We set up per-cpu work structures for vmstat and schedule them on
each cpu when they go online only to re-schedule them on the general
work queue when they first run.
This doesn't seem right - how do we ever guarantee that vmstat_update runs
on all cpus? Either I've missed something or our vm stats are off and per-cpu
pages are not drained as frequently as we think they are.
Fix it by re-scheduling the work item on the same cpu it first ran on.
Tested on x86 on 8 way SMP VM.
Signed-off-by: Gilad Ben-Yossef <gilad@xxxxxxxxxxxxx>
CC: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
CC: Mel Gorman <mel@xxxxxxxxx>
CC: Christoph Lameter <cl@xxxxxxxxx>
CC: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
CC: Andi Kleen <ak@xxxxxxxxxxxxxxx>
CC: linux-mm@xxxxxxxxx
---
diff --git a/mm/vmstat.c b/mm/vmstat.c
index f600557..b396044 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -1144,8 +1144,10 @@ int sysctl_stat_interval __read_mostly = HZ;
static void vmstat_update(struct work_struct *w)
{
- refresh_cpu_vm_stats(smp_processor_id());
- schedule_delayed_work(&__get_cpu_var(vmstat_work),
+ int cpu = smp_processor_id();
+
+ refresh_cpu_vm_stats(cpu);
+ schedule_delayed_work_on(cpu, &__get_cpu_var(vmstat_work),
round_jiffies_relative(sysctl_stat_interval));
}
--
Gilad Ben-Yossef
Chief Coffee Drinker
gilad@xxxxxxxxxxxxx
Israel Cell: +972-52-8260388
US Cell: +1-973-8260388
http://benyossef.com
"If you take a class in large-scale robotics, can you end up in a
situation where the homework eats your dog?"
-- Jean-Baptiste Queru
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href