blob: 6dd8d5f3a3ac6d223e4830d489b5e52c0f8773e3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Christoph Lameter18004c52012-07-06 15:25:12 -050071 * The global cache-chain is protected by the mutex 'slab_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700118#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
Mel Gorman381760e2012-07-31 16:44:30 -0700120#include <net/sock.h>
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500126#include <trace/events/kmem.h>
127
Mel Gorman072bb0a2012-07-31 16:43:58 -0700128#include "internal.h"
129
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800130#include "slab.h"
131
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
Mel Gorman072bb0a2012-07-31 16:43:58 -0700160/*
161 * true if a page was allocated from pfmemalloc reserves for network-based
162 * swap
163 */
164static bool pfmemalloc_active __read_mostly;
165
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 * struct array_cache
168 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 * Purpose:
170 * - LIFO ordering, to hand out cache-warm objects from _alloc
171 * - reduce the number of linked list operations
172 * - reduce spinlock operations
173 *
174 * The limit is stored in the per-cpu structure to reduce the data cache
175 * footprint.
176 *
177 */
178struct array_cache {
179 unsigned int avail;
180 unsigned int limit;
181 unsigned int batchcount;
182 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700183 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700184 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800185 * Must have this definition in here for the proper
186 * alignment of array_cache. Also simplifies accessing
187 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700188 *
189 * Entries should not be directly dereferenced as
190 * entries belonging to slabs marked pfmemalloc will
191 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800192 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193};
194
Mel Gorman072bb0a2012-07-31 16:43:58 -0700195#define SLAB_OBJ_PFMEMALLOC 1
196static inline bool is_obj_pfmemalloc(void *objp)
197{
198 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
199}
200
201static inline void set_obj_pfmemalloc(void **objp)
202{
203 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
204 return;
205}
206
207static inline void clear_obj_pfmemalloc(void **objp)
208{
209 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
210}
211
Andrew Mortona737b3e2006-03-22 00:08:11 -0800212/*
213 * bootstrap: The caches do not work without cpuarrays anymore, but the
214 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 */
216#define BOOT_CPUCACHE_ENTRIES 1
217struct arraycache_init {
218 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800219 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220};
221
222/*
Christoph Lametere498be72005-09-09 13:03:32 -0700223 * Need this for bootstrapping a per node allocator.
224 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200225#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000226static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700227#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200228#define SIZE_AC MAX_NUMNODES
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000229#define SIZE_NODE (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
Christoph Lametered11d9e2006-06-30 01:55:45 -0700231static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000232 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700233static void free_block(struct kmem_cache *cachep, void **objpp, int len,
234 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300235static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000236static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700237
Ingo Molnare0a42722006-06-23 02:03:46 -0700238static int slab_early_init = 1;
239
Christoph Lametere3366012013-01-10 19:14:18 +0000240#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000241#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700242
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000243static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700244{
245 INIT_LIST_HEAD(&parent->slabs_full);
246 INIT_LIST_HEAD(&parent->slabs_partial);
247 INIT_LIST_HEAD(&parent->slabs_free);
248 parent->shared = NULL;
249 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800250 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700251 spin_lock_init(&parent->list_lock);
252 parent->free_objects = 0;
253 parent->free_touched = 0;
254}
255
Andrew Mortona737b3e2006-03-22 00:08:11 -0800256#define MAKE_LIST(cachep, listp, slab, nodeid) \
257 do { \
258 INIT_LIST_HEAD(listp); \
Christoph Lameter6a673682013-01-10 19:14:19 +0000259 list_splice(&(cachep->node[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700260 } while (0)
261
Andrew Mortona737b3e2006-03-22 00:08:11 -0800262#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
263 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700264 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
265 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
266 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
267 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269#define CFLGS_OFF_SLAB (0x80000000UL)
270#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
271
272#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273/*
274 * Optimization question: fewer reaps means less probability for unnessary
275 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100277 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 * which could lock up otherwise freeable slabs.
279 */
280#define REAPTIMEOUT_CPUC (2*HZ)
281#define REAPTIMEOUT_LIST3 (4*HZ)
282
283#if STATS
284#define STATS_INC_ACTIVE(x) ((x)->num_active++)
285#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
286#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
287#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700288#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800289#define STATS_SET_HIGH(x) \
290 do { \
291 if ((x)->num_active > (x)->high_mark) \
292 (x)->high_mark = (x)->num_active; \
293 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294#define STATS_INC_ERR(x) ((x)->errors++)
295#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700296#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700297#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800298#define STATS_SET_FREEABLE(x, i) \
299 do { \
300 if ((x)->max_freeable < i) \
301 (x)->max_freeable = i; \
302 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
304#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
305#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
306#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
307#else
308#define STATS_INC_ACTIVE(x) do { } while (0)
309#define STATS_DEC_ACTIVE(x) do { } while (0)
310#define STATS_INC_ALLOCED(x) do { } while (0)
311#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700312#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313#define STATS_SET_HIGH(x) do { } while (0)
314#define STATS_INC_ERR(x) do { } while (0)
315#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700316#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700317#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800318#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319#define STATS_INC_ALLOCHIT(x) do { } while (0)
320#define STATS_INC_ALLOCMISS(x) do { } while (0)
321#define STATS_INC_FREEHIT(x) do { } while (0)
322#define STATS_INC_FREEMISS(x) do { } while (0)
323#endif
324
325#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326
Andrew Mortona737b3e2006-03-22 00:08:11 -0800327/*
328 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800330 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 * the end of an object is aligned with the end of the real
332 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800333 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800335 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500336 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
337 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800338 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800340static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800342 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343}
344
David Woodhouseb46b8f12007-05-08 00:22:59 -0700345static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346{
347 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700348 return (unsigned long long*) (objp + obj_offset(cachep) -
349 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350}
351
David Woodhouseb46b8f12007-05-08 00:22:59 -0700352static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
354 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
355 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500356 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700357 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400358 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500359 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700360 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361}
362
Pekka Enberg343e0d72006-02-01 03:05:50 -0800363static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
365 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500366 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367}
368
369#else
370
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800371#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700372#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
373#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
375
376#endif
377
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700378#define OBJECT_FREE (0)
379#define OBJECT_ACTIVE (1)
380
381#ifdef CONFIG_DEBUG_SLAB_LEAK
382
383static void set_obj_status(struct page *page, int idx, int val)
384{
385 int freelist_size;
386 char *status;
387 struct kmem_cache *cachep = page->slab_cache;
388
389 freelist_size = cachep->num * sizeof(unsigned int);
390 status = (char *)page->freelist + freelist_size;
391 status[idx] = val;
392}
393
394static inline unsigned int get_obj_status(struct page *page, int idx)
395{
396 int freelist_size;
397 char *status;
398 struct kmem_cache *cachep = page->slab_cache;
399
400 freelist_size = cachep->num * sizeof(unsigned int);
401 status = (char *)page->freelist + freelist_size;
402
403 return status[idx];
404}
405
406#else
407static inline void set_obj_status(struct page *page, int idx, int val) {}
408
409#endif
410
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700412 * Do not go above this order unless 0 objects fit into the slab or
413 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 */
David Rientjes543585c2011-10-18 22:09:24 -0700415#define SLAB_MAX_ORDER_HI 1
416#define SLAB_MAX_ORDER_LO 0
417static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700418static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800420static inline struct kmem_cache *virt_to_cache(const void *obj)
421{
Christoph Lameterb49af682007-05-06 14:49:41 -0700422 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500423 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800424}
425
Joonsoo Kim8456a642013-10-24 10:07:49 +0900426static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800427 unsigned int idx)
428{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900429 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800430}
431
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800432/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500433 * We want to avoid an expensive divide : (offset / cache->size)
434 * Using the fact that size is a constant for a particular cache,
435 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800436 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
437 */
438static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900439 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800440{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900441 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800442 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800443}
444
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800446 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447
448/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000449static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800450 .batchcount = 1,
451 .limit = BOOT_CPUCACHE_ENTRIES,
452 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500453 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800454 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455};
456
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700457#define BAD_ALIEN_MAGIC 0x01020304ul
458
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200459#ifdef CONFIG_LOCKDEP
460
461/*
462 * Slab sometimes uses the kmalloc slabs to store the slab headers
463 * for other slabs "off slab".
464 * The locking for this is tricky in that it nests within the locks
465 * of all other slabs in a few places; to deal with this special
466 * locking we put on-slab caches into a separate lock-class.
467 *
468 * We set lock class for alien array caches which are up during init.
469 * The lock annotation will be lost if all cpus of a node goes down and
470 * then comes back up during hotplug
471 */
472static struct lock_class_key on_slab_l3_key;
473static struct lock_class_key on_slab_alc_key;
474
Peter Zijlstra83835b32011-07-22 15:26:05 +0200475static struct lock_class_key debugobj_l3_key;
476static struct lock_class_key debugobj_alc_key;
477
478static void slab_set_lock_classes(struct kmem_cache *cachep,
479 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
480 int q)
481{
482 struct array_cache **alc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000483 struct kmem_cache_node *n;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200484 int r;
485
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000486 n = cachep->node[q];
487 if (!n)
Peter Zijlstra83835b32011-07-22 15:26:05 +0200488 return;
489
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000490 lockdep_set_class(&n->list_lock, l3_key);
491 alc = n->alien;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200492 /*
493 * FIXME: This check for BAD_ALIEN_MAGIC
494 * should go away when common slab code is taught to
495 * work even without alien caches.
