[PATCH 1/2] userfaultfd: validate the whole range before changing write protection

[PATCH 1/2] userfaultfd: validate the whole range before changing write protection

[Umang Pokhriyal]

mwriteprotect_range() checks each VMA as it walks the range.  When a later
VMA is not registered for write protection, or is a hugetlb VMA the range
is not aligned to, the call fails after the earlier VMAs were already
changed.  When unprotecting, userfaultfd_writeprotect() then skips
wake_userfault(), so threads blocked on those pages are not woken although
the pages are writable again.

An unprivileged process can trigger this with a range that spans an
unregistered VMA.  It hasn't been seen in practice; it was found while
reviewing the userfaultfd range ioctls with an LLM.

Validate the whole range first, as mrwprotect_range() already does.

Fixes: a1b92a3f1498 ("mm/userfaultfd: support WP on multiple VMAs")
Cc: stable@vger.kernel.org
Signed-off-by: Umang Pokhriyal <umangpokhriyall@gmail.com>
---
 mm/userfaultfd.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
index 87da92526..b8a9d9574 100644
--- a/mm/userfaultfd.c
+++ b/mm/userfaultfd.c
@@ -1105,21 +1105,31 @@ static int mwriteprotect_range(struct userfaultfd_ctx *ctx, unsigned long start,
 	if (atomic_read(&ctx->mmap_changing))
 		goto out_unlock;
 
+	/*
+	 * Validate every spanned VMA before changing any of them, so an
+	 * invalid range is rejected without changing part of it.
+	 */
 	err = -ENOENT;
 	for_each_vma_range(vmi, dst_vma, end) {
-
 		if (!userfaultfd_wp(dst_vma)) {
 			err = -ENOENT;
-			break;
+			goto out_unlock;
 		}
 
 		if (vma_is_hugetlb(dst_vma)) {
-			err = -EINVAL;
 			page_mask = vma_kernel_pagesize(dst_vma) - 1;
-			if ((start & page_mask) || (len & page_mask))
-				break;
+			if ((start & page_mask) || (len & page_mask)) {
+				err = -EINVAL;
+				goto out_unlock;
+			}
 		}
+		err = 0;
+	}
+	if (err)
+		goto out_unlock;
 
+	vma_iter_set(&vmi, start);
+	for_each_vma_range(vmi, dst_vma, end) {
 		_start = max(dst_vma->vm_start, start);
 		_end = min(dst_vma->vm_end, end);
 
-- 
2.53.0

Re: [PATCH 1/2] userfaultfd: validate the whole range before changing write protection

[Mike Rapoport]

On Mon, Nov 02, 2026 at 04:06:20PM +0530, Umang Pokhriyal wrote:
> mwriteprotect_range() checks each VMA as it walks the range.  When a later
> VMA is not registered for write protection, or is a hugetlb VMA the range
> is not aligned to, the call fails after the earlier VMAs were already
> changed.  When unprotecting, userfaultfd_writeprotect() then skips
> wake_userfault(), so threads blocked on those pages are not woken although
> the pages are writable again.
> 
> An unprivileged process can trigger this with a range that spans an
> unregistered VMA.  It hasn't been seen in practice; it was found while
> reviewing the userfaultfd range ioctls with an LLM.
> 
> Validate the whole range first, as mrwprotect_range() already does.
> 
> Fixes: a1b92a3f1498 ("mm/userfaultfd: support WP on multiple VMAs")
> Cc: stable@vger.kernel.org
> Signed-off-by: Umang Pokhriyal <umangpokhriyall@gmail.com>
> ---
>  mm/userfaultfd.c | 20 +++++++++++++++-----
>  1 file changed, 15 insertions(+), 5 deletions(-)
> 
> diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> index 87da92526..b8a9d9574 100644
> --- a/mm/userfaultfd.c
> +++ b/mm/userfaultfd.c
> @@ -1105,21 +1105,31 @@ static int mwriteprotect_range(struct userfaultfd_ctx *ctx, unsigned long start,
>  	if (atomic_read(&ctx->mmap_changing))
>  		goto out_unlock;
>  
> +	/*
> +	 * Validate every spanned VMA before changing any of them, so an
> +	 * invalid range is rejected without changing part of it.
> +	 */
>  	err = -ENOENT;

Nit: this assignment is now redundant, isn't it?

>  	for_each_vma_range(vmi, dst_vma, end) {
> -
>  		if (!userfaultfd_wp(dst_vma)) {
>  			err = -ENOENT;
> -			break;
> +			goto out_unlock;
>  		}
>  
>  		if (vma_is_hugetlb(dst_vma)) {
> -			err = -EINVAL;

Same here.

>  			page_mask = vma_kernel_pagesize(dst_vma) - 1;
> -			if ((start & page_mask) || (len & page_mask))
> -				break;
> +			if ((start & page_mask) || (len & page_mask)) {
> +				err = -EINVAL;
> +				goto out_unlock;
> +			}
>  		}
> +		err = 0;
> +	}
> +	if (err)
> +		goto out_unlock;

If we a breaking on any error to out_unlock, why do we need err = 0 and
this check?

>  
> +	vma_iter_set(&vmi, start);
> +	for_each_vma_range(vmi, dst_vma, end) {
>  		_start = max(dst_vma->vm_start, start);
>  		_end = min(dst_vma->vm_end, end);
>  
> -- 
> 2.53.0
> 

-- 
Sincerely yours,
Mike.

Re: [PATCH 1/2] userfaultfd: validate the whole range before changing write protection

[Umang Pokhriyal]

On Wed, Nov 4, 2026 at 12:24 AM Mike Rapoport <rppt@kernel.org> wrote:

> If we a breaking on any error to out_unlock, why do we need err = 0 and
> this check?

We don't, thanks. v2 drops them and the -ENOENT inside the loop. I kept
the one before the loop, because for a range with no VMA both loops are
skipped, and without it the call would return the -EAGAIN left over from
the mmap_changing check instead of -ENOENT.

I'll also drop the selftest.