Re: [PATCH 42/46] mm/huge_memory: split huge PUDs when adjusting VMA boundaries

[David Hildenbrand (Arm)]

On 10/29/26 12:45, Kiryl Shutsemau wrote:
> On Tue, Oct 27, 2026 at 12:18:02AM +0100, David Hildenbrand (Arm) wrote:
>> PFN mappings such as VFIO PCI BAR mappings can use PUD leaves while
>> allowing VMA splits at page granularity. However, vma_adjust_trans_huge()
>> only handles PMDs.
>>
>> After faulting a writable PUD-sized VFIO BAR mapping, changing its first
>> page to read-only with mprotect() makes the PMD walk treat the PUD leaf
>> as a PMD table. Applying MADV_DONTFORK to the first PMD-sized chunk leaves
>> the PUD spanning two VMAs: the PMD-aligned boundary skips splitting.
>>
>> Walk the page tables once at each adjusted boundary and split a huge
>> PUD or PMD leaf as needed. Handle PUD leaves even when the boundary is
>> PMD-aligned.
>>
>> Fixes: f9e54c3a2f5b7 ("vfio/pci: implement huge_fault support")
>> Cc: stable@xxxxxxxxxxxxxxx
>> Signed-off-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
>> ---
>>  mm/huge_memory.c | 70 ++++++++++++++++++++++++++++++++++++++++++++++++--------
>>  1 file changed, 60 insertions(+), 10 deletions(-)
>>
>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>> index 08e58cda7d37d..d090642ac8c2b 100644
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -3503,16 +3503,66 @@ void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
>>  	__split_huge_pmd(vma, pmd, address, freeze);
>>  }
>>  
>> -static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
>> +static bool pud_split_required(const struct vm_area_struct *vma,
>> +		unsigned long address)
>>  {
>> +	if (!IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD))
>> +		return false;
>> +	if (IS_ALIGNED(address, HPAGE_PUD_SIZE))
>> +		return false;
>>  	/*
>> -	 * If the new address isn't hpage aligned and it could previously
>> -	 * contain an hugepage: check if we need to split an huge pmd.
>> +	 * Only selected VMAs support mapping PUDs for now. Lift the
>> +	 * restriction once this changes.
>>  	 */
>> -	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
>> -	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
>> -			 ALIGN(address, HPAGE_PMD_SIZE)))
>> -		split_huge_pmd_address(vma, address, false);
>> +	if (!vma_is_dax(vma) && !vma_test_any(vma, VMA_PFNMAP_BIT, VMA_MIXEDMAP_BIT))
>> +		return false;
>> +	return range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PUD_SIZE),
>> +			    ALIGN(address, HPAGE_PUD_SIZE));
>> +}
>> +
>> +static bool pmd_split_required(const struct vm_area_struct *vma,
>> +		unsigned long address)
>> +{
>> +	if (IS_ALIGNED(address, HPAGE_PMD_SIZE))
>> +		return false;
>> +
>> +	return range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
>> +			    ALIGN(address, HPAGE_PMD_SIZE));
>> +}
>> +
>> +static void split_huge_leaf(struct vm_area_struct *vma,
>> +		unsigned long address)
>> +{
>> +	const bool split_pud = pud_split_required(vma, address);
>> +	const bool split_pmd = pmd_split_required(vma, address);
>> +	pgd_t *pgd;
>> +	p4d_t *p4d;
>> +	pud_t *pud;
>> +	pmd_t *pmd;
>> +
>> +	/*
>> +	 * mmap, VMA and rmap write locks protect the walk; the split helpers
>> +	 * take the page table locks.
>> +	 */
>> +	if (!split_pud && !split_pmd)
>> +		return;
>> +
>> +	pgd = pgd_offset(vma->vm_mm, address);
>> +	if (!pgd_present(*pgd))
>> +		return;
>> +
>> +	p4d = p4d_offset(pgd, address);
>> +	if (!p4d_present(*p4d))
>> +		return;
>> +
>> +	pud = pud_offset(p4d, address);
>> +	if (split_pud && pud_leaf(*pud))
>> +		__split_huge_pud(vma, pud, address);
>> +	if (!split_pmd || !pud_present(*pud))
>> +		return;
> 
> This looks wrong.
> 
> We should bail out here if PUD is was a leaf and wasn't split.

__split_huge_pud() must split a leave. Either by zapping it (what we do right
now) or by populating page tables (as in Usama's PoC).

We can throw in here a

	if (WARN_ON_ONCE(!pud_leaf(pud)))
		return;

But I don't see how that would be required.

If we run into an un-splittable PUD something else is completely messed up.

(this is not hugetlb code :) )

-- 
Cheers,

David