[PATCH] 1/5 VM changes: zone-pressure.patch

[PATCH] 1/5 VM changes: zone-pressure.patch

[Nikita Danilov]

Add zone->pressure field. It estimates scanning intensity for this zone and
is calculated as exponentially decaying average of the scanning priority
required to free enough pages in this zone (mm/vmscan.c:zone_adj_pressure()).

zone->pressure is supposed to be used in stead of scanning priority by
vmscan.c. This is used by later patches in a series.



diff -puN include/linux/mmzone.h~zone-pressure include/linux/mmzone.h
--- i386/include/linux/mmzone.h~zone-pressure	Wed Jul  9 12:24:50 2003
+++ i386-god/include/linux/mmzone.h	Wed Jul  9 12:24:50 2003
@@ -84,11 +84,23 @@ struct zone {
 	atomic_t		refill_counter;
 	unsigned long		nr_active;
 	unsigned long		nr_inactive;
-	int			all_unreclaimable; /* All pages pinned */
+	int			all_unreclaimable:1; /* All pages pinned */
 	unsigned long		pages_scanned;	   /* since last reclaim */
 
 	ZONE_PADDING(_pad2_)
 
+ 	/*
+	 * measure of scanning intensity for this zone. It is calculated
+	 * as exponentially decaying average of the scanning priority
+	 * required to free enough pages in this zone
+	 * (zone_adj_pressure()).
+	 *
+	 *     0                    --- low pressure
+	 *
+	 *     (DEF_PRIORITY << 10) --- high pressure
+	 */
+	int pressure;
+
 	/*
 	 * free areas of different sizes
 	 */
diff -puN mm/vmscan.c~zone-pressure mm/vmscan.c
--- i386/mm/vmscan.c~zone-pressure	Wed Jul  9 12:24:50 2003
+++ i386-god/mm/vmscan.c	Wed Jul  9 12:24:50 2003
@@ -80,6 +80,22 @@ static long total_memory;
 #endif
 
 /*
+ * exponentially decaying average
+ */
+static inline int expavg(int avg, int val, int order)
+{
+	return ((val - avg) >> order) + avg;
+}
+
+static void zone_adj_pressure(struct zone * zone, int priority)
+{
+	int pass;
+
+	pass = DEF_PRIORITY - priority;
+	zone->pressure = expavg(zone->pressure, pass << 10, 1);
+}
+
+/*
  * The list of shrinker callbacks used by to apply pressure to
  * ageable caches.
  */
@@ -794,8 +810,10 @@ shrink_caches(struct zone *classzone, in
 		ret += shrink_zone(zone, max_scan, gfp_mask,
 				to_reclaim, &nr_mapped, ps, priority);
 		*total_scanned += max_scan + nr_mapped;
-		if (ret >= nr_pages)
+		if (ret >= nr_pages) {
+			zone_adj_pressure(zone, priority);
 			break;
+		}
 	}
 	return ret;
 }
@@ -824,6 +842,7 @@ int try_to_free_pages(struct zone *cz,
 	int ret = 0;
 	const int nr_pages = SWAP_CLUSTER_MAX;
 	int nr_reclaimed = 0;
+	struct zone *zone;
 	struct reclaim_state *reclaim_state = current->reclaim_state;
 
 	inc_page_state(allocstall);
@@ -860,6 +879,8 @@ int try_to_free_pages(struct zone *cz,
 	}
 	if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
 		out_of_memory();
+	for (zone = cz; zone >= cz->zone_pgdat->node_zones; -- zone)
+		zone_adj_pressure(zone, -1);
 out:
 	return ret;
 }
@@ -907,8 +928,10 @@ static int balance_pgdat(pg_data_t *pgda
 				to_reclaim = min(to_free, SWAP_CLUSTER_MAX*8);
 			} else {			/* Zone balancing */
 				to_reclaim = zone->pages_high-zone->free_pages;
-				if (to_reclaim <= 0)
+				if (to_reclaim <= 0) {
+					zone_adj_pressure(zone, priority);
 					continue;
+				}
 			}
 			all_zones_ok = 0;
 			max_scan = zone->nr_inactive >> priority;
@@ -932,6 +955,14 @@ static int balance_pgdat(pg_data_t *pgda
 			break;
 		blk_congestion_wait(WRITE, HZ/10);
 	}
+	if (priority < 0) {
+		for (i = 0; i < pgdat->nr_zones; i++) {
+			struct zone *zone = pgdat->node_zones + i;
+
+			if (zone->free_pages < zone->pages_high)
+				zone_adj_pressure(zone, -1);
+		}
+	}
 	return nr_pages - to_free;
 }
 

_

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Andrew Morton]

Nikita Danilov <Nikita@Namesys.COM> wrote:
>
> Add zone->pressure field. It estimates scanning intensity for this zone and
>  is calculated as exponentially decaying average of the scanning priority
>  required to free enough pages in this zone (mm/vmscan.c:zone_adj_pressure()).
> 
>  zone->pressure is supposed to be used in stead of scanning priority by
>  vmscan.c. This is used by later patches in a series.
> 

OK, fixes a bug.

> 
>  diff -puN include/linux/mmzone.h~zone-pressure include/linux/mmzone.h
>  --- i386/include/linux/mmzone.h~zone-pressure	Wed Jul  9 12:24:50 2003
>  +++ i386-god/include/linux/mmzone.h	Wed Jul  9 12:24:50 2003
>  @@ -84,11 +84,23 @@ struct zone {
>   	atomic_t		refill_counter;
>   	unsigned long		nr_active;
>   	unsigned long		nr_inactive;
>  -	int			all_unreclaimable; /* All pages pinned */
>  +	int			all_unreclaimable:1; /* All pages pinned */

I'll revert this change.  Once we start adding bitfields in there they all
need common locking and it gets messy.  integers are simpler.

