[PATCH] Numa-aware slab allocator V5
authorChristoph Lameter <clameter@engr.sgi.com>
Fri, 9 Sep 2005 20:03:32 +0000 (13:03 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 9 Sep 2005 20:57:48 +0000 (13:57 -0700)
The NUMA API change that introduced kmalloc_node was accepted for
2.6.12-rc3.  Now it is possible to do slab allocations on a node to
localize memory structures.  This API was used by the pageset localization
patch and the block layer localization patch now in mm.  The existing
kmalloc_node is slow since it simply searches through all pages of the slab
to find a page that is on the node requested.  The two patches do a one
time allocation of slab structures at initialization and therefore the
speed of kmalloc node does not matter.

This patch allows kmalloc_node to be as fast as kmalloc by introducing node
specific page lists for partial, free and full slabs.  Slab allocation
improves in a NUMA system so that we are seeing a performance gain in AIM7
of about 5% with this patch alone.

More NUMA localizations are possible if kmalloc_node operates in an fast
way like kmalloc.

Test run on a 32p systems with 32G Ram.

w/o patch
Tasks    jobs/min  jti  jobs/min/task      real       cpu
    1      485.36  100       485.3640     11.99      1.91   Sat Apr 30 14:01:51 2005
  100    26582.63   88       265.8263     21.89    144.96   Sat Apr 30 14:02:14 2005
  200    29866.83   81       149.3342     38.97    286.08   Sat Apr 30 14:02:53 2005
  300    33127.16   78       110.4239     52.71    426.54   Sat Apr 30 14:03:46 2005
  400    34889.47   80        87.2237     66.72    568.90   Sat Apr 30 14:04:53 2005
  500    35654.34   76        71.3087     81.62    714.55   Sat Apr 30 14:06:15 2005
  600    36460.83   75        60.7681     95.77    853.42   Sat Apr 30 14:07:51 2005
  700    35957.00   75        51.3671    113.30    990.67   Sat Apr 30 14:09:45 2005
  800    33380.65   73        41.7258    139.48   1140.86   Sat Apr 30 14:12:05 2005
  900    35095.01   76        38.9945    149.25   1281.30   Sat Apr 30 14:14:35 2005
 1000    36094.37   74        36.0944    161.24   1419.66   Sat Apr 30 14:17:17 2005

w/patch
Tasks    jobs/min  jti  jobs/min/task      real       cpu
    1      484.27  100       484.2736     12.02      1.93   Sat Apr 30 15:59:45 2005
  100    28262.03   90       282.6203     20.59    143.57   Sat Apr 30 16:00:06 2005
  200    32246.45   82       161.2322     36.10    282.89   Sat Apr 30 16:00:42 2005
  300    37945.80   83       126.4860     46.01    418.75   Sat Apr 30 16:01:28 2005
  400    40000.69   81       100.0017     58.20    561.48   Sat Apr 30 16:02:27 2005
  500    40976.10   78        81.9522     71.02    696.95   Sat Apr 30 16:03:38 2005
  600    41121.54   78        68.5359     84.92    834.86   Sat Apr 30 16:05:04 2005
  700    44052.77   78        62.9325     92.48    971.53   Sat Apr 30 16:06:37 2005
  800    41066.89   79        51.3336    113.38   1111.15   Sat Apr 30 16:08:31 2005
  900    38918.77   79        43.2431    134.59   1252.57   Sat Apr 30 16:10:46 2005
 1000    41842.21   76        41.8422    139.09   1392.33   Sat Apr 30 16:13:05 2005

These are measurement taken directly after boot and show a greater
improvement than 5%.  However, the performance improvements become less
over time if the AIM7 runs are repeated and settle down at around 5%.

Links to earlier discussions:
http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2
http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2

Changelog V4-V5:
- alloc_arraycache and alloc_aliencache take node parameter instead of cpu
- fix initialization so that nodes without cpus are properly handled.
- simplify code in kmem_cache_init
- patch against Andrews temp mm3 release
- Add Shai to credits
- fallback to __cache_alloc from __cache_alloc_node if the node's cache
  is not available yet.

Changelog V3-V4:
- Patch against 2.6.12-rc5-mm1
- Cleanup patch integrated
- More and better use of for_each_node and for_each_cpu
- GCC 2.95 fix (do not use [] use [0])
- Correct determination of INDEX_AC
- Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes.
- Remove list3_data and list3_data_ptr macros for better readability

Changelog V2-V3:
- Made to patch against 2.6.12-rc4-mm1
- Revised bootstrap mechanism so that larger size kmem_list3 structs can be
  supported. Do a generic solution so that the right slab can be found
  for the internal structs.
- use for_each_online_node

Changelog V1-V2:
- Batching for freeing of wrong-node objects (alien caches)
- Locking changes and NUMA #ifdefs as requested by Manfred

Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com>
Signed-off-by: Shai Fultheim <Shai@Scalex86.org>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Cc: Manfred Spraul <manfred@colorfullife.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/slab.c

index d7c4443991fe8e29e16dfaffa30a1711c6844c30..a041c5378dfa59895c411d07b88cf69153275a11 100644 (file)
--- a/mm/slab.c
+++ b/mm/slab.c
  *
  *     At present, each engine can be growing a cache.  This should be blocked.
  *
+ * 15 March 2005. NUMA slab allocator.
+ *     Shai Fultheim <shai@scalex86.org>.
+ *     Shobhit Dayal <shobhit@calsoftinc.com>
+ *     Alok N Kataria <alokk@calsoftinc.com>
+ *     Christoph Lameter <christoph@lameter.com>
+ *
+ *     Modified the slab allocator to be node aware on NUMA systems.
+ *     Each node has its own list of partial, free and full slabs.
+ *     All object allocations for a node occur from node specific slab lists.
  */
 
 #include       <linux/config.h>
 #include       <linux/module.h>
 #include       <linux/rcupdate.h>
 #include       <linux/string.h>
+#include       <linux/nodemask.h>
 
 #include       <asm/uaccess.h>
 #include       <asm/cacheflush.h>
@@ -212,6 +222,7 @@ struct slab {
        void                    *s_mem;         /* including colour offset */
        unsigned int            inuse;          /* num of objs active in slab */
        kmem_bufctl_t           free;
+       unsigned short          nodeid;
 };
 
 /*
@@ -239,7 +250,6 @@ struct slab_rcu {
 /*
  * struct array_cache
  *
- * Per cpu structures
  * Purpose:
  * - LIFO ordering, to hand out cache-warm objects from _alloc
  * - reduce the number of linked list operations
@@ -254,6 +264,13 @@ struct array_cache {
        unsigned int limit;
        unsigned int batchcount;
        unsigned int touched;
+       spinlock_t lock;
+       void *entry[0];         /*
+                                * Must have this definition in here for the proper
+                                * alignment of array_cache. Also simplifies accessing
+                                * the entries.
+                                * [0] is for gcc 2.95. It should really be [].
+                                */
 };
 
 /* bootstrap: The caches do not work without cpuarrays anymore,
@@ -266,34 +283,83 @@ struct arraycache_init {
 };
 
 /*
- * The slab lists of all objects.
- * Hopefully reduce the internal fragmentation
- * NUMA: The spinlock could be moved from the kmem_cache_t
- * into this structure, too. Figure out what causes
- * fewer cross-node spinlock operations.
+ * The slab lists for all objects.
  */
 struct kmem_list3 {
        struct list_head        slabs_partial;  /* partial list first, better asm code */
        struct list_head        slabs_full;
        struct list_head        slabs_free;
        unsigned long   free_objects;
-       int             free_touched;
        unsigned long   next_reap;
-       struct array_cache      *shared;
+       int             free_touched;
+       unsigned int    free_limit;
+       spinlock_t      list_lock;
+       struct array_cache      *shared;        /* shared per node */
+       struct array_cache      **alien;        /* on other nodes */
 };
 
