Re: [PATCH 3/3] mm/huge_memory: only compare PFNs of huge entries on huge fault insert

From: David Hildenbrand (Arm)

Date: Mon Nov 16 2026 - 15:09:25 EST


On 11/16/26 17:55, Lorenzo Stoakes (ARM) wrote:
> On Mon, Nov 16, 2026 at 05:01:52PM +0100, David Hildenbrand (Arm) wrote:
>> On 11/15/26 21:58, Lorenzo Stoakes (ARM) wrote:
>>> A driver that implements both ->fault and ->huge_fault must anticipate
>>> concurrent use of both.
>>>
>>> This can happen, for example, if a user calls prctl(PR_SET_THP_DISABLE)
>>> concurrent with faulting, as that only holds the mmap write lock and thus
>>> does not exclude faulting threads stabilised by VMA lock.
>>>
>>> insert_pmd() and insert_pud() have issues with this as they both assume
>>> that, if the page table entry being populated is not empty, it must a
>>> present huge entry.
>>>
>>> This is problematic as if it is a write fault, it compares its PFN against
>>> the one being inserted, warning if they differ, on the assumption that the
>>> entry is a huge leaf.
>>>
>>> Since the entry may be a page table, this is invalid and can result in
>>> spurious warnings.
>>>
>>> Equally, the entry might be non-present. Nothing can install such an entry
>>> in the VMAs that reach this code today, but there is no reason to rely upon
>>> that.
>>>
>>> Therefore only examine the PFN if the entry is both present and huge, and
>>> otherwise simply return VM_FAULT_NOPAGE so the fault is retried and handled
>>> at the appropriate level.
>>>
>>> This is what already happens for a read fault, for which the existing entry
>>> is not examined at all.
>>>
>>> Target the commit which introduced the first driver which both mixes fault
>>> orders and passes the write flag for the fix (since this commit, both
>>> drm_gem_shmem and panthor are prey to this issue).
>>>
>>> Reported-by: Eliav Farber <farbere@xxxxxxxxxx>
>>> Closes: https://lore.kernel.org/all/20261112051440.16140-4-farbere@xxxxxxxxxx/
>>> Fixes: 7bfe3b8ea6e3 ("Drivers: hv: Introduce mshv_vtl driver")
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> Signed-off-by: Lorenzo Stoakes (ARM) <ljs@xxxxxxxxxx>
>>> ---
>>> mm/huge_memory.c | 8 ++++++++
>>> 1 file changed, 8 insertions(+)
>>>
>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>> index 0d589be136d2..da632c3e834b 100644
>>> --- a/mm/huge_memory.c
>>> +++ b/mm/huge_memory.c
>>> @@ -1643,6 +1643,10 @@ static vm_fault_t insert_pmd(struct vm_area_struct *vma, unsigned long addr,
>>> const unsigned long pfn = fop.is_folio ? folio_pfn(fop.folio) :
>>> fop.pfn;
>>>
>>> + /* We may have raced a fault which installed a page table. */
>>> + if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd))
>>
>> Ok, we hold the PMD lock, so that cannot change concurrently.
>>
>> Given that we are in THP code, I wonder if we could just test for pmd_leaf()
>> instead.
>>
>> Same below for PUDs.
>
> That's what the original did, but as the commit message says I don't want
> to rely on the entry not being a softleaf here, in case we do something
> later that changes that.

Oh, I was talking about:

if (!pmd_present(*pmd) || !pmd_leaf(*pmd))

Which is our common way of saying

if (!(pmd_present(*pmd) && pmd_leaf(*pmd)))

"This is not a page table".

Anyhow, getting rid of pmd_trans_huge() doesn't have highest priority right now.
--
Cheers,

David