Re: [PATCH V7 6/9] iommufd: pfn reader for file mappings

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Nov 14, 2024 at 03:03:36PM +1100, Alexey Kardashevskiy wrote:
> > > I though KVM had its own private interface to guestmemfd?
> > 
> > At the moment it is essentially this:
> > 
> > folio = filemap_grab_folio(file_inode(file)->i_mapping, index);
> 
> Turns out the complete sequence is:
> 
> folio = filemap_grab_folio(file_inode(file)->i_mapping, index);
> folio_add_pins(folio, 1);
> folio_put(folio);
> 
> so then unpin_user_page_range_dirty_lock() => gup_fast_unpin_user_pages() =>
> gup_put_folio() work properly.

I'm surprised given how forceful people were that pinning should not
be done there. Please confirm with guestmemfd experts..

> Later on, hopefully, guest_memfd will developed ability to mmap() and then
> the existing iommufd mapping API will "just work", I suppose. I'll carry on
> the above for now. Thanks,

I think that is not possible, the entire point of guest memfd is that
private memory is never mapped.

For the AMD case we still need to program the address of private
memory into the hypervisor iommu page tables, so the FD path will have
to used to make that connection.

Jason