Re: [v1 resend 03/12] mm/thp: zone_device awareness in THP handling code

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

On Fri, Jul 04, 2025 at 09:35:02AM +1000, Balbir Singh wrote:
> Make THP handling code in the mm subsystem for THP pages
> aware of zone device pages. Although the code is
> designed to be generic when it comes to handling splitting
> of pages, the code is designed to work for THP page sizes
> corresponding to HPAGE_PMD_NR.
> 
> Modify page_vma_mapped_walk() to return true when a zone
> device huge entry is present, enabling try_to_migrate()
> and other code migration paths to appropriately process the
> entry
> 
> pmd_pfn() does not work well with zone device entries, use
> pfn_pmd_entry_to_swap() for checking and comparison as for
> zone device entries.
> 
> try_to_map_to_unused_zeropage() does not apply to zone device
> entries, zone device entries are ignored in the call.
> 
> Cc: Karol Herbst <kherbst@redhat.com>
> Cc: Lyude Paul <lyude@redhat.com>
> Cc: Danilo Krummrich <dakr@kernel.org>
> Cc: David Airlie <airlied@gmail.com>
> Cc: Simona Vetter <simona@ffwll.ch>
> Cc: "Jérôme Glisse" <jglisse@redhat.com>
> Cc: Shuah Khan <shuah@kernel.org>
> Cc: David Hildenbrand <david@redhat.com>
> Cc: Barry Song <baohua@kernel.org>
> Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
> Cc: Ryan Roberts <ryan.roberts@arm.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Peter Xu <peterx@redhat.com>
> Cc: Zi Yan <ziy@nvidia.com>
> Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
> Cc: Jane Chu <jane.chu@oracle.com>
> Cc: Alistair Popple <apopple@nvidia.com>
> Cc: Donet Tom <donettom@linux.ibm.com>
> 
> Signed-off-by: Balbir Singh <balbirs@nvidia.com>
> ---
>  mm/huge_memory.c     | 153 +++++++++++++++++++++++++++++++------------
>  mm/migrate.c         |   2 +
>  mm/page_vma_mapped.c |  10 +++
>  mm/pgtable-generic.c |   6 ++
>  mm/rmap.c            |  19 +++++-
>  5 files changed, 146 insertions(+), 44 deletions(-)
> 
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index ce130225a8e5..e6e390d0308f 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1711,7 +1711,8 @@ int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,

As mentioned in a reply to the cover letter [1], this code is crashing for
us (Intel) when we fork and then read device pages. I’ve suggested a fix
in an another reply [2] and will send Nvidia’s stakeholders a complete
patch with the all necessary fixes to stabilize our code — more on that
below.

[1] https://lore.kernel.org/linux-mm/aHmJ+L3fCc0tju7A@lstrano-desk.jf.intel.com/
[2] https://lore.kernel.org/linux-mm/aHrsdvjjliBBdVQm@lstrano-desk.jf.intel.com/#t

