[Rd] sending signals to embedded R

[Luke Tierney]

On Sun, 6 May 2007 deepayan.sarkar at gmail.com wrote:

> On 5/5/07, Luke Tierney <luke at stat.uiowa.edu> wrote:
>> On Sat, 5 May 2007, Prof Brian Ripley wrote:
>> 
>> > On Sat, 5 May 2007, Luke Tierney wrote:
>> >
>> >> On Sat, 5 May 2007, Prof Brian Ripley wrote:
>> >>
>> >>> On Fri, 4 May 2007, Luke Tierney wrote:
>> >>>
>> >>>> On Fri, 4 May 2007, Deepayan Sarkar wrote:
>> >>>>
>> >>>>> On 5/4/07, Prof Brian Ripley <ripley at stats.ox.ac.uk> wrote:
>> >>>>>> On Fri, 4 May 2007, Deepayan Sarkar wrote:
>> >>>>>>
>> >>>>>>> one thing I haven't been able to figure out from R-exts is how to
>> >>>>>>> interrupt a calculation running inside an embedded R. C code inside
>> R
>> >>>>>>> calls R_CheckUserInterrupt() intermittently to check for 
>> interrupts,
>> >>>>>>> but how does my GUI tell R that the user wants it interrupted?
>> >>>>>>
>> >>>>>> Well, the intention is that you send an interrupt, which hardly 
>> needs
>> >>>>>> to
>> >>>>>> be in the manual.
>> >>>>>
>> >>>>> I didn't mean to imply that it does. I'm just new to signals and
>> >>>>> things that should be obvious aren't.
>> >>>>>
>> >>>>> Basically kill(2) seems to be the right thing to use, but I wasn't
>> >>>>> sure what the PID needs to be. Turns out sending SIGINT to my GUI 
>> from
>> >>>>> a shell interrupts R, so raise(SIGINT) should be enough.
>> >>>>
>> >>>> The tricky bit here is figuring out who does the sending.  It you have
>> >>>> a separate thread/process for the GUI and R then that is fine (though
>> >>>> may raise other issues).  If it is a single thread then you need your
>> >>>> event processing to get an occasional look in to recognise the user
>> >>>> action that triggers an interrupt. The Windows version handles this by
>> >>>> having R_CheckUserInterrupt() do a limited amount of event processing
>> >>>> (you need to be careful in GUI events have R actions associated with
>> >>>> them).  I believe the Mac version is similar though it has been a
>> >>>
>> >>> I was assuming that Deepayan's GUI (which seems to need Qt4, BTW, so I
>> was
>> >>> unable to compile it)
>
> Why is that a problem? As far as I can tell, Qt4 can peacefully
> coexist with Qt3, and while compiling you just need to use the right
> qmake (this is currently hardcoded in the configure script
> unfortunately).
>
> Qt4 has been around for a while now, and it's GPL on Windows/Mac as
> well as X11, so it seemed like the natural choice.
>
>> >>> worked via the R-Unix eventloop, in which case it
>> >>> gets some CPU time from time to time.
>> >>
>> >> I was assuming that as well.  But my recollection is that on unix the
>> >> event loop is only run from within the console reader.  On Windows
>> >> (and Mac OS X I believe) some event processing also happens in
>> >> R_CheckUserInterrupt(); on Windows there is also some more in some
>> >> blocking library calls, like socket reads as I recall.  But unless
>> >> things have changed since I last looked none of that happens on unix.
>> >>
>> >>>
>> >>> gnomeGUI has an interrupt menu item with action 'onintr', which may 
>> well
>> >>> be what Deepayan is looking for: the only reason that package still
>> exists
>> >>> is to provide example code.  (Not that it was ever properly integrated
>> >>> with the R event loop.)
>> >>
>> >> It does have some sort of interrupt device (I can't recall if it is a
>> >> menu item or a butto and I can't seem to build a working gnomeGUI to
>> >> check). And I believe if you try to use that item (or button?) during
>> >> a long-running computation you can't because the events won't be
>> >> looked at until R gets back to a console read, at which point the
>> >> events will be processed and you jump to the top level (where you
>> >> already are).
>> >
>> > That belief is correct (it has a menu item and a button), but my final
>> > parenthetical remark was that gnomeGUI was not wedged into the event 
>> loop.
>> >
>> >>> If the issue is what happens when the user Ctrl-C's in the GUI console,
>> >>> that depends on what the GUI toolkit does with keyboard input: if it
>> >>> generates a SIGINT this should just work, but otherwise the keyboard
>> >>> handler needs to be told to call onintr() one way or another.
>> >>
>> >> Again only if the GUI gets a chance to look at the keyboard input,
>> >> which I don't think we currently give it.
>> >
>> > We builtin the ability for a front-end to register handlers with the R
>> event
>> > loop, including a polling handler (and that is how we can have a Tcl/Tk
>> front
>> > end).  That postdates gnomeGUI, which runs the Gtk event-loop, not R's.
>> >
>> 
>> I had forgotten about that -- thanks for the reminder.
>> 
>> However, R_PolledEvents is only called from a limited set of places
>> now (including the socket reading code to keep things responsive
>> during blocking reads).  But it is not called from the interupt
>> checking code, which means if a user does something equivalent to
>>
>>     while (TRUE) {}
>> 
>> there is not point where events get looked at to see a user interrupt
>> action. The current definition of R_CheckUserInterrupt is
>> 
>> void R_CheckUserInterrupt(void)
>> {
>>      R_CheckStack();
>>      /* This is the point where GUI systems need to do enough event
>>         processing to determine whether there is a user interrupt event
>>         pending.  Need to be careful not to do too much event
>>         processing though: if event handlers written in R are allowed
>>         to run at this point then we end up with concurrent R
>>         evaluations and that can cause problems until we have proper
>>         concurrency support. LT */
>> #if  ( defined(HAVE_AQUA) || defined(Win32) )
>>      R_ProcessEvents();
>> #else
>>      if (R_interrupts_pending)
>>          onintr();
>> #endif /* Win32 */
>> }
>> 
>> So only on Windows or Mac do we do event processing.
>
> Is there a reason R_ProcessEvents cannot be set on Unix but can on
> Mac? It doesn't seem user-settable on Windows, but whatever the built
> in default is seems to handle the Qt event loop. And for that matter,
> why is it possible to set the file.edit callback on Mac but not Linux?
> This seems arbitrary, and no explanation is given (that I could find).

