> If reclaim fails to make sufficient progress, the priority is raised.
> Once the priority is higher, kswapd starts waiting on congestion.
> However, on systems with large numbers of high-order atomics due to
> crappy network cards, it's important that kswapd keep working in
> parallel to save their sorry ass.
>
> This patch takes into account the order kswapd is reclaiming at before
> waiting on congestion. The higher the order, the longer it is before
> kswapd considers itself to be in trouble. The impact is that kswapd
> works harder in parallel rather than depending on direct reclaimers or
> atomic allocations to fail.
>
> Signed-off-by: Mel Gorman <mel@xxxxxxxxx>
> ---
> mm/vmscan.c | 14 ++++++++++++--
> 1 files changed, 12 insertions(+), 2 deletions(-)
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index ffa1766..5e200f1 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1946,7 +1946,7 @@ static int sleeping_prematurely(int order, long remaining)
> static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
> {
> int all_zones_ok;
> - int priority;
> + int priority, congestion_priority;
> int i;
> unsigned long total_scanned;
> struct reclaim_state *reclaim_state = current->reclaim_state;
> @@ -1967,6 +1967,16 @@ static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
> */
> int temp_priority[MAX_NR_ZONES];
>
> + /*
> + * When priority reaches congestion_priority, kswapd will sleep
> + * for a short time while congestion clears. The higher the
> + * order being reclaimed, the less likely kswapd will go to
> + * sleep as high-order allocations are harder to reclaim and
> + * stall direct reclaimers longer
> + */
> + congestion_priority = DEF_PRIORITY - 2;
> + congestion_priority -= min(congestion_priority, sc.order);
This calculation mean
sc.order congestion_priority scan-pages
---------------------------------------------------------
0 10 1/1024 * zone-mem
1 9 1/512 * zone-mem
2 8 1/256 * zone-mem
3 7 1/128 * zone-mem
4 6 1/64 * zone-mem
5 5 1/32 * zone-mem
6 4 1/16 * zone-mem
7 3 1/8 * zone-mem
8 2 1/4 * zone-mem
9 1 1/2 * zone-mem
10 0 1 * zone-mem
11+ 0 1 * zone-mem
I feel this is too agressive. The intention of this congestion_wait()
is to prevent kswapd use 100% cpu time. but the above promotion seems
break it.
example,
ia64 have 256MB hugepage (i.e. order=14). it mean kswapd never sleep.
example2,
order-3 (i.e. PAGE_ALLOC_COSTLY_ORDER) makes one of most inefficent
reclaim, because it doesn't use lumpy recliam.
I've seen 128GB size zone, it mean 1/128 = 1GB. oh well, kswapd definitely
waste cpu time 100%.
> +
> loop_again:
> total_scanned = 0;
> sc.nr_reclaimed = 0;
> @@ -2092,7 +2102,7 @@ loop_again:
> * OK, kswapd is getting into trouble. Take a nap, then take
> * another pass across the zones.
> */
> - if (total_scanned && priority < DEF_PRIORITY - 2)
> + if (total_scanned && priority < congestion_priority)
> congestion_wait(BLK_RW_ASYNC, HZ/10);
Instead, How about this?
---
mm/vmscan.c | 13 ++++++++++++-
1 files changed, 12 insertions(+), 1 deletions(-)
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 64e4388..937e90d 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1938,6 +1938,7 @@ static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
* free_pages == high_wmark_pages(zone).
*/
int temp_priority[MAX_NR_ZONES];
+ int has_under_min_watermark_zone = 0;
loop_again:
total_scanned = 0;
@@ -2057,6 +2058,15 @@ loop_again:
if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
sc.may_writepage = 1;
+
+ /*
+ * We are still under min water mark. it mean we have
+ * GFP_ATOMIC allocation failure risk. Hurry up!
+ */
+ if (!zone_watermark_ok(zone, order, min_wmark_pages(zone),
+ end_zone, 0))
+ has_under_min_watermark_zone = 1;
+
}
if (all_zones_ok)
break; /* kswapd: all done */
@@ -2064,7 +2074,8 @@ loop_again:
* OK, kswapd is getting into trouble. Take a nap, then take
* another pass across the zones.
*/
- if (total_scanned && priority < DEF_PRIORITY - 2)
+ if (total_scanned && (priority < DEF_PRIORITY - 2) &&
+ !has_under_min_watermark_zone)
congestion_wait(BLK_RW_ASYNC, HZ/10);
/*
--
1.6.2.5
--
To unsubscribe from this list: send the line "unsubscribe kernel-testers" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