496 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
497 * for alloc_alien_cache,
498 */
499 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
500 return;
501 for_each_node(r) {
502 if (alc[r])
503 lockdep_set_class(&alc[r]->lock, alc_key);
504 }
505}
506
507static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
508{
509 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
510}
511
512static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
513{
514 int node;
515
516 for_each_online_node(node)
517 slab_set_debugobj_lock_classes_node(cachep, node);
518}
519
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200520static void init_node_lock_keys(int q)
521{
Christoph Lametere3366012013-01-10 19:14:18 +0000522 int i;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200523
Christoph Lameter97d06602012-07-06 15:25:11 -0500524 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200525 return;
526
Christoph Lameter0f8f8092013-07-02 12:12:10 -0700527 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000528 struct kmem_cache_node *n;
Christoph Lametere3366012013-01-10 19:14:18 +0000529 struct kmem_cache *cache = kmalloc_caches[i];
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200530
Christoph Lametere3366012013-01-10 19:14:18 +0000531 if (!cache)
Pekka Enberg00afa752009-12-27 14:33:14 +0200532 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200533
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000534 n = cache->node[q];
535 if (!n || OFF_SLAB(cache))
Christoph Lametere3366012013-01-10 19:14:18 +0000536 continue;
537
538 slab_set_lock_classes(cache, &on_slab_l3_key,
Peter Zijlstra83835b32011-07-22 15:26:05 +0200539 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200540 }
541}
542
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800543static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
544{
Christoph Lameter6a673682013-01-10 19:14:19 +0000545 if (!cachep->node[q])
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800546 return;
547
548 slab_set_lock_classes(cachep, &on_slab_l3_key,
549 &on_slab_alc_key, q);
550}
551
552static inline void on_slab_lock_classes(struct kmem_cache *cachep)
553{
554 int node;
555
556 VM_BUG_ON(OFF_SLAB(cachep));
557 for_each_node(node)
558 on_slab_lock_classes_node(cachep, node);
559}
560
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200561static inline void init_lock_keys(void)
562{
563 int node;
564
565 for_each_node(node)
566 init_node_lock_keys(node);
567}
568#else
569static void init_node_lock_keys(int q)
570{
571}
572
573static inline void init_lock_keys(void)
574{
575}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200576
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800577static inline void on_slab_lock_classes(struct kmem_cache *cachep)
578{
579}
580
581static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
582{
583}
584
Peter Zijlstra83835b32011-07-22 15:26:05 +0200585static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
586{
587}
588
589static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
590{
591}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200592#endif
593
Tejun Heo1871e522009-10-29 22:34:13 +0900594static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
Pekka Enberg343e0d72006-02-01 03:05:50 -0800596static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597{
598 return cachep->array[smp_processor_id()];
599}
600
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700601static size_t calculate_freelist_size(int nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602{
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700603 size_t freelist_size;
604
605 freelist_size = nr_objs * sizeof(unsigned int);
606 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
607 freelist_size += nr_objs * sizeof(char);
608
609 if (align)
610 freelist_size = ALIGN(freelist_size, align);
611
612 return freelist_size;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800613}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Andrew Mortona737b3e2006-03-22 00:08:11 -0800615/*
616 * Calculate the number of objects and left-over bytes for a given buffer size.
617 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800618static void cache_estimate(unsigned long gfporder, size_t buffer_size,
619 size_t align, int flags, size_t *left_over,
620 unsigned int *num)
621{
622 int nr_objs;
623 size_t mgmt_size;
624 size_t slab_size = PAGE_SIZE << gfporder;
625
626 /*
627 * The slab management structure can be either off the slab or
628 * on it. For the latter case, the memory allocated for a
629 * slab is used for:
630 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900631 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800632 * - Padding to respect alignment of @align
633 * - @buffer_size bytes for each object
634 *
635 * If the slab management structure is off the slab, then the
636 * alignment will already be calculated into the size. Because
637 * the slabs are all pages aligned, the objects will be at the
638 * correct alignment when allocated.
639 */
640 if (flags & CFLGS_OFF_SLAB) {
641 mgmt_size = 0;
642 nr_objs = slab_size / buffer_size;
643
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800644 } else {
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700645 int extra_space = 0;
646
647 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
648 extra_space = sizeof(char);
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800649 /*
650 * Ignore padding for the initial guess. The padding
651 * is at most @align-1 bytes, and @buffer_size is at
652 * least @align. In the worst case, this result will
653 * be one greater than the number of objects that fit
654 * into the memory allocation when taking the padding
655 * into account.
656 */
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700657 nr_objs = (slab_size) /
658 (buffer_size + sizeof(unsigned int) + extra_space);
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800659
660 /*
661 * This calculated number will be either the right
662 * amount, or one greater than what we want.
663 */
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700664 if (calculate_freelist_size(nr_objs, align) >
665 slab_size - nr_objs * buffer_size)
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800666 nr_objs--;
667
Joonsoo Kimd35426e2014-06-23 13:22:06 -0700668 mgmt_size = calculate_freelist_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800670 *num = nr_objs;
671 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672}
673
Christoph Lameterf28510d2012-09-11 19:49:38 +0000674#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700675#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676
Andrew Mortona737b3e2006-03-22 00:08:11 -0800677static void __slab_error(const char *function, struct kmem_cache *cachep,
678 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679{
680 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800681 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030683 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000685#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
Paul Menage3395ee02006-12-06 20:32:16 -0800687/*
688 * By default on NUMA we use alien caches to stage the freeing of
689 * objects allocated from other nodes. This causes massive memory
690 * inefficiencies when using fake NUMA setup to split memory into a
691 * large number of small nodes, so it can be disabled on the command
692 * line
693 */
694
695static int use_alien_caches __read_mostly = 1;
696static int __init noaliencache_setup(char *s)
697{
698 use_alien_caches = 0;
699 return 1;
700}
701__setup("noaliencache", noaliencache_setup);
702
David Rientjes3df1ccc2011-10-18 22:09:28 -0700703static int __init slab_max_order_setup(char *str)
704{
705 get_option(&str, &slab_max_order);
706 slab_max_order = slab_max_order < 0 ? 0 :
707 min(slab_max_order, MAX_ORDER - 1);
708 slab_max_order_set = true;
709
710 return 1;
711}
712__setup("slab_max_order=", slab_max_order_setup);
713
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800714#ifdef CONFIG_NUMA
715/*
716 * Special reaping functions for NUMA systems called from cache_reap().
717 * These take care of doing round robin flushing of alien caches (containing
718 * objects freed on different nodes from which they were allocated) and the
719 * flushing of remote pcps by calling drain_node_pages.
720 */
Tejun Heo1871e522009-10-29 22:34:13 +0900721static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800722
723static void init_reap_node(int cpu)
724{
725 int node;
726
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700727 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800728 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800729 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800730
Tejun Heo1871e522009-10-29 22:34:13 +0900731 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800732}
733
734static void next_reap_node(void)
735{
Christoph Lameter909ea962010-12-08 16:22:55 +0100736 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800737
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800738 node = next_node(node, node_online_map);
739 if (unlikely(node >= MAX_NUMNODES))
740 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100741 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800742}
743
744#else
745#define init_reap_node(cpu) do { } while (0)
746#define next_reap_node(void) do { } while (0)
747#endif
748
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749/*
750 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
751 * via the workqueue/eventd.
752 * Add the CPU number into the expiration time to minimize the possibility of
753 * the CPUs getting into lockstep and contending for the global cache chain
754 * lock.
755 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400756static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
Tejun Heo1871e522009-10-29 22:34:13 +0900758 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
760 /*
761 * When this gets called from do_initcalls via cpucache_init(),
762 * init_workqueues() has already run, so keventd will be setup
763 * at that time.
764 */
David Howells52bad642006-11-22 14:54:01 +0000765 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800766 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700767 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800768 schedule_delayed_work_on(cpu, reap_work,
769 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 }
771}
772
Christoph Lametere498be72005-09-09 13:03:32 -0700773static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300774 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800776 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 struct array_cache *nc = NULL;
778
Pekka Enberg83b519e2009-06-10 19:40:04 +0300779 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100780 /*
781 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300782 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100783 * cache the pointers are not cleared and they could be counted as
784 * valid references during a kmemleak scan. Therefore, kmemleak must
785 * not scan such objects.
786 */
787 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 if (nc) {
789 nc->avail = 0;
790 nc->limit = entries;
791 nc->batchcount = batchcount;
792 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700793 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 }
795 return nc;
796}
797
Joonsoo Kim8456a642013-10-24 10:07:49 +0900798static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700799{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700800 return PageSlabPfmemalloc(page);
801}
802
803/* Clears pfmemalloc_active if no slabs have pfmalloc set */
804static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
805 struct array_cache *ac)
806{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000807 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
Joonsoo Kim8456a642013-10-24 10:07:49 +0900808 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700809 unsigned long flags;
810
811 if (!pfmemalloc_active)
812 return;
813
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000814 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900815 list_for_each_entry(page, &n->slabs_full, lru)
816 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700817 goto out;
818
Joonsoo Kim8456a642013-10-24 10:07:49 +0900819 list_for_each_entry(page, &n->slabs_partial, lru)
820 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700821 goto out;
822
Joonsoo Kim8456a642013-10-24 10:07:49 +0900823 list_for_each_entry(page, &n->slabs_free, lru)
824 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700825 goto out;
826
827 pfmemalloc_active = false;
828out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000829 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700830}
831
Mel Gorman381760e2012-07-31 16:44:30 -0700832static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700833 gfp_t flags, bool force_refill)
834{
835 int i;
836 void *objp = ac->entry[--ac->avail];
837
838 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
839 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000840 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700841
842 if (gfp_pfmemalloc_allowed(flags)) {
843 clear_obj_pfmemalloc(&objp);
844 return objp;
845 }
846
847 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700848 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700849 /* If a !PFMEMALLOC object is found, swap them */
850 if (!is_obj_pfmemalloc(ac->entry[i])) {
851 objp = ac->entry[i];
852 ac->entry[i] = ac->entry[ac->avail];
853 ac->entry[ac->avail] = objp;
854 return objp;
855 }
856 }
857
858 /*
859 * If there are empty slabs on the slabs_free list and we are
860 * being forced to refill the cache, mark this one !pfmemalloc.
