Message-ID: <7E859D31-9C00-4B08-84C8-2EC34D2B06FE@nvidia.com>
Date: Thu, 25 Sep 2025 11:27:38 -0400
From: Zi Yan <ziy@...dia.com>
To: Alistair Popple <apopple@...dia.com>
Cc: David Hildenbrand <david@...hat.com>, Balbir Singh <balbirs@...dia.com>,
 linux-kernel@...r.kernel.org, linux-mm@...ck.org, damon@...ts.linux.dev,
 dri-devel@...ts.freedesktop.org, Joshua Hahn <joshua.hahnjy@...il.com>,
 Rakie Kim <rakie.kim@...com>, Byungchul Park <byungchul@...com>,
 Gregory Price <gourry@...rry.net>, Ying Huang <ying.huang@...ux.alibaba.com>,
 Oscar Salvador <osalvador@...e.de>,
 Lorenzo Stoakes <lorenzo.stoakes@...cle.com>,
 Baolin Wang <baolin.wang@...ux.alibaba.com>,
 "Liam R. Howlett" <Liam.Howlett@...cle.com>, Nico Pache <npache@...hat.com>,
 Ryan Roberts <ryan.roberts@....com>, Dev Jain <dev.jain@....com>,
 Barry Song <baohua@...nel.org>, Lyude Paul <lyude@...hat.com>,
 Danilo Krummrich <dakr@...nel.org>, David Airlie <airlied@...il.com>,
 Simona Vetter <simona@...ll.ch>, Ralph Campbell <rcampbell@...dia.com>,
 Mika Penttilä <mpenttil@...hat.com>,
 Matthew Brost <matthew.brost@...el.com>,
 Francois Dugast <francois.dugast@...el.com>
Subject: Re: [v6 01/15] mm/zone_device: support large zone device private
 folios

On 24 Sep 2025, at 19:45, Alistair Popple wrote:

