Delphi 7, Windows 7, event handler, re-entrent code - delphi

Question

I've got some very old code (15+yr) that used to run ok, on older slower machines with older software versions. It doesn't work so well now because if fails a race condition. This is a general question: tell me why I should have known and expected the failure in this code, so that I can recognise the pattern in other code:
procedure TMainform.portset(iComNumber:word);
begin
windows.outputdebugstring(pchar('portset ' + inttostr(icomnumber)));
with mainform.comport do
try
if open then open := False; // close port
comnumber:=iComNumber;
baud:=baudrate[baudbox.itemindex];
parity:=pNone;
databits:=8;
stopbits:=1;
open:=true;
flushinbuffer;
flushoutbuffer;
if open then mainform.statusb.Panels[5].text:=st[1,langnum] {Port open}
else mainform.statusb.Panels[5].text:=st[2,langnum]; {port set OK}
except
on E: exception do begin
windows.OutputDebugString('exception in portset');
mainform.statusb.Panels[5].text:=st[3,langnum];
beep;
beep;
end;
end;
windows.outputdebugstring('portset exit');
end;
Note that flushinbuffer is protected with EnterCriticalSection(); AFAIK Nothing else is protected, and AFAIK there are no message handling sections. BUT
When this code is called from a click event, it gets part way through, then is interupted by a paint event.
The only tracing I have done is with outputdebugstring. I can see the first string repeated on entry before the second string is shown on exit. Is that real, or is it an illusion?
The trace looks like this:
4.2595 [4680] graph form click event
4.2602 [4680] portset 1 'from click event handler'
4.2606 [4680] graph form paint event
4.2608 [4680] portset 1 'from paint event handler'
4.2609 [4680] portset exit
4.3373 [4680] portset exit
This is a race condition: The paint event handler of the form is called before the click event handler code finishes, which causes failures. Serial code is AsyncPro. No thread code. Yes, there is more code, no it doesn't do anything in particular before "portset 1" but it does write to a form before it gets there:
with graphform do begin
if not waitlab.Visible then begin
waitlab.visible:=true;
waitprogress.position:=0;
waitprogress.visible:=true;
waitprogress.max:=214;
end;
end;
mainform.Statusb.panels[5].text:=gcap[10,langnum];
Don't hold back: What is it doing wrong, what should I be looking for?

Answer 1

This is expected behaviour - opening or closing a TApdComPort will service the message queue, specifically by calling a function it names SafeYield:
function SafeYield : LongInt;
{-Allow other processes a chance to run}
var
Msg : TMsg;
begin
SafeYield := 0;
if PeekMessage(Msg, 0, 0, 0, PM_REMOVE) then begin
if Msg.Message = wm_Quit then
{Re-post quit message so main message loop will terminate}
PostQuitMessage(Msg.WParam)
else begin
TranslateMessage(Msg);
DispatchMessage(Msg);
end;
{Return message so caller can act on message if necessary}
SafeYield := MAKELONG(Msg.Message, Msg.hwnd);
end;
end;
The TApdComPort is an async component - the com port is managed on background threads and opening or closing the port requires either starting or signaling those threads to stop. While waiting for them to free the component services the message queue in case it takes some time for things to synchronize (for example) :
if Assigned(ComThread) then
begin
{Force the comm thread to wake...}
FSerialEvent.SetEvent;
{... and wait for it to die}
ResetEvent(GeneralEvent);
while (ComThread <> nil) do
SafeYield;
end;
You haven't really show us enough of your own code to say why this is problematic in your case, however. I think David's point about com ports being manipulated in a paint handler is valid... we need to see the broader picture and what, exactly, the problem is that you are having.

Answer 2

A standard paint event cannot happen on its own, it can only be triggered by message retrieval. So the only way the code you showed could be interrupted the way you describe is if either the Serial component itself, or an event handler you have assigned to it, is doing something that pumps the calling thread's message queue for new messages.

Answer 3

Since you are closing the port in the beginning of your event handler, if there is any chance of triggering the event twice (i.e. by calling Application.ProcessMessages anywhere from your code, or calling TMainform.portset() directly from a worker thread), the new instance will close your port while the older one tries to communicate trough it, which will result in an error. AFAIS there are two solutions:
The faster but least bearable one is to protect your entire function with a Mutex (or event which is not a syncronisation object but can be used as one), but this only hides the coding error you have made.
The more pro solution is to find where the race condition gets raised, then fix your code. You can do it by searching all references to Application.ProcessMessages() and TMainform.portset(), and make sure that they won't get called paralelly. If no reference can be found on either mentioned function, the problem could still be caused by running multiple instances of your code ('cause it will not create multiple com ports :) ).

Answer 4

Remy Lebeau gets the credit for answering the question, because, as I asked for, it was a general reply to a general question. But it would have been inadequate without his comments in response to Uwe Raabe.
And what conclusively demonstrated that Remy Lebeau was correct was the exceptional answer from J, pointing out the specific point where the code failed.
Thanks also to David Heffernan for asking "why does code that responds to WM_PAINT call portset", which also makes a general point. And yes, the quick fix was just to block the path from the paint event handler to the comms code, but I'd done that without recognising the more general point.
I'll be having a look at the comms code, to see if there are more problems like this, and I'll be looking at the event handlers, to see if there are more problems like this, so thanks to everyone who read and considered the question.