[Bug]: mtree_spec generates incorrect spec file if the root directory has a file and a directory

[sin-ack]

What happened?

I'm using a vendored version of py_layer at work (from https://github.com/aspect-build/bazel-examples/blob/fba56194cd925b8d12b0339becbfc10bb9504c26/oci_python_image/py_layer.bzl). py_layer uses mtree_spec and tar from aspect_bazel_lib to create multiple tar files which are then passed to oci_image. As the input, I'm passing a py_binary like this:

py_binary(
    name = "project_bin",
    srcs = ["project/__main__.py"],
    deps = [":project_lib"],
    imports = ["."],
    main = "project/__main__.py",
)

py_oci_image(
    name = "image",
    base = "@distroless_python",
    binary = ":project_bin",
    entrypoint = ["/usr/bin/python3", "/project_bin"],
)

However, this doesn't work properly. The reason is because the mtree spec generated by aspect_bazel_lib looks like this:

project uid=0 gid=0 time=1672560000 mode=0755 type=dir
project_bin uid=0 gid=0 time=1672560000 mode=0755 type=file content=bazel-out/k8-fastbuild/...
project/__main__.py uid=0 gid=0 time=1672560000 mode=0755 type=file content=project/__main__.py

This unfortunately hits some (frankly very bizarre) rules in the mtree file format, specifically this one:

Relative
If the first whitespace-delimited word has no / characters, it is the name of a file in the current directory. Any relative entry that describes a directory changes the current directory.

This means that when bsdtar processes the first line, it "switches directory" to project/, and creates project/project_bin in the tar file rather than project_bin.

Additionally, we can't use /project_bin because mtree treats leading slashes as "special":

Special
Lines beginning with / are special commands that influence the interpretation of later lines.

./project_bin works, but shows up as ./project_bin in tar tf output as well (not sure if that breaks anything). It seems the only proper way to place files at the root directory is to make sure that files at root are always listed before the first directory entry.

Version

Development (host) and target OS/architectures: Gentoo Linux amd64

Output of bazel --version: 7.1.2

Version of the Aspect rules, or other relevant rules from your
WORKSPACE or MODULE.bazel file:

bazel_dep(name = "aspect_bazel_lib", version = "2.7.3")

Language(s) and/or frameworks involved: Python

How to reproduce

Run the following commands:

$ mkdir repro
$ cd repro
$ mkdir foo
$ touch foo/__main__.py
$ touch bar
$ cat > mtree.spec <<EOF
foo uid=0 gid=0 time=1672560000 mode=0755 type=dir
bar uid=0 gid=0 time=1672560000 mode=0644 type=file content=bar
foo/__main__.py uid=0 gid=0 time=1672560000 mode=0644 type=file content=foo/__main__.py
EOF
$ bsdtar -c -f repro.tar @mtree.spec
$ tar -tf repro.tar
foo/
foo/bar
foo/__main__.py

Any other information?

No response

[gregmagolan]

cc @thesayyn

[thesayyn]

That makes sense, IMHO non-breaking way to fix this is to add trailing slash to the dir entries so they are not interpreted as any special entry. Do you want to make a pr for it?

[sin-ack]

Sure!