Re: [PATCH] mm/khugepaged: Fix ->anon_vma race

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Wed, Jan 11, 2023 at 02:33:51PM +0100, Jann Horn wrote:
> If an ->anon_vma is attached to the VMA, collapse_and_free_pmd() requires
> it to be locked. retract_page_tables() bails out if an ->anon_vma is
> attached, but does this check before holding the mmap lock (as the comment
> above the check explains).
> 
> If we racily merge an existing ->anon_vma (shared with a child process)
> from a neighboring VMA, subsequent rmap traversals on pages belonging to
> the child will be able to see the page tables that we are concurrently
> removing while assuming that nothing else can access them.
> 
> Repeat the ->anon_vma check once we hold the mmap lock to ensure that there
> really is no concurrent page table access.
> 
> Reported-by: Zach O'Keefe <zokeefe@google.com>
> Fixes: f3f0e1d2150b ("khugepaged: add support of collapse for tmpfs/shmem pages")
> Cc: stable@vger.kernel.org
> Signed-off-by: Jann Horn <jannh@google.com>
> ---
> zokeefe@ pointed out to me that the current code (after my last round of patches)
> can hit a lockdep assert by racing, and after staring at it a bit I've
> convinced myself that this is a real, preexisting bug.
> (I haven't written a reproducer for it though. One way to hit it might be
> something along the lines of:
> 
>  - set up a process A with a private-file-mapping VMA V1
>  - let A fork() to create process B, thereby copying V1 in A to V1' in B
>  - let B extend the end of V1'
>  - let B put some anon pages into the extended part of V1'

At this point V1' gets it's own ->anon_vma, not connected to V1, right?

>  - let A map a new private-file-mapping VMA V2 directly behind V1, without
>    an anon_vma
> [race begins here]
>   - in A's thread 1: begin retract_page_tables() on V2, run through first
>     ->anon_vma check
>   - in A's thread 2: run __anon_vma_prepare() on V2 and ensure that it
>     merges the anon_vma of V1 (which implies V1 and V2 must be mapping the
>     same file at compatible offsets)
>   - in B: trigger rmap traversal on anon page in V1'

I don't follow the race. rmap on V1' will not reach V1.

>  mm/khugepaged.c | 14 +++++++++++++-
>  1 file changed, 13 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 5cb401aa2b9d..0bfed37f3a3b 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1644,7 +1644,7 @@ static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff,
>  		 * has higher cost too. It would also probably require locking
>  		 * the anon_vma.
>  		 */
> -		if (vma->anon_vma) {
> +		if (READ_ONCE(vma->anon_vma)) {
>  			result = SCAN_PAGE_ANON;
>  			goto next;
>  		}

This makes perfect sense. At least for readability. But I think
false-negative should not lead to bad results.

> @@ -1672,6 +1672,18 @@ static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff,
>  		result = SCAN_PTE_MAPPED_HUGEPAGE;
>  		if ((cc->is_khugepaged || is_target) &&
>  		    mmap_write_trylock(mm)) {
> +			/*
> +			 * Re-check whether we have an ->anon_vma, because
> +			 * collapse_and_free_pmd() requires that either no
> +			 * ->anon_vma exists or the anon_vma is locked.
> +			 * We already checked ->anon_vma above, but that check
> +			 * is racy because ->anon_vma can be populated under the
> +			 * mmap lock in read mode.
> +			 */
> +			if (vma->anon_vma) {
> +				result = SCAN_PAGE_ANON;
> +				goto unlock_next;
> +			}

This is totally wrong direction. Or I don't understand the race.

At this point we already paid nearly all price of of pagetable retraction.
I don't see any correctness reason to stop here, except for the assert.

I think lockdep assert in collapse_and_free_pmd() is wrong and has to be
dropped.

>  			/*
>  			 * When a vma is registered with uffd-wp, we can't
>  			 * recycle the pmd pgtable because there can be pte
> 
> base-commit: 7dd4b804e08041ff56c88bdd8da742d14b17ed25
> -- 
> 2.39.0.314.g84b9a713c41-goog
> 

-- 
  Kiryl Shutsemau / Kirill A. Shutemov