Re: [PATCH V5 5/9] iommufd: folio subroutines

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Oct 23, 2024 at 07:21:05AM +0000, Tian, Kevin wrote:
> > From: Steve Sistare <steven.sistare@oracle.com>
> > Sent: Wednesday, October 23, 2024 5:01 AM
> > 
> > +static int batch_from_folios(struct pfn_batch *batch, struct folio ***folios_p,
> > +			     unsigned long *offset_p, unsigned long npages)
> > +{
> > +	int rc = 0;
> > +	struct folio **folios = *folios_p;
> > +	unsigned long offset = *offset_p;
> > +
> > +	while (npages) {
> > +		struct folio *folio = *folios;
> > +		unsigned long nr = folio_nr_pages(folio) - offset;
> > +		unsigned long pfn = page_to_pfn(folio_page(folio, offset));
> > +
> > +		nr = min(nr, npages);
> > +		npages -= nr;
> > +
> > +		if (!batch_add_pfn_num(batch, pfn, nr))
> > +			break;
> > +		if (nr > 1) {
> > +			rc = folio_add_pins(folio, nr - 1);
> > +			if (rc) {
> > +				batch_remove_pfn_num(batch, nr);
> > +				goto out;
> > +			}
> > +		}
> 
> forgot to ask one thing here. Could you add a comment here why
> additional pins is required only when batching more than 1 page?

Something like:

 The unpin path during unmap can unpin slices of a single folio, and
 broadly does not keep track of where the folio boundaries were. Since
 it treats everything as a single page transform the refcount on the
 folio from a per-folio ref to a per-page ref. This makes the unmap
 path uniform regardless of how the memory was pinned.

Jason