test: add a test for issue #43645

2026-09-03 10:18:29 -07:00
parent 566436f1c6
commit 369bb5862b
1 changed files with 78 additions and 0 deletions
@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: LGPL-2.1-or-later
+
+set -eux
+set -o pipefail
+
+# shellcheck source=test/units/util.sh
+. "$(dirname "$0")"/util.sh
+
+# FailureAction=exit made the manager exit 0 when the unit failed before any of its
+# processes could run.
+# Issue: https://github.com/systemd/systemd/issues/43645
+
+UNITS="$(mktemp -d)"
+XDG="$(mktemp -d)"
+CG="/sys/fs/cgroup$(cut -d: -f3 /proc/self/cgroup)/issue-43645"
+
+cleanup() {
+    rm -rf "$UNITS" "$XDG" || :
+    rmdir "$CG/app.slice" "$CG" 2>/dev/null || :
+}
+
+trap cleanup EXIT
+
+cp /usr/lib/systemd/user/{basic,exit,paths,shutdown,sockets,timers}.target \
+   /usr/lib/systemd/user/systemd-exit.service "$UNITS/"
+
+cat >"$UNITS/issue-43645.target" <<EOF
+[Unit]
+Requires=issue-43645.service
+After=issue-43645.service
+EOF
+
+cat >"$UNITS/issue-43645.service" <<EOF
+[Unit]
+FailureAction=exit
+
+[Service]
+Type=oneshot
+Slice=app.slice
+ExecStart=false
+EOF
+
+# Move into a cgroup of our own, so that the manager we exec adopts it as its root.
+cat >"$UNITS/wrapper.sh" <<EOF
+#!/usr/bin/env bash
+set -eu
+mkdir -p "$CG"
+echo \$\$ >"$CG/cgroup.procs"
+exec "\$@"
+EOF
+chmod +x "$UNITS/wrapper.sh"
+
+# Runs a nested user manager, so that we can observe what it exits with.
+run_manager() {
+    local rc=0
+
+    rm -rf "$XDG" || :
+    mkdir -p "$XDG"
+    chmod 0700 "$XDG"
+
+    XDG_RUNTIME_DIR="$XDG" SYSTEMD_UNIT_PATH="$UNITS" \
+        timeout 60 "$UNITS/wrapper.sh" /usr/lib/systemd/systemd --user --unit=issue-43645.target || rc=$?
+
+    return $rc
+}
+
+# ExecStart= runs and fails: its exit status is propagated.
+assert_rc 1 run_manager
+
+# Now forbid the slice the unit lands in any descendants, so that creating the unit's
+# own cgroup fails. An ENOMEM fails the very same mkdir().
+mkdir -p "$CG/app.slice"
+echo 0 >"$CG/app.slice/cgroup.max.descendants"
+
+# The unit now fails before ExecStart= could run: no exit status to propagate, but the
+# manager must still not report success.
+assert_rc 255 run_manager