[RFC] Importing from a patch-oriented SCM

I am drafting an import script to turn a GNU Arch into a GIT archive. Importing the branches and commits increamentally is reasonably straightforward -- or so it seems so far. Note: the repository manipulation is based on cvsimport -- so my knowledge of the git repo internals is still pertty close to zero.

Each patchset has a unique identifier, and can carry metadata with the identifiers of the patches it "includes". If you are using gnu arch, when you merge across branches, it'll know to skip a particular patchset if it has been applied already. AFAICT there is no such concept in GIT, and I wonder what to do with all this metadata about merges.

My proto-plan is to keep track of merged stuff (in a cache file somewhere), and if a particular merge means that the branches are fully merged up to the last patch of the series (if no commits from the source branch have been skipped) mark it as a merge in GIT.

If the merges have been done out-of-order, that may show up in the latest merge. For example, branch A and B of the same project each have 10 commits from the branching point. If a merge A -> B does commits 1,2,3,7,8 it gets imported to git as a merge up to commit "3", although there is more there. The next merge, which does 4,5,6,10 will show up as a merge of commit 8.

Yuk. 

If I remember correctly, Junio added some stuff in the merge & rebase code that will identify if a particular patch has been seen and applied, and skip it even if it's a bit out of order. But I don't know what that is based on, and whether I can somehow maximize the chances of the patch being identified as already merged across branches. If it's based on the initial commit identifier being carried through (does that travel with commits when you merge?) I stand a small chance. Otherwise, I'm lost.

Suggestions? 

cheers,

martin

Martin Langhoff <martin.langhoff@gmail.com> writes:

I think you are talking about git-patch-id.

f97672225b3b1a2ca57cfc16f49239bed1efcd87
Author: Linus Torvalds <torvalds@ppc970.osdl.org>
Date:   Thu Jun 23 15:06:04 2005 -0700

    Add "git-patch-id" program to generate patch ID's.

    A "patch ID" is nothing but a SHA1 of the diff associated
    with a patch, with whitespace and line numbers ignored.  As
    such, it's "reasonably stable", but at the same time also
    reasonably unique, ie two patches that have the same "patch
    ID" are almost guaranteed to be the same thing.

    IOW, you can use this thing to look for likely duplicate
    commits.

On 8/19/05, Junio C Hamano <junkio@cox.net> wrote:

Is this used at commit time, and stored somewhere (doesn't seem to be) or do you select older patches from the destination branch at merge time?

If you only compare patches since the last merge, patches that were merged but somehow unreported will fall into a black hole and cause a conflict going forward anyway. Hmm. That seems to be a problem I won't be able to avoid if merges happen out-of-order.

I'll try and work out how it's being used during the merge (pointers/hints welcome) and see if I can do something smart w it. Thanks!

cheers,

martin

On Fri, 19 Aug 2005, Martin Langhoff wrote:

If a patch is applied verbatim, or a merge results in no conflicts (i.e., only offsets), then you can run git-patch-id on the diff caused by it and compare the result with the git-patch-id of the diff caused by your local change to see if you've found it. Of course, if there was any modification to the patch or a conflict was resolved, you won't see a match, but that's plausibly correct anyway: you don't know whether the content change that resulted from your patch really matched the change you wanted to make.

They might cause conflicts, but they're relatively unlikely to require manual intervention, because the merging mechanism in git is stronger than the one in arch (by virtue of identifying a common ancestor), and will recognize when a section of changes made by both sides is the same and produce a warning rather than a conflict. That's how the rebase stuff can identify that your rebased patch is empty (when upstream applies your patch): the content change that it would make has been made.

	-Daniel
*This .sig left intentionally blank*

Hi,

On Fri, 19 Aug 2005, Martin Langhoff wrote:

You should include the metadata in the commit object. If the information is about parents, they should be parents in git, too. If the information is something else, you should convert it to readable text and put it in the comment part of the commit object.

The usual way of git is to use a 3-way merge: given a common ancestor, try to apply the changes between the ancestor and the second branch to the first branch. And yes, this does not take history into account. Originally, I wanted to write an "intelligent" merge, which turns the history into patches and tries to merge these, but ultimately I got convinced that this is too complicated to be worthwhile.

Ciao, Dscho