Re: Why we should avoid new submodules if possible

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Sep 28, 2022 at 07:15:53AM -0600, Warner Losh wrote:
> 
> 
> On Wed, Sep 28, 2022 at 7:09 AM Daniel P. Berrangé <berrange@redhat.com> wrote:
> 
>     On Wed, Sep 28, 2022 at 05:53:17AM -0400, Michael S. Tsirkin wrote:
>     > On Wed, Sep 28, 2022 at 10:37:14AM +0100, Daniel P. Berrangé wrote:
>     > > On Wed, Sep 28, 2022 at 05:26:42AM -0400, Michael S. Tsirkin wrote:
>     > > > On Tue, Jun 28, 2022 at 12:21:39PM +0200, Thomas Huth wrote:
>     > > > > On 28/06/2022 12.03, Michael S. Tsirkin wrote:
>     > > > > [...]
>     > > > > > For biosbits if we are going this route then I feel a submodule
>     is much
>     > > > > > better.  It records which version exactly each qemu version
>     wants.
>     > > > >
>     > > > > As far as I know, you can also specify the version when using pip,
>     can't
>     > > > > you? So that's not really an advantage here.
>     > > > >
>     > > > > On the contrary, submodules have a couple of disadvantages that I
>     really
>     > > > > dislike:
>     > > > >
>     > > > > - submodules do not get updated automatically when doing a "git
>     checkout",
>     > > > > we have to update them via a script instead. This causes e.g.
>     trouble if you
>     > > > > rsync your source tree to a machine that has no access to the
>     internet and
>     > > > > you forgot to update the submodule before the sync
>     > > > >
>     > > > > - the content of submodules is not added to the tarballs that get
>     created on
>     > > > > the git forges automatically. There were lots of requests from
>     users in the
>     > > > > past that tried to download a tarball from github and then wondered
>     why they
>     > > > > couldn't compile QEMU.
>     > > > >
>     > > > > - we include the submodule content in our release tarballs, so
>     people get
>     > > > > the impression that hte submodule content is part of the QEMU
>     sources. This
>     > > > > has two disadvantages:
>     > > > >  * We already got bug reports for the code in the submodule,
>     > > > >    where people did not understand that they should report that
>     > > > >    rather to the original project instead (i.e. you ship it - you
>     > > > >    own it)
>     > > > >  * People get the impression that QEMU is a huge monster
>     > > > >    application if they count the number of code lines, run
>     > > > >    their code scanner tools on the tarball contents, etc.
>     > > > >    Remember "nemu", for example, where one of the main complaints
>     > > > >    was that QEMU has too many lines of code?
>     > > > >
>     > > > > - If programs includes code via submodules, this gets a higher
>     > > > >   burder for distro maintainers, since they have to patch each
>     > > > >   and every package when there is a bug, instead of being able to
>     > > > >   fix it in one central place.
>     > > > >
>     > > > > So in my opinion we should avoid new submodules if there is an
>     alternative.
>     > > > >
>     > > > >  Thomas
>     > > >
>     > > > So looking at the latest proposals downloading files from CI,
>     > > > checksumming them etc etc. No auto checkout, not added automatically
>     > > > either, right?
>     > > >
>     > > > This seems to be the only difference:
>     > > > - we include the submodule content in our release tarballs
>     > >
>     > > That's just one of the issues with submodules. Working with them in
>     general
>     > > is not a pleasant experiance.
>     >
>     > This is what I asked about at the maintainers summit.
>     > I'd like to map the issues and see if there's anything
>     > we can do to solve them. In particular we will likely
>     > keep using submodules where we historically did
>     > so it's time well spent.
>     >
>     > I agree generally but the big question is what to replace these with.
>     > Below I assume the replacement is a script such as avocado or pytest
>     > with its own hashing, calling wget internally etc etc.
>     >
>     >
>     > > Thomas pointed out some of the issues, such
>     > > as 'git checkout' ignoring submodules, requiring extra steps to sync
>     them.
>     >
>     >
>     > Not different from a home grown SCM as part of test script, right?
> 
>     We're not building a home grown SCM as part of a test script, so
>     this answer is irrelevant.
> 
>     > > There's also the perenial problem that developers frequently send
>     > > patches that mistakenly include submodule changes,
>     >
>     > OK, so the thing to do would be to look for ways to exclude submodule
>     changes
>     > from git commits.
> 
>     If someone wants to make git suck less with submodules great, but needs
>     someone to actually do the work.
> 
> 
> A big part of the problem is knowing which of the following commands I have to
> do to undo the uncommitted changes, the committed changes, the staged changes,
> etc:
> 
> git submodule update --init --recursive
> git submodule update --init
> git submodule foreach --recursive git reset --hard
> git submodule foreach --recursive git clean -xfd
> git submodule foreach --recursive git clean -xfg
> 
> (all of these are in my history, I honestly don't know the difference between
> the last two).
> And each 'oops' takes time away from upstreaming bsd-user I don't really have
> that
> much of. I've wasted hours on this over the past year between all the different
> ways
> it can screw up.
> 
> To be fair, it is a relatively small fraction of the time, but as you can tell
> from the
> animation in my email it inspires much passion.
> 
> Warner
>  

OK this is understandable.

I think for what we do in qemu (never change upstream really)
	git submodule update --checkout testsub
	git update-index --skip-worktree testsub
is more or less what most people want.

This assumes you have a script that basically does
	git submodule deinit XYZ
	git submodule update --init XYZ
	git update-index --skip-worktree XYZ
each time, and do not need to work on the submodule proper.




>     > > I'd really like to see us doing more to eliminate as much use of
>     submodules
>     > > as is possible over time.p
>     >
>     > Or try to fix the problems, right?
> 
>     Again needs someone to actually make it happen.
> 
>     Meanwhile  QEMU already has an integrated test harness in the form
>     of Avocado that does everything needed. If Avocado had just been
>     used for this biosbits test in the first place, the test would
>     likely have already been merged to QEMU, instead of us having this
>     never ending debate on how to re-invent an alternative to what
>     already avocado does.
> 
>     With regards,
>     Daniel
>     --
>     |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange
>     :|
>     |: https://libvirt.org         -o-            https://fstop138.berrange.com
>     :|
>     |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange
>     :|
> 
> 
>