-#define LIST3_INIT(parent) \
-       { \
-               .slabs_full     = LIST_HEAD_INIT(parent.slabs_full), \
-               .slabs_partial  = LIST_HEAD_INIT(parent.slabs_partial), \
-               .slabs_free     = LIST_HEAD_INIT(parent.slabs_free) \
+/*
+ * Need this for bootstrapping a per node allocator.
+ */
+#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
+struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
+#define        CACHE_CACHE 0
+#define        SIZE_AC 1
+#define        SIZE_L3 (1 + MAX_NUMNODES)
+
+/*
+ * This function may be completely optimized away if
+ * a constant is passed to it. Mostly the same as
+ * what is in linux/slab.h except it returns an
+ * index.
+ */
+static inline int index_of(const size_t size)
+{
+       if (__builtin_constant_p(size)) {
+               int i = 0;
+
+#define CACHE(x) \
+       if (size <=x) \
+               return i; \
+       else \
+               i++;
+#include "linux/kmalloc_sizes.h"
+#undef CACHE
+               {
+                       extern void __bad_size(void);
+                       __bad_size();
+               }
        }
-#define list3_data(cachep) \
-       (&(cachep)->lists)
+       return 0;
+}
+
+#define INDEX_AC index_of(sizeof(struct arraycache_init))
+#define INDEX_L3 index_of(sizeof(struct kmem_list3))
+
+static inline void kmem_list3_init(struct kmem_list3 *parent)
+{
+       INIT_LIST_HEAD(&parent->slabs_full);
+       INIT_LIST_HEAD(&parent->slabs_partial);
+       INIT_LIST_HEAD(&parent->slabs_free);
+       parent->shared = NULL;
+       parent->alien = NULL;
+       spin_lock_init(&parent->list_lock);
+       parent->free_objects = 0;
+       parent->free_touched = 0;
+}
 
-/* NUMA: per-node */
-#define list3_data_ptr(cachep, ptr) \
-               list3_data(cachep)
+#define MAKE_LIST(cachep, listp, slab, nodeid) \
+       do {    \
+               INIT_LIST_HEAD(listp);          \
+               list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
+       } while (0)
+
+#define        MAKE_ALL_LISTS(cachep, ptr, nodeid)                     \
+       do {                                    \
+       MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);  \
+       MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
+       MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);  \
+       } while (0)
 
 /*
  * kmem_cache_t
@@ -306,13 +372,12 @@ struct kmem_cache_s {
        struct array_cache      *array[NR_CPUS];
        unsigned int            batchcount;
        unsigned int            limit;
-/* 2) touched by every alloc & free from the backend */
-       struct kmem_list3       lists;
-       /* NUMA: kmem_3list_t   *nodelists[MAX_NUMNODES] */
+       unsigned int            shared;
        unsigned int            objsize;
+/* 2) touched by every alloc & free from the backend */
+       struct kmem_list3       *nodelists[MAX_NUMNODES];
        unsigned int            flags;  /* constant flags */
        unsigned int            num;    /* # of objs per slab */
-       unsigned int            free_limit; /* upper limit of objects in the lists */
        spinlock_t              spinlock;
 
 /* 3) cache_grow/shrink */
@@ -349,6 +414,7 @@ struct kmem_cache_s {
        unsigned long           errors;
        unsigned long           max_freeable;
        unsigned long           node_allocs;
+       unsigned long           node_frees;
        atomic_t                allochit;
        atomic_t                allocmiss;
        atomic_t                freehit;
@@ -384,6 +450,7 @@ struct kmem_cache_s {
                                } while (0)
 #define        STATS_INC_ERR(x)        ((x)->errors++)
 #define        STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
+#define        STATS_INC_NODEFREES(x)  ((x)->node_frees++)
 #define        STATS_SET_FREEABLE(x, i) \
                                do { if ((x)->max_freeable < i) \
                                        (x)->max_freeable = i; \
@@ -402,6 +469,7 @@ struct kmem_cache_s {
 #define        STATS_SET_HIGH(x)       do { } while (0)
 #define        STATS_INC_ERR(x)        do { } while (0)
 #define        STATS_INC_NODEALLOCS(x) do { } while (0)
+#define        STATS_INC_NODEFREES(x)  do { } while (0)
 #define        STATS_SET_FREEABLE(x, i) \
                                do { } while (0)
 
@@ -534,9 +602,9 @@ static struct arraycache_init initarray_generic =
 
 /* internal cache of cache description objs */
 static kmem_cache_t cache_cache = {
-       .lists          = LIST3_INIT(cache_cache.lists),
        .batchcount     = 1,
        .limit          = BOOT_CPUCACHE_ENTRIES,
+       .shared         = 1,
        .objsize        = sizeof(kmem_cache_t),
        .flags          = SLAB_NO_REAP,
        .spinlock       = SPIN_LOCK_UNLOCKED,
@@ -557,7 +625,6 @@ static struct list_head cache_chain;
  * SLAB_RECLAIM_ACCOUNT turns this on per-slab
  */
 atomic_t slab_reclaim_pages;
-EXPORT_SYMBOL(slab_reclaim_pages);
 
 /*
  * chicken and egg problem: delay the per-cpu array allocation
@@ -565,7 +632,8 @@ EXPORT_SYMBOL(slab_reclaim_pages);
  */
 static enum {
        NONE,
-       PARTIAL,
+       PARTIAL_AC,
+       PARTIAL_L3,
        FULL
 } g_cpucache_up;
 
@@ -574,11 +642,7 @@ static DEFINE_PER_CPU(struct work_struct, reap_work);
 static void free_block(kmem_cache_t* cachep, void** objpp, int len);
 static void enable_cpucache (kmem_cache_t *cachep);
 static void cache_reap (void *unused);
-
-static inline void **ac_entry(struct array_cache *ac)
-{
-       return (void**)(ac+1);
-}
+static int __node_shrink(kmem_cache_t *cachep, int node);
 
 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
 {
@@ -676,48 +740,160 @@ static void __devinit start_cpu_timer(int cpu)
        }
 }
 
-static struct array_cache *alloc_arraycache(int cpu, int entries,
+static struct array_cache *alloc_arraycache(int node, int entries,
                                                int batchcount)
 {
        int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
        struct array_cache *nc = NULL;
 
-       if (cpu == -1)
-               nc = kmalloc(memsize, GFP_KERNEL);
-       else
-               nc = kmalloc_node(memsize, GFP_KERNEL, cpu_to_node(cpu));
-
+       nc = kmalloc_node(memsize, GFP_KERNEL, node);
        if (nc) {
                nc->avail = 0;
                nc->limit = entries;
                nc->batchcount = batchcount;
                nc->touched = 0;
+               spin_lock_init(&nc->lock);
        }
        return nc;
 }
 
+#ifdef CONFIG_NUMA
+static inline struct array_cache **alloc_alien_cache(int node, int limit)
+{
+       struct array_cache **ac_ptr;
+       int memsize = sizeof(void*)*MAX_NUMNODES;
+       int i;
+
+       if (limit > 1)
+               limit = 12;
+       ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
+       if (ac_ptr) {
+               for_each_node(i) {
+                       if (i == node || !node_online(i)) {
+                               ac_ptr[i] = NULL;
+                               continue;
+                       }
+                       ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
+                       if (!ac_ptr[i]) {
+                               for (i--; i <=0; i--)
+                                       kfree(ac_ptr[i]);
+                               kfree(ac_ptr);
+                               return NULL;
+                       }
+               }
+       }
+       return ac_ptr;
+}
+
+static inline void free_alien_cache(struct array_cache **ac_ptr)
+{
+       int i;
+
+       if (!ac_ptr)
+               return;
+
+       for_each_node(i)
+               kfree(ac_ptr[i]);
+
+       kfree(ac_ptr);
+}
+
+static inline void __drain_alien_cache(kmem_cache_t *cachep, struct array_cache *ac, int node)
+{
+       struct kmem_list3 *rl3 = cachep->nodelists[node];
+
+       if (ac->avail) {
+               spin_lock(&rl3->list_lock);
+               free_block(cachep, ac->entry, ac->avail);
+               ac->avail = 0;
+               spin_unlock(&rl3->list_lock);
+       }
+}
+
+static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
+{
+       int i=0;
+       struct array_cache *ac;
+       unsigned long flags;
+
+       for_each_online_node(i) {
+               ac = l3->alien[i];
+               if (ac) {
+                       spin_lock_irqsave(&ac->lock, flags);
+                       __drain_alien_cache(cachep, ac, i);
+                       spin_unlock_irqrestore(&ac->lock, flags);
+               }
+       }
+}
+#else
+#define alloc_alien_cache(node, limit) do { } while (0)
+#define free_alien_cache(ac_ptr) do { } while (0)
+#define drain_alien_cache(cachep, l3) do { } while (0)
+#endif
+
 static int __devinit cpuup_callback(struct notifier_block *nfb,
                                  unsigned long action, void *hcpu)
 {
        long cpu = (long)hcpu;
        kmem_cache_t* cachep;
+       struct kmem_list3 *l3 = NULL;
+       int node = cpu_to_node(cpu);
+       int memsize = sizeof(struct kmem_list3);
+       struct array_cache *nc = NULL;
 
        switch (action) {
        case CPU_UP_PREPARE:
                down(&cache_chain_sem);
+               /* we need to do this right in the beginning since
+                * alloc_arraycache's are going to use this list.
+                * kmalloc_node allows us to add the slab to the right
+                * kmem_list3 and not this cpu's kmem_list3
+                */
+
                list_for_each_entry(cachep, &cache_chain, next) {
-                       struct array_cache *nc;
+                       /* setup the size64 kmemlist for cpu before we can
+                        * begin anything. Make sure some other cpu on this
+                        * node has not already allocated this
+                        */
+                       if (!cachep->nodelists[node]) {
+                               if (!(l3 = kmalloc_node(memsize,
+                                               GFP_KERNEL, node)))
+                                       goto bad;
+                               kmem_list3_init(l3);
+                               l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
+                                 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
+
+                               cachep->nodelists[node] = l3;
+                       }
+
+                       spin_lock_irq(&cachep->nodelists[node]->list_lock);
+                       cachep->nodelists[node]->free_limit =
+                               (1 + nr_cpus_node(node)) *
+                               cachep->batchcount + cachep->num;
+                       spin_unlock_irq(&cachep->nodelists[node]->list_lock);
+               }
 