> On 2025-09-25 at 03:49 +1000, Zi Yan <ziy@...dia.com> wrote...
>> On 24 Sep 2025, at 7:04, David Hildenbrand wrote:
>>
>>> On 23.09.25 05:47, Balbir Singh wrote:
>>>> On 9/19/25 23:26, Zi Yan wrote:
>>>>> On 19 Sep 2025, at 1:01, Balbir Singh wrote:
>>>>>
>>>>>> On 9/18/25 12:49, Zi Yan wrote:
>>>>>>> On 16 Sep 2025, at 8:21, Balbir Singh wrote:
>>>>>>>
>>>>>>>> Add routines to support allocation of large order zone device folios
>>>>>>>> and helper functions for zone device folios, to check if a folio is
>>>>>>>> device private and helpers for setting zone device data.
>>>>>>>>
>>>>>>>> When large folios are used, the existing page_free() callback in
>>>>>>>> pgmap is called when the folio is freed, this is true for both
>>>>>>>> PAGE_SIZE and higher order pages.
>>>>>>>>
>>>>>>>> Zone device private large folios do not support deferred split and
>>>>>>>> scan like normal THP folios.
>>>>>>>>
>>>>>>>> Signed-off-by: Balbir Singh <balbirs@...dia.com>
>>>>>>>> Cc: David Hildenbrand <david@...hat.com>
>>>>>>>> Cc: Zi Yan <ziy@...dia.com>
>>>>>>>> Cc: Joshua Hahn <joshua.hahnjy@...il.com>
>>>>>>>> Cc: Rakie Kim <rakie.kim@...com>
>>>>>>>> Cc: Byungchul Park <byungchul@...com>
>>>>>>>> Cc: Gregory Price <gourry@...rry.net>
>>>>>>>> Cc: Ying Huang <ying.huang@...ux.alibaba.com>
>>>>>>>> Cc: Alistair Popple <apopple@...dia.com>
>>>>>>>> Cc: Oscar Salvador <osalvador@...e.de>
>>>>>>>> Cc: Lorenzo Stoakes <lorenzo.stoakes@...cle.com>
>>>>>>>> Cc: Baolin Wang <baolin.wang@...ux.alibaba.com>
>>>>>>>> Cc: "Liam R. Howlett" <Liam.Howlett@...cle.com>
>>>>>>>> Cc: Nico Pache <npache@...hat.com>
>>>>>>>> Cc: Ryan Roberts <ryan.roberts@....com>
>>>>>>>> Cc: Dev Jain <dev.jain@....com>
>>>>>>>> Cc: Barry Song <baohua@...nel.org>
>>>>>>>> Cc: Lyude Paul <lyude@...hat.com>
>>>>>>>> Cc: Danilo Krummrich <dakr@...nel.org>
>>>>>>>> Cc: David Airlie <airlied@...il.com>
>>>>>>>> Cc: Simona Vetter <simona@...ll.ch>
>>>>>>>> Cc: Ralph Campbell <rcampbell@...dia.com>
>>>>>>>> Cc: Mika Penttilä <mpenttil@...hat.com>
>>>>>>>> Cc: Matthew Brost <matthew.brost@...el.com>
>>>>>>>> Cc: Francois Dugast <francois.dugast@...el.com>
>>>>>>>> ---
>>>>>>>>   include/linux/memremap.h | 10 +++++++++-
>>>>>>>>   mm/memremap.c            | 34 +++++++++++++++++++++-------------
>>>>>>>>   mm/rmap.c                |  6 +++++-
>>>>>>>>   3 files changed, 35 insertions(+), 15 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/include/linux/memremap.h b/include/linux/memremap.h
>>>>>>>> index e5951ba12a28..9c20327c2be5 100644
>>>>>>>> --- a/include/linux/memremap.h
>>>>>>>> +++ b/include/linux/memremap.h
>>>>>>>> @@ -206,7 +206,7 @@ static inline bool is_fsdax_page(const struct page *page)
>>>>>>>>   }
>>>>>>>>
>>>>>>>>   #ifdef CONFIG_ZONE_DEVICE
>>>>>>>> -void zone_device_page_init(struct page *page);
>>>>>>>> +void zone_device_folio_init(struct folio *folio, unsigned int order);
>>>>>>>>   void *memremap_pages(struct dev_pagemap *pgmap, int nid);
>>>>>>>>   void memunmap_pages(struct dev_pagemap *pgmap);
>>>>>>>>   void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap);
>>>>>>>> @@ -215,6 +215,14 @@ struct dev_pagemap *get_dev_pagemap(unsigned long pfn);
>>>>>>>>   bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn);
>>>>>>>>
>>>>>>>>   unsigned long memremap_compat_align(void);
>>>>>>>> +
>>>>>>>> +static inline void zone_device_page_init(struct page *page)
>>>>>>>> +{
>>>>>>>> +	struct folio *folio = page_folio(page);
>>>>>>>> +
>>>>>>>> +	zone_device_folio_init(folio, 0);
>>>>>>>
>>>>>>> I assume it is for legacy code, where only non-compound page exists?
>>>>>>>
>>>>>>> It seems that you assume @page is always order-0, but there is no check
>>>>>>> for it. Adding VM_WARN_ON_ONCE_FOLIO(folio_order(folio) != 0, folio)
>>>>>>> above it would be useful to detect misuse.
>>>>>>>
>>>>>>>> +}
>>>>>>>> +
>>>>>>>>   #else
>>>>>>>>   static inline void *devm_memremap_pages(struct device *dev,
>>>>>>>>   		struct dev_pagemap *pgmap)
>>>>>>>> diff --git a/mm/memremap.c b/mm/memremap.c
>>>>>>>> index 46cb1b0b6f72..a8481ebf94cc 100644
>>>>>>>> --- a/mm/memremap.c
>>>>>>>> +++ b/mm/memremap.c
>>>>>>>> @@ -416,20 +416,19 @@ EXPORT_SYMBOL_GPL(get_dev_pagemap);
>>>>>>>>   void free_zone_device_folio(struct folio *folio)
>>>>>>>>   {
>>>>>>>>   	struct dev_pagemap *pgmap = folio->pgmap;
>>>>>>>> +	unsigned long nr = folio_nr_pages(folio);
>>>>>>>> +	int i;
>>>>>>>>
>>>>>>>>   	if (WARN_ON_ONCE(!pgmap))
>>>>>>>>   		return;
>>>>>>>>