861 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000862 n = cachep->node[numa_mem_id()];
863 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900864 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900865 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700866 clear_obj_pfmemalloc(&objp);
867 recheck_pfmemalloc_active(cachep, ac);
868 return objp;
869 }
870
871 /* No !PFMEMALLOC objects available */
872 ac->avail++;
873 objp = NULL;
874 }
875
876 return objp;
877}
878
Mel Gorman381760e2012-07-31 16:44:30 -0700879static inline void *ac_get_obj(struct kmem_cache *cachep,
880 struct array_cache *ac, gfp_t flags, bool force_refill)
881{
882 void *objp;
883
884 if (unlikely(sk_memalloc_socks()))
885 objp = __ac_get_obj(cachep, ac, flags, force_refill);
886 else
887 objp = ac->entry[--ac->avail];
888
889 return objp;
890}
891
892static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700893 void *objp)
894{
895 if (unlikely(pfmemalloc_active)) {
896 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700897 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700898 if (PageSlabPfmemalloc(page))
899 set_obj_pfmemalloc(&objp);
900 }
901
Mel Gorman381760e2012-07-31 16:44:30 -0700902 return objp;
903}
904
905static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
906 void *objp)
907{
908 if (unlikely(sk_memalloc_socks()))
909 objp = __ac_put_obj(cachep, ac, objp);
910
Mel Gorman072bb0a2012-07-31 16:43:58 -0700911 ac->entry[ac->avail++] = objp;
912}
913
Christoph Lameter3ded1752006-03-25 03:06:44 -0800914/*
915 * Transfer objects in one arraycache to another.
916 * Locking must be handled by the caller.
917 *
918 * Return the number of entries transferred.
919 */
920static int transfer_objects(struct array_cache *to,
921 struct array_cache *from, unsigned int max)
922{
923 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700924 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800925
926 if (!nr)
927 return 0;
928
929 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
930 sizeof(void *) *nr);
931
932 from->avail -= nr;
933 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800934 return nr;
935}
936
Christoph Lameter765c4502006-09-27 01:50:08 -0700937#ifndef CONFIG_NUMA
938
939#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000940#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700941
Pekka Enberg83b519e2009-06-10 19:40:04 +0300942static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700943{
944 return (struct array_cache **)BAD_ALIEN_MAGIC;
945}
946
947static inline void free_alien_cache(struct array_cache **ac_ptr)
948{
949}
950
951static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
952{
953 return 0;
954}
955
956static inline void *alternate_node_alloc(struct kmem_cache *cachep,
957 gfp_t flags)
958{
959 return NULL;
960}
961
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800962static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700963 gfp_t flags, int nodeid)
964{
965 return NULL;
966}
967
968#else /* CONFIG_NUMA */
969
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800970static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800971static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800972
Pekka Enberg83b519e2009-06-10 19:40:04 +0300973static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700974{
975 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800976 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700977 int i;
978
979 if (limit > 1)
980 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800981 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700982 if (ac_ptr) {
983 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800984 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700985 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300986 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700987 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800988 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700989 kfree(ac_ptr[i]);
990 kfree(ac_ptr);
991 return NULL;
992 }
993 }
994 }
995 return ac_ptr;
996}
997
Pekka Enberg5295a742006-02-01 03:05:48 -0800998static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700999{
1000 int i;
1001
1002 if (!ac_ptr)
1003 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001004 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001005 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001006 kfree(ac_ptr);
1007}
1008
Pekka Enberg343e0d72006-02-01 03:05:50 -08001009static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001010 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001011{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001012 struct kmem_cache_node *n = cachep->node[node];
Christoph Lametere498be72005-09-09 13:03:32 -07001013
1014 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001015 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001016 /*
1017 * Stuff objects into the remote nodes shared array first.
1018 * That way we could avoid the overhead of putting the objects
1019 * into the free lists and getting them back later.
1020 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001021 if (n->shared)
1022 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001023
Christoph Lameterff694162005-09-22 21:44:02 -07001024 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001025 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001026 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001027 }
1028}
1029
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001030/*
1031 * Called from cache_reap() to regularly drain alien caches round robin.
1032 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001033static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001034{
Christoph Lameter909ea962010-12-08 16:22:55 +01001035 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001036
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001037 if (n->alien) {
1038 struct array_cache *ac = n->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001039
1040 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001041 __drain_alien_cache(cachep, ac, node);
1042 spin_unlock_irq(&ac->lock);
1043 }
1044 }
1045}
1046
Andrew Mortona737b3e2006-03-22 00:08:11 -08001047static void drain_alien_cache(struct kmem_cache *cachep,
1048 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001049{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001050 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001051 struct array_cache *ac;
1052 unsigned long flags;
1053
1054 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001055 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001056 if (ac) {
1057 spin_lock_irqsave(&ac->lock, flags);
1058 __drain_alien_cache(cachep, ac, i);
1059 spin_unlock_irqrestore(&ac->lock, flags);
1060 }
1061 }
1062}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001063
Ingo Molnar873623d2006-07-13 14:44:38 +02001064static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001065{
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001066 int nodeid = page_to_nid(virt_to_page(objp));
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001067 struct kmem_cache_node *n;
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001068 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001069 int node;
1070
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001071 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001072
1073 /*
1074 * Make sure we are not freeing a object from another node to the array
1075 * cache on this cpu.
1076 */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001077 if (likely(nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001078 return 0;
1079
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001080 n = cachep->node[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001081 STATS_INC_NODEFREES(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001082 if (n->alien && n->alien[nodeid]) {
1083 alien = n->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001084 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001085 if (unlikely(alien->avail == alien->limit)) {
1086 STATS_INC_ACOVERFLOW(cachep);
1087 __drain_alien_cache(cachep, alien, nodeid);
1088 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001089 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001090 spin_unlock(&alien->lock);
1091 } else {
Christoph Lameter6a673682013-01-10 19:14:19 +00001092 spin_lock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001093 free_block(cachep, &objp, 1, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001094 spin_unlock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001095 }
1096 return 1;
1097}
Christoph Lametere498be72005-09-09 13:03:32 -07001098#endif
1099
David Rientjes8f9f8d92010-03-27 19:40:47 -07001100/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001101 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001102 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001103 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001104 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001105 * already in use.
1106 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001107 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001108 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001109static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001110{
1111 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001112 struct kmem_cache_node *n;
Christoph Lameter6744f0872013-01-10 19:12:17 +00001113 const int memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001114
Christoph Lameter18004c52012-07-06 15:25:12 -05001115 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001116 /*
1117 * Set up the size64 kmemlist for cpu before we can
1118 * begin anything. Make sure some other cpu on this
1119 * node has not already allocated this
1120 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001121 if (!cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001122 n = kmalloc_node(memsize, GFP_KERNEL, node);
1123 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001124 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 kmem_cache_node_init(n);
1126 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
David Rientjes8f9f8d92010-03-27 19:40:47 -07001127 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1128
1129 /*
1130 * The l3s don't come and go as CPUs come and
Christoph Lameter18004c52012-07-06 15:25:12 -05001131 * go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001132 * protection here.
1133 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001134 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001135 }
1136
Christoph Lameter6a673682013-01-10 19:14:19 +00001137 spin_lock_irq(&cachep->node[node]->list_lock);
1138 cachep->node[node]->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001139 (1 + nr_cpus_node(node)) *
1140 cachep->batchcount + cachep->num;
Christoph Lameter6a673682013-01-10 19:14:19 +00001141 spin_unlock_irq(&cachep->node[node]->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001142 }
1143 return 0;
1144}
1145
Wanpeng Li0fa81032013-07-04 08:33:22 +08001146static inline int slabs_tofree(struct kmem_cache *cachep,
1147 struct kmem_cache_node *n)
1148{
1149 return (n->free_objects + cachep->num - 1) / cachep->num;
1150}
1151
Paul Gortmaker0db06282013-06-19 14:53:51 -04001152static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001154 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001155 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001156 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301157 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001158
Christoph Lameter18004c52012-07-06 15:25:12 -05001159 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001160 struct array_cache *nc;
1161 struct array_cache *shared;
1162 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001163
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001164 /* cpu is dead; no one can alloc from it. */
1165 nc = cachep->array[cpu];
1166 cachep->array[cpu] = NULL;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001167 n = cachep->node[node];
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001168
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001169 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001170 goto free_array_cache;
1171
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001172 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001173
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001174 /* Free limit for this kmem_cache_node */
1175 n->free_limit -= cachep->batchcount;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001176 if (nc)
1177 free_block(cachep, nc->entry, nc->avail, node);
1178
Rusty Russell58463c12009-12-17 11:43:12 -06001179 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001180 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001181 goto free_array_cache;
1182 }
1183
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001184 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 if (shared) {
1186 free_block(cachep, shared->entry,
1187 shared->avail, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001188 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001189 }
1190
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001191 alien = n->alien;
1192 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001193
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001194 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001195
1196 kfree(shared);
1197 if (alien) {
1198 drain_alien_cache(cachep, alien);
1199 free_alien_cache(alien);
1200 }
1201free_array_cache:
1202 kfree(nc);
1203 }
1204 /*
1205 * In the previous loop, all the objects were freed to
1206 * the respective cache's slabs, now we can go ahead and
1207 * shrink each nodelist to its limit.
