[PATCH] mm/khugepaged: Don't collapse uffd-minor-registered VMAs

[PATCH] mm/khugepaged: Don't collapse uffd-minor-registered VMAs

[James Houghton]

Userfaultfd minor faults provides userspace with the ability to manually
install PTEs with UFFDIO_CONTINUE. Right now, khugepaged collapse can
map holes in the VMA when a naturally-aligned THP is present without
explicit action from userspace.

This is a problem, as it bypasses userfaultfd minor faults that
userspace is expecting to handle.

If userspace implements post-copy live migration using userfaultfd minor
faults, this situation is currently possible:
1. The VMA for guest memory is userfaultfd-minor-registered and nothing
   is mapped in the page tables.
2. A stale copy of a page is present in a naturally-aligned THP (from
   pre-copy live migration).
3. khugepaged collapses the mapping of the THP, installs a PMD.
4. The VM now has access to the stale contents => VM is broken.
5. After installing the correct contents, userspace attempts to map the
   page with UFFDIO_CONTINUE; it gets EEXIST, indicating that something
   unexpectedly mapped the page.

The naturally-aligned THP case is the only case where this is a problem.
khugepaged otherwise requires all PTEs to be present for
userfaultfd-registered VMAs (i.e., max none PTEs is 0), which is
correct. This check is essentially bypassed for naturally-aligned THPs.

To deal with this issue, completely disallow collapsing in
userfaultfd-minor-registered VMAs. This is slightly pessimistic; it
would be nice to allow MADV_COLLAPSE to work if all PTEs are in fact
present, but that seems more complex than it is worth.

Fixes: 58ac9a8993a1 ("mm/khugepaged: attempt to map file/shmem-backed pte-mapped THPs by pmds")
Cc: <stable@vger.kernel.org> # 6.1
Signed-off-by: James Houghton <jthoughton@google.com>
---
This was caught with manual review while diagnosing a related issue
that came up with in Google's live migration testing.

I've uploaded a mostly-AI-generated reproducer here[1]. As long as
/sys/kernel/mm/transparent_hugepage/shmem_enabled is not set to 'deny',
the repro should work.

[1] https://gist.github.com/48ca/d399bf534158e80241fb4937ef1ff664
---
 mm/khugepaged.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index b237f6e7662a..66f956d3dd67 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -2804,6 +2804,15 @@ static enum scan_result collapse_single_pmd(unsigned long addr,
 		goto end;
 	}
 
+	/*
+	 * Userfaultfd-minor-registered VMAs should not be collapsed, as
+	 * userspace is expecting to explicitly install PTEs.
+	 */
+	if (userfaultfd_minor(vma)) {
+		result = SCAN_PTE_UFFD;
+		goto end;
+	}
+
 	file = get_file(vma->vm_file);
 	pgoff = linear_page_index(vma, addr);
 

base-commit: 26260251022fbc2f248a3d747a9b2b961b18d2d8
-- 
2.55.0.897.gb25b4bd76c-goog

Re: [PATCH] mm/khugepaged: Don't collapse uffd-minor-registered VMAs

[Lance Yang]

On Fri, Aug 28, 2026 at 12:50:04AM +0000, James Houghton wrote:
>Userfaultfd minor faults provides userspace with the ability to manually
>install PTEs with UFFDIO_CONTINUE. Right now, khugepaged collapse can
>map holes in the VMA when a naturally-aligned THP is present without
>explicit action from userspace.
>
>This is a problem, as it bypasses userfaultfd minor faults that
>userspace is expecting to handle.

One basic question first. Should MADV_COLLAPSE refuse to collapse a
UFFD-minor-registered VMA, regardless of whether all PTEs are present?

I'd leave that to the maintainers :D

Anyway, assuming the answer is yes, I wonder whether the new check is
sufficient. See below.

