[RFC 5/7] Slab defrag core

[Christopher Lameter]

Slab defragmentation may occur:

1. Unconditionally when kmem_cache_shrink is called on a slab cache by the
   kernel calling kmem_cache_shrink.

2. Through the use of the slabinfo command.

3. Per node defrag conditionally when kmem_cache_defrag(<node>) is called
   (can be called from reclaim code with a later patch).

   Defragmentation is only performed if the fragmentation of the slab
   is lower than the specified percentage. Fragmentation ratios are measured
   by calculating the percentage of objects in use compared to the total
   number of objects that the slab page can accomodate.

   The scanning of slab caches is optimized because the
   defragmentable slabs come first on the list. Thus we can terminate scans
   on the first slab encountered that does not support defragmentation.

   kmem_cache_defrag() takes a node parameter. This can either be -1 if
   defragmentation should be performed on all nodes, or a node number.

A couple of functions must be setup via a call to kmem_cache_setup_defrag()
in order for a slabcache to support defragmentation. These are

kmem_defrag_isolate_func (void *isolate(struct kmem_cache *s, void **objects, int nr))

	Must stabilize that the objects and ensure that they will not be freed until
	the migration function is complete. SLUB guarantees that
	the objects are still allocated. However, other threads may be blocked
	in slab_free() attempting to free objects in the slab. These may succeed
	as soon as isolate() returns to the slab allocator. The function must
	be able to detect such situations and void the attempts to free such
	objects (by for example voiding the corresponding entry in the objects
	array).

	No slab operations may be performed in isolate(). Interrupts
	are disabled. What can be done is very limited. The slab lock
	for the page that contains the object is taken. Any attempt to perform
	a slab operation may lead to a deadlock.

	kmem_defrag_isolate_func returns a private pointer that is passed to
	kmem_defrag_kick_func(). Should we be unable to obtain all references
	then that pointer may indicate to the kick() function that it should
	not attempt any object removal or move but simply undo the measure
	that were used to stabilize the object.

kmem_defrag_migrate_func (void migrate(struct kmem_cache *, void **objects, int nr,
			int node, void *get_result))

	After SLUB has stabilzed the objects in a
	slab it will then drop all locks and use migrate() to move objects out
	of the slab. The existence of the object is guaranteed by virtue of
	the earlier obtained references via kmem_defrag_get_func(). The
	callback may perform any slab operation since no locks are held at
	the time of call.

	The callback should remove the object from the slab in some way. This
	may be accomplished by reclaiming the object and then running
	kmem_cache_free() or reallocating it and then running
	kmem_cache_free(). Reallocation is advantageous because the partial
	slabs were just sorted to have the partial slabs with the most objects
	first. Reallocation is likely to result in filling up a slab in
	addition to freeing up one slab. A filled up slab can also be removed
	from the partial list. So there could be a double effect.

	kmem_defrag_migrate_func() does not return a result. SLUB will check
	the number of remaining objects in the slab. If all objects were
	removed then the slab is freed and we have reduced the overall
	fragmentation of the slab cache.

Signed-off-by: Christoph Lameter <cl@xxxxxxxxx>

---
 include/linux/slab.h |    3
 mm/slub.c            |  265 ++++++++++++++++++++++++++++++++++++++++-----------
 2 files changed, 215 insertions(+), 53 deletions(-)

Index: linux/mm/slub.c
===================================================================
--- linux.orig/mm/slub.c
+++ linux/mm/slub.c
@@ -351,6 +351,12 @@ static __always_inline void slab_lock(st
 	bit_spin_lock(PG_locked, &page->flags);
 }

+static __always_inline int slab_trylock(struct page *page)
+{
+	VM_BUG_ON_PAGE(PageTail(page), page);
+	return bit_spin_trylock(PG_locked, &page->flags);
+}
+
 static __always_inline void slab_unlock(struct page *page)
 {
 	VM_BUG_ON_PAGE(PageTail(page), page);
@@ -3946,79 +3952,6 @@ void kfree(const void *x)
 }
 EXPORT_SYMBOL(kfree);

