Re: [PATCH v2] mm/mmap: return early if it can't merge in vma_merge()

[Liam R. Howlett]

* Yajun Deng <yajun.deng@xxxxxxxxx> [240221 04:15]:
> In most cases, the range of the area is valid. But in do_mprotect_pkey(),
> the minimum value of end and vma->vm_end is passed to mprotect_fixup().
> This will lead to the end is less than the end of prev.
> 
> In this case, the curr will be NULL, but the next will be equal to the
> prev. So it will attempt to merge before, the vm_pgoff check will cause
> this case to fail.
> 
> To avoid the process described above and reduce unnecessary operations.
> Add a check to immediately return NULL if the end is less than the end of
> prev.

If it's only one caller, could we stop that caller instead of checking
an almost never case for all callers?  Would this better fit in
vma_modify()?  Although that's not just for this caller at this point.
Maybe there isn't a good place?

Or are there other reasons this may happen and is better done in this
function?

Often, this is called the "punch a hole" scenario; where an operation
creates two entries from the old data and either leaves an empty space
or fills the space with a new VMA.

> 
> Signed-off-by: Yajun Deng <yajun.deng@xxxxxxxxx>
> ---
> v2: remove the case label.
> v1: https://lore.kernel.org/all/20240218085028.3294332-1-yajun.deng@xxxxxxxxx/
> ---
>  mm/mmap.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index 0fccd23f056e..7668854d2246 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -890,6 +890,9 @@ static struct vm_area_struct
>  	if (vm_flags & VM_SPECIAL)
>  		return NULL;
>  
> +	if (prev && end < prev->vm_end)
> +		return NULL;
> +
>  	/* Does the input range span an existing VMA? (cases 5 - 8) */
>  	curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
>  
> -- 
> 2.25.1
>