[PATCH 0/3] Reduce watermark-related problems with the per-cpu allocator V2

[PATCH 0/3] Reduce watermark-related problems with the per-cpu allocator V2

[Mel Gorman]

Changelog since V1
  o Fix for !CONFIG_SMP
  o Correct spelling mistakes
  o Clarify a ChangeLog
  o Only check for counter drift on machines large enough for the counter
    drift to breach the min watermark when NR_FREE_PAGES report the low
    watermark is fine

Internal IBM test teams beta testing distribution kernels have reported
problems on machines with a large number of CPUs whereby page allocator
failure messages show huge differences between the nr_free_pages vmstat
counter and what is available on the buddy lists. In an extreme example,
nr_free_pages was above the min watermark but zero pages were on the buddy
lists allowing the system to potentially livelock unable to make forward
progress unless an allocation succeeds. There is no reason why the problems
would not affect mainline so the following series mitigates the problems
in the page allocator related to to per-cpu counter drift and lists.

The first patch ensures that counters are updated after pages are added to
free lists.

The second patch notes that the counter drift between nr_free_pages and what
is on the per-cpu lists can be very high. When memory is low and kswapd
is awake, the per-cpu counters are checked as well as reading the value
of NR_FREE_PAGES. This will slow the page allocator when memory is low and
kswapd is awake but it will be much harder to breach the min watermark and
potentially livelock the system.

The third patch notes that after direct-reclaim an allocation can
fail because the necessary pages are on the per-cpu lists. After a
direct-reclaim-and-allocation-failure, the per-cpu lists are drained and
a second attempt is made.

Performance tests did not show up anything interesting. A version of this
series that continually called vmstat_update() when memory was low was
tested internally and found to help the counter drift problem. I described
this during LSF/MM Summit and the potential for IPI storms was frowned
upon. An alternative fix is in patch two which uses for_each_online_cpu()
to read the vmstat deltas while memory is low and kswapd is awake. This
should be functionally similar.

This patch should be merged after the patch "vmstat : update
zone stat threshold at onlining a cpu" which is in mmotm as
vmstat-update-zone-stat-threshold-when-onlining-a-cpu.patch .

Are there any objections to merging?

 include/linux/mmzone.h |   13 +++++++++++++
 mm/mmzone.c            |   29 +++++++++++++++++++++++++++++
 mm/page_alloc.c        |   29 +++++++++++++++++++++--------
 mm/vmstat.c            |   15 ++++++++++++++-
 4 files changed, 77 insertions(+), 9 deletions(-)

[PATCH 1/3] mm: page allocator: Update free page counters after pages are placed on the free list

[Mel Gorman]

When allocating a page, the system uses NR_FREE_PAGES counters to determine
if watermarks would remain intact after the allocation was made. This
check is made without interrupts disabled or the zone lock held and so is
race-prone by nature. Unfortunately, when pages are being freed in batch,
the counters are updated before the pages are added on the list. During this
window, the counters are misleading as the pages do not exist yet. When
under significant pressure on systems with large numbers of CPUs, it's
possible for processes to make progress even though they should have been
stalled. This is particularly problematic if a number of the processes are
using GFP_ATOMIC as the min watermark can be accidentally breached and in
extreme cases, the system can livelock.

This patch updates the counters after the pages have been added to the
list. This makes the allocator more cautious with respect to preserving
the watermarks and mitigates livelock possibilities.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
---
 mm/page_alloc.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index a9649f4..97d74a0 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -588,12 +588,12 @@ static void free_pcppages_bulk(struct zone *zone, int count,
 {
 	int migratetype = 0;
 	int batch_free = 0;
+	int freed = count;
 
 	spin_lock(&zone->lock);
 	zone->all_unreclaimable = 0;
 	zone->pages_scanned = 0;
 
-	__mod_zone_page_state(zone, NR_FREE_PAGES, count);
 	while (count) {
 		struct page *page;
 		struct list_head *list;
@@ -621,6 +621,7 @@ static void free_pcppages_bulk(struct zone *zone, int count,
 			trace_mm_page_pcpu_drain(page, 0, page_private(page));
 		} while (--count && --batch_free && !list_empty(list));
 	}
+	__mod_zone_page_state(zone, NR_FREE_PAGES, freed);
 	spin_unlock(&zone->lock);
 }
 
@@ -631,8 +632,8 @@ static void free_one_page(struct zone *zone, struct page *page, int order,
 	zone->all_unreclaimable = 0;
 	zone->pages_scanned = 0;
 
-	__mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 	__free_one_page(page, zone, order, migratetype);
+	__mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
 	spin_unlock(&zone->lock);
 }
 
-- 
1.7.1

[PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

Ordinarily watermark checks are based on the vmstat NR_FREE_PAGES as
it is cheaper than scanning a number of lists. To avoid synchronization
overhead, counter deltas are maintained on a per-cpu basis and drained both
periodically and when the delta is above a threshold. On large CPU systems,
the difference between the estimated and real value of NR_FREE_PAGES can be
very high.  If NR_FREE_PAGES is much higher than number of real free page
in buddy, the VM can allocate pages below min watermark, at worst reducing
the real number of pages to zero.  Even if the OOM killer kills some victim
for freeing memory, it may not free memory if the exit path requires a new
page resulting in livelock.

This patch introduces zone_nr_free_pages() to take a slightly more accurate
estimate of NR_FREE_PAGES while kswapd is awake. The estimate is not perfect
and may result in cache line bounces but is expected to be lighter than the
IPI calls necessary to continually drain the per-cpu counters while kswapd
is awake.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
---
 include/linux/mmzone.h |   13 +++++++++++++
 mm/mmzone.c            |   29 +++++++++++++++++++++++++++++
 mm/page_alloc.c        |    4 ++--
 mm/vmstat.c            |   15 ++++++++++++++-
 4 files changed, 58 insertions(+), 3 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 6e6e626..3984c4e 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -284,6 +284,13 @@ struct zone {
 	unsigned long watermark[NR_WMARK];
 
 	/*
+	 * When free pages are below this point, additional steps are taken
+	 * when reading the number of free pages to avoid per-cpu counter
+	 * drift allowing watermarks to be breached
+	 */
+	unsigned long percpu_drift_mark;
+
+	/*
 	 * We don't know if the memory that we're going to allocate will be freeable
 	 * or/and it will be released eventually, so to avoid totally wasting several
 	 * GB of ram we must reserve some of the lower zone memory (otherwise we risk
@@ -441,6 +448,12 @@ static inline int zone_is_oom_locked(const struct zone *zone)
 	return test_bit(ZONE_OOM_LOCKED, &zone->flags);
 }
 
+#ifdef CONFIG_SMP
+unsigned long zone_nr_free_pages(struct zone *zone);
+#else
+#define zone_nr_free_pages(zone) zone_page_state(zone, NR_FREE_PAGES)
+#endif /* CONFIG_SMP */
+
 /*
  * The "priority" of VM scanning is how much of the queues we will scan in one
  * go. A value of 12 for DEF_PRIORITY implies that we will scan 1/4096th of the
diff --git a/mm/mmzone.c b/mm/mmzone.c
index f5b7d17..69ecbe9 100644
--- a/mm/mmzone.c
+++ b/mm/mmzone.c
@@ -87,3 +87,32 @@ int memmap_valid_within(unsigned long pfn,
 	return 1;
 }
 #endif /* CONFIG_ARCH_HAS_HOLES_MEMORYMODEL */
+
+#ifdef CONFIG_SMP
+/* Called when a more accurate view of NR_FREE_PAGES is needed */
+unsigned long zone_nr_free_pages(struct zone *zone)
+{
+	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
+
+	/*
+	 * While kswapd is awake, it is considered the zone is under some
+	 * memory pressure. Under pressure, there is a risk that
+	 * per-cpu-counter-drift will allow the min watermark to be breached
+	 * potentially causing a live-lock. While kswapd is awake and
+	 * free pages are low, get a better estimate for free pages
+	 */
+	if (nr_free_pages < zone->percpu_drift_mark &&
+			!waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
+		int cpu;
+
+		for_each_online_cpu(cpu) {
+			struct per_cpu_pageset *pset;
+
+			pset = per_cpu_ptr(zone->pageset, cpu);
+			nr_free_pages += pset->vm_stat_diff[NR_FREE_PAGES];
+		}
+	}
+
+	return nr_free_pages;
+}
+#endif /* CONFIG_SMP */
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 97d74a0..bbaa959 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1462,7 +1462,7 @@ int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
 {
 	/* free_pages my go negative - that's OK */
 	long min = mark;
-	long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
+	long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
 	int o;
 
