patch, 6 partsInitial subproject support (RFC?)

Ok, the following is a series of six patches that implement some very low-level plumbing for what I consider sane subproject support.

NOTE! I want to make it very clear that this series of patches does not make subprojects "usable". They are very core plumbing that allows people to think about the issues, and shows how the low-level code could (and in my opinion, should) be done.

Some of the early patches are just cleanups and very basic stuff required to actually get to the meat of it all. I actually think that they are all in a state where they could be applied, if only because they don't actually really *do* anything unless you start generating index files entries (and trees) that have the "gitlink" entries in them.

I've actually done some testing with a repository that has these kinds of subproject pointers in them, and no, it's really not fully fleshed out yet, but yes, I can actually do a commit in one of the subprojects, and when I do that, the "raw" diff literally looks like this:

	[torvalds@woody superproject]$ git diff --raw
	:160000 160000 5813084832d3c680a3436b0253639c94ed55445d 0000000... M    sub-B

and I can do a "git commit -a" in the superproject to commit the new state.

NOTE! This series of six patches does not actually contain everything you need to do that - in particular, this series will not actually connect up the magic to make "git add" (and thus "git commit") actually create the gitlink entries for subprojects. That's another (quite small) patch, but I haven't cleaned it up enough to be submittable yet.

I split my original larger patch up into more manageable pieces, so that you should be able to actually just read the patches themselves and get a reasonable idea about what it's doing, even *without* actually testing it. And obviously, "make test" still completes happily, if only because none of the tests actually trigger any of the new code.

The patches are all fairly small, and the two first ones are really just totally independent cleanups/fixes:

 - diff-lib: use ce_mode_from_stat() rather than messing with modes manually:

	 diff-lib.c |   15 +++------------
	 1 files changed, 3 insertions(+), 12 deletions(-)

 - Avoid overflowing name buffer in deep directory structures:

	 dir.c |    3 +++
	 1 files changed, 3 insertions(+), 0 deletions(-)

 - Add 'resolve_gitlink_ref()' helper function:

	 refs.c |   79 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
	 refs.h |    3 ++
	 2 files changed, 82 insertions(+), 0 deletions(-)

 - Add "S_IFDIRLNK" file mode infrastructure for git links:

	 cache.h |   20 +++++++++++++++++++-
	 1 files changed, 19 insertions(+), 1 deletions(-)

 - Teach "fsck" not to follow subproject links:

	 builtin-fsck.c |    9 ++++++++-
	 tree.c         |   15 ++++++++++++++-
	 2 files changed, 22 insertions(+), 2 deletions(-)

 - Teach core object handling functions about gitlinks:

	 builtin-ls-tree.c |   20 +++++++++++++++++++-
	 cache-tree.c      |    2 +-
	 read-cache.c      |   35 +++++++++++++++++++++++++++++++----
	 sha1_file.c       |    3 +++
	 4 files changed, 54 insertions(+), 6 deletions(-)

and will follow in the next few emails..

			Linus

The diff helpers used to do the magic mode canonicalization and all the other special mode handling by hand ("trust executable bit" and "has symlink support" handling).

That's bogus. Use "ce_mode_from_stat()" that does this all for us.

This is also going to be required when we add support for links to other git repositories.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 diff-lib.c |   15 +++------------
 1 files changed, 3 insertions(+), 12 deletions(-)

This just makes sure that when we do a read_directory(), we check that the filename fits in the buffer we allocated (with a bit of slop)

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 dir.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)

This new function resolves a ref in *another* git repository. It's named for its intended use: to look up the git link to a subproject.

It's not actually wired up to anything yet, but we're getting closer to having fundamental plumbing support for "links" from one git directory to another, which is the basis of subproject support.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 refs.c |   79 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 refs.h |    3 ++
 2 files changed, 82 insertions(+), 0 deletions(-)

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

Can't a subproject be bare?

On Tue, 10 Apr 2007, Alex Riesen wrote:

Not when it is checked out, no. That's what "checked out" means ;)

If a subproject is bare, it never gets resolved, because it's never checked out in a superproject.

So a subproject *can* be bare, but when it's bare it is just a totally regular independent git project, simply by *definition* of not being checked out inside a superproject.

But hey, that was just a design decision of mine, and if people can argue for it being wrong, I don't think I'm married to it ;)

		Linus

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

I didn't actually had a use case in mind as I asked it. After a bit of thinking I could imagine a repo which is used for integration exclusively (no compilation or looking at the files at all).

On Tue, 10 Apr 2007, Alex Riesen wrote:

Well, you also cannot *commit* to a bare repository, so it's a bit pointless for integration reasons. You'd still have to commit all changes somewhere else.

That said, it's definitely designed so that if you want to automate tracking other peoples bare repositories, you can do so: you'd just have to *really* script it with something like

	git update-index --cacheinfo 0160000 <sha1> <dirname>

(which is how you could create those commits to a bare repo too, so it's not like this is really even any different)

		Linus

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

Yes. Subprojects are push-only for storing and reference purposes. Superproject can have integrated data checks in Makefiles.

Nice :)

On Tue, 10 Apr 2007, Alex Riesen wrote:

Well, the *really* nice thing about doing it like this is that you can actually update subprojects without even having them even be *local* to where you do the superproject.

IOW, you could literally build up the superproject by saying that you want to track "all git projects I care about" somewhere else, and do a series of automated

	git ls-remote sub-project-xyzzy tracking-branch-xyzzy | ...

and basically create the "superproject" without ever actually downloading or populating the subprojects at all.

Then, if everything is set up correctly, you can basically use the superproject as an "auto-mirror" - whenever you want to get all the projects you care about, you just clone that superproject, and (once you've taught "git clone" to fetch the subprojects, of course ;^) you'd basically fetch them all from their appropriate locations - without ever having the actual superproject have to even *really* care about it.

So basically, a superproject could be used as just a "gathering point", without having to actually *contain* any of the subprojects. The actual sources for subprojects may be on totally different servers. That's what real distribution is all about.

		Linus

On Tuesday 10 April 2007, Linus Torvalds wrote:

It would be nice if a redirection via a "gitdir = ..." line in .git/link of the subproject (when existing) would be possible. This was part of the light-weight checkout proposal.

In contrast to contrib/workdir/git-new-workdir, this would allow for (to be implemented) magic symlinks to stay intact when moving the submodule directory around.

However, this can be added later.

Josef

PS: I wonder how long it takes to move the official KDE repository over to git ;-)

This just adds the basic helper functions to recognize and work with git tree entries that are links to other git repositories ("subprojects"). They still aren't actually connected up to any of the code-paths, but now all the infrastructure is in place.

The next commit will start actually adding actual subproject support.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 cache.h |   20 +++++++++++++++++++-
 1 files changed, 19 insertions(+), 1 deletions(-)

Since the subprojects don't necessarily even exist in the current tree, much less in the current git repository (they are totally independent repositories), we do not want to try to follow the chain from one git repository to another through a gitlink.

This involves teaching fsck to ignore references to gitlink objects from a tree and from the current index.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---
 builtin-fsck.c |    9 ++++++++-
 tree.c         |   15 ++++++++++++++-
 2 files changed, 22 insertions(+), 2 deletions(-)

Linus Torvalds wrote:

Does this consider the case where the intent of the subprojects are to collate multiple, small projects into one bigger project?

In that case, you might want to keep all of the subprojects in the same git repository.

Sam.

On Thu, 12 Apr 2007, Sam Vilain wrote:

I assume you mean "you might want to keep all of the subprojects' objects in the same git object directory".

And yes, that's absolutely true, but it's technically no different from just using GIT_OBJECT_DIRECTORY to share objects between totally unrelated projects, or using git/alternates to share objects between (probably *less* unrelated repositories, but still clearly individual repos).

So the main point of superproject/subprojects is to allow independence (because independence is what allows it to scale), but there is nothing to say that things *have* to kept totally isolated.

			Linus

Linus Torvalds wrote:

Would that be the only distinction?

Would submodules be descended into for object reachability questions?

I'm particularly interested in repositories with, say, thousands of submodules but only a few hundred meg. I really want to avoid the situation where each of those submodules gets checked or descended into separately for updates etc.

Sam.

On Thu, 12 Apr 2007, Sam Vilain wrote:

I think we'll eventually want that *regardless* of how the object handling is done (a kind of "cross-submodule boundary check"), but I think that's actually outside of the scope of the current fsck.

The current fsck goes to great lengths to make sure that the internal consistency of a repository is good. That's also why it takes so long, and why it is such an expensive operation to do (notably when you do a "--full" check).

In contrast, the "cross-submodule boundary check" is a much cheaper operation, *if* you have already verified that the projects are internally consistent. It literally boils down to doing a very simplified commit chain walker that only parses tree objects and simply spits out the SHA1's of the sub-tree commits (and their location in the tree), and then a separate phase that just verifies those against the submodules.

And that separate phase - once you've done the fsck for all the *individual* repositories - is truly trivial. It's literally just a matter of "is that SHA1 a valid commit object". That's *cheap*.

See?

So I think that the way to verify a superproject is:

 - fsck each and every project totally independently. This is something 
   you have to do *anyway*.

 - either as you fsck, or as a separate phase after the fsck, just 
   traverse the trees and spit out "these are the SHA1's of subprojects"

 - finally, just go through the list of SHA1's (after every project has 
   been fsck'd) and verify that they exist (since if they exist, they will 
   have everything that is reachable from them, as that's one of the 
   things that the *local* fsck verifies)

Notice? At no point do you actually need to do a "global fsck". You can do totally independent local fsck's, and then a really cheap test of connectedness once those fsck's have completed.

The reason a *full* global fsck is so expensive is that it would have an absolutely humungous working set, and effectively keep everything in memory through it all. Doing it in stages ("fsck smaller individiual trees separately") is actually the same amount of absolute work, but the working set never grows, so it scales much better.

(fsck'ing projects individually also happens to allow you to do the sub-project fsck's in parallel across multiple CPU's or multiple machines, so it actually scales much better that way too - but the big problem tends to be excessive memory use, so the "SMP parallel version" only makes sense if you have tons of memory and can afford to do these things at the same time!)

