This document walks you through the process of implementing a new API in libvirt. It uses as an example the addition of an API for separating maximum from current vcpu usage of a domain, over the course of a fifteen-patch series. Remember that new API consists of any new public functions, as well as the addition of flags or extensions of XML used by existing functions. The example in this document adds both new functions and an XML extension. Not all libvirt API additions require quite as many patches.
Before you begin coding, it is critical that you propose your changes on the libvirt mailing list and get feedback on your ideas to make sure what you're proposing fits with the general direction of the project. Even before doing a proof of concept implementation, send an email giving an overview of the functionality you think should be added to libvirt. Someone may already be working on the feature you want. Also, recognize that everything you write is likely to undergo significant rework as you discuss it with the other developers, so don't wait too long before getting feedback. In the vcpu example below, list feedback was first requested here and resulted in several rounds of improvements before coding began. In turn, this example is slightly rearranged from the actual order of the commits.
Adding a new API to libvirt is not difficult, but there are quite a few steps. This document assumes that you are familiar with C programming and have checked out the libvirt code from the source code repository and successfully built the existing tree. Instructions on how to check out and build the code can be found at:
http://libvirt.org/downloads.html
Once you have a working development environment, the steps to create a new API are:
It is, of course, possible to implement the pieces in any order, but if the development tasks are completed in the order listed, the code will compile after each step. Given the number of changes required, verification after each step is highly recommended.
Submit new code in the form shown in the example code: one patch per step. That's not to say submit patches before you have working functionality--get the whole thing working and make sure you're happy with it. Then use git or some other version control system that lets you rewrite your commit history and break patches into pieces so you don't drop a big blob of code on the mailing list in one go. Also, you should follow the upstream tree, and rebase your series to adapt your patches to work with any other changes that were accepted upstream during your development.
Don't mix anything else into the patches you submit. The patches should be the minimal changes required to implement the functionality you're adding. If you notice a bug in unrelated code (i.e., code you don't have to touch to implement your API change) during development, create a patch that just addresses that bug and submit it separately.
With that said, let's begin.
The first task is to define the public API. If the new API involves an XML extension, you have to enhance the RelaxNG schema and document the new elements or attributes:
docs/schemas/domain.rng
docs/formatdomain.html.in
If the API extension involves a new function, you have to add a declaration in the public header, and arrange to export the function name (symbol) so other programs can link against the libvirt library and call the new function:
include/libvirt/libvirt.h.in
src/libvirt_public.syms
This task is in many ways the most important to get right, since once the API has been committed to the repository, it's libvirt's policy never to change it. Mistakes in the implementation are bugs that you can fix. Make a mistake in the API definition and you're stuck with it, so think carefully about the interface and don't be afraid to rework it as you go through the process of implementing it.
See 0001-add-to-xml.patch and 0002-add-new-public-API.patch for example code.
Each public API call is associated with a driver, such as a host virtualization driver, a network virtualization driver, a storage virtualization driver, a state driver, or a device monitor. Adding the internal API is ordinarily a matter of adding a new member to the struct representing one of these drivers.
Of course, it's possible that the new API will involve the creation of an entirely new driver type, in which case the changes will include the creation of a new struct type to represent the new driver type.
The driver structs are defined in:
src/driver.h
To define the internal API, first typedef the driver function
prototype and then add a new field for it to the relevant driver
struct. Then, update all existing instances of the driver to
provide a NULL stub for the new function.
See 0003-define-internal-driver-API.patch
Implementing the public API is largely a formality in which we wire up public API to the internal driver API. The public API implementation takes care of some basic validity checks before passing control to the driver implementation. In RFC 2119 vocabulary, this function:
The public API calls are implemented in:
src/libvirt.c
See 0004-implement-the-public-APIs.patch
Implementing the remote protocol is essentially a straightforward exercise which is probably most easily understood by referring to the existing code and the example patch. It involves several related changes, including the regeneration of derived files, with further details below.
See 0005-implement-the-remote-protocol.patch
Defining the wire protocol involves making additions to:
src/remote/remote_protocol.x
First, create two new structs for each new function that you're adding to the API. One struct describes the parameters to be passed to the remote function, and a second struct describes the value returned by the remote function. The one exception to this rule is that functions that return only 0 or -1 for status do not require a struct for returned data.
Second, add values to the remote_procedure enum for each new function added to the API.
