From f2f3d5af8fc25700134ec34a9a24729b57e79e65 Mon Sep 17 00:00:00 2001
From: darken <darken@darken.eu>
Date: Thu, 21 May 2026 18:00:03 +0200
Subject: [PATCH] fix: stop collectd before daemon-reload to free runtime tmpfs

Bookworm Pi-OS feeders default the runtime tmpfs to 43M and graphs1090's collectd in-memory RRD cache can occupy 34M of it. Systemd refuses daemon-reload below its 16M safety buffer, which under set -e aborts the entire upgrade before the feed bridge runs. Stop collectd to free its cache; the script reboots at its tail so collectd auto-starts on next boot. No-op when collectd is absent.
---
 test/update-airplanes-rootfs-smoke.sh | 13 +++++++++++++
 update-airplanes.sh                   | 14 ++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/test/update-airplanes-rootfs-smoke.sh b/test/update-airplanes-rootfs-smoke.sh
index d15cc2b..3d1dc12 100644
--- a/test/update-airplanes-rootfs-smoke.sh
+++ b/test/update-airplanes-rootfs-smoke.sh
@@ -359,6 +359,19 @@ assert_success_state() {
         || fail "release-channel = $(cat "$ROOT_DIR/etc/airplanes/release-channel"), want $expected_release_channel"
     assert_contains "$TAR1090_LOG" '^tar1090 install$'
     assert_contains "$COMMAND_LOG" '^apt-get install '
+    # Pre-flight before daemon-reload: stop collectd to free runtime tmpfs
+    # space so systemd's 16M safety buffer does not abort daemon-reload on
+    # bookworm Pi-OS feeders. Pins the behavior so the pre-flight cannot
+    # silently regress. The stub returns 0 for is-active so stop is always
+    # invoked under the test. Order matters — stop must precede reload.
+    assert_contains "$COMMAND_LOG" '^systemctl is-active --quiet collectd$'
+    assert_contains "$COMMAND_LOG" '^systemctl stop collectd$'
+    local stop_line reload_line
+    stop_line="$(grep -n '^systemctl stop collectd$' "$COMMAND_LOG" | head -1 | cut -d: -f1)"
+    reload_line="$(grep -n '^systemctl daemon-reload$' "$COMMAND_LOG" | head -1 | cut -d: -f1)"
+    if [[ -z "$stop_line" || -z "$reload_line" || "$stop_line" -ge "$reload_line" ]]; then
+        fail "expected 'systemctl stop collectd' before 'systemctl daemon-reload' (stop=${stop_line:-missing}, reload=${reload_line:-missing})"
+    fi
     assert_contains "$COMMAND_LOG" '^systemctl daemon-reload$'
     assert_contains "$COMMAND_LOG" '^systemctl enable airplanes-first-run.service readsb.service airplanes-mlat.service airplanes-feed.service pingfail.service$'
     assert_contains "$COMMAND_LOG" '^systemctl mask autogain1090.timer$'
diff --git a/update-airplanes.sh b/update-airplanes.sh
index 3a7e701..45f6473 100755
--- a/update-airplanes.sh
+++ b/update-airplanes.sh
@@ -196,6 +196,20 @@ fi
 rm -f "$(airplanes_path /etc/systemd/system/dhcpcd.service.d/wait.conf)"
 
 
+# Pre-flight: systemd refuses `daemon-reload` when the runtime tmpfs has
+# less than ~16M free (a safety buffer added in v249). On bookworm Pi-OS
+# feeders the runtime tmpfs defaults to ~43M and graphs1090's collectd
+# can occupy ~34M of it via its in-memory RRD cache, which drops us
+# below the safety buffer. With `set -e`, the daemon-reload failure
+# would abort the entire upgrade (no feed bridge, no service enables).
+# Stop collectd to free its runtime cache; this script reboots at its
+# tail so collectd auto-starts on next boot — no explicit restart
+# needed. No-op when collectd is absent (manual / non-graphs1090
+# installs).
+if systemctl is-active --quiet collectd 2>/dev/null; then
+    systemctl stop collectd || true
+fi
+
 systemctl daemon-reload
 
 # enable services