[PATCH 0/1] MR7174: mfplat: Introduce a MF startup count.

[Conor McCarthy (＠cmccarthy)]

In cases where MF is already shut down, simply forwarding MFShutdown() to
RtwqShutdown() will corrupt the Rtwq lock count if async result objects
still exist, because they hold a lock. JR East Train Simulator does this.

[PATCH 1/1] mfplat: Introduce a MF startup count.

[Conor McCarthy]

From: Conor McCarthy

In cases where MF is already shut down, simply forwarding MFShutdown() to
RtwqShutdown() will corrupt the Rtwq lock count if async result objects
still exist, because they hold a lock. JR East Train Simulator does this.
---
 dlls/mfplat/main.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/dlls/mfplat/main.c b/dlls/mfplat/main.c
index 4ed607686f1..9b73a5b9ef0 100644
--- a/dlls/mfplat/main.c
+++ b/dlls/mfplat/main.c
@@ -70,6 +70,8 @@ struct local_handler
     IMFActivate *activate;
 };

+static LONG startup_count;
+
 static CRITICAL_SECTION local_handlers_section = { NULL, -1, 0, 0, 0, 0 };

static struct list local_scheme_handlers = LIST_INIT(local_scheme_handlers);
@@ -1598,6 +1600,8 @@ HRESULT WINAPI MFStartup(ULONG version, DWORD flags)
     if (version != MF_VERSION_XP && version != MF_VERSION_WIN7)
         return MF_E_BAD_STARTUP_VERSION;

+    InterlockedIncrement(&startup_count);
+
     RtwqStartup();

return S_OK;
@@ -1610,6 +1614,15 @@ HRESULT WINAPI MFShutdown(void)
 {
     TRACE("\n");

+    /* NOTE: if startup_count is zero, calling MFStartup() and MFShutdown()
+     * simultaneously from different threads is inherently a race condition due
+     * to the requirement that MFShutdown() have no effect if startup_count is
+     * zero. No synchronisation is needed here. */
+    if (startup_count <= 0)
+        return S_OK;
+
+    InterlockedDecrement(&startup_count);
+
     RtwqShutdown();

return S_OK;

[Nikolay Sivov (＠nsivov)]

Is it possible we need to do that in rtworkq instead? It would be best if there was a way to verify which side this belongs to. According to docs mfplat is supposed to initialize winsocks too, so maybe it belongs to both sides.