Re: [v1 resend 08/12] mm/thp: add split during migration support

[Zi Yan <ziy@nvidia.com>]

On 17 Jul 2025, at 23:33, Matthew Brost wrote:

> On Thu, Jul 17, 2025 at 09:25:02PM -0400, Zi Yan wrote:
>> On 17 Jul 2025, at 20:41, Matthew Brost wrote:
>>
>>> On Thu, Jul 17, 2025 at 07:04:48PM -0400, Zi Yan wrote:
>>>> On 17 Jul 2025, at 18:24, Matthew Brost wrote:
>>>>
>>>>> On Thu, Jul 17, 2025 at 07:53:40AM +1000, Balbir Singh wrote:
>>>>>> On 7/17/25 02:24, Matthew Brost wrote:
>>>>>>> On Wed, Jul 16, 2025 at 07:19:10AM -0400, Zi Yan wrote:
>>>>>>>> On 16 Jul 2025, at 1:34, Matthew Brost wrote:
>>>>>>>>
>>>>>>>>> On Sun, Jul 06, 2025 at 11:47:10AM +1000, Balbir Singh wrote:
>>>>>>>>>> On 7/6/25 11:34, Zi Yan wrote:
>>>>>>>>>>> On 5 Jul 2025, at 21:15, Balbir Singh wrote:
>>>>>>>>>>>
>>>>>>>>>>>> On 7/5/25 11:55, Zi Yan wrote:
>>>>>>>>>>>>> On 4 Jul 2025, at 20:58, Balbir Singh wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>>> On 7/4/25 21:24, Zi Yan wrote:
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> s/pages/folio
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Thanks, will make the changes
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Why name it isolated if the folio is unmapped? Isolated folios often mean
>>>>>>>>>>>>>>> they are removed from LRU lists. isolated here causes confusion.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Ack, will change the name
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>   *
>>>>>>>>>>>>>>>>   * It calls __split_unmapped_folio() to perform uniform and non-uniform split.
>>>>>>>>>>>>>>>>   * It is in charge of checking whether the split is supported or not and
>>>>>>>>>>>>>>>> @@ -3800,7 +3799,7 @@ bool uniform_split_supported(struct folio *folio, unsigned int new_order,
>>>>>>>>>>>>>>>>   */
>>>>>>>>>>>>>>>>  static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>>>>>>>>>>>>>  		struct page *split_at, struct page *lock_at,
>>>>>>>>>>>>>>>> -		struct list_head *list, bool uniform_split)
>>>>>>>>>>>>>>>> +		struct list_head *list, bool uniform_split, bool isolated)
>>>>>>>>>>>>>>>>  {
>>>>>>>>>>>>>>>>  	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
>>>>>>>>>>>>>>>>  	XA_STATE(xas, &folio->mapping->i_pages, folio->index);
>>>>>>>>>>>>>>>> @@ -3846,14 +3845,16 @@ static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>>>>>>>>>>>>>  		 * is taken to serialise against parallel split or collapse
>>>>>>>>>>>>>>>>  		 * operations.
