I've attached a kernel patch (pci-gart.c.patch.gz), which is believed to resolve this problem, without any noticeable performance impact. The patch should be applied to $SRC/arch/x86_64/kernel/pci-gart.c
For those running RHEL-4.4-x86_64, Redhat has also provided the following kernel RPM which includes this same patch:
Please note that you should *NOT* boot with the iommu kernel parameter when you are using this patch, as doing so would mask the bug (and potential fix). Please feel free to redistribute this patch, however note that it has not yet been accepted into any official kernel tree/branch.
Testing & feedback would be appreciated.