-                       nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount);
+               /* Now we can go ahead with allocating the shared array's
+                 & array cache's */
+               list_for_each_entry(cachep, &cache_chain, next) {
+                       nc = alloc_arraycache(node, cachep->limit,
+                                       cachep->batchcount);
                        if (!nc)
                                goto bad;
-
-                       spin_lock_irq(&cachep->spinlock);
                        cachep->array[cpu] = nc;
-                       cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
-                                               + cachep->num;
-                       spin_unlock_irq(&cachep->spinlock);
 
+                       l3 = cachep->nodelists[node];
+                       BUG_ON(!l3);
+                       if (!l3->shared) {
+                               if (!(nc = alloc_arraycache(node,
+                                       cachep->shared*cachep->batchcount,
+                                       0xbaadf00d)))
+                                       goto  bad;
+
+                               /* we are serialised from CPU_DEAD or
+                                 CPU_UP_CANCELLED by the cpucontrol lock */
+                               l3->shared = nc;
+                       }
                }
                up(&cache_chain_sem);
                break;
@@ -732,13 +908,51 @@ static int __devinit cpuup_callback(struct notifier_block *nfb,
 
                list_for_each_entry(cachep, &cache_chain, next) {
                        struct array_cache *nc;
+                       cpumask_t mask;
 
+                       mask = node_to_cpumask(node);
                        spin_lock_irq(&cachep->spinlock);
                        /* cpu is dead; no one can alloc from it. */
                        nc = cachep->array[cpu];
                        cachep->array[cpu] = NULL;
-                       cachep->free_limit -= cachep->batchcount;
-                       free_block(cachep, ac_entry(nc), nc->avail);
+                       l3 = cachep->nodelists[node];
+
+                       if (!l3)
+                               goto unlock_cache;
+
+                       spin_lock(&l3->list_lock);
+
+                       /* Free limit for this kmem_list3 */
+                       l3->free_limit -= cachep->batchcount;
+                       if (nc)
+                               free_block(cachep, nc->entry, nc->avail);
+
+                       if (!cpus_empty(mask)) {
+                                spin_unlock(&l3->list_lock);
+                                goto unlock_cache;
+                        }
+
+                       if (l3->shared) {
+                               free_block(cachep, l3->shared->entry,
+                                               l3->shared->avail);
+                               kfree(l3->shared);
+                               l3->shared = NULL;
+                       }
+                       if (l3->alien) {
+                               drain_alien_cache(cachep, l3);
+                               free_alien_cache(l3->alien);
+                               l3->alien = NULL;
+                       }
+
+                       /* free slabs belonging to this node */
+                       if (__node_shrink(cachep, node)) {
+                               cachep->nodelists[node] = NULL;
+                               spin_unlock(&l3->list_lock);
+                               kfree(l3);
+                       } else {
+                               spin_unlock(&l3->list_lock);
+                       }
+unlock_cache:
                        spin_unlock_irq(&cachep->spinlock);
                        kfree(nc);
                }
@@ -754,6 +968,25 @@ bad:
 
 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
 
+/*
+ * swap the static kmem_list3 with kmalloced memory
+ */
+static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list,
+               int nodeid)
+{
+       struct kmem_list3 *ptr;
+
+       BUG_ON(cachep->nodelists[nodeid] != list);
+       ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
+       BUG_ON(!ptr);
+
+       local_irq_disable();
+       memcpy(ptr, list, sizeof(struct kmem_list3));
+       MAKE_ALL_LISTS(cachep, ptr, nodeid);
+       cachep->nodelists[nodeid] = ptr;
+       local_irq_enable();
+}
+
 /* Initialisation.
  * Called after the gfp() functions have been enabled, and before smp_init().
  */
@@ -762,6 +995,13 @@ void __init kmem_cache_init(void)
        size_t left_over;
        struct cache_sizes *sizes;
        struct cache_names *names;
+       int i;
+
+       for (i = 0; i < NUM_INIT_LISTS; i++) {
+               kmem_list3_init(&initkmem_list3[i]);
+               if (i < MAX_NUMNODES)
+                       cache_cache.nodelists[i] = NULL;
+       }
 
        /*
         * Fragmentation resistance on low memory - only use bigger
@@ -770,21 +1010,24 @@ void __init kmem_cache_init(void)
        if (num_physpages > (32 << 20) >> PAGE_SHIFT)
                slab_break_gfp_order = BREAK_GFP_ORDER_HI;
 
-       
        /* Bootstrap is tricky, because several objects are allocated
         * from caches that do not exist yet:
         * 1) initialize the cache_cache cache: it contains the kmem_cache_t
         *    structures of all caches, except cache_cache itself: cache_cache
         *    is statically allocated.
-        *    Initially an __init data area is used for the head array, it's
-        *    replaced with a kmalloc allocated array at the end of the bootstrap.
+        *    Initially an __init data area is used for the head array and the
+        *    kmem_list3 structures, it's replaced with a kmalloc allocated
+        *    array at the end of the bootstrap.
         * 2) Create the first kmalloc cache.
-        *    The kmem_cache_t for the new cache is allocated normally. An __init
-        *    data area is used for the head array.
-        * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
+        *    The kmem_cache_t for the new cache is allocated normally.
+        *    An __init data area is used for the head array.
+        * 3) Create the remaining kmalloc caches, with minimally sized
+        *    head arrays.
         * 4) Replace the __init data head arrays for cache_cache and the first
         *    kmalloc cache with kmalloc allocated arrays.
-        * 5) Resize the head arrays of the kmalloc caches to their final sizes.
+        * 5) Replace the __init data for kmem_list3 for cache_cache and
+        *    the other cache's with kmalloc allocated memory.
+        * 6) Resize the head arrays of the kmalloc caches to their final sizes.
         */
 
        /* 1) create the cache_cache */
@@ -793,6 +1036,7 @@ void __init kmem_cache_init(void)
        list_add(&cache_cache.next, &cache_chain);
        cache_cache.colour_off = cache_line_size();
        cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
+       cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
 
        cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
 
@@ -810,15 +1054,33 @@ void __init kmem_cache_init(void)
        sizes = malloc_sizes;
        names = cache_names;
 
+       /* Initialize the caches that provide memory for the array cache
+        * and the kmem_list3 structures first.
+        * Without this, further allocations will bug
+        */
+
+       sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
+                               sizes[INDEX_AC].cs_size, ARCH_KMALLOC_MINALIGN,
+                               (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
+
+       if (INDEX_AC != INDEX_L3)
+               sizes[INDEX_L3].cs_cachep =
+                       kmem_cache_create(names[INDEX_L3].name,
+                               sizes[INDEX_L3].cs_size, ARCH_KMALLOC_MINALIGN,
+                               (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
+
        while (sizes->cs_size != ULONG_MAX) {
-               /* For performance, all the general caches are L1 aligned.
+               /*
+                * For performance, all the general caches are L1 aligned.
                 * This should be particularly beneficial on SMP boxes, as it
                 * eliminates "false sharing".
                 * Note for systems short on memory removing the alignment will
-                * allow tighter packing of the smaller caches. */
-               sizes->cs_cachep = kmem_cache_create(names->name,
-                       sizes->cs_size, ARCH_KMALLOC_MINALIGN,
-                       (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
+                * allow tighter packing of the smaller caches.
+                */
+               if(!sizes->cs_cachep)
+                       sizes->cs_cachep = kmem_cache_create(names->name,
+                               sizes->cs_size, ARCH_KMALLOC_MINALIGN,
+                               (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
 