			Linus

On Wed, 11 Apr 2007, Linus Torvalds wrote:

would it make sense to have a --multiple-project option for fsck that would let you specify multiple 'projects' that share a object set and have the default checking not do the reachability checks that cause problems in this case?

Then people can share the objects if they want to and still do a full check, but would get warned that the full check would take a lot of time. which is not a big problem for a housekeeping thing that's run infrequently to find unreachable objects (which is something that should seldom happen in a well managed project)

David Lang

On Wed, 11 Apr 2007, David Lang wrote:

Well, the thing is, sharing object directories actually makes things *harder* to check, rather than easier.

It can be a nice space optimization, and yes, if there really is a lot of shared state, it can make it much cheaper to do some of the checks, but right now we have absolutely *no* way for fsck to then do the reachability check, because there is no way to tell fsck where all the refs are (since now the refs come in from multiple repositories!)

So the individual objects get cheaper to fsck (no need to fsck shared objects over and over again), but the reachability gets much harder to fsck.

It's not an insurmountable problem, or even necessarily a very large one, but it boils down to one very basic issue:

 - nobody seems to actually *use* the shared object directory model!

The thing is, with pack-files and alternates directories, a lot of the original reasons for shared object directories simply don't exist..

		Linus

On Wed, 11 Apr 2007, Linus Torvalds wrote:

this is why I was suggesting a --multiple-project option to let you tell fsck about all of the repositories that it needs to look for refs in.

agreed.

I suspect that if it coudl be checked it would be used more, especially with the subproject support.

David Lang

On Wed, 11 Apr 2007, David Lang wrote:

Well, just from a personal observation:
 - I would *personally* actually refuse to share objects with anybody 
   else.

I just find the idea too scary. Somebody doing something bad to their object store by mistake (running "git prune" without realizing that there are *my* objects there too, or just deciding that they want to play with the object directory by hand, or running a new fancy experimental importer that has a subtle bug wrt object handling or anything like that).

I'll endorse use "alternates" files, but partly because I know the main project is safe (any alternates usage is in the "satellite" clones anyway, and they will never write to the alternate object directory), and partly because at least for the kernel, we don't have branches that get reset in the main project, so there's no reason to fear that a "git repack -a -d" will ever screw up any of the satellite repositories even by mistake.

But for git projects, even alternates isn't safe, in case somebody bases their own work on a version of "pu" that eventually goes away (even with reflogs, pruning *eventually* takes place).

So I tend to think that alternates and shared object directories are really for "temporary" stuff, or for *managed* repositories that are at git *hosting* sites (eg repo.or.cz), and where there is some other safety involved, ie users don't actually access the object directories directly in any way.

So I've at least personally come to the conclusion that for a *developer* (as opposed to a hosting site!), shared object directories just never make sense. The downsides are just too big. Even alternates is something where you just need to be fairly careful!

		Linus

Linus Torvalds <torvalds@linux-foundation.org> writes:

Actually that is not even true for repo.or.cz -- the site lets people to create *forks* of the main project, and I recall it is implemented in terms of alternates.

That's one of the reasons I never asked to take over git.git repository there. I have alt-git.git instead, which does not allow forks.

On Wed, 11 Apr 2007, Linus Torvalds wrote:

I was actually thinking that hosting sites (and things like gitorrent) would be the ones that would get the most benifit from shareing objects. the amount saved for any individual developer is probably fairly minor (and the individual developer could run a script to look across their objects and hard-link them togeather if they care about the space)

David Lang

On 4/11/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

These arguments all seem pretty convincing to me -- maybe the problem is that I'm not a "*developer*" right now. Instead I'm part of a multi-developer *site*. Below I talk about a possible way we could use git without changing it (since I recognize this would be a minority usage pattern).

We use perforce to manage a mixed hardware/software project (I'm the 55GB check-out guy, remember?). We have at least 3 different kinds of data with different usage patterns, and using perforce for everything in one centralized server was not the best solution.

Each user ("client") has their own worktree and the perforce repository is on a shared central server. You can consider perforce to have the equivalent of git's index, but it is stored on the server, in one file ("db.have") covering all clients. Obviously that becomes a bottleneck -- and recently db.have got larger than the total cache RAM on the server, which really slowed things down until we moved to a larger server. But repository architecture aside, the real problem has been perforce's usability. Frequently one contributor, having gotten ahead of the team, needs to share this more recent work with only a few people. This could be done with p4 branching, but this is really clunky. So instead the work is pushed out (submitted) to everyone, causing instability; this is partially remedied by doing it in smaller chunks. Another perforce problem is that tagging consumes a lot of server space (and may slow things down as well).

Some of this data will stay in perforce, some will move into revision control built-in to some of our other tools, and I'd like to try to move some of it into git. The main attraction for the last group is the lightweight branching that would allow early/tentative work to be easily shared. I think the subproject work currently being discussed is going to be very helpful as well -- the perforce equivalent is chaotic.

We could give each user a work tree and an object repository, and then have a "release" repository. Unfortunately, this would be slower to use than the current perforce "solution": users would check in to their local repository, at the speed of gzip, anyone checking it out would do so at the speed of gzip, and all work would need to be resubmitted (using perforce jargon here) to the central repo, again at the speed of gzip. Currently, people either submit or check out from the central repo, and it's all done at the speed of a network copy. This speed issue is important because of the size of a commit we'd like to share (but not yet release): about 40 files, half of them control files of several KB each, 1/4 of them design files of several MB each, and the last 1/4 detailed design files 100X larger. These 40 files will reference (include) 50 others of several KB each sprinkled through-out the hierarchy, a few of which might have changed. And yes, almost all of these are generated files, but the generation time, and the instability of the tool and script environment, preclude forcing the other users to regenerate them, like you would with a .o file.

So, there are 2 alternative set-ups. In one, everyone uses a shared object repository (everyone's .git/objects is a symlink to it). In this repository, objects/. , objects/?? , objects/pack , and objects/info all have "sticky" set, and we do the appropriate machinations to make all files read-only. There would be an additional phantom user "git" who owns the shared object repository (the only user whose .git/objects is not a symlink). Users would commit to their own repositories, which would write data to the shared object repository and update their refs (e.g. HEAD). To "release", push to the ~git repository. This push would be like a current push -- fast-forward only, figure out the list of objects that need to be transmitted -- but instead of transmitting the objects, change their ownership to ~git and then update ~git's refs. Since users can share local commits, maybe the ~git ownership change should happen at commit time. This all seems do-able without change in git; instead I'd add a few bash wrapper scripts (and see below for fsck and pack/prune).

Another setup is like the previous, but make the central repo have its own hidden object repository. You would push to it using the standard git command.

Finally, users could run git-fsck [with misleading output]; they could run git-prune{,-packed}, but these commands wouldn't be able to delete anything. If we don't want users to pack, then ~git/.git/objects/pack would be writable only by ~git. So basically, normal people wouldn't do the things in this paragraph.

To do meaningful and safe fsck/prune on the shared repository as ~git, I'd add some scripting. If you require all users' GIT_DIR's to look like /home/USER/*/.git , then you can get all their refs and do a meaningful fsck. If not, you could do a fsck --unreachable as ~git and filter the result by date and/or type. (This sort of corresponds to abandoned changesets in perforce.) Once you have an fsck method you like, its filtered output (i.e., --unreachable objects you want to keep) can be fed to git-prune.

Care would also be required with git-repack/git-prune-packed, but it seems mostly addressable with scheduling.

If I proceed down this path, I'd like to implement this procedure without any change in git's .c or .sh files. It's clear this is a minority use and should not depend on anything being maintained for it inside git. I would write a few bash scripts and a README/HOWTO for possible inclusion in contrib.

BTW, has anyone ever thought of writing an "Administrator's Manual" for git?

Thanks,

On Thu, 12 Apr 2007, Dana How wrote:

Yes.

The issues for hosting sites are very different from the issues of individual developers having their own git repositories, and I agree 100% that both alternates and shared object directories make tons of sense for hosting.

I hope it wouldn't even be a minority usage pattern. I am a firm believer that distributed SCM's and git in particular makes a lot more sense for source control hosting than CVS or SVN do. I'm really disappointed with things like sourceforge, and part of the problem is literally that a centralized SCM is really *fundamentally* wrong for a hosting entity.

Using a distributed SCM just makes _so_ much more sense for hosting projects, and I've actually very much wanted to try to make sure that git can help people who host things.

It's not my *own* primary use, but I think it's a very important usage pattern, even though it's very different froma "normal developer" private sandbox case.

So I think your case is really very interesting. I'd love to help figure out how to help you guys with git, but because it's not how I personally work, I can really just try to help when you actually hit a problem - you'll have to figure out what your usage patterns actually are on your own ;)

And btw, I think the shared object model really works very well, but I think it has to be paired with some stricter rules than people who use their own repos tend to have. For example, end-point developers have become very used to rebasing and generally rewriting history (or just resetting to an older state), and that's something that works find in a "local repository" setup, but it's also the kinds of patterns that can really screw you in a hosted and shared-object environment.

As to your two setups: I would suggest you go with the "hidden" shared version (ie people use the remote access pull/push to a server, and the *server* uses a shared object repository for multiple repositories), rather than having a user-visible globally shared object directory. Even with sticky bits and controlled group access etc, I think it's just safer to have that extra level of indirection.

(Partly because a globally visible shared object directory also implies that you'd use a networked filesystem, and I suspect a lot of developers would actually be a lot happier having their own development repositories on their own local disks, or at least some "group disk", rather than have one big and performance-critical network share. Even if you use some competent NetApp box and a modern network filesystem, it's just one less critical infrastructure piece that needs to be really beefy).

		Linus

Linus Torvalds wrote:

Would it not make sense for a hosting environment to say, if you are using alternates, or shared object directories, then you need to include *all* the refs in *all* the projects if you ever do an fsck?

I'm not sure how well git will scale in this case, although it just should be a matter of how well git scales to dealing with a single project with tens of thousands of refs/tags/etc. The only problem might be in passing all those refs/tags to fsck in one go. STDIN, I guess?

