Intel driver (using EXA) crashes system when starting compiz

Binary package hint: xserver-xorg-core

I have detailed the problem extensively upstream, so I'll post a summary here:

I have discovered that the patch "120_fedora_disable_offscreen_pixmaps.diff" causes a system freeze when using a combination of the Intel driver with EXA enabled (XAA is fine) and compiz. It has been reproduced on two of my systems using 855GM and 865G respectively).

I recompiled xserver-xorg-core with the aforementioned patch disabled, and all is well. I would appreciate if the Ubuntu Xorg maintainers could revise this patch or drop it entirely. Everything is detailed upstream, which I will tag here.

This is xserver-xorg-core patched to exclude the offending patch; if anyone wants to test this package against the official one they are free to do so.

Please see the discussion linked upstream; Michel Danzer has confirmed the aforementioned patch to be the cause of the crashes.

Confirmed, this does fix it for me (with an ATI radeon 9000), however it seems like EXA is incredibly sluggish for me, especially with epiphany with compiz. Good job finding the offending patch.

I'm going to hold off on dropping this patch until I have a better understanding of why it was added in the first place, since I would assume it was added because some people experience issues without it. Timo remarked that he had attempted to drop a number of patches including this one for the Feisty release, but ended up restoring them due to issues found. So I assume the issue this fixes is still present. However, perhaps if further research is done we can explore dropping it for Gutsy+1.

Also, it sounds as if xserver 1.4 is right around the corner, so it is also possible that contains a better fix. After that has been incorporated into Ubuntu it would be worthwhile to re-test this bug.

Bryce,

I understand your position and mostly agree. I believe the offending patch is a workaround allowing compiz to work correctly without 'Option "XaaNoOffscreenPixmaps' being present in your xorg.conf. User of ATI and Intel cards have reported crashes (in the case of ATI) and freezes (for Intel) with EXA, and using the patched X server attached above solves the problem.

Ideally this patch should be "revised" in the sense that whatever it does, it only does it when XAA is active; I would assume this will resolve this bug without sacrificing functionality for either acceleration architecture.

Another note: if Xorg7.3 (and hopefully the latest DRM modules to enable the new Mesa TTM memory manager), EXA will be substantially faster than before due to Michel Danzer's aiglx-zero-copy-tfp patches, but XAA won't benefit. So there's a good reason to revise this patch, aside from increasing system stability.

I vote for the removal of this patch, as XAA is broken for me - see bug #127101

Thomas,

Try adding 'Option "MigrationHeuristic" "greedy' to 'Section "Device" of your xorg.conf, it may improve performance with EXA - however, some operations may still be slower than XAA. There are further optimizations in the works for EXA, but the real strength of the architecture will lie in its compositing abilities and speed in the near future.

Bryce,

Just in case you didn't look at the upstream bug, I marked it there as "NOTOURBUG" myself, which Launchpad "translated" as Invalid - don't let that put you off. All the relevant logs for my system are posted there, and for your reference, Michel Danzer brought the patch to my attention in an off-topic bug, here: http://bugs.freedesktop.org/show_bug.cgi?id=11432#c15

I reported bug #127101, which was marked as a duplicate of this bug.
I'm trying to reproduce my bug in an up-to-date fresh gutsy tribe 4 install, but I haven't been able to get the "evicting pixmaps" message. In my previous fresh tribe 3 install I just needed to enable compiz and switch VT, but now I can't see the message and everything has been working fine. Is there a way to force the execution of the code that shows that message?

Note: I haven't modified my xorg.conf, so I'm using XAA.

Conn,

In upstream bug <https://bugs.freedesktop.org/show_bug.cgi?id=11626> you said:

"Unfortunately, restarting X/switching to VTs while Compiz is active causes my system to hang (but this happens also with XAA, so it's a matter for another bug)"

That's what I reported at bug #127101 (using XAA), so it's probably not a duplicate of this bug. Please check. Thanks.

Juan,

I was the one that marked your bug as a duplicate, so I unmarked it again for your convenience. If you're not experiencing this issue anymore, perhaps it's fine, but Michel Danzer called the aforementioned patch "buggy", and if you read the patch it is actually what generates the "(II) XAA: Evicting pixmaps" message.

You may want to keep an eye on this bug as well as yours, as I do suspect the issue is still related.