>>>>>>>>>>>>>>>>  		 */
>>>>>>>>>>>>>>>> -		anon_vma = folio_get_anon_vma(folio);
>>>>>>>>>>>>>>>> -		if (!anon_vma) {
>>>>>>>>>>>>>>>> -			ret = -EBUSY;
>>>>>>>>>>>>>>>> -			goto out;
>>>>>>>>>>>>>>>> +		if (!isolated) {
>>>>>>>>>>>>>>>> +			anon_vma = folio_get_anon_vma(folio);
>>>>>>>>>>>>>>>> +			if (!anon_vma) {
>>>>>>>>>>>>>>>> +				ret = -EBUSY;
>>>>>>>>>>>>>>>> +				goto out;
>>>>>>>>>>>>>>>> +			}
>>>>>>>>>>>>>>>> +			anon_vma_lock_write(anon_vma);
>>>>>>>>>>>>>>>>  		}
>>>>>>>>>>>>>>>>  		end = -1;
>>>>>>>>>>>>>>>>  		mapping = NULL;
>>>>>>>>>>>>>>>> -		anon_vma_lock_write(anon_vma);
>>>>>>>>>>>>>>>>  	} else {
>>>>>>>>>>>>>>>>  		unsigned int min_order;
>>>>>>>>>>>>>>>>  		gfp_t gfp;
>>>>>>>>>>>>>>>> @@ -3920,7 +3921,8 @@ static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>>>>>>>>>>>>>  		goto out_unlock;
>>>>>>>>>>>>>>>>  	}
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> -	unmap_folio(folio);
>>>>>>>>>>>>>>>> +	if (!isolated)
>>>>>>>>>>>>>>>> +		unmap_folio(folio);
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>  	/* block interrupt reentry in xa_lock and spinlock */
>>>>>>>>>>>>>>>>  	local_irq_disable();
>>>>>>>>>>>>>>>> @@ -3973,14 +3975,15 @@ static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>  		ret = __split_unmapped_folio(folio, new_order,
>>>>>>>>>>>>>>>>  				split_at, lock_at, list, end, &xas, mapping,
>>>>>>>>>>>>>>>> -				uniform_split);
>>>>>>>>>>>>>>>> +				uniform_split, isolated);
>>>>>>>>>>>>>>>>  	} else {
>>>>>>>>>>>>>>>>  		spin_unlock(&ds_queue->split_queue_lock);
>>>>>>>>>>>>>>>>  fail:
>>>>>>>>>>>>>>>>  		if (mapping)
>>>>>>>>>>>>>>>>  			xas_unlock(&xas);
>>>>>>>>>>>>>>>>  		local_irq_enable();
>>>>>>>>>>>>>>>> -		remap_page(folio, folio_nr_pages(folio), 0);
>>>>>>>>>>>>>>>> +		if (!isolated)
>>>>>>>>>>>>>>>> +			remap_page(folio, folio_nr_pages(folio), 0);
>>>>>>>>>>>>>>>>  		ret = -EAGAIN;
>>>>>>>>>>>>>>>>  	}
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> These "isolated" special handlings does not look good, I wonder if there
>>>>>>>>>>>>>>> is a way of letting split code handle device private folios more gracefully.
>>>>>>>>>>>>>>> It also causes confusions, since why does "isolated/unmapped" folios
>>>>>>>>>>>>>>> not need to unmap_page(), remap_page(), or unlock?
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> There are two reasons for going down the current code path
>>>>>>>>>>>>>
>>>>>>>>>>>>> After thinking more, I think adding isolated/unmapped is not the right
>>>>>>>>>>>>> way, since unmapped folio is a very generic concept. If you add it,
>>>>>>>>>>>>> one can easily misuse the folio split code by first unmapping a folio
>>>>>>>>>>>>> and trying to split it with unmapped = true. I do not think that is
>>>>>>>>>>>>> supported and your patch does not prevent that from happening in the future.
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> I don't understand the misuse case you mention, I assume you mean someone can
>>>>>>>>>>>> get the usage wrong? The responsibility is on the caller to do the right thing
>>>>>>>>>>>> if calling the API with unmapped
>>>>>>>>>>>
>>>>>>>>>>> Before your patch, there is no use case of splitting unmapped folios.
>>>>>>>>>>> Your patch only adds support for device private page split, not any unmapped
>>>>>>>>>>> folio split. So using a generic isolated/unmapped parameter is not OK.
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> There is a use for splitting unmapped folios (see below)
>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>> You should teach different parts of folio split code path to handle
>>>>>>>>>>>>> device private folios properly. Details are below.
>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> 1. if the isolated check is not present, folio_get_anon_vma will fail and cause
>>>>>>>>>>>>>>    the split routine to return with -EBUSY
>>>>>>>>>>>>>
>>>>>>>>>>>>> You do something below instead.
>>>>>>>>>>>>>
>>>>>>>>>>>>> if (!anon_vma && !folio_is_device_private(folio)) {
>>>>>>>>>>>>> 	ret = -EBUSY;
>>>>>>>>>>>>> 	goto out;
>>>>>>>>>>>>> } else if (anon_vma) {
>>>>>>>>>>>>> 	anon_vma_lock_write(anon_vma);
>>>>>>>>>>>>> }
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> folio_get_anon() cannot be called for unmapped folios. In our case the page has
>>>>>>>>>>>> already been unmapped. Is there a reason why you mix anon_vma_lock_write with
>>>>>>>>>>>> the check for device private folios?
