Re: [RFC 01/11] mm/zone_device: support large zone device private folios

[Balbir Singh <balbirs@nvidia.com>]

On 7/8/25 23:37, David Hildenbrand wrote:
> On 06.03.25 05:42, 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.
>>
>> Signed-off-by: Balbir Singh <balbirs@nvidia.com>
>> ---
>>   include/linux/memremap.h | 22 +++++++++++++++++-
>>   mm/memremap.c            | 50 +++++++++++++++++++++++++++++-----------
>>   2 files changed, 58 insertions(+), 14 deletions(-)
>>
>> diff --git a/include/linux/memremap.h b/include/linux/memremap.h
>> index 4aa151914eab..11d586dd8ef1 100644
>> --- a/include/linux/memremap.h
>> +++ b/include/linux/memremap.h
>> @@ -169,6 +169,18 @@ static inline bool folio_is_device_private(const struct folio *folio)
>>       return is_device_private_page(&folio->page);
>>   }
>>   +static inline void *folio_zone_device_data(const struct folio *folio)
>> +{
>> +    VM_BUG_ON_FOLIO(!folio_is_device_private(folio), folio);
>> +    return folio->page.zone_device_data;
>> +}
> 
> Not used.
> 
>> +
>> +static inline void folio_set_zone_device_data(struct folio *folio, void *data)
>> +{
>> +    VM_BUG_ON_FOLIO(!folio_is_device_private(folio), folio);
>> +    folio->page.zone_device_data = data;
>> +}
>> +
> 
> Not used.
> 
> Move both into the patch where they are actually used.
> 

Ack

>>   static inline bool is_pci_p2pdma_page(const struct page *page)
>>   {
>>       return IS_ENABLED(CONFIG_PCI_P2PDMA) &&
>> @@ -199,7 +211,7 @@ static inline bool folio_is_fsdax(const struct folio *folio)
>>   }
>>     #ifdef CONFIG_ZONE_DEVICE
>> -void zone_device_page_init(struct page *page);
>> +void init_zone_device_folio(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);
>> @@ -209,6 +221,14 @@ 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);
>> +
>> +    init_zone_device_folio(folio, 0);
>> +}
>> +
>>   #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 2aebc1b192da..7d98d0a4c0cd 100644
>> --- a/mm/memremap.c
>> +++ b/mm/memremap.c
>> @@ -459,20 +459,21 @@ EXPORT_SYMBOL_GPL(get_dev_pagemap);
>>   void free_zone_device_folio(struct folio *folio)
>>   {
>>       struct dev_pagemap *pgmap = folio->pgmap;
>> +    unsigned int nr = folio_nr_pages(folio);
>> +    int i;
>> +    bool anon = folio_test_anon(folio);
> 
> You can easily get rid of this (see below).
> 
>> +    struct page *page = folio_page(folio, 0);
> 
> Please inline folio_page(folio, 0) below instead.

Sure, is that preferred to taking a struct page ref?

> 
>>         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));
>> +    WARN_ON_ONCE(folio_test_large(folio) && !anon);
>> +
>> +    for (i = 0; i < nr; i++) {
>> +        if (anon)
>> +            __ClearPageAnonExclusive(folio_page(folio, i));
>>       }
> 
> if (folio_test_anon(folio)) {
>     for (i = 0; i < nr; i++)
>         __ClearPageAnonExclusive(folio_page(folio, i));
> } else {
>     VM_WARN_ON_ONCE(folio_test_large(folio));
> }
> 

Ack

>>         /*
>> @@ -496,10 +497,19 @@ void free_zone_device_folio(struct folio *folio)
>>         switch (pgmap->type) {
>>       case MEMORY_DEVICE_PRIVATE:
>> +        if (folio_test_large(folio)) {
>> +            folio_unqueue_deferred_split(folio);
> 
> Is deferred splitting even a thing for device-private?
> 
> Should we ever queue them for deferred splitting?
> 

Not really, but wanted to do the right thing in the tear down path, I can remove these bits

>> +
>> +            percpu_ref_put_many(&folio->pgmap->ref, nr - 1);
> 
> Looks like we instead want a helper put_dev_pagemap_refs(pgmap, nr) below instead
> 
>> +        }
>> +        pgmap->ops->page_free(page);
>> +        put_dev_pagemap(pgmap);
>> +        page->mapping = NULL;
>> +        break;
>>       case MEMORY_DEVICE_COHERENT:
>>           if (WARN_ON_ONCE(!pgmap->ops || !pgmap->ops->page_free))
>>               break;
>> -        pgmap->ops->page_free(folio_page(folio, 0));
>> +        pgmap->ops->page_free(page);
>>           put_dev_pagemap(pgmap);
>>           break;
>>   @@ -523,14 +533,28 @@ void free_zone_device_folio(struct folio *folio)
>>       }
>>   }
>>   -void zone_device_page_init(struct page *page)
>> +void init_zone_device_folio(struct folio *folio, unsigned int order)
>>   {
>> +    struct page *page = folio_page(folio, 0);
>> +
>> +    VM_BUG_ON(order > MAX_ORDER_NR_PAGES);
> 
> VM_WARN_ON_ONCE() or anything else that is not *BUG, please.
> 

Ack

>> +
>> +    WARN_ON_ONCE(order && order != HPAGE_PMD_ORDER);
> 
> Why do we need that limitation?
> 

mTHP is not yet supported in the series. We could keep this routine more generic
and not need the checks, but I added them to prevent unsupported order usage

>> +
>>       /*
>>        * 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);
>> +
>> +    /*
>> +     * Only PMD level migration is supported for THP migration
>> +     */
> 
> I don't understand how that comment interacts with the code below. This is basic large folio initialization.
> 
> Drop the comment, or move it above the HPAGE_PMD_ORDER check with a clear reason why that limitation excists.
>

Ack

 
>> +    if (order > 1) {
>> +        prep_compound_page(page, order);
>> +        folio_set_large_rmappable(folio);
>> +    }
>>   }
>> -EXPORT_SYMBOL_GPL(zone_device_page_init);
>> +EXPORT_SYMBOL_GPL(init_zone_device_folio);
> 
> 

Thanks for the review
Balbir