See also https://bugzilla.redhat.com/show_bug.cgi?id=658587

Created attachment 40973 [details] [review]
Possible fix

Haven't tested this thoroughly, as I'm not convinced early-out is really the right way to handle Xinerama here, but I think this should fix the crashes at least.  Feel free to try it out.

Hi,

It's not pretty, but James' patch fixes the crash for me.

Thanks,
Colin Harrison

*** Bug 31675 has been marked as a duplicate of this bug. ***

James, have you sent this to xorg-devel for review / comment?  I didn't see it in a quick check of emails from you in December.

*** Bug 36060 has been marked as a duplicate of this bug. ***

*** Bug 34007 has been marked as a duplicate of this bug. ***

Here's another ping for this bad crasher. :)

I've sent it to xorg-devel.  Please review and comment in that thread

Discussion regarding why my patch ("Possible fix"), and a proposed patch from Adam Jackson, aren't functionally correct from a long time ago:

< jamesjones> ajax: Did that patch I attached to the font loading crash bug look at all correct?
< ajax> oh man, i seriously just finished writing a patch for that
< ajax> if you beat me to it i'm going to feel very lame
< jamesjones> sorry
< jamesjones> I'm not at all sure it's correct.
< ajax> oh wow, i see what you did.
< jamesjones> It seems like the whole idea of bailing out if you need to sleep ruins the point of Xinerama looping
< ajax> yeah, i don't think that's right
< ajax> one second...
< jamesjones> but I was too lazy to dig deeper since that patch was enough to unblock my testing.
< ajax> jamesjones: give this one a try? http://fpaste.org/BqJf/raw/
< ajax> i'm a bit ill at how that whole thing works, honestly.
< jamesjones> Give me a second, rebuilding
< jamesjones> But yeah, it doesn't make sense to me.
< ajax> at any rate it fixes my testcase of xset fp+ unix/:7100 && xlsfonts
< jturney> heh, it removes code, so it must be right :)
< jamesjones> Hehe
< ajax> nothing like having a library where your app has to implement functions with particular names and signatures to work
< jamesjones> but why the looping, if it's OK to return Success when the operation isn't done on screen[n] where n>0?
< ajax> jamesjones: it's not quite that.
< ajax> most of the font ops don't have any screen state, so you only have to call them once.
< ajax> PolyText and ImageTexto sleep do have to hit every screen, so you want to make sure that they only bump the client's sleep count once if they d
< ajax> so the minimal thing would have been to only change the closures for {Poly,Image}Text to not keep a sleep count
< ajax> but if you're doing that you may as well do them all
< jamesjones> right, I figured it was something like that, but then 1) Why loop the others 2) Why is it OK to early-out in PolyText/ImageText?  Does the looping get restarted somewhere I'm not seeing when the client wakes up?
< jamesjones> I suppose I should just load up xinerama and step through it
< ajax> the others aren't actually looped by xinerama
< ajax> (which i had missed the first time around, i just assumed they were)
< jamesjones> ah, k
< jamesjones> that makes more sense.
< ajax> the text rendering routines push themselves onto the work queue if they don't have a font they need because the font server hasn't gotten back to them
< ajax> check out the context around ClientSleep in doImageText, it's quite horrific
< ajax> the way they wake up is that the font server's fd selects readable, and the handler for that notices which request tat reply was for and pokes the client back into the work queue
< ajax> i think that's right anyway?  i kind of want to not know.
< jamesjones> it sounds right if there's only one screen
< ajax> well, if there's multiple screens, you might run through the sleep pattern multiple times
< jamesjones> Yeah, but it looks like each screen would loop in and block, only sleeping once
< jamesjones> then when it woke up, only the first screen would be serviced.
< jamesjones> *loop in and detect the need to sleep rather
< ajax> ew.  yeah.
< ajax> okay, so
< jamesjones> Also, it seems like with your patch, those other screens would leak their closure structs
< ajax> i think the way to fix that is to fix PanoramiX*Text* check for sleep too
< jamesjones> which is what I was attempting to avoid in mine.
< jamesjones> Yeah, that sounds like it has a better chance at correctness.
< jamesjones> Or I was thinking something like looping at a lower level, so there would be xinerama data in teh closure, but that sounded like it would violate all kinds of design stuff
< jamesjones> Your patch does avoid the crash I was seeing, FWIW
< ajax> progress
< ajax> thanks for testing.  i'll fix the xinerama thing and send the patches off to the list
< jamesjones> np, sounds good
< ajax> every once in a while i dream about working in a language where closures aren't something you have to hand roll every time
< jamesjones> If there weren't a bunch entrenched old languages I happen to know, I'd have to do something creative with my time.  Couldn't have that.
< ajax> aaaaaargh
< ajax> if you change PanoramiX*Text* to be restartable then there's a race between each screen where some other client on the same drawable or gc could modify your state
< ajax> so there goes the atomicity invariant

*** Bug 48900 has been marked as a duplicate of this bug. ***

Please note that those of us who live with Ubuntu's LTS world will be arriving at 12.04 over the coming months.  This bug will jump right out at you if you run an environment with many stations and non-standard collections of fonts.  It sure blew *me* right out of the water...

See also http://patchwork.freedesktop.org/patch/14507/

Created attachment 118203 [details] [review]
Allow font library to return Suspended more than once per request

fonts: Continue when font calls return Suspended more than once
    
    Patch 3ab6cd31cbdf8095b2948034fce5fb645422d8da fixed Xinerama
    interactions with font servers by not putting clients to sleep
    multiple times. However, it introduced additional changes dealing with
    libXfont routine returning Suspended more than once for the same
    request. This additional change was to abandon processing of the
    current request and free the closure data by jumping to
    'xinerama_sleep' in each of the functions.
    
    Font library functions shouldn't return Suspended more than once,
    except for ListFontsWithInfo, which produces multiple replies, and
    thus ends up returning Suspended many times during processing.
    
    With the jump to xinerama_sleep occurring after the first reply was
    processed, the closure for the request was freed and future calls into
    the ListFontsWithInfo callback resulted in dereferencing freed
    memory.
    
    This patch removes the added branches, reverting the code to its
    previous behaviour, which permitted multiple Suspended returns and
    simply waited for the client to be signaled again so that the callback
    could continue processing the request.
    
    Signed-off-by: Keith Packard <keithp@keithp.com>
    Cc: Adam Jackson <ajax@redhat.com>

Thanks for taking a look at this.

This appears to fix my testcase above, so Tested-by:

I don't know if this is likely to regress bug #3040, or how to test that.

That patch was committed:

commit f9a04d19aef77bf787b8d322305a6971d24a6ba1
Author: Keith Packard <keithp@keithp.com>
Date:   Mon Sep 21 07:16:12 2015 +0100

As such, closing this ticket.