rtworkq: don't hold queues_section while shutting down the system queues

MFShutdown() held queues_section while waiting for pool callbacks, and a
source reader callback unlocking its user queue waited for the same
lock, hanging the Kingdom Hearts 358/2 Days port after its first
cutscene. Shut the system queues down outside the lock, like
unlock_user_queue() already does for user queues.

https://github.com/GloriousEggroll/proton-ge-custom/issues/899

Co-Authored-By: Claude Opus 5.5 <[email protected]>

patches/protonprep-valve-staging.sh

@@ -645,6 +645,10 @@ apply_all_in_dir() {
     echo "WINE: -WAYLAND- winewayland: animate animated cursors (game cursors static on Wine-Wayland, The Sims 3)"
     apply_patch "../patches/wine-hotfixes/pending/winewayland-animate-animated-cursors.patch"
 
+    # https://github.com/GloriousEggroll/proton-ge-custom/issues/899
+    echo "WINE: -GAME FIXES- rtworkq: don't hold queues_section while shutting down the system queues (MFShutdown deadlock, Kingdom Hearts 358/2 Days port hangs after the first cutscene)"
+    apply_patch "../patches/wine-hotfixes/pending/rtworkq-dont-hold-queues-section-during-system-queue-shutdown.patch"
+
     # https://github.com/GloriousEggroll/proton-ge-custom/issues/881
     echo "WINE: -HOTFIX- win32u: defer destroying NVIDIA Wayland host Vulkan instances while another is alive (NVIDIA driver clears its global libwayland table, EA app crash on winewayland)"
     apply_patch "../patches/wine-hotfixes/pending/win32u-defer-nvidia-wayland-host-instance-destroy.patch"

patches/wine-hotfixes/pending/rtworkq-dont-hold-queues-section-during-system-queue-shutdown.patch

@@ -0,0 +1,73 @@
+From: GloriousEggroll <[email protected]>
+Date: Mon, 2 Nov 2026 20:00:00 -0700
+Subject: [PATCH] rtworkq: Don't hold queues_section while shutting down the
+ system queues.
+
+shutdown_system_queues() held queues_section across shutdown_queue(),
+which waits for the running pool callbacks. A callback that unlocks a
+user queue takes queues_section too, so MFShutdown() deadlocked with it:
+
+  thread A: MFShutdown -> RtwqShutdown -> shutdown_system_queues
+            (holds queues_section, waits for the callback)
+  thread B: source reader callback -> RtwqUnlockWorkQueue
+            (waits for queues_section)
+
+unlock_user_queue() already shuts its queue down outside the lock.
+
+Shut the standard queue down first, without the lock. grab_queue() then
+fails, and it now rechecks that under the lock, so no new system queue
+is created behind us; taking the lock once waits for one still being
+created. Then shut the other system queues down, also without the lock.
+
+https://github.com/GloriousEggroll/proton-ge-custom/issues/899
+---
+--- a/dlls/rtworkq/queue.c
++++ b/dlls/rtworkq/queue.c
+@@ -665,6 +665,18 @@
+         struct queue_desc desc;
+ 
+         EnterCriticalSection(&queues_section);
++        /* Don't create a system queue after shutdown_system_queues() went past it. */
++        if (!system_queues[SYS_QUEUE_STANDARD].pool)
++        {
++            LeaveCriticalSection(&queues_section);
++            return RTWQ_E_SHUTDOWN;
++        }
++        if (queue->pool)
++        {
++            LeaveCriticalSection(&queues_section);
++            *ret = queue;
++            return S_OK;
++        }
+         switch (queue_id)
+         {
+             case RTWQ_CALLBACK_QUEUE_IO:
+@@ -1310,17 +1322,26 @@
+     unsigned int i;
+     HRESULT hr;
+ 
++    /* Shutting down a pool queue waits for its running callbacks, which may
++     * need queues_section themselves, e.g. to unlock a user queue. Don't hold
++     * it while waiting.
++     *
++     * Once the standard queue is shut down, grab_queue() fails and creates no
++     * more system queues. Taking the lock waits for one that is still being
++     * created, so that it is shut down below. */
++    shutdown_queue(&system_queues[SYS_QUEUE_STANDARD]);
++
+     EnterCriticalSection(&queues_section);
++    LeaveCriticalSection(&queues_section);
+ 
+     for (i = 0; i < ARRAY_SIZE(system_queues); ++i)
+     {
++        if (i == SYS_QUEUE_STANDARD) continue;
+         shutdown_queue(&system_queues[i]);
+     }
+ 
+     if (FAILED(hr = CoDecrementMTAUsage(mta_cookie)))
+         WARN("Failed to uninitialize MTA, hr %#lx.\n", hr);
+-
+-    LeaveCriticalSection(&queues_section);
+ }
+ 
+ HRESULT WINAPI RtwqShutdown(void)