On 11/17/26 00:06, Tal Zussman wrote:
> On 11/16/26 3:45 PM, David Hildenbrand (Arm) wrote:
>> On 11/16/26 21:02, Tal Zussman wrote:
>>> Device migration replaces a PTE that maps the zero page with a new PTE
>>> that is built only from the VMA flags, so its userfaultfd WP or RWP
>>> protection is lost. The same applies to the huge zero PMD. Writes and
>>> accesses to the page are then no longer reported to userfaultfd.
>>>
>>> To fix it, do not migrate a uffd-protected zero page. Skip it during
>>> collection, and check again under the page table lock before inserting,
>>> as another thread can protect the page after it was collected.
>>>
>>> The alternative would be to carry the uffd state over to the new entry,
>>> but a uffd-protected zero page in a range that a driver migrates should
>>> be rare.
>>>
>>> Fixes: 292924b26024 ("userfaultfd: wp: apply _PAGE_UFFD_WP bit")
>>> Fixes: a30b48bf1b24 ("mm/migrate_device: implement THP migration of zone 
>>> device pages")
>>> Cc: Thomas Hellström <[email protected]>
>>> Cc: Lyude Paul <[email protected]>
>>> Cc: Danilo Krummrich <[email protected]>
>>> Cc: Felix Kuehling <[email protected]>
>>> Cc: Jason Gunthorpe <[email protected]>
>>> Cc: Leon Romanovsky <[email protected]>
>>> Cc: [email protected]
>>> Cc: [email protected]
>>> Cc: [email protected]
>>> Cc: [email protected]
>>> Reviewed-by: Alistair Popple <[email protected]>
>>> Signed-off-by: Tal Zussman <[email protected]>
>>> ---
>>>  mm/migrate_device.c | 23 +++++++++++++++++++++--
>>>  1 file changed, 21 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/mm/migrate_device.c b/mm/migrate_device.c
>>> index 9e19fadfa65c..140b99f72508 100644
>>> --- a/mm/migrate_device.c
>>> +++ b/mm/migrate_device.c
>>> @@ -159,7 +159,16 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, 
>>> unsigned long start,
>>>  
>>>             folio = pmd_folio(*pmdp);
>>>             if (is_huge_zero_folio(folio)) {
>>> +                   /*
>>> +                    * Nothing carries the uffd state of the zero page
>>> +                    * over, so skip it if its uffd bit is set.
>>> +                    */
>>> +                   bool protected = userfaultfd_protected(walk->vma) &&
>>> +                                    pmd_uffd(*pmdp);
>>
>> Now that I read it 3 times in a row below, these needs a helper.
>>
>> userfaultfd_pte_protected() ? userfaultfd_pte?
>>
>> That might help you to get rid of the local variable (that should have been
>> called something that includes uffd and should have been const)
>>
>>
>> Maybe just add a simple helper for the pmd as well and use it below.
>>
> 
> Will do. I'll do userfaultfd_pte_protected() and 
> userfaultfd_huge_pmd_protected()
> to match the existing wp and rwp helpers.
> 
> Re: the local variable, I'll rid of it and just call spin_unlock() in
> both branches, which is a little ugly.

The clean thing to do would be to read the pmd once using pmdp_get() :)

-- 
Cheers,

David

Reply via email to