Re: [PATCH 1/4] mm/migrate: Add migrate_device_split_page

[Matthew Brost <matthew.brost@intel.com>]

On Thu, Jan 08, 2026 at 11:56:03AM +1100, Balbir Singh wrote:
> On 1/8/26 08:03, Zi Yan wrote:
> > On 7 Jan 2026, at 16:15, Matthew Brost wrote:
> > 
> >> On Wed, Jan 07, 2026 at 03:38:35PM -0500, Zi Yan wrote:
> >>> On 7 Jan 2026, at 15:20, Zi Yan wrote:
> >>>
> >>>> +THP folks
> >>>
> >>> +willy, since he commented in another thread.
> >>>
> >>>>
> >>>> On 16 Dec 2025, at 15:10, Francois Dugast wrote:
> >>>>
> >>>>> From: Matthew Brost <matthew.brost@intel.com>
> >>>>>
> >>>>> Introduce migrate_device_split_page() to split a device page into
> >>>>> lower-order pages. Used when a folio allocated as higher-order is freed
> >>>>> and later reallocated at a smaller order by the driver memory manager.
> >>>>>
> >>>>> Cc: Andrew Morton <akpm@linux-foundation.org>
> >>>>> Cc: Balbir Singh <balbirs@nvidia.com>
> >>>>> Cc: dri-devel@lists.freedesktop.org
> >>>>> Cc: linux-mm@kvack.org
> >>>>> Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> >>>>> Signed-off-by: Francois Dugast <francois.dugast@intel.com>
> >>>>> ---
> >>>>>  include/linux/huge_mm.h |  3 +++
> >>>>>  include/linux/migrate.h |  1 +
> >>>>>  mm/huge_memory.c        |  6 ++---
> >>>>>  mm/migrate_device.c     | 49 +++++++++++++++++++++++++++++++++++++++++
> >>>>>  4 files changed, 56 insertions(+), 3 deletions(-)
> >>>>>
> >>>>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> >>>>> index a4d9f964dfde..6ad8f359bc0d 100644
> >>>>> --- a/include/linux/huge_mm.h
> >>>>> +++ b/include/linux/huge_mm.h
> >>>>> @@ -374,6 +374,9 @@ int __split_huge_page_to_list_to_order(struct page *page, struct list_head *list
> >>>>>  int folio_split_unmapped(struct folio *folio, unsigned int new_order);
> >>>>>  unsigned int min_order_for_split(struct folio *folio);
> >>>>>  int split_folio_to_list(struct folio *folio, struct list_head *list);
> >>>>> +int __split_unmapped_folio(struct folio *folio, int new_order,
> >>>>> +			   struct page *split_at, struct xa_state *xas,
> >>>>> +			   struct address_space *mapping, enum split_type split_type);
> >>>>>  int folio_check_splittable(struct folio *folio, unsigned int new_order,
> >>>>>  			   enum split_type split_type);
> >>>>>  int folio_split(struct folio *folio, unsigned int new_order, struct page *page,
> >>>>> diff --git a/include/linux/migrate.h b/include/linux/migrate.h
> >>>>> index 26ca00c325d9..ec65e4fd5f88 100644
> >>>>> --- a/include/linux/migrate.h
> >>>>> +++ b/include/linux/migrate.h
> >>>>> @@ -192,6 +192,7 @@ void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns,
> >>>>>  			unsigned long npages);
> >>>>>  void migrate_device_finalize(unsigned long *src_pfns,
> >>>>>  			unsigned long *dst_pfns, unsigned long npages);
> >>>>> +int migrate_device_split_page(struct page *page);
> >>>>>
> >>>>>  #endif /* CONFIG_MIGRATION */
> >>>>>
> >>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >>>>> index 40cf59301c21..7ded35a3ecec 100644
> >>>>> --- a/mm/huge_memory.c
> >>>>> +++ b/mm/huge_memory.c
> >>>>> @@ -3621,9 +3621,9 @@ static void __split_folio_to_order(struct folio *folio, int old_order,
> >>>>>   * Return: 0 - successful, <0 - failed (if -ENOMEM is returned, @folio might be
> >>>>>   * split but not to @new_order, the caller needs to check)
> >>>>>   */
> >>>>> -static int __split_unmapped_folio(struct folio *folio, int new_order,
> >>>>> -		struct page *split_at, struct xa_state *xas,
> >>>>> -		struct address_space *mapping, enum split_type split_type)
> >>>>> +int __split_unmapped_folio(struct folio *folio, int new_order,
> >>>>> +			   struct page *split_at, struct xa_state *xas,
> >>>>> +			   struct address_space *mapping, enum split_type split_type)
> >>>>>  {
> >>>>>  	const bool is_anon = folio_test_anon(folio);
> >>>>>  	int old_order = folio_order(folio);
> >>>>> diff --git a/mm/migrate_device.c b/mm/migrate_device.c
> >>>>> index 23379663b1e1..eb0f0e938947 100644
> >>>>> --- a/mm/migrate_device.c
> >>>>> +++ b/mm/migrate_device.c
> >>>>> @@ -775,6 +775,49 @@ int migrate_vma_setup(struct migrate_vma *args)
> >>>>>  EXPORT_SYMBOL(migrate_vma_setup);
> >>>>>
> >>>>>  #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
> >>>>> +/**
> >>>>> + * migrate_device_split_page() - Split device page
> >>>>> + * @page: Device page to split
> >>>>> + *
> >>>>> + * Splits a device page into smaller pages. Typically called when reallocating a
> >>>>> + * folio to a smaller size. Inherently racy—only safe if the caller ensures
> >>>>> + * mutual exclusion within the page's folio (i.e., no other threads are using
> >>>>> + * pages within the folio). Expected to be called a free device page and
> >>>>> + * restores all split out pages to a free state.
> >>>>> + */
> >>>
> >>> Do you mind explaining why __split_unmapped_folio() is needed for a free device
> >>> page? A free page is not supposed to be a large folio, at least from a core
> >>> MM point of view. __split_unmapped_folio() is intended to work on large folios
> >>> (or compound pages), even if the input folio has refcount == 0 (because it is
> >>> frozen).
> >>>
> >>
> >> Well, then maybe this is a bug in core MM where the freed page is still
> >> a THP. Let me explain the scenario and why this is needed from my POV.
> >>
> >> Our VRAM allocator in Xe (and several other DRM drivers) is DRM buddy.
> >> This is a shared pool between traditional DRM GEMs (buffer objects) and
> >> SVM allocations (pages). It doesn’t have any view of the page backing—it
> >> basically just hands back a pointer to VRAM space that we allocate from.
> >> From that, if it’s an SVM allocation, we can derive the device pages.
> >>
> >> What I see happening is: a 2M buddy allocation occurs, we make the
> >> backing device pages a large folio, and sometime later the folio
> >> refcount goes to zero and we free the buddy allocation. Later, the buddy
> >> allocation is reused for a smaller allocation (e.g., 4K or 64K), but the
> >> backing pages are still a large folio. Here is where we need to split
> > 
> > I agree with you that it might be a bug in free_zone_device_folio() based
> > on my understanding. Since zone_device_page_init() calls prep_compound_page()
> > for >0 orders, but free_zone_device_folio() never reverse the process.
> > 
> > Balbir and Alistair might be able to help here.
> 
> I agree it's an API limitation
> 
> > 
> > I cherry picked the code from __free_frozen_pages() to reverse the process.
> > Can you give it a try to see if it solve the above issue? Thanks.
> > 
> > From 3aa03baa39b7e62ea079e826de6ed5aab3061e46 Mon Sep 17 00:00:00 2001
> > From: Zi Yan <ziy@nvidia.com>
> > Date: Wed, 7 Jan 2026 16:49:52 -0500
> > Subject: [PATCH] mm/memremap: free device private folio fix
> > Content-Type: text/plain; charset="utf-8"
> > 
> > Signed-off-by: Zi Yan <ziy@nvidia.com>
> > ---
> >  mm/memremap.c | 15 +++++++++++++++
> >  1 file changed, 15 insertions(+)
> > 
> > diff --git a/mm/memremap.c b/mm/memremap.c
> > index 63c6ab4fdf08..483666ff7271 100644
> > --- a/mm/memremap.c
> > +++ b/mm/memremap.c
> > @@ -475,6 +475,21 @@ void free_zone_device_folio(struct folio *folio)
> >  		pgmap->ops->folio_free(folio);
> >  		break;
> >  	}
> > +
> > +	if (nr > 1) {
> > +		struct page *head = folio_page(folio, 0);
> > +
> > +		head[1].flags.f &= ~PAGE_FLAGS_SECOND;
> > +#ifdef NR_PAGES_IN_LARGE_FOLIO
> > +		folio->_nr_pages = 0;
> > +#endif
> > +		for (i = 1; i < nr; i++) {
> > +			(head + i)->mapping = NULL;
> > +			clear_compound_head(head + i);
> 
> I see that your skipping the checks in free_page_tail_prepare()? IIUC, we should be able
> to invoke it even for zone device private pages
> 
> > +		}
> > +		folio->mapping = NULL;
> 
> This is already done in free_zone_device_folio()
> 
> > +		head->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
> 
> I don't think this is required for zone device private folios, but I suppose it
> keeps the code generic
> 

