Case insensitive branch names

On the Git User's list, Diego J. reported that:

'When I "checkout" a branch using different Upper Case/Lower Case than the original, the branch doesn't show in "git branch [--list]"' [1]

While case sensitivity for filenames is a common issue on Windows and the like, I haven't seen any discussion regarding ref name sensitivity - any pointers to past discussions?

In particular, if I have a branch 'nocase', and ask to checkout 'NoCase', then git(-for-Windows) will happily say that 'NoCase' has been checked out, but then it gets confusing.

We are both, not on any branch, and yet not detached. In fact HEAD points to NoCase (which does not exist with that case setting). The checkout is actually of 'nocase', and 'git branch' does not mark any branch ref current (with the *). [see test script: 2].

While this is the way it is, should the 'git checkout' in some way check for the branch case sensitivity issue?

For example, maybe simply correcting the checkout name to that as recorded in the File System (but that has lots of code impact because the path and name are distinct elements of the branch struct), or maybe a warning could be emitted (config?), or even an err/die.

My gut feeling was that, at least for refs, that something should be done. My delving into the complexities of the code for 'git checkout' got me to refs.c ~#L1603 where the requested path is lstat'd. Perhaps this point is a special case for the mingw_lstat where the true pathname can be compared to the requested pathname (or at least the refs/heads/name part) to detect this mis-checkout case?

Thoughts? Is it a rabbit hole others have explored? Is it even worth pursuing?

-- Philip

[1] https://groups.google.com/forum/#!topic/git-users/EryCnwKL4_E [2] Script git init casetest cd casetest/

git add -A & git commit -m'first' git status git branch case git branch # two branches, * on master. git status git checkout Case # successful checkout, uses branch name as given! git branch # note that no branch is marked as current, but neither are we detached head.. git status cat .git/HEAD # HEAD records the requested name, not the branch that was actually checked out! # User left confused...

=== output Philip@PhilipOakley MINGW32 /c $ bash -v ./casetest.sh

git init casetest Initialized empty Git repository in C:/casetest/.git/ cd casetest/

git add -A & git commit -m'first'
[master (root-commit) 988c046] first
1 file changed, 0 insertions(+), 0 deletions(-)
create mode 100644 temp.txt
git status
On branch master
nothing to commit, working directory clean
git branch case
git branch
case
* master
# two branches, * on master.
git status
On branch master
nothing to commit, working directory clean
git checkout Case
Switched to branch 'Case'
# successful checkout, uses branch name as given!
git branch
case
master
# note that no branch is marked as current, but neither are we detached 
head..
git status
On branch Case
nothing to commit, working directory clean
cat .git/HEAD
ref: refs/heads/Case
# HEAD records the requested name, not the branch that was actually checked 
out!
# User left confused..

On Mon, Dec 21, 2015 at 6:01 PM, Philip Oakley <philipoakley@iee.org> wrote:

Multiple ref backend [1] should solve this.It's supposed to help handle a huge number refs. But a side effect from not using file system for refs is we can handle ref case-sensitivity much easier. And I believe David had that in mind he chose lmdb (because of Mac, which shares the same problem with Windows)

[1] http://thread.gmane.org/gmane.comp.version-control.git/281925

Duy Nguyen <pclouds@gmail.com> writes:

Yup, I had the same reaction. Instead of restricting the namespace of branches even on systems that do not have this problem, use a ref backend that is not limited by the underlying filesystem. A much better solution.

In addition to the LMDB backend, it might not be a bad idea to add another filesystem-based backend that encodes the refnames safely on case insensitive or case destroying filesystem. That way, those who do not want an extra dependency but do want case sensitive refnames would have an option, and having two non-default backends with quite different semantics may be a good way to ensure that the API for refs backend is kept sane.

On 2015-12-21 18.37, Junio C Hamano wrote:

This has been reported (probably a couple of times), one copy I have here was under "Branch Name Case Sensitivity" around Feb/March 2014.

The lstat() in refs.c can not be fixed, as the underlying OS/FS thinks that lstat("nocase") == lstat("NoCase") and open("nocase") == NoCase").

The the "real name" will not be detected, unless somebody does opendir(), readdir() and closedir(). This is expensive (in terms of execution time), and nobody has tried to do something.

One cheap solution would be to run "git pack-refs" internally, either from C-code inside Git itself, or from a script.

From: "Torsten Bögershausen" <tboegi@web.de>

My thought was more that on a case preserving FS we (that FS Git implementation) might warn if they have ended up not on a valid branch name. It felt wrong that the checkout reported success.

A suitable case sensitive, case preserving backend would solve it for those using it, which will be a help. Though those that don't will still need to be careful.

Found the thread http://thread.gmane.org/gmane.comp.version-control.git/242768/focus=242846 though I try to avoid having a 'don't do that' response for some of these potentially hazardous cases (though the cost benefit may not justify it;-)

That's probably a key bit of info I didn't know - that the true file name can't be known without that overhead. Various stackoverflow Q&As [1] show that others have also had some difficulties...

-- Philip [1] Windows methods of getting true file name (allegedly) http://stackoverflow.com/questions/4763117/how-can-i-obtain-the-case-sensitive-path-on-windows/4763137#4763137 http://stackoverflow.com/questions/478826/c-sharp-filepath-recasing/479198#479198 http://stackoverflow.com/questions/74451/getting-actual-file-name-with-proper-casing-on-windows