[libcamera-devel] versioning

Sun Aug 14 23:44:05 CEST 2022

Hi Laurent,

I think the SONAME is not the primary reason for tagging/versioning.
Having API versions is mainly used to check compatibility for other
projects. That means that the version information has to be encoded in
some project file, e.g. *.pc files. Using a SONAME is really just useful
if you want to have multiple versions of a library installed or if you
want to make sure that you only link a specific version.

Best,
Christian

Am 14.08.22 um 21:01 schrieb Laurent Pinchart:
> Hello,
>
> On Sun, Aug 14, 2022 at 06:01:43PM +0100, Kieran Bingham via libcamera-devel wrote:
>> Quoting Christian Rauch (2022-08-14 17:25:48)
>>> Am 14.08.22 um 16:12 schrieb Kieran Bingham:
>>>> Quoting Christian Rauch via libcamera-devel (2022-08-13 22:14:44)
>>>>> Hi all,
>>>>>
>>>>> I am aware that the libcamera API is unstable and because of this, there
>>>>> is no versioning. I am using libcamera in another project, and it would
>>>>> be helpful if I could check some version constraints in order to prevent
>>>>> compiling with an incompatible API.
>>>>
>>>> I've been long advocating that we should have a version number, even
>>>> with an unstable API. The last time I tried to push on this topic we
>>>> added 0.0.0 at the very first commit which gives us some degree of
>>>> patchlevel versioning for every commit.  It has taken a long time to
>>>> identify clear reasons /why/ we should have more specific versions, when
>>>> we have a clear unstable API, and we refer to the latest.
>>>>
>>>> However as you have reported here, with more API breakages recently -
>>>> having other pacakges compiling against libcamera, they likely want to
>>>> check what is installed. It's not always under the control and can
>>>> depend upon the installation of the distro.
>>>>
>>>>> Would you be willing to start versioning the minor or patch-level of
>>>>> libcamera? I imagine that the major version will stay at "0" until the
>>>>
>>>> I believe the answer is 'yes' now. But we still need to implement the
>>>> 'how'.
>>>>
>>>> Personally - I would like to use something like abi-compliance-checker
>>>> to automatically detect any ABI/API change, and automatically increment
>>>> the minor patch level. libabigail looks interesting and relevant too:
>>>> https://developers.redhat.com/blog/2020/04/02/how-to-write-an-abi-compliance-checker-using-libabigail#
>>>>
>>>> That would give us a version scheme of:
>>>>
>>>>  0.a[bp]i.patch
>>>>
>>>> Where other packages and applications could then use the second value to
>>>> determine if they are compatible, (or to make conditional changes to be
>>>> compilation compatible).
>>>>
>>>> The patch level of course would just be the number of patches since that
>>>> change, and the leading version 0, would not be incremented until we
>>>> were at a point that we could be comfortable stating we are 'ABI stable'.
>>>>
>>>>> API is stable, but that either the minor or patch-level could be
>>>>> incremented (0.1, 0.2, ...) every time the API changes, or simply on a
>>>>> monthly or quarterly basis.
>>>>
>>>> Every time I've suggested we should start making release versions, I
>>>> always face push back of "What decrees a release, when should it be
>>>> made?".
>>>>
>>>> To me time based milestones are acceptable, but don't make it
>>>> easy/straightforwards to perform the version specific compilation
>>>> changes required.
>>>>
>>>> I think they help users understand how new or old the library is though,
>>>> and at least that is more helpful than our current versioning scheme
>>>> where comparing 074fa98ac4e against 18d61deb3c0 is just not human
>>>> readable, and doesn't convey any useful information. We have tagged our
>>>> first commit as 0.0.0 which does now at least provide version strings
>>>> generated by utils/gen-version.sh:
>>>>
>>>>  - 0.0.0+2819-18d61deb
>>>>  - 0.0.0+3459-074fa98a
>>>>
>>>> So we could already say we could (manually) identify API/ABI changes
>>>> between the patch level numbers after the + (2819, to 3459) - but that's
>>>> quite labour intensive otherwise and prone to error, which is why I'd
