Add mitigation for potential ssh stop hang

psiphon/common/crypto/ssh/mux.go

@@ -190,6 +190,29 @@ func (m *mux) loop() {
 		err = m.onePacket()
 	}
 
+	// [Psiphon]
+	//
+	// goroutine profile dumps from hanging shutdowns show that the mux.loop
+	// termination can be blocked in the channel.close loop below, while
+	// other goroutines calling channel write are blocked in
+	// handshakeTransport.writePacket, at handshakeTransport.writeCond.Wait,
+	// awaiting the completion of a new KEX -- which may be required once
+	// writeBytesLeft has been exhausted. At this point, when closing the ssh
+	// client, the underlying network connection has been closed and the KEX
+	// will never complete. As a workaround for scenarios where
+	// handshakeTransport.kexLoop may never call writeCond.Broadcast and
+	// unblock the channel writes, call Broadcast here. It's expected that
+	// the unblocked channel writes will go on to fail, assuming the
+	// underlying network connection is indeed closed; otherwise there's a
+	// chance that blocked channel writes will proceed under the previous KEX.
+	//
+	// Possibly related upstream x/crypto/ssh change:
+	// https://github.com/golang/crypto/commit/7292932d45d55c7199324ab0027cc86e8198aa22.
+
+	if h, ok := m.conn.(*handshakeTransport); ok {
+		h.writeCond.Broadcast()
+	}
+
 	for _, ch := range m.chanList.dropAll() {
 		ch.close()
 	}