1208 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001209 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001210 n = cachep->node[node];
1211 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001212 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001213 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001214 }
1215}
1216
Paul Gortmaker0db06282013-06-19 14:53:51 -04001217static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001218{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001219 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001220 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001221 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001222 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001224 /*
1225 * We need to do this right in the beginning since
1226 * alloc_arraycache's are going to use this list.
1227 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001228 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001229 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001230 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001231 if (err < 0)
1232 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001233
1234 /*
1235 * Now we can go ahead with allocating the shared arrays and
1236 * array caches
1237 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001238 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001239 struct array_cache *nc;
1240 struct array_cache *shared = NULL;
1241 struct array_cache **alien = NULL;
1242
1243 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001244 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001245 if (!nc)
1246 goto bad;
1247 if (cachep->shared) {
1248 shared = alloc_arraycache(node,
1249 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001250 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001251 if (!shared) {
1252 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001253 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001254 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001255 }
1256 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001257 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001258 if (!alien) {
1259 kfree(shared);
1260 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001261 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001262 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001263 }
1264 cachep->array[cpu] = nc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001265 n = cachep->node[node];
1266 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001267
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001268 spin_lock_irq(&n->list_lock);
1269 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001270 /*
1271 * We are serialised from CPU_DEAD or
1272 * CPU_UP_CANCELLED by the cpucontrol lock
1273 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001274 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001275 shared = NULL;
1276 }
1277#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001278 if (!n->alien) {
1279 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 alien = NULL;
1281 }
1282#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001283 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001284 kfree(shared);
1285 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001286 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1287 slab_set_debugobj_lock_classes_node(cachep, node);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08001288 else if (!OFF_SLAB(cachep) &&
1289 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1290 on_slab_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001291 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001292 init_node_lock_keys(node);
1293
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001294 return 0;
1295bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001296 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001297 return -ENOMEM;
1298}
1299
Paul Gortmaker0db06282013-06-19 14:53:51 -04001300static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001301 unsigned long action, void *hcpu)
1302{
1303 long cpu = (long)hcpu;
1304 int err = 0;
1305
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001307 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001308 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001309 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001310 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001311 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 break;
1313 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001314 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 start_cpu_timer(cpu);
1316 break;
1317#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001318 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001319 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001320 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001321 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001322 * held so that if cache_reap() is invoked it cannot do
1323 * anything expensive but will only modify reap_work
1324 * and reschedule the timer.
1325 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001326 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001327 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001328 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001329 break;
1330 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001331 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001332 start_cpu_timer(cpu);
1333 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001335 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001336 /*
1337 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001338 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001339 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001340 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001341 * structure is usually allocated from kmem_cache_create() and
1342 * gets destroyed at kmem_cache_destroy().
1343 */
Simon Arlott183ff222007-10-20 01:27:18 +02001344 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001345#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001347 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001348 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001349 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001350 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001353 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354}
1355
Paul Gortmaker0db06282013-06-19 14:53:51 -04001356static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001357 &cpuup_callback, NULL, 0
1358};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
David Rientjes8f9f8d92010-03-27 19:40:47 -07001360#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1361/*
1362 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1363 * Returns -EBUSY if all objects cannot be drained so that the node is not
1364 * removed.
1365 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001366 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001367 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001368static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001369{
1370 struct kmem_cache *cachep;
1371 int ret = 0;
1372
Christoph Lameter18004c52012-07-06 15:25:12 -05001373 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001374 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001375
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001376 n = cachep->node[node];
1377 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001378 continue;
1379
Wanpeng Li0fa81032013-07-04 08:33:22 +08001380 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001381
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001382 if (!list_empty(&n->slabs_full) ||
1383 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001384 ret = -EBUSY;
1385 break;
1386 }
1387 }
1388 return ret;
1389}
1390
1391static int __meminit slab_memory_callback(struct notifier_block *self,
1392 unsigned long action, void *arg)
1393{
1394 struct memory_notify *mnb = arg;
1395 int ret = 0;
1396 int nid;
1397
1398 nid = mnb->status_change_nid;
1399 if (nid < 0)
1400 goto out;
1401
1402 switch (action) {
1403 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001404 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001405 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001406 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001407 break;
1408 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001409 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001410 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001411 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001412 break;
1413 case MEM_ONLINE:
1414 case MEM_OFFLINE:
1415 case MEM_CANCEL_ONLINE:
1416 case MEM_CANCEL_OFFLINE:
1417 break;
1418 }
1419out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001420 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001421}
1422#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1423
Christoph Lametere498be72005-09-09 13:03:32 -07001424/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001425 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001426 */
Christoph Lameter6744f0872013-01-10 19:12:17 +00001427static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001428 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001429{
Christoph Lameter6744f0872013-01-10 19:12:17 +00001430 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001431
Christoph Lameter6744f0872013-01-10 19:12:17 +00001432 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001433 BUG_ON(!ptr);
1434
Christoph Lameter6744f0872013-01-10 19:12:17 +00001435 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001436 /*
1437 * Do not assume that spinlocks can be initialized via memcpy:
1438 */
1439 spin_lock_init(&ptr->list_lock);
1440
Christoph Lametere498be72005-09-09 13:03:32 -07001441 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001442 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001443}
1444
Andrew Mortona737b3e2006-03-22 00:08:11 -08001445/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001446 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1447 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001448 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001449static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001450{
1451 int node;
1452
1453 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001454 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001455 cachep->node[node]->next_reap = jiffies +
Pekka Enberg556a1692008-01-25 08:20:51 +02001456 REAPTIMEOUT_LIST3 +
1457 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1458 }
1459}
1460
1461/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001462 * The memory after the last cpu cache pointer is used for the
Christoph Lameter6a673682013-01-10 19:14:19 +00001463 * the node pointer.
Christoph Lameter3c583462012-11-28 16:23:01 +00001464 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001465static void setup_node_pointer(struct kmem_cache *cachep)
Christoph Lameter3c583462012-11-28 16:23:01 +00001466{
Christoph Lameter6a673682013-01-10 19:14:19 +00001467 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
Christoph Lameter3c583462012-11-28 16:23:01 +00001468}
1469
1470/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001471 * Initialisation. Called after the page allocator have been initialised and
1472 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 */
1474void __init kmem_cache_init(void)
1475{
Christoph Lametere498be72005-09-09 13:03:32 -07001476 int i;
1477
Joonsoo Kim68126702013-10-24 10:07:42 +09001478 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1479 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001480 kmem_cache = &kmem_cache_boot;
Christoph Lameter6a673682013-01-10 19:14:19 +00001481 setup_node_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001482
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001483 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001484 use_alien_caches = 0;
1485
Christoph Lameter3c583462012-11-28 16:23:01 +00001486 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001487 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001488
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001489 set_up_node(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490
1491 /*
1492 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001493 * page orders on machines with more than 32MB of memory if
1494 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001496 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001497 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 /* Bootstrap is tricky, because several objects are allocated
1500 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001501 * 1) initialize the kmem_cache cache: it contains the struct
1502 * kmem_cache structures of all caches, except kmem_cache itself:
1503 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001504 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001505 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001506 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001508 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001509 * An __init data area is used for the head array.
1510 * 3) Create the remaining kmalloc caches, with minimally sized
1511 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001512 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001514 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001515 * the other cache's with kmalloc allocated memory.
