Windows: Symlinks are not treated as a directory in gitignore
When there is a gitignore entry with a trailing slash, git_index_add_all will add symlinks that match the gitignore patterns on Windows but not other platforms.
Reproduction steps
The following Rust program demonstrates the problem. Sorry for only including a Rust reproduction, hopefully it is easy to see the equivalent C api calls (it is mostly 1:1). I couldn't figure out how to build a project with Visual Studio linking to libgit2 (if there are docs somewhere on how to do that, I'd be happy to make a C example).
use git2::*;
use std::fs;
fn main() {
// Create a repo with a gitignore file and a symlink.
fs::create_dir_all("repo/src").unwrap();
fs::write("repo/src/samplefile", "test").unwrap();
#[cfg(windows)]
{
std::os::windows::fs::symlink_dir("src", "repo/src2").unwrap();
}
#[cfg(unix)]
{
std::os::unix::fs::symlink("src", "repo/src2").unwrap();
}
fs::write("repo/.gitignore", "/src2/").unwrap();
let repo = Repository::init("repo").unwrap();
// Add all and commit.
let mut index = repo.index().unwrap();
index
.add_all(["*"], git2::IndexAddOption::DEFAULT, None)
.unwrap();
index.write().unwrap();
let tree_id = index.write_tree().unwrap();
let tree_oid = repo.find_tree(tree_id).unwrap();
let sig = repo.signature().unwrap();
let oid = repo
.commit(Some("HEAD"), &sig, &sig, "initial commit", &tree_oid, &[])
.unwrap();
// Check the contents of the commit.
let commit = repo.find_commit(oid).unwrap();
let mut names = Vec::new();
commit
.tree()
.unwrap()
.walk(TreeWalkMode::PreOrder, |_name, entry| {
names.push(entry.name().unwrap().to_string());
TreeWalkResult::Ok
})
.unwrap();
names.sort();
assert_eq!(names, [".gitignore", "samplefile", "src"]);
}
In this example, there is a file src/samplefile and a symlink src2 -> src. The .gitignore has a pattern /src2/ intending to prevent src2 from being added.
Expected behavior
git_index_add_all will only add .gitignore and src/samplefile.
Actual behavior
On Windows, it also adds the src2 symlink, causing the final assert to fail.
This seems to only happen with patterns with a trailing slash. A summary of the the gitignore pattern behavior:
/src2/andsrc2/fail on Windows./src2andsrc2work as expected on all platforms.
I've also tested with core.symlinks=true or false, it doesn't seem to affect it.
Version of libgit2 (release number or SHA1)
2a0d0bd19b5d13e2ab7f3780e094404828cbb9a7
Operating system(s) tested
Windows, macOS, Linux
Thanks @ehuss - you need not apologize for including rust repro steps.
Windows junction points / symlinks are ... complex, so apologies for the question. When you say that it's a symlink - how was this created? mklink /D ...?
Yea, CreateSymbolicLinkW with SYMBOLIC_LINK_FLAG_DIRECTORY, which I'm pretty certain is equivalent to mklink /d.
I am investigating the issue for gitoxide and wanted to share some findings that might be relevant for libgit2 as well.
When reproducing the issue on MacOS with this script…
git init excluded-symlinks-to-dir
(cd excluded-symlinks-to-dir
cat <<EOF >.gitignore
src1
src2/
EOF
git add .gitignore && git commit -m "init"
mkdir src
>src/file
ln -s src src1
ln -s src src2
)
…and running git status in `excluded-symlinks-to-dir…
❯ git status
On branch main
Untracked files:
(use "git add <file>..." to include in what will be committed)
src/
src2
nothing added to commit but untracked files present (use "git add" to track)
…I would have expected that git (git version 2.39.3 (Apple Git-146)) will only show src as untracked as it will resolve src2 as directory which matches src2/ in the .gitignore file.
Thus it seems that Git also doesn't handle this case. Arguments can probably be made for and against this behaviour, such that…
- …symlinks to directories are always considered files OR
- …symlinks to directories are considered directories
…by the exclude/gitignore machinery of the directory walk.
It seems that libgit2 changes the behaviour that I can reproduce with Git.
I would expect us to match git's behavior on POSIXy systems. But I think that windows is just different. There's several different types of things that are "links", all (IIRC) implemented as "reparse points", and potentially all treated a little bit differently. We should probably match git for windows behavior here, but I don't think that there's been some well-designed, thoughtful plan around this, I think it's just "whatever mingw does". So it will most likely be a disappointing change.