Re: "git rm" seems to do recursive removal even without "-r"

From: Junio C Hamano <gitster@pobox.com>
Date: Oct 8, 2017, 04:20 UTC
Message-ID: <xmqqy3oms22q.fsf@gitster.mtv.corp.google.com>
In-Reply-To: <alpine.LFD.2.21.1710071749240.16818@localhost.localdomain>

"Robert P. J. Day" <rpjday@crashcourse.ca> writes:

Show 7 quoted lines

> ... so if, in the kernel source
> tree, i ran:
>
>   $ git rm \*.c
>
> i would end up removing *all* 25,569 "*.c" files in the kernel source
> repository.

Yes, as that is exactly what the command line asks Git to do.

If you said

    $ git rm *.c

then the shell expands the glob and all Git sees is that you want to remove a.c b.c d.c ...; if you said "git rm -r *.c", unless b.c is not a directory, these and only these files are removed.

Show 7 quoted lines

>   however, let's say i wanted to remove, recursively, all files with a
> *precise* (non-globbed) name, such as "Makefile". so i, naively, run:
>
>   $ git rm Makefile
>
> guess what ... the lack of globbing means i remove only the single
> Makefile at the top of the working directory.

Again, that is exactly what you asked Git to do.

    $ git rm $(find . -name Makefile -print)

would of course one way to remove all Makefiles. If you let POSIX shell glob, i.e.

    $ git rm */Makefile

the asterisk would not expand nothing but a single level, so it may remove fs/Makefile, but not fs/ext4/Makefile (some shells allow "wildmatch" expansion so "git rm **/Makefile" may catch the latter with such a shell).

By letting Git see the glob, i.e.

    $ git rm Makefile \*/Makefile

you would let Git to go over the paths it knows/cares about to find ones that match the pathspec pattern and remove them (but not recursively, even if you had a directory whose name is Makefile; for that, you would use "-r").

Earlier some people mentioned "Unix newbie" in the thread, but I do not think this is about Unix. In general, Unix tools do not perform grobbing themselves but expect the user to tell the shell to do so before the tools see the arguments. In that sense, I do think the combination of "-r" and globbing pathspec may produce a result that looks confusing at first glance.

"git rm [-r] <pathspec>..."

 (1) walks the paths it knows/cares about, rejecting ones that do
     not match the <pathspec>;

 (2) decides to remove the ones that match; and

 (3) when it is asked to recursively remove, the ones that are
     directories are removed together with its contents.  If it was
     not asked to go recursive, it refuses to act on directories.

where (1) and (2) are not something the tool needs to worry about---what is given from the command line is the only set of paths that the tool is asked to operate on. These two steps are quite unlike regular Unix tools.

Once you decide to give the tool the flexibility that come from taking pathspec, however, steps (1) and (2) do have to happen inside the tool. And great power takes some understanding of the tool on the part of the user to exercise. I suspect that the occasion you would need to use "-r" with "git rm" is a lot less frequent than a plain "rm".

Of course, there is no confusion if you do not quote wildcards. By quoting wildcards and not letting the shell to expand them, the user tells Git that the fact _it_ sees the asterisk is because the user is doing so on purpose---so that Git would find paths that match the pattern.

Hope this clarifies and helps.