Git Rev News: Edition 28 (June 14th, 2017)

Welcome to the 28th edition of Git Rev News, a digest of all things Git. For our goals, the archives, the way we work, and how to contribute or to subscribe, see the Git Rev News page on

This edition covers what happened during the month of May 2017.



John Shahid sent an email saying:

Currently with 2.13 you cannot run git push --recurse-submodules=on-demand if the parent repo is on a different branch than the sub repos, e.g. parent repo is on “develop” and sub-repo on “master”.

John pointed to a test case he had created to show the problem, and he also had bisected it to a recent commit by Brandon Williams that teaches git push --recurse-submodules to recursively propagate the remote and refspec it is passed to the pushes performed in the submodules.

Junio Hamano, the Git maintainer, couldn’t tell from the report if the new behavior was a regression or a good thing, so he asked Brandon about his opinion on this.

Stefan Beller then explained that before Brandon’s patch, Git “just invoked git push in the submodule, assuming the user had setup a remote tracking branch and a push strategy such that git push would do the right thing”.

John then described in more details the previous behavior:

We have a parent repo on a branch called “develop” and a submodule on a branch called “master”. Prior to git version 2.13 if we had an unpushed commit in the submodule and ran git push origin develop --recurse-submodules=on-demand git would happily push the “develop” branch of the parent repo and the “master” branch of the submodule

and the new one:

After the change mentioned in my previous email, git would propagate the refspec from the parent repo to the submodule, i.e. it would try to push a branch called “develop” in the submodule which would error since no branch with that name exist in the submodule.

Brandon replied that:

Yeah my patches would definitely break that kind of workflow because they assumed that if you actually provided a refspec + --recurse that you would want it propagated down.

Jonathan Nieder then replied to John’s original email by explaining in details the context of Brandon’s patch, the goal of many of the recent submodule patch series from Brandon and others being “to allow thinking of a repository as a single unit, including submodules, most of the time”.

Jonathan also asked John:

Can you say more about how this change affects you? Would you be able to push without a refspec, or do you e.g. have scripting that was relying on the existing loose behavior?

John replied that he “loves the idea of having tools treat the repo as a unit”, but he “thinks propagating the refspec when no refspec is passed on the CLI is very surprising action that should be corrected”.

He also wrote that pushing without a refspec would work for him (if no refspec is propagated).

Unfortunately after that email at the end of May, it looks like not much happened, so recently John replied to his own mail:

bump. it’s been a while and I’m still not clear what the next steps are. I’m happy to send a patch but I would like to get a consensus first.

Eric Rannaud emailed the mailing list:

It used to be possible to run a sequence like:

  foo() { echo X; }
  export -f foo
  git rebase --exec foo HEAD~10

Since upgrading to 2.13.0, I had to update my scripts to run:

  git rebase --exec "bash -c foo" HEAD~10

Eric had bisected this to a commit that switched the interactive rebase to use the rebase–helper builtin.

An interactive rebase is usually a rebase that is passed the -i or --interactive option, though the --exec option automatically switches on an interactive rebase too.

The interactive rebase used to be implemented with a “git-rebase–” shell script, but, since the commit that Eric found, in most cases the interactive rebase would use the git rebase--helper builtin command.

The git rebase--helper builtin command had been made by porting shell code from “git-rebase–” to C.

Jeff King, alias Peff, replied to Eric that a git rebase --exec STRING command still uses a shell to run STRING, but that the behavior change may come from an optimization that was made in the prepare_shell_cmd() function in “run-command.c” to skip shell invocations for “simple” commands.

So Peff asked Eric to add an extraneous semi-colon in the STRING part to confirm that the problem comes from the optimization, as, with a semi-colon in it, the STRING part would not be considered a “simple” command.

Peff then explained that the optimization in prepare_shell_cmd() is quite old, but it affects git rebase --exec only now because it is now using C code. And this optimization fails only when people are using exported functions which is not so common, as not all shells support them.

Peff also suggested, and later sent a patch, to look for BASH_FUNC_* in the environment and disable the optimization in this case, though that wouldn’t work in all cases.

Junio Hamano replied to Peff, describing Peff’s analysis as “brilliant” and suggesting documenting the semi-colon trick.

Peff agreed with Junio on documenting it along with the optimization to skip shell invocations for “simple” commands, if it was decided not to try to fix the problem.

Meanwhile Johannes Schindelin, alias Dscho, who authored the patch that made the interactive rebase use the rebase–helper builtin, replied to Eric, saying that relying on exported functions to work in git rebase --exec was relying on an implementation detail.

