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

From: Hugh Dickins
Date: Sun Jun 15 2014 - 20:52:06 EST


On Wed, 11 Jun 2014, Chen Yucong wrote:
> On Tue, 2014-06-10 at 16:33 -0700, Andrew Morton wrote:
> > > break;
> > >
> > > if (nr_file > nr_anon) {
> > > - unsigned long scan_target =
> > targets[LRU_INACTIVE_ANON] +
> > >
> > - targets[LRU_ACTIVE_ANON]
> > + 1;
> > > + nr_to_scan = nr_file - ratio * nr_anon;
> > > + percentage = nr[LRU_FILE] * 100 / nr_file;
> >
> > here, nr_file and nr_anon are derived from the contents of nr[]. But
> > nr[] was modified in the for_each_evictable_lru() loop, so its
> > contents
> > now may differ from what was in targets[]?
>
> nr_to_scan is used for recording the number of pages that should be
> scanned to keep original *ratio*.
>
> We can assume that the value of (nr_file > nr_anon) is true, nr_to_scan
> should be distribute to nr[LRU_ACTIVE_FILE] and nr[LRU_INACTIVE_FILE] in
> proportion.
>
> nr_file = nr[LRU_ACTIVE_FILE] + nr[LRU_INACTIVE_FILE];
> percentage = nr[LRU_FILE] / nr_file;
>
> Note that in comparison with *old* percentage, the "new" percentage has
> the different meaning. It is just used to divide nr_so_scan pages
> appropriately.

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

I have not reviewed your logic at all, but soon hit a divide-by-zero
crash on mmotm: it needs some such fix as below.

Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
---

mm/vmscan.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)

--- mmotm/mm/vmscan.c 2014-06-12 17:46:36.632008452 -0700
+++ linux/mm/vmscan.c 2014-06-12 18:55:18.832425713 -0700
@@ -2122,11 +2122,12 @@ static void shrink_lruvec(struct lruvec
nr_to_scan = nr_file - ratio * nr_anon;
percentage = nr[LRU_FILE] * 100 / nr_file;
lru = LRU_BASE;
- } else {
+ } else if (ratio) {
nr_to_scan = nr_anon - nr_file / ratio;
percentage = nr[LRU_BASE] * 100 / nr_anon;
lru = LRU_FILE;
- }
+ } else
+ break;

/* Stop scanning the smaller of the LRU */
nr[lru] = 0;
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/