libvirt architecture

Currently libvirt supports 2 kind of virtualization, and its internal structure is based on a driver model which simplifies adding new engines:

Xen support

When running in a Xen environment, programs using libvirt have to execute in "Domain 0", which is the primary Linux OS loaded on the machine. That OS kernel provides most if not all of the actual drivers used by the set of domains. It also runs the Xen Store, a database of information shared by the hypervisor, the kernels, the drivers and the xen daemon. Xend. The xen daemon supervise the control and execution of the sets of domains. The hypervisor, drivers, kernels and daemons communicate though a shared system bus implemented in the hypervisor. The figure below tries to provide a view of this environment:

The library can be initialized in 2 ways depending on the level of privilege of the embedding program. If it runs with root access, virConnectOpen() can be used, it will use three different ways to connect to the Xen infrastructure:

• a connection to the Xen Daemon though an HTTP RPC layer
• a read/write connection to the Xen Store
• use Xen Hypervisor calls
• when used as non-root libvirt connect to a proxy daemon running as root and providing read-only support

The library will usually interact with the Xen daemon for any operation changing the state of the system, but for performance and accuracy reasons may talk directly to the hypervisor when gathering state information at least when possible (i.e. when the running program using libvirt has root privilege access).

If it runs without root access virConnectOpenReadOnly() should be used to connect to initialize the library. It will then fork a libvirt_proxy program running as root and providing read_only access to the API, this is then only useful for reporting and monitoring.

QEmu and KVM support

The model for QEmu and KVM is completely similar, basically KVM is based on QEmu for the process controlling a new domain, only small details differs between the two. In both case the libvirt API is provided by a controlling process forked by libvirt in the background and which launch and control the QEmu or KVM process. That program called libvirt_qemud talks though a specific protocol to the library, and connects to the console of the QEmu process in order to control and report on its status. Libvirt tries to expose all the emulations models of QEmu, the selection is done when creating the new domain, by specifying the architecture and machine type targeted.

The code controlling the QEmu process is available in the qemud/ directory.

Driver based architecture

As the previous section explains, libvirt can communicate using different channels with the current hypervisor, and should also be able to use different kind of hypervisor. To simplify the internal design, code, ease maintenance and simplify the support of other virtualization engine the internals have been structured as one core component, the libvirt.c module acting as a front-end for the library API and a set of hypervisor drivers defining a common set of routines. That way the Xen Daemon access, the Xen Store one, the Hypervisor hypercall are all isolated in separate C modules implementing at least a subset of the common operations defined by the drivers present in driver.h:

• xend_internal: implements the driver functions though the Xen Daemon
• xs_internal: implements the subset of the driver available though the Xen Store
• xen_internal: provide the implementation of the functions possible via direct hypervisor access
• proxy_internal: provide read-only Xen access via a proxy, the proxy code is in the proxy/ directory.
• xm_internal: provide support for Xen defined but not running domains.
• qemu_internal: implement the driver functions for QEmu and KVM virtualization engines. It also uses a qemud/ specific daemon which interacts with the QEmu process to implement libvirt API.
• test: this is a test driver useful for regression tests of the front-end part of libvirt.

Note that a given driver may only implement a subset of those functions, (for example saving a Xen domain state to disk and restoring it is only possible though the Xen Daemon), in that case the driver entry points for unsupported functions are initialized to NULL.