Bug 30420 - Xserver freeze with all OpenGL applications on SandyBridge
Summary: Xserver freeze with all OpenGL applications on SandyBridge
Alias: None
Product: Mesa
Classification: Unclassified
Component: Drivers/DRI/i965 (show other bugs)
Version: git
Hardware: x86 (IA32) Linux (All)
: medium blocker
Assignee: Wang Zhenyu
QA Contact:
Depends on:
Reported: 2010-09-28 06:47 UTC by Matthias Hopf
Modified: 2010-09-30 02:35 UTC (History)
2 users (show)

See Also:
i915 platform:
i915 features:

Backtrace of frozen Xserver (15.62 KB, text/plain)
2010-09-28 06:47 UTC, Matthias Hopf
Backtrace of glxgears (8.51 KB, text/plain)
2010-09-28 06:48 UTC, Matthias Hopf
gzipped output of intel_gpu_dump (97.88 KB, application/octet-stream)
2010-09-28 09:40 UTC, Matthias Hopf

Description Matthias Hopf 2010-09-28 06:47:02 UTC
Created attachment 39009 [details]
Backtrace of frozen Xserver

On a SandyBridge [8086:0106] (internal reference at SuSE: NUE837) running

- kernel       2.6.36-rc5
- xorg server  1.9.0
- intel driver git 55b5fe8
- libdrm       git 7ec9a1e
- Mesa         7.9-RC1

glxinfo shows me

  direct rendering: Yes
  OpenGL renderer string: Mesa DRI Unknown Intel Chipset GEM 20100330 DEVELOPMENT x86/MMX/SSE2

/var/log/Xorg.0.log shows that DRI2 is initialized successfully.

glxgears just shows a black window, and X is frozen (ssh does work, though). I'm seeing exactly ZERO messages in /var/log/messages or /var/log/Xorg.0.log, there's nothing logged there.

X seems to hang in drm_intel_gem_bo_map_gtt(), and glxgears in a poll (probably waiting for the Xserver). Attaching backtraces for both the Xserver and glxgears.
Comment 1 Matthias Hopf 2010-09-28 06:48:12 UTC
Created attachment 39010 [details]
Backtrace of glxgears

Missing the symbols of glxgears itself, though.
Comment 2 Chris Wilson 2010-09-28 09:16:13 UTC
cat /sys/kernel/dri/0/i915/i915_gem_interrupts would presumably show X waiting on the rendering to the buffer to be complete before it could transfer it to the GTT domain. intel_gpu_dump would be useful to confirm if the ring was still processing or was idle.
Comment 3 Matthias Hopf 2010-09-28 09:39:39 UTC
I assume you mean /sys/kernel/debug/dri/0/i915_gem_interrupt ;-)

North Display Interrupt enable:         8c248080
North Display Interrupt identity:       00000000
North Display Interrupt mask:           73dbffff
South Display Interrupt enable:         00000f00
South Display Interrupt identity:       00000000
South Display Interrupt mask:           fffff0ff
Graphics Interrupt enable:              00000010
Graphics Interrupt identity:            00000000
Graphics Interrupt mask:                ffffffef
Interrupts received: 914
Current sequence:    0
Waiter sequence:     0
IRQ sequence:        0

Sorry, but I don't speak i965 fluent enough to make anything out of this...

Will attach the output of intel_gpu_dump.
Comment 4 Matthias Hopf 2010-09-28 09:40:30 UTC
Created attachment 39021 [details]
gzipped output of intel_gpu_dump
Comment 5 Wang Zhenyu 2010-09-28 17:31:18 UTC
Could you test with current mesa master? We've not been able to make sandybridge into 7.9 release.
Comment 6 Matthias Hopf 2010-09-29 09:00:21 UTC
Will do.

But note that the driver drop we received from Intel *specifically* for Sandybridge support is based on the 7.9 branch...
Comment 7 Wang Zhenyu 2010-09-29 17:54:01 UTC
yeah, we've been late for 7.9, and also for master, sorry for that.

I'm not sure if Ian would pick up more sandybridge fixes for 7.9, but we'll have a release note for sandybridge support in Q3 release if it requires anything not in stable releases.
Comment 8 Gordon Jin 2010-09-29 18:40:58 UTC
This should have been fixed in 7.9-rc2.

Ian cherry-picked >20 zhenyu's snb patches from mesa master to 7.9 branch about 24 hours ago then made it as 7.9-rc2.

So 7.9-rc2 has all sandybridge code on master except this one (which was committed to master after that time, and shouldn't matter much):

Commit: 73dab75b4165f7d2214a68d4ba8e3cb7aab9b4ac
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=73dab75b4165f7d2214a68d4ba8e3cb7aab9b4ac

Author: Zhenyu Wang <zhenyuw@linux.intel.com>
Date:   Sun Sep 26 13:15:39 2010 +0800

i965: fallback lineloop on sandybridge for now

Until we fixed GS hang issue.

With mesa 7.9-rc2, glxgears runs fine on my side, and running all mesa demo shows only 2 fbo cases hang (bug#30445).
Comment 9 Matthias Hopf 2010-09-30 02:35:34 UTC
Yes, 7.9-RC2 works.

It's still pretty slow (glxgears fullscreen 30fps), presumably because the back-to-front buffer copy is still not accelerated.

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.