 	if (alloc_flags & ALLOC_HIGH)
@@ -2424,7 +2424,7 @@ void show_free_areas(void)
 			" all_unreclaimable? %s"
 			"\n",
 			zone->name,
-			K(zone_page_state(zone, NR_FREE_PAGES)),
+			K(zone_nr_free_pages(zone)),
 			K(min_wmark_pages(zone)),
 			K(low_wmark_pages(zone)),
 			K(high_wmark_pages(zone)),
diff --git a/mm/vmstat.c b/mm/vmstat.c
index f389168..696cab2 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -138,11 +138,24 @@ static void refresh_zone_stat_thresholds(void)
 	int threshold;
 
 	for_each_populated_zone(zone) {
+		unsigned long max_drift, tolerate_drift;
+
 		threshold = calculate_threshold(zone);
 
 		for_each_online_cpu(cpu)
 			per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 							= threshold;
+
+		/*
+		 * Only set percpu_drift_mark if there is a danger that
+		 * NR_FREE_PAGES reports the low watermark is ok when in fact
+		 * the min watermark could be breached by an allocation
+		 */
+		tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
+		max_drift = num_online_cpus() * threshold;
+		if (max_drift > tolerate_drift)
+			zone->percpu_drift_mark = high_wmark_pages(zone) +
+					max_drift;
 	}
 }
 
@@ -813,7 +826,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 		   "\n        scanned  %lu"
 		   "\n        spanned  %lu"
 		   "\n        present  %lu",
-		   zone_page_state(zone, NR_FREE_PAGES),
+		   zone_nr_free_pages(zone),
 		   min_wmark_pages(zone),
 		   low_wmark_pages(zone),
 		   high_wmark_pages(zone),
-- 
1.7.1

[PATCH 3/3] mm: page allocator: Drain per-cpu lists after direct reclaim allocation fails

[Mel Gorman]

When under significant memory pressure, a process enters direct reclaim
and immediately afterwards tries to allocate a page. If it fails and no
further progress is made, it's possible the system will go OOM. However,
on systems with large amounts of memory, it's possible that a significant
number of pages are on per-cpu lists and inaccessible to the calling
process. This leads to a process entering direct reclaim more often than
it should increasing the pressure on the system and compounding the problem.

This patch notes that if direct reclaim is making progress but
allocations are still failing that the system is already under heavy
pressure. In this case, it drains the per-cpu lists and tries the
allocation a second time before continuing.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
---
 mm/page_alloc.c |   20 ++++++++++++++++----
 1 files changed, 16 insertions(+), 4 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index bbaa959..750e1dc 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1847,6 +1847,7 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
 	struct page *page = NULL;
 	struct reclaim_state reclaim_state;
 	struct task_struct *p = current;
+	bool drained = false;
 
 	cond_resched();
 
@@ -1865,14 +1866,25 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
 
 	cond_resched();
 
-	if (order != 0)
-		drain_all_pages();
+	if (unlikely(!(*did_some_progress)))
+		return NULL;
 
-	if (likely(*did_some_progress))
-		page = get_page_from_freelist(gfp_mask, nodemask, order,
+retry:
+	page = get_page_from_freelist(gfp_mask, nodemask, order,
 					zonelist, high_zoneidx,
 					alloc_flags, preferred_zone,
 					migratetype);
+
+	/*
+	 * If an allocation failed after direct reclaim, it could be because
+	 * pages are pinned on the per-cpu lists. Drain them and try again
+	 */
+	if (!page && !drained) {
+		drain_all_pages();
+		drained = true;
+		goto retry;
+	}
+
 	return page;
 }
 
-- 
1.7.1

Re: [PATCH 0/3] Reduce watermark-related problems with the per-cpu allocator V2

[Christoph Lameter]

On Mon, 23 Aug 2010, Mel Gorman wrote:

> Internal IBM test teams beta testing distribution kernels have reported
> problems on machines with a large number of CPUs whereby page allocator
> failure messages show huge differences between the nr_free_pages vmstat
> counter and what is available on the buddy lists. In an extreme example,
> nr_free_pages was above the min watermark but zero pages were on the buddy
> lists allowing the system to potentially livelock unable to make forward
> progress unless an allocation succeeds. There is no reason why the problems
> would not affect mainline so the following series mitigates the problems
> in the page allocator related to to per-cpu counter drift and lists.

The maximum time for which the livelock can exists is the vm stat
interval. By default the counters are brought up to date at least once per
second or if a certain delta was violated. Drifts are controlled by the
delta configuration.

Re: [PATCH 1/3] mm: page allocator: Update free page counters after pages are placed on the free list

[Christoph Lameter]

Reviewed-by: Christoph Lameter <cl@linux.com>

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Mon, 23 Aug 2010, Mel Gorman wrote:

> This patch introduces zone_nr_free_pages() to take a slightly more accurate
> estimate of NR_FREE_PAGES while kswapd is awake. The estimate is not perfect
> and may result in cache line bounces but is expected to be lighter than the
> IPI calls necessary to continually drain the per-cpu counters while kswapd
> is awake.

The delta of the counters could also be reduced to increase accuracy.
See refresh_zone_stat_thresholds().

Also would it be possible to add the summation function to vmstat? It may
be useful elsewhere.

A new function like

	zone_page_state_snapshot()

or so?

Re: [PATCH 0/3] Reduce watermark-related problems with the per-cpu allocator V2

[Mel Gorman]

On Mon, Aug 23, 2010 at 07:45:25AM -0500, Christoph Lameter wrote:
> On Mon, 23 Aug 2010, Mel Gorman wrote:
> 
> > Internal IBM test teams beta testing distribution kernels have reported
> > problems on machines with a large number of CPUs whereby page allocator
> > failure messages show huge differences between the nr_free_pages vmstat
> > counter and what is available on the buddy lists. In an extreme example,
> > nr_free_pages was above the min watermark but zero pages were on the buddy
> > lists allowing the system to potentially livelock unable to make forward
> > progress unless an allocation succeeds. There is no reason why the problems
> > would not affect mainline so the following series mitigates the problems
> > in the page allocator related to to per-cpu counter drift and lists.
> 
> The maximum time for which the livelock can exists is the vm stat
> interval. By default the counters are brought up to date at least once per
> second or if a certain delta was violated. Drifts are controlled by the
> delta configuration.
> 

While there is a maximum time (2 seconds I think) the drift can exist
in, a machine under enough pressure can make a mess of the watermarks
during that time. If it wasn't the case, these livelocks with 0 pages
free wouldn't be happening.

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Mon, Aug 23, 2010 at 07:56:40AM -0500, Christoph Lameter wrote:
> On Mon, 23 Aug 2010, Mel Gorman wrote:
> 
> > This patch introduces zone_nr_free_pages() to take a slightly more accurate
> > estimate of NR_FREE_PAGES while kswapd is awake. The estimate is not perfect
> > and may result in cache line bounces but is expected to be lighter than the
> > IPI calls necessary to continually drain the per-cpu counters while kswapd
> > is awake.
> 
> The delta of the counters could also be reduced to increase accuracy.
> See refresh_zone_stat_thresholds().
> 

True, but I thought that would introduce a constant performance penalty
for a corner case which I didn't like.

> Also would it be possible to add the summation function to vmstat? It may
> be useful elsewhere.
> 
> A new function like
> 
> 	zone_page_state_snapshot()
> 
> or so?
> 

We could if there is another counter that results in bad system
behaviour due to counter drift. As NR_FREE_PAGES seemed to be the only
one, zone_nr_free_pages() seemed adequate. If such a helper did exist,
zone_nr_free_pages() would be a simple wrapper around it. The
indirection didn't seem necessary at this point though.