Rogan

On Fri, 13 Apr 2007, Rogan Dawes wrote:

Yes. And it shouldn't be hard to add support to do it. It's just not been done.

A lot of git programs already take refs on stdin, but fsck just doesn't do it (it can do it from the command line, but you'd run out of command line space very quickly).

More natural would be to just list all the git repos by git repo pathname (and there, usually the command line probably *is* long enough), but somebody would just have to do it. It's probably not very much code: just iterate over each repo both when adding refs and when actually doing the fsck itself.

For a real shared object directory, passing the refs to stdin (and teaching fsck about a "--stdin" flag) would be consistent with what we do for many other commands, so yes, that would work.

However, fsck actually tends to want not just the refs, but actually things like the index files and reflog files too, because those add other reachability info, which is why it's probably more natural to just give fsck the list of related repositories and let it figure them out.

That's also what you'd want to do for "alternates", since now there is no longer a single object directory either, but multiple separate (but related) ones.

Somebody would just have to write the code.. The basic rules are really all in "git/builtin-fsck.c": cmd_fsck(). Hint hint.

			Linus

On 4/12/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

For clarity I should have written *office* instead of *site* to describe my situation, but the mention of a NetApp below indicates no actual confusion occurred.

We did go down the local disk route, but after two significant losses of individuals' work, it was decreed that (perforce) work trees must be on the NetApp. So we already made the investment in beefiness -- for different reasons -- and I need to conform to these decisions for the moment.

After reliability, the other big criterion (especially with our penchant for large files) will be speed. With perforce, users now see submit={1 copy to server}, sync={1 copy from server}. In the short term I can't get away with changing this to submit={copy working to indiv repo, copy indiv repo to shared repo} and sync={copy shared repo to indiv repo, copy indiv repo to working}, because at first everyone will be trying to emulate what they did in perforce.

So probably I'll start out with either a very small testgroup, or one shared object repository with sticky/group tricks on the NetApp. Once git's collaboration advantages are apparent, I'll switch to the hidden repository model which I prefer as well. And hopefully these collaboration advantages will also mean people will commit more often and local disks can come back into favor -- and then the "extra" local repo file copy operations will be less noticeable.

In any event, I have some scripting to do to learn more about our usage patterns and pushing our datasets throught git. I also need to finish the pack-splitting patch (after 64b index goes in). Finally, before all that, I'll be out of the country for the next ~10 days...

Thanks,

On 4/11/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

Cool -- my previous email makes me either a git idiot or a git pioneer!

So I'll think through my usage model some more and look over the fsck source.

Until then,

Linus Torvalds wrote:

Well, not if the refs are only gitlinks because there is no checkout.

I think that's just the chicken-and-egg problem. Once this happens I think we'll see people aggregating all sorts of related repositories with this feature, and possibly making much richer histories by tracking portions of their trees as subprojects rather than just a subdirectory.

Sam.

Linus Torvalds <torvalds@linux-foundation.org> writes:

The small detail in the last step is wrong, though. Even if they EXIST, they may be isolated commits that are note connected to refs, and fsck in the repository would not have warned about unreachable trees from such unconnected commits. So you would need to do a reachability from these commits to the refs in the subproject.

This would be similar to the quick-fetch topic I sent out a couple of patches for, that implements logic to skip fetching objects from your alternate. You would have rev-list --objects traverse from them with "--not --all" in the subproject repository and make sure it does not trigger "I could not list all objects reachable from the commits you wanted because such and such tree/blob are missing".

    That reminds me of one thing I haven't verified.  I am not
    absolutely sure that rev-list --objects makes sure that
    blobs it lists exist (trees are checked as it needs to read
    them, and if they are missing or corrupt it would notice and
    barf).  When it is used for the purpose of this "subproject
    boundary fsck" and the quick-fetch, it should.  Perhaps a
    specialized option to check deeper than usual is needed.  I
    dunno.

This is still true.

On Wed, 11 Apr 2007, Junio C Hamano wrote:

The superproject *is* a ref.

You cannot prune the subprojects on their own. That's the *only* real special rule about subprojects. Exactly because pruning them on their own is not a valid op to do.

It's the same way with an source of "alternate" objects (or a shared object directory) - you'd better not prune them, because other projects may have refs to them that you don't know about locally. So this isn't somethign new to subprojects.

		Linus

Linus Torvalds <torvalds@linux-foundation.org> writes:

But when you fsck the subproject repository in isolation in the earlier step in your procedure, that is not taken into account, is it?

The situation I had in mind was not about pruning, but an earlier fetch, either the native one that unpacks the objects into loose form or a http walker, fetched a commit near the tip but was interrupted/killed before finishing the fetch nor updating the ref. The tip of such an incomplete commit chain would be reported dangling. They are ahead of your refs but they may lack commits and trees to complete the chain back to your refs yet. When the higher-level project points at such a commit, the existence of the commit is not a proof that everything needed to complete the commit is available.

We need to prove that separately, and that was my suggestion to run a "rev-list --objects $those-commits --not --all" in the subproject repository, simlar to what the quick-fetch topic does.

Junio C Hamano <junkio@cox.net> writes:

Ah, forget about this. The HEAD, which is in the tree of the higher-level project, is a ref. Silly me.

On Wed, 11 Apr 2007, Junio C Hamano wrote:

Well, not entirely "silly you".

If you do a "git reset" in the superproject, that will obviously have to rewrite the heads in the subproject.

I do suspect that we should always enable reflogs for the subprojects, so that pruning is safe even for these kinds of situations, but that doesn't resolve all issues.

For example: to manage *cloning* of the extra stuff, you might actually want to have externally visible refs, and while I suspect the main solution will always be to just do good maintenance (ie "don't do 'git bisect' and _never_ rewrite history in the main superproject!!"), I don't think it's out of the question to add other safety nets too..

So for example, while I'm not sure it's necessary, I don't think it would be *wrong* if we might eventually end up having *other* safety features like adding a totally separate "refs/superprojects/xyzzy" ref structure.

Or something like that.. Just to make the refs more visible both externally and internally, and to make it much harder to make stupid mistakes without realizing it.

I suspect a lot of this will depend on just how many mistakes people make. I don't think we've so far had a single problem with alternates files, re-basing, and people then pruning away objects used by other repositories by mistake, so maybe people really don't make those kinds of mistakes.

So maybe we don't need any extra safety nets at all. But who knows..

		Linus

On 4/11/07, Sam Vilain <sam@vilain.net> wrote:

This seems slightly related to the hazy picture I'm forming of how I'd like to use git at our site. Essentially, everyone would have their own working tree with .git directory, but .git/objects is a symlink to a shared object repository. How do you fully run git-fsck on this shared object repository? The actual heads (roots) are distributed amongst many .git/refs directories (I suppose you could do something akin to git-fsck $(cat /somepaths*/.git/refs/*), but that means you know where all the repositories are). So in this setup, maybe I'd want to run fsck twice: the first time checking everything but not complaining about dangling commit objects [but listing them?], and maybe a 2nd finding all these in the users' repos [still need to know where these are]. Please note this is just a thought experiment at this point.

Anyway, git started out with a 1:1 relationship between working tree, index, and object repository. Various things could weaken that -- alternates, subprojects with different relationships to their object repositories, etc. -- so special commands like git fsck which focus mostly on the object repository may need a little tweaking eventually.

This teaches the really fundamental core SHA1 object handling routines about gitlinks. We can compare trees with gitlinks in them (although we can not actually generate patches for them yet - just raw git diffs), and they show up as commits in "git ls-tree".

We also know to compare gitlinks as if they were directories (ie the normal "sort as trees" rules apply).

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
---

Ok, that's it for now.

NOTE NOTE NOTE! I'd like to note once more that this doesn't actually get you working subproject support. Not only do I need to connect up a few more low-level helper functions (things like "git diff" don't know how to generate even rudimentary "subproject X changed" patches, nor can you actually yet *add* subprojects), but quite apart from that low-level stuff, anything more high-level (like "git fetch" and friends) will need to know about subprojects.

In general, think of this like the early git plumbing: it's the early "content-addressable filesystem" part. The actual SCM parts going on top of it are yet to be done.

I'm hoping/expecting that there are more people who have the ability and the interest to work on the higher-level interfaces once the core plumbing support is there. There's still some plumbing to be done, but after that, maybe more people (and maybe the SoC people) can start filling out the higher-level details..

Comments on the patches/approach so far?

 builtin-ls-tree.c |   20 +++++++++++++++++++-
 cache-tree.c      |    2 +-
 read-cache.c      |   35 +++++++++++++++++++++++++++++++----
 sha1_file.c       |    3 +++
 4 files changed, 54 insertions(+), 6 deletions(-)

On Mon, Apr 09, 2007 at 09:20:29PM -0700, Linus Torvalds wrote:

Not that I have time right now to look up the exact context (only read the patch), but I would've expected a "case S_IFDIRLNK:" here?

Gruesse,

On 4/10/07, Frank Lichtenheld <frank@lichtenheld.de> wrote:

No, the st_mode comes directly from file system. It knows nothing about dirlinks.

On Tue, 10 Apr 2007, Frank Lichtenheld wrote:

So we have this strange (and worrying) dualism inside git: we use the same macros *both* for "stat data" *and* for "git-internal file modes".

So sometimes a mode is the result of a [l]stat() call like above, and then a gitlink is just a directory and we use S_IFDIR. And if it comes from the index, then it uses the internal git representation, and is S_IFDIRLNK.

I'm not very happy about it, but I'm actually most unhappy about it since I could imagine that the constants themselves are different on different OS's (eg VMS - a Unix-related OS will use the same constants for historical reasons).

In this particular place (index-path), we obviously not only have a stat() result, but more importantly, we never come here for a "normal" directory, since a normal directory would have been expanded into its component paths by the "read_directory()" logic.

So that interaction with directory expansion is somewhat non-obvious: normal directories are expanded recursively into the files they contain, while git directories end up being visible to internals as real directories, and are turned into gitlinks by code like the above.

