Re: [RFC V1 0/4] iommufd live update

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Aug 21, 2024 at 01:54:04PM -0400, Steven Sistare wrote:

> memfd_pin_folios() returns an array of folios which must be saved and
> later passed back to unpin_folios().   That's a slight bummer, since
> currently you save memory by retrieving PFNs from the iommu as needed,
> rather than saving all of them explicitly (with the exception of pinned_pfns).

You wouldn't save the folios directly. The existing PFN mechanism is
fine and already works with folios memory.

The mismatch betwen gup returning pages and memfd_pin_folios returning
folios is fixed up by adjusting the folio refcount back to page
granularity, then treating it the same as a bunch of pages and using
unpin_user_pages() as normal.

gup.c already has code that can do this adjustment it would probably
need to be exported. Something like:

folio_refs_to_pages(folio, len);

Then don't change very much in iommufd, just stuff the folio's PFNs
into the existing reader mechanism the same as the pin_user_pages()
flow after fixing the refcount.

Jason