Re: [PATCH] kho: initialize tail pages for higher order folios properly

[Mike Rapoport <rppt@kernel.org>]

On Mon, Jun 09, 2025 at 04:07:50PM -0400, Pasha Tatashin wrote:
> On Mon, Jun 9, 2025 at 3:36 PM Mike Rapoport <rppt@kernel.org> wrote:
> >
> > Hi Pratyush,
> >
> > On Fri, Jun 06, 2025 at 06:23:06PM +0200, Pratyush Yadav wrote:
> > > Hi Mike,
> > >
> > > On Fri, Jun 06 2025, Mike Rapoport wrote:
> > >
> > > > On Thu, Jun 05, 2025 at 07:11:41PM +0200, Pratyush Yadav wrote:
> > > >> From: Pratyush Yadav <ptyadav@amazon.de>
> > > >>
> > > >> --- a/kernel/kexec_handover.c
> > > >> +++ b/kernel/kexec_handover.c
> > > >> @@ -157,11 +157,21 @@ static int __kho_preserve_order(struct kho_mem_track *track, unsigned long pfn,
> > > >>  }
> > > >>
> > > >>  /* almost as free_reserved_page(), just don't free the page */
> > > >> -static void kho_restore_page(struct page *page)
> > > >> +static void kho_restore_page(struct page *page, unsigned int order)
> > > >>  {
> > > >> -  ClearPageReserved(page);
> > > >
> > > > So now we don't clear PG_Reserved even on order-0 pages? ;-)
> > >
> > > We don't need to. As I mentioned in the commit message as well,
> > > PG_Reserved is never set for KHO pages since they are reserved with
> > > MEMBLOCK_RSRV_NOINIT, so memmap_init_reserved_pages() skips over them.
> >
> > You are right, I missed it.
> >
> > > That said, while reading through some of the code, I noticed another
> > > bug: because KHO reserves the preserved pages as NOINIT, with
> > > CONFIG_DEFERRED_STRUCT_PAGE_INIT == n, all the pages get initialized
> > > when memmap_init_range() is called from setup_arch (paging_init() on
> > > x86). This happens before kho_memory_init(), so the KHO-preserved pages
> > > are not marked as reserved to memblock yet.
> > >
> > > With deferred page init, some pages might not get initialized early, and
> > > get initialized after kho_memory_init(), by which time the KHO-preserved
> > > pages are marked as reserved. So, deferred_init_maxorder() will skip
> > > over those pages and leave them uninitialized.
> > >
> > > So we need to either also call init_deferred_page(), or remove the
> > > memblock_reserved_mark_noinit() call in deserialize_bitmap(). And TBH, I
> > > am not sure why KHO pages even need to be marked noinit in the first
> > > place. Probably the only benefit would be if a large chunk of memory is
> > > KHO-preserved, the pages can be initialized later on-demand, reducing
> > > bootup time a bit.
> >
> > One benefit is performance indeed, because in not deferred case the
> > initialization of reserved pages in memmap_init_reserved_pages() is really
> > excessive.
> >
> > But more importantly, if we remove memblock_reserved_mark_noinit(), with
> > CONFIG_DEFERRED_STRUCT_PAGE_INIT we'd loose page->private because the
> > struct page will be cleared after kho_mem_deserialize().
> >
> > > What do you think? Should we drop noinit or call init_deferred_page()?
> > > FWIW, my preference is to drop noinit, since init_deferred_page() is
> > > __meminit and we would have to make sure it doesn't go away after boot.
> >
> > We can't drop noinit and calling init_deferred_page() after boot just won't
> > work because it uses memblock to find the page's node and memblock is gone
> > after init.
> >
> > The simplest short-term solution is to disable KHO when
> > CONFIG_DEFERRED_STRUCT_PAGE_INIT is set and then find an efficient way to
> > make it all work together.
> 
> This is what I've done in LUOv3 WIP:
> https://github.com/soleen/linux/commit/3059f38ac0a39a397873759fb429bd5d1f8ea681

I think it should be the other way around, KHO should depend on
!DEFERRED_STRUCT_PAGE_INIT.
 
> We will need to teah KHO to work with deferred struct page init. I
> suspect, we could init preserved struct pages and then skip over them
> during deferred init.

We could, but with that would mean we'll run this before SMP and it's not
desirable. Also, init_deferred_page() for a random page requires
finding its node with early_pfn_to_nid() that's also suboptimal. 
 
> Pasha
> 
> >
> > --
> > Sincerely yours,
> > Mike.

-- 
Sincerely yours,
Mike.