Stop refusing to run `git bisect` from a subdirectory

[dscho]

We never bothered to prevent git rebase to run in a subdirectory, and that would have the same ramifications as running git bisect from a subdirectory. So the latter's insistence on being run from the top-level directory does not actually make sense.

[nirmalhk7]

I would love to take this issue, but I still cant find how does Git realise the subdirectory the user is present in. Is there any C file, script or documentation that I could refer for this?

[dscho]

I would love to take this issue, but I still cant find how does Git realise the subdirectory the user is present in.

The easiest method is to use Git itself, in a worktree that contains a clone of https://github.com/git/git:

git grep "You need to run this command from the toplevel"

This will show you that the message comes from here: https://github.com/git/git/blob/v2.24.1/git-sh-setup.sh#L373

This line is within the git_dir_init() function, which is called directly afterwards (again, use git grep git_dir_init to figure this out, there is also a call in git-mergetool.sh, but as you can guess from the file name, that's not related to git bisect).

Now, the really crucial thing is that that message is produced only when this condition holds true:

	if [ -z "$SUBDIRECTORY_OK" ]

This is interesting, right? If the variable SUBDIRECTORY_OK is empty, then we run into this message, otherwise we do not. Let's search for this variable name:

$ git grep SUBDIRECTORY_OK
Documentation/doc-diff:SUBDIRECTORY_OK=1
Documentation/git-sh-setup.txt:given by `usage()` shell function.  `SUBDIRECTORY_OK` can be set
contrib/hooks/setgitperms.perl:#     SUBDIRECTORY_OK=1 . git-sh-setup
contrib/hooks/setgitperms.perl:#     SUBDIRECTORY_OK=1 . git-sh-setup
contrib/rerere-train.sh:SUBDIRECTORY_OK=Yes
git-instaweb.sh:SUBDIRECTORY_OK=Yes
git-legacy-stash.sh:SUBDIRECTORY_OK=Yes
git-merge-one-file.sh:SUBDIRECTORY_OK=Yes
git-mergetool.sh:SUBDIRECTORY_OK=Yes
git-request-pull.sh:SUBDIRECTORY_OK='Yes'
git-sh-setup.sh:        if [ -z "$SUBDIRECTORY_OK" ]
git-submodule.sh:SUBDIRECTORY_OK=Yes
t/t2300-cd-to-toplevel.sh:SUBDIRECTORY_OK=1

You will see that there are quite a few scripts that define this, but not git-bisect.sh. Let's look at one example how this is used: https://github.com/git/git/blob/v2.24.1/git-instaweb.sh#L23. So all we need to do is to define SUBDIRECTORY_OK=Yes just before sourcing git-sh-setup! That should be easy: just insert that line before https://github.com/git/git/blob/v2.24.1/git-bisect.sh#L35 and you should be set.

Actually not all that set. For the commit message, and also to build confidence in your patch, you should

• run the test suite (you can do this in a PR either here or in https://github.com/git/git)
• more importantly research where the SUBDIRECTORY_OK comes from (use git blame git-sh-setup.sh or git log -L :git_dir_init:git-sh-setup.sh and read the commit messages) and why git-bisect.sh does not have it (look specifically for such an explanation in those commit messages, you may also want to see where e.g. git-instaweb.sh got its SUBDIRECTORY_OK from.

[harry-hov]

Hey @nirmalhk7,
Are you still on it ??
If not, I would like to take this one.

[nirmalhk7]

Since I'm having trouble building Git project (some imap-send problem), I wouldn't mind if someone else picks this issue up

[sangu09]

Hey @dscho ! I found this issue is still present. I went through the above files and got some background on what needs to be done. Can I work on this?

[sangu09]

Oh yes, also I know that we are in a process to change bisect code from shell to C so is it that this issue might go on till hold until we complete the migration?

[dscho]

Go right ahead! And no, I don't think it should wait for the conversion to C to be complete, it should even help that transition.

[sangu09]

Oh okay! Thanks! I am on it!

[adlternative]

according to #765,this problem is not so easy to solve...I thought I needed to record the absolute path of a git root directory so that we could switch...

[newren]

The difficulties surrounding this issue ("What if the directory the user is in goes away?") are largely removed with da81d47 ("Merge branch 'en/keep-cwd'", 2022-01-05). In fact, one of the responses to that series when it was submitted was, well shouldn't we also make "git bisect" stop erroring out when running from a subdirectory? See https://lore.kernel.org/git/CABPp-BHhtZzq3rqbaHm62LOkA6cnzRC6cKokGP7s2ZsQzqjJrw@mail.gmail.com/.