Bug 10365

Created attachment 9250 [details]
xorg log file

My guess would be the amount of aperture space reserved for the DRM memory manager is too small. Or maybe Thomas has an idea.

agpgart says the aperture is "256M @ 0xb0000000" (see below), and the xorg log file says it's only asking for 32M.

------------------------------------------
(II) intel(0): Will try to reserve 32768 kiB of AGP aperture space
        for the DRM memory manager.
------------------------------------------

More details in relation to agp stuff:
-------------------------------------------
$ dmesg | grep -i agp
Linux agpgart interface v0.102 (c) Dave Jones
agpgart: Detected an Intel 915GM Chipset.
agpgart: Detected 7932K stolen memory.
agpgart: AGP aperture is 256M @ 0xb0000000
 [<f8858682>] intel_i830_free_by_type+0x42/0x70 [intel_agp]
 [<c024b352>] agp_free_memory+0x72/0x90
 [<c0249dc5>] agpioc_deallocate_wrap+0x15/0x20
 [<c024a0f0>] agp_ioctl+0x230/0x410
 [<f8858682>] intel_i830_free_by_type+0x42/0x70 [intel_agp]
 [<c024b352>] agp_free_memory+0x72/0x90
 [<c0249dc5>] agpioc_deallocate_wrap+0x15/0x20
 [<c024a0f0>] agp_ioctl+0x230/0x410
 [<f8858682>] intel_i830_free_by_type+0x42/0x70 [intel_agp]
 [<c024b352>] agp_free_memory+0x72/0x90
 [<c0249dc5>] agpioc_deallocate_wrap+0x15/0x20
 [<c024a0f0>] agp_ioctl+0x230/0x410
agpgart: pg_start == 0x00000038,intel_i830_private.gtt_entries == 0x000007bf
agpgart: Trying to insert into local/stolen memory
[drm:drm_agp_bind_ttm] *ERROR* AGP Bind memory failed
-------------------------------------------------------------

I'll also attach my xorg.conf.

Created attachment 9251 [details]
xorg config

Debian is probably not using a compatible AGP driver and probably an old DRM.

Please upgrade to currend drm git and either a 2.6.21 kernel or fetch an agpgart driver from 
git://git.freedesktop.org/git/mesa/linux-agp-compat.

Quoting from my version note above:

-----------------------------------------------------------
Kernel: 2.6.21-rc4
drm: current from git://anongit.freedesktop.org/git/mesa/drm
agp: current from git://git.freedesktop.org/git/mesa/linux-agp-compat
-----------------------------------------------------------

(In reply to comment #6)
> Quoting from my version note above:
> 
> -----------------------------------------------------------
> Kernel: 2.6.21-rc4
> drm: current from git://anongit.freedesktop.org/git/mesa/drm
> agp: current from git://git.freedesktop.org/git/mesa/linux-agp-compat
> -----------------------------------------------------------
> 

Sorry,
It's me not reading properly at this late hour.

This is the root of the problem:
(II) intel(0): 0x00000000-0x0001ffff: ring buffer (128 kB)
(II) intel(0): 0x00020000-0x00027fff: logical 3D context (32 kB)
(II) intel(0): 0x00028000-0x00037fff: xaa scratch (64 kB)
(II) intel(0): 0x00038000-0x00237fff: DRI memory manager (2048 kB)

The memory manager is placed in the stolen memory area but it needs a fixed 
area of unused aperture space, and cannot be moved. Ever. That's why it was originally placed at the very end of the aperture.
 
Need to figure out why AGP crashes, though.

/Thomas

 

> Need to figure out why AGP crashes, though.
> 
> /Thomas
> 

Hmm,
Seems like the intel xorg driver is trying to free an agp area it never allocated, and agpgart is not doing proper input checking.

This is all probably due to recent changes in aperture space allocation strategies in xf86-video-intel.

/Thomas

Added myself to the CC list since I'm having the exact same problem and interested in a resolution.

I think this is a regression that should be looked into before release.

(In reply to comment #10)
> I think this is a regression that should be looked into before release.
> 

I just wanted to say that I've tried this with the intel driver from git head + xserver 1.3RC3, as well as mesa-git. It's still a no-go. I'm on a 855GM. I really hope this gets fixed before release...

If I can help in anyway, please let me know.

Using the dri drivers from mesa-git solves this problem, although 3D-performance is noticeably worse. Compiz, e.g. is no longer usable (under AIGLX). It feels like some of the actions are done in software mode (high CPU usage when scrolling, etc.)

Closing as comment#12 says mesa-git fixes the reported issue.