1516 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 */
1518
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001519 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520
Eric Dumazet8da34302007-05-06 14:49:29 -07001521 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001522 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001523 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001524 create_boot_cache(kmem_cache, "kmem_cache",
1525 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Christoph Lameter6744f0872013-01-10 19:12:17 +00001526 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001527 SLAB_HWCACHE_ALIGN);
1528 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529
1530 /* 2+3) create the kmalloc caches */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531
Andrew Mortona737b3e2006-03-22 00:08:11 -08001532 /*
1533 * Initialize the caches that provide memory for the array cache and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001534 * kmem_cache_node structures first. Without this, further allocations will
Andrew Mortona737b3e2006-03-22 00:08:11 -08001535 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001536 */
1537
Christoph Lametere3366012013-01-10 19:14:18 +00001538 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1539 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001540
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001541 if (INDEX_AC != INDEX_NODE)
1542 kmalloc_caches[INDEX_NODE] =
1543 create_kmalloc_cache("kmalloc-node",
1544 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001545
Ingo Molnare0a42722006-06-23 02:03:46 -07001546 slab_early_init = 0;
1547
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 /* 4) Replace the bootstrap head arrays */
1549 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001550 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001551
Pekka Enberg83b519e2009-06-10 19:40:04 +03001552 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001553
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001554 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001555 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001556 /*
1557 * Do not assume that spinlocks can be initialized via memcpy:
1558 */
1559 spin_lock_init(&ptr->lock);
1560
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001561 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001562
Pekka Enberg83b519e2009-06-10 19:40:04 +03001563 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001564
Christoph Lametere3366012013-01-10 19:14:18 +00001565 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001566 != &initarray_generic.cache);
Christoph Lametere3366012013-01-10 19:14:18 +00001567 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001568 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001569 /*
1570 * Do not assume that spinlocks can be initialized via memcpy:
1571 */
1572 spin_lock_init(&ptr->lock);
1573
Christoph Lametere3366012013-01-10 19:14:18 +00001574 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001576 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001577 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001578 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579
Mel Gorman9c09a952008-01-24 05:49:54 -08001580 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001581 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001582
Christoph Lametere3366012013-01-10 19:14:18 +00001583 init_list(kmalloc_caches[INDEX_AC],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001584 &init_kmem_cache_node[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001585
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001586 if (INDEX_AC != INDEX_NODE) {
1587 init_list(kmalloc_caches[INDEX_NODE],
1588 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001589 }
1590 }
1591 }
1592
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001593 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001594}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001595
Pekka Enberg8429db52009-06-12 15:58:59 +03001596void __init kmem_cache_init_late(void)
1597{
1598 struct kmem_cache *cachep;
1599
Christoph Lameter97d06602012-07-06 15:25:11 -05001600 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001601
Pekka Enberg8429db52009-06-12 15:58:59 +03001602 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001603 mutex_lock(&slab_mutex);
1604 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001605 if (enable_cpucache(cachep, GFP_NOWAIT))
1606 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001607 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001608
Michael Wang947ca182012-09-05 10:33:18 +08001609 /* Annotate slab for lockdep -- annotate the malloc caches */
1610 init_lock_keys();
1611
Christoph Lameter97d06602012-07-06 15:25:11 -05001612 /* Done! */
1613 slab_state = FULL;
1614
Andrew Mortona737b3e2006-03-22 00:08:11 -08001615 /*
1616 * Register a cpu startup notifier callback that initializes
1617 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 */
1619 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
David Rientjes8f9f8d92010-03-27 19:40:47 -07001621#ifdef CONFIG_NUMA
1622 /*
1623 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001624 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001625 */
1626 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1627#endif
1628
Andrew Mortona737b3e2006-03-22 00:08:11 -08001629 /*
1630 * The reap timers are started later, with a module init call: That part
1631 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 */
1633}
1634
1635static int __init cpucache_init(void)
1636{
1637 int cpu;
1638
Andrew Mortona737b3e2006-03-22 00:08:11 -08001639 /*
1640 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 */
Christoph Lametere498be72005-09-09 13:03:32 -07001642 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001643 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001644
1645 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001646 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 return 0;
1648}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649__initcall(cpucache_init);
1650
Rafael Aquini8bdec192012-03-09 17:27:27 -03001651static noinline void
1652slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1653{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001654 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001655 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001656 unsigned long flags;
1657 int node;
1658
1659 printk(KERN_WARNING
1660 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1661 nodeid, gfpflags);
1662 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001663 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001664
1665 for_each_online_node(node) {
1666 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1667 unsigned long active_slabs = 0, num_slabs = 0;
1668
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001669 n = cachep->node[node];
1670 if (!n)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001671 continue;
1672
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001673 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001674 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001675 active_objs += cachep->num;
1676 active_slabs++;
1677 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001678 list_for_each_entry(page, &n->slabs_partial, lru) {
1679 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001680 active_slabs++;
1681 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001682 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001683 num_slabs++;
1684
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001685 free_objects += n->free_objects;
1686 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001687
1688 num_slabs += active_slabs;
1689 num_objs = num_slabs * cachep->num;
1690 printk(KERN_WARNING
1691 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1692 node, active_slabs, num_slabs, active_objs, num_objs,
1693 free_objects);
1694 }
1695}
1696
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697/*
1698 * Interface to system's page allocator. No need to hold the cache-lock.
1699 *
1700 * If we requested dmaable memory, we will get it. Even if we
1701 * did not request dmaable memory, we might get it, but that
1702 * would be relatively rare and ignorable.
1703 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001704static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1705 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706{
1707 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001708 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001709
Glauber Costaa618e892012-06-14 16:17:21 +04001710 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001711 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1712 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001713
Linus Torvalds517d0862009-06-16 19:50:13 -07001714 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001715 if (!page) {
1716 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1717 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001719 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001721 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001722 if (unlikely(page->pfmemalloc))
1723 pfmemalloc_active = true;
1724
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001725 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001727 add_zone_page_state(page_zone(page),
1728 NR_SLAB_RECLAIMABLE, nr_pages);
1729 else
1730 add_zone_page_state(page_zone(page),
1731 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001732 __SetPageSlab(page);
1733 if (page->pfmemalloc)
1734 SetPageSlabPfmemalloc(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001735 memcg_bind_pages(cachep, cachep->gfporder);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001736
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001737 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1738 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1739
1740 if (cachep->ctor)
1741 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1742 else
1743 kmemcheck_mark_unallocated_pages(page, nr_pages);
1744 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001745
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001746 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747}
1748
1749/*
1750 * Interface to system's page release.
1751 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001752static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001754 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001756 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001757
Christoph Lameter972d1a72006-09-25 23:31:51 -07001758 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1759 sub_zone_page_state(page_zone(page),
1760 NR_SLAB_RECLAIMABLE, nr_freed);
1761 else
1762 sub_zone_page_state(page_zone(page),
1763 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001764
Joonsoo Kima57a4982013-10-24 10:07:44 +09001765 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001766 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001767 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001768 page_mapcount_reset(page);
1769 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001770
1771 memcg_release_pages(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 if (current->reclaim_state)
1773 current->reclaim_state->reclaimed_slab += nr_freed;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001774 __free_memcg_kmem_pages(page, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775}
1776
1777static void kmem_rcu_free(struct rcu_head *head)
1778{
Joonsoo Kim68126702013-10-24 10:07:42 +09001779 struct kmem_cache *cachep;
1780 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781
Joonsoo Kim68126702013-10-24 10:07:42 +09001782 page = container_of(head, struct page, rcu_head);
1783 cachep = page->slab_cache;
1784
1785 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786}
1787
1788#if DEBUG
1789
1790#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001791static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001792 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001794 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001796 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001798 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 return;
1800
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001801 *addr++ = 0x12345678;
1802 *addr++ = caller;
1803 *addr++ = smp_processor_id();
1804 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 {
1806 unsigned long *sptr = &caller;
1807 unsigned long svalue;
1808
1809 while (!kstack_end(sptr)) {
1810 svalue = *sptr++;
1811 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001812 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 size -= sizeof(unsigned long);
1814 if (size <= sizeof(unsigned long))
1815 break;
1816 }
1817 }
1818
1819 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001820 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821}
1822#endif
1823
Pekka Enberg343e0d72006-02-01 03:05:50 -08001824static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001826 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001827 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828
1829 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001830 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831}
1832
1833static void dump_line(char *data, int offset, int limit)
1834{
1835 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001836 unsigned char error = 0;
1837 int bad_count = 0;
1838
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001839 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001840 for (i = 0; i < limit; i++) {
1841 if (data[offset + i] != POISON_FREE) {
1842 error = data[offset + i];
1843 bad_count++;
1844 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001845 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001846 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1847 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001848
1849 if (bad_count == 1) {
1850 error ^= POISON_FREE;
1851 if (!(error & (error - 1))) {
1852 printk(KERN_ERR "Single bit error detected. Probably "
1853 "bad RAM.\n");
1854#ifdef CONFIG_X86
1855 printk(KERN_ERR "Run memtest86+ or a similar memory "
1856 "test tool.\n");
1857#else
1858 printk(KERN_ERR "Run a memory test tool.\n");
1859#endif
1860 }
1861 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862}
1863#endif
1864
1865#if DEBUG
1866
Pekka Enberg343e0d72006-02-01 03:05:50 -08001867static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868{
1869 int i, size;
1870 char *realobj;
1871
1872 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001873 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001874 *dbg_redzone1(cachep, objp),
1875 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 }
1877
1878 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001879 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1880 *dbg_userword(cachep, objp),
1881 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001883 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001884 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001885 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 int limit;
1887 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001888 if (i + limit > size)
1889 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 dump_line(realobj, i, limit);
1891 }
1892}
1893
Pekka Enberg343e0d72006-02-01 03:05:50 -08001894static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895{
1896 char *realobj;
1897 int size, i;
1898 int lines = 0;
1899
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001900 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001901 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001903 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001905 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 exp = POISON_END;
1907 if (realobj[i] != exp) {
1908 int limit;
1909 /* Mismatch ! */
1910 /* Print header */
1911 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001912 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001913 "Slab corruption (%s): %s start=%p, len=%d\n",
1914 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 print_objinfo(cachep, objp, 0);
1916 }
1917 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001918 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001920 if (i + limit > size)
1921 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 dump_line(realobj, i, limit);
1923 i += 16;
1924 lines++;
1925 /* Limit to 5 lines */
1926 if (lines > 5)
1927 break;
1928 }
1929 }
1930 if (lines != 0) {
1931 /* Print some data about the neighboring objects, if they
1932 * exist:
1933 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001934 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001935 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
Joonsoo Kim8456a642013-10-24 10:07:49 +09001937 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001939 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001940 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001942 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 print_objinfo(cachep, objp, 2);
1944 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001945 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001946 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001947 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001949 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 print_objinfo(cachep, objp, 2);
1951 }
1952 }
1953}
1954#endif
1955
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001957static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1958 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001959{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 int i;
1961 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001962 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963
1964 if (cachep->flags & SLAB_POISON) {
1965#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001966 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001967 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001968 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001969 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 else
1971 check_poison_obj(cachep, objp);
1972#else
1973 check_poison_obj(cachep, objp);
1974#endif
1975 }
1976 if (cachep->flags & SLAB_RED_ZONE) {
1977 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1978 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001979 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1981 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001982 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001985}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001987static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1988 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001989{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001990}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991#endif
1992
Randy Dunlap911851e2006-03-22 00:08:14 -08001993/**
1994 * slab_destroy - destroy and release all objects in a slab
1995 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001996 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001997 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001998 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001999 * Before calling the slab must have been unlinked from the cache. The
2000 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002001 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002002static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002003{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002004 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002005
Joonsoo Kim8456a642013-10-24 10:07:49 +09002006 freelist = page->freelist;
2007 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09002009 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010
Joonsoo Kim68126702013-10-24 10:07:42 +09002011 /*
2012 * RCU free overloads the RCU head over the LRU.
