mm: migration: take a reference to the anon_vma before migrating

This patchset is a memory compaction mechanism that reduces external
fragmentation memory by moving GFP_MOVABLE pages to a fewer number of
pageblocks.  The term "compaction" was chosen as there are is a number of
mechanisms that are not mutually exclusive that can be used to defragment
memory.  For example, lumpy reclaim is a form of defragmentation as was
slub "defragmentation" (really a form of targeted reclaim).  Hence, this
is called "compaction" to distinguish it from other forms of
defragmentation.

In this implementation, a full compaction run involves two scanners
operating within a zone - a migration and a free scanner.  The migration
scanner starts at the beginning of a zone and finds all movable pages
within one pageblock_nr_pages-sized area and isolates them on a
migratepages list.  The free scanner begins at the end of the zone and
searches on a per-area basis for enough free pages to migrate all the
pages on the migratepages list.  As each area is respectively migrated or
exhausted of free pages, the scanners are advanced one area.  A compaction
run completes within a zone when the two scanners meet.

This method is a bit primitive but is easy to understand and greater
sophistication would require maintenance of counters on a per-pageblock
basis.  This would have a big impact on allocator fast-paths to improve
compaction which is a poor trade-off.

It also does not try relocate virtually contiguous pages to be physically
contiguous.  However, assuming transparent hugepages were in use, a
hypothetical khugepaged might reuse compaction code to isolate free pages,
split them and relocate userspace pages for promotion.

Memory compaction can be triggered in one of three ways.  It may be
triggered explicitly by writing any value to /proc/sys/vm/compact_memory
and compacting all of memory.  It can be triggered on a per-node basis by
writing any value to /sys/devices/system/node/nodeN/compact where N is the
node ID to be compacted.  When a process fails to allocate a high-order
page, it may compact memory in an attempt to satisfy the allocation
instead of entering direct reclaim.  Explicit compaction does not finish
until the two scanners meet and direct compaction ends if a suitable page
becomes available that would meet watermarks.

The series is in 14 patches.  The first three are not "core" to the series
but are important pre-requisites.

Patch 1 reference counts anon_vma for rmap_walk_anon(). Without this
	patch, it's possible to use anon_vma after free if the caller is
	not holding a VMA or mmap_sem for the pages in question. While
	there should be no existing user that causes this problem,
	it's a requirement for memory compaction to be stable. The patch
	is at the start of the series for bisection reasons.
Patch 2 merges the KSM and migrate counts. It could be merged with patch 1
	but would be slightly harder to review.
Patch 3 skips over unmapped anon pages during migration as there are no
	guarantees about the anon_vma existing. There is a window between
	when a page was isolated and migration started during which anon_vma
	could disappear.
Patch 4 notes that PageSwapCache pages can still be migrated even if they
	are unmapped.
Patch 5 allows CONFIG_MIGRATION to be set without CONFIG_NUMA
Patch 6 exports a "unusable free space index" via debugfs. It's
	a measure of external fragmentation that takes the size of the
	allocation request into account. It can also be calculated from
	userspace so can be dropped if requested
Patch 7 exports a "fragmentation index" which only has meaning when an
	allocation request fails. It determines if an allocation failure
	would be due to a lack of memory or external fragmentation.
Patch 8 moves the definition for LRU isolation modes for use by compaction
Patch 9 is the compaction mechanism although it's unreachable at this point
Patch 10 adds a means of compacting all of memory with a proc trgger
Patch 11 adds a means of compacting a specific node with a sysfs trigger
Patch 12 adds "direct compaction" before "direct reclaim" if it is
	determined there is a good chance of success.
Patch 13 adds a sysctl that allows tuning of the threshold at which the
	kernel will compact or direct reclaim
Patch 14 temporarily disables compaction if an allocation failure occurs
	after compaction.

Testing of compaction was in three stages.  For the test, debugging,
preempt, the sleep watchdog and lockdep were all enabled but nothing nasty
popped out.  min_free_kbytes was tuned as recommended by hugeadm to help
fragmentation avoidance and high-order allocations.  It was tested on X86,
X86-64 and PPC64.

Ths first test represents one of the easiest cases that can be faced for
lumpy reclaim or memory compaction.

1. Machine freshly booted and configured for hugepage usage with
	a) hugeadm --create-global-mounts
	b) hugeadm --pool-pages-max DEFAULT:8G
	c) hugeadm --set-recommended-min_free_kbytes
	d) hugeadm --set-recommended-shmmax

	The min_free_kbytes here is important. Anti-fragmentation works best
	when pageblocks don't mix. hugeadm knows how to calculate a value that
	will significantly reduce the worst of external-fragmentation-related
	events as reported by the mm_page_alloc_extfrag tracepoint.

2. Load up memory
	a) Start updatedb
	b) Create in parallel a X files of pagesize*128 in size. Wait
	   until files are created. By parallel, I mean that 4096 instances
	   of dd were launched, one after the other using &. The crude
	   objective being to mix filesystem metadata allocations with
	   the buffer cache.
	c) Delete every second file so that pageblocks are likely to
	   have holes
	d) kill updatedb if it's still running

	At this point, the system is quiet, memory is full but it's full with
	clean filesystem metadata and clean buffer cache that is unmapped.
	This is readily migrated or discarded so you'd expect lumpy reclaim
	to have no significant advantage over compaction but this is at
	the POC stage.

