Re: [PATCH] fixup: mm/huge_memory.c: introduce folio_split_unmapped

[Balbir Singh <balbirs@nvidia.com>]

On 11/20/25 20:32, David Hildenbrand (Red Hat) wrote:
> On 11/20/25 10:25, Balbir Singh wrote:
>> On 11/20/25 20:09, David Hildenbrand (Red Hat) wrote:
>>> On 11/20/25 04:07, Balbir Singh wrote:
>>>> Code refactoring of __folio_split() via helper
>>>> __folio_freeze_and_split_unmapped() caused a regression with clang-20
>>>> with CONFIG_SHMEM=n, the compiler was not able to optimize away the
>>>> call to shmem_uncharge() due to changes in nr_shmem_dropped.
>>>> Fix this by checking for shmem_mapping() prior to calling
>>>> shmem_uncharge(), shmem_mapping() returns false when CONFIG_SHMEM=n.
>>>>
>>>> smatch also complained about parameter end being used without
>>>> initialization, which is a false positive, but keep the tool happy
>>>> by sending in initialized parameters. end is initialized to 0.
>>>>
>>>> Add detailed documentation comments for folio_split_unmapped()
>>>>
>>>> Cc: Andrew Morton <akpm@linux-foundation.org>
>>>> 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>
>>>>
>>>> Signed-off-by: Balbir Singh <balbirs@nvidia.com>
>>>> ---
>>>>    mm/huge_memory.c | 32 ++++++++++++++++++++++----------
>>>>    1 file changed, 22 insertions(+), 10 deletions(-)
>>>>
>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>>> index 78a31a476ad3..c4267a0f74df 100644
>>>> --- a/mm/huge_memory.c
>>>> +++ b/mm/huge_memory.c
>>>> @@ -3751,6 +3751,7 @@ static int __folio_freeze_and_split_unmapped(struct folio *folio, unsigned int n
>>>>        int ret = 0;
>>>>        struct deferred_split *ds_queue;
>>>>    +    VM_WARN_ON_ONCE(!mapping && end != 0);
>>>
>>> You could drop the "!= 0"
>>
>> Ack
>>
>> VM_WARN_ONE(!mapping && end);
>>
>>>
>>>>        /* Prevent deferred_split_scan() touching ->_refcount */
>>>>        ds_queue = folio_split_queue_lock(folio);
>>>>        if (folio_ref_freeze(folio, 1 + extra_pins)) {
>>>> @@ -3919,7 +3920,7 @@ static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>        int nr_shmem_dropped = 0;
>>>>        int remap_flags = 0;
>>>>        int extra_pins, ret;
>>>> -    pgoff_t end;
>>>> +    pgoff_t end = 0;
>>>>        bool is_hzp;
>>>>          VM_WARN_ON_ONCE_FOLIO(!folio_test_locked(folio), folio);
>>>> @@ -4049,7 +4050,7 @@ static int __folio_split(struct folio *folio, unsigned int new_order,
>>>>          local_irq_enable();
>>>>    -    if (nr_shmem_dropped)
>>>> +    if (mapping && shmem_mapping(mapping) && nr_shmem_dropped)
>>>>            shmem_uncharge(mapping->host, nr_shmem_dropped);
>>>
>>> That looks questionable. We shouldn't add runtime check to handle buildtime things.
>>>
>>> Likely what you want is instead
>>>
>>> if (IS_ENABLED(CONFIG_SHMEM) && nr_shmem_dropped)
>>>      shmem_uncharge()
>>>
>>
>> shmem_mapping() returns false for CONFIG_SHMEM=n and shmem_mapping() checks that the mapping
>> is indeed for shmem ops before uncharging. Happy to change it if you like,
>> your version is more readable
> Good point, but the questionable thing is that it looks like nr_shmem_dropped
> could be set for non-shmem mappings, when it's really just a compiler thing.
> 
> What about handling it through a proper stub so we can keep this calling code simple?
> 
> diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
> index 5b368f9549d67..e38cb01031200 100644
> --- a/include/linux/shmem_fs.h
> +++ b/include/linux/shmem_fs.h
> @@ -136,11 +136,15 @@ static inline bool shmem_hpage_pmd_enabled(void)
>  
>  #ifdef CONFIG_SHMEM
>  extern unsigned long shmem_swap_usage(struct vm_area_struct *vma);
> +extern void shmem_uncharge(struct inode *inode, long pages);
>  #else
>  static inline unsigned long shmem_swap_usage(struct vm_area_struct *vma)
>  {
>         return 0;
>  }
> +static inline void shmem_uncharge(struct inode *inode, long pages)
> +{
> +}
>  #endif
>  extern unsigned long shmem_partial_swap_usage(struct address_space *mapping,
>                                                 pgoff_t start, pgoff_t end);
> @@ -194,7 +198,6 @@ static inline pgoff_t shmem_fallocend(struct inode *inode, pgoff_t eof)
>  }
>  
>  extern bool shmem_charge(struct inode *inode, long pages);
> -extern void shmem_uncharge(struct inode *inode, long pages);
>  
>  #ifdef CONFIG_USERFAULTFD
>  #ifdef CONFIG_SHMEM
> 
> 

Agreed, I would like to let this patch proceed and then immediately follow up patch
along the lines of CONFIG_SHMEM as separate independent patch (independent of this
series). What do you think?

Balbir
changes.