[dpdk-dev] [PATCH v6 00/11] Add ABI compatibility checks to the meson build

[Bruce Richardson]

On Thu, Dec 19, 2019 at 10:58:35PM +0100, David Marchand wrote:
> Hello Kevin,
> 
> On Fri, Dec 13, 2019 at 5:41 PM Kevin Laatz <kevin.laatz at intel.com> wrote:
> >
> > With the recent changes made to stabilize ABI versioning in DPDK, it will
> > become increasingly important to check patches for ABI compatibility. We
> > propose adding the ABI compatibility checking to be done as part of the
> > build.
> >
> > The advantages to adding the ABI compatibility checking to the build are
> > two-fold. Firstly, developers can easily check their patches to make sure
> > they don’t break the ABI without adding any extra steps. Secondly, it
> > makes the integration into existing CI seamless since there are no extra
> > scripts to make the CI run. The build will run as usual and if an
> > incompatibility is detected in the ABI, the build will fail and show the
> > incompatibility. As an added bonus, enabling the ABI compatibility checks
> > does not impact the build speed.
> >
> > The proposed solution works as follows:
> > 1. Generate the ABI dump of the baseline. This can be done with the new
> >    script added in this set. This step will only need to be done when the
> >    ABI version changes (so once a year) and can be added to master so it
> >    exists by default. This step can be skipped if the dump files for the
> >    baseline already exist.
> > 2. Build with meson. If there is an ABI incompatibility, the build will
> >    fail and print the incompatibility information.
> >
> > The patches in this set include the ABI dump file generating script, the
> > dump files for drivers and libs, the meson option required to
> > enable/disable the checks, and the required meson changes to run the
> > compatibility checks during the build.
> >
> > Note: This patch set depends on: http://patches.dpdk.org/patch/63765/.  The
> > generated .dump files in this patch set are based on the changes in the
> > patch "build: fix soname info for 19.11 compatibility". If a decision is
> > made to use a different format for the sonames, then a new version of this
> > patch set will be required as the .dump files will need to be regenerated.
> >
> > Note: The following driver dump files are not included in these patches:
> >         common/mvep:    missing dependency, "libmusdk"
> >         net/mvneta:     missing dependency, "libmusdk"
> >         net/mvpp2:      missing dependency, "libmusdk"
> >         net/nfb:        missing dependency, "libnfb"
> >         crypto/mvsam:   missing dependency, "libmusdk"
> >
> > They have not been included as I do not have access to these dependencies.
> > Please feel free to add them if you can! (Maintainers of the above Cc'ed).
> 
> - I asked for the dump files, but I can see that it is impractical.
> 
> The dump files are huge. I did not expect that :-).

Yes, they are big alright, but on the other hand, they also don't change
very much (we hope!)

> The dump files are architecture specific and maintaining multi arch
> dumps would be even bigger than just what you sent for x86_64.
> (not even considering the changes in ABI if some configuration items
> have an impact...).

Good point, we missed that when looking at this.

> 
> As you pointed out, people who don't have all dependencies won't
> create/update those dump files.

Well, the creation should be a once-off, the comparison is what is done
regularly and needs the build tools.

> Dealing with ABI updates (thinking of bumping the ABI version) is
> likely a maintainer job, but it will be a source of issues and we
> (maintainers) might miss some updates especially for drivers we can't
> build.
> 
> 
> - Why do we restrict the checks to meson?
> The make build framework is going to disappear ok, but we can't leave
> it untested.
> People still rely on it.
> 
Because as you point out below, checking the ABI is technically orthogonal
to building the DPDK, so we didn't see the payoff in adding support to two
build systems as being worth the additional effort.

> Checking the ABI is orthogonal to building DPDK.
> Dumping the ABI and checking it against objects can be done externally.
> 

True, but the advantage of doing so as part of each and every build is that
any ABI break is caught by the original developer before he even submits
his patch to the CI. As with so many things, the earlier in the process
that something can be run the better it is.

> 
> - For those reasons, I have been trying an alternate approach [1]: in
> Travis, generate a reference dump based on the first ancestor tag,
> then build the proposed patches.
> All contributions get picked up by Aaron robot and would have to pass
> through this check.

Yes, the alternative to having the checks done at build time is to have
them done as part of the CI, though I'd personally perfer the former.

> 
> As an exercise, I tried to integrate Eelco patch [2], that moves
> symbols from EXPERIMENTAL to a stable ABI. The check passes fine.
> I also tried to bump the ABI major version. The check fails, as
> expected, but I prepared a way to bypass such failures for the
> releases where we will explicitely break ABI.
> 

IMHO: we should not bypass such failures, but instead update our reference
ABI dumps. This is one reason why having the ABI spec in the git repo would
work well, any patches that change ABI would also include the updates to
the dump files.

> For maintainers that integrate patches or developers that get a CI
> failure and want to fix it, we would need to help them to:
> * generate dumps on a reference version, so I would tend to write some
> documentation since playing with the current sources would be too
> dangerous from my pov,

This should be a one-off reference dump archived somewhere. Each maintainer
should not have his own copy, I think.

> * run the checks, like adding the check in the
> devtools/test-*-build.sh scripts that already exist, with a new
> configuration item to point at the dumps per target,
> 

Where do we store the dumps then? Do they get stored in a separate git
repo?

> Those last two points are still to be done.
> 
> WDYT?
> 
Makes sense. I still prefer a solution where we keep the offical ABI in git
alongside the source code, but I realise that doing so for multiple
archtectures is likely to get to be very big. However, since these are
plain text files though, they should compress well when stored in the git
repo itself or when being pushed/pulled. [And again, the deltas for these
should be pretty tiny, even when we do update the ABI].

/Bruce