Merge pull request 'Repin upstream source to jellyfin-ha-src main' (#15) from benvin/repin-ha-fixes into main

Reviewed-on: #15

Dockerfile.runtime

@@ -27,7 +27,7 @@ RUN apt-get update \
 # Versions are the newest each plugin publishes whose targetAbi <= the pinned
 # Jellyfin server version (12.0.0):
 #   LDAP Authentication 24.0.0.0  targetAbi 12.0.0.0
-#   SSO Authentication  4.0.0.4   targetAbi 10.11.0.0 (newest release; loads on 12.0)
+#   SSO Authentication  5.0.0.0   targetAbi 12.0.0.0
 # sha256 pins make each fetch reproducible.
 FROM --platform=linux/amd64 debian:bookworm-slim AS plugins
 
@@ -37,10 +37,12 @@ RUN apt-get update \
 
 ARG LDAP_URL=http://artifactapi.k8s.syd1.au.unkin.net/api/v1/remote/jellyfin/files/plugin/ldap-authentication/ldap-authentication_24.0.0.0.zip
 ARG LDAP_SHA256=3be1f9d6a6ce9ea375e556dd30136d178a8dbe35cbe866d30d3451dc3ff7e804
-# LDAP is served through artifactapi remote. SSO is served through the artifactapi
+# Both are served through artifactapi remotes over http: this stage's base image
-# github proxy, which the CI build network can reach (github is not directly reachable).
+# trusts only public CAs and artifactapi presents an internal-CA certificate, so
-ARG SSO_URL=http://artifactapi.k8s.syd1.au.unkin.net/api/v1/remote/github/9p4/jellyfin-plugin-sso/releases/download/v4.0.0.4/sso-authentication_4.0.0.4.zip
+# https here fails to verify. The sha256 pins below supply the integrity guarantee.
-ARG SSO_SHA256=c09f16ba31059a434ddd7f811e4f9608d4b4c4514cc80a5bf1ca33bee61e1107
+# SSO is the in-house unkin fork (valkey-backed OAuth state, shared across replicas).
+ARG SSO_URL=http://artifactapi.k8s.syd1.au.unkin.net/api/v2/remotes/jellyfin-plugins/files/unkin/jellyfin-plugin-sso/5.0.0.0/sso-authentication_5.0.0.0.zip
+ARG SSO_SHA256=7e5f09cc4c81dce35edca650d74ed0680f425ca65c6221bb75456de9b8557e14
 
 WORKDIR /plugins
 RUN set -eu; \
@@ -50,8 +52,8 @@ RUN set -eu; \
     unzip -oq ldap.zip -d "LDAP Authentication_24.0.0.0"; \
     curl -fsSL "$SSO_URL" -o sso.zip; \
     echo "$SSO_SHA256  sso.zip" | sha256sum -c -; \
-    mkdir -p "SSO Authentication_4.0.0.4"; \
+    mkdir -p "SSO Authentication_5.0.0.0"; \
-    unzip -oq sso.zip -d "SSO Authentication_4.0.0.4"; \
+    unzip -oq sso.zip -d "SSO Authentication_5.0.0.0"; \
     rm -f ldap.zip sso.zip
 
 # ── Runtime stage ─────────────────────────────────────────────────────────────

UPSTREAM_REF

@@ -1 +1 @@
-10fed9409db4d65a101bdcbadaf86c8f2180a901
+6691b785c30715ec4561b00b7bd0fd112478d3e3

docker-entrypoint.sh

@@ -1,24 +1,63 @@
 #!/bin/sh
-# Sync image-baked plugins into the /config (datadir) plugins directory on every
+# Sync image-baked plugins into the /config (datadir) plugins directory on
-# start. /config is a PVC that overlays the image, so plugins baked into the
+# start. /config is an RWX volume shared across replicas, so plugins baked
-# image are invisible until copied in here. Removing any existing versioned dir
+# into the image are invisible until copied in here -- and this must be safe
-# of the same plugin first lets the image version win across restarts/downgrades.
+# 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