Re: [PATCH v1 1/2] mm: Clear uffd-wp PTE/PMD state on mremap()

[Ryan Roberts <ryan.roberts@arm.com>]

I think there might be a bug in this after all...


On 07/01/2025 14:47, Ryan Roberts wrote:
> When mremap()ing a memory region previously registered with userfaultfd
> as write-protected but without UFFD_FEATURE_EVENT_REMAP, an
> inconsistency in flag clearing leads to a mismatch between the vma flags
> (which have uffd-wp cleared) and the pte/pmd flags (which do not have
> uffd-wp cleared). This mismatch causes a subsequent mprotect(PROT_WRITE)
> to trigger a warning in page_table_check_pte_flags() due to setting the
> pte to writable while uffd-wp is still set.
> 
> Fix this by always explicitly clearing the uffd-wp pte/pmd flags on any
> such mremap() so that the values are consistent with the existing
> clearing of VM_UFFD_WP. Be careful to clear the logical flag regardless
> of its physical form; a PTE bit, a swap PTE bit, or a PTE marker. Cover
> PTE, huge PMD and hugetlb paths.
> 
> Co-developed-by: Mikołaj Lenczewski <miko.lenczewski@arm.com>
> Signed-off-by: Mikołaj Lenczewski <miko.lenczewski@arm.com>
> Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
> Closes: https://lore.kernel.org/linux-mm/810b44a8-d2ae-4107-b665-5a42eae2d948@arm.com/
> Fixes: 63b2d4174c4a ("userfaultfd: wp: add the writeprotect API to userfaultfd ioctl")
> Cc: stable@vger.kernel.org
> ---
>  include/linux/userfaultfd_k.h | 12 ++++++++++++
>  mm/huge_memory.c              | 12 ++++++++++++
>  mm/hugetlb.c                  | 14 +++++++++++++-
>  mm/mremap.c                   | 32 +++++++++++++++++++++++++++++++-
>  4 files changed, 68 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/userfaultfd_k.h b/include/linux/userfaultfd_k.h
> index cb40f1a1d081..75342022d144 100644
> --- a/include/linux/userfaultfd_k.h
> +++ b/include/linux/userfaultfd_k.h
> @@ -247,6 +247,13 @@ static inline bool vma_can_userfault(struct vm_area_struct *vma,
>  	    vma_is_shmem(vma);
>  }
>  
> +static inline bool vma_has_uffd_without_event_remap(struct vm_area_struct *vma)
> +{
> +	struct userfaultfd_ctx *uffd_ctx = vma->vm_userfaultfd_ctx.ctx;
> +
> +	return uffd_ctx && (uffd_ctx->features & UFFD_FEATURE_EVENT_REMAP) == 0;
> +}
> +
>  extern int dup_userfaultfd(struct vm_area_struct *, struct list_head *);
>  extern void dup_userfaultfd_complete(struct list_head *);
>  void dup_userfaultfd_fail(struct list_head *);
> @@ -402,6 +409,11 @@ static inline bool userfaultfd_wp_async(struct vm_area_struct *vma)
>  	return false;
>  }
>  
> +static inline bool vma_has_uffd_without_event_remap(struct vm_area_struct *vma)
> +{
> +	return false;
> +}
> +
>  #endif /* CONFIG_USERFAULTFD */
>  
>  static inline bool userfaultfd_wp_use_markers(struct vm_area_struct *vma)
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index c89aed1510f1..2654a9548749 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -2212,6 +2212,16 @@ static pmd_t move_soft_dirty_pmd(pmd_t pmd)
>  	return pmd;
>  }
>  
> +static pmd_t clear_uffd_wp_pmd(pmd_t pmd)
> +{
> +	if (pmd_present(pmd))
> +		pmd = pmd_clear_uffd_wp(pmd);
> +	else if (is_swap_pmd(pmd))
> +		pmd = pmd_swp_clear_uffd_wp(pmd);
> +
> +	return pmd;
> +}
> +
>  bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
>  		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
>  {
> @@ -2250,6 +2260,8 @@ bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
>  			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
>  		}
>  		pmd = move_soft_dirty_pmd(pmd);
> +		if (vma_has_uffd_without_event_remap(vma))
> +			pmd = clear_uffd_wp_pmd(pmd);
>  		set_pmd_at(mm, new_addr, new_pmd, pmd);
>  		if (force_flush)
>  			flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 354eec6f7e84..cdbc55d5384f 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -5454,6 +5454,7 @@ static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
>  			  unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
>  			  unsigned long sz)
>  {
> +	bool need_clear_uffd_wp = vma_has_uffd_without_event_remap(vma);
>  	struct hstate *h = hstate_vma(vma);
>  	struct mm_struct *mm = vma->vm_mm;
>  	spinlock_t *src_ptl, *dst_ptl;
> @@ -5470,7 +5471,18 @@ static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
>  		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
>  
>  	pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
> -	set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
> +
> +	if (need_clear_uffd_wp && pte_marker_uffd_wp(pte))
> +		huge_pte_clear(mm, new_addr, dst_pte, sz);

