From 837d678dc264c797c16f81cf56f615f7544891c1 Mon Sep 17 00:00:00 2001
From: Joonsoo Kim <js1304@gmail.com>
Date: Thu, 16 Aug 2012 00:02:40 +0900
Subject: [PATCH] slub: remove one code path and reduce lock contention in
 __slab_free()

When we try to free object, there is some of case that we need
to take a node lock. This is the necessary step for preventing a race.
After taking a lock, then we try to cmpxchg_double_slab().
But, there is a possible scenario that cmpxchg_double_slab() is failed
with taking a lock. Following example explains it.

CPU A               CPU B
need lock
...                 need lock
...                 lock!!
lock..but spin      free success
spin...             unlock
lock!!
free fail

In this case, retry with taking a lock is occured in CPU A.
I think that in this case for CPU A,
"release a lock first, and re-take a lock if necessary" is preferable way.

There are two reasons for this.

First, this makes __slab_free()'s logic somehow simple.
With this patch, 'was_frozen = 1' is "always" handled without taking a lock.
So we can remove one code path.

Second, it may reduce lock contention.
When we do retrying, status of slab is already changed,
so we don't need a lock anymore in almost every case.
"release a lock first, and re-take a lock if necessary" policy is
helpful to this.

Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Acked-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
---
 mm/slub.c | 36 +++++++++++++++---------------------
 1 file changed, 15 insertions(+), 21 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index a0d698467f70..e7aec2001ae5 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2459,7 +2459,6 @@ static void __slab_free(struct kmem_cache *s, struct page *page,
 	void *prior;
 	void **object = (void *)x;
 	int was_frozen;
-	int inuse;
 	struct page new;
 	unsigned long counters;
 	struct kmem_cache_node *n = NULL;
@@ -2472,13 +2471,17 @@ static void __slab_free(struct kmem_cache *s, struct page *page,
 		return;
 
 	do {
+		if (unlikely(n)) {
+			spin_unlock_irqrestore(&n->list_lock, flags);
+			n = NULL;
+		}
 		prior = page->freelist;
 		counters = page->counters;
 		set_freepointer(s, object, prior);
 		new.counters = counters;
 		was_frozen = new.frozen;
 		new.inuse--;
-		if ((!new.inuse || !prior) && !was_frozen && !n) {
+		if ((!new.inuse || !prior) && !was_frozen) {
 
 			if (!kmem_cache_debug(s) && !prior)
 
@@ -2503,7 +2506,6 @@ static void __slab_free(struct kmem_cache *s, struct page *page,
 
 			}
 		}
-		inuse = new.inuse;
 
 	} while (!cmpxchg_double_slab(s, page,
 		prior, counters,
@@ -2529,25 +2531,17 @@ static void __slab_free(struct kmem_cache *s, struct page *page,
                 return;
         }
 
-	/*
-	 * was_frozen may have been set after we acquired the list_lock in
-	 * an earlier loop. So we need to check it here again.
-	 */
-	if (was_frozen)
-		stat(s, FREE_FROZEN);
-	else {
-		if (unlikely(!inuse && n->nr_partial > s->min_partial))
-                        goto slab_empty;
+	if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
+		goto slab_empty;
 
-		/*
-		 * Objects left in the slab. If it was not on the partial list before
-		 * then add it.
-		 */
-		if (unlikely(!prior)) {
-			remove_full(s, page);
-			add_partial(n, page, DEACTIVATE_TO_TAIL);
-			stat(s, FREE_ADD_PARTIAL);
-		}
+	/*
+	 * Objects left in the slab. If it was not on the partial list before
+	 * then add it.
+	 */
+	if (kmem_cache_debug(s) && unlikely(!prior)) {
+		remove_full(s, page);
+		add_partial(n, page, DEACTIVATE_TO_TAIL);
+		stat(s, FREE_ADD_PARTIAL);
 	}
 	spin_unlock_irqrestore(&n->list_lock, flags);
 	return;