Re: [PATCH] mm: khugepaged: fix memory leak in collapse_file rollback path

[Dev Jain <dev.jain@arm.com>]

On 23/11/25 6:57 pm, Shardul Bankar wrote:
> When collapse_file() fails after xas_create_range() succeeds, the
> rollback path does not clean up pre-allocated XArray nodes stored in
> xas->xa_alloc. These nodes are allocated by xas_nomem() when
> xas_create() fails with GFP_NOWAIT and need to be freed.
>
> The leak occurs because:
> 1. xas_create_range() may call xas_nomem() which allocates a node
>     and stores it in xas->xa_alloc
> 2. If the collapse operation fails later, the rollback path jumps
>     to the 'rollback:' label
> 3. The rollback path cleans up folios but does not call xas_destroy()
>     to free the pre-allocated nodes in xas->xa_alloc
>
> Fix this by calling xas_destroy(&xas) at the beginning of the rollback
> path to free any pre-allocated nodes. This is safe because xas_destroy()
> only frees nodes in xas->xa_alloc that were never inserted into the
> XArray tree.
>
> Note: This fixes the leak of pre-allocated nodes. A separate fix will
> be needed to clean up empty nodes that were inserted into the tree by
> xas_create_range() but never populated.

No "fix" is needed in this case, the empty nodes are there in the tree
and there is no leak.

>
> Link: https://syzkaller.appspot.com/bug?id=a274d65fc733448ed518ad15481ed575669dd98c
> Fixes: cae106dd67b9 ("mm/khugepaged: refactor collapse_file control flow")
> Signed-off-by: Shardul Bankar <shardul.b@mpiricsoftware.com>
> ---
>   mm/khugepaged.c | 1 +
>   1 file changed, 1 insertion(+)
>
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index abe54f0043c7..f2fe7924afa0 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -2230,6 +2230,7 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
>   	goto out;
>   
>   rollback:
> +	xas_destroy(&xas);
>   	/* Something went wrong: roll back page cache changes */
>   	if (nr_none) {
>   		xas_lock_irq(&xas);

I believe that the correct explanation of this patch is - after xas_nomem, if someone else
expands the tree and the retry for xas_create_range() trivially succeeds without using the
node allocated from xas_nomem, then we need to free that node. Although now I am worried
about other users of xas_nomem() which do not call xas_destroy() on failure path...

I think this should work,

Reviewed-by: Dev Jain <dev.jain@arm.com>