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

[Mike Rapoport <rppt@kernel.org>]

On Wed, Jun 11, 2025 at 04:01:52PM +0200, Pratyush Yadav wrote:
> On Wed, Jun 11 2025, Mike Rapoport wrote:
> 
> > On Wed, Jun 11, 2025 at 09:14:55AM -0400, Pasha Tatashin wrote:
> >> On Wed, Jun 11, 2025 at 9:06 AM Pratyush Yadav <pratyush@kernel.org> wrote:
> >> >
> >> > On Tue, Jun 10 2025, Pasha Tatashin wrote:
> >> >
> >> > >> > > I think it should be the other way around, KHO should depend on
> >> > >> > > !DEFERRED_STRUCT_PAGE_INIT.
> >> > >> >
> >> > >> > Agreed, and this is what I first tried, but that does not work, there
> >> > >> > is some circular dependency breaking the build. If you feel
> >> > >> > adventurous you can try that :-)
> >> > >>
> >> > >> Hmm, weird, worked for me :/
> >> >
> >> > Worked for me as well.
> >> >
> >> > >
> >> > > I am super confused, it did not work for me over weekend, and now it
> >> > > is working. Even `make menuconfig` would not work. Anyways, I will put
> >> > > it in the appropriate place.
> >> > >
> >> > >>
> >> > >> > > > 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
> >> > >> >
> >> > >> > We already run KHO init before smp_init:
> >> > >> > start_kernel() -> mm_core_init() -> kho_memory_init() ->
> >> > >> > kho_restore_folio() -> struct pages must be already initialized here!
> >> > >> >
> >> > >> > While deferred struct pages are initialized:
> >> > >> > start_kernel() -> rest_init() -> kernel_init() ->
> >> > >> > kernel_init_freeable() -> page_alloc_init_late() ->
> >> > >> > deferred_init_memmap()
> >> > >> >
> >> > >> > If the number of preserved pages that is needed during early boot is
> >> > >> > relatively small, that it should not be an issue to pre-initialize
> >> > >> > struct pages for them before deferred struct pages are initialized. We
> >> > >> > already pre-initialize some  "struct pages" that are needed during
> >> > >> > early boot before the reset are initialized, see deferred_grow_zone()
> >> > >>
> >> > >> deferred_grow_zone() takes a chunk in the beginning of uninitialized range,
> >> > >> with kho we are talking about some random pages. If we preinit them early,
> >> > >> deferred_init_memmap() will overwrite them.
> >> > >
> >> > > Yes, this is why I am saying that we would need to skip the KHO
> >> > > initialized "struct pages" somehow during deferred initialization. If
> >> > > we create an ordered by PFN list of early-initialized KHO struct
> >> > > pages, skipping during deferred initialization could be done
> >> > > efficiently.
> >> >
> >> > Or keep things simple and don't use any KHO struct pages during early
> >> > init. You can access the page itself, just don't use its struct page.
> >> >
> >> > Currently the only user of kho_restore_folio() during init is
> >> > kho_memory_init(). The FDT is accessed by doing
> >> > phys_to_virt(kho_in.fdt_phys) anyway, so there is really no need for
> >> > restoring the folio so early. It can be done later, for example when LUO
> >> > does the finish event, to clean up and free the folio.
> >> 
> >> Good suggestion, however, KHO does not have any sophisticated users
> >> that we are going to be adding as part of the live update work in the
> >> future: IR, KVM, early VCPU threads, and so on. So, while today, this
> >> might work, in the future, I am not sure if we should expect struct
> >> pages are not accessed until after deferred initialization or simply
> >> fix it once and for all.
> >
> > KHO already accesses stuct page early and uses page->private for order.
> > Since preserved memory is reserved in memblock, deferred init of struct
> > pages won't touch those pages, we just need to make sure they are properly 
> 
> Not strictly true. Some of them might have been initialized from
> free_area_init() -> memmap_init() (the ones not eligible for deferred
> init), which happens before KHO makes its memblock reservations.
> 
> > initialized at some point. If we don't expect many kho_restore_folio()
> > before page_alloc_init_late() we can use init_deferred_page() for early
> > accesses.
> 
> I tried doing this when looking into this initially, but it doesn't work
> for some reason.
> 
>     static void kho_restore_page(struct page *page, unsigned int order)
>     {
>     	unsigned int i, nr_pages = (1 << order);
>     
>     	/* Head page gets refcount of 1. */
>     	init_deferred_page(page_to_pfn(page), NUMA_NO_NODE);


This would do

	if (early_page_initialised(pfn, nid))
		return;

	__init_page_from_nid(pfn, nid);

and I'm really surprised it didn't crash in early_page_initialised()
because of NUMA_NO_NODE :)

What might work here is 

	pfn = page_to_pfn(page);
	__init_page_from_nid(pfn, early_pfn_to_nid(pfn));

>     	set_page_count(page, 1);
>     
>     	/* For higher order folios, tail pages get a page count of zero. */
>     	for (i = 1; i < nr_pages; i++) {
>     		init_deferred_page(page_to_pfn(page + i), NUMA_NO_NODE);
>     		set_page_count(page + i, 0);
>     	}
>     
>     [...]
> 
> results in:
> 
>     [    0.644032] page:(____ptrval____) is uninitialized and poisoned
>     [    0.644679] page dumped because: VM_BUG_ON_PAGE(PagePoisoned(page))
>     [    0.645376] ------------[ cut here ]------------
>     [    0.645883] kernel BUG at ./include/linux/mm.h:1512!
>     [...]
>     [    0.647924] RIP: 0010:__pageblock_pfn_to_page+0x166/0x180
>     [...]
>     [    0.647924]  <TASK>
>     [    0.647924]  set_zone_contiguous+0x6b/0x90
>     [    0.647924]  page_alloc_init_late+0x356/0x370
>     [    0.647924]  kernel_init_freeable+0x12d/0x190
>     [    0.647924]  ? __pfx_kernel_init+0x10/0x10
>     [    0.647924]  kernel_init+0x1a/0x130
> 
> didn't dig any deeper on why it happens...
> 
> -- 
> Regards,
> Pratyush Yadav

-- 
Sincerely yours,
Mike.