-#define SHRINK_PROMOTE_MAX 32
-
-/*
- * kmem_cache_shrink discards empty slabs and promotes the slabs filled
- * up most to the head of the partial lists. New allocations will then
- * fill those up and thus they can be removed from the partial lists.
- *
- * The slabs with the least items are placed last. This results in them
- * being allocated from last increasing the chance that the last objects
- * are freed in them.
- */
-int __kmem_cache_shrink(struct kmem_cache *s)
-{
-	int node;
-	int i;
-	struct kmem_cache_node *n;
-	struct page *page;
-	struct page *t;
-	struct list_head discard;
-	struct list_head promote[SHRINK_PROMOTE_MAX];
-	unsigned long flags;
-	int ret = 0;
-
-	flush_all(s);
-	for_each_kmem_cache_node(s, node, n) {
-		INIT_LIST_HEAD(&discard);
-		for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
-			INIT_LIST_HEAD(promote + i);
-
-		spin_lock_irqsave(&n->list_lock, flags);
-
-		/*
-		 * Build lists of slabs to discard or promote.
-		 *
-		 * Note that concurrent frees may occur while we hold the
-		 * list_lock. page->inuse here is the upper limit.
-		 */
-		list_for_each_entry_safe(page, t, &n->partial, lru) {
-			int free = page->objects - page->inuse;
-
-			/* Do not reread page->inuse */
-			barrier();
-
-			/* We do not keep full slabs on the list */
-			BUG_ON(free <= 0);
-
-			if (free == page->objects) {
-				list_move(&page->lru, &discard);
-				n->nr_partial--;
-			} else if (free <= SHRINK_PROMOTE_MAX)
-				list_move(&page->lru, promote + free - 1);
-		}
-
-		/*
-		 * Promote the slabs filled up most to the head of the
-		 * partial list.
-		 */
-		for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
-			list_splice(promote + i, &n->partial);
-
-		spin_unlock_irqrestore(&n->list_lock, flags);
-
-		/* Release empty slabs */
-		list_for_each_entry_safe(page, t, &discard, lru)
-			discard_slab(s, page);
-
-		if (slabs_node(s, node))
-			ret = 1;
-	}
-
-	return ret;
-}
-
 #ifdef CONFIG_MEMCG
 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
 {
@@ -4325,14 +4258,271 @@ static inline void *alloc_scratch(void)
 		GFP_KERNEL);
 }