                /* Inc off-slab bufctl limit until the ceiling is hit. */
                if (!(OFF_SLAB(sizes->cs_cachep))) {
@@ -837,24 +1099,47 @@ void __init kmem_cache_init(void)
        /* 4) Replace the bootstrap head arrays */
        {
                void * ptr;
-               
+
                ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
+
                local_irq_disable();
                BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
-               memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
+               memcpy(ptr, ac_data(&cache_cache),
+                               sizeof(struct arraycache_init));
                cache_cache.array[smp_processor_id()] = ptr;
                local_irq_enable();
-       
+
                ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
+
                local_irq_disable();
-               BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
-               memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
+               BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
+                               != &initarray_generic.cache);
+               memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
                                sizeof(struct arraycache_init));
-               malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
+               malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
+                                               ptr;
                local_irq_enable();
        }
+       /* 5) Replace the bootstrap kmem_list3's */
+       {
+               int node;
+               /* Replace the static kmem_list3 structures for the boot cpu */
+               init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
+                               numa_node_id());
+
+               for_each_online_node(node) {
+                       init_list(malloc_sizes[INDEX_AC].cs_cachep,
+                                       &initkmem_list3[SIZE_AC+node], node);
+
+                       if (INDEX_AC != INDEX_L3) {
+                               init_list(malloc_sizes[INDEX_L3].cs_cachep,
+                                               &initkmem_list3[SIZE_L3+node],
+                                               node);
+                       }
+               }
+       }
 
-       /* 5) resize the head arrays to their final sizes */
+       /* 6) resize the head arrays to their final sizes */
        {
                kmem_cache_t *cachep;
                down(&cache_chain_sem);
@@ -870,7 +1155,6 @@ void __init kmem_cache_init(void)
         * that initializes ac_data for all new cpus
         */
        register_cpu_notifier(&cpucache_notifier);
-       
 
        /* The reap timers are started later, with a module init call:
         * That part of the kernel is not yet operational.
@@ -885,10 +1169,8 @@ static int __init cpucache_init(void)
         * Register the timers that return unneeded
         * pages to gfp.
         */
-       for (cpu = 0; cpu < NR_CPUS; cpu++) {
-               if (cpu_online(cpu))
-                       start_cpu_timer(cpu);
-       }
+       for_each_online_cpu(cpu)
+               start_cpu_timer(cpu);
 
        return 0;
 }
@@ -1167,6 +1449,20 @@ static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
        }
 }
 
+/* For setting up all the kmem_list3s for cache whose objsize is same
+   as size of kmem_list3. */
+static inline void set_up_list3s(kmem_cache_t *cachep, int index)
+{
+       int node;
+
+       for_each_online_node(node) {
+               cachep->nodelists[node] = &initkmem_list3[index+node];
+               cachep->nodelists[node]->next_reap = jiffies +
+                       REAPTIMEOUT_LIST3 +
+                       ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
+       }
+}
+
 /**
  * kmem_cache_create - Create a cache.
  * @name: A string which is used in /proc/slabinfo to identify this cache.
@@ -1320,7 +1616,7 @@ kmem_cache_create (const char *name, size_t size, size_t align,
                size += BYTES_PER_WORD;
        }
 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
-       if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
+       if (size >= malloc_sizes[INDEX_L3+1].cs_size && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
                cachep->dbghead += PAGE_SIZE - size;
                size = PAGE_SIZE;
        }
@@ -1422,10 +1718,6 @@ next:
                cachep->gfpflags |= GFP_DMA;
        spin_lock_init(&cachep->spinlock);
        cachep->objsize = size;
-       /* NUMA */
-       INIT_LIST_HEAD(&cachep->lists.slabs_full);
-       INIT_LIST_HEAD(&cachep->lists.slabs_partial);
-       INIT_LIST_HEAD(&cachep->lists.slabs_free);
 
        if (flags & CFLGS_OFF_SLAB)
                cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
@@ -1444,11 +1736,43 @@ next:
                         * the cache that's used by kmalloc(24), otherwise
                         * the creation of further caches will BUG().
                         */
-                       cachep->array[smp_processor_id()] = &initarray_generic.cache;
-                       g_cpucache_up = PARTIAL;
+                       cachep->array[smp_processor_id()] =
+                               &initarray_generic.cache;
+
+                       /* If the cache that's used by
+                        * kmalloc(sizeof(kmem_list3)) is the first cache,
+                        * then we need to set up all its list3s, otherwise
+                        * the creation of further caches will BUG().
+                        */
+                       set_up_list3s(cachep, SIZE_AC);
+                       if (INDEX_AC == INDEX_L3)
+                               g_cpucache_up = PARTIAL_L3;
+                       else
+                               g_cpucache_up = PARTIAL_AC;
                } else {
-                       cachep->array[smp_processor_id()] = kmalloc(sizeof(struct arraycache_init),GFP_KERNEL);
+                       cachep->array[smp_processor_id()] =
+                               kmalloc(sizeof(struct arraycache_init),
+                                               GFP_KERNEL);
+
+                       if (g_cpucache_up == PARTIAL_AC) {
+                               set_up_list3s(cachep, SIZE_L3);
+                               g_cpucache_up = PARTIAL_L3;
+                       } else {
+                               int node;
+                               for_each_online_node(node) {
+
+                                       cachep->nodelists[node] =
+                                               kmalloc_node(sizeof(struct kmem_list3),
+                                                               GFP_KERNEL, node);
+                                       BUG_ON(!cachep->nodelists[node]);
+                                       kmem_list3_init(cachep->nodelists[node]);
+                               }
+                       }
                }
+               cachep->nodelists[numa_node_id()]->next_reap =
+                       jiffies + REAPTIMEOUT_LIST3 +
+                       ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
+
                BUG_ON(!ac_data(cachep));
                ac_data(cachep)->avail = 0;
                ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
@@ -1456,13 +1780,8 @@ next:
                ac_data(cachep)->touched = 0;
                cachep->batchcount = 1;
                cachep->limit = BOOT_CPUCACHE_ENTRIES;
-               cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
-                                       + cachep->num;
        } 
 
-       cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
-                                       ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
-
        /* Need the semaphore to access the chain. */
        down(&cache_chain_sem);
        {
@@ -1519,13 +1838,23 @@ static void check_spinlock_acquired(kmem_cache_t *cachep)
 {
 #ifdef CONFIG_SMP
        check_irq_off();
-       BUG_ON(spin_trylock(&cachep->spinlock));
+       assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
 #endif
 }
+
+static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
+{
+#ifdef CONFIG_SMP
+       check_irq_off();
+       assert_spin_locked(&cachep->nodelists[node]->list_lock);
+#endif
+}
+
 #else
 #define check_irq_off()        do { } while(0)
 #define check_irq_on() do { } while(0)
 #define check_spinlock_acquired(x) do { } while(0)
+#define check_spinlock_acquired_node(x, y) do { } while(0)
 #endif
 
 /*
@@ -1547,7 +1876,7 @@ static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
 }
 
 static void drain_array_locked(kmem_cache_t* cachep,
-                               struct array_cache *ac, int force);
+                               struct array_cache *ac, int force, int node);
 
 static void do_drain(void *arg)
 {
@@ -1556,59 +1885,82 @@ static void do_drain(void *arg)
 
        check_irq_off();
        ac = ac_data(cachep);
-       spin_lock(&cachep->spinlock);
-       free_block(cachep, &ac_entry(ac)[0], ac->avail);
-       spin_unlock(&cachep->spinlock);
+       spin_lock(&cachep->nodelists[numa_node_id()]->list_lock);
+       free_block(cachep, ac->entry, ac->avail);
+       spin_unlock(&cachep->nodelists[numa_node_id()]->list_lock);
        ac->avail = 0;
 }
 
 static void drain_cpu_caches(kmem_cache_t *cachep)
 {
+       struct kmem_list3 *l3;
+       int node;
+
        smp_call_function_all_cpus(do_drain, cachep);
        check_irq_on();
        spin_lock_irq(&cachep->spinlock);
-       if (cachep->lists.shared)
-               drain_array_locked(cachep, cachep->lists.shared, 1);
+       for_each_online_node(node)  {
+               l3 = cachep->nodelists[node];
+               if (l3) {
+                       spin_lock(&l3->list_lock);
+                       drain_array_locked(cachep, l3->shared, 1, node);
+                       spin_unlock(&l3->list_lock);
+                       if (l3->alien)
+                               drain_alien_cache(cachep, l3);
+               }
+       }
        spin_unlock_irq(&cachep->spinlock);
 }
 
-
-/* NUMA shrink all list3s */
-static int __cache_shrink(kmem_cache_t *cachep)
+static int __node_shrink(kmem_cache_t *cachep, int node)
 {
        struct slab *slabp;
+       struct kmem_list3 *l3 = cachep->nodelists[node];
        int ret;
 
-       drain_cpu_caches(cachep);
-
-       check_irq_on();
-       spin_lock_irq(&cachep->spinlock);
-
-       for(;;) {
+       for (;;) {
                struct list_head *p;
 
-               p = cachep->lists.slabs_free.prev;
-               if (p == &cachep->lists.slabs_free)
+               p = l3->slabs_free.prev;
+               if (p == &l3->slabs_free)
                        break;
 
