Repository navigation
Description
Kiri version and system
0.2.1 on macOS 27.2.1 (26B131), Apple silicon. Docker CLI 29.8.2 (API 1.47).
What you did
Minimal reproduction, no DDEV needed, plain alpine:
# fails:
docker --context kiri run --rm --user 1000:1000 -w /does/not/exist alpine pwd
# both of these succeed, isolating the trigger to (non-root + missing workdir):
docker --context kiri run --rm -w /does/not/exist alpine pwd # root + missing dir: OK
docker --context kiri run --rm --user 1000:1000 -w /tmp alpine pwd # non-root + existing dir: OK
It also surfaces through DDEV (ddev start), where the ddev-webserver container runs as a non-root user with a WorkingDir that is not present in the image: the db container comes up healthy, the web container exits 127.
What happened
kiri: cannot start container: working directory: No such file or directory (os error 2)
The container exits 127 immediately. Via DDEV the same error appears as service web has exited; state=exited after Failed to chown Mutagen volume.
What you expected instead
Docker creates a missing WorkingDir (like mkdir -p) and starts the container. The same commands on OrbStack and Docker Desktop succeed. On OrbStack the auto-created directory is root:root, mode 0755, and the non-root process enters it fine (chdir needs only +x on the path).
Anything else
Root cause looks like a privilege-ordering issue in crates/kiri-init/src/linux/container.rs, finish_child() (at commit 7846fe0):
1343 setgroups / setgid / setuid # drop root -> container user
1346 chdir(p.cwd) # fails: dir missing
1347 // Docker creates the whole working directory path, like mkdir -p.
1348 fs::create_dir_all(p.cwd) # fails: now unprivileged, cannot mkdir under /
1349 chdir(p.cwd) # still fails
1350 fail("working directory", ...) # exit 127
The create_dir_all recovery runs after the setuid, so for a non-root container it has no permission to create the directory. The intent in the comment is right; the placement is after the privilege drop.
Suggested fix: create the directory while still root (before the setgroups/setgid/setuid block), then keep a single chdir after the drop. Matching Docker, leave it root:root 0755 (no chown):
// while still root, before setgroups/setgid/setuid:
// SAFETY: Docker creates a missing WorkingDir (mkdir -p), left root-owned 0755,
// so a non-root container can still chdir into it.
unsafe {
let _ = fs::create_dir_all(
std::ffi::OsStr::from_encoded_bytes_unchecked(p.cwd.as_bytes())
);
}
// ... existing setgroups / setgid / setuid ...
if libc::chdir(p.cwd.as_ptr()) != 0 {
fail("working directory", io::Error::last_os_error());
}I have not built/tested this patch locally (guest kiri-init cross-build), so treat it as a suggestion; the reproduction above is on a real machine.
Activity
verify: libracoder#370 workdir cases and e2e
Containers: a missing working directory is made as root, so a non-roo…
libracoder commentedon Nov 16, 2026
Thanks for tracking down the cause. It was exactly the ordering you described. Fixed in #400: Kiri now creates a missing working directory as root, mode 0755, before it switches to the container's user, as Docker does.
docker run --rm -u 1000:1000 -w /does/not/exist alpine pwdnow prints the directory, and DDEV's web container should start. It's set for 0.2.2. Checked on macOS, Linux and WSL 2.libracoder commentedon Nov 16, 2026
Update: the fix is now in #403, which replaces #400. It also covers
--read-onlycontainers:docker run --read-only -u 1000:1000 -w /does/not/exist alpine pwdstarts too.