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

Re: [PATCH 0/3] Reduce watermark-related problems with the per-cpu allocator V2

[Christoph Lameter]

On Mon, 23 Aug 2010, Mel Gorman wrote:

> > The maximum time for which the livelock can exists is the vm stat
> > interval. By default the counters are brought up to date at least once per
> > second or if a certain delta was violated. Drifts are controlled by the
> > delta configuration.
> >
>
> While there is a maximum time (2 seconds I think) the drift can exist
> in, a machine under enough pressure can make a mess of the watermarks
> during that time. If it wasn't the case, these livelocks with 0 pages
> free wouldn't be happening.

So because we go way beyond the watermarks we reach a state in which a
livelock exists that does not go away when the counters are finally
updated?

Re: [PATCH 0/3] Reduce watermark-related problems with the per-cpu allocator V2

[Mel Gorman]

On Mon, Aug 23, 2010 at 08:38:25AM -0500, Christoph Lameter wrote:
> On Mon, 23 Aug 2010, Mel Gorman wrote:
> 
> > > The maximum time for which the livelock can exists is the vm stat
> > > interval. By default the counters are brought up to date at least once per
> > > second or if a certain delta was violated. Drifts are controlled by the
> > > delta configuration.
> > >
> >
> > While there is a maximum time (2 seconds I think) the drift can exist
> > in, a machine under enough pressure can make a mess of the watermarks
> > during that time. If it wasn't the case, these livelocks with 0 pages
> > free wouldn't be happening.
> 
> So because we go way beyond the watermarks we reach a state in which a
> livelock exists that does not go away when the counters are finally
> updated?
> 

That appears to be the case. The system has already gotten into a state
where there are 0 pages free. Just because the NR_FREE_PAGES counter
gets updated to reflect the accurate count of 0 does not mean the system
can recover from it.

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Mon, 23 Aug 2010, Mel Gorman wrote:

> > The delta of the counters could also be reduced to increase accuracy.
> > See refresh_zone_stat_thresholds().
> True, but I thought that would introduce a constant performance penalty
> for a corner case which I didn't like.

Sure, an increased frequency of updates would increase the chance of
bouncing cachelines. But the bouncing cacheline scenario for the vm
counters was tuned for applications that continually allocate pages in
parallel.

When the vm gets into a state where continual reclaim is necessary then
the counters are not that frequently updated. If the machine is already
slowing down due to reclaim then the vm can likely affort more frequent
counter updates.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Mon, Aug 23, 2010 at 08:41:56AM -0500, Christoph Lameter wrote:
> On Mon, 23 Aug 2010, Mel Gorman wrote:
> 
> > > The delta of the counters could also be reduced to increase accuracy.
> > > See refresh_zone_stat_thresholds().
> > True, but I thought that would introduce a constant performance penalty
> > for a corner case which I didn't like.
> 
> Sure, an increased frequency of updates would increase the chance of
> bouncing cachelines. But the bouncing cacheline scenario for the vm
> counters was tuned for applications that continually allocate pages in
> parallel.
> 
> When the vm gets into a state where continual reclaim is necessary then
> the counters are not that frequently updated. If the machine is already
> slowing down due to reclaim then the vm can likely affort more frequent
> counter updates.
> 

Ok, but is that better than this patch? Decreasing the size of the window by
reducing the threshold still leaves a window. There is still a small amount
of drift by summing up all the deltas but you get a much more accurate count
at the point of time it was important to know.

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Mon, 23 Aug 2010, Mel Gorman wrote:

> > When the vm gets into a state where continual reclaim is necessary then
> > the counters are not that frequently updated. If the machine is already
> > slowing down due to reclaim then the vm can likely affort more frequent
> > counter updates.
> >
>
> Ok, but is that better than this patch? Decreasing the size of the window by
> reducing the threshold still leaves a window. There is still a small amount
> of drift by summing up all the deltas but you get a much more accurate count
> at the point of time it was important to know.

In order to make that decision we would need to know what deltas make a
significant difference. Would be also important to know if there are any
other counters that have issues. If so then the reduction of the
thresholds is addressing these problems in a number of counters.

I have no objection against this approach here but it may just be bandaid
on a larger issue that could be approached in a cleaner way.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Mon, Aug 23, 2010 at 11:04:38AM -0500, Christoph Lameter wrote:
> On Mon, 23 Aug 2010, Mel Gorman wrote:
> 
> > > When the vm gets into a state where continual reclaim is necessary then
> > > the counters are not that frequently updated. If the machine is already
> > > slowing down due to reclaim then the vm can likely affort more frequent
> > > counter updates.
> > >
> >
> > Ok, but is that better than this patch? Decreasing the size of the window by
> > reducing the threshold still leaves a window. There is still a small amount
> > of drift by summing up all the deltas but you get a much more accurate count
> > at the point of time it was important to know.
> 
> In order to make that decision we would need to know what deltas make a
> significant difference.

A delta on the NR_FREE_PAGES is the obvious problem. The page allocation
failure report I saw clearly stated that free was a value above min watermark
where as the buddy lists just as clearly showed that the number of pages on
the list were 0.

> Would be also important to know if there are any
> other counters that have issues.

I am not aware of similar issues with another counter where drift causes
the system to make the wrong decision, are you?

> If so then the reduction of the
> thresholds is addressing these problems in a number of counters.
> 
> I have no objection against this approach here but it may just be bandaid
> on a larger issue that could be approached in a cleaner way.
> 

Unfortunately, I do not have access to a machine large enough to investigate
around this area. All I have to go on is a few bug reports showing the delta
problem with NR_FREE_PAGES and test results in a patch functionally similar
to this patch showing that the livelock problem went away.

At best all we can do is keep an eye out for problems one large machines
that could be explained by counter drift. If such a bug is found with a
reporter with regular access to the machine for test kernels, we can
investigate if reducing the thresholds fix the problem without affecting
general performance.

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

Re: [PATCH 3/3] mm: page allocator: Drain per-cpu lists after direct reclaim allocation fails

[KOSAKI Motohiro]

> When under significant memory pressure, a process enters direct reclaim
> and immediately afterwards tries to allocate a page. If it fails and no
> further progress is made, it's possible the system will go OOM. However,
> on systems with large amounts of memory, it's possible that a significant
> number of pages are on per-cpu lists and inaccessible to the calling
> process. This leads to a process entering direct reclaim more often than
> it should increasing the pressure on the system and compounding the problem.
> 
> This patch notes that if direct reclaim is making progress but
> allocations are still failing that the system is already under heavy
> pressure. In this case, it drains the per-cpu lists and tries the
> allocation a second time before continuing.
> 
> Signed-off-by: Mel Gorman <mel@csn.ul.ie>
> Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
> Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> ---
>  mm/page_alloc.c |   20 ++++++++++++++++----
>  1 files changed, 16 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index bbaa959..750e1dc 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1847,6 +1847,7 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
>  	struct page *page = NULL;
>  	struct reclaim_state reclaim_state;
>  	struct task_struct *p = current;
> +	bool drained = false;
>  
>  	cond_resched();
>  
> @@ -1865,14 +1866,25 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
>  
>  	cond_resched();
>  
> -	if (order != 0)
> -		drain_all_pages();
> +	if (unlikely(!(*did_some_progress)))
> +		return NULL;
>  
> -	if (likely(*did_some_progress))
> -		page = get_page_from_freelist(gfp_mask, nodemask, order,
> +retry:
> +	page = get_page_from_freelist(gfp_mask, nodemask, order,
>  					zonelist, high_zoneidx,
>  					alloc_flags, preferred_zone,
>  					migratetype);
> +
> +	/*
> +	 * If an allocation failed after direct reclaim, it could be because
> +	 * pages are pinned on the per-cpu lists. Drain them and try again
> +	 */
> +	if (!page && !drained) {
> +		drain_all_pages();
> +		drained = true;
> +		goto retry;
> +	}
> +
>  	return page;

I haven't read all of this patch series. (iow, this mail is luckly on top
of my mail box now) but at least I think this one is correct and good.

	Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Fri, Sep 03, 2010 at 03:55:37PM -0700, Andrew Morton wrote:
> On Fri,  3 Sep 2010 10:08:45 +0100
> Mel Gorman <mel@csn.ul.ie> wrote:
> 
> > From: Christoph Lameter <cl@linux.com>
> > 
> > Ordinarily watermark checks are based on the vmstat NR_FREE_PAGES as
> > it is cheaper than scanning a number of lists. To avoid synchronization
> > overhead, counter deltas are maintained on a per-cpu basis and drained both
> > periodically and when the delta is above a threshold. On large CPU systems,
> > the difference between the estimated and real value of NR_FREE_PAGES can
> > be very high. If NR_FREE_PAGES is much higher than number of real free page
> > in buddy, the VM can allocate pages below min watermark, at worst reducing
> > the real number of pages to zero. Even if the OOM killer kills some victim
> > for freeing memory, it may not free memory if the exit path requires a new
> > page resulting in livelock.
> > 
> > This patch introduces a zone_page_state_snapshot() function (courtesy of
> > Christoph) that takes a slightly more accurate of an arbitrary vmstat counter.
> > It is used to read NR_FREE_PAGES while kswapd is awake to avoid the watermark
> > being accidentally broken.  The estimate is not perfect and may result
> > in cache line bounces but is expected to be lighter than the IPI calls
> > necessary to continually drain the per-cpu counters while kswapd is awake.
> > 
> 
> The "is kswapd awake" heuristic seems fairly hacky.  Can it be
> improved, made more deterministic? 

It could be removed but the problem is that the snap version of the
function could be continually used on large systems that are using
almost all physical memory but not under any memory pressure. kswapd
being awake seemed a reasonable proxy indicator that the system is under
pressure.

> Exactly what state are we looking
> for here?
> 

We want to know when the system is in a state where it is both under
pressure and in danger of breaching the watermark due to per-cpu counter
drift.

> 
> > +/*
> > + * More accurate version that also considers the currently pending
> > + * deltas. For that we need to loop over all cpus to find the current
> > + * deltas. There is no synchronization so the result cannot be
> > + * exactly accurate either.
> > + */
> > +static inline unsigned long zone_page_state_snapshot(struct zone *zone,
> > +					enum zone_stat_item item)
> > +{
> > +	long x = atomic_long_read(&zone->vm_stat[item]);
> > +
> > +#ifdef CONFIG_SMP
> > +	int cpu;
> > +	for_each_online_cpu(cpu)
> > +		x += per_cpu_ptr(zone->pageset, cpu)->vm_stat_diff[item];
> > +
> > +	if (x < 0)
> > +		x = 0;
> > +#endif
> > +	return x;
> > +}
> 
> aka percpu_counter_sum()!
> 
> Can someone remind me why per_cpu_pageset went and reimplemented
> percpu_counters rather than just using them?
> 

It's not an exact fit. Christoph answered this and I do not have
anything additional to say.

> >  extern unsigned long global_reclaimable_pages(void);
> >  extern unsigned long zone_reclaimable_pages(struct zone *zone);
> >  
> > diff --git a/mm/mmzone.c b/mm/mmzone.c
> > index f5b7d17..e35bfb8 100644
> > --- a/mm/mmzone.c
> > +++ b/mm/mmzone.c
> > @@ -87,3 +87,24 @@ int memmap_valid_within(unsigned long pfn,
> >  	return 1;
> >  }
> >  #endif /* CONFIG_ARCH_HAS_HOLES_MEMORYMODEL */
> > +
> > +#ifdef CONFIG_SMP
> > +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> > +unsigned long zone_nr_free_pages(struct zone *zone)
> > +{
> > +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> > +
> > +	/*
> > +	 * While kswapd is awake, it is considered the zone is under some
> > +	 * memory pressure. Under pressure, there is a risk that
> > +	 * per-cpu-counter-drift will allow the min watermark to be breached
> > +	 * potentially causing a live-lock. While kswapd is awake and
> > +	 * free pages are low, get a better estimate for free pages
> > +	 */
> > +	if (nr_free_pages < zone->percpu_drift_mark &&
> > +			!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
> > +		return zone_page_state_snapshot(zone, NR_FREE_PAGES);
> > +
> > +	return nr_free_pages;
> > +}
> 
> Is this really the best way of doing it?  The way we usually solve
> this problem (and boy, was this bug a newbie mistake!) is:
> 
> 	foo = percpu_counter_read(x);
> 
> 	if (foo says something bad) {
> 		/* Bad stuff: let's get a more accurate foo */
> 		foo = percpu_counter_sum(x);
> 	}
> 
> 	if (foo still says something bad)
> 		do_bad_thing();
> 
> In other words, don't do all this stuff with percpu_drift_mark and the
> kswapd heuristic.

The percpu_drift_mark and the kswapd heuristic correspond to your "foo
says something bad" above. The drift mark is detecting we're in
potential danger and the kswapd check is telling us we are both in
danger and there is memory pressure. Even if we were using the percpu
counters, it wouldn't eliminate the need for percpu_drift_mark and the
kswapd heuristic, right?

> Just change zone_watermark_ok() to use the more
> accurate read if it's about to return "no".
> 

It could be too late by then. By the tiome zone_watermark_ok() is about
to return no, we could have already breached the watermark by a
significant amount due to the per-cpu counter drift.

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Fri, 3 Sep 2010, Andrew Morton wrote:

> > percpu counters must always be added up when their value is determined.
>
> Nope.  That's the difference between percpu_counter_read() and
> percpu_counter_sum().

Hmmm... Okay you can fold them therefore. That is analogous to what we do
in the _snapshot function now.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Andrew Morton]

On Fri, 3 Sep 2010 18:17:46 -0500 (CDT)
Christoph Lameter <cl@linux.com> wrote:

> On Fri, 3 Sep 2010, Andrew Morton wrote:
> 
> > Can someone remind me why per_cpu_pageset went and reimplemented
> > percpu_counters rather than just using them?
> 
> The vm counters are per zone and per cpu and have a flow from per cpu /
> zone deltas to zone counters and then also into global counters.

hm.  percpu counters would require overflow-time hooks to do that. 
Might be worth looking at.

> > Is this really the best way of doing it?  The way we usually solve
> > this problem (and boy, was this bug a newbie mistake!) is:
> >
> > 	foo = percpu_counter_read(x);
> >
> > 	if (foo says something bad) {
> > 		/* Bad stuff: let's get a more accurate foo */
> > 		foo = percpu_counter_sum(x);
> > 	}
> >
> > 	if (foo still says something bad)
> > 		do_bad_thing();
> >
> > In other words, don't do all this stuff with percpu_drift_mark and the
> > kswapd heuristic.  Just change zone_watermark_ok() to use the more
> > accurate read if it's about to return "no".
> 
> percpu counters must always be added up when their value is determined.

Nope.  That's the difference between percpu_counter_read() and
percpu_counter_sum().

> This seems to be a special case here where Mel does not want to have to
> cost to bring the counters up to date nor reduce the delta/time limits to
> get some more accuracy but wants take some sort of snapshot of the whole
> situation for this particular case.

My suggestion didn't actually have anything to do with percpu_counters.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Fri, 3 Sep 2010, Andrew Morton wrote:

> Can someone remind me why per_cpu_pageset went and reimplemented
> percpu_counters rather than just using them?

The vm counters are per zone and per cpu and have a flow from per cpu /
zone deltas to zone counters and then also into global counters.

> Is this really the best way of doing it?  The way we usually solve
> this problem (and boy, was this bug a newbie mistake!) is:
>
> 	foo = percpu_counter_read(x);
>
> 	if (foo says something bad) {
> 		/* Bad stuff: let's get a more accurate foo */
> 		foo = percpu_counter_sum(x);
> 	}
>
> 	if (foo still says something bad)
> 		do_bad_thing();
>
> In other words, don't do all this stuff with percpu_drift_mark and the
> kswapd heuristic.  Just change zone_watermark_ok() to use the more
> accurate read if it's about to return "no".

percpu counters must always be added up when their value is determined. We
cannot really affort that for the VM. Counters are always available
without looping over all cpus.

vm counters are continually kept up to date (but may have delta limited by
time and counter values).

