From 716e10e3fa86878015d8ea7822a4e448e5dbb9b1 Mon Sep 17 00:00:00 2001
From: Joe DiPrima <epilectrik@gmail.com>
Date: Mon, 10 Aug 2026 13:30:41 -0500
Subject: [PATCH] serial write hardening: bound the overlapped-write wait
 (500ms + CancelIo + 3s sick window) -- a dropped USB-serial link parked the
 game forever in GetOverlappedResult (ALPHA-MR minidump: SendPacket <- RIO
 ctor); txTO= counter joins the [rio] health line

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 engine/MUNGA_L4/L4PCSPAK.cpp | 53 +++++++++++++++++++++++++++++++++---
 engine/MUNGA_L4/L4RIO.cpp    |  3 ++
 2 files changed, 52 insertions(+), 4 deletions(-)

diff --git a/engine/MUNGA_L4/L4PCSPAK.cpp b/engine/MUNGA_L4/L4PCSPAK.cpp
index 144e85d..cab66c4 100644
--- a/engine/MUNGA_L4/L4PCSPAK.cpp
+++ b/engine/MUNGA_L4/L4PCSPAK.cpp
@@ -410,6 +410,20 @@ int PCSerialPacket::ReceivePacket(BYTE *destPtr)			// returns len (zero if none)
 // Returns:
 //		void
 //-------------------------------------------------------------------------
+//
+// Serial-write health (see the bounded wait below): after a write timeout
+// the link is SICK for 3 s and writes fast-drop -- one stall, not one per
+// poll.  BTPcspakTxTimeouts feeds the [rio] telemetry.
+//
+static unsigned gPcspakTxTimeout   = 0;
+static DWORD    gPcspakTxSickUntil = 0;
+
+extern "C" unsigned
+	BTPcspakTxTimeouts()
+{
+	return gPcspakTxTimeout;
+}
+
 void PCSerialPacket::SendPacket(BYTE *srcPtr)				//sends packet
 {
 	//--------------------------------------------------------------
@@ -418,6 +432,11 @@ void PCSerialPacket::SendPacket(BYTE *srcPtr)				//sends packet
 	if(!enabled)							//if disabled, don't even try
 		return;
 
+	if (gPcspakTxSickUntil != 0 && GetTickCount() < gPcspakTxSickUntil)
+	{
+		return;								// adapter sick: drop, don't stall
+	}
+
 	BYTE cmd = *srcPtr;						//get the command byte
 	cmd &= 0x7F;							//remove MSB
 	int length = txLengthPtr[cmd];			//get the length in al
@@ -459,11 +478,37 @@ void PCSerialPacket::SendPacket(BYTE *srcPtr)				//sends packet
          return;
       }
    }
-	DWORD bytesWritten;
-	GetOverlappedResult(hComm, &overlapSend, &bytesWritten, TRUE);
-	if(bytesWritten != length + 1)
+	//
+	// BOUNDED completion wait (2026-08-10, ALPHA-MR field minidump).  The
+	// RIO's "COM1" is a USB-serial adapter; when its USB link drops with a
+	// write in flight the overlapped write NEVER completes, and the original
+	// GetOverlappedResult(..., TRUE) parked the whole game at 0 CPU forever
+	// (cdb-proven: NtWaitForSingleObject <- GetOverlappedResult <-
+	// SendPacket <- RIO ctor).  Now: 500 ms bound (9600 baud needs ~21 ms),
+	// CancelIo on expiry, and a 3 s SICK window that fast-drops writes so a
+	// dead adapter costs one stall, not one per 50 ms poll -- the game stays
+	// at frame rate with a degraded RIO and the keyboard fallback alive.
+	// gPcspakTxTimeout feeds the [rio] health line (txTO=).
+	//
+	DWORD bytesWritten = 0;
+	DWORD wait_result = WaitForSingleObject(overlapSend.hEvent, 500);
+	if (wait_result == WAIT_OBJECT_0)
 	{
-		DEBUG_STREAM << "Not all data was written to serial port!\n" << std::flush;
+		GetOverlappedResult(hComm, &overlapSend, &bytesWritten, FALSE);
+		gPcspakTxSickUntil = 0;
+		if (bytesWritten != (DWORD)(length + 1))
+		{
+			DEBUG_STREAM << "Not all data was written to serial port!\n"
+				<< std::flush;
+		}
+	}
+	else
+	{
+		// No print here -- serial-path discipline; the counter surfaces in
+		// the [rio] health line (txTO=) from the glass-tick printer.
+		CancelIo(hComm);
+		++gPcspakTxTimeout;
+		gPcspakTxSickUntil = GetTickCount() + 3000;
 	}
 	delete buffer;
 }
diff --git a/engine/MUNGA_L4/L4RIO.cpp b/engine/MUNGA_L4/L4RIO.cpp
index 71adb88..2b34826 100644
--- a/engine/MUNGA_L4/L4RIO.cpp
+++ b/engine/MUNGA_L4/L4RIO.cpp
@@ -1267,6 +1267,8 @@ static unsigned gRioRRetryMirror   = 0;
 static unsigned gRioRAbandonMirror = 0;
 static unsigned gRioRFullMirror    = 0;
 
+extern "C" unsigned BTPcspakTxTimeouts();	// L4PCSPAK.cpp: bounded-write stat
+
 //
 // The printer half -- called from BTGlassPanels_Tick beside the [glassperf]
 // report (a stream site that prints every second without incident), NEVER
@@ -1326,6 +1328,7 @@ extern "C" void
 			<< " age=" << (gRioLastReplyTick != 0
 							? (now - gRioLastReplyTick) : 0)
 			<< "ms stalls=" << gRioStallCount
+			<< " txTO=" << BTPcspakTxTimeouts()
 			<< " lineErr=" << gRioLineErrMirror
 			<< " overrun=" << gRioOverrunMirror
 			<< " abandon=" << gRioAbandonMirror