overlay/oxide/plugins/RunicGateway.cs

@@ -67,6 +67,25 @@ namespace Oxide.Plugins
         /// <summary>The longest line accepted from the sidecar. Over-long lines are discarded.</summary>
         private const int MaxInboundLineBytes = 1024 * 1024;
 
+        /// <summary>
+        ///     How long to wait for the TCP connect itself before giving up and backing off.
+        ///
+        ///     <para>
+        ///     It exists because <c>TcpClient.Connect</c> is blocking and has no timeout of its
+        ///     own, and the OS default is tens of seconds. A host that REFUSES answers instantly —
+        ///     which is every loopback test, and why this was invisible for two phases — but a host
+        ///     that <b>drops</b> does not answer at all: a firewall, a wrong address, a machine
+        ///     that is off. The link thread then sits in the connect, deaf to everything, and
+        ///     <c>Unload</c> waits out its whole join on the game's MAIN THREAD.
+        ///     </para>
+        ///
+        ///     <para>
+        ///     Generous against any real network and short against a hang. Loopback, which is the
+        ///     supported deployment, connects in well under a millisecond.
+        ///     </para>
+        /// </summary>
+        private const int ConnectTimeoutMs = 3000;
+
         /// <summary>
         ///     How often the boards are re-sent (PROTOCOL.md §8.3) and the per-player tallies are
         ///     flushed (§8.6). One cadence serves both because both answer the same question —
@@ -84,6 +103,18 @@ namespace Oxide.Plugins
         private static readonly ConcurrentQueue<string> Outbound = new ConcurrentQueue<string>();
         private static readonly AutoResetEvent Wake = new AutoResetEvent(false);
 
+        /// <summary>
+        ///     Set once, by <c>Unload</c>, and never reset. It is what lets a thread blocked on a
+        ///     CONNECT give up immediately instead of at the end of its timeout.
+        ///
+        ///     <para>
+        ///     It cannot be <see cref="Wake"/>: that handle also means "the queue has something in
+        ///     it", which <c>Emit</c> signals constantly, so waiting on it during a connect would
+        ///     abandon the attempt every time a hook fired. Two meanings, two handles.
+        ///     </para>
+        /// </summary>
+        private readonly ManualResetEvent _stopping = new ManualResetEvent(false);
+
         private PluginConfig _config;
 
         private Thread _link;
@@ -223,10 +254,19 @@ namespace Oxide.Plugins
 
         private void Unload()
         {
-            // Best-effort, and it has to happen BEFORE `_running` goes false, because `Emit`
+            // Attempted before `_running` goes false, because `Emit` refuses to enqueue once it
-            // refuses to enqueue once it does. A tally lost here is one interval of one player's
+            // does — but be honest about what this achieves: the writer stops on the same flag a
-            // gathering; flushing costs nothing and the ordering is the only reason it works at
+            // few lines below and the queue is cleared after the join, so a frame enqueued here
-            // all.
+            // only survives if the writer happens to be mid-drain. It usually is not.
+            //
+            // That is left as it is rather than fixed with a drain-before-exit, because draining
+            // means waiting on a socket from `Unload`, which runs on the MAIN THREAD — the exact
+            // stall phase 1 removed when a 1.9-second reload showed up as
+            // `Calling 'Unload' took 1918ms`. A wedged peer would make it worse than the bug.
+            //
+            // What is actually lost is under a minute of one player's gathering, and only on a
+            // manual plugin reload: a real shutdown flushes at `OnServerShutdown`, and a player
+            // leaving flushes at their disconnect.
             try
             {
                 FlushTallies();
@@ -243,6 +283,7 @@ namespace Oxide.Plugins
             }
 
             _running = false;
+            _stopping.Set();
             Wake.Set();
 
             Thread t = _link;
@@ -385,7 +426,7 @@ namespace Oxide.Plugins
 
                     client = new TcpClient();
                     client.NoDelay = true;
-                    client.Connect(host, port);
+                    Connect(client, host, port);
 
                     NetworkStream stream = client.GetStream();
                     stream.WriteTimeout = 5000; // a wedged peer must surface as an error, not a hang
@@ -479,6 +520,48 @@ namespace Oxide.Plugins
             _connected = false;
         }
 
+        /// <summary>
+        ///     Connects with a bound, and with a way out.
+        ///
+        ///     <para>
+        ///     The wait ends on whichever comes first: the connect completing, the plugin being
+        ///     unloaded, or <see cref="ConnectTimeoutMs"/>. Without the middle one a reload against
+        ///     an unreachable sidecar costs the game's main thread the whole join timeout —
+        ///     observed as <c>hook 'Unload' took longer than 100ms [2002ms]</c> beside
+        ///     <c>link thread did not stop cleanly</c>, which is phase 1's stall arriving by a
+        ///     different road.
+        ///     </para>
+        /// </summary>
+        private void Connect(TcpClient client, string host, int port)
+        {
+            IAsyncResult pending = client.BeginConnect(host, port, null, null);
+
+            int signalled = WaitHandle.WaitAny(
+                new WaitHandle[] { pending.AsyncWaitHandle, _stopping }, ConnectTimeoutMs);
+
+            if (signalled != 0)
+            {
+                // Closing the client is what abandons the in-flight connect; the socket would
+                // otherwise stay open behind an IAsyncResult nobody is going to end.
+                try
+                {
+                    client.Close();
+                }
+                catch
+                {
+                }
+
+                throw new System.IO.IOException(
+                    signalled == 1
+                        ? "unloading"
+                        : $"connect to {host}:{port} timed out after {ConnectTimeoutMs}ms");
+            }
+
+            // Throws the real socket error — a refusal, an unknown host — which is what the caller
+            // logs. Only reached when the connect actually completed or failed on its own.
+            client.EndConnect(pending);
+        }
+
         private void WriteLoop(NetworkStream stream)
         {
             while (_running && !_dead)
@@ -696,6 +779,12 @@ namespace Oxide.Plugins
                 { "serverId", _config != null ? _config.ServerId : "main" }
             };
 
+            // Resolved lazily while it is still unknown, which covers the case above on a
+            // framework that never fires `OnServerInitialized` — and costs one static property
+            // read per frame only until the answer exists.
+            if (string.IsNullOrEmpty(_wipeId))
+                _wipeId = ResolveWipeId();
+
             if (!string.IsNullOrEmpty(_wipeId))
                 frame["wipeId"] = _wipeId;
 
@@ -1035,6 +1124,17 @@ namespace Oxide.Plugins
         {
             MarkHook("OnServerInitialized");
 
+            // Re-read the wipe, and this is not belt-and-braces: at a real server
+            // start `Init` runs BEFORE the save is loaded, so `SaveCreatedTime` is not yet
+            // meaningful and the id resolved there is null for the entire session. This hook is
+            // the first moment the world is actually there.
+            //
+            // It was invisible on a development rig for the reason such things usually are: a
+            // hot-reloaded plugin reads an already-loaded world and gets the right answer every
+            // time. It took a server that BOOTED with the plugin installed — which is every real
+            // one — to show `wipeId=none` on a server with a save sitting on disk.
+            _wipeId = ResolveWipeId();
+
             // Started here rather than in Init: this hook fires both at the end of server startup
             // and immediately on a hotload, so it is the one place that runs exactly once per
             // load with the world available either way.