>>>>>>>>>>>
>>>>>>>>>>> Oh, I did not notice that anon_vma = folio_get_anon_vma(folio) is also
>>>>>>>>>>> in if (!isolated) branch. In that case, just do
>>>>>>>>>>>
>>>>>>>>>>> if (folio_is_device_private(folio) {
>>>>>>>>>>> ...
>>>>>>>>>>> } else if (is_anon) {
>>>>>>>>>>> ...
>>>>>>>>>>> } else {
>>>>>>>>>>> ...
>>>>>>>>>>> }
>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>> People can know device private folio split needs a special handling.
>>>>>>>>>>>>>
>>>>>>>>>>>>> BTW, why a device private folio can also be anonymous? Does it mean
>>>>>>>>>>>>> if a page cache folio is migrated to device private, kernel also
>>>>>>>>>>>>> sees it as both device private and file-backed?
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> FYI: device private folios only work with anonymous private pages, hence
>>>>>>>>>>>> the name device private.
>>>>>>>>>>>
>>>>>>>>>>> OK.
>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>> 2. Going through unmap_page(), remap_page() causes a full page table walk, which
>>>>>>>>>>>>>>    the migrate_device API has already just done as a part of the migration. The
>>>>>>>>>>>>>>    entries under consideration are already migration entries in this case.
>>>>>>>>>>>>>>    This is wasteful and in some case unexpected.
>>>>>>>>>>>>>
>>>>>>>>>>>>> unmap_folio() already adds TTU_SPLIT_HUGE_PMD to try to split
>>>>>>>>>>>>> PMD mapping, which you did in migrate_vma_split_pages(). You probably
>>>>>>>>>>>>> can teach either try_to_migrate() or try_to_unmap() to just split
>>>>>>>>>>>>> device private PMD mapping. Or if that is not preferred,
>>>>>>>>>>>>> you can simply call split_huge_pmd_address() when unmap_folio()
>>>>>>>>>>>>> sees a device private folio.
>>>>>>>>>>>>>
>>>>>>>>>>>>> For remap_page(), you can simply return for device private folios
>>>>>>>>>>>>> like it is currently doing for non anonymous folios.
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> Doing a full rmap walk does not make sense with unmap_folio() and
>>>>>>>>>>>> remap_folio(), because
>>>>>>>>>>>>
>>>>>>>>>>>> 1. We need to do a page table walk/rmap walk again
>>>>>>>>>>>> 2. We'll need special handling of migration <-> migration entries
>>>>>>>>>>>>    in the rmap handling (set/remove migration ptes)
>>>>>>>>>>>> 3. In this context, the code is already in the middle of migration,
>>>>>>>>>>>>    so trying to do that again does not make sense.
>>>>>>>>>>>
>>>>>>>>>>> Why doing split in the middle of migration? Existing split code
>>>>>>>>>>> assumes to-be-split folios are mapped.
>>>>>>>>>>>
>>>>>>>>>>> What prevents doing split before migration?
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> The code does do a split prior to migration if THP selection fails
>>>>>>>>>>
>>>>>>>>>> Please see https://lore.kernel.org/lkml/20250703233511.2028395-5-balbirs@nvidia.com/
>>>>>>>>>> and the fallback part which calls split_folio()
>>>>>>>>>>
>>>>>>>>>> But the case under consideration is special since the device needs to allocate
>>>>>>>>>> corresponding pfn's as well. The changelog mentions it:
>>>>>>>>>>
>>>>>>>>>> "The common case that arises is that after setup, during migrate
>>>>>>>>>> the destination might not be able to allocate MIGRATE_PFN_COMPOUND
>>>>>>>>>> pages."
>>>>>>>>>>
>>>>>>>>>> I can expand on it, because migrate_vma() is a multi-phase operation
>>>>>>>>>>
>>>>>>>>>> 1. migrate_vma_setup()
>>>>>>>>>> 2. migrate_vma_pages()
>>>>>>>>>> 3. migrate_vma_finalize()
>>>>>>>>>>
>>>>>>>>>> It can so happen that when we get the destination pfn's allocated the destination
>>>>>>>>>> might not be able to allocate a large page, so we do the split in migrate_vma_pages().