This is checking if the source huge_pte is a uffd-wp marker and clearing the
destination if so. The destination could have previously held arbitrary valid
mappings, I guess?

But huge_pte_clear() does not call page_table_check_pte_clear(). So any previous
valid mapping will not have it's page_table_check ref count decremented?

I think it should be replaced with:
		huge_ptep_get_and_clear(mm, new_addr, dst_pte);

Since there is no huge_ptep_clear().

The tests I wrote are always mremapping into PROT_NONE space so they don't hit
this condition. If people agree this is a bug, I'll send out a fix.

Thanks,
Ryan

> +	else {
> +		if (need_clear_uffd_wp) {
> +			if (pte_present(pte))
> +				pte = huge_pte_clear_uffd_wp(pte);
> +			else if (is_swap_pte(pte))
> +				pte = pte_swp_clear_uffd_wp(pte);
> +		}
> +		set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
> +	}
>  
>  	if (src_ptl != dst_ptl)
>  		spin_unlock(src_ptl);
> diff --git a/mm/mremap.c b/mm/mremap.c
> index 60473413836b..cff7f552f909 100644
> --- a/mm/mremap.c
> +++ b/mm/mremap.c
> @@ -138,6 +138,7 @@ static int move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
>  		struct vm_area_struct *new_vma, pmd_t *new_pmd,
>  		unsigned long new_addr, bool need_rmap_locks)
>  {
> +	bool need_clear_uffd_wp = vma_has_uffd_without_event_remap(vma);
>  	struct mm_struct *mm = vma->vm_mm;
>  	pte_t *old_pte, *new_pte, pte;
>  	pmd_t dummy_pmdval;
> @@ -216,7 +217,18 @@ static int move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
>  			force_flush = true;
>  		pte = move_pte(pte, old_addr, new_addr);
>  		pte = move_soft_dirty_pte(pte);
> -		set_pte_at(mm, new_addr, new_pte, pte);
> +
> +		if (need_clear_uffd_wp && pte_marker_uffd_wp(pte))
> +			pte_clear(mm, new_addr, new_pte);
> +		else {
> +			if (need_clear_uffd_wp) {
> +				if (pte_present(pte))
> +					pte = pte_clear_uffd_wp(pte);
> +				else if (is_swap_pte(pte))
> +					pte = pte_swp_clear_uffd_wp(pte);
> +			}
> +			set_pte_at(mm, new_addr, new_pte, pte);
> +		}
>  	}
>  
>  	arch_leave_lazy_mmu_mode();
> @@ -278,6 +290,15 @@ static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
>  	if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
>  		return false;
>  
> +	/* If this pmd belongs to a uffd vma with remap events disabled, we need
> +	 * to ensure that the uffd-wp state is cleared from all pgtables. This
> +	 * means recursing into lower page tables in move_page_tables(), and we
> +	 * can reuse the existing code if we simply treat the entry as "not
> +	 * moved".
> +	 */
> +	if (vma_has_uffd_without_event_remap(vma))
> +		return false;
> +
>  	/*
>  	 * We don't have to worry about the ordering of src and dst
>  	 * ptlocks because exclusive mmap_lock prevents deadlock.
> @@ -333,6 +354,15 @@ static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
>  	if (WARN_ON_ONCE(!pud_none(*new_pud)))
>  		return false;
>  
> +	/* If this pud belongs to a uffd vma with remap events disabled, we need
> +	 * to ensure that the uffd-wp state is cleared from all pgtables. This
> +	 * means recursing into lower page tables in move_page_tables(), and we
> +	 * can reuse the existing code if we simply treat the entry as "not
> +	 * moved".
> +	 */
> +	if (vma_has_uffd_without_event_remap(vma))
> +		return false;
> +
>  	/*
>  	 * We don't have to worry about the ordering of src and dst
>  	 * ptlocks because exclusive mmap_lock prevents deadlock.