available. The DataEngine has hardcoded connection to qemu:///system, which is a connection to global session of local libvirtd. The protocol specifies hypervisor type. To obtain list of all available domains, one must query active (running) and inactive domains separately. It's also worth noting, that domains in Libvirt can be uniquely identified by their name (though that can change during runtime) and UUID (persistent). Only running domains have ID. virDomainGetID() for inactive domain returns -1. The ID is incremental, but it reset to zero after every libvirtd restart.

(LibvirtEngine::init())

    /* Get number of active and inactive domains */
    int numActiveDomains = virConnectNumOfDomains(m_connection);
    int numInactiveDomains = virConnectNumOfDefinedDomains(m_connection);

    char **inactiveDomains = (char **) malloc(sizeof(char *) * numInactiveDomains);
    int *activeDomains = (int *) malloc(sizeof(int) * numActiveDomains);

    /* Get IDs of active domains and names of inactive domains */
    numActiveDomains = virConnectListDomains(m_connection, activeDomains, numActiveDomains);
    numInactiveDomains = virConnectListDefinedDomains(m_connection, inactiveDomains, numInactiveDomains);

    for (int i = 0 ; i < numActiveDomains ; i++) {
        /* Lookup the actual domain object by it's ID */
        virDomainPtr domain = virDomainLookupByID(m_connection, activeDomains[i]);
        /* Create a new Plasma DataContainer for each domain */
        addSource(new LibvirtSource(domain, m_connection, this));
    }

    for (int i = 0 ; i < numInactiveDomains ; i++) {
        /* Lookup the actual domain by it's name */
        virDomainPtr domain = virDomainLookupByName(m_connection, inactiveDomains[i]);
        addSource(new LibvirtSource(domain, m_connection, this));
        free(inactiveDomains[i]);
    }

The code above creates a new DataContainer (Source) for every domain. The sources are then exposed to visualizers via addSource().

The next task is to populate the sources with data from domains. We do it in LibvirtSource::dataUpdateRequested() slot, which is called whenever anyone wants the Source to update. Elementary domain information are simple to retrieve, there are dedicated functions for it, like virDomainGetName(), virDomainGetID() or virDomainGetState(). For example memory configuration (soft and hard limit for VM memory allocation) can be obtained only from the XML description. Libvirt stores configuration of every domain in a XML file which can be retrieved via virDomainGetXMLDesc(). To obtain a user's text description of a domain, there is virDomainGetMetaData() function, but it throws a runtime warning when the domain does not have any description (quite common case), therefor it's better to manually look into the XML for <description> tag and silently continue when it's not there. The XML format is documented on Libvirt homepage.

It should be possible to query the domain for memory usage (e.g. amount of used memory within the VM), but it requires some cgroups settings which didn't work for me, so the DataEngine does not report it. It is possible to get information about CPU usage though - virDomainGetInfo() returns number of virtual CPUs allocated to the domain and total CPU time used by all vCPUs (see LibvirtSource::updateCPUUsage()).

As mentioned in the beginning of the article, every DataEngine Source can also have a Service. Our DataEngine have a few simple services for changing state of a domain and it's memory size and CPU count.

To change a domain state, there is a specific method for every state to be invoked:

• Start: virDomainCreate(), virDomainCreateWithFlags(). Note that virDomainCreateXML() will create and start a new temporary domain (a full XML description must be provided) which will be deleted upon shutdown.
• Pause: virDomainSuspend()
• Unpause: virDomainResume()
• Reboot: virDomainReboot()
• Shut off: virDomainShutdown(), virDomainShutdownFlags() can specify whether to send an ACPI event or initiate shutdown through guest agent
• Force off: virDomainDestroy()
• Create image: virDomainSave() will save the entire domain state into provided file
• Restore image: virDomainRestore() will restore domain from a file

To change number of virtual CPUs, virDomainSetVcpusFlags() is used. The flags specify whether the change should affect only current session (if the domain is active at the moment) or whether it should be persistently written to domain configuration. When VIR_DOMAIN_VCPU_MAXIMUM flag is provided, the hard limit is modified. virDomainSetMemoryFlags() works analogically for domain memory allocation and the hard limit is set through virDomainSetMaxMemory().

It is of course possible to alter any other preferences. Some can be changed through dedicated API calls (for example storage pool has quite an extensive API), some only by altering domain's XML description. To update domain's XML, virDomainDefineXML() can be used. If the XML is missing UUID or has UUID of non-existent domain a new domain will be created. Finally, domains can be removed by virDomainUndefine().

The Libvirt API is much more extensive, for full description see the official API reference and the Application Development Guide. If you were interested in a more detailed use of the Libvirt API, check Virt-Manager, which is written in Python and utilizes almost everything Libvirt can offer.

To see all the code above in context and in action, you can build the example application. Tarball with sources is available here, git repository can be cloned from git://anongit.kde.org/scratch/dvratil/plasma-virt-monitor.

       

Obrázky

Komentáře

[David Rehza]

Tento článek se neslučuje s kvalitou.