On Mon, Aug 03, 2026 at 07:22:06PM +0100, Matt Evans wrote:
The second problem is this isn't just a limitation to not do an mmap() with offset approaching 1TB, because you could run into trouble just mapping, say, a 2TB BAR and then splitting it by unmapping a hole in the middle: the second VMA now has a huge offset.
It seems a proper fix is easy though:
unsigned long vma_off = (vma->vm_pgoff - priv->vma_pgoff_adjust) << PAGE_SHIFT; /* No masking! */
Then in vfio_pci_core_mmap_prep_dmabuf(),
priv->vma_pgoff_adjust = vma->vm_pgoff;
I.e., if vma_pgoff_adjust just includes the region index up high, it cancels out the (same) index in the vm_pgoff. Thus regular mmap should work (up to the 1TB offset), and DMABUF VMAs can have arbitrarily large offsets.
This is what I was expecting when I pointed at this stuff in the first place :)
Jason