-               slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
+               slabp = list_entry(l3->slabs_free.prev, struct slab, list);
 #if DEBUG
                if (slabp->inuse)
                        BUG();
 #endif
                list_del(&slabp->list);
 
-               cachep->lists.free_objects -= cachep->num;
-               spin_unlock_irq(&cachep->spinlock);
+               l3->free_objects -= cachep->num;
+               spin_unlock_irq(&l3->list_lock);
                slab_destroy(cachep, slabp);
-               spin_lock_irq(&cachep->spinlock);
+               spin_lock_irq(&l3->list_lock);
        }
-       ret = !list_empty(&cachep->lists.slabs_full) ||
-               !list_empty(&cachep->lists.slabs_partial);
-       spin_unlock_irq(&cachep->spinlock);
+       ret = !list_empty(&l3->slabs_full) ||
+               !list_empty(&l3->slabs_partial);
        return ret;
 }
 
+static int __cache_shrink(kmem_cache_t *cachep)
+{
+       int ret = 0, i = 0;
+       struct kmem_list3 *l3;
+
+       drain_cpu_caches(cachep);
+
+       check_irq_on();
+       for_each_online_node(i) {
+               l3 = cachep->nodelists[i];
+               if (l3) {
+                       spin_lock_irq(&l3->list_lock);
+                       ret += __node_shrink(cachep, i);
+                       spin_unlock_irq(&l3->list_lock);
+               }
+       }
+       return (ret ? 1 : 0);
+}
+
 /**
  * kmem_cache_shrink - Shrink a cache.
  * @cachep: The cache to shrink.
@@ -1645,6 +1997,7 @@ EXPORT_SYMBOL(kmem_cache_shrink);
 int kmem_cache_destroy(kmem_cache_t * cachep)
 {
        int i;
+       struct kmem_list3 *l3;
 
        if (!cachep || in_interrupt())
                BUG();
@@ -1672,15 +2025,17 @@ int kmem_cache_destroy(kmem_cache_t * cachep)
        if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
                synchronize_rcu();
 
-       /* no cpu_online check required here since we clear the percpu
-        * array on cpu offline and set this to NULL.
-        */
-       for (i = 0; i < NR_CPUS; i++)
+       for_each_online_cpu(i)
                kfree(cachep->array[i]);
 
        /* NUMA: free the list3 structures */
-       kfree(cachep->lists.shared);
-       cachep->lists.shared = NULL;
+       for_each_online_node(i) {
+               if ((l3 = cachep->nodelists[i])) {
+                       kfree(l3->shared);
+                       free_alien_cache(l3->alien);
+                       kfree(l3);
+               }
+       }
        kmem_cache_free(&cache_cache, cachep);
 
        unlock_cpu_hotplug();
@@ -1690,8 +2045,8 @@ int kmem_cache_destroy(kmem_cache_t * cachep)
 EXPORT_SYMBOL(kmem_cache_destroy);
 
 /* Get the memory for a slab management obj. */
-static struct slab* alloc_slabmgmt(kmem_cache_t *cachep,
-                       void *objp, int colour_off, unsigned int __nocast local_flags)
+static struct slab* alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
+                       int colour_off, unsigned int __nocast local_flags)
 {
        struct slab *slabp;
        
@@ -1722,7 +2077,7 @@ static void cache_init_objs(kmem_cache_t *cachep,
        int i;
 
        for (i = 0; i < cachep->num; i++) {
-               voidobjp = slabp->s_mem+cachep->objsize*i;
+               void *objp = slabp->s_mem+cachep->objsize*i;
 #if DEBUG
                /* need to poison the objs? */
                if (cachep->flags & SLAB_POISON)
@@ -1799,6 +2154,7 @@ static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nod
        size_t           offset;
        unsigned int     local_flags;
        unsigned long    ctor_flags;
+       struct kmem_list3 *l3;
 
        /* Be lazy and only check for valid flags here,
         * keeping it out of the critical path in kmem_cache_alloc().
@@ -1830,6 +2186,7 @@ static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nod
 
        spin_unlock(&cachep->spinlock);
 
+       check_irq_off();
        if (local_flags & __GFP_WAIT)
                local_irq_enable();
 
@@ -1841,8 +2198,9 @@ static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nod
         */
        kmem_flagcheck(cachep, flags);
 
-
-       /* Get mem for the objs. */
+       /* Get mem for the objs.
+        * Attempt to allocate a physical page from 'nodeid',
+        */
        if (!(objp = kmem_getpages(cachep, flags, nodeid)))
                goto failed;
 
@@ -1850,6 +2208,7 @@ static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nod
        if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
                goto opps1;
 
+       slabp->nodeid = nodeid;
        set_slab_attr(cachep, slabp, objp);
 
        cache_init_objs(cachep, slabp, ctor_flags);
@@ -1857,13 +2216,14 @@ static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nod
        if (local_flags & __GFP_WAIT)
                local_irq_disable();
        check_irq_off();
-       spin_lock(&cachep->spinlock);
+       l3 = cachep->nodelists[nodeid];
+       spin_lock(&l3->list_lock);
 
        /* Make slab active. */
-       list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
+       list_add_tail(&slabp->list, &(l3->slabs_free));
        STATS_INC_GROWN(cachep);
-       list3_data(cachep)->free_objects += cachep->num;
-       spin_unlock(&cachep->spinlock);
+       l3->free_objects += cachep->num;
+       spin_unlock(&l3->list_lock);
        return 1;
 opps1:
        kmem_freepages(cachep, objp);
@@ -1969,7 +2329,6 @@ static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
        kmem_bufctl_t i;
        int entries = 0;
        
-       check_spinlock_acquired(cachep);
        /* Check slab's freelist to see if this obj is there. */
        for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
                entries++;
@@ -2012,10 +2371,11 @@ retry:
                 */
                batchcount = BATCHREFILL_LIMIT;
        }
-       l3 = list3_data(cachep);
+       l3 = cachep->nodelists[numa_node_id()];
+
+       BUG_ON(ac->avail > 0 || !l3);
+       spin_lock(&l3->list_lock);
 
-       BUG_ON(ac->avail > 0);
-       spin_lock(&cachep->spinlock);
        if (l3->shared) {
                struct array_cache *shared_array = l3->shared;
                if (shared_array->avail) {
@@ -2023,8 +2383,9 @@ retry:
                                batchcount = shared_array->avail;
                        shared_array->avail -= batchcount;
                        ac->avail = batchcount;
-                       memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
-                                       sizeof(void*)*batchcount);
+                       memcpy(ac->entry,
+                               &(shared_array->entry[shared_array->avail]),
+                               sizeof(void*)*batchcount);
                        shared_array->touched = 1;
                        goto alloc_done;
                }
@@ -2051,7 +2412,8 @@ retry:
                        STATS_SET_HIGH(cachep);
 
                        /* get obj pointer */
-                       ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
+                       ac->entry[ac->avail++] = slabp->s_mem +
+                               slabp->free*cachep->objsize;
 
                        slabp->inuse++;
                        next = slab_bufctl(slabp)[slabp->free];
@@ -2073,12 +2435,12 @@ retry:
 must_grow:
        l3->free_objects -= ac->avail;
 alloc_done:
-       spin_unlock(&cachep->spinlock);
+       spin_unlock(&l3->list_lock);
 
        if (unlikely(!ac->avail)) {
                int x;
-               x = cache_grow(cachep, flags, -1);
-               
+               x = cache_grow(cachep, flags, numa_node_id());
+
                // cache_grow can reenable interrupts, then ac could change.
                ac = ac_data(cachep);
                if (!x && ac->avail == 0)       // no objects in sight? abort
@@ -2088,7 +2450,7 @@ alloc_done:
                        goto retry;
        }
        ac->touched = 1;
-       return ac_entry(ac)[--ac->avail];
+       return ac->entry[--ac->avail];
 }
 
 static inline void
@@ -2160,7 +2522,7 @@ static inline void *__cache_alloc(kmem_cache_t *cachep, unsigned int __nocast fl
        if (likely(ac->avail)) {
                STATS_INC_ALLOCHIT(cachep);
                ac->touched = 1;
-               objp = ac_entry(ac)[--ac->avail];
+               objp = ac->entry[--ac->avail];
        } else {
                STATS_INC_ALLOCMISS(cachep);
                objp = cache_alloc_refill(cachep, flags);
@@ -2172,33 +2534,104 @@ static inline void *__cache_alloc(kmem_cache_t *cachep, unsigned int __nocast fl
        return objp;
 }
 
