Re: [Dri-devel] Re: More DRI breakage

[Dave J. <da...@co...>]

On Fri, Apr 25, 2003 at 08:02:29PM -0700, Linus Torvalds wrote:
 > but my suspicion is that either the AGP or the DRI part leaves 
 > something pending, so that when the hardware is initialized next time 
 > around it will end up using the wrong pages for the frame buffer or 
 > similar.

It's a *lot* of work, (I guess an afternoon or so), but it would rule
out AGP if you backed out everything that happened to the agpgart
driver since circa 2.5.20 or so, and then fixed up any API changes.
Alternatively, drop back to an older kernel, and drop the newer DRM
stuff in.

With both DRI and AGPGART being quite large complicated bits of code,
ruling one or the other out could cut down on search time a lot.

Also, Andrew, don't you carry Manfreds 'unmap page after free' code in -mm?
That could be worth a try perhaps. Long shot, but right now we could use
all the debugging tools we can get.

 > The fact that I'm pretty much up-to-date wrt DRI CVS (I did another merge 
 > today), and Keith doesn't see anything makes me wonder if AGP is the real 
 > culprit. I've tried to walk through the AGP code too - and I've asked Dave 
 > to check it out - but haven't found anything really suspicious there. 

I've gone over pretty much all of it the last few days whilst merging a
large set of changes from Christoph Hellwig, and neither of us noticed
anything odd.
 
 > There's some other corraborating evidence that AGP may be the culprit: I'm
 > definitely not the only one seeing this, judging by some other reports on
 > linux-kernel. One was "Dell Latitude X200, i830M graphic board" and 
 > claimed to not have DRI even enabled (but had several AGP messages - I 
 > thought most drivers ended up using AGP only if DRI was on??)

Could be that DRI loaded, opened agpgart, then found something was
wrong, and disabled itself ?

		Dave