Merge pull request 'fix: make plugin sync safe for concurrent replica starts' (#12) from benvin/fix-plugin-sync-race into main

Reviewed-on: #12

docker-entrypoint.sh

@@ -1,24 +1,63 @@
 #!/bin/sh
-# Sync image-baked plugins into the /config (datadir) plugins directory on every
-# start. /config is a PVC that overlays the image, so plugins baked into the
-# image are invisible until copied in here. Removing any existing versioned dir
-# of the same plugin first lets the image version win across restarts/downgrades.
+# Sync image-baked plugins into the /config (datadir) plugins directory on
+# start. /config is an RWX volume shared across replicas, so plugins baked
+# into the image are invisible until copied in here -- and this must be safe
+# when several replicas start (or restart) at the same instant:
+#   - skip entirely once the correct version is already in place, so the
+#     steady state (almost every start) never touches the shared volume;
+#   - install a new/changed version via copy-to-staging + atomic rename, so
+#     no reader (another replica, or this container's own jellyfin process)
+#     ever observes a partially-written plugin directory. A replica that
+#     loses the rename race just discards its own copy -- that's success,
+#     not an error;
+#   - drop stale, differently-versioned copies of the same plugin afterwards
+#     so they don't shadow the current one. Best-effort: another replica may
+#     already be doing, or have finished, the same cleanup.
+# Deliberately no locking: a lock held by a replica that dies mid-sync would
+# wedge every future start on this volume, which is worse than the race it
+# would prevent.
 set -eu
 
 BAKED_DIR=/usr/share/jellyfin/plugins-baked
 PLUGIN_DIR=/config/plugins
+STAGING_DIR="$PLUGIN_DIR/.sync-tmp"
+REPLICA=$(hostname)
 
 if [ -d "$BAKED_DIR" ]; then
-  mkdir -p "$PLUGIN_DIR"
+  mkdir -p "$PLUGIN_DIR" "$STAGING_DIR"
   for src in "$BAKED_DIR"/*; do
     [ -d "$src" ] || continue
-    name=$(basename "$src")   # e.g. "LDAP Authentication_22.0.0.0"
+    name=$(basename "$src")   # e.g. "LDAP Authentication_24.0.0.0"
     base=${name%_*}           # plugin name without the trailing _<version>
+    target="$PLUGIN_DIR/$name"
+
+    if [ ! -d "$target" ]; then
+      # Build the new version privately (keyed by this replica's own
+      # hostname, so concurrent replicas never share a staging path), then
+      # move it into place in one atomic rename. mv -T fails with
+      # "Directory not empty" if another replica's rename already won --
+      # that's fine, our copy just becomes garbage we discard.
+      staging="$STAGING_DIR/$REPLICA.$name"
+      rm -rf "$staging"
+      cp -a "$src" "$staging"
+      if mv_err=$(mv -T "$staging" "$target" 2>&1); then
+        :
+      elif [ -d "$target" ]; then
+        rm -rf "$staging"
+      else
+        echo "docker-entrypoint: failed to install plugin $name: $mv_err" >&2
+        exit 1
+      fi
+    fi
+
+    # Remove any other version of this plugin so it can't shadow the one
+    # above. Another replica may be racing the same cleanup, or have already
+    # finished it -- an entry that's already gone is success, not an error.
     for existing in "$PLUGIN_DIR/$base"_*; do
-      [ -e "$existing" ] && rm -rf "$existing"
+      [ -e "$existing" ] || continue
+      [ "$existing" = "$target" ] && continue
+      rm -rf "$existing" 2>/dev/null || true
     done
-    rm -rf "$PLUGIN_DIR/$name"
-    cp -a "$src" "$PLUGIN_DIR/$name"
   done
 fi