-/* 
- * NUMA: different approach needed if the spinlock is moved into
- * the l3 structure
+#ifdef CONFIG_NUMA
+/*
+ * A interface to enable slab creation on nodeid
  */
+static void *__cache_alloc_node(kmem_cache_t *cachep, int flags, int nodeid)
+{
+       struct list_head *entry;
+       struct slab *slabp;
+       struct kmem_list3 *l3;
+       void *obj;
+       kmem_bufctl_t next;
+       int x;
+
+       l3 = cachep->nodelists[nodeid];
+       BUG_ON(!l3);
+
+retry:
+       spin_lock(&l3->list_lock);
+       entry = l3->slabs_partial.next;
+       if (entry == &l3->slabs_partial) {
+               l3->free_touched = 1;
+               entry = l3->slabs_free.next;
+               if (entry == &l3->slabs_free)
+                       goto must_grow;
+       }
+
+       slabp = list_entry(entry, struct slab, list);
+       check_spinlock_acquired_node(cachep, nodeid);
+       check_slabp(cachep, slabp);
+
+       STATS_INC_NODEALLOCS(cachep);
+       STATS_INC_ACTIVE(cachep);
+       STATS_SET_HIGH(cachep);
+
+       BUG_ON(slabp->inuse == cachep->num);
+
+       /* get obj pointer */
+       obj =  slabp->s_mem + slabp->free*cachep->objsize;
+       slabp->inuse++;
+       next = slab_bufctl(slabp)[slabp->free];
+#if DEBUG
+       slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
+#endif
+       slabp->free = next;
+       check_slabp(cachep, slabp);
+       l3->free_objects--;
+       /* move slabp to correct slabp list: */
+       list_del(&slabp->list);
+
+       if (slabp->free == BUFCTL_END) {
+               list_add(&slabp->list, &l3->slabs_full);
+       } else {
+               list_add(&slabp->list, &l3->slabs_partial);
+       }
+
+       spin_unlock(&l3->list_lock);
+       goto done;
+
+must_grow:
+       spin_unlock(&l3->list_lock);
+       x = cache_grow(cachep, flags, nodeid);
 
+       if (!x)
+               return NULL;
+
+       goto retry;
+done:
+       return obj;
+}
+#endif
+
+/*
+ * Caller needs to acquire correct kmem_list's list_lock
+ */
 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
 {
        int i;
-
-       check_spinlock_acquired(cachep);
-
-       /* NUMA: move add into loop */
-       cachep->lists.free_objects += nr_objects;
+       struct kmem_list3 *l3;
 
        for (i = 0; i < nr_objects; i++) {
                void *objp = objpp[i];
                struct slab *slabp;
                unsigned int objnr;
+               int nodeid = 0;
 
                slabp = GET_PAGE_SLAB(virt_to_page(objp));
+               nodeid = slabp->nodeid;
+               l3 = cachep->nodelists[nodeid];
                list_del(&slabp->list);
                objnr = (objp - slabp->s_mem) / cachep->objsize;
+               check_spinlock_acquired_node(cachep, nodeid);
                check_slabp(cachep, slabp);
+
+
 #if DEBUG
                if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
-                       printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
-                                               cachep->name, objp);
+                       printk(KERN_ERR "slab: double free detected in cache "
+                                       "'%s', objp %p\n", cachep->name, objp);
                        BUG();
                }
 #endif
@@ -2206,24 +2639,23 @@ static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
                slabp->free = objnr;
                STATS_DEC_ACTIVE(cachep);
                slabp->inuse--;
+               l3->free_objects++;
                check_slabp(cachep, slabp);
 
                /* fixup slab chains */
                if (slabp->inuse == 0) {
-                       if (cachep->lists.free_objects > cachep->free_limit) {
-                               cachep->lists.free_objects -= cachep->num;
+                       if (l3->free_objects > l3->free_limit) {
+                               l3->free_objects -= cachep->num;
                                slab_destroy(cachep, slabp);
                        } else {
-                               list_add(&slabp->list,
-                               &list3_data_ptr(cachep, objp)->slabs_free);
+                               list_add(&slabp->list, &l3->slabs_free);
                        }
                } else {
                        /* Unconditionally move a slab to the end of the
                         * partial list on free - maximum time for the
                         * other objects to be freed, too.
                         */
-                       list_add_tail(&slabp->list,
-                               &list3_data_ptr(cachep, objp)->slabs_partial);
+                       list_add_tail(&slabp->list, &l3->slabs_partial);
                }
        }
 }
@@ -2231,36 +2663,38 @@ static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
 static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
 {
        int batchcount;
+       struct kmem_list3 *l3;
 
        batchcount = ac->batchcount;
 #if DEBUG
        BUG_ON(!batchcount || batchcount > ac->avail);
 #endif
        check_irq_off();
-       spin_lock(&cachep->spinlock);
-       if (cachep->lists.shared) {
-               struct array_cache *shared_array = cachep->lists.shared;
+       l3 = cachep->nodelists[numa_node_id()];
+       spin_lock(&l3->list_lock);
+       if (l3->shared) {
+               struct array_cache *shared_array = l3->shared;
                int max = shared_array->limit-shared_array->avail;
                if (max) {
                        if (batchcount > max)
                                batchcount = max;
-                       memcpy(&ac_entry(shared_array)[shared_array->avail],
-                                       &ac_entry(ac)[0],
+                       memcpy(&(shared_array->entry[shared_array->avail]),
+                                       ac->entry,
                                        sizeof(void*)*batchcount);
                        shared_array->avail += batchcount;
                        goto free_done;
                }
        }
 
-       free_block(cachep, &ac_entry(ac)[0], batchcount);
+       free_block(cachep, ac->entry, batchcount);
 free_done:
 #if STATS
        {
                int i = 0;
                struct list_head *p;
 
-               p = list3_data(cachep)->slabs_free.next;
-               while (p != &(list3_data(cachep)->slabs_free)) {
+               p = l3->slabs_free.next;
+               while (p != &(l3->slabs_free)) {
                        struct slab *slabp;
 
                        slabp = list_entry(p, struct slab, list);
@@ -2272,12 +2706,13 @@ free_done:
                STATS_SET_FREEABLE(cachep, i);
        }
 #endif
-       spin_unlock(&cachep->spinlock);
+       spin_unlock(&l3->list_lock);
        ac->avail -= batchcount;
-       memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
+       memmove(ac->entry, &(ac->entry[batchcount]),
                        sizeof(void*)*ac->avail);
 }
 
+
 /*
  * __cache_free
  * Release an obj back to its cache. If the obj has a constructed
@@ -2292,14 +2727,46 @@ static inline void __cache_free(kmem_cache_t *cachep, void *objp)
        check_irq_off();
        objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
 
+       /* Make sure we are not freeing a object from another
+        * node to the array cache on this cpu.
+        */
+#ifdef CONFIG_NUMA
+       {
+               struct slab *slabp;
+               slabp = GET_PAGE_SLAB(virt_to_page(objp));
+               if (unlikely(slabp->nodeid != numa_node_id())) {
+                       struct array_cache *alien = NULL;
+                       int nodeid = slabp->nodeid;
+                       struct kmem_list3 *l3 = cachep->nodelists[numa_node_id()];
+
+                       STATS_INC_NODEFREES(cachep);
+                       if (l3->alien && l3->alien[nodeid]) {
+                               alien = l3->alien[nodeid];
+                               spin_lock(&alien->lock);
+                               if (unlikely(alien->avail == alien->limit))
+                                       __drain_alien_cache(cachep,
+                                                       alien, nodeid);
+                               alien->entry[alien->avail++] = objp;
+                               spin_unlock(&alien->lock);
+                       } else {
+                               spin_lock(&(cachep->nodelists[nodeid])->
+                                               list_lock);
+                               free_block(cachep, &objp, 1);
+                               spin_unlock(&(cachep->nodelists[nodeid])->
+                                               list_lock);
+                       }
+                       return;
+               }
+       }
+#endif
        if (likely(ac->avail < ac->limit)) {
                STATS_INC_FREEHIT(cachep);
-               ac_entry(ac)[ac->avail++] = objp;
+               ac->entry[ac->avail++] = objp;
                return;
        } else {
                STATS_INC_FREEMISS(cachep);
                cache_flusharray(cachep, ac);
-               ac_entry(ac)[ac->avail++] = objp;
+               ac->entry[ac->avail++] = objp;
        }
 }
 
