Fix race condition

Author: royitaqi

lldb/tools/lldb-dap/DAP.cpp

@@ -898,6 +898,26 @@ bool DAP::HandleObject(const Message &M) {
 }
 
 void DAP::SendTerminatedEvent() {
+  // The folloiwng is to prevent races between the request handler thread and
+  // the event thread, where we are being asked to disconnect while the debugee
+  // process exits.
+  //
+  // The request handler thread, when entering this method, would have already
+  // acquired the API mutex (in BaseRequestHandler::Run). The next thing it
+  // tries to do is to acquire the call_once() mutex below.
+  //
+  // A deadlock will happen if the event thread happens to acquire the
+  // call_once() mutex first, then go into SBDebugger::Destroy() and eventually
+  // Target::Destroy(), where it tries to acquire the API mutex.
+  //
+  // To avoid this deadlock, we require that both threads acquire the API mutex
+  // first. Whoever gets it can go through the call_once() without issue.
+  //
+  // This is during the termination of a debug session, so performance loss
+  // introduced by this additional lock are hopefully not too critical.
+  lldb::SBMutex lock = GetAPIMutex();
+  std::lock_guard<lldb::SBMutex> guard(lock);
+
   // Prevent races if the process exits while we're being asked to disconnect.
   llvm::call_once(terminated_event_flag, [&] {
     RunTerminateCommands();