I tried the attached package and it does indeed work. Seems very important to fix this issue properly, as I've heard claims that gutsy wants to enable 3D effects by default. That's not going to happen if movies can't play, which won't happen until EXA is smoothed out, which won't happen until showstopper bugs like this one are fixed. Frankly there doesn't seem to be enough time. Maybe next release.

sorry. wrong click.

Why next release? Until now there seems to be no negative efects, maybe this fix just needs more testing to make sure there are no secondary effects.

What I meant is that EXA isn't tested by many. Those who have report it is slower, and there are some remaining bugs. Consider that newer intel cards haven't been tested at all because of bugs like these. Anyways, I suppose this is another thread of discussion. Sorry.

liminal,

I believe this should be fixed as soon as possible purely due to the severity of the bug. If I'm using the EXA acceleration architecture and try to run compiz, the machine hard-locks, and I have to force a power-down, potentially damaging the hard-drive and hardware (happens on my laptop and desktop system with Intel graphics). Frankly it seems crazy to keep a patch that causes such a regression.

As for performance, EXA is overall slower with some operations, but faster with others, especially RENDER acceleration. When a certain class of patches is included in mesa/xserver-xorg-core/xserver-xorg-video-* (called aiglx-zero-copy-tfp), compiz & EXA will be much smoother than XAA & compiz. However, these patches may not make it to gutsy in time.

The offending patch is to prevent the need for "XaaNoOffscreenPixmaps" in your xorg.conf with XAA, so it's a necessary patch - but it needs to be *debugged* and fixed so that it doesn't conflict with EXA.

using the package attached to this bug, i can now use radeon+exa+compiz. previously i was also having compiz crash X when using EXA.

Marking Confirmed due to user reports and confirmation upstream

Hi,
since I upgraded to Gutsy yesterday, my system now hangs up (apparentely random) with the error:

(II) XAA: Evicting pixmaps

in the /var/log/Xorg.0.log
I have an Intel i810 ( Intel Corporation Mobile 915GM/GMS/910GML Express)

the xserver-xorg-core version is 2:1.3.0.0.dfsg-6ubuntu3

I've just updated to xserver-xorg-core_1.3.0.0.dfsg-12ubuntu1, and still had to remove the patch 120_fedora_disable_offscreen_pixmaps.diff to be able to use compiz with exa acceleration. Here is the patched package, in case anyone else wants it.

With the update to xorg-server_1.3.0.0.dfsg-6ubuntu3, EXA now works perfectly. No more need to rebuild xorg-server without patch 120.

Unfortunately, the problem is NOT fixed on my system. I'm attaching logs* of my system after invoking "compiz --replace". The screen blanks, shows corruption and then restarts. Sometimes my system will freeze (same as when I first reported this bug) - usually when I try to run compiz twice in succession.

Relevant excerpt from Xorg.0.log:

Backtrace:
0: /usr/bin/X(xf86SigHandler+0x81) [0x80c9521]
1: [0xffffe420]

Fatal server error:
Caught signal 11. Server aborting

(II) AIGLX: Suspending AIGLX clients for VT switch
(II) intel(0): xf86UnbindGARTMemory: unbind key 0
(II) intel(0): xf86UnbindGARTMemory: unbind key 1
(II) intel(0): xf86UnbindGARTMemory: unbind key 2
(II) intel(0): xf86UnbindGARTMemory: unbind key 3
(II) intel(0): xf86UnbindGARTMemory: unbind key 4
(II) intel(0): xf86UnbindGARTMemory: unbind key 5
(II) intel(0): xf86UnbindGARTMemory: unbind key 6
(II) intel(0): xf86UnbindGARTMemory: unbind key 7
(II) intel(0): xf86UnbindGARTMemory: unbind key 8
(II) intel(0): xf86UnbindGARTMemory: unbind key 9

* Attaching in .tar.bz2 archive to avoid excessive attachments.

Bryce,

I tried your "testing" 2:1.3.0.0.dfsg-12ubuntu4 release and the problem remains. The backtrace shows a different address, but that may not be relevant. Here it is:

Backtrace:
0: /usr/bin/X(xf86SigHandler+0x81) [0x80c76f1]
1: [0xffffe420]

I can confirm that the patched xserver-xorg-core package to remove offending patch, fixes the crash.

