Re: [PATCH 10/13] PCI/P2PDMA: support compound page in p2pmem_alloc_mmap()

[Jason Gunthorpe <jgg@ziepe.ca>]

On Mon, Dec 22, 2025 at 10:04:17AM -0700, Logan Gunthorpe wrote:
> I would have expected this code to allocate an appropriately aligned
> block of the p2p memory based on the requirements of the current
> mapping, not based on alignment requirements established when the device
> is probed.

Yeah, I think this is not right too.

I think the flow has become confused by trying to set a static
vmemmap_shift when creating the pgmap. That is not how something like
this should work at all.

Instead the basic idea should be that each mmap systemcall will
determine what folio order it would like to have, it will allocate an
aligned range of physical from the genpool, and then it will alter the
folios in that range into a single high order folio.

Finally the high order folio is installed in one shot with the mm
dealing with placing it optimally in the right page table levels.

You could use a heuristic (ie I'm 2M size aligned or 1G size aligned)
or maybe use the MAP_HUGE_2M/MAP_HUGE_1G flags, or something else
perhaps.

Don't follow what DAX did, this doesn't have the limitations DAX had
to work with.

I also don't think drivers should be open coding the
vm_insert_folio_xx() stuff, the mm should have a helper to accept a
folio of any order, the VA and the phys, then install it optimally. So
don't export vm_insert_folio_pmd()/etc please.

Finally, Peter Xu has been working on the issue of setting the
alignment of VMAs when they will be used to hold large aligned folios,
that would help this be more useful by avoiding the need for MAP_FIXED:

https://lore.kernel.org/kvm/20251204151003.171039-1-peterx@redhat.com/

Assuming the folio size can be determined early enough in the VMA
process, though Lorenzo's recent refactorying here into mmap_prepare
may be helpful.

Jason