data corruption with nvidia chipsets and IDE/SATA drives // memory hole mapping related bug?!
I'd like to mention what might be a new twist on this problem. We are seeing the same kind of 4k-block data corruption on multiple Tyan dual-Opteron boards (S3870) with a ServerWorks chipset, not Nvidia. I wonder if it really an Nvidia-specific issue. The Nvidia boards are a lot more popular, so maybe it's just more likely to be seen with Nvidia. I added a more detailed comment about this to the kernel bug report: http://bugzilla.kernel.org/show_bug.cgi?id=7768#c15.