@@ -2369,81 +2836,30 @@ out:
  * Identical to kmem_cache_alloc, except that this function is slow
  * and can sleep. And it will allocate memory on the given node, which
  * can improve the performance for cpu bound structures.
+ * New and improved: it will now make sure that the object gets
+ * put on the correct node list so that there is no false sharing.
  */
 void *kmem_cache_alloc_node(kmem_cache_t *cachep, int flags, int nodeid)
 {
-       int loop;
-       void *objp;
-       struct slab *slabp;
-       kmem_bufctl_t next;
-
-       if (nodeid == -1)
-               return kmem_cache_alloc(cachep, flags);
-
-       for (loop = 0;;loop++) {
-               struct list_head *q;
-
-               objp = NULL;
-               check_irq_on();
-               spin_lock_irq(&cachep->spinlock);
-               /* walk through all partial and empty slab and find one
-                * from the right node */
-               list_for_each(q,&cachep->lists.slabs_partial) {
-                       slabp = list_entry(q, struct slab, list);
-
-                       if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid ||
-                                       loop > 2)
-                               goto got_slabp;
-               }
-               list_for_each(q, &cachep->lists.slabs_free) {
-                       slabp = list_entry(q, struct slab, list);
+       unsigned long save_flags;
+       void *ptr;
 
-                       if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid ||
-                                       loop > 2)
-                               goto got_slabp;
-               }
-               spin_unlock_irq(&cachep->spinlock);
+       if (nodeid == numa_node_id() || nodeid == -1)
+               return __cache_alloc(cachep, flags);
 
-               local_irq_disable();
-               if (!cache_grow(cachep, flags, nodeid)) {
-                       local_irq_enable();
-                       return NULL;
-               }
-               local_irq_enable();
+       if (unlikely(!cachep->nodelists[nodeid])) {
+               /* Fall back to __cache_alloc if we run into trouble */
+               printk(KERN_WARNING "slab: not allocating in inactive node %d for cache %s\n", nodeid, cachep->name);
+               return __cache_alloc(cachep,flags);
        }
-got_slabp:
-       /* found one: allocate object */
-       check_slabp(cachep, slabp);
-       check_spinlock_acquired(cachep);
 
-       STATS_INC_ALLOCED(cachep);
-       STATS_INC_ACTIVE(cachep);
-       STATS_SET_HIGH(cachep);
-       STATS_INC_NODEALLOCS(cachep);
-
-       objp = slabp->s_mem + slabp->free*cachep->objsize;
-
-       slabp->inuse++;
-       next = slab_bufctl(slabp)[slabp->free];
-#if DEBUG
-       slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
-#endif
-       slabp->free = next;
-       check_slabp(cachep, slabp);
-
-       /* move slabp to correct slabp list: */
-       list_del(&slabp->list);
-       if (slabp->free == BUFCTL_END)
-               list_add(&slabp->list, &cachep->lists.slabs_full);
-       else
-               list_add(&slabp->list, &cachep->lists.slabs_partial);
-
-       list3_data(cachep)->free_objects--;
-       spin_unlock_irq(&cachep->spinlock);
+       cache_alloc_debugcheck_before(cachep, flags);
+       local_irq_save(save_flags);
+       ptr = __cache_alloc_node(cachep, flags, nodeid);
+       local_irq_restore(save_flags);
+       ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, __builtin_return_address(0));
 
-       objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp,
-                                       __builtin_return_address(0));
-       return objp;
+       return ptr;
 }
 EXPORT_SYMBOL(kmem_cache_alloc_node);
 
@@ -2513,11 +2929,18 @@ void *__alloc_percpu(size_t size, size_t align)
        if (!pdata)
                return NULL;
 
-       for (i = 0; i < NR_CPUS; i++) {
-               if (!cpu_possible(i))
-                       continue;
-               pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL,
-                                               cpu_to_node(i));
+       /*
+        * Cannot use for_each_online_cpu since a cpu may come online
+        * and we have no way of figuring out how to fix the array
+        * that we have allocated then....
+        */
+       for_each_cpu(i) {
+               int node = cpu_to_node(i);
+
+               if (node_online(node))
+                       pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
+               else
+                       pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
 
                if (!pdata->ptrs[i])
                        goto unwind_oom;
@@ -2607,11 +3030,11 @@ free_percpu(const void *objp)
        int i;
        struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
 
-       for (i = 0; i < NR_CPUS; i++) {
-               if (!cpu_possible(i))
-                       continue;
+       /*
+        * We allocate for all cpus so we cannot use for online cpu here.
+        */
+       for_each_cpu(i)
                kfree(p->ptrs[i]);
-       }
        kfree(p);
 }
 EXPORT_SYMBOL(free_percpu);
@@ -2629,6 +3052,64 @@ const char *kmem_cache_name(kmem_cache_t *cachep)
 }
 EXPORT_SYMBOL_GPL(kmem_cache_name);
 
+/*
+ * This initializes kmem_list3 for all nodes.
+ */
+static int alloc_kmemlist(kmem_cache_t *cachep)
+{
+       int node;
+       struct kmem_list3 *l3;
+       int err = 0;
+
+       for_each_online_node(node) {
+               struct array_cache *nc = NULL, *new;
+               struct array_cache **new_alien = NULL;
+#ifdef CONFIG_NUMA
+               if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
+                       goto fail;
+#endif
+               if (!(new = alloc_arraycache(node, (cachep->shared*
+                               cachep->batchcount), 0xbaadf00d)))
+                       goto fail;
+               if ((l3 = cachep->nodelists[node])) {
+
+                       spin_lock_irq(&l3->list_lock);
+
+                       if ((nc = cachep->nodelists[node]->shared))
+                               free_block(cachep, nc->entry,
+                                                       nc->avail);
+
+                       l3->shared = new;
+                       if (!cachep->nodelists[node]->alien) {
+                               l3->alien = new_alien;
+                               new_alien = NULL;
+                       }
+                       l3->free_limit = (1 + nr_cpus_node(node))*
+                               cachep->batchcount + cachep->num;
+                       spin_unlock_irq(&l3->list_lock);
+                       kfree(nc);
+                       free_alien_cache(new_alien);
+                       continue;
+               }
+               if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
+                                               GFP_KERNEL, node)))
+                       goto fail;
+
+               kmem_list3_init(l3);
+               l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
+                       ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
+               l3->shared = new;
+               l3->alien = new_alien;
+               l3->free_limit = (1 + nr_cpus_node(node))*
+                       cachep->batchcount + cachep->num;
+               cachep->nodelists[node] = l3;
+       }
+       return err;
+fail:
+       err = -ENOMEM;
+       return err;
+}
+
 struct ccupdate_struct {
        kmem_cache_t *cachep;
        struct array_cache *new[NR_CPUS];
@@ -2641,7 +3122,7 @@ static void do_ccupdate_local(void *info)
 
        check_irq_off();
        old = ac_data(new->cachep);
-       
+
        new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
        new->new[smp_processor_id()] = old;
 }
@@ -2651,54 +3132,43 @@ static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
                                int shared)
 {
        struct ccupdate_struct new;
-       struct array_cache *new_shared;
-       int i;
+       int i, err;
 
        memset(&new.new,0,sizeof(new.new));
-       for (i = 0; i < NR_CPUS; i++) {
-               if (cpu_online(i)) {
-                       new.new[i] = alloc_arraycache(i, limit, batchcount);
-                       if (!new.new[i]) {
-                               for (i--; i >= 0; i--) kfree(new.new[i]);
-                               return -ENOMEM;
-                       }
-               } else {
-                       new.new[i] = NULL;
+       for_each_online_cpu(i) {
+               new.new[i] = alloc_arraycache(cpu_to_node(i), limit, batchcount);
+               if (!new.new[i]) {
+                       for (i--; i >= 0; i--) kfree(new.new[i]);
+                       return -ENOMEM;
                }
        }
        new.cachep = cachep;
 
        smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
-       
+
        check_irq_on();
        spin_lock_irq(&cachep->spinlock);
        cachep->batchcount = batchcount;
        cachep->limit = limit;
-       cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
+       cachep->shared = shared;
        spin_unlock_irq(&cachep->spinlock);
 
-       for (i = 0; i < NR_CPUS; i++) {
+       for_each_online_cpu(i) {
                struct array_cache *ccold = new.new[i];
                if (!ccold)
                        continue;
-               spin_lock_irq(&cachep->spinlock);
-               free_block(cachep, ac_entry(ccold), ccold->avail);
-               spin_unlock_irq(&cachep->spinlock);
+               spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
+               free_block(cachep, ccold->entry, ccold->avail);
+               spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
                kfree(ccold);
        }
-       new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d);
-       if (new_shared) {
-               struct array_cache *old;
 
-               spin_lock_irq(&cachep->spinlock);
-               old = cachep->lists.shared;
-               cachep->lists.shared = new_shared;
-               if (old)
-                       free_block(cachep, ac_entry(old), old->avail);
-               spin_unlock_irq(&cachep->spinlock);
-               kfree(old);
+       err = alloc_kmemlist(cachep);
+       if (err) {
+               printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
+                               cachep->name, -err);
+               BUG();
        }
-
        return 0;
 }
 
