On 30 Apr 2019, at 23:09 , Paul Murrell <paul at stat.auckland.ac.nz> wrote:
Hi Peter
Yes, that looks roughly right to me. I would be in favour of your option (b), partly because it is probably easiest and partly because that retains the basic graphics device startup logic pattern that is replicated across all(?) graphics devices.
Paul
On 28/04/19 11:39 AM, peter dalgaard wrote:
I had a look at the current code, and AFAICT it has essentially the same structure as it did back then. I think it may have finally dawned upon me what the issue really is:
The logic is that in Rf_addX11Device, we have
if (!X11DeviceDriver(dev, display, width, height,
ps, gamma, colormodel, maxcubesize,
bgcolor, canvascolor, sfonts, res,
xpos, ypos, title, useCairo, antialias, family)) {
free(dev);
errorcall(call, _("unable to start device %s"), devname);
}
dd = GEcreateDevDesc(dev);
GEaddDevice2(dd, devname);
i.e., we start the device driver, and if it fails, we throw away the "dev" structure and call it a day. If it succeeds, we proceed to create a device descriptor structure and add it to the list of open devices.
This approach means that X11DeviceDriver() cannot do anything that potentially accesses the dd structure because it isn't there yet, and the things it cannot do apparently includes calling R_ProcessX11Events(). [To be completely sure that this is actually still true, I'd need to have a closer look at what handleEvent() does.]
So to fix things, it would seem that you could (a) add the device before attempting to start the driver, preparing to back it out if the driver fails to start, or (b) add a call to R_ProcessX11Events() _after_ the GEaddDevice2(dd, devname). Option (b) is probably the easiest.
Paul: Does this analysis look roughly right?
-pd
On 26 Apr 2019, at 01:23 , frederik at ofb.net wrote:
Thanks Professor Dalgard.
If you have a different way to fix the bug then I'd be happy to test
it.
Or whatever. I understand that maybe some data was being referenced
before it had been initialized. I could also support moving the
R_ProcessEvents call in another place, but it seems one would also
like to generate some kind of warning message, at the location of the
bad reference, rather than segfaulting. Was it not possible to
identify this location? I'm guessing that Valgrind is a bit more
mature now than it was in 2001...?
Frederick
On Wed, Apr 24, 2019 at 03:12:55PM +0200, peter dalgaard wrote:
OK, so I did the archaeology anyway....
This was the story, R-core November 29, 2001. Part of thread "X11 still segfaults".
------------>>
.....
Gah. I've been too tired today. Why did that take me so long?
The culprit seems to be
R_ProcessEvents((void*) NULL)
in newX11DeviceDriver
This gets called *before* this stuff at the end of Rf_addX11Device
dd = GEcreateDevDesc(dev);
addDevice((DevDesc*) dd);
initDisplayList((DevDesc*) dd);
and it is that "dd" that gets called by Rf_playDisplayList. Removing
the offending line stops the segfaulting, seemingly with no ill
effects.
I'm not really sure what the use of that line ever was; it might be
necessary to make the call somewhere later, but it appears to have
been possible to race past it before receiving any events all the
time.
I also changed a couple of spots missing dd->newDevStruct=1
Will commit in a moment.
<<------------
And the following day, in "graphics saga part III", we had
------------->>
...
I can't make it happen in 1.3.1 but...
It is probably not unrelated to the R_ProcessEvents line that
I took out, but that was definitely wrong. However, one might reenable
it if one could change this bit of code
if (!(ptr_X11DeviceDriver)((DevDesc*)(dev), display, width, height, ps, gamma,
colormodel, maxcubesize, canvascolor)) {
free(dev);
errorcall(gcall, "unable to start device %s", devname);
}
gsetVar(install(".Device"), mkString(devname), R_NilValue);
dd = GEcreateDevDesc(dev);
addDevice((DevDesc*) dd);
initDisplayList((DevDesc*) dd);
and put the if-clause last. A cursory clance through the three
functions that are being called didn't reveal anything that would rely
on having opened the device driver first.