+/*
+ * Move all objects in the given slab.
+ *
+ * If the target node is the current node then the object is moved else
+ * where on the same node. Which is an effective way of defragmentation
+ * since the current slab page with its object is exempt from allocation.
+ *
+ * The scratch area passed to list function is sufficient to hold
+ * struct listhead times objects per slab. We use it to hold void ** times
+ * objects per slab plus a bitmap for each object.
+ */
+static void kmem_cache_move(struct page *page, void *scratch, int node)
+{
+	void **vector = scratch;
+	void *p;
+	void *addr = page_address(page);
+	struct kmem_cache *s;
+	unsigned long *map;
+	int count;
+	void *private;
+	unsigned long flags;
+	unsigned long objects;
+
+	local_irq_save(flags);
+	slab_lock(page);
+
+	BUG_ON(!PageSlab(page));	/* Must be s slab page */
+	BUG_ON(!page->frozen);	/* Slab must have been frozen earlier */
+
+	s = page->slab_cache;
+	objects = page->objects;
+	map = scratch + objects * sizeof(void **);
+
+	/* Determine used objects */
+	bitmap_fill(map, objects);
+	for (p = page->freelist; p; p = get_freepointer(s, p))
+		__clear_bit(slab_index(p, s, addr), map);
+
+	/* Build vector of pointers to objects */
+	count = 0;
+	memset(vector, 0, objects * sizeof(void **));
+	for_each_object(p, s, addr, objects)
+		if (test_bit(slab_index(p, s, addr), map))
+			vector[count++] = p;
+
+	if (s->isolate)
+		private = s->isolate(s, vector, count);
+	else
+		/*
+		 * Objects do not need to be isolated.
+		 */
+		private = NULL;
+
+	/*
+	 * Pinned the objects. Now we can drop the slab lock. The slab
+	 * is frozen so it cannot vanish from under us nor will
+	 * allocations be performed on the slab. However, unlocking the
+	 * slab will allow concurrent slab_frees to proceed. So
+	 * the subsystem must have a way to tell from the content
+	 * of the object that it was freed.
+	 *
+	 * If neither RCU nor ctor is being used then the object
+	 * may be modified by the allocator after being freed
+	 * which may disrupt the ability of the migrate function
+	 * to tell if the object is free or not.
+	 */
+	slab_unlock(page);
+	local_irq_restore(flags);
+
+	/*
+	 * Perform callbacks to move the objects.
+	 */
+	s->migrate(s, vector, count, node, private);
+}
+
+/*
+ * Move slab objects on a particular node of the cache.
+ * Release slabs with zero objects and tryg to call the move function for
+ * slabs with less than the configured percentage of objects allocated.
+ *
+ * Returns the number of slabs left on the node after the operation.
+ */
+static unsigned long __move(struct kmem_cache *s, int node,
+		int target_node, int ratio)
+{
+	unsigned long flags;
+	struct page *page, *page2;
+	LIST_HEAD(move_list);
+	struct kmem_cache_node *n = get_node(s, node);
+
+	if (node == target_node && n->nr_partial <= 1)
+		/*
+		 * Trying to reduce fragmentataion on a node but there is
+		 * only a single or no partial slab page. This is already
+		 * the optimal object density that we can reach
+		 */
+		goto out;
+
+	spin_lock_irqsave(&n->list_lock, flags);
+	list_for_each_entry_safe(page, page2, &n->partial, lru) {
+		if (!slab_trylock(page))
+			/* Busy slab. Get out of the way */
+			continue;
+
+		if (page->inuse) {
+			if (page->inuse > ratio * page->objects / 100) {
+				slab_unlock(page);
+				/*
+				 * Skip slab because the object density
+				 * in the slab page is high enough
+				*/
+				continue;
+			}
+
+			list_move(&page->lru, &move_list);
+			if (s->migrate) {
+				/* Remove page from being considered for allocations */
+				n->nr_partial--;
+				page->frozen = 1;
+			}
+			slab_unlock(page);
+		} else {
+			/* Empty slab page */
+			list_del(&page->lru);
+			n->nr_partial--;
+			slab_unlock(page);
+			discard_slab(s, page);
+		}
+	}
+
+	if (!s->migrate)
+		/*
+		 * No defrag method. By simply putting the zaplist at the
+		 * end of the partial list we can let them simmer longer
+		 * and thus increase the chance of all objects being
+		 * reclaimed.
+		 *
+		 * We have effectively sorted the partial list and put
+		 * the slabs with more objects first. As soon as they
+		 * are allocated they are going to be removed from the
+		 * partial list.
+		 */
+		list_splice(&move_list, n->partial.prev);
+
+
+	spin_unlock_irqrestore(&n->list_lock, flags);
+
+	if (s->migrate && !list_empty(&move_list)) {
+		void **scratch = alloc_scratch();
+		struct page *page;
+		struct page *page2;
+
+		if (scratch) {
+			/* Try to remove / move the objects left */
+			list_for_each_entry(page, &move_list, lru) {
+				if (page->inuse)
+					kmem_cache_move(page, scratch, target_node);
+			}
+			kfree(scratch);
+		}
+
+		/* Inspect results and dispose of pages */
+		spin_lock_irqsave(&n->list_lock, flags);
+		list_for_each_entry_safe(page, page2, &move_list, lru) {
+			list_del(&page->lru);
+			slab_lock(page);
+			page->frozen = 0;
+
+			if (page->inuse) {
+				/*
+				 * Objects left in slab page move it to
+				 * the tail of the partial list to
+				 * increase the change that the freeing
+				 * of the remaining objects will
+				 * free the slab page
+				 */
+				n->nr_partial++;
+				list_add_tail(&page->lru, &n->partial);
+				slab_unlock(page);
+
+			} else {
+				slab_unlock(page);
+				discard_slab(s, page);
+			}
+		}
+		spin_unlock_irqrestore(&n->list_lock, flags);
+	}
+out:
+	return atomic_long_read(&n->nr_slabs);
+}
+
+/*
+ * Defrag slabs conditional on the amount of fragmentation in a page.
+ */
+int kmem_cache_defrag(int node)
+{
+	struct kmem_cache *s;
+	unsigned long left = 0;
+
+	/*
+	 * kmem_cache_defrag may be called from the reclaim path which may be
+	 * called for any page allocator alloc. So there is the danger that we
+	 * get called in a situation where slub already acquired the slub_lock
+	 * for other purposes.
+	 */
+	if (!mutex_trylock(&slab_mutex))
+		return 0;
+
+	list_for_each_entry(s, &slab_caches, list) {
+		/*
+		 * Defragmentable caches come first. If the slab cache is not
+		 * defragmentable then we can stop traversing the list.
+		 */
+		if (!s->migrate)
+			break;
+
+		if (node == -1) {
+			int nid;
+
+			for_each_node_state(nid, N_NORMAL_MEMORY)
+				if (s->node[nid]->nr_partial > MAX_PARTIAL)
+					left += __move(s, nid, nid, s->defrag_ratio);
+		} else
+			left  += __move(s, node, node, 100);
+
+	}
+	mutex_unlock(&slab_mutex);
+	return left;
+}
+EXPORT_SYMBOL(kmem_cache_defrag);
+
+/*
+ * kmem_cache_shrink reduces the memory footprint of a slab cache
+ * by as much as possible. This works by removing empty slabs from
+ * the partial list, migrating slab objects to denser slab pages
+ * (if the slab cache supports that) or reorganizing the partial
+ * list so that denser slab pages come first and less dense
+ * allocated slab pages are at the end.
+ */
+int __kmem_cache_shrink(struct kmem_cache *s)
+{
+	int node;
+	int left = 0;
+
+	flush_all(s);
+	for_each_node_state(node, N_NORMAL_MEMORY)
+		left += __move(s, node, node, 100);
+
+	return 0;
+}
+EXPORT_SYMBOL(__kmem_cache_shrink);
+
 void kmem_cache_setup_mobility(struct kmem_cache *s,
 	kmem_isolate_func isolate, kmem_migrate_func migrate)
 {
 	int max_objects = oo_objects(s->max);

 	/*
-	 * Defragmentable slabs must have a ctor otherwise objects may be
-	 * in an undetermined state after they are allocated.
+	 * Mobile objects must have a ctor otherwise the
+	 * object may be in an undefined state on allocation.
+	 *
+	 * Since the object may need to be inspected by the
+	 * migration function at any time after allocation we
+	 * must ensure that the object always has a defined
+	 * state.
 	 */
 	BUG_ON(!s->ctor);