Re: [libvirt PATCH 000/351] port libvirt to Meson build system

On Tue, Jul 28, 2020 at 10:00:20AM +0200, Pavel Hrdina wrote: > On Mon, Jul 27, 2020 at 06:11:11PM +0200, Peter Krempa wrote: > > On Fri, Jul 17, 2020 at 15:02:10 +0100, Daniel Berrange wrote: > > > On Thu, Jul 16, 2020 at 03:44:25PM +0200, Pavel Hrdina wrote: > > > > On Thu, Jul 16, 2020 at 01:59:00PM +0100, Daniel P. Berrangé wrote: > > > > > Personally I'd really like to avoid squashing them, because splitting > > > > > up big patches is not merely to benefit the initial pre-merge review, > > > > > but to also benefit people who need to debug stuff that's already > > > > > merged and understand the scope of the intended change. So being able > > > > > to look back at the changes in isolation after commit is still a big > > > > > plus point. > > > > > > > > I would like to avoid squashing the patches as well and in most cases I > > > > would object to it as well. I only suggested that to not break git > > > > bisect. > > > > > > > > If we don't care about git bisect and the fact that we would not be able > > > > to build libvirt correctly within these patches I'm OK with pushing it > > > > without squashing. > > > > > > git bisect reliabity is key, so I reluctantly think we'll need to > > > squash. I don't want to hit a pathc in this series with a bisect > > > and be unable to continue the bisect due to inability to build the > > > code. > > > > I agree. It's definitely necessary that the build is complete at any > > point in time. > > > > I'm reluctantly willing to accept that the build fails with an > > appropriate error message until the build system is able to build > > everything if we opt for commiting a patchset for simplicity. What's > > off-limits is if build "succeeds", but is incomplete due to missing > > steps in the implementation. I'm not going to want to guess which part > > is already built or which isn't. > > > > Given that the rewrite is a singularity anyways it doesn't really matter > > that we will not be able to bisect problems caused by the build system > > across the boundary. > > So based on all the comments we have these options for pushing this > series: > > 1) Squash it into single commit. > > Pros: - no issues with git bisect > > Cons: - we will not have the history of changes > > 2) Keep the patches as they are and running meson build & ninja will > not fail. > > Pros: - full history of changes where each commit removes the > relevant bits from autotools > > - git bisect is not broken as compilation will not fail > > Cons: - meson build && ninja will not produce complete libvirt > binaries and there is no autogen.sh > > - script used for git bisect will have to detect if tested > binaries are compiled or use git bisect skip So basically any time git bisect lands on one of the meson conversion patches, we'll need to use "git bisect skip". You might have todo that multiple times. It'll make bisect less efficient, but assuming the problem was not part of the meson series, git will eventually show you the broken commit. The problem here is remembering which commits are ones which need to be skipped..... > 3) Keep the patches as they are but error out in meson until the > conversion is complete. The error can be used to detect if git > bisect is withing the meson rewrite. > > Pros: - full history of changes where each commit removes the > relevant bits from autotools > > - git bisect is not broken if failed compilation is not an > issue and marked as git bisect skip > > Cons: - meson build will fail and there is no autogen.sh so no > way how to compile libvirt even partially > > - script used for git bisect will have to skip failed > compilation with an option to check for specific error ..So explicitly failing the meson build is a significant improvement. We could have - Meson build is forced to fail out of the box - An option "force_incomplete_build" to turn off the fail - When failing prints an error message "This commit is part of the meson conversion and does not build a complete libvirt. If bisecting, use "git bisect skip" to continue, or "-Dforce_incomplete_build=true" to perform a partial build" So with that we would have full history, and git bisect would be able to identify problems in ANY commit that is NOT part of the meson series, except the single commit that is immediately either side of the meson series. That should be viable I think.

> 4) Rework the series to have patches adding meson bits without > removing anything from autotools and drop the autotools files in > a single commit once the meson rewrite is complete > > Pros: - full history of changes > > - git bisect not broken because autogen.sh && make will > work the whole time until meson build && ninja is ready > > Cons: - no reference of the meson changes to autotools code > > - additional work for me to redo the patches