* prevent connection write collisions
When multiple calls are made on the same connection, it's possible for
the writes to collide with each other. This adds a write mutex when
communicating with the libvirt daemon.
libvirt follows a convention where structure types used in its API are
named "remote_nonnull_domain", and optional values for those structures
are called "remote_domain". The generator was translating these into go
names like "NonnullDomain" and "Domain". In go this seems unnatural, and
doesn't match the way the pre-generator version of go-libvirt named
things. So this commit changes the names: "remote_nonnull_domain" will
now be "Domain" in go; "remote_domain" will be "OptDomain". This pattern
is applied to all types.
- Update libvirt.go so that all libvirt calls now go through the
generated routines.
- Remove some libvirt routines that had the same name as generated ones,
leave the rest as convenience routines.
- Fix the handling of Optional-values (the declarations of which in the
.x file look like pointers)
The generated wrappers have an argument for every field in their "Args"
struct, and return everything in their "Ret" struct (these structs are
defined in the protocol file, and identified by procedure name).
Marshaling and unmarshaling is handled inside the generated procedures.
- Add a yacc-based parser and a hand-written lexer to read the
remote_protocol.x file from libvirt's sources.
- Use the new parser to generate the constants used to communicate with
libvirt.
* Handle TypedParamStrings
Add handling for TypedParamStrings, which are different from the other
TypedParam... types in that the decoded value is variable-sized.
* Add Get/SetBlockIoTune to go-libvirt API.
This adds two libvirt entry points to the go-libvirt API:
virDomainSetBlockIoTune and virDomainGetBlockIoTune. These can be used
to control block device throttling for a VM.
This adds a new method, Capabilities(), which returns an XML
definition of the hypervisor's system capabilities, much like
`virsh capabilities`.
I've only included integration tests for these two methods due to the
large amount of data returned by the capabilities call.
This adds support for dumping a domain's XML definition, akin to `virsh
dumpxml <domain>`. The returned data is quite large so I've included an
integration test rather than adding a huge blob of hex to `libvirttest`.
Further testing reveals we need some additional padding on the request
passed to libvirt when performing migrations. I believe this is due to
the structure alignment normally provided automatically by the compiler.
In testing, this fixes an error I was seeing: "Unable to decode message
paylaod", returned from libvirt when performing migrations.
This adds basic support for domain migrations from one hypervisor to
another. Migration options, e.g., live, tunneled, compressed, etc..,
are specified by the constants described `constants.Migrate*`.
Two unknowns remain, Libvirt specifies `RemoteParameters` and `CookieIn`.
In testing both values are always set to 0 by `virsh` and the source
does not provide clear definitions of their purpose. For now, using
the same zero'd values used by `virsh` will be Good Enough.
When performing QEMU monitor commands, libvirt will return StatusOK even
when the underlying QEMU process fails to perform the command.
This modifies Run() to check for QEMU errors.
I'm not entirely happy with the hacky modifications to the test library
to handle this scenario. The test framework is in obvious need for a
complete refactor. For now this will have to work.
This splits out the mock libvirt server for testing within
other packages. Constants from the main libvirt package
have been moved to a separate package, constants, for shared access.