Once these changes are in place, it's necessary to run 'make rpcgen' in the src directory to create the .c and .h files required by the remote protocol code. This must be done on a Linux host using the GLibC rpcgen program. Other rpcgen versions may generate code which results in bogus compile time warnings. This regenerates the following files:
daemon/remote_dispatch_args.h
daemon/remote_dispatch_prototypes.h
daemon/remote_dispatch_table.h
src/remote/remote_protocol.c
src/remote/remote_protocol.h
Implementing the uses the rpcgen generated .h files. The remote method calls go in:
src/remote/remote_internal.c
Each remote method invocation does the following:
Implementing the server side of the remote function call is simply a matter of deserializing the parameters passed in from the remote caller and passing them to the corresponding internal API function. The server side dispatchers are implemented in:
daemon/remote.c
Again, this step uses the .h files generated by make rpcgen.
After all three pieces of the remote protocol are complete, and the generated files have been updated, it will be necessary to update the file:
src/remote_protocol-structs
This file should only have new lines added; modifications to existing lines probably imply a backwards-incompatible API change.
See 0005-implement-the-remote-protocol.patch
Sometimes, a new API serves as a superset of existing API, by adding more granularity in what can be managed. When this is the case, it makes sense to share a common implementation by making the older API become a trivial wrapper around the new API, rather than duplicating the common code. This step should not introduce any semantic differences for the old API, and is not necessary if the new API has no relation to existing API.
See 0006-make-old-API-trivially-wrap-to-new-API.patch
All new API should be manageable from the virsh command line shell. This proves that the API is sufficient for the intended purpose, and helps to identify whether the proposed API needs slight changes for easier usage. However, remember that virsh is used to connect to hosts running older versions of libvirtd, so new commands should have fallbacks to an older API if possible; implementing the virsh hooks at this point makes it very easy to test these fallbacks. Also remember to document virsh additions.
A virsh command is composed of a few pieces of code. You need to define an array of vshCmdInfo structs for each new command that contain the help text and the command description text. You also need an array of vshCmdOptDef structs to describe the command options. Once you have those pieces in place you can write the function implementing the virsh command. Finally, you need to add the new command to the commands[] array. The following files need changes:
tools/virsh.c
tools/virsh.pod
See 0007-add-virsh-support.patch
So, after all that, we get to the fun part. All functionality in libvirt is implemented inside a driver. Thus, here is where you implement whatever functionality you're adding to libvirt. You'll either need to add additional files to the src directory or extend files that are already there, depending on what functionality you're adding.
If the new API is applicable to more than one driver, it may make sense to provide some utility routines, or to factor some of the work into the dispatcher, to avoid reimplementing the same code in every driver. In the example code, this involved adding a member to the virDomainDefPtr struct for mapping between the XML API addition and the in-memory representation of a domain, along with updating all clients to use the new member. Up to this point, there have been no changes to existing semantics, and the new APIs will fail unless they are used in the same way as the older API wrappers.
See 0008-support-new-xml.patch
The remaining patches should only touch one driver at a time. It is possible to implement all changes for a driver in one patch, but for review purposes it may still make sense to break things into simpler steps. Here is where the new APIs finally start working.
In the example patches, three separate drivers are supported: test, qemu, and xen. It is always a good idea to patch the test driver in addition to the target driver, to prove that the API can be used for more than one driver. The example updates the test driver in one patch:
See 0009-support-all-flags-in-test-driver.patch
The qemu changes were easier to split into two phases, one for updating the mapping between the new XML and the hypervisor command line arguments, and one for supporting all possible flags of the new API:
See 0010-improve-vcpu-support-in-qemu-command-line.patch and 0011-complete-vcpu-support-in-qemu-driver.patch
Finally, the example breaks the xen driver changes across four patches. One maps the XML changes to the hypervisor command, the next two are independently implementing the getter and setter APIs, and the last one provides cleanup of code that was rendered dead by the new API.
See 0012-improve-vcpu-support-in-xen-command-line.patch, 0013-improve-getting-xen-vcpu-counts.patch, 0014-improve-setting-xen-vcpu-counts.patch, and 0015-remove-dead-xen-code.patch
The exact details of the example code are probably uninteresting unless you're concerned with virtual cpu management.
Once you have working functionality, run make check and make syntax-check on each patch of the series before submitting patches. It may also be worth writing tests for the libvirt-TCK testsuite to exercise your new API, although those patches are not kept in the libvirt repository.