2013 * slab_page has been overloeaded over the LRU,
2014 * however it is not used from now on so that
2015 * we can use it safely.
2016 */
2017 head = (void *)&page->rcu_head;
2018 call_rcu(head, kmem_rcu_free);
2019
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002021 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 }
Joonsoo Kim68126702013-10-24 10:07:42 +09002023
2024 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09002025 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09002026 * although actual page can be freed in rcu context
2027 */
2028 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09002029 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030}
2031
2032/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002033 * calculate_slab_order - calculate size (page order) of slabs
2034 * @cachep: pointer to the cache that is being created
2035 * @size: size of objects to be created in this cache.
2036 * @align: required alignment for the objects.
2037 * @flags: slab allocation flags
2038 *
2039 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002040 *
2041 * This could be made much more intelligent. For now, try to avoid using
2042 * high order pages for slabs. When the gfp() functions are more friendly
2043 * towards high-order requests, this should be changed.
2044 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002045static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002046 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002047{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002048 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002049 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002050 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002051
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002052 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002053 unsigned int num;
2054 size_t remainder;
2055
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002056 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002057 if (!num)
2058 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002059
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002060 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002061 size_t freelist_size_per_obj = sizeof(unsigned int);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002062 /*
2063 * Max number of objs-per-slab for caches which
2064 * use off-slab slabs. Needed to avoid a possible
2065 * looping condition in cache_grow().
2066 */
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002067 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
2068 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002069 offslab_limit = size;
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002070 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002071
2072 if (num > offslab_limit)
2073 break;
2074 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002075
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002076 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002077 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002078 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002079 left_over = remainder;
2080
2081 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002082 * A VFS-reclaimable slab tends to have most allocations
2083 * as GFP_NOFS and we really don't want to have to be allocating
2084 * higher-order pages when we are unable to shrink dcache.
2085 */
2086 if (flags & SLAB_RECLAIM_ACCOUNT)
2087 break;
2088
2089 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002090 * Large number of objects is good, but very large slabs are
2091 * currently bad for the gfp()s.
2092 */
David Rientjes543585c2011-10-18 22:09:24 -07002093 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002094 break;
2095
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002096 /*
2097 * Acceptable internal fragmentation?
2098 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002099 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002100 break;
2101 }
2102 return left_over;
2103}
2104
Pekka Enberg83b519e2009-06-10 19:40:04 +03002105static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002106{
Christoph Lameter97d06602012-07-06 15:25:11 -05002107 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002108 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002109
Christoph Lameter97d06602012-07-06 15:25:11 -05002110 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002111 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002112 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002113 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002114 * of by the caller of __kmem_cache_create
2115 */
2116 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2117 slab_state = PARTIAL;
2118 } else if (slab_state == PARTIAL) {
2119 /*
2120 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002121 * that's used by kmalloc(24), otherwise the creation of
2122 * further caches will BUG().
2123 */
2124 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2125
2126 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002127 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2128 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002129 * otherwise the creation of further caches will BUG().
2130 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002131 set_up_node(cachep, SIZE_AC);
2132 if (INDEX_AC == INDEX_NODE)
2133 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002134 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002135 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002136 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002137 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002138 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002139 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002140
Christoph Lameter97d06602012-07-06 15:25:11 -05002141 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002142 set_up_node(cachep, SIZE_NODE);
2143 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002144 } else {
2145 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002146 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002147 cachep->node[node] =
Christoph Lameter6744f0872013-01-10 19:12:17 +00002148 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002149 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002150 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002151 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002152 }
2153 }
2154 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002155 cachep->node[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002156 jiffies + REAPTIMEOUT_LIST3 +
2157 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2158
2159 cpu_cache_get(cachep)->avail = 0;
2160 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2161 cpu_cache_get(cachep)->batchcount = 1;
2162 cpu_cache_get(cachep)->touched = 0;
2163 cachep->batchcount = 1;
2164 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002165 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002166}
2167
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002168/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002169 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002170 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 *
2173 * Returns a ptr to the cache on success, NULL on failure.
2174 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002175 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 * The flags are
2178 *
2179 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2180 * to catch references to uninitialised memory.
2181 *
2182 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2183 * for buffer overruns.
2184 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2186 * cacheline. This can be beneficial if you're counting cycles as closely
2187 * as davem.
2188 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002189int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002190__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002192 size_t left_over, freelist_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002193 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002194 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002195 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198#if FORCED_DEBUG
2199 /*
2200 * Enable redzoning and last user accounting, except for caches with
2201 * large objects, if the increased size would increase the object size
2202 * above the next power of two: caches with object sizes just above a
2203 * power of two have a significant amount of internal fragmentation.
2204 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002205 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2206 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002207 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 if (!(flags & SLAB_DESTROY_BY_RCU))
2209 flags |= SLAB_POISON;
2210#endif
2211 if (flags & SLAB_DESTROY_BY_RCU)
2212 BUG_ON(flags & SLAB_POISON);
2213#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214
Andrew Mortona737b3e2006-03-22 00:08:11 -08002215 /*
2216 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 * unaligned accesses for some archs when redzoning is used, and makes
2218 * sure any on-slab bufctl's are also correctly aligned.
2219 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002220 if (size & (BYTES_PER_WORD - 1)) {
2221 size += (BYTES_PER_WORD - 1);
2222 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 }
2224
Pekka Enbergca5f9702006-09-25 23:31:25 -07002225 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002226 * Redzoning and user store require word alignment or possibly larger.
2227 * Note this will be overridden by architecture or caller mandated
2228 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002229 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002230 if (flags & SLAB_STORE_USER)
2231 ralign = BYTES_PER_WORD;
2232
2233 if (flags & SLAB_RED_ZONE) {
2234 ralign = REDZONE_ALIGN;
2235 /* If redzoning, ensure that the second redzone is suitably
2236 * aligned, by adjusting the object size accordingly. */
2237 size += REDZONE_ALIGN - 1;
2238 size &= ~(REDZONE_ALIGN - 1);
2239 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002240
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002241 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002242 if (ralign < cachep->align) {
2243 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002245 /* disable debug if necessary */
2246 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002247 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002248 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002249 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002251 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252
Pekka Enberg83b519e2009-06-10 19:40:04 +03002253 if (slab_is_available())
2254 gfp = GFP_KERNEL;
2255 else
2256 gfp = GFP_NOWAIT;
2257
Christoph Lameter6a673682013-01-10 19:14:19 +00002258 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260
Pekka Enbergca5f9702006-09-25 23:31:25 -07002261 /*
2262 * Both debugging options require word-alignment which is calculated
2263 * into align above.
2264 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002267 cachep->obj_offset += sizeof(unsigned long long);
2268 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 }
2270 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002271 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002272 * the real object. But if the second red zone needs to be
2273 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002275 if (flags & SLAB_RED_ZONE)
2276 size += REDZONE_ALIGN;
2277 else
2278 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 }
2280#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002281 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002282 && cachep->object_size > cache_line_size()
2283 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2284 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 size = PAGE_SIZE;
2286 }
2287#endif
2288#endif
2289
Ingo Molnare0a42722006-06-23 02:03:46 -07002290 /*
2291 * Determine if the slab management is 'on' or 'off' slab.
2292 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002293 * it too early on. Always use on-slab management when
2294 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002295 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002296 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2297 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 /*
2299 * Size is large, assume best to place the slab management obj
2300 * off-slab (should allow better packing of objs).
2301 */
2302 flags |= CFLGS_OFF_SLAB;
2303
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002304 size = ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002306 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002308 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002309 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002310
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002311 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312
2313 /*
2314 * If the slab has been placed off-slab, and we have enough space then
2315 * move it on-slab. This is at the expense of any extra colouring.
2316 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002317 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002319 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 }
2321
2322 if (flags & CFLGS_OFF_SLAB) {
2323 /* really off slab. No need for manual alignment */
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002324 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302325
2326#ifdef CONFIG_PAGE_POISONING
2327 /* If we're going to use the generic kernel_map_pages()
2328 * poisoning, then it's going to smash the contents of
2329 * the redzone and userword anyhow, so switch them off.
2330 */
2331 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2332 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2333#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 }
2335
2336 cachep->colour_off = cache_line_size();
2337 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002338 if (cachep->colour_off < cachep->align)
2339 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002340 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002341 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002343 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002344 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002345 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002346 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002347 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002349 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002350 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002351 /*
2352 * This is a possibility for one of the malloc_sizes caches.
2353 * But since we go off slab only for object size greater than
2354 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2355 * this should not happen at all.
2356 * But leave a BUG_ON for some lucky dude.
2357 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002358 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002359 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002361 err = setup_cpu_cache(cachep, gfp);
2362 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002363 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002364 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002365 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366
Peter Zijlstra83835b32011-07-22 15:26:05 +02002367 if (flags & SLAB_DEBUG_OBJECTS) {
2368 /*
2369 * Would deadlock through slab_destroy()->call_rcu()->
2370 * debug_object_activate()->kmem_cache_alloc().