Paul?
(I might try it locally, but I'm not sure I should commit anything.)
<<-----------
It seems that the suggestion was never followed up on?
-pd
On 24 Apr 2019, at 11:42 , peter dalgaard <pdalgd at gmail.com> wrote:
I don't recall exactly what I did 18 years ago eiher and I likely don't have the time to dig into the archives and reconstruct.
I can imagine that the issue had to do with the protocol around creating and mapping windows. Presumably the segfault comes from looking for events on a window that hasn't been created yet, or has already been destroyed, leading to a NULL reference somewhere. I have a vague recollection that the issue was window manager dependent (in 2001 probably not twm, more likely xvwm on RedHat if it was affecting me).
A proper fix should go via proper understanding of the X11 protocol - uncommenting a line is as bad as commenting it in the 1st place.... So more like "wait for window to exist THEN process events" -- but the 1st part may be WM specific, etc.
I recall docs being quite obtuse, and the X11 "mechanism not policy" credo doesn't help as WMs are not obliged to (say) send notifications, so you can end up stalling, waiting for events that never happen.
It is entirely possible that there is stuff in here that I didn't understand properly at the time, and still don't!
- pd
On 24 Apr 2019, at 02:30 , Paul Murrell <paul at stat.auckland.ac.nz> wrote:
Hi
Sorry, I can't offer an explanation for the commented-out line.
However, regarding your final question of avoiding the R-core bottleneck, you do have the option of creating a third-party graphics device package. See, for example, the 'tikzDevice' and 'svglite' packages on CRAN. Does that provide you with a way forward ?
Paul
On 20/04/2019 5:27 p.m., frederik at ofb.net wrote:
Dear R Devel,
I know that someone put this line in src/modules/X11/devX11.c:2824 for
a reason, because commenting it out causes R to miss an important
ConfigureNotify event in my window manager. The result is that plots
are initially drawn off the window borders, unreadable.
R_ProcessX11Events((void*) NULL);
Unfortunately for me, this line is commented in the standard release
of R, it has "#if BUG ... #endif" around it.
I guess it is also unfortunate for anyone who uses the same window
manager as I do, namely i3, which I think is pretty popular among Unix
power users these days; not to mention other full-screen window
managers which probably exhibit the same bug in R.
Maybe everyone on the Core team uses twm as their window manager? Or
RStudio on Windows? Which would be sad because then we're not
representing an important user demographic, namely those who prefer
software which is modern and powerful, yet simple to understand and
modify; fully configurable and interoperable and so on.
I first reported this bug 3 years ago. In doing research for my bug
report, I found that the line was commented out by Peter Dalgaard in
2001 with the explanation "X11 segfault fix - I hope".
I don't know what the way forward is. Obviously the Core Team has
reason to say, "look, this isn't very important, it's been broken
since 2001, maybe fixing it will cause the undocumented segfault bug
to reappear, clearly no one here uses your window manager". Do I have
to submit a correctness proof for the proposed change? What do I do?
https://bugs.r-project.org/bugzilla/show_bug.cgi?id=16702
As mentioned in my bug report, I checked using gdb that
ConfigureNotify is indeed being received by the call to
R_ProcessX11Events() when it is uncommented. I haven't experienced any
segfaults.
It's good that Peter left evidence that "R_ProcessX11Events" was being
called 18 years ago from X11DeviceDriver(). If he had deleted the
line, rather than commenting it, then discovering the reason for the
window rendering bug would have been much harder for me.
However, the downside is that now it is not just a matter of inserting
the line where it belongs; I also feel a bit like I have to explain
why it was initially removed. But although I've given it some thought,
I still have no idea.
Somewhat tangentially, I am wondering if there is some way that we
could make the development of R's graphics code proceed at a faster
rate, for example by pulling it out into a separate module, so that
people could offer alternative implementations via CRAN etc., rather
than having R Core be the bottleneck. Would this make sense? Has it
already been done?
Thank you,
Frederick