3. In increments, attempt to allocate 5% of memory as hugepages.
	   Measure how long it took, how successful it was, how many
	   direct reclaims took place and how how many compactions. Note
	   the compaction figures might not fully add up as compactions
	   can take place for orders other than the hugepage size

X86				vanilla		compaction
Final page count                    913                916 (attempted 1002)
pages reclaimed                   68296               9791

X86-64				vanilla		compaction
Final page count:                   901                902 (attempted 1002)
Total pages reclaimed:           112599              53234

PPC64				vanilla		compaction
Final page count:                    93                 94 (attempted 110)
Total pages reclaimed:           103216              61838

There was not a dramatic improvement in success rates but it wouldn't be
expected in this case either.  What was important is that fewer pages were
reclaimed in all cases reducing the amount of IO required to satisfy a
huge page allocation.

The second tests were all performance related - kernbench, netperf, iozone
and sysbench.  None showed anything too remarkable.

The last test was a high-order allocation stress test.  Many kernel
compiles are started to fill memory with a pressured mix of unmovable and
movable allocations.  During this, an attempt is made to allocate 90% of
memory as huge pages - one at a time with small delays between attempts to
avoid flooding the IO queue.

                                             vanilla   compaction
Percentage of request allocated X86               98           99
Percentage of request allocated X86-64            95           98
Percentage of request allocated PPC64             55           70

This patch:

rmap_walk_anon() does not use page_lock_anon_vma() for looking up and
locking an anon_vma and it does not appear to have sufficient locking to
ensure the anon_vma does not disappear from under it.

This patch copies an approach used by KSM to take a reference on the
anon_vma while pages are being migrated.  This should prevent rmap_walk()
running into nasty surprises later because anon_vma has been freed.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

3 files changed, 40 insertions, 5 deletions

diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index d25bd22..567d43f 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -29,6 +29,9 @@ struct anon_vma {
 #ifdef CONFIG_KSM
 	atomic_t ksm_refcount;
 #endif
+#ifdef CONFIG_MIGRATION
+	atomic_t migrate_refcount;
+#endif
 	/*
 	 * NOTE: the LSB of the head.next is set by
 	 * mm_take_all_locks() _after_ taking the above lock. So the
@@ -81,6 +84,26 @@ static inline int ksm_refcount(struct anon_vma *anon_vma)
 	return 0;
 }
 #endif /* CONFIG_KSM */
+#ifdef CONFIG_MIGRATION
+static inline void migrate_refcount_init(struct anon_vma *anon_vma)
+{
+	atomic_set(&anon_vma->migrate_refcount, 0);
+}
+
+static inline int migrate_refcount(struct anon_vma *anon_vma)
+{
+	return atomic_read(&anon_vma->migrate_refcount);
+}
+#else
+static inline void migrate_refcount_init(struct anon_vma *anon_vma)
+{
+}
+
+static inline int migrate_refcount(struct anon_vma *anon_vma)
+{
+	return 0;
+}
+#endif /* CONFIG_MIGRATE */
 
 static inline struct anon_vma *page_anon_vma(struct page *page)
 {
diff --git a/mm/migrate.c b/mm/migrate.c
index 5938db5..b768a1d 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -543,6 +543,7 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 	int rcu_locked = 0;
 	int charge = 0;
 	struct mem_cgroup *mem = NULL;
+	struct anon_vma *anon_vma = NULL;
 
 	if (!newpage)
 		return -ENOMEM;
@@ -599,6 +600,8 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
 	if (PageAnon(page)) {
 		rcu_read_lock();
 		rcu_locked = 1;
+		anon_vma = page_anon_vma(page);
+		atomic_inc(&anon_vma->migrate_refcount);
 	}
 
 	/*
@@ -638,6 +641,15 @@ skip_unmap:
 	if (rc)
 		remove_migration_ptes(page, page);
 rcu_unlock:
+
+	/* Drop an anon_vma reference if we took one */
+	if (anon_vma && atomic_dec_and_lock(&anon_vma->migrate_refcount, &anon_vma->lock)) {
+		int empty = list_empty(&anon_vma->head);
+		spin_unlock(&anon_vma->lock);
+		if (empty)
+			anon_vma_free(anon_vma);
+	}
+
 	if (rcu_locked)
 		rcu_read_unlock();
 uncharge:
diff --git a/mm/rmap.c b/mm/rmap.c
index 0feeef8..f522cb0 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -250,7 +250,8 @@ static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
 	list_del(&anon_vma_chain->same_anon_vma);
 
 	/* We must garbage collect the anon_vma if it's empty */
-	empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma);
+	empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma) &&
+					!migrate_refcount(anon_vma);
 	spin_unlock(&anon_vma->lock);
 
 	if (empty)
@@ -275,6 +276,7 @@ static void anon_vma_ctor(void *data)
 
 	spin_lock_init(&anon_vma->lock);
 	ksm_refcount_init(anon_vma);
+	migrate_refcount_init(anon_vma);
 	INIT_LIST_HEAD(&anon_vma->head);
 }
 
@@ -1355,10 +1357,8 @@ static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
 	/*
 	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
 	 * because that depends on page_mapped(); but not all its usages
-	 * are holding mmap_sem, which also gave the necessary guarantee
-	 * (that this anon_vma's slab has not already been destroyed).
-	 * This needs to be reviewed later: avoiding page_lock_anon_vma()
-	 * is risky, and currently limits the usefulness of rmap_walk().
+	 * are holding mmap_sem. Users without mmap_sem are required to
+	 * take a reference count to prevent the anon_vma disappearing
 	 */
 	anon_vma = page_anon_vma(page);
 	if (!anon_vma)