		Linus

On Tuesday 10 April 2007, Linus Torvalds wrote:

So this does mean that the SHA1 of a gitlink entry corresponds to the commit in the subproject?

I wonder if it is not useful to be able to add some attribute(s) to a gitlink, i.e. first reference a gitlink object in the superproject, which then references the submodule commit, and also holds some further attributes. These attributes can not be put into the subproject, as it should be independent.

An example for such an attribute would be a subproject name/ID. An argument for this: The user should be able to specify some policies for submodules, like "do not clone/checkout this submodule". But the path where the submodule resides in a given commit is not useful here, as a submodule can reside at different paths in the history of the supermodule.

Josef

On 4/10/07, Josef Weidendorfer <Josef.Weidendorfer@gmx.de> wrote:

Right.

These attributes can be put into a file in superproject tree and checked in at the same as the gitlink. No real need for introducing another object type (right now there is no gitlink object type, just an entry in tree with special mode).

On Tuesday 10 April 2007, Alex Riesen wrote:

Like... .gitattributes ? ;-) Ok, this could work; however, there of course is the possibility of inconsistencies when e.g. manually moving subprojects around.

How is consistency ensured for .gitattributes ? I see that for .gitignore consistency, the user is responsible.

Josef

On Tue, 10 Apr 2007, Josef Weidendorfer wrote:

Yes.

The special "link" object has come up before, and I actually thought I'd do it that way first, but there were a few reasons why I didn't:

 - I tend to like "minimal", and the patches I sent out really are pretty 
   minimal, in the sense that they introduce just _one_ new concept, in 
   one place (it's basically a "tree entry" - so it shows up in tree 
   reading and writing, and nowhere else. The index, of course, is the 
   staging area for trees, so the index was also affected, but that was 
   really a very direct result of that "it's a new tree entry" thing).

 - in a "link" object, the only thing that would normally *change* is 
   really just the commit SHA1. Everything else is really pretty static. 
   As such, I decided that it's just a waste of a perfectly fine object to 
   have several thousands of the "link" objects that really only differ in 
   the pointer to the commit.

 - the "static" part, which you might as well have somewhere else, tends 
   to be stuff that you would need to be able to override locally, and as 
   such it does *not* really have a global meaning that is useful 
   historically.

   For example, the things that you'd want to associate with the gitlink 
   are things like "where would I find the repository that the commit is 
   part of" and "what is a description of that submodule" and "what are 
   the relationships between the submodules". These are things that aren't 
   necessarily even totally independent: in CVS, for example, you have 
   module names that are really not submodules themselves, but are really 
   just aliases for *collections* of submodules.

   So a 1:1 link object simply wouldn't make much sense anyway, and you'd 
   want to override those defaults with site-specific ones (maybe there is 
   a "canonical" address for the submodule repository, but if you have a 
   copy of it locally on-site, when you clone, you'd rather use the 
   *local* copy over the standard site, for example).

So all of this just made me say:
 - the tree entry just contains the commit ID of the subproject, and 
   *nothing* else.
 - any incidental data probably isn't 1:1 with tree entries anyway (both 
   over time: you have tree entries being updated with new commit ID's, 
   but the incidental data does *not* change, and over "space": different 
   repositories might want to use their local preferences for incidental 
   rules)
 - which all implies that the extra information should go in a separate 
   file that actually describes the modules.

In fact, it shouldn't be _one_ separate file: it should be at least two, since you'd want to have the *defaults* (which get cloned along with the superproject) in a revision-controlled file, and then have local *extra* information that is local.

This is exactly the same as the situation with the ".gitignore" file (which is revision-controlled and cloned with the respository) and the ".git/ignore" file (which is repository-local).

I've been thinking either ".gitmodules" (and ".git/modules") or to just extend the ".git/config" file parser to *also* parse a version-controlled ".gitconfig" file, and just describe the modules there. The config file really has pretty nice syntax, and I think module descriptions in many ways end up similar to remote branch descriptions, so it would fit in there, I think.

(But there's nothing that says that the ".gitmodules" file couldn't just use the same parser as the git config file, so I don't really strongly care either way. I just think it would be nice to be able to say

	[module "kdelibs"]
		dir = kdelibs
		url = git://git.kde.org/kdelibs
		description = "Basic KDE libraries module"

	[module "base"]
		alias = "kdelibs", "kdebase", "kdenetwork"

or whatever. You get the idea..)

		Linus

On Tuesday 2007, April 10, Linus Torvalds wrote:

Would it be nicer if .gitmodules were line-based to aid in merging?

Andy

On Tue, 10 Apr 2007, Andy Parkins wrote:

I seriously doubt you'll ever be merging or changing this a lot. So I don't think it's a huge concern.

		Linus

Linus Torvalds <torvalds@linux-foundation.org> writes:

I think Andy's comment comes from our earlier discussion on the other in-tree configuration, .gitattributes file.

We were talking about using in-tree .gitattributes for deciding if we apply crlf to each paths and other things like which 3-way file-level merge backend to apply, and need to make the system gracefully degrade even when in-tree .gitattributes have conflict markers during a merge. And for that purpose, it is certainly easier to arrange "pick each line, while ignoring <<< or === or >>>, and if there are conflicting duplicates do something sensible about them", if the file is line oriented.

But I do not think the .gitmodules thing needs that. If we have conflicting (or non-conflicting for that matter) submodule moves, that's a _MAJOR_ project re-organization, and I do not think we would even want to automatically descend into submodules for merging or checking-out when we have such a situation in the higher level project.

On Tue, 10 Apr 2007, Junio C Hamano wrote:

100% agreed. 

Also, note that while the ".gitmodules" (or whatever) file will be required to do things like "git pull", the basic tree-level logic that I sent out obviously doesn't need/use .gitmodules at all.

So there's a very real issue where a repository with submodules still "works", even with a .gitmodules file that is totally scrogged and doesn't have the right information (yet), it's just that it may simply not be able to do all the operations because it cannot figure out where to pull missing subproject data from etc..

So there is no reason to believe that we need to magically and automatically resolve conflicts - if conflicts happen, functionality is reduced, but it's not reduced so much that you cannot use the tree and try to resolve them (which is important, btw, since often before you commit your fix for the conflicts you'd want to *test* that fix, so we definitely don't want these kinds of files to be so central that it gets hard to get normal work done without them).

It really boils down to the same design issue: the way I think submodules should work is that they are very loosely coupled with the supermodule. The fact that the ".gitmodules" file isn't *that* critical comes largely from that loose coupling.

		Linus

Linus Torvalds wrote:

Whoa... "missing" subproject data?

Surely, unless you're doing lightweight/shallow clones, if you have a gitlink you've also got the dependent repository? Otherwise the reachability rule will be broken.

Sam.

hoi :)

On Thu, Apr 12, 2007 at 12:12:59PM +1200, Sam Vilain wrote:

With submodules you actually have a natural cutting point where you can say: no, I don't want to get that. So for submodules the reachability rule is a little bit more relaxed.

And when you fetch the superproject you now need some way to fetch the new submodule objects. They may be in the same upstream repository but it may make sense to have this configurable.

[ Dang. Power failure in the middle of writing emails. Can't remember 
  which one was lost. Am rewriting some of this reply in abbreviated form.  ]

On Thu, 12 Apr 2007, Sam Vilain wrote:

Absolutely. Not just subproject data. The whole subproject is often missing.

If I fetch the KDE superproject, I generally do *not* want every single subproject. In fact, I'd likely just want one or two subprojects.

The notion that all subprojects are populated is a *bug*. I would personally refuse to use such a setup. Even CVS can handle that just fine, we certainly don't want to be worse than CVS here.

If you just track a project, it's quite common to only check out the "src" module, and *not* fetch things like the "validation" or "test" module if you're just following along.

Or you might fetch the "kdebase" module, but that sure doesn't mean that you want all the other ones (kdevelop source code? full kdelibs sources? If I'm only interested in kwin and some other random app? No thanks!).

The reachability rule *must* be breakable. That's why fsck currently doesn't care AT ALL.

It's much better to break that rule than to even check it! I'd rather leave fsck like it is now, than to *ever* fix it, if the "fix" involves "you have to always fetch all submodules to shut fsck up".

		Linus

Linus Torvalds wrote:

Ok, but couldn't this be considered a variation of a lightweight checkout?

The only reason I'm worried about this is the case where the superproject contains *thousands* of subprojects. Eg, a superproject for all repo.or.cz projects. Say in a day 200 projects get updated with a few commits - do you have to do 200 pulls or just one? But maybe that problem can be solved in another way, or maybe it won't really hurt so much in practice and still be faster/more efficient than rsync mirroring.

This is especially the case in concert with gittorrent, which will need modifications to support sharing multiple repositories (not that that's a huge issue, given there's no implementation yet).

Well fsck can be fixed easily enough to not descend, like lightweight checkouts.

What I really want to avoid is the situation where you can't checkout, even though you didn't indicate a shallow/lightweight clone.

What else might this decision impact? Obviously with a smaller base you have fewer delta targets, though that's probably not a real issue.

Sam.

On Tue, 10 Apr 2007, Andy Parkins wrote:

this is very similar to the problem I asked about with merging config files a couple weeks ago. the answer then was that when we get .gitattributes we should be able to specify content specific merge programs that could deal with this sort of thing on a per-file basis. That sounds like the answer to your concern as well, rather then makeing things order dependant and otherwise harder to read to make it able to be merged with the current tools (which assume line-based order-dependant content)

David Lang

Andy Parkins <andyparkins@gmail.com> writes:

I personally feel that if there are cases that merge conflict is hard to resolve, there is something wrong in the communication between project members. In other words, merging this *should* be hard.

Really, if somebody wants to have project X at directory sub/X/ and somebody else wants the same at directory X/, merging the modules file would be the least of your concern -- resulting toplevel would not build correctly until you decide which tree hierarchy should be picked, and later exchange of results among project members would not be usable easily to half the people who picked the hierarchy differently from you did.

On Tuesday 10 April 2007, Linus Torvalds wrote:

So when moving the kdelibs submodule around, you would have to update the .gitmodules file.

