* [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved
@ 2026-11-17 21:24 Eric Li
2026-11-18 23:45 ` Jose A. Perez de Azpillaga
` (2 more replies)
0 siblings, 3 replies; 4+ messages in thread
From: Eric Li @ 2026-11-17 21:24 UTC (permalink / raw)
To: Andrew Morton
Cc: Mike Rapoport, Peter Xu, David Hildenbrand, Suren Baghdasaryan,
Lokesh Gidra, Kairui Song, linux-mm, linux-kernel
move_pages_ptes() gets the folio of a present source PTE and then checks
PageAnonExclusive() on the first page of that folio. PG_anon_exclusive
is per page, so for a PTE-mapped large folio this can give the wrong
answer for the page that is actually being moved.
This happens after fork(). fork() clears the flag on every page of the
folio. Once the child exits, a write fault in the parent can reuse the
large folio, and do_wp_page() sets the flag again only on the faulting
page. This case only exists since commit 1da190f4d0a6 ("mm:
Copy-on-Write (COW) reuse support for PTE-mapped THP"). Before that,
the same state could already come from fork() after MADV_DONTFORK on
part of the folio, or from reclaim when try_to_unmap() replaces the
earlier pages with swap entries and then fails on a later one.
Moving that page with UFFDIO_MOVE then fails with -EBUSY, even though
the folio could be split and the page moved. Callers that fall back to
UFFDIO_COPY on -EBUSY end up copying instead. It also goes wrong the
other way: if the first page is exclusive and the page being moved is
not, the folio is split for nothing before the retry sees the shared
page and returns -EBUSY.
Check the flag on the page mapped by the source PTE and get the folio
from that page. Locking, refcounting and the split-and-retry path stay
the same. The batched path only handles small folios and the PMD path
already checks the page the PMD maps, so they don't need this.
Tested in an x86_64 KVM guest on mm-unstable with only 64K anonymous
mTHP enabled: map a 64K-aligned region backed by one large folio, fork,
let the child exit, write to the sixth page and move it. Without this
patch all 20 moves failed with -EBUSY. With it all 20 worked and the
data was correct. Moving the first page right after fork() still fails
with -EBUSY, as it should. Writing to the first page instead and moving
the sixth fails with -EBUSY either way, but without this patch the folio
was also split all 20 times. The move tests in uffd-unit-tests pass
with and without the patch (the two swap tests were skipped, there is
no swap in the guest). I didn't reproduce the MADV_DONTFORK or reclaim
cases and didn't test arm64.
I found this with Claude Code while debugging an out-of-tree swap
backend, and it wrote the first version of the fix. This version, the
changelog and the reproducer were also done with AI help. I've reviewed
all of it.
Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
Link: https://lore.kernel.org/all/20261116110122.309748-1-nakixii@gmail.com/
Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Signed-off-by: Eric Li <nakixii@gmail.com>
---
Notes:
v2:
- Say in the changelog that the fork() case only exists since
1da190f4d0a6, and describe the needless split when only the first
page is exclusive (Mike)
- Mention fork() after MADV_DONTFORK on part of the folio as an older
way to get there (David)
- Check the page before taking its folio (Mike)
- Sign off with my real name (Mike)
- Add Mike's Acked-by
v1: https://lore.kernel.org/all/20261117191853.1314834-1-nakixii@gmail.com/
mm/userfaultfd.c | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
index d65184d9a..e270b5722 100644
--- a/mm/userfaultfd.c
+++ b/mm/userfaultfd.c
@@ -1557,6 +1557,7 @@ static long move_pages_ptes(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd
*/
if (!src_folio) {
struct folio *folio;
+ struct page *page;
bool locked;
/*
@@ -1570,13 +1571,20 @@ static long move_pages_ptes(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd
goto out;
}
- folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
- if (!folio || !PageAnonExclusive(&folio->page)) {
+ /*
+ * PG_anon_exclusive is per page. The first page of a
+ * large folio can be shared while the page mapped here
+ * is still exclusive, so check the page being moved.
+ */
+ page = vm_normal_page(src_vma, src_addr, orig_src_pte);
+ if (!page || !PageAnonExclusive(page)) {
spin_unlock(src_ptl);
ret = -EBUSY;
goto out;
}
+ folio = page_folio(page);
+
locked = folio_trylock(folio);
/*
* We avoid waiting for folio lock with a raised
base-commit: 81188836663676f8503f3603cfe56e88264c4d50
--
2.43.0
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved
2026-11-17 21:24 [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved Eric Li
@ 2026-11-18 23:45 ` Jose A. Perez de Azpillaga
2026-11-18 23:46 ` Jose A. Perez de Azpillaga
2026-11-19 8:02 ` David Hildenbrand (Arm)
2 siblings, 0 replies; 4+ messages in thread
From: Jose A. Perez de Azpillaga @ 2026-11-18 23:45 UTC (permalink / raw)
To: Eric Li, Andrew Morton
Cc: Mike Rapoport, Peter Xu, David Hildenbrand, Suren Baghdasaryan,
Lokesh Gidra, Kairui Song, linux-mm, linux-kernel
On 11/17/26 10:24 PM, Eric Li wrote:
> move_pages_ptes() gets the folio of a present source PTE and then checks
> PageAnonExclusive() on the first page of that folio. PG_anon_exclusive
> is per page, so for a PTE-mapped large folio this can give the wrong
> answer for the page that is actually being moved.