>
>If userspace implements post-copy live migration using userfaultfd minor
>faults, this situation is currently possible:
>1. The VMA for guest memory is userfaultfd-minor-registered and nothing
>   is mapped in the page tables.
>2. A stale copy of a page is present in a naturally-aligned THP (from
>   pre-copy live migration).
>3. khugepaged collapses the mapping of the THP, installs a PMD.
>4. The VM now has access to the stale contents => VM is broken.
>5. After installing the correct contents, userspace attempts to map the
>   page with UFFDIO_CONTINUE; it gets EEXIST, indicating that something
>   unexpectedly mapped the page.
>
>The naturally-aligned THP case is the only case where this is a problem.
>khugepaged otherwise requires all PTEs to be present for
>userfaultfd-registered VMAs (i.e., max none PTEs is 0), which is
>correct. This check is essentially bypassed for naturally-aligned THPs.
>
>To deal with this issue, completely disallow collapsing in
>userfaultfd-minor-registered VMAs. This is slightly pessimistic; it
>would be nice to allow MADV_COLLAPSE to work if all PTEs are in fact
>present, but that seems more complex than it is worth.
>
>Fixes: 58ac9a8993a1 ("mm/khugepaged: attempt to map file/shmem-backed pte-mapped THPs by pmds")
>Cc: <stable@vger.kernel.org> # 6.1
>Signed-off-by: James Houghton <jthoughton@google.com>
>---
>This was caught with manual review while diagnosing a related issue
>that came up with in Google's live migration testing.
>
>I've uploaded a mostly-AI-generated reproducer here[1]. As long as
>/sys/kernel/mm/transparent_hugepage/shmem_enabled is not set to 'deny',
>the repro should work.
>
>[1] https://gist.github.com/48ca/d399bf534158e80241fb4937ef1ff664
>---
> mm/khugepaged.c | 9 +++++++++
> 1 file changed, 9 insertions(+)
>
>diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>index b237f6e7662a..66f956d3dd67 100644
>--- a/mm/khugepaged.c
>+++ b/mm/khugepaged.c
>@@ -2804,6 +2804,15 @@ static enum scan_result collapse_single_pmd(unsigned long addr,
> 		goto end;
> 	}
> 
>+	/*
>+	 * Userfaultfd-minor-registered VMAs should not be collapsed, as
>+	 * userspace is expecting to explicitly install PTEs.
>+	 */
>+	if (userfaultfd_minor(vma)) {
>+		result = SCAN_PTE_UFFD;
>+		goto end;
>+	}

Assume UFFDIO_REGISTER_MODE_MINOR completes after collapse_single_pmd()
drops the mmap read lock and before it reacquires it.

Doesn't this still leave a registration race, no?


int madvise_collapse(struct vm_area_struct *vma, unsigned long start,
		     unsigned long end, bool *lock_dropped)
{
...
	cc->is_khugepaged = false;
...
		result = collapse_single_pmd(addr, vma, &mmap_unlocked, cc);
...
}

static enum scan_result collapse_single_pmd(unsigned long addr,
		struct vm_area_struct *vma, bool *lock_dropped,
		struct collapse_control *cc)
{
...
	if (userfaultfd_minor(vma)) {
		result = SCAN_PTE_UFFD;
		goto end;
	}
...
	mmap_read_unlock(mm);
	*lock_dropped = true;
...
	if (result == SCAN_PTE_MAPPED_HUGEPAGE) {
		mmap_read_lock(mm);
		if (collapse_test_exit_or_disable(mm))
			result = SCAN_ANY_PROCESS;
		else
			result = try_collapse_pte_mapped_thp(mm, addr,
							     !cc->is_khugepaged);
...
		mmap_read_unlock(mm);
	}
...
}

static enum scan_result try_collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
		bool install_pmd)
{
...
	struct vm_area_struct *vma = vma_lookup(mm, haddr);
...
	if (!vma || !vma->vm_file ||
	    !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
		return SCAN_VMA_CHECK;
...
	if (userfaultfd_protected(vma))
		return SCAN_PTE_UFFD;
...
	result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
	switch (result) {
	case SCAN_SUCCEED:
		break;
	case SCAN_NO_PTE_TABLE:
...
		goto maybe_install_pmd;
	default:
		goto drop_folio;
	}
...
maybe_install_pmd:
	/* step 5: install pmd entry */
	result = install_pmd
			? set_huge_pmd(vma, haddr, pmd, folio, &folio->page)
			: SCAN_SUCCEED;
...
}


static inline bool userfaultfd_minor(struct vm_area_struct *vma)
{
	return vma_test_any_mask(vma, VMA_UFFD_MINOR);
}

static inline bool userfaultfd_protected(struct vm_area_struct *vma)
{
	return userfaultfd_wp(vma) || userfaultfd_rwp(vma);
}

Emm ... userfaultfd_protected() only covers WP and RWP. MADV_COLLAPSE
passes install_pmd=true, so the SCAN_NO_PTE_TABLE case can still reach
set_huge_pmd() after UFFDIO_REGISTER_MODE_MINOR has completed ...