I have to take back comment #20 - I hadn't noticed that I didn't update xorg-server because apt thought my version was the higher one. I still need to rebuild xorg-server without patch 120 to be able to use EXA and compiz on my 945GM chipset laptop.

I can confirm both that:
 - this bug affects the HP6710b (running i965 chipset),
 - the patched xserver-xorg-core.deb supplied by conn resolves the issue

thanks!
Dave Smylie

Unfortunately, that patch is needed in order to have a usable compiz/composite experience with XAA, and besides not that many drivers support zero-copy-tfp with EXA yet. This should be re-evaluated when hardy opens. Mesa has r3xx (ati) support for z-c-tfp, and I've tested r200. Don't know what the situation is with intel, though.

I beg to differ, that patch is not needed - it is only a alternative to adding 'option "XAANoOffScreenPixmaps"' to xorg.conf. So, to avoid entering a line in xorg.conf if you still use XAA (soon to be obsolete), everyone using EXA (which is the future direction of X acceleration) has to rebuild his own xserver-xorg. Seems like a big difference in scale.

I am glad to try this patch but I am using Gutsy AMD64. It makes me cannot use the .i386.deb file. Can anyone please provide the .amd64.ded? Or give me a link guiding me to rebuild my xserver-xorg.

You need to get the source "apt-get source xorg-server", the build dependencies "sudo apt-get build-dep xorg-server", then cd into xorg-server-1.3.0.0.dfsg, edit debian/patches/series and comment or remove the line that starts with "120", and rebuild the deb - "fakeroot debian/rules binary". When it finishes, you'll have a amd64.deb package in the dir above.

Many thanks to Jose !!!

I have a firm confirmation that removing the aforementioned patch and rebuilding the xorg-server make me have EXA running smoothly.

Besides, I found that the video played with XV plugin enabled really bad, picture is flickering too much, CPU is used up to 50%. Without XV, video is smooth enough, CPU is used about 20%. Does it make sense? Does EXA not support XV?

PS. I use Compaq laptop with Core 2 Duo 7300, Intel X3100 graphics card.

Thank you, Jose for the excellent build instructions!

I also can confirm that removing the patch fixes the xserver crash when running KDE4's kwin with EXA on an i965.

Confirmed on Dell XPS M1330 with Intel I965 graphics card (blacklisted by compiz as PCI ID 8086:2a02).

Conn, many thanks for the excellent work with the work-around.

When enabling EXA and compiz I get some flaws on window borders that may be a problem with shadows but I guess that is not directly related to this thread.

PS. A hint for others that want to try out the fix: On Ubuntu Gutsy 7.10 you can edit /usr/bin/compiz and remove the PCI ID of your graphics card from the compiz blacklist. (When you run compiz from the command line you it will tell you the PCI ID of your card.)

Well, I was happy for a while... Unfortunately, Firefox (and probably other applications as well) crashes with the work-around applied. Looking forward to xserver 1.4 in Ubuntu...

Hi,

From hardy's xorg-server 2:1.4.1~git20071119-1ubuntu1 upload:

   * debian/patches/120_fedora_xserver-xaa-evict-pixmaps.patch:
     - New version of the hack to copy textures from video memory. Shouldn't
       break EXA anymore.

We have two choices now:
1. Ensure exa functions in hardy, and mark this bug as wontfix for gutsy, or;
2. Backport this patch to gutsy's xserver-xorg-core (arguably it can be called a stability fix).

Thoughts anyone?

Conn: doesn't crash on hardy (tested on i945), and I doubt it can be backported fully, since it would mean backporting the EXA changes too.

Timo,

Thanks. I didn't take a look at the new patch yet, so I'll take your word for it. Shall we mark this bug as closed/fixed?

Conn: Yes, EXA is just becoming really usable and I doubt there will be backports for gutsy since it might break other setups. So I'm marking this as fixed in Hardy.

If you have a chance to try Hardy and it still hangs, please reopen and we'll consider other options!

Hi! I had the "XAA: Evicting pixmaps" problem in Gutsy using the xserver-xorg-video-intel driver with i945 chipset. I upgraded to Hardy and now I get these two last lines in my Xorg.0.log file:

(WW) intel(0): fbc already enabled on plane a, not enabling on plane a
(II) intel(0): fbc disabled on plane a

Then it hangs so I have to do a hard reboot, same as in Gutsy