On Thu, Aug 14, 2025 at 10:22:33AM -0300, Jason Gunthorpe wrote:
> On Thu, Aug 07, 2025 at 01:44:13AM +0000, Pasha Tatashin wrote:
> > +int kho_unpreserve_phys(phys_addr_t phys, size_t size)
> > +{
>
> Why are we adding phys apis? Didn't we talk about this before and
> agree not to expose these?
>
> The places using it are goofy:
>
> +static int luo_fdt_setup(void)
> +{
> + fdt_out = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
> + get_order(LUO_FDT_SIZE));
>
> + ret = kho_preserve_phys(__pa(fdt_out), LUO_FDT_SIZE);
>
> + WARN_ON_ONCE(kho_unpreserve_phys(__pa(fdt_out), LUO_FDT_SIZE));
>
> It literally allocated a page and then for some reason switches to
> phys with an open coded __pa??
>
> This is ugly, if you want a helper to match __get_free_pages() then
> make one that works on void * directly. You can get the order of the
> void * directly from the struct page IIRC when using GFP_COMP.
>
> Which is perhaps another comment, if this __get_free_pages() is going
> to be a common pattern (and I guess it will be) then the API should be
> streamlined alot more:
>
> void *kho_alloc_preserved_memory(gfp, size);
> void kho_free_preserved_memory(void *);
This looks backwards to me. KHO should not deal with memory allocation,
it's responsibility to preserve/restore memory objects it supports.
For __get_free_pages() the natural KHO API is kho_(un)preserve_pages().
With struct page/mesdesc we always have page_to_<specialized object> from
one side and page_to_pfn from the other side.
Then folio and phys/virt APIS just become a thin wrappers around the _page
APIs. And down the road we can add slab and maybe vmalloc.
Once folio won't overlap struct page, we'll have a hard time with only
kho_preserve_folio() for memory that's not actually folio (i.e. anon and
page cache)
> Which can wrapper the get_free_pages and the preserve logic and gives
> a nice path to possibly someday supporting non-PAGE_SIZE allocations.
>
> Jason
>
--
Sincerely yours,
Mike.
