Re: [PATCH v2] mm/gup: honour FOLL_PIN in NOMMU __get_user_pages_locked()

[Andrew Morton]

On Fri, 24 Apr 2026 16:19:32 +0200 "David Hildenbrand (Arm)" <david@xxxxxxxxxx> wrote:

> On 4/24/26 15:38, Andrew Morton wrote:
> > On Thu, 23 Apr 2026 16:28:04 +0200 Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > 
> >> The !CONFIG_MMU implementation of __get_user_pages_locked() takes a bare
> >> get_page() reference for each page regardless of foll_flags:
> >> 	if (pages[i])
> >> 		get_page(pages[i]);
> >>
> >> This is reached from pin_user_pages*() with FOLL_PIN set.
> >> unpin_user_page() is shared between MMU and NOMMU configurations and
> >> unconditionally calls gup_put_folio(..., FOLL_PIN), which subtracts
> >> GUP_PIN_COUNTING_BIAS (1024) from the folio refcount.
> >>
> >> This means that pin adds 1, and then unpin will subtract 1024.
> >>
> >> If a user maps a page (refcount 1), registers it 1023 times as an
> >> io_uring fixed buffer (1023 pin_user_pages calls -> refcount 1024), then
> >> unregisters: the first unpin_user_page subtracts 1024, refcount hits 0,
> >> the page is freed and returned to the buddy allocator.  The remaining
> >> 1022 unpins write into whatever was reallocated, and the user's VMA
> >> still maps the freed page (NOMMU has no MMU to invalidate it).
> >> Reallocating the page for an io_uring pbuf_ring then lets userspace
> >> corrupt the new owner's data through the stale mapping.
> >>
> >> Use try_grab_folio() which adds GUP_PIN_COUNTING_BIAS for FOLL_PIN and 1
> >> for FOLL_GET, mirroring the CONFIG_MMU path so pin and unpin are
> >> symmetric.
> > 
> > Battle of the bots?
> > 	https://sashiko.dev/#/patchset/2026042303-vendor-outright-b9d2@gregkh
> 
> It references the
> 
> 	if (pages && !(flags & FOLL_PIN))
> 		flags |= FOLL_GET;
> 
> I'm not sure if there is actual NOMMU code that triggers it. For example,
> uprobes uses that pattern, but I suspect that that's not a thing on NOMMU.
> 
> 
> Probably best to just squash:
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index ad9ded39609c..44bd28cf6e00 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1988,6 +1988,9 @@ static long __get_user_pages_locked(struct mm_struct *mm,
> unsigned long start,
>         if (!nr_pages)
>                 return 0;
> 
> +       if (pages && !(foll_flags & FOLL_PIN))
> +               foll_flags |= FOLL_GET;
> +
>         /*
>          * The internal caller expects GUP to manage the lock internally and the
>          * lock must be released when this returns.

Nothing happened.

Should we proceed with this as-is?

From: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
Subject: mm/gup: honour FOLL_PIN in NOMMU __get_user_pages_locked()
Date: Thu, 23 Apr 2026 16:28:04 +0200

The !CONFIG_MMU implementation of __get_user_pages_locked() takes a bare
get_page() reference for each page regardless of foll_flags:

	if (pages[i])
		get_page(pages[i]);

This is reached from pin_user_pages*() with FOLL_PIN set. 
unpin_user_page() is shared between MMU and NOMMU configurations and
unconditionally calls gup_put_folio(..., FOLL_PIN), which subtracts
GUP_PIN_COUNTING_BIAS (1024) from the folio refcount.

This means that pin adds 1, and then unpin will subtract 1024.

If a user maps a page (refcount 1), registers it 1023 times as an io_uring
fixed buffer (1023 pin_user_pages calls -> refcount 1024), then
unregisters: the first unpin_user_page subtracts 1024, refcount hits 0,
the page is freed and returned to the buddy allocator.  The remaining 1022
unpins write into whatever was reallocated, and the user's VMA still maps
the freed page (NOMMU has no MMU to invalidate it).  Reallocating the page
for an io_uring pbuf_ring then lets userspace corrupt the new owner's data
through the stale mapping.

Use try_grab_folio() which adds GUP_PIN_COUNTING_BIAS for FOLL_PIN and 1
for FOLL_GET, mirroring the CONFIG_MMU path so pin and unpin are
symmetric.

While at it, don't return NULL pointers in the page array, as this is
really not expected for GUP users; instead, just fail and return
-EFAULT.

[david@xxxxxxxxxx: changelog update]
  https://lore.kernel.org/e9c5cf89-fa4c-4b83-ae70-9d3c72542ee9@xxxxxxxxxx
Link: https://lore.kernel.org/2026042303-vendor-outright-b9d2@gregkh
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
Acked-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
Assisted-by: David Hildenbrand <david@xxxxxxxxxx>
Cc: Jason Gunthorpe <jgg@xxxxxxxx>
Cc: John Hubbard <jhubbard@xxxxxxxxxx>
Cc: Peter Xu <peterx@xxxxxxxxxx>
Reported-by: Anthropic
Assisted-by: gkh_clanker_t1000
Cc: <stable@xxxxxxxxxxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
---

 mm/gup.c |   13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

--- a/mm/gup.c~mm-gup-honour-foll_pin-in-nommu-__get_user_pages_locked
+++ a/mm/gup.c
@@ -1983,6 +1983,7 @@ static long __get_user_pages_locked(stru
 	struct vm_area_struct *vma;
 	bool must_unlock = false;
 	vm_flags_t vm_flags;
+	int ret, err = -EFAULT;
 	long i;
 
 	if (!nr_pages)
@@ -2019,8 +2020,14 @@ static long __get_user_pages_locked(stru
 
 		if (pages) {
 			pages[i] = virt_to_page((void *)start);
-			if (pages[i])
-				get_page(pages[i]);
+			if (!pages[i])
+				break;
+			ret = try_grab_folio(page_folio(pages[i]), 1, foll_flags);
+			if (unlikely(ret)) {
+				pages[i] = NULL;
+				err = ret;
+				break;
+			}
 		}
 
 		start = (start + PAGE_SIZE) & PAGE_MASK;
@@ -2031,7 +2038,7 @@ static long __get_user_pages_locked(stru
 		*locked = 0;
 	}
 
-	return i ? : -EFAULT;
+	return i ? : err;
 }
 #endif /* !CONFIG_MMU */
 
_