Re: [v6 04/15] mm/huge_memory: implement device-private THP splitting

[Balbir Singh <balbirs@nvidia.com>]

On 9/23/25 12:09, Zi Yan wrote:
> On 22 Sep 2025, at 21:50, Balbir Singh wrote:
> 
>> On 9/23/25 07:09, Zi Yan wrote:
>>> On 16 Sep 2025, at 8:21, Balbir Singh wrote:
>>>
>>>> Add support for splitting device-private THP folios, enabling fallback
>>>> to smaller page sizes when large page allocation or migration fails.
>>>>
>>>> Key changes:
>>>> - split_huge_pmd(): Handle device-private PMD entries during splitting
>>>> - Preserve RMAP_EXCLUSIVE semantics for anonymous exclusive folios
>>>> - Skip RMP_USE_SHARED_ZEROPAGE for device-private entries as they
>>>>   don't support shared zero page semantics
>>>>
>>>> Signed-off-by: Balbir Singh <balbirs@nvidia.com>
>>>> Cc: David Hildenbrand <david@redhat.com>
>>>> Cc: Zi Yan <ziy@nvidia.com>
>>>> Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
>>>> Cc: Rakie Kim <rakie.kim@sk.com>
>>>> Cc: Byungchul Park <byungchul@sk.com>
>>>> Cc: Gregory Price <gourry@gourry.net>
>>>> Cc: Ying Huang <ying.huang@linux.alibaba.com>
>>>> Cc: Alistair Popple <apopple@nvidia.com>
>>>> Cc: Oscar Salvador <osalvador@suse.de>
>>>> Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
>>>> Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
>>>> Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
>>>> Cc: Nico Pache <npache@redhat.com>
>>>> Cc: Ryan Roberts <ryan.roberts@arm.com>
>>>> Cc: Dev Jain <dev.jain@arm.com>
>>>> Cc: Barry Song <baohua@kernel.org>
>>>> 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: Ralph Campbell <rcampbell@nvidia.com>
>>>> Cc: Mika Penttilä <mpenttil@redhat.com>
>>>> Cc: Matthew Brost <matthew.brost@intel.com>
>>>> Cc: Francois Dugast <francois.dugast@intel.com>
>>>> ---
>>>>  mm/huge_memory.c | 138 +++++++++++++++++++++++++++++++++--------------
>>>>  1 file changed, 98 insertions(+), 40 deletions(-)
>>>>
>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>>> index 78166db72f4d..5291ee155a02 100644
>>>> --- a/mm/huge_memory.c
>>>> +++ b/mm/huge_memory.c
>>>> @@ -2872,16 +2872,18 @@ 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 soft_dirty, uffd_wp = false, young = false, write = 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_WARN_ON(!is_pmd_non_present_folio_entry(*pmd) && !pmd_trans_huge(*pmd));
>>>>
>>>>  	count_vm_event(THP_SPLIT_PMD);
>>>>
>>>> @@ -2929,20 +2931,47 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>>>>  		return __split_huge_zero_page_pmd(vma, haddr, pmd);
>>>>  	}
>>>>
>>>> -	pmd_migration = is_pmd_migration_entry(*pmd);
>>>> -	if (unlikely(pmd_migration)) {
>>>> -		swp_entry_t entry;
>>>>
>>>> +	present = pmd_present(*pmd);
>>>> +	if (is_pmd_migration_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);
>>>> +		swp_entry = pmd_to_swp_entry(old_pmd);
>>>> +		page = pfn_swap_entry_to_page(swp_entry);
>>>> +		folio = page_folio(page);
>>>> +
>>>> +		soft_dirty = pmd_swp_soft_dirty(old_pmd);
>>>> +		uffd_wp = pmd_swp_uffd_wp(old_pmd);
>>>> +
>>>> +		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);
>>>> +		young = is_migration_entry_young(swp_entry);
>>>> +		dirty = is_migration_entry_dirty(swp_entry);
>>>> +	} else if (is_pmd_device_private_entry(*pmd)) {
>>>> +		old_pmd = *pmd;
>>>> +		swp_entry = pmd_to_swp_entry(old_pmd);
>>>> +		page = pfn_swap_entry_to_page(swp_entry);
>>>> +		folio = page_folio(page);
>>>> +
>>>>  		soft_dirty = pmd_swp_soft_dirty(old_pmd);
>>>>  		uffd_wp = pmd_swp_uffd_wp(old_pmd);
>>>> +
>>>> +		write = is_writable_device_private_entry(swp_entry);
>>>> +		anon_exclusive = PageAnonExclusive(page);
>>>> +
>>>> +		if (freeze && anon_exclusive &&
>>>> +		    folio_try_share_anon_rmap_pmd(folio, page))
>>>> +			freeze = false;
>>>
>>> Why is it OK to change the freeze request? OK, it is replicating
>>> the code for present PMD folios. Either add a comment to point
>>> to the explanation in the comment below, or move
>>> “if (is_pmd_device_private_entry(*pmd))“ branch in the else below
>>> to deduplicate this code.
>>
>> Similar to the code for present pages, ideally folio_try_share_anon_rmap_pmd()
>> should never fail.
> 
> anon_exclusive = PageAnonExclusive(page);
> if (freeze && anon_exclusive &&
>     folio_try_share_anon_rmap_pmd(folio, page))
>         freeze = false;
> if (!freeze) {
>         rmap_t rmap_flags = RMAP_NONE;
> 
>         folio_ref_add(folio, HPAGE_PMD_NR - 1);
>         if (anon_exclusive)
>                 rmap_flags |= RMAP_EXCLUSIVE;
>         folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
>                                     vma, haddr, rmap_flags);
> }
> 
> are the same for both device private and present. Can it be deduplicated
> by doing below?
> 
> if (is_pmd_migration_entry(*pmd)) {
> ...
> } else {
> 	if (is_pmd_device_private_entry(*pmd)) {
> 		...
> 	} else if (pmd_present()) {
> 		...
> 	}
> 
> 	/* the above code */
> }
> 
> If not, at least adding a comment in the device private copy of the code
> pointing to the present copy's comment.
> 
>>
>>>
>>>> +		if (!freeze) {
>>>> +			rmap_t rmap_flags = RMAP_NONE;
>>>> +
>>>> +			folio_ref_add(folio, HPAGE_PMD_NR - 1);
>>>> +			if (anon_exclusive)
>>>> +				rmap_flags |= RMAP_EXCLUSIVE;
>>>> +
>>>> +			folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
>>>> +						 vma, haddr, rmap_flags);
>>>> +		}
>>>>  	} else {
>>>>  		/*
>>>>  		 * Up to this point the pmd is present and huge and userland has
>>>> @@ -3026,32 +3055,57 @@ 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) {
>>>> -		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 || !present) {
>>>> +		pte_t entry;
>>>>
>>>> -			VM_WARN_ON(!pte_none(ptep_get(pte + i)));
>>>> -			set_pte_at(mm, addr, pte + i, entry);
>>>> +		if (freeze || is_migration_entry(swp_entry)) {
>>>>
>>> <snip>
>>>> +		} else {
>>> <snip>
>>>>  		}
>>>>  	} else {
>>>>  		pte_t entry;
>>>
>>> David already pointed this out in v5. It can be done such as:
>>>
>>> if (freeze || pmd_migration) {
>>> ...
>>> } else if (is_pmd_device_private_entry(old_pmd)) {
>>> ...
>>
>> No.. freeze can be true for device private entries as well
> 
> When freeze is true, migration entry is installed in place of
> device private entry, since the "if (freeze || pmd_migration)"
> branch is taken. This proposal is same as your code. What is
> the difference?
> 