Well, the above code doesn’t work, but I think it’s the right idea.
clear_compound_head aliases to pgmap, which we don’t want to be NULL. I
believe the individual pages likely need their flags cleared (?), and
this step must be done before calling folio_free and include a barrier,
as the page can be immediately reallocated.

So here’s what I came up with, and it seems to work (for Xe, GPU SVM):

 mm/memremap.c | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/mm/memremap.c b/mm/memremap.c
index 63c6ab4fdf08..ac20abb6a441 100644
--- a/mm/memremap.c
+++ b/mm/memremap.c
@@ -448,6 +448,27 @@ void free_zone_device_folio(struct folio *folio)
            pgmap->type != MEMORY_DEVICE_GENERIC)
                folio->mapping = NULL;

+       if (nr > 1) {
+               struct page *head = folio_page(folio, 0);
+
+               head[1].flags.f &= ~PAGE_FLAGS_SECOND;
+#ifdef NR_PAGES_IN_LARGE_FOLIO
+               folio->_nr_pages = 0;
+#endif
+               for (i = 1; i < nr; i++) {
+                       struct folio *new_folio = (struct folio *)(head + i);
+
+                       (head + i)->mapping = NULL;
+                       (head + i)->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
+
+                       /* Overwrite compound_head with pgmap */
+                       new_folio->pgmap = pgmap;
+               }
+
+               head->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
+               smp_wmb();	/* Changes but be visable before freeing folio */
+       }
+
        switch (pgmap->type) {
        case MEMORY_DEVICE_PRIVATE:
        case MEMORY_DEVICE_COHERENT:

> > +	}
> >  }
> > 
> >  void zone_device_page_init(struct page *page, unsigned int order)
> 
> 
> Otherwise, it seems like the right way to solve the issue.
> 

My question is: why isn’t Nouveau hitting this issue, or your Nvidia
out-of-tree driver (lack of testing, Xe's test suite coverage is quite
good at finding corners).

Also, will this change in behavior break either ofthose drivers?

Matt

> Balbir