>  	if (unlikely(is_swap_pmd(pmd))) {
>  		swp_entry_t entry = pmd_to_swp_entry(pmd);
>  
> -		VM_BUG_ON(!is_pmd_migration_entry(pmd));
> +		VM_BUG_ON(!is_pmd_migration_entry(pmd) &&
> +				!is_device_private_entry(entry));
>  		if (!is_readable_migration_entry(entry)) {
>  			entry = make_readable_migration_entry(
>  							swp_offset(entry));
> @@ -2222,10 +2223,17 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
>  		} else if (thp_migration_supported()) {
>  			swp_entry_t entry;
>  
> -			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
>  			entry = pmd_to_swp_entry(orig_pmd);
>  			folio = pfn_swap_entry_folio(entry);
>  			flush_needed = 0;
> +
> +			VM_BUG_ON(!is_pmd_migration_entry(*pmd) &&
> +					!folio_is_device_private(folio));
> +
> +			if (folio_is_device_private(folio)) {
> +				folio_remove_rmap_pmd(folio, folio_page(folio, 0), vma);
> +				WARN_ON_ONCE(folio_mapcount(folio) < 0);
> +			}
>  		} else
>  			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
>  
> @@ -2247,6 +2255,15 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
>  				folio_mark_accessed(folio);
>  		}
>  
> +		/*
> +		 * Do a folio put on zone device private pages after
> +		 * changes to mm_counter, because the folio_put() will
> +		 * clean folio->mapping and the folio_test_anon() check
> +		 * will not be usable.
> +		 */
> +		if (folio_is_device_private(folio))
> +			folio_put(folio);
> +
>  		spin_unlock(ptl);
>  		if (flush_needed)
>  			tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE);
> @@ -2375,7 +2392,8 @@ int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
>  		struct folio *folio = pfn_swap_entry_folio(entry);
>  		pmd_t newpmd;
>  
> -		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
> +		VM_BUG_ON(!is_pmd_migration_entry(*pmd) &&
> +			  !folio_is_device_private(folio));
>  		if (is_writable_migration_entry(entry)) {
>  			/*
>  			 * A protection check is difficult so
> @@ -2388,9 +2406,11 @@ int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
>  			newpmd = swp_entry_to_pmd(entry);
>  			if (pmd_swp_soft_dirty(*pmd))
>  				newpmd = pmd_swp_mksoft_dirty(newpmd);
> -		} else {
> +		} else if (is_writable_device_private_entry(entry)) {
> +			newpmd = swp_entry_to_pmd(entry);
> +			entry = make_device_exclusive_entry(swp_offset(entry));
> +		} else
>  			newpmd = *pmd;
> -		}
>  
>  		if (uffd_wp)
>  			newpmd = pmd_swp_mkuffd_wp(newpmd);
> @@ -2842,16 +2862,20 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>  	struct page *page;
>  	pgtable_t pgtable;
>  	pmd_t old_pmd, _pmd;
> -	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
> -	bool anon_exclusive = false, dirty = false;
> +	bool young, write, soft_dirty, uffd_wp = false;
> +	bool anon_exclusive = false, dirty = false, present = false;
>  	unsigned long addr;
>  	pte_t *pte;
>  	int i;
> +	swp_entry_t swp_entry;
>  
>  	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
>  	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
>  	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
> -	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd));
> +
> +	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
> +			&& !(is_swap_pmd(*pmd) &&
> +			is_device_private_entry(pmd_to_swp_entry(*pmd))));
>  
>  	count_vm_event(THP_SPLIT_PMD);
>  
> @@ -2899,20 +2923,25 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>  		return __split_huge_zero_page_pmd(vma, haddr, pmd);

This function is causing a crash with the following test case:

- Fault in a 2M GPU page (i.e., an order-9 device folio is in the PMD)
- User space calls munmap on a partial region

A quick explanation of the crash:

- zap_pmd_range() calls __split_huge_pmd()
- zap_nonpresent_ptes() finds multiple PTE swap entries that all point
  to the same large folio; it decrements the refcount multiple times,
  causing a kernel crash

I believe there are likely several other problematic cases in the kernel
as well, but I only deep-dived into the case above.

The solution I came up with is: if a device-private PMD is found, split
the folio. This seems to work.

Rather than include the fix I came up with here, I’ve just sent Nvidia’s
stakeholders a patch titled "mm: Changes for Nvidia's device THP series
to enable device THP in GPU SVM / Xe", which contains all the core MM
changes we made to stabilize our code. Feel free to make that patch
public for discussion or use it however you see fit.

