Yesterday, Rik van Riel, William Lee Irwin and myself were able to discuss this issue of Athlon/AGP instability with AMD. This converation was very enlightening, to say the least.

Let's start with some facts. The issue that is being discussed here is not related to the so-called "INVPLG bug". So, what is the issue, and who's at fault? Yesterday, AMD offered us an explanation of what may be happening to cause AGP instability on Linux systems. This explanation appears to be the exact scenario that bit Windows 2000 in September 2000. AMD characterizes this issue as a cache coherency problem resulting from how other components on the motherboard (the GART in particular) interact with a particular performance-enhancing feature of the Athlon Processor called speculative writes. The GART and the CPU have two different views of memory -- the GART's view does not "see" the contents of the Athlon's internal caches, and herein lies the problem. In certain situations the CPU and the GART can step on each other's toes, resulting in memory corruption. For more information, see my Linux kernel mailing list post that contains an explanation of this particular problem in AMD's own words.

So, is the Linux kernel affected by this problem? It appears so. The GART and the CPU see two different views of memory, and it's the kernel's responsibility to map memory in such a way as to prevent bad interactions. Currently, that isn't happening. Third-party drivers such as the NVIDIA driver for Linux may also have some problems in this area. Now that the Linux kernel development community is aware of the issue, they have started the process of devising a good approach to avoid this memory corruption problem. Disabling extended paging via the mem=nopentium boot option is really not a good long-term solution since it can impact performance, so a better solution is needed.

I guess the big question on everyone's mind is this -- who's to blame for this very real problem? Strangely enough, no one party appears to be at fault here. But now that the problem is out in the open, the solution is clear. The Linux kernel's approach to memory management must become more sophisticated in order to address potential conflicts between the highly-speculative nature of Athlon processors and the non-cache-coherent AGP GART.

I'll be following the resolution of this issue closely and post information here as I get it.