But Peff later disagreed with Dscho saying that the user “expected the user-provided --exec command to be run by a shell, which seems like a reasonable thing for Git to promise (and we already make a similar promise for most user-provided commands that we run)”.

Dscho also sent a long reply to Junio about the reason he has been working on porting shell scripts like “git-rebase–” to C, which contained the following:

But the real truth is: shell scripting is not portable.

Shell scripting is never only shell scripting, of course. A quite undocumented set of utilities is expected to be present for our scripts to run, too: sed, awk, tr, cat, expr, just to name a few.

It does not end there. For example, sed is not equal to sed. BSD sed has different semantics than GNU sed, and we jump through hoops to try to ensure that our shell scripts run with both versions.

Eric then replied to Peff to “clarify if there was any doubt, the semicolon trick does indeed work fine”. He also asked for consistency writing that with an exported foo function the current behavior was:

  git bisect run foo     # works
  git bisect run 'foo;'  # doesn't work

  git rebase --exec foo master^^      # fails
  git rebase --exec 'foo;' master^^   # OK
  git rebase --exec 'foo 1' master^^  # OK

In the meantime Jonathan Nieder, Brandon Williams and then Kevin Daudt chimed in to discuss with Peff and Eric a suggestion by Peff to “speculatively run foo without the shell, and if execve fails to find it, then fall back to running the shell”. They also discussed a bit Peff’s suggestion to look for BASH_FUNC_* in the environment, but it was clear that both of these solutions had a number of problems.

Then Linus Torvalds chimed in too. He suggested to just lookup in $PATH for the first word passed in the --exec string, as this test is better than looking for special characters like semi-colons, to decide if we can avoid calling a shell.

He also suggested using vfork() instead of fork() if we “really want to optimize the code that executes an external program (whether in shell or directly)”.

It looks like Linus’ suggestions might be the best indeed, but it doesn’t look like much has been done to implement them yet.

Developer Spotlight: Brandon Williams

I’m a Software Engineer currently working on a team at Google focused on improving git, particularly making git behave better on large repositories (both mono-repos and repositories with many submodules).

Until about a year or so ago I never really thought I’d be a Software Engineer. Back in high school I took a physics class and a simple C programming class and I became fascinated with circuits and how computers worked. That led me to study Electrical Engineering during my undergrad followed by grad work in Computer Science focusing on Computer Architectures. Due to that schooling, I had various jobs in micro-processor design, varying from doing post-silicon power characterization, to ASIC design developing hardware modules in Verilog, to being an Architect modeling different micro-architectures in simulation.

Once I started at Google I pivoted my focus and now I’m a happy git contributor, though I still love discussing Computer Architecture!

Though I haven’t been around the project long I think some of my most important contribution has been to add submodule support to ls-files and grep.

Most of my contributions (and Stefan’s) are aimed at improving the submodule experience, which means adding submodule support to more and more git commands. Adding that support to ls-files and grep was a bit challenging in that our code base isn’t able to handle working on more than 1 repository at a time, so when you want to do any operation on a submodule you need to kick off a child-process to do the work for you. This ends up being very clunky to communicate needed information between the two processes, it would be so much easier if we could just operate on more than 1 repository in a single process.

So I’m aiming to solve this issue by introducing a ‘Repository Object’ to git. So instead of relying on global state, a reference to a ‘struct repository’ can be passed around allowing for multiple repositories to be open in a single process.

One of the biggest pain points for working in git is the large amount of hard-coded or global state we have. So I’d have a team work on redesigning large parts of the code base to make things more object oriented and to ensure there were good abstractions in place to facilitate changes down the line. This way transitioning to a new hash function, for example, would be less painful.

I’d like to drop anything that makes it difficult to scale git well with large repositories. One big one that has been discussed is improving the protocol to drop the initial ref advertisement. Unfortunately a change like that has the potential for breaking lots of old git clients.

I haven’t really used any git-related tools besides git itself. Though I have written a couple of tools myself to help with developing with git. One to view and explore a blame layer in vim, one to apply patches to a local branch from my mail client, and one to print out some information about a stack of commits I’m working on. I don’t think any of my tools are particularly well written, but they’ve been able to get the job done.


Other News


Light reading

Git tools and sites


This edition of Git Rev News was curated by Christian Couder <>, Thomas Ferris Nicolaisen <>, Jakub Narębski <> and Markus Jansen <> with help from Brandon Williams, Johannes Schindelin, Stefan Beller and Luca Milanesio.