Created attachment 24480 [details]
dmesg output of the carnage
commit 5b5441f8cc119db0d1e03dd35bd06015a26270dd "Increase default GART size to 32MB on >=R300" causes X to fail spectacularly, bringing down all sorts of running processes, such as sshd and hald.
All commits prior to this work. Tested with xserver-1.5.3.
Created attachment 24486 [details]
Xorg.0.log during crash
Looks like it doesn't show much of anything interesting.
So does it still work with Option "GARTSize" "8"? What about "16"?
When the kernel agpgart modules are loaded, what do they say about the aperture size? Assuming it's >= 32 MB: If I'm reading the dmesg correctly, the machine seems to have at least 4 GB of RAM? Does the problem occur if you boot with something like mem=1g or mem=2g?
(In reply to comment #2)
> So does it still work with Option "GARTSize" "8"? What about "16"?'
Yes, setting GARTSize to 8 allows X to start properly. With 16, it fails in the same fashion.
> When the kernel agpgart modules are loaded, what do they say about the aperture
> size? Assuming it's >= 32 MB: If I'm reading the dmesg correctly, the machine
> seems to have at least 4 GB of RAM? Does the problem occur if you boot with
> something like mem=1g or mem=2g?
I tried booting with mem=1g and GARTSize 16. It failed in the same way.
This motherboard is a UP1500, made by Samsung, and never had all the firmware bugs worked out. For instance, the PALcode initializes the AGP GART in such a way that it is unusable on Alpha. Hence, the card must be forced into PCI mode in xorg.conf to even start.
See http://bugzilla.kernel.org/show_bug.cgi?id=12127#c6 for a more detailed analysis. I'll contact the poster of that message to see if he can offer insight into this problem.
So this looks like a DRM or possibly even other kernel issue.
This problem appears to be specific to this motherboard, so I'm closing as NOTOURBUG.