Re: error building GUI on Mac OS-X 10.6.8

[John W. Eaton]

On 08/11/2012 01:06 PM, Michael Goffioul wrote:
> On Sat, Aug 11, 2012 at 5:53 PM, John W. Eaton <address@hidden
> <mailto:address@hidden>> wrote:
>
>     On 11-Aug-2012, Ben Abbott wrote:
>
>     |
>     | On Aug 11, 2012, at 12:13 PM, Michael Goffioul wrote:
>     |
>     | > On Sat, Aug 11, 2012 at 3:53 PM, Ben Abbott <address@hidden
>     <mailto:address@hidden>> wrote:
>     | >> The the "thread" subdirectory references removed
>     | >
>     | > OK, but if you remove moc_QTerminal.cpp and let moc recreate it
>     )when run from gui/src/), does it contain a reference to thread again?
>     |
>     | You are correct. The moc_QTerminal.cpp is recreated but without
>     the "thread" reference.
>     |
>     | I did this a second time to verify.  Now when I ./run-octave the
>     gui comes up!
>     |
>     | So what I did was ...
>     |
>     | (1) Apply the findFirst.diff patch
>     | (2) Commented out out the call to dissociate_terminal in
>     octave_start_gui in gui/src/octave-gui.cc.
>     | (3) Removed qterminal/libqterminal/moc_QTerminal.cpp
>     | (4) Ran "make"
>     |
>     | I'm able to type commands and plot, and don't see any focus
>     problems.  However, the keyboard mapping isn't correct.  When "less"
>     is paging the output, typeing "q" has no effect.  Neither does
>     ctrl-c.  The "Quit" from the MacOS menu bar also does not work.  So
>     I had to kill Octave from the MacOS activity monitor.
>
>     OK, can you try the following change and see if it helps?
>
>     If this change works, then we would need to find a way to call
>     dissociate_terminal earlier.  But I'd prefer to avoid calling it
>     unconditionally.
>
>     Even if this change does work, it might be best to find a way to
>     associate the keyboard (/dev/tty?) with the command window without
>     having to fork.
>
>
> That's what I try to achieved when I looked at it, but didn't succeed.
> The reasoning was the following:
> - to change the controlling terminal, one can call ioctl with TIOCSTTY
> - the above is only possible if the calling process is a session leader
> - to become a session leader, one can use setsid
> - setsid requires the calling process not to be a group leader
> - the recommended way to use setsid is to fork
>
> As I mentioned in one of my previous mail, low-level Linux expert needed.

I think what we need to do is give up the controlling terminal.

I remember that Jacob noted that "more" worked and "less" failed.  I did 
some poking around, and less will try to read keyboard input from 
/dev/tty if it can open that, otherwise it uses stderr (!) which seems 
to be some old Unixy behavior that just happens to work.  Without the 
fork, opening /dev/tty succeeds, but it is not associated with the 
QTerminal window.  After the fork and the call to setsid, opening 
/dev/tty fails and less tries to use file descriptor 2 (stderr) and that 
works, because we've set up the stdin, stdout, and stderr file 
descriptors so that they are connected to the slave process (Octave 
running in the separate thread, connected with the pty).  Forking 
doesn't matter for more because it always reads from file descriptor 2.

On OS X, does the attached patch cause less to work correctly in the 
Octave terminal window?  The call to ioctl is supposed to give up the 
controlling terminal.  It makes less work for me without the fork+setsid 
call.  Unfortunately, exiting Octave now crashes when the QTerminal 
object is deleted.  So although this change may provide a clue, it is 
not the ultimate solution.

The #ifdefs that I used are probably not ideal either.

Also, I now see that with either the fork+setsid call or the calls to 
ioctl, calling

  system ("emacs -nw")

fails with a message about not being able to open /dev/tty.  So neither 
of these solutions really seems right.  We really want to have a 
terminal widget that can run any program and properly handle I/O...

jwe

diffs.txt
Description: Text document