Re: [PATCH v3 3/4] mm/mmap/vma_merge: explicitly assign res, vma, extend invariants

[Vlastimil Babka]

On 3/22/23 08:13, Lorenzo Stoakes wrote:
> Previously, vma was an uninitialised variable which was only definitely
> assigned as a result of the logic covering all possible input cases - for
> it to have remained uninitialised, prev would have to be NULL, and next
> would _have_ to be mergeable.
> 
> The value of res defaults to NULL, so we can neatly eliminate the
> assignment to res and vma in the if (prev) block and ensure that both res
> and vma are both explicitly assigned, by just setting both to prev.
> 
> In addition we add an explanation as to under what circumstances both might
> change, and since we absolutely do rely on addr == curr->vm_start should
> curr exist, assert that this is the case.

Hm replied to v2 of this, sorry, so let me repeat that here:

> Signed-off-by: Lorenzo Stoakes <lstoakes@xxxxxxxxx>

Reviewed-by: Vlastimil Babka <vbabka@xxxxxxx>

Nit suggestion below.


> ---
>  mm/mmap.c | 19 ++++++++++++++-----
>  1 file changed, 14 insertions(+), 5 deletions(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index dbdbb92493b2..2a4f63716231 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -911,7 +911,7 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>  {
>  	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
>  	pgoff_t vma_pgoff;
> -	struct vm_area_struct *curr, *next, *res = NULL;
> +	struct vm_area_struct *curr, *next, *res;
>  	struct vm_area_struct *vma, *adjust, *remove, *remove2;
>  	int err = -1;
>  	bool merge_prev = false;
> @@ -939,14 +939,18 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>  	else
>  		next = NULL;		/* case 5 */
>  
> -	/* verify some invariant that must be enforced by the caller */
> +	/*
> +	 * By default, we return prev. Cases 3, 4, 8 will instead return next
> +	 * and cases 3, 8 will also update vma to point at next.
> +	 */
> +	res = vma = prev;

Later in the function there's a line:

	remove = remove2 = adjust = NULL;

Now it would make sense to move it up here?

> +
> +	/* Verify some invariant that must be enforced by the caller. */
>  	VM_WARN_ON(prev && addr <= prev->vm_start);
> -	VM_WARN_ON(curr && end > curr->vm_end);
> +	VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
>  	VM_WARN_ON(addr >= end);
>  
>  	if (prev) {
> -		res = prev;
> -		vma = prev;
>  		vma_start = prev->vm_start;
>  		vma_pgoff = prev->vm_pgoff;
>  		/* Can we merge the predecessor? */
> @@ -957,6 +961,7 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>  			vma_prev(vmi);
>  		}
>  	}
> +
>  	/* Can we merge the successor? */
>  	if (next && mpol_equal(policy, vma_policy(next)) &&
>  	    can_vma_merge_before(next, vm_flags,
> @@ -997,6 +1002,10 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>  			adj_start = -(prev->vm_end - addr);
>  			err = dup_anon_vma(next, prev);
>  		} else {
> +			/*
> +			 * Note that cases 3 and 8 are the ONLY ones where prev
> +			 * is permitted to be (but is not necessarily) NULL.
> +			 */
>  			vma = next;			/* case 3 */
>  			vma_start = addr;
>  			vma_end = next->vm_end;