Re: [PATCH] mm/vmscan.c: avoid recording the original scan targets in shrink_lruvec()

[Chen Yucong <slaoub@gmail.com>]

On Mon, 2014-06-09 at 21:27 +0800, Chen Yucong wrote:
> Via https://lkml.org/lkml/2013/4/10/334 , we can find that recording the
> original scan targets introduces extra 40 bytes on the stack. This patch
> is able to avoid this situation and the call to memcpy(). At the same time,
> it does not change the relative design idea.
> 
> ratio = original_nr_file / original_nr_anon;
> 
> If (nr_file > nr_anon), then ratio = (nr_file - x) / nr_anon.
>  x = nr_file - ratio * nr_anon;
> 
> if (nr_file <= nr_anon), then ratio = nr_file / (nr_anon - x).
>  x = nr_anon - nr_file / ratio;
> 
Hi Andrew Morton,

I think the patch
 
[PATCH]
mm-vmscanc-avoid-recording-the-original-scan-targets-in-shrink_lruvec-fix.patch

which I committed should be discarded. Because It have some critical
defects.
    1) If we want to solve the divide-by-zero and unfair problems, it
needs to two variables for recording the ratios.
 
    2) For "x = nr_file - ratio * nr_anon", the "x" is likely to be a
negative number. we can assume:

      nr[LRU_ACTIVE_FILE] = 30
      nr[LRU_INACTIVE_FILE] = 30
      nr[LRU_ACTIVE_ANON] = 0
      nr[LRU_INACTIVE_ANON] = 40

      ratio = 60/40 = 3/2

When the value of (nr_reclaimed < nr_to_reclaim) become false, there are
the following results:
      nr[LRU_ACTIVE_FILE] = 15
      nr[LRU_INACTIVE_FILE] = 15
      nr[LRU_ACTIVE_ANON] = 0
      nr[LRU_INACTIVE_ANON] = 25
 
      nr_file = 30
      nr_anon = 25

      x = 30 - 25 * (3/2) = 30 - 37.5 = -7.5.

The result is too terrible. 
   
   3) This method is less accurate than the original, especially for the
qualitative difference between FILE and ANON that is very small.

thx!
cyc