Re: [PATCH] mm/userfaultfd: clear uffd-wp PTE state when re-registering without WP

[Andrew Morton]

On Mon,  1 Jun 2026 16:26:09 +0800 Jianhui Zhou <jianhuizzzzz@xxxxxxxxx> wrote:

> UFFDIO_REGISTER can be issued on a range that is already registered in
> the same userfaultfd context, replacing the VMA's userfaultfd tracking
> mode.  For example, a range can be registered with
> UFFDIO_REGISTER_MODE_WP and later re-registered with
> UFFDIO_REGISTER_MODE_MISSING.
> 
> When the second registration removes VM_UFFD_WP, the VMA flags are
> updated but existing uffd-wp state in page-table entries is left behind.
> That stale state can survive in swap PTEs.  On swapin, do_swap_page()
> restores _PAGE_UFFD_WP from the swap PTE and can then install a writable
> PTE, triggering page_table_check:
> 
>   pte_uffd_wp(pte) && pte_write(pte)
> 
> Handle removal of WP mode through UFFDIO_REGISTER the same way as
> UFFDIO_UNREGISTER: resolve the per-PTE uffd-wp state before dropping
> VM_UFFD_WP from the VMA.
> 
> Also make the same-context fast path require an exact UFFD mode match.
> The old subset check treats MISSING|WP -> MISSING as a no-op, even though
> WP mode is being removed.

Thanks.

> Fixes: f45ec5ff16a7 ("userfaultfd: wp: support swap and page migration")
> Reported-by: syzbot+18d274a59b87cf80e86d@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://syzkaller.appspot.com/bug?extid=18d274a59b87cf80e86d

So we have a WARN which can be triggered by unprivileged userspace.

The patch arrived during -rc6 when people were reviewing and testing
material for the next merge window, rather than looking at new changes.

> --- a/mm/userfaultfd.c
> +++ b/mm/userfaultfd.c
> @@ -2153,13 +2153,21 @@ int userfaultfd_register_range(struct userfaultfd_ctx *ctx,
>  		 * userfaultfd and with the right tracking mode too.
>  		 */
>  		if (vma->vm_userfaultfd_ctx.ctx == ctx &&
> -		    vma_test_all_mask(vma, vma_flags))
> +		    (vma->vm_flags & __VM_UFFD_FLAGS) == vm_flags)
>  			goto skip;
>  
>  		if (vma->vm_start > start)
>  			start = vma->vm_start;
>  		vma_end = min(end, vma->vm_end);
>  
> +		/*
> +		 * Re-registering into the same userfaultfd can remove WP mode.
> +		 * Clear any per-PTE uffd-wp state before dropping VM_UFFD_WP,
> +		 * matching the UFFDIO_UNREGISTER cleanup semantics.
> +		 */
> +		if (userfaultfd_wp(vma) && !(vm_flags & VM_UFFD_WP))
> +			uffd_wp_range(vma, start, vma_end - start, false);
> +
>  		new_vma_flags = vma->flags;
>  		vma_flags_clear_mask(&new_vma_flags, __VMA_UFFD_FLAGS);
>  		vma_flags_set_mask(&new_vma_flags, vma_flags);

I'll queue this for 7.3-rc1, with a cc:stable.  f45ec5ff16a7 was 6
years ago, so there's no rush here.

I don't know if this is the best fix, so I'll await reviewer input
before taking it any further.