Re: RFA: Various Windows (mingw32) additions, mostly relating to select or serial ports

[Daniel Jacobowitz <drow@false.org>]

On Sat, Feb 04, 2006 at 02:38:56PM +0200, Eli Zaretskii wrote:
> > Date: Fri, 3 Feb 2006 17:05:29 -0500
> > From: Daniel Jacobowitz <drow@false.org>
> > 
> > The primary ugly bit of this patch is the select wrapper.  Windows has
> > interfaces for all these things which map to Unix file descriptors, but
> > while the Unix interfaces are actually compatible, the Windows interfaces
> > are just designed along similar principles.  So you can handle a serial port
> > in roughly the same way you handle a console window.... but only roughly.
> > In fact you need to know what sort of device is behind each "file
> > descriptor", in order to handle it appropriately.
> 
> Could you elaborate a bit?  I cannot easily see where that device
> knowledge is present in the patch; it all looks to me like several
> instances of the same code with almost identical flow.
> 
> In particular, I thought WaitForMultipleObjects could handle any kind
> of handle, be it a pipe, a socket, a console, a process, or a file.
> 
> Even if the code is slightly different in that it calls different OS
> API functions, cannot it all be expressed as the same code that uses a
> function table indexed by the interface type?

Nope.  Well, in a sense, that is what I've done - the
serial_wait_handle interface feeds back to a central loop using
WaitForMultipleObjects.

The problem is that, yes, all these objects are HANDLEs, and
WaitForMultipleObjects can wait for many kinds of HANDLEs.  But there's
different things that "waiting for a handle" might mean, and it doesn't
happen to pick the right one.  So what I'm doing is using other
handles, controlled by threads or async I/O functions, to signal the
conditions we're interested in.

The four I implemented are:

- Serial.  Yes, I confess this is more different from the others than
it has to be; the most elegant way was to open the file in "overlapped"
mode, and then issue an overlapped (similar to non-blocking) wait for
EV_RXCHAR.  It could be reworked to use a thread, like the other three,
but it would still need to use WaitCommEvent.  I did this one first.
As far as I can tell you can't wait directly on a serial handle at all,
or if you can, the MSDN documentation doesn't tell you what will
cause it to become signalled.

- Console.  You can wait on a console handle, and Mark's previous code
did so.  However, it turns out, this signals for any event in the
input queue - including key release events!  But reading from the
console will block since that only fetches keypress events.  So,
we need to explicitly look for keypress events and discard other
things.

- Pipes.  There's just no wait function.  You can wait on a connected
pipe, but it's always signalled as far as I've been able to determine
(and it's undocumented).  You can wait on a named pipe, but that
only waits for a connection to be available, not data.

- Sockets.  They're pretty easy actually; you can associate the
socket with an arbitrary event object.  But afterwards you need
to make a socket-specific call to figure out whether you got
a read or an error; otherwise GDB doesn't detect hangups.

> I'd like an explanation about this paradigm:
> 
> > +      wait_events[0] = state->stop_select;
> > +      wait_events[1] = h;
> > +
> > +      event_index = WaitForMultipleObjects (2, wait_events, FALSE, INFINITE);
> > +
> > +      if (event_index == WAIT_OBJECT_0
> > +	  || WaitForSingleObject (state->stop_select, 0) == WAIT_OBJECT_0)
> > +	{
> > +	  CloseHandle (state->stop_select);
> > +	  return 0;
> > +	}
> 
> (You have similar code in gdb_select and elsewhere.)  Why do you need
> to wait for an object again with WaitForSingleObject, after you've
> just waited for it in WaitForMultipleObjects?

WaitForMultipleObjects returns a single result.  However, both of the
input objects could have been signalled before we woke.  The case this
is trying to handle is WaitForMultipleObjects returning
WAIT_OBJECT_0 + 1, and then stop_select being signalled; if someone has
asked for the thread to exit, we should do so immediately, because
the other handles we rely on will have been closed.

(Yes, this isn't 100% perfect on race conditions.  However, it is as
close as I was able to come up with.  I suppose I could kill and
restart the threads for every select...)

> >  - Windows serial support.  This definitely deserves a NEWS entry,
> >    included.
> 
> ...I don't think you included this entry.

Oops.  I added it while I was writing that bullet point.

+* Improved Windows host support
+
+GDB now builds as a cross debugger hosted on i686-mingw32, including
+native console support, and remote communications using either
+network sockets or serial ports.
+


-- 
Daniel Jacobowitz
CodeSourcery