On Wed, Jul 06, 2016 at 09:42:00AM +0100, Mel Gorman wrote:
<snip>
> > > >
> > > > If buffer_head is over limit, old logic force to reclaim highmem but
> > > > this zone_balanced logic will prevent it.
> > > >
> > >
> > > The old logic was always busted on 64-bit because is_highmem would always
> > > be 0. The original intent appears to be that buffer_heads_over_limit
> > > would release the buffers when pages went inactive. There are a number
> >
> > Yes but the difference is in old, it was handled both direct and background
> > reclaim once buffers_heads is over the limit but your change slightly
> > changs it so kswapd couldn't reclaim high zone if any eligible zone
> > is balanced. I don't know how big difference it can make but we saw
> > highmem buffer_head problems several times, IIRC. So, I just wanted
> > to notice it to you. whether it's handled or not, it's up to you.
> >
>
> The last time I remember buffer_heads_over_limit was an NTFS filesystem
> using small sub-page block sizes with a large highmem:lowmem ratio. If a
> similar situation is encountered then a test patch would be something like;
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index dc12af938a8d..a8ebd1871f16 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -3151,7 +3151,7 @@ static int balance_pgdat(pg_data_t *pgdat, int order,
> int classzone_idx)
> * zone was balanced even under extreme pressure when the
> * overall node may be congested.
> */
> - for (i = sc.reclaim_idx; i >= 0; i--) {
> + for (i = sc.reclaim_idx; i >= 0 && !buffer_heads_over_limit;
> i--) {
> zone = pgdat->node_zones + i;
> if (!populated_zone(zone))
> continue;
>
> I'm not going to go with it for now because buffer_heads_over_limit is not
> necessarily a problem unless lowmem is factor. We don't want background
> reclaim to go ahead unnecessarily just because buffer_heads_over_limit.
> It could be distinguished by only forcing reclaim to go ahead on systems
> with highmem.
If you don't think it's a problem, I don't want to insist on it because I don't
have any report/workload right now. Instead, please write some comment in there
for others to understand why kswapd is okay to ignore buffer_heads_over_limit
unlike direct reclaim. Such non-symmetric behavior is really hard to follow
without any description.
