[Intel-gfx] Intel 945GM with >=2.6 and newer kernels: massive performance and or memory leak problems

[Thomas Bächler]

Thomas Bächler schrieb:
> 1) With KMS enabled, 2.6.29-rc1 was already bad and everything before 
> that seemed broken.
> 
> 2) I bisected with the following configuration:
> - xf86-video-intel 2.6.99.902 (with some extra bugfixes applied, so it 
> is something from post-2.6.99.902, but not the tip of the 2.7 branch, if 
> you must know I can look up exactly what was in there).
> - KMS was disabled in all tests
> - I had UXA enabled in all tests
> - For hardware see my first post in this thread, if you need more info 
> please ask.
> 
> This is from git-bisect:
> 
> 0f973f27888e4664b253ab2cf69c67c2eb80ab1b is first bad commit
> commit 0f973f27888e4664b253ab2cf69c67c2eb80ab1b
> Author: Jesse Barnes <jbarnes at virtuousgeek.org>
> Date:   Mon Jan 26 17:10:45 2009 -0800
> 
>     drm/i915: add fence register management to execbuf
> 
>     Adds code to set up fence registers at execbuf time on pre-965 chips as
>     necessary.  Also fixes up a few bugs in the pre-965 tile register 
> support
>     (get_order != ffs).  The number of fences available to the kernel 
> defaults
>     to the hw limit minus 3 (for legacy X front/back/depth), but a new 
> parameter
>     allows userspace to override that as needed.
> 
>     Signed-off-by: Jesse Barnes <jbarnes at virtuousgeek.org>
>     Signed-off-by: Eric Anholt <eric at anholt.net>
>     Signed-off-by: Dave Airlie <airlied at linux.ie>
> 
> :040000 040000 22098cda6a4a56bbcb2420344640881a49f23dcf 
> 70a93a08be654f198c6d8907a6454ff6f0d914d0 M      drivers
> :040000 040000 40d7c115066ce0aed35e97ac9149e027ce076ee0 
> ef6a8fbdc0da2b7b3798a1f939aed3749eda5133 M      include
> 

Just for reference/archiving purposes, this is my status now:

I am using

- drm-intel-next (280b713b5b0fd84cf2469098aee88acbb5de859c)
- The PAT patches from 
http://permalink.gmane.org/gmane.linux.kernel/816080 and 
http://permalink.gmane.org/gmane.linux.kernel/819914
(to avoid flooding the log with "freeing invald memtype")
- xf86-video-intel master branch (9b615a52671aacf34666f90ecfff98651ce6afe2)
- The "Use all fence regs" fix from #20761

With this combination, compositing performance is excellent again 
(feeling better than with 2.6.28 + UXA). I only tested with KMS. 3D 
performance is still nonexistent (tested with google earth), but that is 
not a regression from 2.6.28.


About memory leaks: After a few hours of usage, memory usage is still 
"sane" (other than with the 2.6.3 driver where it would leak memory 
constantly), but I have to investigate that further. However, when I 
come back from hibernation, about 1GB of swap is being used when before 
hibernation less than 500MB of memory was in use (and swap was empty). I 
will investigate that more thoroughly and open a separate thread about it.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 260 bytes
Desc: OpenPGP digital signature
URL: <http://lists.freedesktop.org/archives/intel-gfx/attachments/20090415/7fd3ede2/attachment.sig>