Maybe:

---8<---
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 33c41bc32af8..0eada7265d59 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1893,6 +1893,8 @@ static enum scan_result try_collapse_pte_mapped_thp(struct mm_struct *mm, unsign
 	 */
 	if (userfaultfd_protected(vma))
 		return SCAN_PTE_UFFD;
+	if (userfaultfd_minor(vma))
+		return SCAN_PTE_UFFD;

 	folio = filemap_lock_folio(vma->vm_file->f_mapping,
 			       linear_page_index(vma, haddr));
--

With that, LGTM.

Tested-by: Lance Yang <lance.yang@linux.dev>

Cheers, Lance

Re: [PATCH] mm/khugepaged: Don't collapse uffd-minor-registered VMAs

[Kiryl Shutsemau]

On Fri, Aug 28, 2026 at 05:47:03PM +0800, Lance Yang wrote:
> 
> On Fri, Aug 28, 2026 at 12:50:04AM +0000, James Houghton wrote:
> >Userfaultfd minor faults provides userspace with the ability to manually
> >install PTEs with UFFDIO_CONTINUE. Right now, khugepaged collapse can
> >map holes in the VMA when a naturally-aligned THP is present without
> >explicit action from userspace.
> >
> >This is a problem, as it bypasses userfaultfd minor faults that
> >userspace is expecting to handle.
> 
> One basic question first. Should MADV_COLLAPSE refuse to collapse a
> UFFD-minor-registered VMA, regardless of whether all PTEs are present?

I think the line is not between MADV_COLLAPSE vs khugepaged, it is if
vma->vm_mm is current->mm or not.

In other words, we might want to allow collapse if workload itself asks
for it, but not for khugepaged or *external* process_madvise(MADV_COLLAPSE).
is_khugepaged is not the right proxy.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCH] mm/khugepaged: Don't collapse uffd-minor-registered VMAs

[James Houghton]

On Fri, Aug 28, 2026 at 2:47 AM Lance Yang <lance.yang@linux.dev> wrote:
>
> On Fri, Aug 28, 2026 at 12:50:04AM +0000, James Houghton wrote:
> >Userfaultfd minor faults provides userspace with the ability to manually
> >install PTEs with UFFDIO_CONTINUE. Right now, khugepaged collapse can
> >map holes in the VMA when a naturally-aligned THP is present without
> >explicit action from userspace.
> >
> >This is a problem, as it bypasses userfaultfd minor faults that
> >userspace is expecting to handle.
>
> One basic question first. Should MADV_COLLAPSE refuse to collapse a
> UFFD-minor-registered VMA, regardless of whether all PTEs are present?
>
> I'd leave that to the maintainers :D
>
> Anyway, assuming the answer is yes, I wonder whether the new check is
> sufficient. See below.

I also think the answer is yes.

Theoretically you could optimize this and allow userspace to avoid
calling a bunch of UFFDIO_CONTINUEs and instead call MADV_COLLAPSE,
but let's wait for someone (probably won't be me, at least not for a
couple years) to ask for it. But khugepaged's own collapse routines
(and potentially other processes via process_madvise()) should remain
unchanged; that would reintroduce the bug being fixed here.

> static inline bool userfaultfd_minor(struct vm_area_struct *vma)
> {
>         return vma_test_any_mask(vma, VMA_UFFD_MINOR);
> }
>
> static inline bool userfaultfd_protected(struct vm_area_struct *vma)
> {
>         return userfaultfd_wp(vma) || userfaultfd_rwp(vma);
> }
>
> Emm ... userfaultfd_protected() only covers WP and RWP. MADV_COLLAPSE
> passes install_pmd=true, so the SCAN_NO_PTE_TABLE case can still reach
> set_huge_pmd() after UFFDIO_REGISTER_MODE_MINOR has completed ...
>
> Maybe:
>
> ---8<---
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 33c41bc32af8..0eada7265d59 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1893,6 +1893,8 @@ static enum scan_result try_collapse_pte_mapped_thp(struct mm_struct *mm, unsign
>          */
>         if (userfaultfd_protected(vma))
>                 return SCAN_PTE_UFFD;
> +       if (userfaultfd_minor(vma))
> +               return SCAN_PTE_UFFD;
>
>         folio = filemap_lock_folio(vma->vm_file->f_mapping,
>                                linear_page_index(vma, haddr));
> --
>
> With that, LGTM.
>
> Tested-by: Lance Yang <lance.yang@linux.dev>

Thanks, Lance. This is how I should have written the patch to begin
with. :) It addresses both comments that Sashiko left as well.

I'll send a v2 soon. Thank you!