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Nikita Danilov]

Andrew Morton writes:
 > Nikita Danilov <Nikita@Namesys.COM> wrote:
 > >
 > > Add zone->pressure field. It estimates scanning intensity for this zone and
 > >  is calculated as exponentially decaying average of the scanning priority
 > >  required to free enough pages in this zone (mm/vmscan.c:zone_adj_pressure()).
 > > 
 > >  zone->pressure is supposed to be used in stead of scanning priority by
 > >  vmscan.c. This is used by later patches in a series.
 > > 
 > 
 > OK, fixes a bug.

What bug?

 > 
 > > 
 > >  diff -puN include/linux/mmzone.h~zone-pressure include/linux/mmzone.h
 > >  --- i386/include/linux/mmzone.h~zone-pressure	Wed Jul  9 12:24:50 2003
 > >  +++ i386-god/include/linux/mmzone.h	Wed Jul  9 12:24:50 2003
 > >  @@ -84,11 +84,23 @@ struct zone {
 > >   	atomic_t		refill_counter;
 > >   	unsigned long		nr_active;
 > >   	unsigned long		nr_inactive;
 > >  -	int			all_unreclaimable; /* All pages pinned */
 > >  +	int			all_unreclaimable:1; /* All pages pinned */
 > 
 > I'll revert this change.  Once we start adding bitfields in there they all
 > need common locking and it gets messy.  integers are simpler.

OK.

 > 

Nikita.

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Andrew Morton]

Nikita Danilov <Nikita@Namesys.COM> wrote:
>
>  > OK, fixes a bug.
> 
>  What bug?

Failing to consider mapped pages on the active list until the scanning
priority gets large.

I ran up your five patches on a 256MB box, running `qsbench -m 350'.  It got
all slow then the machine seized up.   I'll poke at it some.

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Nikita Danilov]

Andrew Morton writes:
 > Nikita Danilov <Nikita@Namesys.COM> wrote:
 > >
 > >  > OK, fixes a bug.
 > > 
 > >  What bug?
 > 
 > Failing to consider mapped pages on the active list until the scanning
 > priority gets large.
 > 
 > I ran up your five patches on a 256MB box, running `qsbench -m 350'.  It got
 > all slow then the machine seized up.   I'll poke at it some.
 > 

Thanks for trying them out.

Nikita.

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Martin Schlemmer]

On Wed, 2003-07-09 at 12:42, Andrew Morton wrote:
> Nikita Danilov <Nikita@Namesys.COM> wrote:
> >
> >  > OK, fixes a bug.
> > 
> >  What bug?
> 
> Failing to consider mapped pages on the active list until the scanning
> priority gets large.
> 
> I ran up your five patches on a 256MB box, running `qsbench -m 350'.  It got
> all slow then the machine seized up.   I'll poke at it some.
> 

P4 2.4HT with 1GB of ram - could be me, but things seem to start a
tad quicker.  If I however push it with two 'make -j12' compiles that
normally works fine, it goes for about 10-15mins, and then OOM killer
kills X, evo, galeon, etc.

Must note that for most of the time free memory stays in the region of
450-650mb and then suddenly things goes for the worse.  Also, after this
there is only about 80MB used ...  If I then start X, etc again, same
thing after 10-15 mins - the box thus never died, just suddenly all free
memory was taken.

If you need any stats, etc, let me know.

BTW, kernel is 2.5.74-bk7.


Regards,

-- 
Martin Schlemmer

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Nikita Danilov]

Andrew Morton writes:
 > Nikita Danilov <Nikita@Namesys.COM> wrote:
 > >
 > >  > OK, fixes a bug.
 > > 
 > >  What bug?
 > 
 > Failing to consider mapped pages on the active list until the scanning
 > priority gets large.
 > 
 > I ran up your five patches on a 256MB box, running `qsbench -m 350'.  It got
 > all slow then the machine seized up.   I'll poke at it some.
 > 

My understanding of this is following:

1. kswapd/balance_pgdat runs through all priorities, but fails to rise
zone->free_pages above zone->pages_min. So it exet. Which is strange.

2. nobody is going to wake it up, because all memory allocators are
looping indefinitely in that newly introduced do_retry loop inside
__alloc_pages.

I think either wakeup_kswapd() should be called in __alloc_pages()
before goto rebalance, or kswapd should loop until it restores all zones
to the balance (actually, comment before balance_pgdat() says it does),
or both.

Nikita.

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Andrew Morton]

Nikita Danilov <Nikita@Namesys.COM> wrote:
>
>  +	if (priority < 0) {
>  +		for (i = 0; i < pgdat->nr_zones; i++) {
>  +			struct zone *zone = pgdat->node_zones + i;
>  +
>  +			if (zone->free_pages < zone->pages_high)
>  +				zone_adj_pressure(zone, -1);
>  +		}
>  +	}

What is this bit doing?

Re: [PATCH] 1/5 VM changes: zone-pressure.patch

[Nikita Danilov]

Andrew Morton writes:
 > Nikita Danilov <Nikita@Namesys.COM> wrote:
 > >
 > >  +	if (priority < 0) {
 > >  +		for (i = 0; i < pgdat->nr_zones; i++) {
 > >  +			struct zone *zone = pgdat->node_zones + i;
 > >  +
 > >  +			if (zone->free_pages < zone->pages_high)
 > >  +				zone_adj_pressure(zone, -1);
 > >  +		}
 > >  +	}
 > 
 > What is this bit doing?

kswapd failed to balance some zone after going up to the maximal
priority (0), increase ->pressure on this zone.

In other words: zone->pressure is average of the scanning priority
required to free enough pages in this zone. As no scanning priority was
enough to free pages, zone->pressure should be extra high.

Nikita.