The R_PRocessEvents callback may be settable on MacOS but I'm not sure
it's used -- at least a quick grep didn't reveal its use anywhere
outside the gnuwin32 code.

It would be good to unify the Mac and *nix mechanisms here since the
OS underpinings are now so similar, but it will have to get high
enough on someone's priority list to happen.

Adding the polled events callback is probably not unreasonable at
least as a temporary measure.  It might be useful to think about an
alternate interface that allows code to distinguish a call frm the
interrupt check, which might want to be very fast and not do any more
than necessary to check for an interrupt event, and calbacks in
blocking situations like socket reads.

>
>> We could add a
>> R_PolledEvents() call in the #else bit to support this, though the
>> cautions in the comment do need to be kept in mind.
>
> Yes, that's what I ended up trying, and it seems to work reasonably
> well (with one caveat, see below).
>
> My R_PolledEvents calls qApp->processEvents, and if I set the Esc key
> to call kill(getpid(), SIGINT) or onintr(), then I can interrupt some
> things, like Sys.sleep. But without the patch to make
> R_CheckUserInterrupt call R_PolledEvents (or runHandlers) many other
> things, like all plotting routines, cannot be interrupted. So, it
> would indeed be very helpful to have R_PolledEvents called on Linux,
> or allow the setting of the R_ProcessEvents callback.
>
> The problem I'm having with this solution is that whenever I interrupt
> a graphics command, R crashes. This is true for commands being
> evaluated by R_tryEval, but not those run from the REPL (for example,
> if I make the call inside a debug() environment, interrupting it
> causes no problems). As far as I can tell, this is only a problem with
> graphics; other commands can be interrupted even when run using
> R_tryEval().

That sounds like a longjmp being done to a place that doesn't exist --
maybe a threading issue in Qt.  See what gdb tells you about where the
crash is occurring. It might be different for onintr and kill.  You
might also try just setting the R_interrupts_pending flag from the
interrupt event handler rather than calling onintr (which probably
longjmp's) or kill (which may be doing something you don't want if
other threads with other signal handlers are involved).

Best,

luke

>
> Running R in a separate thread is a better solution to the GUI
> responsiveness issue (and maybe interrupts can be cleanly sent too).
> Unfortunately, Qt's threading paradigm doesn't like it when the
> process running in the secondary thread (R in this case) tries to run
> Qt widgets of its own (e.g. a Qt graphics device). There may be a
> general solution to this, but making use of R_CheckUserInterrupt() is
> the easier solution for now.
>
> -Deepayan
>
> [...]
>
>> > So my assumption 'worked via the R-Unix eventloop' was that a handler
>> > (probably a polling handler) had been wedged in the eventloop.
>> > That was in contrast to running under a separate thread.
>> >
>> >> The UI provided by a shell running in a separate process may not have
>> >> a 'G' but it does have its advantages :-)
>> >
>> > Or a separate thread, as Rterm.exe does.  Really RGui should also run in 
>> a
>> > separate thread, but when Guido did so, it did not work under Windows 95:
>> if
>> > we ever give up support for pre-NT Windows I will take a look again at
>> this.
>> >
>> > I guess my underlying point is that rather than run the GUI from
>> > R_ProcessEvents (as RGui is), on Unix you can run it from an eventloop
>> > handler.
>> >
>> > Brian
>> >
>> >>
>> >> Best,
>> >>
>> >> luke
>> >>
>> >>>> while since I looked at that. I don't believe the unix version of
>> >>>> R_CheckUserInterrupt() does not provide hooks for installing such
>> >>>> checking (we have talked about this off an on but I don't believe
>> >>>> anything happened -- could be wrong there though).
>> >>>>
>> >>>> If Qt allows this one option may be to have events on your nterrupt
>> >>>> widget managed by a small thread that does nothing other than send a
>> >>>> signal to the main thread if the widget is clicked.
>> >>>>
>> >>>> Best,
>> >>>>
>> >>>> luke
>

-- 
Luke Tierney
Chair, Statistics and Actuarial Science
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
      Actuarial Science
241 Schaeffer Hall                  email:      luke at stat.uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu