Bug 20567

[  646.646286] Xorg          D 00000064     0  1207   1190
[  646.646286]  f5ec9ea0 00003046 b1130ad6 00000064 00000000 f68c4850 f652eae0 c09ffb24
[  646.646286]  c0a02a80 c0a02a80 c0a02a80 c0a02a80 f652eae0 f652ed34 c1cf7a80 00000000
[  646.646286]  00000000 c046928b f6462e2c f6462dd0 c1cf7a90 00003046 00000000 00000000
[  646.646286] Call Trace:
[  646.646286]  [<c046928b>] ? might_fault+0x43/0x80
[  646.646286]  [<c0579d33>] ? i915_gem_throttle_ioctl+0x17/0x3c
[  646.646286]  [<c0783c3c>] mutex_lock_nested+0x14f/0x249
[  646.646286]  [<c0579d33>] ? i915_gem_throttle_ioctl+0x17/0x3c
[  646.646286]  [<c0579d33>] i915_gem_throttle_ioctl+0x17/0x3c
[  646.646286]  [<c0566dd5>] drm_ioctl+0x1b7/0x22f
[  646.646286]  [<c0579d1c>] ? i915_gem_throttle_ioctl+0x0/0x3c
[  646.646286]  [<c0486711>] vfs_ioctl+0x50/0x69
[  646.646286]  [<c0486c4e>] do_vfs_ioctl+0x460/0x499
[  646.646286]  [<c0402c7c>] ? restore_sigcontext+0x100/0x143
[  646.646286]  [<c0486cb3>] sys_ioctl+0x2c/0x45
[  646.646286]  [<c0403152>] syscall_call+0x7/0xb


another backtrace, this was with the latest git driver+libdrm

Upstream code (with xf86-video-intel master branch) test result:
most clutter cases (e.g. test-layout) works fine.
some clutter case (test-multistage) gets X crash, i.e. bug#20286.

while the symptom you see seems X hang, but still there. 

Can you try test-layout?

did you use a system that has KMS enabled, and is running an opengl using compositing window manager?

(In reply to comment #3)
> did you use a system that has KMS enabled, and is running an opengl using
> compositing window manager?
> 

Yes, we are running on Fedora 10, with UXA and KMS enabled, and compiz.

Is your driver version the one moblin QA is testing (today or yesterday)? I haven't heard they met so critical issues.

(In reply to comment #2)
> Upstream code (with xf86-video-intel master branch) test result:
> most clutter cases (e.g. test-layout) works fine.
> some clutter case (test-multistage) gets X crash, i.e. bug#20286.
> 
> while the symptom you see seems X hang, but still there. 
> 
> Can you try test-layout?
> 

test-multistage will crash X in my test with KMS. test-layout works well.

disable memory tiling will avoid X crash, but get poor performance. 

On 915GM, KMS, compiz, all the clutter tests I tried worked fine.

Arjan, please attach the information I requested originally:
http://intellinuxgraphics.org/how_to_report_bug.html
(in particular, "Information to provide".)

>> Starting a clutter based application causes the X server to stop responding
Arjan, do you mean "any" of clutter case can trigger this? If not, please specify one. How about test-layout, as I've asked in comment#2?

OK, let's call this bug about test-multistage, on the assumption that that's testing what really matters.  What we know:

- Only happens with DRI2.
- Disabling tiling fixes it.
- KMS is not a requirement (I reproduce it without KMS).
- Compositing is not a requirement (I reproduce it with bare server + xlogo running)
- Chris Wilson's tiling fixes don't fix it. (got those applied)
- It does appear to be a normal GPU hang (not kernel oops/lockup/app lockup).

Fix posted to the mailing list and in my tree on the 945-fence-fix branch until someone else can review it, given that I'm exhausted here.

commit 7637e349061b5220e47b1df94174989f0d195075
Author: Eric Anholt <eric@anholt.net>
Date:   Tue Mar 10 22:34:49 2009 -0700

    drm/i915: fix 945 fence register writes for fence 8 and above.
    
    The last 8 fence registers sit at a different offset, so when we went to set
    fence number 8 in the lower offset, we instead set PGETBL_CTL, and the GPU
    got all sorts of angry at us.
    
    fd.o bug #20567.  Easily reproducible by running glxgears and killing it abo
    6 times.
    
    Signed-off-by: Eric Anholt <eric@anholt.net>

In latest moblin trunk
xorg-x11-server-1.6.0-1.2.moblin2.src.rpm
xorg-x11-drv-intel-2.6.3-3.1.moblin2.src.rpm
mesa-7.2-6.2.moblin2.src.rpm
libdrm-2.4.5-1.2..moblin2.src.rpm
kernel-2.6.29.rc7-40.1.moblin2.src.rpm

besides test-multistage, following clutter demos crashed too:
test-clutter-cairo-flowers
test-cogl-primitives
test-cogl-tex-foreign
test-multistage
test-perspective
test-scale
test-threads

BTW, when I replaced gfx relateed components of moblin with upstream 
(Xserver: 1.6.99.1, xf86-video-intel: 2.6.99, mesa: 7.5-dev, drm: 2.4.5)
I found that: 1st any of clutter-demos could cause X crash
2nd, the icons of desktop were lost

(In reply to comment #9)
> Fix posted to the mailing list and in my tree on the 945-fence-fix branch until
> someone else can review it, given that I'm exhausted here.
> 
> commit 7637e349061b5220e47b1df94174989f0d195075
> Author: Eric Anholt <eric@anholt.net>
> Date:   Tue Mar 10 22:34:49 2009 -0700
> 
>     drm/i915: fix 945 fence register writes for fence 8 and above.
> 
>     The last 8 fence registers sit at a different offset, so when we went to
> set
>     fence number 8 in the lower offset, we instead set PGETBL_CTL, and the GPU
>     got all sorts of angry at us.
> 
>     fd.o bug #20567.  Easily reproducible by running glxgears and killing it
> abo
>     6 times.
> 
>     Signed-off-by: Eric Anholt <eric@anholt.net>
> 

Thanks Eric, I have test your patch and it fixed this issue.

Thanks for the confirmation!  It's sent off to linus now.