>
> This happens after fork(). fork() clears the flag on every page of the
> folio. Once the child exits, a write fault in the parent can reuse the
> large folio, and do_wp_page() sets the flag again only on the faulting
> page. This case only exists since commit 1da190f4d0a6 ("mm:
> Copy-on-Write (COW) reuse support for PTE-mapped THP"). Before that,
> the same state could already come from fork() after MADV_DONTFORK on
> part of the folio, or from reclaim when try_to_unmap() replaces the
> earlier pages with swap entries and then fails on a later one.
>
> Moving that page with UFFDIO_MOVE then fails with -EBUSY, even though
> the folio could be split and the page moved. Callers that fall back to
> UFFDIO_COPY on -EBUSY end up copying instead. It also goes wrong the
> other way: if the first page is exclusive and the page being moved is
> not, the folio is split for nothing before the retry sees the shared
> page and returns -EBUSY.
>
> Check the flag on the page mapped by the source PTE and get the folio
> from that page. Locking, refcounting and the split-and-retry path stay
> the same. The batched path only handles small folios and the PMD path
> already checks the page the PMD maps, so they don't need this.
>
> Tested in an x86_64 KVM guest on mm-unstable with only 64K anonymous
> mTHP enabled: map a 64K-aligned region backed by one large folio, fork,
> let the child exit, write to the sixth page and move it. Without this
> patch all 20 moves failed with -EBUSY. With it all 20 worked and the
> data was correct. Moving the first page right after fork() still fails
> with -EBUSY, as it should. Writing to the first page instead and moving
> the sixth fails with -EBUSY either way, but without this patch the folio
> was also split all 20 times. The move tests in uffd-unit-tests pass
> with and without the patch (the two swap tests were skipped, there is
> no swap in the guest). I didn't reproduce the MADV_DONTFORK or reclaim
> cases and didn't test arm64.
>
> I found this with Claude Code while debugging an out-of-tree swap
> backend, and it wrote the first version of the fix. This version, the
> changelog and the reproducer were also done with AI help. I've reviewed
> all of it.
>
> Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
> Link: https://lore.kernel.org/all/20261116110122.309748-1-nakixii@gmail.com/
> Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
> Signed-off-by: Eric Li <nakixii@gmail.com>
LGTM:
Reviewed-by: Jose A. Perez de Azpillaga <azpijr@gmail.com>
--
cheers,
jose a. p-a
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved
2026-11-17 21:24 [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved Eric Li
2026-11-18 23:45 ` Jose A. Perez de Azpillaga
@ 2026-11-18 23:46 ` Jose A. Perez de Azpillaga
2026-11-19 8:02 ` David Hildenbrand (Arm)
2 siblings, 0 replies; 4+ messages in thread
From: Jose A. Perez de Azpillaga @ 2026-11-18 23:46 UTC (permalink / raw)
To: Eric Li, Andrew Morton
Cc: Mike Rapoport, Peter Xu, David Hildenbrand, Suren Baghdasaryan,
Lokesh Gidra, Kairui Song, linux-mm, linux-kernel
On 11/17/26 10:24 PM, Eric Li wrote:
> move_pages_ptes() gets the folio of a present source PTE and then checks
> PageAnonExclusive() on the first page of that folio. PG_anon_exclusive
> is per page, so for a PTE-mapped large folio this can give the wrong
> answer for the page that is actually being moved.
>
> This happens after fork(). fork() clears the flag on every page of the
> folio. Once the child exits, a write fault in the parent can reuse the
> large folio, and do_wp_page() sets the flag again only on the faulting
> page. This case only exists since commit 1da190f4d0a6 ("mm:
> Copy-on-Write (COW) reuse support for PTE-mapped THP"). Before that,
> the same state could already come from fork() after MADV_DONTFORK on
> part of the folio, or from reclaim when try_to_unmap() replaces the
> earlier pages with swap entries and then fails on a later one.
>
> Moving that page with UFFDIO_MOVE then fails with -EBUSY, even though
> the folio could be split and the page moved. Callers that fall back to
> UFFDIO_COPY on -EBUSY end up copying instead. It also goes wrong the
> other way: if the first page is exclusive and the page being moved is
> not, the folio is split for nothing before the retry sees the shared
> page and returns -EBUSY.
>
> Check the flag on the page mapped by the source PTE and get the folio
> from that page. Locking, refcounting and the split-and-retry path stay
> the same. The batched path only handles small folios and the PMD path
> already checks the page the PMD maps, so they don't need this.
