Re: [PATCH 1/4] vmstat: remove zone->lock from walk_zones_in_node

[Mel Gorman <mel@csn.ul.ie>]

On Tue, Jan 05, 2010 at 11:04:58AM +0900, KOSAKI Motohiro wrote:
> Hi
> 
> > On Mon, Dec 28, 2009 at 04:47:22PM +0900, KOSAKI Motohiro wrote:
> > > The zone->lock is one of performance critical locks. Then, it shouldn't
> > > be hold for long time. Currently, we have four walk_zones_in_node()
> > > usage and almost use-case don't need to hold zone->lock.
> > > 
> > > Thus, this patch move locking responsibility from walk_zones_in_node
> > > to its sub function. Also this patch kill unnecessary zone->lock taking.
> > > 
> > > Cc: Mel Gorman <mel@csn.ul.ie>
> > > Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
> > > ---
> > >  mm/vmstat.c |    8 +++++---
> > >  1 files changed, 5 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/mm/vmstat.c b/mm/vmstat.c
> > > index 6051fba..a5d45bc 100644
> > > --- a/mm/vmstat.c
> > > +++ b/mm/vmstat.c
> > > @@ -418,15 +418,12 @@ static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
> > >  {
> > >  	struct zone *zone;
> > >  	struct zone *node_zones = pgdat->node_zones;
> > > -	unsigned long flags;
> > >  
> > >  	for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
> > >  		if (!populated_zone(zone))
> > >  			continue;
> > >  
> > > -		spin_lock_irqsave(&zone->lock, flags);
> > >  		print(m, pgdat, zone);
> > > -		spin_unlock_irqrestore(&zone->lock, flags);
> > >  	}
> > >  }
> > >  
> > > @@ -455,6 +452,7 @@ static void pagetypeinfo_showfree_print(struct seq_file *m,
> > >  					pg_data_t *pgdat, struct zone *zone)
> > >  {
> > >  	int order, mtype;
> > > +	unsigned long flags;
> > >  
> > >  	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
> > >  		seq_printf(m, "Node %4d, zone %8s, type %12s ",
> > > @@ -468,8 +466,11 @@ static void pagetypeinfo_showfree_print(struct seq_file *m,
> > >  
> > >  			area = &(zone->free_area[order]);
> > >  
> > > +			spin_lock_irqsave(&zone->lock, flags);
> > >  			list_for_each(curr, &area->free_list[mtype])
> > >  				freecount++;
> > > +			spin_unlock_irqrestore(&zone->lock, flags);
> > > +
> > 
> > It's not clear why you feel this information requires the lock and the
> > others do not.
> 
> I think above list operation require lock to prevent NULL pointer access. but other parts
> doesn't protect anything, because memory-hotplug change them without zone lock.
> 

True. Add a comment explaining that. I considered list_for_each_safe()
but it wouldn't work in all cases.

> 
> > For the most part, I agree that the accuracy of the information is
> > not critical. Assuming partial writes of the data are not a problem,
> > the information is not going to go so badly out of sync that it would be
> > noticable, even if the information is out of date within the zone.
> > 
> > However, inconsistent reads in zoneinfo really could be a problem. I am
> > concerned that under heavy allocation load that that "pages free" would
> > not match "nr_pages_free" for example. Other examples that adding all the
> > counters together may or may not equal the total number of pages in the zone.
> > 
> > Lets say for example there was a subtle bug related to __inc_zone_page_state()
> > that meant that counters were getting slightly out of sync but it was very
> > marginal and/or difficult to reproduce. With this patch applied, we could
> > not be absolutly sure the counters were correct because it could always have
> > raced with someone holding the zone->lock.
> > 
> > Minimally, I think zoneinfo should be taking the zone lock.
> 
> Thanks lots comments. 
> hmm.. I'd like to clarily your point. My point is memory-hotplug don't take zone lock,
> then zone lock doesn't protect anything. so we have two option
> 
> 1) Add zone lock to memroy-hotplug
> 2) Remove zone lock from zoneinfo
> 
> I thought (2) is sufficient. Do you mean you prefer to (1)? Or you prefer to ignore rarely event
> (of cource, memory hotplug is rarely)?
> 

I think (2) will make zoneinfo harder to use for examining all the counters
properly as I explained above. I haven't looked at memory-hotplug in a
while but IIRC, fields like present_pages should be protected by a lock on
the pgdat and a seq lock on the zone. If this is not true at the moment,
it is a problem.

For the free lists, memory hotplug should be taking the zone->lock properly as
the final stage of onlining memory is to walk the sections being hot-added,
init the memmap and then __free_page() each page individually - i.e. the
normal free path.

So, if memory hotplug is not protected by proper locking, it's not intentional.

> 
> > Secondly, has increased zone->lock contention due to reading /proc
> > really been shown to be a problem? The only situation that I can think
> > of is a badly-written monitor program that is copying all of /proc
> > instead of the files of interest. If a monitor program is doing
> > something like that, it's likely to be incurring performance problems in
> > a large number of different areas. If that is not the trigger case, what
> > is?
> 
> Ah no. I haven't observe such issue. my point is removing meaningless lock.
> 

Then I believe the zonelock should be preserved so that all entries in
/proc/zoneinfo are consistent.

> 
> > >  			seq_printf(m, "%6lu ", freecount);
> > >  		}
> > >  		seq_putc(m, '\n');
> > > @@ -709,6 +710,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
> > >  							struct zone *zone)
> > >  {
> > >  	int i;
> > > +
> > 
> > Unnecessary whitespace change.
> 
> Ug. thanks, it's my fault.
> 
> 
> 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab