Re: [PATCH 2.6.14-rc2] fix incorrect mm->hiwater_vm and mm->hiwater_rss

[Frank van Maarseveen <frankvm@frankvm.com>]

On Wed, Sep 21, 2005 at 08:26:28AM -0700, Christoph Lameter wrote:
> On Wed, 21 Sep 2005, Hugh Dickins wrote:
> 
> > On Wed, 21 Sep 2005, Frank van Maarseveen wrote:
> > > This fixes a post 2.6.11 regression in maintaining the mm->hiwater_* counters.
> > 
> > It would be a good idea to CC Christoph Lameter, who I believe was the
> > one who very intentionally moved most of these updates out to timer tick.
> 
> Right and we need to include Jay who introduced these in the first place.
> 
> What is the reason to move these counters back into the performance 
> critical vm paths? We agreed that some inaccuracy in these numbers was 
> acceptable.

Try running the script below after exposing hiwater_vm via
/proc/pid/status as "VmPeak:" (see my other mail in this thread).

========
#!/bin/sh

for f in /proc/*/status
do
	awk '
		BEGIN{ state = 0; }
		/^Name:/	{ name = $2;	++state; }
		/^Pid:/		{ pid = $2;	++state; }
		/^VmPeak:/	{ x = $2;	++state; }
		/^VmSize:/	{ y = $2;	++state; }
		END{ if (state == 4 && x < y)
			printf("pid %d (%s): %d < %d\n", pid, name, x, y);
		}
	    ' <$f
done
========

On 2.6.12.2 it prints:

pid 13170 (sh): 3312 < 3336
pid 16621 (sh): 2916 < 2996
pid 1696 (getty): 1576 < 1580
pid 2338 (getty): 1576 < 1580
pid 2344 (getty): 1444 < 1580
pid 24497 (getty): 1576 < 1580
pid 27929 (knews.sh): 2984 < 2988
pid 6207 (rsh-title): 1156 < 2980
pid 6208 (rsh): 1952 < 2036
pid 16749 (awk): 1836 < 2368

On 2.6.13 it is much worse.

-- 
Frank