Embedded Linux development with GIT

Hi,

I have a situation where we have a local GIT repository that is based on v2.6.17. We initially added the source tarball to an empty repository and then started applying changes to it. Looking back, that might not have been the best idea 400 commits later.

My goal is to eventually bring our repository closer to mainline revisions so as to make it easier to actually contribute back to the community. So how can I fix my repository so as to give it visibility to Linus' kernel?

Here is my initial thoughts:

1) Clone kernel.org kernel and it is Master
2) Create a local Head based on 2.6.17 and call it Local
3) Pull my existing heavily patched repository into the Local branch and merge

Is it possible then to see our 400 odd commits then in the Local branch on top of 2.6.17 so that we can see not only our history but also the history that came before? Then as Master advances we can see about backporting and bringing our code close enough to mainline kernel to actually be able to contribute back to the community and submit patches. Is this realistic approach. I am unsure of the GIT commands that I need to do this?

Thanks,

Sean

On 7/5/07, Sean Kelley <svk.sweng@gmail.com> wrote:

If I understand correctly, what you might want to do is keep your current master branch, pull in the mainline kernel to a separate branch, and rebase your master on the mainline branch.

git checkout master git rebase mainline will find the last common commit between two branches, then plop all the mainline commits on top, then proceed to merge in your changes on the master branch, one-by-one, stopping if there are any conflicts and allowing you to resolve as they happen. Then your 400 commits will be on top and you can rebase whenever and still see them on top.

-Dan "Puzzles" Chokola

On 7/5/07, Sean Kelley <svk.sweng@gmail.com> wrote:

Better:

  $ (cd $OLDREPO && git format-patch --stdout -k first..) | git am -k

or, if it is in the same repo, assuming it starts with sha1 "old-beginning"

  $ git format-patch --stdout -k old-beginning.. | git am -k

Merging of unrelated histories is slow (but works).

Yes. Look at how OLPC (www.laptop.org) does this with their kernel.

aside from the already mentioned git format-patch and git am: git apply, git log, git show, gitk, and come to think of it - all of the rest too, except maybe for importing programs.

Sean Kelley wrote:

That is possible using a graft:

  $ echo "$x $(git rev-parse v2.6.17^0)" >> .git/info/grafts

where $x is the SHA1 of the first commit you made on top of the imported tarball. This way you have spliced your history with Linus's history. (This is a strictly _local_ matter! Every clone of your history must repeat the game!)

Now, Linus will not be able to pull from your faked history because he doesn't know about the graft. In order to fix that, you can run git-filter-branch from current git's master branch to rewrite your history:

  $ git filter-branch new-master v2.6.17..master

Read the man page of git-filter-branch, and understand the implications before you publish the result.

-- Hannes

Hi,

On Thu, 5 Jul 2007, Johannes Sixt wrote:

Yes, this is also how I would do it.

Now, here I disagree slightly. If you merge just once, subsequent merges will be possible even without that graft.

So if you merge with some newer Linux version, all your cloners get the benefit.

Except if you merge with a more recent version of Linux.

However, I doubt that such a distant (in terms of time!) merge would appeal to Linus. I guess you have to rebase on top of Linus' version _anyway_.

This is a way to fix your history, yes. Note that filter-branch is not yet in an official release of Git, and so you either have to wait for 1.5.3, or you get filter-branch from git.git's "next" branch (just picking this one script should work fine, if you have at least 1.5.1 installed).

Note that this rewrites the history, so all the disadvantages of rebasing with pulling apply here, too.

But as stated above, I think you have to rebase eventually anyway, if you go for inclusion in Linus' tree. In that case, the filter-branch is unnecessary.

Ciao, Dscho