>>>> really like to see API/ABI breakages identifiable through something like
>>>>
>>>>  - 0.25.245-18d61deb
>>>>  - 0 29.15-074fa98a
>>>>
>>>> Which we can then parse and extract more useful information from.
>>>>
>>>> I would envisage the automatic abi checker to be able to provide the
>>>> 0.<abi> tags automatically on every change if run on every commit (and
>>>> could be back dated right to the beginning of the tree history) ...
>>>> however the point I have difficulty with is how we get those tags to
>>>> then convey the correct information about the release in external
>>>> systems.
>>>>
>>>> Is a 'tag' sufficient? or do we have to (for every 'release') manually
>>>> edit the meson.build version string, or make other manual explicit
>>>> changes.
>>>>
>>>> If we can do something around simply adding a tag on every abi/api
>>>> version change, - I think that would provide the cleanest/simplest way
>>>> forwards for now.
>>>>
>>>> Any ideas or extended comments?
>
> As I see it, the main reason to tag releases, from a distribution point
> of view, is to increase the SONAME of the library. This will prevent
> binaries linked against one version of libcamera to run with an
> incompatible version. This can probably be done without modifying
> meson.build, based on a similar mechanism as gen-version.sh. The project
> version (in the root meson.build file) and the libcamera shared library
> version can also be generated through a similar mechanism, using
> run_command() directly in the project() function call for the former as
> the project() function must be called before anything else. See
> https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/meson.build for
> instance.
>
> Until we tag v1.0, we should use the ABI version as the SONAME. Ideally,
> I'd like to use libcamera.so.0.X as the SONAME for now, not
> libcamera.so.X, to allow using libcamera.so.1 once we release v1.0. I
> don't know if that's a valid SONAME though.
>
> Automatic ABI breakage checks is nice to have, but I don't think it's a
> strict requirement to tag anything. We can start with a time-based
> tagging scheme (weekly for instance), with v0.0.X where X will be an
> increasing index. Once ABI checks are in place, this could become
> v0.X.Y. The exact versioning scheme doesn't matter much to me, as long
> as it doesn't corner us for the v1.0 release.
>
> Kieran, if you'd like to implement now a weekly release scheme that
> would use the latest tag or the .tarball-version file to populate the
> project version, library version and SONAME using a v0.0.X versioning
> scheme without ABI checking, I'll review the patches.
>
>>> I have seen CMake projects generating the version information from git.
>>> This somehow works that a git command is executed as part of the build
>>> process and then the stdout is parsed. Of course, this only works when
>>> the source is checked out as a git repo. But I think something similar
>>> could be done in meson. Then, you only need to create a new git tag and
>>
>> We already do this.
>>
>> https://git.libcamera.org/libcamera/libcamera.git/tree/meson.build#n13
>>
>> meson calls utils/gen-version.sh which generates a version string such
>> as :
>>
>>   0.0.0+3829-f88d73af
>>
>> But it only works when the source is checked out as a git repo indeed.
>>
>> How does that compare for your current needs against the discussions
>> above?
>>
>>> the next time libcamera is compiled, it would generate the version.
>>>
>>>>> This way, projects could check for a specific version before compiling
>>>>> or search for a specific "API version". It wouldn't guarantee that the
>>>>> API is compatible between versions, it's just useful to know which
>>>>> version you have to use to compile a project.
>>>>
>>>> Agreed!
>>>>
>>>> It's also clear from other recent discussions that we should bump the
>>>> SONAME for every ABI breakage, so again - if this can be handled through
>>>> automatically detecting those breakages - I think that would help too.
>>>> Having the SONAME as 0.<abi-version> as above would be a great benefit
>>>> already I think.
>