This seems to be a special case here where Mel does not want to have to
cost to bring the counters up to date nor reduce the delta/time limits to
get some more accuracy but wants take some sort of snapshot of the whole
situation for this particular case.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Andrew Morton]

On Fri,  3 Sep 2010 10:08:45 +0100
Mel Gorman <mel@csn.ul.ie> wrote:

> From: Christoph Lameter <cl@linux.com>
> 
> Ordinarily watermark checks are based on the vmstat NR_FREE_PAGES as
> it is cheaper than scanning a number of lists. To avoid synchronization
> overhead, counter deltas are maintained on a per-cpu basis and drained both
> periodically and when the delta is above a threshold. On large CPU systems,
> the difference between the estimated and real value of NR_FREE_PAGES can
> be very high. If NR_FREE_PAGES is much higher than number of real free page
> in buddy, the VM can allocate pages below min watermark, at worst reducing
> the real number of pages to zero. Even if the OOM killer kills some victim
> for freeing memory, it may not free memory if the exit path requires a new
> page resulting in livelock.
> 
> This patch introduces a zone_page_state_snapshot() function (courtesy of
> Christoph) that takes a slightly more accurate of an arbitrary vmstat counter.
> It is used to read NR_FREE_PAGES while kswapd is awake to avoid the watermark
> being accidentally broken.  The estimate is not perfect and may result
> in cache line bounces but is expected to be lighter than the IPI calls
> necessary to continually drain the per-cpu counters while kswapd is awake.
> 

The "is kswapd awake" heuristic seems fairly hacky.  Can it be
improved, made more deterministic?  Exactly what state are we looking
for here?


> +/*
> + * More accurate version that also considers the currently pending
> + * deltas. For that we need to loop over all cpus to find the current
> + * deltas. There is no synchronization so the result cannot be
> + * exactly accurate either.
> + */
> +static inline unsigned long zone_page_state_snapshot(struct zone *zone,
> +					enum zone_stat_item item)
> +{
> +	long x = atomic_long_read(&zone->vm_stat[item]);
> +
> +#ifdef CONFIG_SMP
> +	int cpu;
> +	for_each_online_cpu(cpu)
> +		x += per_cpu_ptr(zone->pageset, cpu)->vm_stat_diff[item];
> +
> +	if (x < 0)
> +		x = 0;
> +#endif
> +	return x;
> +}

aka percpu_counter_sum()!

Can someone remind me why per_cpu_pageset went and reimplemented
percpu_counters rather than just using them?

>  extern unsigned long global_reclaimable_pages(void);
>  extern unsigned long zone_reclaimable_pages(struct zone *zone);
>  
> diff --git a/mm/mmzone.c b/mm/mmzone.c
> index f5b7d17..e35bfb8 100644
> --- a/mm/mmzone.c
> +++ b/mm/mmzone.c
> @@ -87,3 +87,24 @@ int memmap_valid_within(unsigned long pfn,
>  	return 1;
>  }
>  #endif /* CONFIG_ARCH_HAS_HOLES_MEMORYMODEL */
> +
> +#ifdef CONFIG_SMP
> +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> +unsigned long zone_nr_free_pages(struct zone *zone)
> +{
> +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> +
> +	/*
> +	 * While kswapd is awake, it is considered the zone is under some
> +	 * memory pressure. Under pressure, there is a risk that
> +	 * per-cpu-counter-drift will allow the min watermark to be breached
> +	 * potentially causing a live-lock. While kswapd is awake and
> +	 * free pages are low, get a better estimate for free pages
> +	 */
> +	if (nr_free_pages < zone->percpu_drift_mark &&
> +			!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
> +		return zone_page_state_snapshot(zone, NR_FREE_PAGES);
> +
> +	return nr_free_pages;
> +}

Is this really the best way of doing it?  The way we usually solve
this problem (and boy, was this bug a newbie mistake!) is:

	foo = percpu_counter_read(x);

	if (foo says something bad) {
		/* Bad stuff: let's get a more accurate foo */
		foo = percpu_counter_sum(x);
	}

	if (foo still says something bad)
		do_bad_thing();

In other words, don't do all this stuff with percpu_drift_mark and the
kswapd heuristic.  Just change zone_watermark_ok() to use the more
accurate read if it's about to return "no".

[PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

From: Christoph Lameter <cl@linux.com>

Ordinarily watermark checks are based on the vmstat NR_FREE_PAGES as
it is cheaper than scanning a number of lists. To avoid synchronization
overhead, counter deltas are maintained on a per-cpu basis and drained both
periodically and when the delta is above a threshold. On large CPU systems,
the difference between the estimated and real value of NR_FREE_PAGES can
be very high. If NR_FREE_PAGES is much higher than number of real free page
in buddy, the VM can allocate pages below min watermark, at worst reducing
the real number of pages to zero. Even if the OOM killer kills some victim
for freeing memory, it may not free memory if the exit path requires a new
page resulting in livelock.

This patch introduces a zone_page_state_snapshot() function (courtesy of
Christoph) that takes a slightly more accurate of an arbitrary vmstat counter.
It is used to read NR_FREE_PAGES while kswapd is awake to avoid the watermark
being accidentally broken.  The estimate is not perfect and may result
in cache line bounces but is expected to be lighter than the IPI calls
necessary to continually drain the per-cpu counters while kswapd is awake.

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
---
 include/linux/mmzone.h |   13 +++++++++++++
 include/linux/vmstat.h |   22 ++++++++++++++++++++++
 mm/mmzone.c            |   21 +++++++++++++++++++++
 mm/page_alloc.c        |    4 ++--
 mm/vmstat.c            |   15 ++++++++++++++-
 5 files changed, 72 insertions(+), 3 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 6e6e626..3984c4e 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -284,6 +284,13 @@ struct zone {
 	unsigned long watermark[NR_WMARK];
 
 	/*
+	 * When free pages are below this point, additional steps are taken
+	 * when reading the number of free pages to avoid per-cpu counter
+	 * drift allowing watermarks to be breached
+	 */
+	unsigned long percpu_drift_mark;
+
+	/*
 	 * We don't know if the memory that we're going to allocate will be freeable
 	 * or/and it will be released eventually, so to avoid totally wasting several
 	 * GB of ram we must reserve some of the lower zone memory (otherwise we risk
@@ -441,6 +448,12 @@ static inline int zone_is_oom_locked(const struct zone *zone)
 	return test_bit(ZONE_OOM_LOCKED, &zone->flags);
 }
 
+#ifdef CONFIG_SMP
+unsigned long zone_nr_free_pages(struct zone *zone);
+#else
+#define zone_nr_free_pages(zone) zone_page_state(zone, NR_FREE_PAGES)
+#endif /* CONFIG_SMP */
+
 /*
  * The "priority" of VM scanning is how much of the queues we will scan in one
  * go. A value of 12 for DEF_PRIORITY implies that we will scan 1/4096th of the
diff --git a/include/linux/vmstat.h b/include/linux/vmstat.h
index 7f43ccd..eaaea37 100644
--- a/include/linux/vmstat.h
+++ b/include/linux/vmstat.h
@@ -170,6 +170,28 @@ static inline unsigned long zone_page_state(struct zone *zone,
 	return x;
 }
 
+/*
+ * More accurate version that also considers the currently pending
+ * deltas. For that we need to loop over all cpus to find the current
+ * deltas. There is no synchronization so the result cannot be
+ * exactly accurate either.
+ */
+static inline unsigned long zone_page_state_snapshot(struct zone *zone,
+					enum zone_stat_item item)
+{
+	long x = atomic_long_read(&zone->vm_stat[item]);
+
+#ifdef CONFIG_SMP
+	int cpu;
+	for_each_online_cpu(cpu)
+		x += per_cpu_ptr(zone->pageset, cpu)->vm_stat_diff[item];
+
+	if (x < 0)
+		x = 0;
+#endif
+	return x;
+}
+
 extern unsigned long global_reclaimable_pages(void);
 extern unsigned long zone_reclaimable_pages(struct zone *zone);
 
diff --git a/mm/mmzone.c b/mm/mmzone.c
index f5b7d17..e35bfb8 100644
--- a/mm/mmzone.c
+++ b/mm/mmzone.c
@@ -87,3 +87,24 @@ int memmap_valid_within(unsigned long pfn,
 	return 1;
 }
 #endif /* CONFIG_ARCH_HAS_HOLES_MEMORYMODEL */
+
+#ifdef CONFIG_SMP
+/* Called when a more accurate view of NR_FREE_PAGES is needed */
+unsigned long zone_nr_free_pages(struct zone *zone)
+{
+	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
+
+	/*
+	 * While kswapd is awake, it is considered the zone is under some
+	 * memory pressure. Under pressure, there is a risk that
+	 * per-cpu-counter-drift will allow the min watermark to be breached
+	 * potentially causing a live-lock. While kswapd is awake and
+	 * free pages are low, get a better estimate for free pages
+	 */
+	if (nr_free_pages < zone->percpu_drift_mark &&
+			!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
+		return zone_page_state_snapshot(zone, NR_FREE_PAGES);
+
+	return nr_free_pages;
+}
+#endif /* CONFIG_SMP */
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 97d74a0..bbaa959 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1462,7 +1462,7 @@ int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
 {
 	/* free_pages my go negative - that's OK */
 	long min = mark;
-	long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
+	long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
 	int o;
 
