Suspend/Resumes fails early with r600(RV710) in "unpin not necessary"

[Manuel Krause]

Hi! & Thank you very much for your recent answer!

On 2012-03-20 10:06, Michel Dänzer wrote:
> On Die, 2012-03-20 at 02:29 +0100, Manuel Krause wrote:
>> On 2012-03-20 00:12, Manuel Krause wrote:
>>>>> Mar 14 01:50:51 COOLHEAD kernel: agpgart-via 0000:00:00.0:
>>>>> Refused to change power state, currently in D0
>>>>
>>>> Does radeon.agpmode=-1 help at all?
>>>
>>> As this may reduce agp rate I don't like to try it. ;-)
>>> I'll come back on that later.
>>
>> I set that on kernel command line at first and removed "options
>> radeon agpmode=4" from my /etc/modprobe.d/99-local.conf.
>>
>> That went well for the first resume cycle. The second resume just
>> shortly after the first failed without any possibility to record
>> anything like logs (VT switching impossible, hard lock).
>>
>> Then I went to /etc/X11/xorg.conf setting the # in front of:
>> #	Option		"BusType"		"AGP"
>> #	Option		"AGPMode"		"4"
>>
>> That now worked for 4 subsequent hibernate/resume cycles [...]
>
> These options have no effect with KMS (you can verify that in the
> Xorg.0.log file), the failure after the second resume was probably a
> different problem or just a fluke.

Yes, you're right in both. The fluke only comes back once when 
re-switchting from AGP to PCIE or vice versa.

>> I can't see any performance impact.
>>
>> Any clues?
>
> Looks like it boils down to 'AGP sucks'. Moreover, VIA AGP bridges were
> kind of infamous for instability. As PCIe mode seems to work well for
> you, it's probably best to leave it at that.

I know of the infamous instability at least. And out of personal 
experience. I previously had the non-Tualatin-capable board 
version (P3V4X vs. TUV4X now) before with a Nvidia gfx-card. But 
even that combo was able to run AGP 4x with SideBandAddressing on 
with FastWrites off (run with the proprietary drivers in those days).
The TUV4X is much more matured, obviously, as it was able to 
operate the nvidia @ AGP 4x +SBA +FW without any crashes (what 
shouldn't sound like an ad for a competitor now, but as an 
encouragement for your work!).

So, I'm left with some questions now:
- Why does hibernation only work with PCIE forced mode?
- If I don't see performance regressions in PCIE forced mode,
    is the AGP mode coded to max. performance at all?
    (As I thought until now, AGP offered higher transfer rates
    than 33MHz PCI ?!)
- Are there "hidden brakes" inside the kernel, that disallow
    AGP 4x, FW and SBA in combination of my chipset with
    this ATI HD4350 card from HIS ? Can I unlock them manually?
- Are there "tunables" of that PCIE forced mode?

I would be glad, if you answered these questions,

best regards,

Manuel
[now running self-compiled kernel 3.3.0 from openSUSE sources]