Re: [RFC PATCH 1/1] mm/vmalloc.c: correct logic errors when insert vmap_area

From: Michal Hocko
Date: Wed Oct 12 2016 - 10:46:24 EST


[Let's CC Nick who has written this code]

On Wed 12-10-16 22:30:13, zijun_hu wrote:
> From: zijun_hu <zijun_hu@xxxxxxx>
>
> the KVA allocator organizes vmap_areas allocated by rbtree. in order to
> insert a new vmap_area @i_va into the rbtree, walk around the rbtree from
> root and compare the vmap_area @t_va met on the rbtree against @i_va; walk
> toward the left branch of @t_va if @i_va is lower than @t_va, and right
> branch if higher, otherwise handle this error case since @i_va has overlay
> with @t_va; however, __insert_vmap_area() don't follow the desired
> procedure rightly, moreover, it includes a meaningless else if condition
> and a redundant else branch as shown by comments in below code segments:
> static void __insert_vmap_area(struct vmap_area *va)
> {
> as a internal interface parameter, we assume vmap_area @va has nonzero size
> ...
> if (va->va_start < tmp->va_end)
> p = &(*p)->rb_left;
> else if (va->va_end > tmp->va_start)
> p = &(*p)->rb_right;
> this else if condition is always true and meaningless due to
> va->va_end > va->va_start >= tmp_va->va_end > tmp_va->va_start normally
> else
> BUG();
> this BUG() is meaningless too due to never be reached normally
> ...
> }
>
> it looks like the else if condition and else branch are canceled. no errors
> are caused since the vmap_area @va to insert as a internal interface
> parameter doesn't have overlay with any one on the rbtree normally. however
> __insert_vmap_area() looks weird and really has several logic errors as
> pointed out above when it is viewed as a separate function.

I have tried to read this several times but I am completely lost to
understand what the actual bug is and how it causes vmap_area sorting to
misbehave. So is this a correctness issue, performance improvement or
theoretical fix for an incorrect input?

> fix by walking around vmap_area rbtree as described above to insert
> a vmap_area.
>
> BTW, (va->va_end == tmp_va->va_start) is consider as legal case since it
> indicates vmap_area @va left neighbors with @tmp_va tightly.
>
> Fixes: db64fe02258f ("mm: rewrite vmap layer")
> Signed-off-by: zijun_hu <zijun_hu@xxxxxxx>
> ---
> mm/vmalloc.c | 8 ++++----
> 1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 5daf3211b84f..8b80931654b7 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -321,10 +321,10 @@ static void __insert_vmap_area(struct vmap_area *va)
>
> parent = *p;
> tmp_va = rb_entry(parent, struct vmap_area, rb_node);
> - if (va->va_start < tmp_va->va_end)
> - p = &(*p)->rb_left;
> - else if (va->va_end > tmp_va->va_start)
> - p = &(*p)->rb_right;
> + if (va->va_end <= tmp_va->va_start)
> + p = &parent->rb_left;
> + else if (va->va_start >= tmp_va->va_end)
> + p = &parent->rb_right;
> else
> BUG();
> }
> --
> 1.9.1

--
Michal Hocko
SUSE Labs