I like it.

Josef

On Tue, 10 Apr 2007, Josef Weidendorfer wrote:

Right. The assumption here is:
 - submodules almost never actually change. You might add a new one 
   occasionally, and once a decade you might do some bigger 
   re-organization, but in general it's pretty much static.
 - when you do move submodules around, it's probably a big flag-day anyway 
   (ie I would expect that it's a big reorg, and that you'd quite likely 
   expect developers to have to re-check out their tree if you did major 
   surgery).

That's certainly how it works under CVS. I bet we can make it much nicer than CVS, but the point is, people really don't expect submodules to be something that you move around very dynamically. You want to be *able* to move them around, but it's not a normal operation.

The advantage with splitting things out like this is that it allows you much more flexibility than something automatic and deeply integrated does.

You can still edit the modules setup even if you yourself might not even have that particular module checked out! That may sound insane, but it's actually *required* for things like "oh, the standard server for that module went away, I need to edit the module settings to get it from xyz instead".

		Linus

Linus Torvalds wrote:

Also, in the Perl 5 Perforce conversion there are a number of "submodules" (ie, bundled modules with their own history) that move around a lot. In some tree representations used during the conversion process they might even appear twice in a given tree with differing versions.

Sam.

On Thu, 12 Apr 2007, Sam Vilain wrote:

That should actually be something that is fairly natural to handle with the current git submodule design - there's absolutely no problem with having the same subproject showing up in multiple different places in the tree (and each place obviously will have its own commit).

However, it causes some questions at two points:

 - What do you do in the ".gitmodules" file, where you describe the 
   submodule setup?

   This is not so much a _problem_ as a "how do you want to handle it" 
   issue.

   Would people want such a module to show up as "one module" that is just 
   visible in the tree in multiple places? Or do people prefer to think of 
   of it as completely separate modules that just happen to have the same 
   base repository?

   I don't think it's clear that one or the other is the "right way" to 
   see things, and I don't think git really should care. I suspect it's 
   more likely to be a detail that some importer script just has to 
   resolve one way or the other.

   The core git infrastructure needs to be able to have one module show up 
   in multiple places over time anyway, so I don't think there is any real 
   reason not to allow the same module to show up in multiple places even 
   within one single commit.. (Ie it's really mostly about the .gitmodules 
   file *syntax* - but if we use the config file syntax, it's actually 
   very natural to allow multiple entries for the module directory name)

   At the same time, there are reasons why you might want to consider them 
   separate modules too - maybe you want to *descibe* them separately, and 
   maybe one of the copies is used for "legacy support", and you might be 
   in a situation where you want to check out only one of the copies and 
   not the other (and thus describing them as two *different* modules 
   rather than two versions of the *same* module actually makes sense!).

   So I think this is something where we are technically neutral, but 
   where we may have non-technical issues to choose one representation 
   over another (and those issues may have more to do with the *importer* 
   than with any git issues - if importing from Perforce, it probably 
   makes most sense to make the import behave as much as possible the way 
   Perforce did in that case, and I have *no* idea what that is ;)

 - After a conversion is done, and you're no longer talking about a 
   historical archive, but a "going forward" concern, exactly how 
   automatic is subproject movement going to be, and what are downstream 
   developers that pull these things supposed to do when a subproject that 
   they have checked out is moved?

   This is mostly a UI issue. I suspect that the initial answer is: "you 
   may have to un-check-out a subproject, then pull the superproject, and 
   then re-check-it-out to get it in the new location". Simply because 
   it's going to be a lot easier to do than actually having "git pull" 
   notice when subprojects move.

   IOW, that is more of a "just how nice do we want to be to people", and 
   I _think_ the answer is: "as nice as possible, but some things are more 
   important than others, and some things might take longer before they 
   are really pleasant to do" ;)

Hmm?

		Linus

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

I like this concept.

I guess this file could also cover the case where the superproject is only interested in a small subset of the subproject. For example if I only uses some header-files in a library and want "/lib1/src/interface" in the subproject end up as "/includes/lib1" in the superproject. Could single files be handled in a similar way?

Although this is just an example, external links shouldn't be specified in the same configuration file as project internal things (which should be version-controlled). If the url configuration gets overwritten with checkouts there will be problems bisecting if the url changes over time.

hoi :)

On Thu, Apr 12, 2007 at 02:42:43AM +0200, Torgil Svensson wrote:

Conceptionally this information would have to be part of the supermodule tree (after all it changes how your tree is set up).

I think it makes more sense to make users think about which part of their tree can be reused and make them choose submodule boundaries wisely so that the above partial-checkout is not needed.

Most of the time we may not need to add any per-submodule URL information anyway. If you fetch a new supermodule version, you can get the new submodule from the same source (or from a per-submodule source which can be determined by looking at and munching the supermodule URL).

On 4/12/07, Martin Waitz <tali@admingilde.org> wrote:

I agree. This could be included in the module config file which in turn is version-controlled.

Sometimes you can't control upstream projects the way you want it. Also, splitting up projects for the potential need of future superprojects has several obvious disadvantages (multiple changelogs, versions etc). I don't see the subfolder checkout thing as a problem since the core plumbing in Linus's implementation doesn't care what's beneath the commit link. The subfolder checkout can "easily" be done in a porcelain.

It's more problematic if you want to cherry-pick individual files in a subproject. Here, I think the tight connection between links and directories to be too restrictive. Why does a subproject commit-link have to be represented as a folder?

//Torgil

Josef Weidendorfer wrote:

I mentioned this briefly on another strand of this thread, but I think that the simplest way to do this would be to just make refs/subproject/* populate itself sensibly when you commit in the superproject.

I mentioned refs/subprojects/path/branch before, but I think it would probably be the sort of thing that should be in the .git/config

Sam.

hoi :)

thanks Linus for your nice implementation. Your core code is so much nicer than my hacked-up prototype :-).

I only had little time to actually have a look at it but the core is very similiar to my approach and I'll try to rebase some of my code on top of yours in the following days.

The only thing I disagree with you is in using HEAD of the submodule:

On Mon, Apr 09, 2007 at 09:20:29PM -0700, Linus Torvalds wrote:

Always using HEAD of the submodule makes branches in the submodule useless.

Whenever you do a checkout in the supermodule you also have to update the submodule and this update has to change the same thing which is read above. Updating the branch which HEAD points to is dangerous. You could overwrite some unrelated branch just because the user forgot to switch back to his supermodule-tracking-branch. The user would always have to make sure that all the submodules are in the correct state for an update of the supermodule. Updating HEAD directly is possible now and may make some sense, but you still get problems when you want to switch to some temporary branch in the submodule. You have no chance to get back to the original supermodule version and now your temporary submodule branch gets shown as the new submodule version which should be part of the supermodule. The submodule version which is stored in the supermodules tree is kind of a hidden/remote reference/branch. When working on a remote branch we first create a local working branch and then sync it with the remote one. I think that it makes sense to use the same model for submodules: have one local branch in the submodule which is used for all work that is done in the supermodule context.

So my advice is: Always read and write one dedicated branch (hardcoded "master" or configurable) when the supermodule wants to access a submodule.

Then you have two type of branches: You can branch the supermodule and have you own branch of the entire project with all submodules. Use this if you want to commit your work on the submodule into the supermodule. You can also branch the submodule to effectively disconnect the submodule from the supermodule temporarily. You can use this to do some experimental/debugging stuff which should not yet go into the supermodule. Once you want this branch to show up in the supermodule, just merge it to "master" and commit it to the supermodule (and now its in the supermodule branch, the submodule branch is not needed any more).

See also the discussion about it in the messages around http://marc.info/?l=git&m=116636334226668&w=2

On 4/11/07, Martin Waitz <tali@admingilde.org> wrote:

In this case it does not correspond to the working tree anymore. HEAD is the "closest" to working tree of submodule.

hoi :)

On Wed, Apr 11, 2007 at 10:29:24AM +0200, Alex Riesen wrote:

yes.

This has been discussed in length already. Please have a look at the archives.

Your working tree now contains a complete git repository which has features which are not available for normal files. Notable, you have the possibility to create branches in the submodule. If you insist in using HEAD you throw away those submodule capabilities.

On 4/11/07, Martin Waitz <tali@admingilde.org> wrote:

"Yes" what? It should _not_ correspond to HEAD?

I should. But at least a short summary of the reasons would be nice.

In this (a very special, I believe) case, why not use git update-index --cacheinfo?

hoi :)

On Wed, Apr 11, 2007 at 10:49:18AM +0200, Alex Riesen wrote:

Not neccessarily, yes.

Branches in the submodule make no sense unless they are independent from supermodule branches. And then changing to another branch in the submodule automatically means that your current submodule working directory should be independent to the supermodule.

git-status in the supermodule should of course warn when a submodule is on a different branch, so that you don't accidently loose submodule commits which did not get committed to the supermodule.

I think misunderstood each other. For me branching is not special case.

Martin Waitz <tali@admingilde.org> writes:

Why? If you are working in the parent module (e.g integration) and notice breakage due to a bug in a submodule, it is very plausible that you would want to cd into the directory you have the submodule checked out, which has its own .git/ as its repository, and perform a fix-up there, with the goal of coming up with a commit usable by the parent project pointed at by the HEAD of the submodule repository. And while working toward that goal, you will use branches, rebase, rewind or use StGIT there in that submodule repository. It does not forbid you from using any of these things -- as long as you end up with a good commit at HEAD that the supermodule can use.

Once you come up with a suitable commit sitting at HEAD of the submodule repository, you cd up to the parent module. Top-level git-diff would notice that the commit recorded at the submodule path has been updated (because you now have a good commit at HEAD of the submodule repository, while earlier the one in your index was a dud).

So it is not clear to me what your argument about throwing away capabilities is.

hoi :)

On Wed, Apr 11, 2007 at 02:15:36AM -0700, Junio C Hamano wrote:

that's perfectly fine. I only require one more thing: make sure that your commit is on one dedicated branch (simply by merging your working/rebased/whatever branch into the dedicated one) and not on some random one.