>>>>>>>>   	mem_cgroup_uncharge(folio);
>>>>>>>>
>>>>>>>> -	/*
>>>>>>>> -	 * Note: we don't expect anonymous compound pages yet. Once supported
>>>>>>>> -	 * and we could PTE-map them similar to THP, we'd have to clear
>>>>>>>> -	 * PG_anon_exclusive on all tail pages.
>>>>>>>> -	 */
>>>>>>>>   	if (folio_test_anon(folio)) {
>>>>>>>> -		VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
>>>>>>>> -		__ClearPageAnonExclusive(folio_page(folio, 0));
>>>>>>>> +		for (i = 0; i < nr; i++)
>>>>>>>> +			__ClearPageAnonExclusive(folio_page(folio, i));
>>>>>>>> +	} else {
>>>>>>>> +		VM_WARN_ON_ONCE(folio_test_large(folio));
>>>>>>>>   	}
>>>>>>>>
>>>>>>>>   	/*
>>>>>>>> @@ -456,8 +455,8 @@ void free_zone_device_folio(struct folio *folio)
>>>>>>>>   	case MEMORY_DEVICE_COHERENT:
>>>>>>>>   		if (WARN_ON_ONCE(!pgmap->ops || !pgmap->ops->page_free))
>>>>>>>>   			break;
>>>>>>>> -		pgmap->ops->page_free(folio_page(folio, 0));
>>>>>>>> -		put_dev_pagemap(pgmap);
>>>>>>>> +		pgmap->ops->page_free(&folio->page);
>>>>>>>> +		percpu_ref_put_many(&folio->pgmap->ref, nr);
>>>>>>>>   		break;
>>>>>>>>
>>>>>>>>   	case MEMORY_DEVICE_GENERIC:
>>>>>>>> @@ -480,14 +479,23 @@ void free_zone_device_folio(struct folio *folio)
>>>>>>>>   	}
>>>>>>>>   }
>>>>>>>>
>>>>>>>> -void zone_device_page_init(struct page *page)
>>>>>>>> +void zone_device_folio_init(struct folio *folio, unsigned int order)
>>>>>>>>   {
>>>>>>>> +	struct page *page = folio_page(folio, 0);
>>>>>>>
>>>>>>> It is strange to see a folio is converted back to page in
>>>>>>> a function called zone_device_folio_init().
>>>>>>>
>>>>>>>> +
>>>>>>>> +	VM_WARN_ON_ONCE(order > MAX_ORDER_NR_PAGES);
>>>>>>>> +
>>>>>>>>   	/*
>>>>>>>>   	 * Drivers shouldn't be allocating pages after calling
>>>>>>>>   	 * memunmap_pages().
>>>>>>>>   	 */
>>>>>>>> -	WARN_ON_ONCE(!percpu_ref_tryget_live(&page_pgmap(page)->ref));
>>>>>>>> -	set_page_count(page, 1);
>>>>>>>> +	WARN_ON_ONCE(!percpu_ref_tryget_many(&page_pgmap(page)->ref, 1 << order));
>>>>>>>> +	folio_set_count(folio, 1);
>>>>>>>>   	lock_page(page);
>>>>>>>> +
>>>>>>>> +	if (order > 1) {
>
> Why is this only called for order > 1 rather than order > 0 ?
>
>>>>>>>> +		prep_compound_page(page, order);
>>>>>>>> +		folio_set_large_rmappable(folio);
>>>>>>>> +	}
>>>>>>>
>>>>>>> OK, so basically, @folio is not a compound page yet when zone_device_folio_init()
>>>>>>> is called.
>>>>>>>
>>>>>>> I feel that your zone_device_page_init() and zone_device_folio_init()
>>>>>>> implementations are inverse. They should follow the same pattern
>>>>>>> as __alloc_pages_noprof() and __folio_alloc_noprof(), where
>>>>>>> zone_device_page_init() does the actual initialization and
>>>>>>> zone_device_folio_init() just convert a page to folio.
>>>>>>>
>>>>>>> Something like:
>>>>>>>
>>>>>>> void zone_device_page_init(struct page *page, unsigned int order)
>>>>>>> {
>>>>>>> 	VM_WARN_ON_ONCE(order > MAX_ORDER_NR_PAGES);
>>>>>>>
>>>>>>> 	/*
>>>>>>> 	 * Drivers shouldn't be allocating pages after calling
>>>>>>> 	 * memunmap_pages().
>>>>>>> 	 */
>>>>>>>
>>>>>>>      WARN_ON_ONCE(!percpu_ref_tryget_many(&page_pgmap(page)->ref, 1 << order));
>>>>>>> 	
>>>>>>> 	/*
>>>>>>> 	 * anonymous folio does not support order-1, high order file-backed folio
>>>>>>> 	 * is not supported at all.
>>>>>>> 	 */
>
> I guess that answers my question :-)
>
>>>>>>> 	VM_WARN_ON_ONCE(order == 1);
>>>>>>>
>>>>>>> 	if (order > 1)
>>>>>>> 		prep_compound_page(page, order);
>>>>>>>
>>>>>>> 	/* page has to be compound head here */
>>>>>>> 	set_page_count(page, 1);
>>>>>>> 	lock_page(page);
>>>>>>> }
>>>>>>>
>>>>>>> void zone_device_folio_init(struct folio *folio, unsigned int order)
>>>>>>> {
>>>>>>> 	struct page *page = folio_page(folio, 0);
>>>>>>>
>>>>>>> 	zone_device_page_init(page, order);
>>>>>>> 	page_rmappable_folio(page);
>>>>>>> }
>>>>>>>
>>>>>>> Or
>>>>>>>
>>>>>>> struct folio *zone_device_folio_init(struct page *page, unsigned int order)
>>>>>>> {
>>>>>>> 	zone_device_page_init(page, order);
>>>>>>> 	return page_rmappable_folio(page);
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>> Then, it comes to free_zone_device_folio() above,
>>>>>>> I feel that pgmap->ops->page_free() should take an additional order
>>>>>>> parameter to free a compound page like free_frozen_pages().
>>>>>>>
>>>>>>>
>>>>>>> This is my impression after reading the patch and zone device page code.
>>>>>>>
>>>>>>> Alistair and David can correct me if this is wrong, since I am new to