>>>>>>>>>>
>>>>>>>>>> The pages have been unmapped and collected in migrate_vma_setup()
>>>>>>>>>>
>>>>>>>>>> The next patch in the series 9/12 (https://lore.kernel.org/lkml/20250703233511.2028395-10-balbirs@nvidia.com/)
>>>>>>>>>> tests the split and emulates a failure on the device side to allocate large pages
>>>>>>>>>> and tests it in 10/12 (https://lore.kernel.org/lkml/20250703233511.2028395-11-balbirs@nvidia.com/)
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Another use case I’ve seen is when a previously allocated high-order
>>>>>>>>> folio, now in the free memory pool, is reallocated as a lower-order
>>>>>>>>> page. For example, a 2MB fault allocates a folio, the memory is later
>>>>>>>>
>>>>>>>> That is different. If the high-order folio is free, it should be split
>>>>>>>> using split_page() from mm/page_alloc.c.
>>>>>>>>
>>>>>>>
>>>>>>> Ah, ok. Let me see if that works - it would easier.
>>>>>>>
>>>>>
>>>>> This suggestion quickly blows up as PageCompound is true and page_count
>>>>> here is zero.
>>>>
>>>> OK, your folio has PageCompound set. Then you will need __split_unmapped_foio().
>>>>
>>>>>
>>>>>>>>> freed, and then a 4KB fault reuses a page from that previously allocated
>>>>>>>>> folio. This will be actually quite common in Xe / GPU SVM. In such
>>>>>>>>> cases, the folio in an unmapped state needs to be split. I’d suggest a
>>>>>>>>
>>>>>>>> This folio is unused, so ->flags, ->mapping, and etc. are not set,
>>>>>>>> __split_unmapped_folio() is not for it, unless you mean free folio
>>>>>>>> differently.
>>>>>>>>
>>>>>>>
>>>>>>> This is right, those fields should be clear.
>>>>>>>
>>>>>>> Thanks for the tip.
>>>>>>>
>>>>>> I was hoping to reuse __split_folio_to_order() at some point in the future
>>>>>> to split the backing pages in the driver, but it is not an immediate priority
>>>>>>
>>>>>
>>>>> I think we need something for the scenario I describe here. I was to
>>>>> make __split_huge_page_to_list_to_order with a couple of hacks but it
>>>>> almostly certainig not right as Zi pointed out.
>>>>>
>>>>> New to the MM stuff, but play around with this a bit and see if I can
>>>>> come up with something that will work here.
>>>>
>>>> Can you try to write a new split_page function with __split_unmapped_folio()?
>>>> Since based on your description, your folio is not mapped.
>>>>
>>>
>>> Yes, page->mapping is NULL in this case - that was part of the hacks to
>>> __split_huge_page_to_list_to_order (more specially __folio_split) I had
>>> to make in order to get something working for this case.
>>>
>>> I can try out something based on __split_unmapped_folio and report back.
>>
>> mm-new tree has an updated __split_unmapped_folio() version, it moves
>> all unmap irrelevant code out of __split_unmaped_folio(). You might find
>> it easier to reuse.
>>
>> See: https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/tree/mm/huge_memory.c?h=mm-new#n3430
>>
>
> Will take a look. It is possible some of the issues we are hitting are
> due to working on drm-tip + pulling in core MM patches in this series on
> top of that branch then missing some other patches in mm-new. I'll see
> if ww can figure out a work flow to have the latest and greatest from
> both drm-tip and the MM branches.
>
> Will these changes be in 6.17?

Hopefully yes. mm patches usually go from mm-new to mm-unstable
to mm-stable to mainline. If not, we will figure it out. :)

>
>> I am about to update the code with v4 patches. I will cc you, so that
>> you can get the updated __split_unmaped_folio().
>>
>> Feel free to ask questions on folio split code.
>>
>
> Thanks.
>
> Matt
>
>> Best Regards,
>> Yan, Zi


Best Regards,
Yan, Zi