Today with folios being an overlay on struct page, folio_alloc() is simple; we allocate a compound page and cast the result to a struct folio pointer.

Once we live in a MatthewWilcox/Memdescs world, we need to do two allocations. First, we allocate a struct folio from slab. Then we call into the MatthewWilcox/BuddyAllocator to allocate the memory. I envisage a function like this:

struct folio *folio_alloc(gfp_t gfp, unsigned int order)
{
        struct page *page;
        struct folio *folio = kmem_cache_alloc(folio_slab, gfp);

        if (!folio)
                return NULL;
        page = alloc_page_desc(gfp, order, MEMDESC_TYPE_FOLIO, folio);
        if (!page) {
                kmem_cache_free(folio_slab, folio);
                return NULL;
        }
        folio->pfn = page_to_pfn(page);
        return folio;
}