 	if (alloc_flags & ALLOC_HIGH)
@@ -2424,7 +2424,7 @@ void show_free_areas(void)
 			" all_unreclaimable? %s"
 			"\n",
 			zone->name,
-			K(zone_page_state(zone, NR_FREE_PAGES)),
+			K(zone_nr_free_pages(zone)),
 			K(min_wmark_pages(zone)),
 			K(low_wmark_pages(zone)),
 			K(high_wmark_pages(zone)),
diff --git a/mm/vmstat.c b/mm/vmstat.c
index f389168..696cab2 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -138,11 +138,24 @@ static void refresh_zone_stat_thresholds(void)
 	int threshold;
 
 	for_each_populated_zone(zone) {
+		unsigned long max_drift, tolerate_drift;
+
 		threshold = calculate_threshold(zone);
 
 		for_each_online_cpu(cpu)
 			per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 							= threshold;
+
+		/*
+		 * Only set percpu_drift_mark if there is a danger that
+		 * NR_FREE_PAGES reports the low watermark is ok when in fact
+		 * the min watermark could be breached by an allocation
+		 */
+		tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
+		max_drift = num_online_cpus() * threshold;
+		if (max_drift > tolerate_drift)
+			zone->percpu_drift_mark = high_wmark_pages(zone) +
+					max_drift;
 	}
 }
 
@@ -813,7 +826,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 		   "\n        scanned  %lu"
 		   "\n        spanned  %lu"
 		   "\n        present  %lu",
-		   zone_page_state(zone, NR_FREE_PAGES),
+		   zone_nr_free_pages(zone),
 		   min_wmark_pages(zone),
 		   low_wmark_pages(zone),
 		   high_wmark_pages(zone),
-- 
1.7.1

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Wed, Sep 01, 2010 at 07:43:41PM -0500, Christoph Lameter wrote:
> On Tue, 31 Aug 2010, Mel Gorman wrote:
> 
> > +#ifdef CONFIG_SMP
> > +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> > +unsigned long zone_nr_free_pages(struct zone *zone)
> > +{
> > +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> 
> You cannot call zone_page_state here because zone_page_state clips the
> counter at zero. The nr_free_pages needs to reflect the unclipped state
> and then the deltas need to be added. Then the clipping at zero can be
> done.
> 

Good point. This justifies the use of a generic helper that is co-located
with vmstat.h. I've taken your zone_page_state_snapshot() patch, am using
the helper to take a more accurate reading of NR_FREE_PAGES and preparing
for a test.  Thanks Christoph.

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Wed, 1 Sep 2010, Christoph Lameter wrote:

> The effect needs to be the same as retrieving a global or
> zone ZVC counter. Which is currently implemented in the following way:
>
> static inline unsigned long zone_page_state(struct zone *zone,
>                                         enum zone_stat_item item)
> {
>         long x = atomic_long_read(&zone->vm_stat[item]);
> #ifdef CONFIG_SMP
>         if (x < 0)
>                 x = 0;
> #endif
>         return x;
> }
>

Here is a patch that defined a snapshot function that works in the same
way:

Subject: Add a snapshot function for vm statistics

Add a snapshot function that can more accurately determine
the current value of a zone counter.

Signed-off-by: Christoph Lameter <cl@linux.com>


Index: linux-2.6/include/linux/vmstat.h
===================================================================
--- linux-2.6.orig/include/linux/vmstat.h	2010-09-01 19:45:23.506071189 -0500
+++ linux-2.6/include/linux/vmstat.h	2010-09-01 19:53:02.978979081 -0500
@@ -170,6 +170,28 @@
 	return x;
 }

+/*
+ * More accurate version that also considers the currently pending
+ * deltas. For that we need to loop over all cpus to find the current
+ * deltas. There is no synchronization so the result cannot be
+ * exactly accurate either.
+ */
+static inline unsigned long zone_page_state_snapshot(struct zone *zone,
+					enum zone_stat_item item)
+{
+	int cpu;
+	long x = atomic_long_read(&zone->vm_stat[item]);
+
+#ifdef CONFIG_SMP
+	for_each_online_cpu(cpu)
+		x += per_cpu_ptr(zone->pageset, cpu)->vm_stat_diff[item];
+
+	if (x < 0)
+		x = 0;
+#endif
+	return x;
+}
+
 extern unsigned long global_reclaimable_pages(void);
 extern unsigned long zone_reclaimable_pages(struct zone *zone);

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[KOSAKI Motohiro]

> On Tue, 31 Aug 2010, Mel Gorman wrote:
> 
> > +#ifdef CONFIG_SMP
> > +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> > +unsigned long zone_nr_free_pages(struct zone *zone)
> > +{
> > +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> 
> You cannot call zone_page_state here because zone_page_state clips the
> counter at zero. The nr_free_pages needs to reflect the unclipped state
> and then the deltas need to be added. Then the clipping at zero can be
> done.

Good spotting. you are right.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Tue, 31 Aug 2010, Mel Gorman wrote:

> +#ifdef CONFIG_SMP
> +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> +unsigned long zone_nr_free_pages(struct zone *zone)
> +{
> +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);

You cannot call zone_page_state here because zone_page_state clips the
counter at zero. The nr_free_pages needs to reflect the unclipped state
and then the deltas need to be added. Then the clipping at zero can be
done.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Thu, 2 Sep 2010, KOSAKI Motohiro wrote:

> > 	nr_free_pages += delta;
> > 	if (nr_free_pages < 0)
> > 		nr_free_pages = 0;
>
> nr_free_pages is unsined. this wouldn't works ;)

The VM counters are signed and must be signed otherwise the deferred
update scheme would cause desasters. For treatment in the page allocator
these may be converted to unsigned.

The effect needs to be the same as retrieving a global or
zone ZVC counter. Which is currently implemented in the following way:

static inline unsigned long zone_page_state(struct zone *zone,
                                        enum zone_stat_item item)
{
        long x = atomic_long_read(&zone->vm_stat[item]);
#ifdef CONFIG_SMP
        if (x < 0)
                x = 0;
#endif
        return x;
}

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[KOSAKI Motohiro]

> On Wed, 1 Sep 2010, Mel Gorman wrote:
> 
> > > > >         if (delta < 0 && abs(delta) > nr_free_pages)
> > > > >                 delta = -nr_free_pages;
> > >
> > > Not sure what the point here is. If the delta is going below zero then
> > > there was a concurrent operation updating the counters negatively while
> > > we summed up the counters.
> >
> > The point is if the negative delta is greater than the current value of
> > nr_free_pages then nr_free_pages would underflow when delta is applied to it.
> 
> Ok. then
> 
> 	nr_free_pages += delta;
> 	if (nr_free_pages < 0)
> 		nr_free_pages = 0;

nr_free_pages is unsined. this wouldn't works ;)

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Wed, 1 Sep 2010, Mel Gorman wrote:

> > > >         if (delta < 0 && abs(delta) > nr_free_pages)
> > > >                 delta = -nr_free_pages;
> >
> > Not sure what the point here is. If the delta is going below zero then
> > there was a concurrent operation updating the counters negatively while
> > we summed up the counters.
>
> The point is if the negative delta is greater than the current value of
> nr_free_pages then nr_free_pages would underflow when delta is applied to it.

Ok. then

	nr_free_pages += delta;
	if (nr_free_pages < 0)
		nr_free_pages = 0;

> > would be correct.
>
> Lets say the reading at the start for nr_free_pages is 120 and the delta is
> -20, then the estimated true value of nr_free_pages is 100. If we used your
> logic, the estimate would be 120. Maybe I'm missing what you're saying.

Well yes the sum of the counter needs to be checked not just the sum of
the deltas. This is the same as the counter determination in vmstat.h

> > See also handling of counter underflow in
> > vmstat.h:zone_page_state().
>
> I'm not seeing the relation. zone_nr_free_pages() is trying to
> reconcile the reading from zone_page_state() with the contents of
> vm_stat_diff[].