Again: for your above example this is not neccessary and using HEAD
would indeed be perfectly fine.

But you also have to update the submodule when you do a checkout in the supermodule. So what do you update? Updating 'HEAD' is not very concrete, please have a look at my initial mail to Linus.

What is stored in the supermodule? It stores a reference to a specific point in the history of the submodule. As such I am convinced that the right counterpart inside the submodule is a refs/heads/whatever, and not the branch selector HEAD. You can have other branches next to the one which is tracked by the supermodule. If you always update HEAD you don't have a clear distinction between the branch which is tracked and other branches.

If the supermodule just updates some random submodule branch I happen to use at the time of a supermodule pull then submodule branches are of much lower value. Suddenly you have to make sure for yourself that the correct branch gets updated. For me, different branches should be independent and I want git to always update the correct one.

Martin Waitz <tali@admingilde.org> writes:

Because 'submodule' is a project on its own, it can make progress while the parent project is still using the stable commit. Think of this:

 - Your application uses product of another project as a
   library (e.g. you are doing video application and embedding
   ffmpeg).
   
 - Your 'master' commit records a commit in the library
   subproject.  Maybe library subproject declared stable 1.0 and
   that is what you used to integrate.

 - But being an independent project on its own, the library
   project can make progress, outside the context of this
   aggregated work (i.e. your application).  Next time you do:

	$ cd ffmpeg ; git fetch

   there may not be any branch that points at the exact "stable 1.0"
   commit.

When you do a "checkout -f --recurse-into-subprojects" from the toplevel, I suspect that you would need to detach HEAD in the subproject repository grafted in your application tree to move it to the exact commit the toplevel project (i.e. your application) wants, and match the working tree to that commit. The toplevel simply should _not_ have to care what branch that commit comes from.

hoi :)

On Wed, Apr 11, 2007 at 01:01:17PM -0700, Junio C Hamano wrote:

yes.

But why does everybody want to detach the submodule HEAD, instead of creating one 'special' branch which holds the commit which is used by the supermodule?

If you then want to switch to another submodule branch you loose the reference that comes from the supermodule.

I want to create the extra branch exactly _because_ there is independent work going on in the submodule (or the project it is based on). As you can switch between detached HEAD and an independent branch you can also switch between the 'supermodule branch' and independent branches -- only that you can easily switch back if you have an branch of your own.

BTW: I also think that your --recurse-into-subprojects should
be implied.
If you check out one index entry, you should be able to read it
back afterwards.  That is a nice property everyone expects from
normal files and we should try to keep that for submodules.
When checkout_entry wants to touch a submodule we can simply rewrite
the 'supermodule branch' in the submodule.  If HEAD happens to point
to it we also read-tree the submodule.
This is easy to understand and implement and I have some good experience
with this model.

On Thu, 12 Apr 2007, Martin Waitz wrote:

I don't think "everybody" wants it.

But the point is, *regardless* of whether you want a "detached HEAD" or you want a "'special' branch", you should always use HEAD to look up the commit, and using HEAD *allows* both (ie just make HEAD a symref to the 'special' branch if you want that behaviour).

And if you *do* use a special branch, HEAD *must* match that special branch anyway, since when you commit in the supermodule, the only behaviour that makes sense is to commit the currently checked out state!

And that is entirely appropriate.

But that still means that HEAD must point to that branch (when in the submodule), since that branch must be the one that is checked out. If it isn't the branch that is checked out, normal operations like "git diff" etc wouldn't make sense from the supermodule.

And that is why *regardless* of whether you use a special branch or not, HEAD is the right thing to look up.

		Linus

On Wednesday 2007 April 11 09:06, Martin Waitz wrote:

I know we've had this discussion before, but I'm going to bring it up again - mainly because Linus's implementation exactly matches what I envisaged when we originally spoke of this. I think in your "Updating the branch which HEAD points to is dangerous" section, the main thing you're not taking into account is that git can make detached checkouts. Updating HEAD is not dangerous - updating refs is; and I don't think anyone is proposing that a submodule ref should ever be updated by a supermodule.

I think you're also too strongly focussed on the idea that the supermodule tracks submodule branches - it cannot branches are not part of "the" repository they point at "a" repository. References are outside the repository pointing in, and hence the supermodule cannot refer to them at its core.

Now, if you check out a revision in the supermodule, that's going to look up 
the submodule revision stored in the DIRLINK tree entry which will recurse 
into the submodule and checkout that revision - almost certainly as a 
detached HEAD.  There are three possibilities then:
 - The submodule revision is in the past and no submodule branch points at it
 - The submodule revision is current and a submodule branch points at it
 - The submodule revision is current and multiple submodule branches point at 
   it
The supermodule checkout will have to make a decision whether to update the 
submodule HEAD (in one case it's obvious: a revision in the past has to be 
detached HEAD as there is no suitable branch).  It's also possible that the 
single submodule branch case is easy - undetach HEAD; however I don't think 
that is universally correct.

I know you're very much in favour of making branches in the submodule correspond to branches in the supermodule, but I just don't see a way of making it work - the supermodule cannot know about submodule branches, branches are not part of the repository, they just point at the repository. My branches could be different from your branches.

It may be that some handy configuration settings and some clever porcelain could keep them in sync for your working repository - but it's never going to be the case that checking out "master" in the supermodule can be universally resolved to mean "checkout master in the submodule".

The way submodules should be treated is that the whole submodule is analogous to a single repository-tracked file - that's essentially what a submodule is in the end but the content of the "file" is the submodule revision.

There is one difference from ordinary files, a submodule has two "modified" 
states, not one:
 1. HEAD of submodule is different from DIRLINK revision
 2. Submodule is dirty

In state (1) the submodule has to have git-add run on it in the supermodule, just as you would with a modified file, to get it into the index (or not if you don't want to commit that change). In state (2) it should be impossible to git-add, because the state of the submodule doesn't represent something that could be restored - there is nothing reasonable that could be written to the DIRLINK tree object. This is certainly a porcelain issue, because it's only really a warning that "git-add" isn't doing what you think it's doing when the submodule is dirty.

Now, if you change branch in the submodule, the supermodule will see that as a change in the submodule (as it should). If you changed back, it will be restored and the supermodule will again see it as unchanged. If you commit on the submodule, the supermodule will see that as a change and you'll have to git-add the submodule and commit in the supermodule. The submodule is on whatever branch it is on - at all times.

The only time I can see this causing difficulties is when you want to checkout the tip of a submodule branch - how is the supermodule to know when it is correct to change HEAD from being detached to being attached? I suppose it's got to be config-based; and out-of-tree config at that.

Andy

hoi :)

On Wed, Apr 11, 2007 at 10:47:00AM +0100, Andy Parkins wrote:

Then we already agree on the most important part. My argument is mostly against updating the ref which is behind HEAD, not HEAD per se. And I haven't thought about using detached HEADs until I wrote the mail.

No, that may be an misunderstanding because my very first prototype really did track branches. In the meantime I changed my mind, my current prototypes all track submodule commits directly. But in doing so we create a branch of its own: remember, a branch in git is just a moving reference into the history. Such a reference can be stored in .git/refs/heads or it can be stored in the index/tree of the supermodule. The difference is not really big. So we do not track a branch, but we create a branch by tracking.

I don't like to guess which branches to update. I'd prefer to just unconditionally update one specific one.

That would not work, you are right. Please see my above comment about tracking & branches.

Wholeheartedly agreed.

Again, doing things conditionally here just adds to confusion. Just have one dedicated branch and be done with it.

On Wed, 11 Apr 2007, Martin Waitz wrote:

Well, I don't actually see much choice. HEAD is just shorthand for "whatever is checked out".

No. 

Branches in submodules actually in many ways are *more* important than branches in supermodules - it's just that with the CVS mentality, you would never actually see that, because CVS obviously doesn't really support such a notion.

So I'd argue that branches in submodules give you:

 - you can develop the submodule *independently* of the supermodule, but 
   still be able to easily merge back and forth.

   Quite often, the submodule would be developed entirely _outside_ of the 
   supermodule, and the "branch" that gets the most development would thus
   actually be the "vendor branch", entirely outside the supermodule. Call 
   that the "main" branch or whatever, inside the supermodule it would 
   often be something like the remote "remotes/origin/master" branch.

   So inside the supermodule, the HEAD would generally point to something 
   that is *not* necessarily the "main development" branch, because the 
   supermodule maintainer would quite logically and often have his own 
   modifications to the original project on that branch. It migth be a 
   detached branch, or just a local branch inside the submodule.

 - branches inside submodules are *also* very useful even inside the 
   supermodule, ie they again allow topic work to be fetched into the
   submodule *without* having to actually be part of the supermodule,
   or as a way to track a certain experimental branch of the supermodule.

   I suspect that most supermodule usage is as an "integrator" branch, 
   which means that the supermodule tends to follow the "main 
   development", and the whole point of the supermodule is largely to have 
   a collection of "stable things that work together". 

   In contrast, branches within submodules are useful for doing all the 
   development that is *not* yet ready to be committed to the supermodule, 
   exactly because it's not yet been tested in the full "make World" kind 
   of situation.

I suspect (but will not guarantee) that the right approach is that a supermodule checkout usually just uses a "detached HEAD" setup. Within the context of the supermodule, only the actual commit SHA1 matters, not what branch it was developed on (side note: I haven't decided if we should allow the SHA1 to be a signed tag object too - the current patches obviously don't care since they never follow the SHA1 anyway, and it might be a good idea).

So I strongly suspect (and that is what the patch series embodies) that as far as the supermodule is concerned, it should *not* matter at all what branch the subproject was on. The subproject can use branches for development, and the supermodule really doesn't care what the local branchname was when a commit was made - because branch-names are *local* things, and a branch that is called "experimental" in one environment might be called "master" in another.

So once the commit hits the superproject, the branch identities just go away (only as far as the superproject is concerned, of course - the subproject still stays with whatever branches it has), and the only thing that matters is the commit SHA1.

I would strongly suggest that the *superproject* never really change the status of the subproject HEAD, except it updates it for "pull/reset", and then it just would use whatever the subproject decided to use.