>>>>>>> zone device page code.
>>>>>>> 	
>>>>>>
>>>>>> Thanks, I did not want to change zone_device_page_init() for several
>>>>>> drivers (outside my test scope) that already assume it has an order size of 0.
>
> It's a trivial change, so I don't think avoiding changes to other drivers should
> be a concern.
>
>>>>>
>>>>> But my proposed zone_device_page_init() should still work for order-0
>>>>> pages. You just need to change call site to add 0 as a new parameter.
>>>>>
>>>>
>>>> I did not want to change existing callers (increases testing impact)
>>>> without a strong reason.
>>>>
>>>>>
>>>>> One strange thing I found in the original zone_device_page_init() is
>>>>> the use of page_pgmap() in
>>>>> WARN_ON_ONCE(!percpu_ref_tryget_many(&page_pgmap(page)->ref, 1 << order)).
>>>>> page_pgmap() calls page_folio() on the given page to access pgmap field.
>>>>> And pgmap field is only available in struct folio. The code initializes
>>>>> struct page, but in middle it suddenly finds the page is actually a folio,
>>>>> then treat it as a page afterwards. I wonder if it can be done better.
>>>>>
>>>>> This might be a question to Alistair, since he made the change.
>
> Hello! I might be him :)
>
> I think this situation is just historical - when I originally wrote
> zone_device_page_init() the pgmap was stored on the page rather than the folio.
> That only changed fairly recently with commit 82ba975e4c43 ("mm: allow compound
> zone device pages").
>
> The reason pgmap is now only available on the folio is described in the
> commit log. The TLDR is switching FS DAX to use compound pages required
> page->compound_head to be available for use, and that was being shared
> with page->pgmap. So the solution was to move pgmap to the folio freeing up
> page->compound_head for use on tail pages.
>
> The whole percpu pgmap->ref could actually now go away - I've debated removing
> it but haven't found the motivation as it provides a small advantage on driver
> tear down. Basically it just tracks how many pages are allocated in the pgmap
> so drivers could use that to determine if they need to trigger migrations before
> tearing down the pgmap.
>
> The alternative is just to loop over every page in the pgmap to ensure the
> folio/page refcounts are 0 before tear down.
>
>>>>>
>>>>
>>>> I'll let him answer it :)
>>>
>>> Not him, but I think this goes back to my question raised in my other reply: When would we allocate "struct folio" in the future.
>>>
>>> If it's "always" then actually most of the zone-device code would only ever operate on folios and never on pages in the future.
>>>
>>> I recall during a discussion at LSF/MM I raised that, and the answer was (IIRC) that we will allocate "struct folio" as we will initialize the memmap for dax.
>
> Sounds about right.
>
>>> So essentially, we'd always have folios and would never really have to operate on pages.
>
> Yeah, I think I mentioned to Matthew at LSF/MM that I thought ZONE_DEVICE (and
> in particular ZONE_DEVICE_PRIVATE) might be a good candidate to experiment with
> removing struct pages entirely and switching to memdesc's or whatever. Because
> we should, in theory at least, only need to operate on folio's. But I'm still a
> little vague on the details how that would actually work. It's been on my TODO
> list for a while, so myabe I will try and look at it for LPC as a healthy bit of
> conference driven development.
>
>> Hmm, then what is the point of having “struct folio”, which originally is
>> added to save compound_head() calls, where everything is a folio in device
>> private world? We might need DAX people to explain the rationale of
>> “always struct folio”.
>
> Longer term isn't there an aim to remove struct page? So I assumed moving to

Right. But my current impression based on my code reading and this patchset
is that every device private page is a folio. To form a high order folio,
each device private folio is converted to page, prep_compound*()’d, then
converted back to folio. Based on what you said above, this weird conversion
might be temporary until the code is switched to memdesc.

I am looking forward to more details on how device private will be switched
to memdesc from you. :)

> folio's was part of that effort. As you say though many of the clean-ups thus
> far related to switching ZONE_DEVICE to folios have indeed just been about
> removing compound_head() calls.



Best Regards,
Yan, Zi

Powered by blists - more mailing lists