>
> Tested in an x86_64 KVM guest on mm-unstable with only 64K anonymous
> mTHP enabled: map a 64K-aligned region backed by one large folio, fork,
> let the child exit, write to the sixth page and move it. Without this
> patch all 20 moves failed with -EBUSY. With it all 20 worked and the
> data was correct. Moving the first page right after fork() still fails
> with -EBUSY, as it should. Writing to the first page instead and moving
> the sixth fails with -EBUSY either way, but without this patch the folio
> was also split all 20 times. The move tests in uffd-unit-tests pass
> with and without the patch (the two swap tests were skipped, there is
> no swap in the guest). I didn't reproduce the MADV_DONTFORK or reclaim
> cases and didn't test arm64.
>
> I found this with Claude Code while debugging an out-of-tree swap
> backend, and it wrote the first version of the fix. This version, the
> changelog and the reproducer were also done with AI help. I've reviewed
> all of it.
>
> Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
> Link: https://lore.kernel.org/all/20261116110122.309748-1-nakixii@gmail.com/
> Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
> Signed-off-by: Eric Li <nakixii@gmail.com>
LGTM:
Reviewed-by: Jose A. Perez de Azpillaga <azpijr@gmail.com>
--
cheers,
jose a. p-a
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved
2026-11-17 21:24 [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved Eric Li
2026-11-18 23:45 ` Jose A. Perez de Azpillaga
2026-11-18 23:46 ` Jose A. Perez de Azpillaga
@ 2026-11-19 8:02 ` David Hildenbrand (Arm)
2 siblings, 0 replies; 4+ messages in thread
From: David Hildenbrand (Arm) @ 2026-11-19 8:02 UTC (permalink / raw)
To: Eric Li, Andrew Morton
Cc: Mike Rapoport, Peter Xu, Suren Baghdasaryan, Lokesh Gidra,
Kairui Song, linux-mm, linux-kernel
On 11/17/26 22:24, Eric Li wrote:
> move_pages_ptes() gets the folio of a present source PTE and then checks
> PageAnonExclusive() on the first page of that folio. PG_anon_exclusive
> is per page, so for a PTE-mapped large folio this can give the wrong
> answer for the page that is actually being moved.
>
> This happens after fork(). fork() clears the flag on every page of the
> folio. Once the child exits, a write fault in the parent can reuse the
> large folio, and do_wp_page() sets the flag again only on the faulting
> page. This case only exists since commit 1da190f4d0a6 ("mm:
> Copy-on-Write (COW) reuse support for PTE-mapped THP"). Before that,
> the same state could already come from fork() after MADV_DONTFORK on
> part of the folio, or from reclaim when try_to_unmap() replaces the
> earlier pages with swap entries and then fails on a later one.
>
> Moving that page with UFFDIO_MOVE then fails with -EBUSY, even though
> the folio could be split and the page moved. Callers that fall back to
> UFFDIO_COPY on -EBUSY end up copying instead. It also goes wrong the
> other way: if the first page is exclusive and the page being moved is
> not, the folio is split for nothing before the retry sees the shared
> page and returns -EBUSY.
>
> Check the flag on the page mapped by the source PTE and get the folio
> from that page. Locking, refcounting and the split-and-retry path stay
> the same. The batched path only handles small folios and the PMD path
> already checks the page the PMD maps, so they don't need this.
>
> Tested in an x86_64 KVM guest on mm-unstable with only 64K anonymous
> mTHP enabled: map a 64K-aligned region backed by one large folio, fork,
> let the child exit, write to the sixth page and move it. Without this
> patch all 20 moves failed with -EBUSY. With it all 20 worked and the
> data was correct. Moving the first page right after fork() still fails
> with -EBUSY, as it should. Writing to the first page instead and moving
> the sixth fails with -EBUSY either way, but without this patch the folio
> was also split all 20 times. The move tests in uffd-unit-tests pass
> with and without the patch (the two swap tests were skipped, there is
> no swap in the guest). I didn't reproduce the MADV_DONTFORK or reclaim
> cases and didn't test arm64.
>
> I found this with Claude Code while debugging an out-of-tree swap
> backend, and it wrote the first version of the fix. This version, the
> changelog and the reproducer were also done with AI help. I've reviewed
> all of it.
>
> Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
> Link: https://lore.kernel.org/all/20261116110122.309748-1-nakixii@gmail.com/
> Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
> Signed-off-by: Eric Li <nakixii@gmail.com>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
--
Cheers,
David
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2026-11-19 8:02 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-17 21:24 [PATCH v2] mm/userfaultfd: check exclusivity of the page being moved Eric Li
2026-11-18 23:45 ` Jose A. Perez de Azpillaga
2026-11-18 23:46 ` Jose A. Perez de Azpillaga
2026-11-19 8:02 ` David Hildenbrand (Arm)
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®