The subproject HEAD policy would be entirely under the control of the subproject. If the subproject wants to use a branch to track the superproject, go wild: have a real branch that is called "my-integration" and make HEAD a symref to that (and thus any work in the superproject will update that branch - something that is visible when you pull directly from that subproject!)

But quite often, I suspect that a subproject would just use a detached HEAD. The subproject may have branches of its own, of course, but you can think of HEAD as not being connected to any of it's "own" branches, but simply being the "superproject branch". That's a fairly accurate picture of reality, and using "detached HEAD" sounds like a very natural thing to do in that situation.

So I really think you can do both, and I think using HEAD inside the superproject gives you exactly that flexibility - you can decide on a per-subproject basis whether HEAD should track a real local branch in a subproject, or whether it should be detached.

(Side note: if you do *not* use detatched HEAD, I suspect the .gitmodules file could also contain the branchname to be used for the subproject tracking, but I think that's a detail, and quite debatable)

So the main reasons I don't think that is a good idea are:

 - it's less flexible: see above on why you might want to use a dedicated 
   branch *or* just detached HEAD, and why you might want to choose your 
   own name for the dedicated branch.

 - it's also going to be quite confusing when the superproject sees 
   something *else* than what is actually checked out. This is an equally 
   strong argument for just using HEAD - when we actually implement a

	 git diff --subproject

   flag that recurses into the subproject, if you don't use HEAD inside 
   the subproject, that suddenly becomes a *very* confusing thing.

In other words, I really think HEAD is absolutely the right thing to use, but that said, I obviously wrote "resolve_gitlink_ref()" so that it can take any ref-name, and we *can* change that later, or make it a per-module config option or whatever.

		Linus

Linus Torvalds wrote:

To discuss this detail, what about keeping refs, such as refs/submodules/branch/path/* (or some other convention) which are updated on commit? Then you can also easily clone just the submodule.

Sam.

hoi :)

On Wed, Apr 11, 2007 at 08:16:10AM -0700, Linus Torvalds wrote:

I fully agree with you about the importance of submodule branches. In fact, I want to make them even more important and useable!

And by the way, I long forgot about CVS ;-)

I fully agree.

I fully agree. You are just so much better in describing things than I am...

If you use a detached HEAD then you can no longer switch back to it once you used some other (independent) branch (for testing or whatever). This is my main argument: If you just update some 'special' refs/heads/from-supermodule (or whatever, maybe get it from .gitmodules/config) you can still switch between branches, making them more useful IMHO.

If we create some other way to easily get to the commit referenced by the index of the supermodule then a detached HEAD is ok for me, too. But why create two things (this not-yet-existing way to get the supermodule index entry, plus submodules HEAD) for the same thing? Why not simply create a new refs/heads/whatever? This is easy and everybody knows how to work with it.

Fully agree.

Please don't confuse my "I always want to use one dedicated branch" with "I always want to use one special branch from the submodule project". This refs/heads/whatever I am talking about is _purely_ for ease of use of the submodule inside the supermodule. It is in no way linked to the branchnames that are used by the submodule project. Well, besides that you can merge back and forth between them, of course.

Fully agree.

So you now have this nice "my-integration" branch lying next to other independent (not-supermodule-related) branches. If you want to _switch_ to one of these unrelated branches you obviously have to change HEAD, and suddenly your unrelated branches are considered to be part of the supermodule (ok, not yet part of its index of course, but now all supermodule operations would work on this unrelated branch).

I want to preserve these unrelated branches and see them as a strong feature. Branches in submodules should be independent from the supermodule _because_ the supermodule has no notion of which branch is used.

Only that you loose your nice detached HEAD view once you start using those nice branches inside your submodule.

In terms of flexibility it is important what you can do with the submodule. Being able to use branches just like in a normal repository ("switch the branch to go to an other, unrelated branch") is a plus for me.

A detached HEAD does not give the same level of flexibility as a real head.

Well, the user explicitly expressed his intent to switch to another branch! In a normal repository you are not confused about the working directory not being in sync with "master", and we always prominently state which branch you are on. Of course this has to be clear for submodules, too. So if you do git-status in the supermodule it should print some "submodule is on different branch"-dirty marker.

At least I had some situations where I wanted to use something like this: use some experimental brach which should not be directly touched by the supermodule. Instead provide a method ("git merge from-supermodule") to sync your working branch with new stuff from the supermodule.

This is right. Suddenly we have one more player in the field which you can diff against.

Before submodules: tree <-> index <-> working file

submodules always using HEAD: tree <-> index <-> submodule HEAD <-> submodule working dir

submodules using some dedicated branch: tree <-> index <-> subm. "from-supermodule" <-> subm. HEAD <-> subm. wd

I haven't thought about which diff really makes sense in which situation.

On Apr 11, 2007, at 7:54 PM, Martin Waitz wrote:

Why can't can't we extend checkout with an option to look for an enclosing git project, find the gitlink in the index, and check out that commit? That allows you to return to the original state without needing to bother with new special branches.

And instead of recording the path in a .gitmodules file, why not a list of git directories we search for the commit? Allows moving of subprojects without suddenly breaking configuration files. When we find the appropriate git dir, we can use a .gitlink file or symlinks to attach the directory to it's repository.

I dislike moving git in the direction of enforcing more policy instead of less, and of making it less capable of handling content movement instead of more.

~~ Brian

On Thursday 12 April 2007, you wrote:

The supermodule checkout could create a .git/SUPER_HEAD for this. OK, that is a special kind of reference.

Or introduce "git --super ..." with works with the superproject. Form a submodule directory, a "git --super checkout ." could reset the submodule checkout.

Josef

On Mon, 9 Apr 2007, Linus Torvalds wrote:

Here is, for your enjoyment, the last patch I used to actually test this all. I do *not* submit it as a patch for actual inclusion - the other patches in the series are, I think, ready to actually be merged. This one is not.

It's broken for a few reasons:

 - it allows you to do "git add subproject" to add the subproject to the 
   index (and then use "git commit" to commit it), but even something as 
   simple as "git commit -a" doesn't work right, because the sequence that 
   "git commit -a" uses to update the index doesn't work with the current 
   state of the plumbing (ie the

	git-diff-files --name-only -z |
		git-update-index --remove -z --stdin

   thing doesn't work right.

 - even for "git add", the logic isn't really right. It should take the 
   old index state into account to decide if it wants to add it as a 
   subproject. 

so this patch really isn't very good, but it allows people who are interested to perhaps actually test something. For example, my test repo was actually created with this:

	[torvalds@woody superproject]$ git log --raw
	commit 649ad968bdd79cb3b0f50feb819b7e9b134d3a1a
	Author: Linus Torvalds <torvalds@woody.linux-foundation.org>
	Date:   Mon Apr 9 21:36:53 2007 -0700
	
	    This commits the modification to sub-project B
	
	:160000 160000 5813084832d3c680a3436b0253639c94ed55445d 17d246a35f27a46762328281eb6e9d4558f91e9d M      sub-B

	commit f3c55ffcc000a8c0fecc6801e8909d084e3d419e
	Author: Linus Torvalds <torvalds@woody.linux-foundation.org>
	Date:   Mon Apr 9 16:12:29 2007 -0700
	
	    Superproject with two subprojects
	
	:000000 160000 0000000... c0daf4c85d48879ab450a6a887bbb241eb0de00a A    sub-A
	:000000 160000 0000000... 5813084832d3c680a3436b0253639c94ed55445d A    sub-B

	commit 45eb14edb43b10e3d3ac7a495a1ec861e85dc36f
	Author: Linus Torvalds <torvalds@woody.linux-foundation.org>
	Date:   Mon Apr 9 15:36:24 2007 -0700
	
	    Add top-level Makefile for super-project
	
	:000000 100644 0000000... 57e8394... A  Makefile

so you can see how things look at a low level (ie a "gitlink" is just a tree entry with mode 0160000, and the SHA1 is just the SHA1 of the HEAD commit in the subproject)

		Linus

---

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

The other thing which will be missed a lot (I miss it that much) is a subproject-recursive git-commit and git-status. It is very possible that the default should be different for the git-commit and git-status: git-commit is likely to have it off whereas git-status will very much depend on how fast the usual response is (or wished for). An integrator on very fast machine may like it on for both, a subproject developer can have it off for both (to avoid accidental commits and generally being not interested in anything besides his code), an occasional person can have the status defaulting to on and commit to off - to avoid accidental commits in subprojects which are just tracked.

A separate config option and a command-line switch, probably.

On Tue, 10 Apr 2007, Alex Riesen wrote:

Note that I was definitely planning on adding them too, but they are at a higher level.

So the long-term plan is/was to add a flag to "git diff" (and "git ls-tree" etc) to say "recurse into subprojects".

You cound perhaps even make that flag the default with some .git/config option, if your superproject is small enough.

But this series of 6 (and the seventh ugly hack) is literally meant for just the really core object-handling stuff, and even there it's not really complete.

For example, you cannot even clone a superproject yet, simply because git-upload-pack doesn't know that it's not supposed to follow the gitlink things etc. So there's a lot of details left even for the really *core* stuff, but I wanted to post the series of six patches because those six patches are actually enough to reach the point where you can start looking at individual problems (like "git upload-pack") and fix them incrementally.

So I'd like this to be merged somewhere, not because "it works" or "it's complete", but because it's in a shape where I think a lot of people can start fixing small details.

For example, with just two smallish updates:
 - teach "git upload-pack" not to try to follow gitlinks
 - teach "git read-tree" to check out a git-link as just an empty 
   subdirectory
you should already be pretty close to being able to clone a superproject. 
You'd still have to clone the subprojects one-by-one manually, and that 
would be more of a porcelain'ish issue to teach git clone to fetch 
submodules too (with some ".gitmodules" file that contains the rules for 
that!)

But no, I didn't do any of that. I literally did just the "tree object format change" to support the *notion* of gitlinks - not all the pieces to then actually *implement* the notion are done by a long shot.

