There's a fencepost error in the GART IOMMU handling on x86_64
in the unmap path.  When testing to see if the bus address is
within the IOMMU window and needs to be unmapped, the start of
the first page *beyond* the window also passes the test.  This
can cause the first doubleword of the next page beyond the gatt
table to be smashed to zero, with unpredictable results depending
on what that page is used for.