Re: vendor/illumos merges

From: Mark Johnston <markj_at_freebsd.org>
Date: Wed, 19 May 2021 15:08:49 -0400
On Wed, May 19, 2021 at 09:04:57PM +0200, Ulrich Spörlein wrote:
> On Tue, May 18, 2021 at 7:49 PM Mark Johnston <markj_at_freebsd.org> wrote:
> 
> > On Sun, Apr 25, 2021 at 04:02:22PM +0200, Ulrich Spörlein wrote:
> > > On Sat, 2021-04-24 at 11:08:58 -0400, Mark Johnston wrote:
> > > >On Sat, Apr 24, 2021 at 12:44:40PM +0200, Ulrich Spörlein wrote:
> > > >> On Fri, 2021-04-23 at 17:26:33 -0400, Mark Johnston wrote:
> > > >> >Hi,
> > > >> >
> > > >> >Now that FreeBSD uses OpenZFS as the upstream for ZFS,
> > vendor/illumos is
> > > >> >mostly unused.  However, we still use illumos as an upstream for CTF
> > > >> >tools and DTrace, though there haven't been any imports in a while.
> > > >> >
> > > >> >illumos has put a lot of work into their CTF toolchain, and I'd like
> > to
> > > >> >import that.  There are a couple of snags that I'd appreciate some
> > > >> >guidance on.
> > > >> >
> > > >> >First, I believe I should delete now-unused ZFS code from the vendor
> > > >> >branch and merge the result to main.  I did this locally and got an
> > > >> >empty merge, which is what I'd expect.  Is there any problem with
> > this?
> > > >>
> > > >> Why would you record this empty merge? If you clean up vendor/foo,
> > just
> > > >> do that but don't merge a no-op back into main (nothing changed, after
> > > >> all).
> > > >
> > > >Ok, I guess there is no reason to merge that change separately.  It
> > > >will end up being merged with subsequent imports though.
> > > >
> > > >> >Second, with Subversion we had both vendor/illumos and
> > > >> >vendor-sys/illumos, and now we just have the former, seemingly with
> > > >> >sys/* bits imported from vendor-sys.  Some of the upstream commits
> > touch
> > > >> >both userspace and kernel bits, but the merge targets for these in
> > > >> >FreeBSD are different: cddl/contrib/opensolaris vs.
> > > >> >sys/cddl/contrib/opensolaris.  How should I merge into main in this
> > > >> >case?  I don't really see any options other than to split each
> > offending
> > > >> >upstream commit into two parts, one for userspace and one for the
> > > >> >kernel, and merge them separately.
> > > >> >
> > > >> >If it helps to look at the branch where I staged the upstream
> > commits,
> > > >> >I've pushed it to vendor/illumos2 in
> > https://github.com/markjdb/freebsd
> > > >> >.
> > > >>
> > > >> Can you clarify why the merging of the two might be an issue? Note
> > that
> > > >> unlike subversion, in git there's no "merge a certain subtree"
> > handling,
> > > >> all that is recorded is a tree of some form and then a set of parents
> > or
> > > >> ancestor commits. (git is a content tracker, not really a VCS :)
> > > >>
> > > >> I was under the impression that userland and kernel imports/merges
> > need
> > > >> to happen at the same time anyway, so I assume you would import all
> > the
> > > >> bits under vendor/foo in 1 commit and then merge them in 1 commit into
> > > >> main. Is that not how it goes?
> > > >
> > > >How can I do that with git subtree merge?  Suppose an illumos commit
> > > >modifies cmd/dtrace/foo.c (userspace) and uts/common/dtrace/foo.c
> > > >(kernel).  That maps to cddl/contrib/opensolaris/cmd/dtrace/foo.c and
> > > >sys/cddl/contrib/opensolaris/uts/common/dtrace/foo.c in FreeBSD,
> > > >respectively.  So to do a subtree merge, I need to use distinct prefixes
> > > >depending on whether I'm importing userspace or kernel changes.  When
> > > >they are mixed together, it's not clear to me how I can merge at all.
> > > >
> > > >I see that for OpenZFS we keep all code, including userspace code, under
> > > >sys/contrib/openzfs, so it doesn't have this problem.
> > >
> > > I don't think you want a subtree merge, especially as things are
> > > scattered all over the place. Also note that none of this subtree magic
> > > is in any way recorded in the git data, all it does is help you with the
> > > 3-way merges (or whatever).
> > >
> > > So I would do:
> > > - import whatever you need into contrib/foo, commit normally.
> > > - munge /usr/src to have every kernel and userland stuff (not sure what
> > > other merge tools exist, just make sure to copy over file deletions as
> > > well :). You could rsync --del two times with the right source/dest
> > > pairs, or export a diff/patch from step 1 and apply it under the right
> > > prefixes. test, test, test.
> > > - write out this tree to git using: git write-tree
> > > - then commit this using: git commit-tree -m "my message" -p HEAD -p
> > > origin/vendor/illumos <tree hash from previous command>
> > > - bump main to point to that hash using git update-ref
> > > - git log --graph and inspect the hell out of this
> > > - git push, then curse that we disallow merge commits and you need to
> > > `git pull --rebase` to advance to the latest published head and that
> > > might mess up your merge commit pretty bad :(
> > >
> > >
> > > Maybe 2x git subtree merge + then rewriting and squashing them into 1
> > > would work. But I fear it will record 3 parents, not 2 parents.
> > >
> > > Whatever you do, maybe please push to your private Github clone or our
> > > dev repo first and tell us where to look, so we can inspect whether it
> > > looks ok.
> >
> > I followed your suggestions and have what looks like a clean result.
> > Basically I did two subtree merges and committed the result.
> >
> > I pushed it here:
> > https://github.com/markjdb/freebsd/tree/ff/illumos-merge
> > The merge is the second-last commit on that branch.
> >
> > Rebasing the merge is painful, partly because there are some old ZFS
> > commits in the illumos vendor branch which git tries to merge.  Rename
> > detection leads to some really weird conflicts as well.  I'm not sure
> > how to handle this: there's a lot of room to make mistakes when
> > rebasing, so I would want to be careful and do extra testing before
> > pushing the final merge, but during that window it's likely that I will
> > end up having to rebase again.
> >
> > I should perhaps remove all ZFS bits from the vendor branch, and merge
> > it into main first before importing other stuff.
> >
> 
> I see only 1 merge commit in there, is that the expected outcome?

Yes.  I should probably do this as two separate merge commits as I
suggested above, but I wanted to know if I had done anything incorrectly
in this attempt before spending more time on it.  The branch isn't ready
to merge yet in any case, there are some upstream bugs to squash.

> The output of
> git show -p --first-parent 23a19903267ee799cbddc35eb3e6f978ac1b4f38
> looks mostly sane though, does it look like the changes that you'd like to
> bring into main?

Yes, that's correct.

> Warner, could you please also have a look?
> 
> Cheers
> Uli
Received on Wed May 19 2021 - 19:08:49 UTC

Original text of this message