Re: [PATCH mm-unstable v18 06/14] mm/khugepaged: generalize collapse_huge_page for mTHP collapse

[Lance Yang <lance.yang@linux.dev>]

On 2026/6/1 14:54, David Hildenbrand (Arm) wrote:
> On 6/1/26 05:28, Lance Yang wrote:
>>
>> On Sun, May 31, 2026 at 10:00:17PM +0200, David Hildenbrand (Arm) wrote:
>>> On 5/31/26 11:39, Lance Yang wrote:
>>>>
>>>>
>>>> Emm ... is it safe to use map_anon_folio_pte_nopf() here?
>>>>
>>>> At this point pmdp_collapse_flush() has cleared the PMD from the page
>>>> tables. The PTE table we are updating is only reachable through the saved
>>>> old PMD value, _pmd, until pmd_populate() below.
>>>>
>>>> map_anon_folio_pte_nopf() does set_ptes() and then calls
>>>> update_mmu_cache_range(). Documentation/core-api/cachetlb.rst describes
>>>> that hook as:
>>>>
>>>> "
>>>> 	At the end of every page fault, this routine is invoked to tell
>>>> 	the architecture specific code that translations now exists
>>>> 	in the software page tables for address space "vma->vm_mm"
>>>> 	at virtual address "address" for "nr" consecutive pages.
>>>> "
>>>>
>>>> But that does not seem true here yet, since the PTE table is not
>>>> reachable from vma->vm_mm when update_mmu_cache_range() is called.
>>>>
>>>> Should we avoid calling update_mmu_cache_range() until after the PTE
>>>> table is reinstalled with pmd_populate()?
>>>
>>> I recall that update_mmu_cache* users mostly care about updating folios flags,
>>> for the folio derived from the PTE ... or flushing caches for the user address.
>>>
>>> So intuitively I would say "the architecture code doesn't care that the PMD
>>> table will only be visible to HW shortly after". The important thing should be
>>> that it will definetly happen, and that nothing else is curently there or can be
>>> there?
>>
>> Ah, fair point.
>>
>> I was mostly worried about arch hooks that walk vma->vm_mm again, rather
>> than only using the pte pointer passed in. For example, mips does:
> 
> Right, a re-walk would be the real problem.
> 
>>
>>    update_mmu_cache_range()
>>      -> __update_tlb()
>>        -> pgd_offset(vma->vm_mm, address)
>>        -> pte_offset_map(...)
>>
>> and __update_tlb() has this assumption:
>>
>> 		/*
>> 		 * update_mmu_cache() is called between pte_offset_map_lock()
>> 		 * and pte_unmap_unlock(), so we can assume that ptep is not
>> 		 * NULL here: and what should be done below if it were NULL?
>> 		 */
>>
>> So if khugepaged happens to run with current->active_mm == vma->vm_mm
>> here, could __update_tlb() hit the none PMD, get NULL from
>> pte_offset_map(), and then dereference it?
> 
> Likely yes -- that MIPS code is horrible. And the comment in MIPS code
> even spells that out. :(
> 
> Do you know about other code like that, or is MIPS the only one doing a
> re-walk and crossing fingers?
> 
>>
>> Just wanted to raise it since some arch code may still have assumptions
>> like this, and the always-enable-mTHP work is getting closer ...
> 
> Right. I assume set_pte_at() couldn't trigger something similar (re-walk) in arch code,
> because we simply provide the ptep. update_mmu_cache_range() only consumes the pte.
> 
>>
>> Probably very very very hard to hit, though :)
> 
> Delaying update_mmu_cache_range() is nasty, as we'd have to make sure that
> nobody can interfere in the meantime ... and the PMD lock will not be sufficient.
> 
> Maybe we could reinstall the page table with the cleared (none) entries while
> still holding the PTL?
> 
> Thinking out loud:
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 5ba298d420b7..e39b750b1e6f 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1413,13 +1413,17 @@ static enum scan_result collapse_huge_page(struct mm_struct *mm, unsigned long s
>                  map_anon_folio_pmd_nopf(folio, pmd, vma, pmd_addr);
>          } else {
>                  /*
> -                * set_ptes is called in map_anon_folio_pte_nopf with the
> -                * pmd_ptl lock still held; this is safe as the PMD is expected
> -                * to be none. The pmd entry is then repopulated below.
> +                * Re-insert the page table with the cleared entries, but
> +                * hold the PTL, such that no one can mess with the re-installed
> +                * page table until we updated the temporarily-cleared entries
> +                * through map_anon_folio_pte_nopf().
>                   */
> -               map_anon_folio_pte_nopf(folio, pte, vma, start_addr, /*uffd_wp=*/ false);
> -               smp_wmb(); /* make PTEs visible before PMD. See pmd_install() */

One small thing, I think we should probably keep the smp_wmb(), and just
move it before the earlier pmd_populate().

IIUC, the ordering we want is still:

   clear old PTEs
   smp_wmb()
   pmd_populate()

so another CPU cannot walk through the re-installed PMD and still observe
the old PTEs, right?

> +               if (pte_ptl != pmd_ptl)
> +                       spin_lock(pte_ptl);
>                  pmd_populate(mm, pmd, pmd_pgtable(_pmd));
> +               map_anon_folio_pte_nopf(folio, pte, vma, start_addr, /*uffd_wp=*/ false);
> +               if (pte_ptl != pmd_ptl)
> +                       spin_unlock(pte_ptl);
>          }
>          spin_unlock(pmd_ptl);
>   

Cheers, Lance