Re: [PATCH 1/3] mm/memory: remove racy huge PMD assertion from walk_to_pmd()

From: David Hildenbrand (Arm)

Date: Mon Nov 16 2026 - 15:06:38 EST


On 11/16/26 18:10, Lorenzo Stoakes (ARM) wrote:
> On Mon, Nov 16, 2026 at 04:56:25PM +0100, David Hildenbrand (Arm) wrote:
>> On 11/15/26 21:58, Lorenzo Stoakes (ARM) wrote:
>>> From: Eliav Farber <farbere@xxxxxxxxxx>
>>>
>>> walk_to_pmd() asserts that the PMD entry it returns is not huge, but does
>>> so without holding the PMD lock, so the assertion is only valid if nothing
>>> can concurrently install a huge PMD entry in the range.
>>>
>>> This is not the case for a VMA whose driver implements both ->fault and
>>> ->huge_fault and which invokes vmf_insert_pfn() from the former.
>>>
>>> The existence of prctl(PR_SET_THP_DISABLE) is problematic here, as it only
>>> takes the mmap write lock and faults taken under the VMA lock are
>>> unaffected by this.
>>>
>>> This can result in two concurrent threads faulting the same thing
>>> disagreeing as to whether to attempt a huge fault and thus ->fault and
>>> ->huge_fault running at the same time and triggering this assertion.
>>>
>>> So just remove the assert.
>>>
>>> Target the commit which introduced the first driver able to do this for the
>>> fix - vfio-pci. DAX also mixes fault orders within a VMA, but serialises
>>> them using its own mapping entry lock.
>>>
>>> Link: https://lore.kernel.org/all/20261112051440.16140-2-farbere@xxxxxxxxxx/
>>> Fixes: f9e54c3a2f5b ("vfio/pci: implement huge_fault support")
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> Signed-off-by: Eliav Farber <farbere@xxxxxxxxxx>
>>> [ ljs: rewrote commit message ]
>>> Signed-off-by: Lorenzo Stoakes (ARM) <ljs@xxxxxxxxxx>
>>> ---
>>> mm/memory.c | 1 -
>>> 1 file changed, 1 deletion(-)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 3a20d64fd276..e03110906aa8 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -2384,7 +2384,6 @@ static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
>>> if (!pmd)
>>> return NULL;
>>>
>>> - VM_BUG_ON(pmd_trans_huge(*pmd));
>>> return pmd;
>>> }
>>>
>>>
>>
>> Could this also stumble over a huge PUD somehow? Should we just prepare it for
>> any leaves and bail out if we find some?
>
> No, walk_to_pmd() can only come from the order-0 path right now and
> __handle_mm_fault() makes sure that can't happen in the fault path:

That implies that get_locked_pte() and all insert_pages interfaces can only be
called from that path in __handle_mm_fault().

x86 an s390x use get_locked_pte() for some other paths, so I wouldn't want to
trust on that :)

(did I say that I hate walk_to_pmd and get_locked_pte? But you likely guessed it
at this point lol)

>
> vmf.pmd = pmd_alloc(mm, vmf.pud, address);
> if (!vmf.pmd)
> return VM_FAULT_OOM;
>
> /* Huge pud page fault raced with pmd_alloc? */
> if (pud_trans_unstable(vmf.pud))
> goto retry_pud;
>
> So by the time any ->fault path runs, the PUD is stable.
>
> We don't have to worry about the mmap paths as at that point nothing has
> faulted in and it's all fresh.
>
> So it's all safe as things stand.
>
> But could update for future-proofing like:
>
> /* PUD leaf might have been installed. */
> if (pud_trans_unstable(pud))
> return NULL;

Yeah, that makes sense. I was just wondering whether we'd actually want to
return a proper error then instead, because this would not be an -ENOMEM case.

--
Cheers,

David