Both are determinations of a counter value. The global or zone counters
can also temporarily go below zero due to deferred updates. If
this happens then 0 will be returned(!). zonr_nr_free_pages need to work
in the same way.

> > As I have said before: I would rather have the
> > counter handling in one place to avoid creating differences in counter
> > handling.
> >
>
> And I'd rather not hurt the paths for every counter unnecessarily
> without good cause. I can move zone_nr_free_pages() to mm/vmstat.c if
> you'd prefer?

Generalize it on the way please to work with any counter?

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Wed, Sep 01, 2010 at 03:16:59PM -0500, Christoph Lameter wrote:
> On Wed, 1 Sep 2010, KOSAKI Motohiro wrote:
> 
> > > How about the following? It records a delta and checks if delta is negative
> > > and would cause underflow.
> > >
> > > unsigned long zone_nr_free_pages(struct zone *zone)
> > > {
> > >         unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> > >         long delta = 0;
> > >
> > >         /*
> > >          * While kswapd is awake, it is considered the zone is under some
> > >          * memory pressure. Under pressure, there is a risk that
> > >          * per-cpu-counter-drift will allow the min watermark to be breached
> > >          * potentially causing a live-lock. While kswapd is awake and
> > >          * free pages are low, get a better estimate for free pages
> > >          */
> > >         if (nr_free_pages < zone->percpu_drift_mark &&
> > >                         !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
> > >                 int cpu;
> > >
> > >                 for_each_online_cpu(cpu) {
> > >                         struct per_cpu_pageset *pset;
> > >
> > >                         pset = per_cpu_ptr(zone->pageset, cpu);
> > >                         delta += pset->vm_stat_diff[NR_FREE_PAGES];
> > >                 }
> > >         }
> > >
> > >         /* Watch for underflow */
> > >         if (delta < 0 && abs(delta) > nr_free_pages)
> > >                 delta = -nr_free_pages;
> 
> Not sure what the point here is. If the delta is going below zero then
> there was a concurrent operation updating the counters negatively while
> we summed up the counters.

The point is if the negative delta is greater than the current value of
nr_free_pages then nr_free_pages would underflow when delta is applied to it.

> It is then safe to assume a value of zero. We
> cannot really be more accurate than that.
> 
> so
> 
> 	if (delta < 0)
> 		delta = 0;
> 
> would be correct.

Lets say the reading at the start for nr_free_pages is 120 and the delta is
-20, then the estimated true value of nr_free_pages is 100. If we used your
logic, the estimate would be 120. Maybe I'm missing what you're saying.

> See also handling of counter underflow in
> vmstat.h:zone_page_state().

I'm not seeing the relation. zone_nr_free_pages() is trying to
reconcile the reading from zone_page_state() with the contents of
vm_stat_diff[].

> As I have said before: I would rather have the
> counter handling in one place to avoid creating differences in counter
> handling.
> 

And I'd rather not hurt the paths for every counter unnecessarily
without good cause. I can move zone_nr_free_pages() to mm/vmstat.c if
you'd prefer?

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

On Wed, 1 Sep 2010, KOSAKI Motohiro wrote:

> > How about the following? It records a delta and checks if delta is negative
> > and would cause underflow.
> >
> > unsigned long zone_nr_free_pages(struct zone *zone)
> > {
> >         unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> >         long delta = 0;
> >
> >         /*
> >          * While kswapd is awake, it is considered the zone is under some
> >          * memory pressure. Under pressure, there is a risk that
> >          * per-cpu-counter-drift will allow the min watermark to be breached
> >          * potentially causing a live-lock. While kswapd is awake and
> >          * free pages are low, get a better estimate for free pages
> >          */
> >         if (nr_free_pages < zone->percpu_drift_mark &&
> >                         !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
> >                 int cpu;
> >
> >                 for_each_online_cpu(cpu) {
> >                         struct per_cpu_pageset *pset;
> >
> >                         pset = per_cpu_ptr(zone->pageset, cpu);
> >                         delta += pset->vm_stat_diff[NR_FREE_PAGES];
> >                 }
> >         }
> >
> >         /* Watch for underflow */
> >         if (delta < 0 && abs(delta) > nr_free_pages)
> >                 delta = -nr_free_pages;

Not sure what the point here is. If the delta is going below zero then
there was a concurrent operation updating the counters negatively while
we summed up the counters. It is then safe to assume a value of zero. We
cannot really be more accurate than that.

so

	if (delta < 0)
		delta = 0;

would be correct. See also handling of counter underflow in
vmstat.h:zone_page_state(). As I have said before: I would rather have the
counter handling in one place to avoid creating differences in counter
handling.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[KOSAKI Motohiro]

> On Wed, Sep 01, 2010 at 08:37:41AM +0900, KOSAKI Motohiro wrote:
> > > +#ifdef CONFIG_SMP
> > > +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> > > +unsigned long zone_nr_free_pages(struct zone *zone)
> > > +{
> > > +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> > > +
> > > +	/*
> > > +	 * While kswapd is awake, it is considered the zone is under some
> > > +	 * memory pressure. Under pressure, there is a risk that
> > > +	 * per-cpu-counter-drift will allow the min watermark to be breached
> > > +	 * potentially causing a live-lock. While kswapd is awake and
> > > +	 * free pages are low, get a better estimate for free pages
> > > +	 */
> > > +	if (nr_free_pages < zone->percpu_drift_mark &&
> > > +			!waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
> > > +		int cpu;
> > > +
> > > +		for_each_online_cpu(cpu) {
> > > +			struct per_cpu_pageset *pset;
> > > +
> > > +			pset = per_cpu_ptr(zone->pageset, cpu);
> > > +			nr_free_pages += pset->vm_stat_diff[NR_FREE_PAGES];
> > 
> > If my understanding is correct, we have no lock when reading pset->vm_stat_diff.
> > It mean nr_free_pages can reach negative value at very rarely race. boundary
> > check is necessary?
> > 
> 
> True, well spotted.
> 
> How about the following? It records a delta and checks if delta is negative
> and would cause underflow.
> 
> unsigned long zone_nr_free_pages(struct zone *zone)
> {
>         unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
>         long delta = 0;
> 
>         /*
>          * While kswapd is awake, it is considered the zone is under some
>          * memory pressure. Under pressure, there is a risk that
>          * per-cpu-counter-drift will allow the min watermark to be breached
>          * potentially causing a live-lock. While kswapd is awake and
>          * free pages are low, get a better estimate for free pages
>          */
>         if (nr_free_pages < zone->percpu_drift_mark &&
>                         !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
>                 int cpu;
> 
>                 for_each_online_cpu(cpu) {
>                         struct per_cpu_pageset *pset;
> 
>                         pset = per_cpu_ptr(zone->pageset, cpu);
>                         delta += pset->vm_stat_diff[NR_FREE_PAGES];
>                 }
>         }
> 
>         /* Watch for underflow */
>         if (delta < 0 && abs(delta) > nr_free_pages)
>                 delta = -nr_free_pages;
> 
>         return nr_free_pages + delta;
> }

Looks good to me :)
	Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>

Thanks.

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

On Wed, Sep 01, 2010 at 08:37:41AM +0900, KOSAKI Motohiro wrote:
> > +#ifdef CONFIG_SMP
> > +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> > +unsigned long zone_nr_free_pages(struct zone *zone)
> > +{
> > +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> > +
> > +	/*
> > +	 * While kswapd is awake, it is considered the zone is under some
> > +	 * memory pressure. Under pressure, there is a risk that
> > +	 * per-cpu-counter-drift will allow the min watermark to be breached
> > +	 * potentially causing a live-lock. While kswapd is awake and
> > +	 * free pages are low, get a better estimate for free pages
> > +	 */
> > +	if (nr_free_pages < zone->percpu_drift_mark &&
> > +			!waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
> > +		int cpu;
> > +
> > +		for_each_online_cpu(cpu) {
> > +			struct per_cpu_pageset *pset;
> > +
> > +			pset = per_cpu_ptr(zone->pageset, cpu);
> > +			nr_free_pages += pset->vm_stat_diff[NR_FREE_PAGES];
> 
> If my understanding is correct, we have no lock when reading pset->vm_stat_diff.
> It mean nr_free_pages can reach negative value at very rarely race. boundary
> check is necessary?
> 

True, well spotted.