Matt

>  	}
>  
> -	pmd_migration = is_pmd_migration_entry(*pmd);
> -	if (unlikely(pmd_migration)) {
> -		swp_entry_t entry;
>  
> +	present = pmd_present(*pmd);
> +	if (unlikely(!present)) {
> +		swp_entry = pmd_to_swp_entry(*pmd);
>  		old_pmd = *pmd;
> -		entry = pmd_to_swp_entry(old_pmd);
> -		page = pfn_swap_entry_to_page(entry);
> -		write = is_writable_migration_entry(entry);
> +
> +		folio = pfn_swap_entry_folio(swp_entry);
> +		VM_BUG_ON(!is_migration_entry(swp_entry) &&
> +				!is_device_private_entry(swp_entry));
> +		page = pfn_swap_entry_to_page(swp_entry);
> +		write = is_writable_migration_entry(swp_entry);
> +
>  		if (PageAnon(page))
> -			anon_exclusive = is_readable_exclusive_migration_entry(entry);
> -		young = is_migration_entry_young(entry);
> -		dirty = is_migration_entry_dirty(entry);
> +			anon_exclusive =
> +				is_readable_exclusive_migration_entry(swp_entry);
>  		soft_dirty = pmd_swp_soft_dirty(old_pmd);
>  		uffd_wp = pmd_swp_uffd_wp(old_pmd);
> +		young = is_migration_entry_young(swp_entry);
> +		dirty = is_migration_entry_dirty(swp_entry);
>  	} else {
>  		/*
>  		 * Up to this point the pmd is present and huge and userland has
> @@ -2996,30 +3025,45 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>  	 * Note that NUMA hinting access restrictions are not transferred to
>  	 * avoid any possibility of altering permissions across VMAs.
>  	 */
> -	if (freeze || pmd_migration) {
> +	if (freeze || !present) {
>  		for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
>  			pte_t entry;
> -			swp_entry_t swp_entry;
> -
> -			if (write)
> -				swp_entry = make_writable_migration_entry(
> -							page_to_pfn(page + i));
> -			else if (anon_exclusive)
> -				swp_entry = make_readable_exclusive_migration_entry(
> -							page_to_pfn(page + i));
> -			else
> -				swp_entry = make_readable_migration_entry(
> -							page_to_pfn(page + i));
> -			if (young)
> -				swp_entry = make_migration_entry_young(swp_entry);
> -			if (dirty)
> -				swp_entry = make_migration_entry_dirty(swp_entry);
> -			entry = swp_entry_to_pte(swp_entry);
> -			if (soft_dirty)
> -				entry = pte_swp_mksoft_dirty(entry);
> -			if (uffd_wp)
> -				entry = pte_swp_mkuffd_wp(entry);
> -
> +			if (freeze || is_migration_entry(swp_entry)) {
> +				if (write)
> +					swp_entry = make_writable_migration_entry(
> +								page_to_pfn(page + i));
> +				else if (anon_exclusive)
> +					swp_entry = make_readable_exclusive_migration_entry(
> +								page_to_pfn(page + i));
> +				else
> +					swp_entry = make_readable_migration_entry(
> +								page_to_pfn(page + i));
> +				if (young)
> +					swp_entry = make_migration_entry_young(swp_entry);
> +				if (dirty)
> +					swp_entry = make_migration_entry_dirty(swp_entry);
> +				entry = swp_entry_to_pte(swp_entry);
> +				if (soft_dirty)
> +					entry = pte_swp_mksoft_dirty(entry);
> +				if (uffd_wp)
> +					entry = pte_swp_mkuffd_wp(entry);
> +			} else {
> +				VM_BUG_ON(!is_device_private_entry(swp_entry));
> +				if (write)
> +					swp_entry = make_writable_device_private_entry(
> +								page_to_pfn(page + i));
> +				else if (anon_exclusive)
> +					swp_entry = make_device_exclusive_entry(
> +								page_to_pfn(page + i));
> +				else
> +					swp_entry = make_readable_device_private_entry(
> +								page_to_pfn(page + i));
> +				entry = swp_entry_to_pte(swp_entry);
> +				if (soft_dirty)
> +					entry = pte_swp_mksoft_dirty(entry);
> +				if (uffd_wp)
> +					entry = pte_swp_mkuffd_wp(entry);
> +			}
>  			VM_WARN_ON(!pte_none(ptep_get(pte + i)));
>  			set_pte_at(mm, addr, pte + i, entry);
>  		}
> @@ -3046,7 +3090,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>  	}
>  	pte_unmap(pte);
>  
> -	if (!pmd_migration)
> +	if (present)
>  		folio_remove_rmap_pmd(folio, page, vma);
>  	if (freeze)
>  		put_page(page);
> @@ -3058,8 +3102,11 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>  void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
>  			   pmd_t *pmd, bool freeze)
>  {
> +
>  	VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE));
> -	if (pmd_trans_huge(*pmd) || is_pmd_migration_entry(*pmd))
> +	if (pmd_trans_huge(*pmd) || is_pmd_migration_entry(*pmd) ||
> +			(is_swap_pmd(*pmd) &&
> +			is_device_private_entry(pmd_to_swp_entry(*pmd))))
>  		__split_huge_pmd_locked(vma, pmd, address, freeze);
>  }
>  
> @@ -3238,6 +3285,9 @@ static void lru_add_split_folio(struct folio *folio, struct folio *new_folio,
>  	VM_BUG_ON_FOLIO(folio_test_lru(new_folio), folio);
>  	lockdep_assert_held(&lruvec->lru_lock);
>  
> +	if (folio_is_device_private(folio))
> +		return;
> +
>  	if (list) {
>  		/* page reclaim is reclaiming a huge page */