I think everybody agrees that we need some kind of subproject support, and the KDE repository certainly shows that subprojects need to be truly independent (because if they aren't, you end up with all the scaling issues that we see now - including something as simple as just "fsck" taking way way too long unless you have 4GB of RAM or more), and this sets the basic rules for that.

But they really are pretty low-level rules. For example, to go back to the KDE thing: we'd also need to teach *importers* to import certain subdirectories as submodules (or have a git->git translator that turns a subdirectory into a separate submodule).

So those are examples of things that obviously need to be done, and that my patches do not address in *any* way. They are really low-level plumbing support, kind of like the old original days when you had to run

	git-update-index ...
	tree=$(git-write-tree)
	commit=$(git-commit-tree -p $parent $tree <$msgfile)

by hand. A few monts later it was "git commit -a", but it started out with just fairly low-level plumbing..

		Linus

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

It is already "merged somewhere": as soon as the patches left landed on vger, it is not possible to loose (and even destroy) them. The feature is just too much sought after.

which also should fix switching between the branches with subprojects.

On Tue, 10 Apr 2007, Alex Riesen wrote:

Well, unless it hits something like Junios 'pu' (or 'next') branch, or somebody (like you?) ends up maintaining a repo with this, it's just unnecessarily hard to have lots of people working together on it..

I'm obviously interested in working on it, but at the same time, I don't expect to be a primary *user* of it, so I'm hoping others will come in and start looking at it.

It looks promising that you're getting involved, but I suspect you may be a bit too optimistic when you say "just too much sought after". We've been *talking* about subprojects for a long long time, and we've had other patches fail. So...

Yes. It would require either git-read-tree or the git-checkout script around it knowing to then also check out the subproject branches.

It's actually not *entirely* obvious what you should do when you switch branches (or even just do a "git reset --hard") in the superproject. The branches in the subprojects are likely to be totally different from the superproject, so as far as I can see, you end up having two choices when you reset a subproject:

 - either basically create a "disconnected HEAD" in the subproject(s) when 
   you switch them around as a consequence of resetting/switching the 
   branch in the superproject.

 - or you'd stay on the same branch in the subproject, and just reset that 
   branch..

 - or you describe the branch name in the ".gitmodules" file in the
   superproject, and use whatever branch in the submodule that is 
   described in the supermodule that you reset/check-out.

 - or possibly other policies.

So there is bound to be various "policy" issues like this worth sorting out. I don't think they matter that deeply.

I would _personally_ tend to like the notion of using ".gitmodules" in the supermodule to describe things like this, exactly because it's a policy decision - not something that git itself should really decide about, but that the supermodule maintainers can just decide to agree on.

But I haven't really even thought about all the things I'd want to have in the .gitmodules. We'd obviously need to list the default URL's for the submodules some way etc, but I haven't really sat down and thought about what all the higher-level porcelain really would need to know.

I suspect that somebody who has used and set up CVS "modules" setups should be thinking about that. I've been a "stupid user" for CVS modules setups, but I've never actually needed to really know how they *work*.

		Linus

On 4/10/07, Linus Torvalds <torvalds@linux-foundation.org> wrote:

The people who need the feature are still using other VCS. Some do not even know about git, the others are more interested in their own projects than in hacking on git (like KDE or Ubuntu people). And then there are commercial projects with thirdparty libraries, components or data. The other VCS' provide the feature, even if they do it wrong and badly (I never could go back in time in my day-work project, always asked myself what was the point of using Perforce at all). So, I suspect it is the people who are unable or unwilling to contribute to git (to anything, really) who need the feature most.

Linus Torvalds <torvalds@linux-foundation.org> writes:

Well, I was planning to apply this directly on 'master' after giving them another pass.

On Tue, 10 Apr 2007, Junio C Hamano wrote:

Goodie. I gave them another pass myself, and noticed a small leak and a stupid copy-paste problem, fixed thus..

		Linus

---

Linus Torvalds <torvalds@linux-foundation.org> writes:

Yeah, I noticed the first one but not the second.  Thanks.

By the way,...

People occasionally ask "how would I make a small fix to a commit that is buried in the history", so let me take a moment to give them a recipe.

Let's say while reviewing the code after applying all of the 6-series, you noticed the above thinko. First find out which commit caused it:

$ git checkout lt/gitlink $ git blame -L229,+7 master.. -- refs.c b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 229) if (!f) b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 230) re.. b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 231) read_packe.. b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 232) ref = refs.. b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 233) retval = -.. b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 234) while (ref.. b60108a1 (Linus Torvalds 2007-04-09 21:14:26 -0700 235) if..

The commit to fix is b60108a1 (this is what I have in my private repo, and I'll be rebuilding the series with this example, so you will never see this commit object name in the end result I'll be pushing out). So I detach the HEAD at that commit and make a fix:

$ git checkout b60108a1 $ edit refs.c $ git diff; # just to make sure $ git commit -a --amend

At this point, the detached HEAD and the original branch look like this:

$ git show-branch lt/gitlink HEAD
! [lt/gitlink] Teach core object handling functions about gitlinks
 * [HEAD] Add 'resolve_gitlink_ref()' helper function
--
 * [HEAD] Add 'resolve_gitlink_ref()' helper function
+  [lt/gitlink] Teach core object handling functions about gitlinks
+  [lt/gitlink^] Teach "fsck" not to follow subproject links
+  [lt/gitlink~2] Add "S_IFDIRLNK" file mode infrastructure for git links
+  [lt/gitlink~3] Add 'resolve_gitlink_ref()' helper function
+* [HEAD^] Avoid overflowing name buffer in deep directory structures

We fixed lt/gitlink~3 and the fixed-up commit is at HEAD. We want to rebase the rest of lt/gitlink on top of HEAD, like this:

$ git rebase HEAD lt/gitlink

This will take us back on lt/gitlink branch, set the tip of the branch to the commit we just made with the fix-up, and the first round will try to apply the change lt/gitlink~3 brings in on top of our HEAD. This _will_ fail, but that is to be expected, as we intend to replace that with what we just amended. Just reset it away and keep going.

$ git reset --hard $ git rebase --skip

Dealing with the other one in read-cache.c can be handled similarly after this. Luckily blame finds out that it is the last in the series (i.e. at the tip of lt/gitlink branch), so usual "fix the topmost commit" procedure applies.

$ edit read-cache.c $ git diff ;# checking... $ git commit -a --amend

On Tue, Apr 10, 2007 at 01:52:46PM -0700, Junio C Hamano wrote:

That recipe looks ummm complicated... What I usually do is:

git format-patch HEAD~4..HEAD git reset --hard HAED~4 patch -p1 < 0004* ...edit... delete diff from 0004* git diff >> 0004* git reset --hard git am 000*

Maybe this is as complicated as your example but this is very simple to deal with. And I do not destroy history or anything.

But that said I do not use topic brances but simply clone my local repository as needed. And I always deal with a linear history.

[I post this mostly to check if this is insane and I need to understand the way you propose to do stuff]

	Sam

Sam Ravnborg <sam@ravnborg.org> writes:

It's really the same. You keep 000* file, I keep them in the original branch and have "git rebase" take care of the details.

On Tue, 10 Apr 2007, Junio C Hamano wrote:

This is definitively good Documentation/howto/ material.

Nicolas

On Tue, Apr 10, 2007 at 05:03:13PM -0400, Nicolas Pitre wrote:

There's actually something similar already in "modifying a single commit" in the "user manual":

http://www.kernel.org/pub/software/scm/git/docs/user-manual.html#id276844

But it uses a throw-away branch instead of the detached head, and uses rebase --onto instead of rebasing and then --skip'ing.

--b.

Junio C Hamano <junkio@cox.net> writes:

Wouldn't

$ git rebase --onto HEAD lt/gitlink~3 lt/gitlink

do the trick in one step?

David Kågedal <davidk@lysator.liu.se> writes:

It is probably more Kosher, and I used to always do that, but it is much longer to type, and I use both perhaps 50%/50% depending on the mood.

When the fix-up only adds stuff, 3-way merge would say that the commit before fixing up (lt/gitlink~3 in our example, which you are explicitly excluding, while I am letting rebase to see it) has already been applied, in which case the procedure would not even stop. The case illustrated in my message which only adds a forgotten line "fclose(f)" falls into that category.

On Wed, Apr 11, 2007 at 02:32:48AM -0700, Junio C Hamano wrote:

Also remembering which commit you amended is a pain sometimes. So I usually do

	git tag base lt/gitlink~3
	git checkout base
	... edit and amend ...
	git rebase --onto HEAD base lt/gitlink
	git tag -d base

But the trick of letting the rebase fail and skipping looks less cumbersome.

--b.

hoi :)

On Tue, Apr 10, 2007 at 03:04:33PM +0200, Alex Riesen wrote:

git-status should really point out if a subproject has any changes, as it does for files. Only that a submodule may have more types of possible changes: has new commits which are not yet in the supermodule index, has an dirty index of its own, dirty working directory.

But for commit it really does not make any sense. The commit in the submodule is totally independent to the commit in the supermodule. You'd want the the submodule commit message to not refer to any supermodule stuff (as you likely want to reuse the submodule in other supermodules), while the supermodule commit is much more high-level and only records that the submodule got changed.

When viewed from the supermodule, a submodule is just part of its tree, just as normal files. So a submodule commit is conceptually similiar to changing a file, and you don't change files while you commit, also ;-).

On 4/11/07, Martin Waitz <tali@admingilde.org> wrote:

Only if I want it to. HEAD change check (which is cheap enough to be done unconditionally) can be done always.

Right. Perhaps not a commit in submodule but a recursive check for working directory changes in submodules. So that you can make that you don't make a superproject commit which cannot be resolved to what you had in all the working directories:

  git commit -a --check-clean-subprojects

hoi :)

On Wed, Apr 11, 2007 at 10:42:57AM +0200, Alex Riesen wrote:

Yes, that's the equivalent of checking normal files. The recursive check for dirty files/index should be configurable.

For -a such a check may even make sense unconditionally. And without -a I don't see any value in such a check. So we can just add that check to -a if we see that dirty submodules are a problem for users.