How about the following? It records a delta and checks if delta is negative
and would cause underflow.

unsigned long zone_nr_free_pages(struct zone *zone)
{
        unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
        long delta = 0;

        /*
         * While kswapd is awake, it is considered the zone is under some
         * memory pressure. Under pressure, there is a risk that
         * per-cpu-counter-drift will allow the min watermark to be breached
         * potentially causing a live-lock. While kswapd is awake and
         * free pages are low, get a better estimate for free pages
         */
        if (nr_free_pages < zone->percpu_drift_mark &&
                        !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
                int cpu;

                for_each_online_cpu(cpu) {
                        struct per_cpu_pageset *pset;

                        pset = per_cpu_ptr(zone->pageset, cpu);
                        delta += pset->vm_stat_diff[NR_FREE_PAGES];
                }
        }

        /* Watch for underflow */
        if (delta < 0 && abs(delta) > nr_free_pages)
                delta = -nr_free_pages;

        return nr_free_pages + delta;
}

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

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[KOSAKI Motohiro]

> +#ifdef CONFIG_SMP
> +/* Called when a more accurate view of NR_FREE_PAGES is needed */
> +unsigned long zone_nr_free_pages(struct zone *zone)
> +{
> +	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
> +
> +	/*
> +	 * While kswapd is awake, it is considered the zone is under some
> +	 * memory pressure. Under pressure, there is a risk that
> +	 * per-cpu-counter-drift will allow the min watermark to be breached
> +	 * potentially causing a live-lock. While kswapd is awake and
> +	 * free pages are low, get a better estimate for free pages
> +	 */
> +	if (nr_free_pages < zone->percpu_drift_mark &&
> +			!waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
> +		int cpu;
> +
> +		for_each_online_cpu(cpu) {
> +			struct per_cpu_pageset *pset;
> +
> +			pset = per_cpu_ptr(zone->pageset, cpu);
> +			nr_free_pages += pset->vm_stat_diff[NR_FREE_PAGES];

If my understanding is correct, we have no lock when reading pset->vm_stat_diff.
It mean nr_free_pages can reach negative value at very rarely race. boundary
check is necessary?

Re: [PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Christoph Lameter]

Reviewed-by: Christoph Lameter <cl@linux.com>

[PATCH 2/3] mm: page allocator: Calculate a better estimate of NR_FREE_PAGES when memory is low and kswapd is awake

[Mel Gorman]

Ordinarily watermark checks are based on the vmstat NR_FREE_PAGES as
it is cheaper than scanning a number of lists. To avoid synchronization
overhead, counter deltas are maintained on a per-cpu basis and drained both
periodically and when the delta is above a threshold. On large CPU systems,
the difference between the estimated and real value of NR_FREE_PAGES can be
very high.  If NR_FREE_PAGES is much higher than number of real free page
in buddy, the VM can allocate pages below min watermark, at worst reducing
the real number of pages to zero.  Even if the OOM killer kills some victim
for freeing memory, it may not free memory if the exit path requires a new
page resulting in livelock.

This patch introduces zone_nr_free_pages() to take a slightly more accurate
estimate of NR_FREE_PAGES while kswapd is awake. The estimate is not perfect
and may result in cache line bounces but is expected to be lighter than the
IPI calls necessary to continually drain the per-cpu counters while kswapd
is awake.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
---
 include/linux/mmzone.h |   13 +++++++++++++
 mm/mmzone.c            |   29 +++++++++++++++++++++++++++++
 mm/page_alloc.c        |    4 ++--
 mm/vmstat.c            |   15 ++++++++++++++-
 4 files changed, 58 insertions(+), 3 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 6e6e626..3984c4e 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -284,6 +284,13 @@ struct zone {
 	unsigned long watermark[NR_WMARK];
 
 	/*
+	 * When free pages are below this point, additional steps are taken
+	 * when reading the number of free pages to avoid per-cpu counter
+	 * drift allowing watermarks to be breached
+	 */
+	unsigned long percpu_drift_mark;
+
+	/*
 	 * We don't know if the memory that we're going to allocate will be freeable
 	 * or/and it will be released eventually, so to avoid totally wasting several
 	 * GB of ram we must reserve some of the lower zone memory (otherwise we risk
@@ -441,6 +448,12 @@ static inline int zone_is_oom_locked(const struct zone *zone)
 	return test_bit(ZONE_OOM_LOCKED, &zone->flags);
 }
 
+#ifdef CONFIG_SMP
+unsigned long zone_nr_free_pages(struct zone *zone);
+#else
+#define zone_nr_free_pages(zone) zone_page_state(zone, NR_FREE_PAGES)
+#endif /* CONFIG_SMP */
+
 /*
  * The "priority" of VM scanning is how much of the queues we will scan in one
  * go. A value of 12 for DEF_PRIORITY implies that we will scan 1/4096th of the
diff --git a/mm/mmzone.c b/mm/mmzone.c
index f5b7d17..69ecbe9 100644
--- a/mm/mmzone.c
+++ b/mm/mmzone.c
@@ -87,3 +87,32 @@ int memmap_valid_within(unsigned long pfn,
 	return 1;
 }
 #endif /* CONFIG_ARCH_HAS_HOLES_MEMORYMODEL */
+
+#ifdef CONFIG_SMP
+/* Called when a more accurate view of NR_FREE_PAGES is needed */
+unsigned long zone_nr_free_pages(struct zone *zone)
+{
+	unsigned long nr_free_pages = zone_page_state(zone, NR_FREE_PAGES);
+
+	/*
+	 * While kswapd is awake, it is considered the zone is under some
+	 * memory pressure. Under pressure, there is a risk that
+	 * per-cpu-counter-drift will allow the min watermark to be breached
+	 * potentially causing a live-lock. While kswapd is awake and
+	 * free pages are low, get a better estimate for free pages
+	 */
+	if (nr_free_pages < zone->percpu_drift_mark &&
+			!waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
+		int cpu;
+
+		for_each_online_cpu(cpu) {
+			struct per_cpu_pageset *pset;
+
+			pset = per_cpu_ptr(zone->pageset, cpu);
+			nr_free_pages += pset->vm_stat_diff[NR_FREE_PAGES];
+		}
+	}
+
+	return nr_free_pages;
+}
+#endif /* CONFIG_SMP */
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 97d74a0..bbaa959 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1462,7 +1462,7 @@ int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
 {
 	/* free_pages my go negative - that's OK */
 	long min = mark;
-	long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
+	long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
 	int o;
 
 	if (alloc_flags & ALLOC_HIGH)
@@ -2424,7 +2424,7 @@ void show_free_areas(void)
 			" all_unreclaimable? %s"
 			"\n",
 			zone->name,
-			K(zone_page_state(zone, NR_FREE_PAGES)),
+			K(zone_nr_free_pages(zone)),
 			K(min_wmark_pages(zone)),
 			K(low_wmark_pages(zone)),
 			K(high_wmark_pages(zone)),
diff --git a/mm/vmstat.c b/mm/vmstat.c
index f389168..696cab2 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -138,11 +138,24 @@ static void refresh_zone_stat_thresholds(void)
 	int threshold;
 
 	for_each_populated_zone(zone) {
+		unsigned long max_drift, tolerate_drift;
+
 		threshold = calculate_threshold(zone);
 
 		for_each_online_cpu(cpu)
 			per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 							= threshold;
+
+		/*
+		 * Only set percpu_drift_mark if there is a danger that
+		 * NR_FREE_PAGES reports the low watermark is ok when in fact
+		 * the min watermark could be breached by an allocation
+		 */
+		tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
+		max_drift = num_online_cpus() * threshold;
+		if (max_drift > tolerate_drift)
+			zone->percpu_drift_mark = high_wmark_pages(zone) +
+					max_drift;
 	}
 }
 
@@ -813,7 +826,7 @@ static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 		   "\n        scanned  %lu"
 		   "\n        spanned  %lu"
 		   "\n        present  %lu",
-		   zone_page_state(zone, NR_FREE_PAGES),
+		   zone_nr_free_pages(zone),
 		   min_wmark_pages(zone),
 		   low_wmark_pages(zone),
 		   high_wmark_pages(zone),
-- 
1.7.1