2371 */
2372 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2373
2374 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002375 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2376 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002377
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002378 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380
2381#if DEBUG
2382static void check_irq_off(void)
2383{
2384 BUG_ON(!irqs_disabled());
2385}
2386
2387static void check_irq_on(void)
2388{
2389 BUG_ON(irqs_disabled());
2390}
2391
Pekka Enberg343e0d72006-02-01 03:05:50 -08002392static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393{
2394#ifdef CONFIG_SMP
2395 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002396 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397#endif
2398}
Christoph Lametere498be72005-09-09 13:03:32 -07002399
Pekka Enberg343e0d72006-02-01 03:05:50 -08002400static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002401{
2402#ifdef CONFIG_SMP
2403 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002404 assert_spin_locked(&cachep->node[node]->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002405#endif
2406}
2407
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408#else
2409#define check_irq_off() do { } while(0)
2410#define check_irq_on() do { } while(0)
2411#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002412#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413#endif
2414
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002415static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002416 struct array_cache *ac,
2417 int force, int node);
2418
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419static void do_drain(void *arg)
2420{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002421 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002423 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424
2425 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002426 ac = cpu_cache_get(cachep);
Christoph Lameter6a673682013-01-10 19:14:19 +00002427 spin_lock(&cachep->node[node]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002428 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002429 spin_unlock(&cachep->node[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 ac->avail = 0;
2431}
2432
Pekka Enberg343e0d72006-02-01 03:05:50 -08002433static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002435 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002436 int node;
2437
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002438 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002440 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002441 n = cachep->node[node];
2442 if (n && n->alien)
2443 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002444 }
2445
2446 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002447 n = cachep->node[node];
2448 if (n)
2449 drain_array(cachep, n, n->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002450 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451}
2452
Christoph Lametered11d9e2006-06-30 01:55:45 -07002453/*
2454 * Remove slabs from the list of free slabs.
2455 * Specify the number of slabs to drain in tofree.
2456 *
2457 * Returns the actual number of slabs released.
2458 */
2459static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002460 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002462 struct list_head *p;
2463 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002464 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
Christoph Lametered11d9e2006-06-30 01:55:45 -07002466 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002467 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002469 spin_lock_irq(&n->list_lock);
2470 p = n->slabs_free.prev;
2471 if (p == &n->slabs_free) {
2472 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002473 goto out;
2474 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Joonsoo Kim8456a642013-10-24 10:07:49 +09002476 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002478 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002480 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002481 /*
2482 * Safe to drop the lock. The slab is no longer linked
2483 * to the cache.
2484 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002485 n->free_objects -= cache->num;
2486 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002487 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002488 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002490out:
2491 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492}
2493
Christoph Lameter18004c52012-07-06 15:25:12 -05002494/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002495static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002496{
2497 int ret = 0, i = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002498 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002499
2500 drain_cpu_caches(cachep);
2501
2502 check_irq_on();
2503 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002504 n = cachep->node[i];
2505 if (!n)
Christoph Lametered11d9e2006-06-30 01:55:45 -07002506 continue;
2507
Wanpeng Li0fa81032013-07-04 08:33:22 +08002508 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002509
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002510 ret += !list_empty(&n->slabs_full) ||
2511 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002512 }
2513 return (ret ? 1 : 0);
2514}
2515
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516/**
2517 * kmem_cache_shrink - Shrink a cache.
2518 * @cachep: The cache to shrink.
2519 *
2520 * Releases as many slabs as possible for a cache.
2521 * To help debugging, a zero exit status indicates all slabs were released.
2522 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002523int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002525 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002526 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002528 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002529 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002530 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002531 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002532 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002533 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534}
2535EXPORT_SYMBOL(kmem_cache_shrink);
2536
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002537int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538{
Christoph Lameter12c36672012-09-04 23:38:33 +00002539 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002540 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002541 int rc = __cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542
Christoph Lameter12c36672012-09-04 23:38:33 +00002543 if (rc)
2544 return rc;
2545
2546 for_each_online_cpu(i)
2547 kfree(cachep->array[i]);
2548
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002549 /* NUMA: free the node structures */
Christoph Lameter12c36672012-09-04 23:38:33 +00002550 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002551 n = cachep->node[i];
2552 if (n) {
2553 kfree(n->shared);
2554 free_alien_cache(n->alien);
2555 kfree(n);
Christoph Lameter12c36672012-09-04 23:38:33 +00002556 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002558 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002561/*
2562 * Get the memory for a slab management obj.
2563 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2564 * always come from malloc_sizes caches. The slab descriptor cannot
2565 * come from the same cache which is getting created because,
2566 * when we are searching for an appropriate cache for these
2567 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2568 * If we are creating a malloc_sizes cache here it would not be visible to
2569 * kmem_find_general_cachep till the initialization is complete.
Joonsoo Kim8456a642013-10-24 10:07:49 +09002570 * Hence we cannot have freelist_cache same as the original cache.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002571 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002572static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002573 struct page *page, int colour_off,
2574 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002576 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002577 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002578
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 if (OFF_SLAB(cachep)) {
2580 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002581 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002582 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002583 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 return NULL;
2585 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002586 freelist = addr + colour_off;
2587 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002589 page->active = 0;
2590 page->s_mem = addr + colour_off;
2591 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592}
2593
Joonsoo Kime7444d92013-10-24 10:07:51 +09002594static inline unsigned int *slab_freelist(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002596 return (unsigned int *)(page->freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597}
2598
Pekka Enberg343e0d72006-02-01 03:05:50 -08002599static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002600 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601{
2602 int i;
2603
2604 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002605 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606#if DEBUG
2607 /* need to poison the objs? */
2608 if (cachep->flags & SLAB_POISON)
2609 poison_obj(cachep, objp, POISON_FREE);
2610 if (cachep->flags & SLAB_STORE_USER)
2611 *dbg_userword(cachep, objp) = NULL;
2612
2613 if (cachep->flags & SLAB_RED_ZONE) {
2614 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2615 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2616 }
2617 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002618 * Constructors are not allowed to allocate memory from the same
2619 * cache which they are a constructor for. Otherwise, deadlock.
2620 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 */
2622 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002623 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624
2625 if (cachep->flags & SLAB_RED_ZONE) {
2626 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2627 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002628 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2630 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002631 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002633 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002634 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002635 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002636 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637#else
2638 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002639 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640#endif
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002641 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime7444d92013-10-24 10:07:51 +09002642 slab_freelist(page)[i] = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644}
2645
Pekka Enberg343e0d72006-02-01 03:05:50 -08002646static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002648 if (CONFIG_ZONE_DMA_FLAG) {
2649 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002650 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002651 else
Glauber Costaa618e892012-06-14 16:17:21 +04002652 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654}
2655
Joonsoo Kim8456a642013-10-24 10:07:49 +09002656static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002657 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002658{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002659 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002660
Joonsoo Kime7444d92013-10-24 10:07:51 +09002661 objp = index_to_obj(cachep, page, slab_freelist(page)[page->active]);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002662 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002663#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002664 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002665#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002666
2667 return objp;
2668}
2669
Joonsoo Kim8456a642013-10-24 10:07:49 +09002670static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002671 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002672{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002673 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002674#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002675 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002676
Matthew Dobson78d382d2006-02-01 03:05:47 -08002677 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002678 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002679
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002680 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002681 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime7444d92013-10-24 10:07:51 +09002682 if (slab_freelist(page)[i] == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002683 printk(KERN_ERR "slab: double free detected in cache "
2684 "'%s', objp %p\n", cachep->name, objp);
2685 BUG();
2686 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002687 }
2688#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002689 page->active--;
Joonsoo Kime7444d92013-10-24 10:07:51 +09002690 slab_freelist(page)[page->active] = objnr;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002691}
2692
Pekka Enberg47768742006-06-23 02:03:07 -07002693/*
2694 * Map pages beginning at addr to the given cache and slab. This is required
2695 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002696 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002697 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002698static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002699 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002701 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002702 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703}
2704
2705/*
2706 * Grow (by 1) the number of slabs within a cache. This is called by
2707 * kmem_cache_alloc() when there are no active objs left in a cache.
2708 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002709static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002710 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002712 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002713 size_t offset;
2714 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002715 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716
Andrew Mortona737b3e2006-03-22 00:08:11 -08002717 /*
2718 * Be lazy and only check for valid flags here, keeping it out of the
2719 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002721 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2722 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002724 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002726 n = cachep->node[nodeid];
2727 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728
2729 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002730 offset = n->colour_next;
2731 n->colour_next++;
2732 if (n->colour_next >= cachep->colour)
2733 n->colour_next = 0;
2734 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002736 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737
2738 if (local_flags & __GFP_WAIT)
2739 local_irq_enable();
2740
2741 /*
2742 * The test for missing atomic flag is performed here, rather than
2743 * the more obvious place, simply to reduce the critical path length
2744 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2745 * will eventually be caught here (where it matters).
2746 */
2747 kmem_flagcheck(cachep, flags);
2748
Andrew Mortona737b3e2006-03-22 00:08:11 -08002749 /*
2750 * Get mem for the objs. Attempt to allocate a physical page from
2751 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002752 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002753 if (!page)
2754 page = kmem_getpages(cachep, local_flags, nodeid);
2755 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 goto failed;
2757
2758 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002759 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002760 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002761 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 goto opps1;
2763
Joonsoo Kim8456a642013-10-24 10:07:49 +09002764 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
Joonsoo Kim8456a642013-10-24 10:07:49 +09002766 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767
2768 if (local_flags & __GFP_WAIT)
2769 local_irq_disable();
2770 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002771 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772
2773 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002774 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002776 n->free_objects += cachep->num;
2777 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002779opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002780 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002781failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 if (local_flags & __GFP_WAIT)
2783 local_irq_disable();
2784 return 0;
2785}
2786
2787#if DEBUG
2788
2789/*
2790 * Perform extra freeing checks:
2791 * - detect bad pointers.
