* false NUMA OOM
@ 2002-09-17 2:50 ` William Lee Irwin III
0 siblings, 0 replies; 6+ messages in thread
From: William Lee Irwin III @ 2002-09-17 2:50 UTC (permalink / raw)
To: linux-mm; +Cc: linux-kernel, akpm
Well, there's an obvious problem. shrink_caches() hammers out_of_memory()
when it has only looked at a single node. Something like this might help.
Totally untested. Problem discovered during 2 simultaneous dbench 512's
on separate 12GB tmpfs fs's on a 32x NUMA-Q with 32GB of RAM.
Against 2.5.35.
Bill
--- mm/vmscan.c.orig 2002-09-16 19:02:11.000000000 -0700
+++ mm/vmscan.c 2002-09-16 19:07:50.000000000 -0700
@@ -519,18 +519,24 @@
shrink_caches(struct zone *classzone, int priority,
int gfp_mask, int nr_pages)
{
+ pg_data_t *pgdat;
struct zone *first_classzone;
struct zone *zone;
+ int type;
first_classzone = classzone->zone_pgdat->node_zones;
- zone = classzone;
- while (zone >= first_classzone && nr_pages > 0) {
- if (zone->free_pages <= zone->pages_high) {
- nr_pages = shrink_zone(zone, priority,
- gfp_mask, nr_pages);
+ for (type = classzone - first_classzone; type >= 0; --type)
+ for_each_pgdat(pgdat) {
+ zone = pgdat->node_zones + type;
+ if (!zone->size)
+ continue;
+ if (zone->free_pages <= zone->pages_high)
+ nr_pages = shrink_zone(zone, priority,
+ gfp_mask, nr_pages);
+ if (nr_pages <= 0)
+ return nr_pages;
}
- zone--;
- }
+
return nr_pages;
}
^ permalink raw reply [flat|nested] 6+ messages in thread* false NUMA OOM
@ 2002-09-17 2:50 ` William Lee Irwin III
0 siblings, 0 replies; 6+ messages in thread
From: William Lee Irwin III @ 2002-09-17 2:50 UTC (permalink / raw)
To: linux-mm; +Cc: linux-kernel, akpm
Well, there's an obvious problem. shrink_caches() hammers out_of_memory()
when it has only looked at a single node. Something like this might help.
Totally untested. Problem discovered during 2 simultaneous dbench 512's
on separate 12GB tmpfs fs's on a 32x NUMA-Q with 32GB of RAM.
Against 2.5.35.
Bill
--- mm/vmscan.c.orig 2002-09-16 19:02:11.000000000 -0700
+++ mm/vmscan.c 2002-09-16 19:07:50.000000000 -0700
@@ -519,18 +519,24 @@
shrink_caches(struct zone *classzone, int priority,
int gfp_mask, int nr_pages)
{
+ pg_data_t *pgdat;
struct zone *first_classzone;
struct zone *zone;
+ int type;
first_classzone = classzone->zone_pgdat->node_zones;
- zone = classzone;
- while (zone >= first_classzone && nr_pages > 0) {
- if (zone->free_pages <= zone->pages_high) {
- nr_pages = shrink_zone(zone, priority,
- gfp_mask, nr_pages);
+ for (type = classzone - first_classzone; type >= 0; --type)
+ for_each_pgdat(pgdat) {
+ zone = pgdat->node_zones + type;
+ if (!zone->size)
+ continue;
+ if (zone->free_pages <= zone->pages_high)
+ nr_pages = shrink_zone(zone, priority,
+ gfp_mask, nr_pages);
+ if (nr_pages <= 0)
+ return nr_pages;
}
- zone--;
- }
+
return nr_pages;
}
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: false NUMA OOM
2002-09-17 2:50 ` William Lee Irwin III
@ 2002-09-17 3:17 ` Andrew Morton
-1 siblings, 0 replies; 6+ messages in thread
From: Andrew Morton @ 2002-09-17 3:17 UTC (permalink / raw)
To: William Lee Irwin III; +Cc: linux-mm, linux-kernel, akpm
William Lee Irwin III wrote:
>
> ...
> + for (type = classzone - first_classzone; type >= 0; --type)
> + for_each_pgdat(pgdat) {
> + zone = pgdat->node_zones + type;
Well you'd want to start with (and prefer) the local node's zones?
I'm also wondering whether one shouldn't just poke a remote kswapd
and wait.
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: false NUMA OOM
@ 2002-09-17 3:17 ` Andrew Morton
0 siblings, 0 replies; 6+ messages in thread
From: Andrew Morton @ 2002-09-17 3:17 UTC (permalink / raw)
To: William Lee Irwin III; +Cc: linux-mm, linux-kernel, akpm
William Lee Irwin III wrote:
>
> ...
> + for (type = classzone - first_classzone; type >= 0; --type)
> + for_each_pgdat(pgdat) {
> + zone = pgdat->node_zones + type;
Well you'd want to start with (and prefer) the local node's zones?
I'm also wondering whether one shouldn't just poke a remote kswapd
and wait.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: false NUMA OOM
2002-09-17 3:17 ` Andrew Morton
@ 2002-09-17 3:29 ` William Lee Irwin III
-1 siblings, 0 replies; 6+ messages in thread
From: William Lee Irwin III @ 2002-09-17 3:29 UTC (permalink / raw)
To: Andrew Morton; +Cc: linux-mm, linux-kernel, akpm
William Lee Irwin III wrote:
+ for (type = classzone - first_classzone; type >= 0; --type)
+ for_each_pgdat(pgdat) {
+ zone = pgdat->node_zones + type;
On Mon, Sep 16, 2002 at 08:17:03PM -0700, Andrew Morton wrote:
> Well you'd want to start with (and prefer) the local node's zones?
> I'm also wondering whether one shouldn't just poke a remote kswapd
> and wait.
I just sort of rearranged what was already there so it wouldn't die
quite so blatantly (i.e. minimal fix). Those are also sound methods.
Cheers,
Bill
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: false NUMA OOM
@ 2002-09-17 3:29 ` William Lee Irwin III
0 siblings, 0 replies; 6+ messages in thread
From: William Lee Irwin III @ 2002-09-17 3:29 UTC (permalink / raw)
To: Andrew Morton; +Cc: linux-mm, linux-kernel, akpm
William Lee Irwin III wrote:
+ for (type = classzone - first_classzone; type >= 0; --type)
+ for_each_pgdat(pgdat) {
+ zone = pgdat->node_zones + type;
On Mon, Sep 16, 2002 at 08:17:03PM -0700, Andrew Morton wrote:
> Well you'd want to start with (and prefer) the local node's zones?
> I'm also wondering whether one shouldn't just poke a remote kswapd
> and wait.
I just sort of rearranged what was already there so it wouldn't die
quite so blatantly (i.e. minimal fix). Those are also sound methods.
Cheers,
Bill
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2002-09-17 3:29 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2002-09-17 2:50 false NUMA OOM William Lee Irwin III
2002-09-17 2:50 ` William Lee Irwin III
2002-09-17 3:17 ` Andrew Morton
2002-09-17 3:17 ` Andrew Morton
2002-09-17 3:29 ` William Lee Irwin III
2002-09-17 3:29 ` William Lee Irwin III
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.