"git rev-parse --show-superproject-working-tree" fails with "fatal: BUG: returned path string doesn't match cwd?" if supermodule has unmerged changes of the submodule reference

Sam McKelvie· Sep 24, 2018, 17:59 UTC · lore

I experienced this problem using git 2.17.1; however, from inspection of the next branch, function get_superproject_working_tree() in submodule.c has not changed in 2 years.

I believe the problem is related to the fact that when a merge of the submodule reference is in progress, "git --stage —full-name <submodule-relative-path>” returns three seperate entries for the submodule (one for each stage) rather than a single entry; e.g.,

$ git ls-files --stage --full-name submodule-child-test 160000 dbbd2766fa330fa741ea59bb38689fcc2d283ac5 1 submodule-child-test 160000 f174d1dbfe863a59692c3bdae730a36f2a788c51 2 submodule-child-test 160000 e6178f3a58b958543952e12824aa2106d560f21d 3 submodule-child-test

The code in get_superproject_working_tree() uses the “-z” option on ls-files, so it expects null-byte termination between entries. However, the computation of super_sub_len:

    		super_sub_len = sb.buf + sb.len - super_sub - 1;

will only work when there is exactly one entry returned. If this line is changed to:

		super_sub_len = strlen(super_sub);

then only the first returned entry is used, and the bug is resolved.

strlen() should be safe to use here because strbuf_read ensures the result buffer is null-terminated.

Thanks, Sam McKelvie

Stefan Beller· Sep 24, 2018, 19:15 UTC · re: Sam McKelvie · lore

Re: "git rev-parse --show-superproject-working-tree" fails with "fatal: BUG: returned path string doesn't match cwd?" if supermodule has unmerged changes of the submodule reference

On Mon, Sep 24, 2018 at 10:59 AM Sam McKelvie <sammck@gmail.com> wrote:

Show 21 quoted lines

>
> I experienced this problem using git 2.17.1; however, from inspection of the next branch, function get_superproject_working_tree() in submodule.c has not changed in 2 years.
>
> I believe the problem is related to the fact that when a merge of the submodule reference is in progress, "git --stage —full-name <submodule-relative-path>” returns three seperate entries for the submodule (one for each stage) rather than a single entry; e.g.,
>
> $ git ls-files --stage --full-name submodule-child-test
> 160000 dbbd2766fa330fa741ea59bb38689fcc2d283ac5 1       submodule-child-test
> 160000 f174d1dbfe863a59692c3bdae730a36f2a788c51 2       submodule-child-test
> 160000 e6178f3a58b958543952e12824aa2106d560f21d 3       submodule-child-test
>
> The code in get_superproject_working_tree() uses the “-z” option on ls-files, so it expects null-byte termination between entries. However, the computation of super_sub_len:
>
>                 super_sub_len = sb.buf + sb.len - super_sub - 1;
>
> will only work when there is exactly one entry returned. If this line is changed to:
>
>                 super_sub_len = strlen(super_sub);
>
> then only the first returned entry is used, and the bug is resolved.
>
> strlen() should be safe to use here because strbuf_read ensures the result buffer is null-terminated.

This is good analysis of the issue. Thanks for writing it up! Would you also mind to send a patch fixing the problem?

I agree that using strlen should work. I do not recall why I did not use it at the time of writing it.

Thanks, Stefan