>  		VM_WARN_ON(folio_test_lru(folio));
> @@ -3252,6 +3302,7 @@ static void lru_add_split_folio(struct folio *folio, struct folio *new_folio,
>  			list_add_tail(&new_folio->lru, &folio->lru);
>  		folio_set_lru(new_folio);
>  	}
> +
>  }
>  
>  /* Racy check whether the huge page can be split */
> @@ -3543,6 +3594,10 @@ static int __split_unmapped_folio(struct folio *folio, int new_order,
>  					((mapping || swap_cache) ?
>  						folio_nr_pages(release) : 0));
>  
> +			if (folio_is_device_private(release))
> +				percpu_ref_get_many(&release->pgmap->ref,
> +							(1 << new_order) - 1);
> +
>  			lru_add_split_folio(origin_folio, release, lruvec,
>  					list);
>  
> @@ -4596,7 +4651,10 @@ int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
>  		return 0;
>  
>  	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
> -	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
> +	if (!folio_is_device_private(folio))
> +		pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
> +	else
> +		pmdval = pmdp_huge_clear_flush(vma, address, pvmw->pmd);
>  
>  	/* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */
>  	anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);
> @@ -4646,6 +4704,17 @@ void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
>  	entry = pmd_to_swp_entry(*pvmw->pmd);
>  	folio_get(folio);
>  	pmde = folio_mk_pmd(folio, READ_ONCE(vma->vm_page_prot));
> +
> +	if (unlikely(folio_is_device_private(folio))) {
> +		if (pmd_write(pmde))
> +			entry = make_writable_device_private_entry(
> +							page_to_pfn(new));
> +		else
> +			entry = make_readable_device_private_entry(
> +							page_to_pfn(new));
> +		pmde = swp_entry_to_pmd(entry);
> +	}
> +
>  	if (pmd_swp_soft_dirty(*pvmw->pmd))
>  		pmde = pmd_mksoft_dirty(pmde);
>  	if (is_writable_migration_entry(entry))
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 767f503f0875..0b6ecf559b22 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -200,6 +200,8 @@ static bool try_to_map_unused_to_zeropage(struct page_vma_mapped_walk *pvmw,
>  
>  	if (PageCompound(page))
>  		return false;
> +	if (folio_is_device_private(folio))
> +		return false;
>  	VM_BUG_ON_PAGE(!PageAnon(page), page);
>  	VM_BUG_ON_PAGE(!PageLocked(page), page);
>  	VM_BUG_ON_PAGE(pte_present(ptep_get(pvmw->pte)), page);
> diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
> index e981a1a292d2..ff8254e52de5 100644
> --- a/mm/page_vma_mapped.c
> +++ b/mm/page_vma_mapped.c
> @@ -277,6 +277,16 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
>  			 * cannot return prematurely, while zap_huge_pmd() has
>  			 * cleared *pmd but not decremented compound_mapcount().
>  			 */
> +			swp_entry_t entry;
> +
> +			if (!thp_migration_supported())
> +				return not_found(pvmw);
> +			entry = pmd_to_swp_entry(pmde);
> +			if (is_device_private_entry(entry)) {
> +				pvmw->ptl = pmd_lock(mm, pvmw->pmd);
> +				return true;
> +			}
> +
>  			if ((pvmw->flags & PVMW_SYNC) &&
>  			    thp_vma_suitable_order(vma, pvmw->address,
>  						   PMD_ORDER) &&
> diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c
> index 567e2d084071..604e8206a2ec 100644
> --- a/mm/pgtable-generic.c
> +++ b/mm/pgtable-generic.c
> @@ -292,6 +292,12 @@ pte_t *___pte_offset_map(pmd_t *pmd, unsigned long addr, pmd_t *pmdvalp)
>  		*pmdvalp = pmdval;
>  	if (unlikely(pmd_none(pmdval) || is_pmd_migration_entry(pmdval)))
>  		goto nomap;
> +	if (is_swap_pmd(pmdval)) {
> +		swp_entry_t entry = pmd_to_swp_entry(pmdval);
> +
> +		if (is_device_private_entry(entry))
> +			goto nomap;
> +	}
>  	if (unlikely(pmd_trans_huge(pmdval)))
>  		goto nomap;
>  	if (unlikely(pmd_bad(pmdval))) {
> diff --git a/mm/rmap.c b/mm/rmap.c
> index bd83724d14b6..da1e5b03e1fe 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -2336,8 +2336,23 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
>  				break;
>  			}
>  #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
> -			subpage = folio_page(folio,
> -				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
> +			/*
> +			 * Zone device private folios do not work well with
> +			 * pmd_pfn() on some architectures due to pte
> +			 * inversion.
> +			 */
> +			if (folio_is_device_private(folio)) {
> +				swp_entry_t entry = pmd_to_swp_entry(*pvmw.pmd);
> +				unsigned long pfn = swp_offset_pfn(entry);
> +
> +				subpage = folio_page(folio, pfn
> +							- folio_pfn(folio));
> +			} else {
> +				subpage = folio_page(folio,
> +							pmd_pfn(*pvmw.pmd)
> +							- folio_pfn(folio));
> +			}
> +
>  			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
>  					!folio_test_pmd_mappable(folio), folio);
>  
> -- 
> 2.49.0
>