Re: [RFC][PATCH] lib: generic percpu counter array

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Thu, 5 Nov 2009 10:20:18 -0500 (EST)
Christoph Lameter <cl@linux-foundation.org> wrote:

> On Thu, 5 Nov 2009, KAMEZAWA Hiroyuki wrote:
> 
> > Anothter major percpu coutner is vm_stat[]. This patch implements
> > vm_stat[] style counter array in lib/percpu_counter.c
> > This is designed for introducing vm_stat[] style counter to memcg,
> > but maybe useful for other people. By using this, counter array
> > using percpu can be implemented easily in compact structure.
> 
> 
> Note that vm_stat support was written that way because we have extreme
> space constraints due to the need to keep statistics per zone and per cpu
> and avoid cache line pressure that would result through the use of big
> integer arrays per zone and per cpu. For a large number of zones and cpus
> this is desastrous.
> 
> If you only need to keep statistics per cpu for an entity then the vmstat
> approach is overkill. A per cpu allocation of a counter is enough.
> 
counter per memcg is required.
Memcg uses its own one but I want to remove it. (it doesn't consider memory
placement.)
What I can use under /lib is percpu_counter, but it's really overkill.

My concern on pure percpu counter is "read" side.
Now, we read counters only via status file and sometimes vmscan will read it.
For supporting dirty_ratio, we need to read them more.
I'll check I can move it to pure percpu counter as you do in mm_counters and
see how read side is affected by for_each_possible_cpu(). Anyway, it's
better than current one.

Thanks,
-Kame

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>