@@ -2756,11 +3226,11 @@ static void enable_cpucache(kmem_cache_t *cachep)
 }
 
 static void drain_array_locked(kmem_cache_t *cachep,
-                               struct array_cache *ac, int force)
+                               struct array_cache *ac, int force, int node)
 {
        int tofree;
 
-       check_spinlock_acquired(cachep);
+       check_spinlock_acquired_node(cachep, node);
        if (ac->touched && !force) {
                ac->touched = 0;
        } else if (ac->avail) {
@@ -2768,9 +3238,9 @@ static void drain_array_locked(kmem_cache_t *cachep,
                if (tofree > ac->avail) {
                        tofree = (ac->avail+1)/2;
                }
-               free_block(cachep, ac_entry(ac), tofree);
+               free_block(cachep, ac->entry, tofree);
                ac->avail -= tofree;
-               memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
+               memmove(ac->entry, &(ac->entry[tofree]),
                                        sizeof(void*)*ac->avail);
        }
 }
@@ -2789,6 +3259,7 @@ static void drain_array_locked(kmem_cache_t *cachep,
 static void cache_reap(void *unused)
 {
        struct list_head *walk;
+       struct kmem_list3 *l3;
 
        if (down_trylock(&cache_chain_sem)) {
                /* Give up. Setup the next iteration. */
@@ -2809,27 +3280,32 @@ static void cache_reap(void *unused)
 
                check_irq_on();
 
-               spin_lock_irq(&searchp->spinlock);
+               l3 = searchp->nodelists[numa_node_id()];
+               if (l3->alien)
+                       drain_alien_cache(searchp, l3);
+               spin_lock_irq(&l3->list_lock);
 
-               drain_array_locked(searchp, ac_data(searchp), 0);
+               drain_array_locked(searchp, ac_data(searchp), 0,
+                               numa_node_id());
 
-               if(time_after(searchp->lists.next_reap, jiffies))
+               if (time_after(l3->next_reap, jiffies))
                        goto next_unlock;
 
-               searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
+               l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
 
-               if (searchp->lists.shared)
-                       drain_array_locked(searchp, searchp->lists.shared, 0);
+               if (l3->shared)
+                       drain_array_locked(searchp, l3->shared, 0,
+                               numa_node_id());
 
-               if (searchp->lists.free_touched) {
-                       searchp->lists.free_touched = 0;
+               if (l3->free_touched) {
+                       l3->free_touched = 0;
                        goto next_unlock;
                }
 
-               tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
+               tofree = (l3->free_limit+5*searchp->num-1)/(5*searchp->num);
                do {
-                       p = list3_data(searchp)->slabs_free.next;
-                       if (p == &(list3_data(searchp)->slabs_free))
+                       p = l3->slabs_free.next;
+                       if (p == &(l3->slabs_free))
                                break;
 
                        slabp = list_entry(p, struct slab, list);
@@ -2842,13 +3318,13 @@ static void cache_reap(void *unused)
                         * searchp cannot disappear, we hold
                         * cache_chain_lock
                         */
-                       searchp->lists.free_objects -= searchp->num;
-                       spin_unlock_irq(&searchp->spinlock);
+                       l3->free_objects -= searchp->num;
+                       spin_unlock_irq(&l3->list_lock);
                        slab_destroy(searchp, slabp);
-                       spin_lock_irq(&searchp->spinlock);
+                       spin_lock_irq(&l3->list_lock);
                } while(--tofree > 0);
 next_unlock:
-               spin_unlock_irq(&searchp->spinlock);
+               spin_unlock_irq(&l3->list_lock);
 next:
                cond_resched();
        }
@@ -2882,7 +3358,7 @@ static void *s_start(struct seq_file *m, loff_t *pos)
                seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
 #if STATS
                seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>"
-                               " <error> <maxfreeable> <freelimit> <nodeallocs>");
+                               " <error> <maxfreeable> <nodeallocs> <remotefrees>");
                seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
 #endif
                seq_putc(m, '\n');
@@ -2917,39 +3393,53 @@ static int s_show(struct seq_file *m, void *p)
        unsigned long   active_objs;
        unsigned long   num_objs;
        unsigned long   active_slabs = 0;
-       unsigned long   num_slabs;
-       const char *name; 
+       unsigned long   num_slabs, free_objects = 0, shared_avail = 0;
+       const char *name;
        char *error = NULL;
+       int node;
+       struct kmem_list3 *l3;
 
        check_irq_on();
        spin_lock_irq(&cachep->spinlock);
        active_objs = 0;
        num_slabs = 0;
-       list_for_each(q,&cachep->lists.slabs_full) {
-               slabp = list_entry(q, struct slab, list);
-               if (slabp->inuse != cachep->num && !error)
-                       error = "slabs_full accounting error";
-               active_objs += cachep->num;
-               active_slabs++;
-       }
-       list_for_each(q,&cachep->lists.slabs_partial) {
-               slabp = list_entry(q, struct slab, list);
-               if (slabp->inuse == cachep->num && !error)
-                       error = "slabs_partial inuse accounting error";
-               if (!slabp->inuse && !error)
-                       error = "slabs_partial/inuse accounting error";
-               active_objs += slabp->inuse;
-               active_slabs++;
-       }
-       list_for_each(q,&cachep->lists.slabs_free) {
-               slabp = list_entry(q, struct slab, list);
-               if (slabp->inuse && !error)
-                       error = "slabs_free/inuse accounting error";
-               num_slabs++;
+       for_each_online_node(node) {
+               l3 = cachep->nodelists[node];
+               if (!l3)
+                       continue;
+
+               spin_lock(&l3->list_lock);
+
+               list_for_each(q,&l3->slabs_full) {
+                       slabp = list_entry(q, struct slab, list);
+                       if (slabp->inuse != cachep->num && !error)
+                               error = "slabs_full accounting error";
+                       active_objs += cachep->num;
+                       active_slabs++;
+               }
+               list_for_each(q,&l3->slabs_partial) {
+                       slabp = list_entry(q, struct slab, list);
+                       if (slabp->inuse == cachep->num && !error)
+                               error = "slabs_partial inuse accounting error";
+                       if (!slabp->inuse && !error)
+                               error = "slabs_partial/inuse accounting error";
+                       active_objs += slabp->inuse;
+                       active_slabs++;
+               }
+               list_for_each(q,&l3->slabs_free) {
+                       slabp = list_entry(q, struct slab, list);
+                       if (slabp->inuse && !error)
+                               error = "slabs_free/inuse accounting error";
+                       num_slabs++;
+               }
+               free_objects += l3->free_objects;
+               shared_avail += l3->shared->avail;
+
+               spin_unlock(&l3->list_lock);
        }
        num_slabs+=active_slabs;
        num_objs = num_slabs*cachep->num;
-       if (num_objs - active_objs != cachep->lists.free_objects && !error)
+       if (num_objs - active_objs != free_objects && !error)
                error = "free_objects accounting error";
 
        name = cachep->name; 
@@ -2961,9 +3451,9 @@ static int s_show(struct seq_file *m, void *p)
                cachep->num, (1<<cachep->gfporder));
        seq_printf(m, " : tunables %4u %4u %4u",
                        cachep->limit, cachep->batchcount,
-                       cachep->lists.shared->limit/cachep->batchcount);
-       seq_printf(m, " : slabdata %6lu %6lu %6u",
-                       active_slabs, num_slabs, cachep->lists.shared->avail);
+                       cachep->shared);
+       seq_printf(m, " : slabdata %6lu %6lu %6lu",
+                       active_slabs, num_slabs, shared_avail);
 #if STATS
        {       /* list3 stats */
                unsigned long high = cachep->high_mark;
@@ -2972,12 +3462,13 @@ static int s_show(struct seq_file *m, void *p)
                unsigned long reaped = cachep->reaped;
                unsigned long errors = cachep->errors;
                unsigned long max_freeable = cachep->max_freeable;
-               unsigned long free_limit = cachep->free_limit;
                unsigned long node_allocs = cachep->node_allocs;
+               unsigned long node_frees = cachep->node_frees;
 
-               seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu",
-                               allocs, high, grown, reaped, errors, 
-                               max_freeable, free_limit, node_allocs);
+               seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
+                               %4lu %4lu %4lu %4lu",
+                               allocs, high, grown, reaped, errors,
+                               max_freeable, node_allocs, node_frees);
        }
        /* cpu stats */
        {
@@ -3056,9 +3547,10 @@ ssize_t slabinfo_write(struct file *file, const char __user *buffer,
                            batchcount < 1 ||
                            batchcount > limit ||
                            shared < 0) {
-                               res = -EINVAL;
+                               res = 0;
                        } else {
-                               res = do_tune_cpucache(cachep, limit, batchcount, shared);
+                               res = do_tune_cpucache(cachep, limit,
+                                                       batchcount, shared);
                        }
                        break;
                }