I read the else if incorrectly, I'll simplify

>>
>>> } else {
>>> /* for present, non freeze case */
>>> }
>>>
>>>> @@ -3076,7 +3130,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>>>>  	}
>>>>  	pte_unmap(pte);
>>>>
>>>> -	if (!pmd_migration)
>>>> +	if (!is_pmd_migration_entry(*pmd))
>>>>  		folio_remove_rmap_pmd(folio, page, vma);
>>>>  	if (freeze)
>>>>  		put_page(page);
>>>> @@ -3089,7 +3143,7 @@ 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_non_present_folio_entry(*pmd))
>>>>  		__split_huge_pmd_locked(vma, pmd, address, freeze);
>>>>  }
>>>>
>>>> @@ -3268,6 +3322,9 @@ 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));
>>>> @@ -3885,8 +3942,9 @@ static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>  	if (nr_shmem_dropped)
>>>>  		shmem_uncharge(mapping->host, nr_shmem_dropped);
>>>>
>>>> -	if (!ret && is_anon)
>>>> +	if (!ret && is_anon && !folio_is_device_private(folio))
>>>>  		remap_flags = RMP_USE_SHARED_ZEROPAGE;
>>>> +
>>>
>>> You should remove this and add
>>>
>>> if (folio_is_device_private(folio))
>>> 	return false;
>>>
>>> in try_to_map_unused_to_zeropage(). Otherwise, no one would know
>>> device private folios need to be excluded from mapping unused to
>>> zero page.
>>>
>>
>> I had that upto v2 and then David asked me to remove it. FYI, this
>> is the only call site for RMP_USE_SHARED_ZEROPAGE
> 
> Can you provide a link?
> 

Please see https://lore.kernel.org/linux-mm/20250306044239.3874247-3-balbirs@nvidia.com/T/

> Even if this is the only call site, there is no guarantee that
> there will be none in the future. I am not sure why we want caller
> to handle this special case. Who is going to tell the next user
> of RMP_USE_SHARED_ZEROPAGE or caller to try_to_map_unused_to_zeropage()
> that device private is incompatible with them?
> 

I don't disagree, but the question was why are device private pages even making
it to try_to_map_unused_to_zeropage()>>
>>>>  	remap_page(folio, 1 << order, remap_flags);
>>>>
>>>>  	/*
>>>> -- 
>>>> 2.50.1
>>>
>>>
>>
>> Thanks for the review
>> Balbir

Thanks,
Balbir