2792 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 */
2794static void kfree_debugcheck(const void *objp)
2795{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 if (!virt_addr_valid(objp)) {
2797 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002798 (unsigned long)objp);
2799 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801}
2802
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002803static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2804{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002805 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002806
2807 redzone1 = *dbg_redzone1(cache, obj);
2808 redzone2 = *dbg_redzone2(cache, obj);
2809
2810 /*
2811 * Redzone is ok.
2812 */
2813 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2814 return;
2815
2816 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2817 slab_error(cache, "double free detected");
2818 else
2819 slab_error(cache, "memory outside object was overwritten");
2820
David Woodhouseb46b8f12007-05-08 00:22:59 -07002821 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002822 obj, redzone1, redzone2);
2823}
2824
Pekka Enberg343e0d72006-02-01 03:05:50 -08002825static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002826 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002829 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002831 BUG_ON(virt_to_cache(objp) != cachep);
2832
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002833 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002835 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002838 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2840 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2841 }
2842 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002843 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844
Joonsoo Kim8456a642013-10-24 10:07:49 +09002845 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846
2847 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002848 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002850 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 if (cachep->flags & SLAB_POISON) {
2852#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002853 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002854 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002855 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002856 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 } else {
2858 poison_obj(cachep, objp, POISON_FREE);
2859 }
2860#else
2861 poison_obj(cachep, objp, POISON_FREE);
2862#endif
2863 }
2864 return objp;
2865}
2866
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867#else
2868#define kfree_debugcheck(x) do { } while(0)
2869#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870#endif
2871
Mel Gorman072bb0a2012-07-31 16:43:58 -07002872static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2873 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874{
2875 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002876 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002878 int node;
2879
Joe Korty6d2144d2008-03-05 15:04:59 -08002880 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002881 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002882 if (unlikely(force_refill))
2883 goto force_grow;
2884retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002885 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 batchcount = ac->batchcount;
2887 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002888 /*
2889 * If there was little recent activity on this cache, then
2890 * perform only a partial refill. Otherwise we could generate
2891 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 */
2893 batchcount = BATCHREFILL_LIMIT;
2894 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002895 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002897 BUG_ON(ac->avail > 0 || !n);
2898 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002899
Christoph Lameter3ded1752006-03-25 03:06:44 -08002900 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002901 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2902 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002903 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002904 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002905
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 while (batchcount > 0) {
2907 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002908 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002910 entry = n->slabs_partial.next;
2911 if (entry == &n->slabs_partial) {
2912 n->free_touched = 1;
2913 entry = n->slabs_free.next;
2914 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 goto must_grow;
2916 }
2917
Joonsoo Kim8456a642013-10-24 10:07:49 +09002918 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002920
2921 /*
2922 * The slab was either on partial or free list so
2923 * there must be at least one object available for
2924 * allocation.
2925 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002926 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002927
Joonsoo Kim8456a642013-10-24 10:07:49 +09002928 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 STATS_INC_ALLOCED(cachep);
2930 STATS_INC_ACTIVE(cachep);
2931 STATS_SET_HIGH(cachep);
2932
Joonsoo Kim8456a642013-10-24 10:07:49 +09002933 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002934 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936
2937 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002938 list_del(&page->lru);
2939 if (page->active == cachep->num)
2940 list_add(&page->list, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09002942 list_add(&page->list, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 }
2944
Andrew Mortona737b3e2006-03-22 00:08:11 -08002945must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002946 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002947alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002948 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949
2950 if (unlikely(!ac->avail)) {
2951 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002952force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002953 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002954
Andrew Mortona737b3e2006-03-22 00:08:11 -08002955 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002956 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002957 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002958
2959 /* no objects in sight? abort */
2960 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 return NULL;
2962
Andrew Mortona737b3e2006-03-22 00:08:11 -08002963 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 goto retry;
2965 }
2966 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002967
2968 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969}
2970
Andrew Mortona737b3e2006-03-22 00:08:11 -08002971static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2972 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973{
2974 might_sleep_if(flags & __GFP_WAIT);
2975#if DEBUG
2976 kmem_flagcheck(cachep, flags);
2977#endif
2978}
2979
2980#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002981static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002982 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983{
Joonsoo Kimd35426e2014-06-23 13:22:06 -07002984 struct page *page;
2985
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002986 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002988 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002990 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002991 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002992 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 else
2994 check_poison_obj(cachep, objp);
2995#else
2996 check_poison_obj(cachep, objp);
2997#endif
2998 poison_obj(cachep, objp, POISON_INUSE);
2999 }
3000 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003001 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002
3003 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003004 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3005 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3006 slab_error(cachep, "double free, or memory outside"
3007 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003008 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003009 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003010 objp, *dbg_redzone1(cachep, objp),
3011 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 }
3013 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3014 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3015 }
Joonsoo Kimd35426e2014-06-23 13:22:06 -07003016
3017 page = virt_to_head_page(objp);
3018 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003019 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003020 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003021 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003022 if (ARCH_SLAB_MINALIGN &&
3023 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003024 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003025 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003026 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 return objp;
3028}
3029#else
3030#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3031#endif
3032
Akinobu Mita773ff602008-12-23 19:37:01 +09003033static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003034{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00003035 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003036 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003037
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003038 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003039}
3040
Pekka Enberg343e0d72006-02-01 03:05:50 -08003041static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003043 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003045 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046
Alok N Kataria5c382302005-09-27 21:45:46 -07003047 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003048
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003049 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003052 objp = ac_get_obj(cachep, ac, flags, false);
3053
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003054 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003055 * Allow for the possibility all avail objects are not allowed
3056 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003057 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003058 if (objp) {
3059 STATS_INC_ALLOCHIT(cachep);
3060 goto out;
3061 }
3062 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003064
3065 STATS_INC_ALLOCMISS(cachep);
3066 objp = cache_alloc_refill(cachep, flags, force_refill);
3067 /*
3068 * the 'ac' may be updated by cache_alloc_refill(),
3069 * and kmemleak_erase() requires its correct value.
3070 */
3071 ac = cpu_cache_get(cachep);
3072
3073out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003074 /*
3075 * To avoid a false negative, if an object that is in one of the
3076 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3077 * treat the array pointers as a reference to the object.
3078 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003079 if (objp)
3080 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003081 return objp;
3082}
3083
Christoph Lametere498be72005-09-09 13:03:32 -07003084#ifdef CONFIG_NUMA
3085/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003086 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003087 *
3088 * If we are in_interrupt, then process context, including cpusets and
3089 * mempolicy, may not apply and should not be used for allocation policy.
3090 */
3091static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3092{
3093 int nid_alloc, nid_here;
3094
Christoph Lameter765c4502006-09-27 01:50:08 -07003095 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003096 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003097 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003098 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003099 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003100 else if (current->mempolicy)
Andi Kleene7b691b2012-06-09 02:40:03 -07003101 nid_alloc = slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003102 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003103 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003104 return NULL;
3105}
3106
3107/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003108 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003109 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003110 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003111 * perform an allocation without specifying a node. This allows the page
3112 * allocator to do its reclaim / fallback magic. We then insert the
3113 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003114 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003115static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003116{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003117 struct zonelist *zonelist;
3118 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003119 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003120 struct zone *zone;
3121 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003122 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003123 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003124 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003125
3126 if (flags & __GFP_THISNODE)
3127 return NULL;
3128
Christoph Lameter6cb06222007-10-16 01:25:41 -07003129 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003130
Mel Gormancc9a6c82012-03-21 16:34:11 -07003131retry_cpuset:
3132 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07003133 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003134
Christoph Lameter3c517a62006-12-06 20:33:29 -08003135retry:
3136 /*
3137 * Look through allowed nodes for objects available
3138 * from existing per node queues.
3139 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003140 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3141 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003142
Mel Gorman54a6eb52008-04-28 02:12:16 -07003143 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter6a673682013-01-10 19:14:19 +00003144 cache->node[nid] &&
3145 cache->node[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003146 obj = ____cache_alloc_node(cache,
3147 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003148 if (obj)
3149 break;
3150 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003151 }
3152
Christoph Lametercfce6602007-05-06 14:50:17 -07003153 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003154 /*
3155 * This allocation will be performed within the constraints
3156 * of the current cpuset / memory policy requirements.
3157 * We may trigger various forms of reclaim on the allowed
3158 * set and go into memory reserves if necessary.
3159 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003160 struct page *page;
3161
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003162 if (local_flags & __GFP_WAIT)
3163 local_irq_enable();
3164 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003165 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003166 if (local_flags & __GFP_WAIT)
3167 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003168 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003169 /*
3170 * Insert into the appropriate per node queues
3171 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003172 nid = page_to_nid(page);
3173 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003174 obj = ____cache_alloc_node(cache,
3175 flags | GFP_THISNODE, nid);
3176 if (!obj)
3177 /*
3178 * Another processor may allocate the
3179 * objects in the slab since we are
3180 * not holding any locks.
3181 */
3182 goto retry;
3183 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003184 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003185 obj = NULL;
3186 }
3187 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003188 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003189
3190 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3191 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003192 return obj;
3193}
3194
3195/*
Christoph Lametere498be72005-09-09 13:03:32 -07003196 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003198static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003199 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003200{
3201 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003202 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003203 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003204 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003205 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003207 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003208 n = cachep->node[nodeid];
3209 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003210
Andrew Mortona737b3e2006-03-22 00:08:11 -08003211retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003212 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003213 spin_lock(&n->list_lock);
3214 entry = n->slabs_partial.next;
3215 if (entry == &n->slabs_partial) {
3216 n->free_touched = 1;
3217 entry = n->slabs_free.next;
3218 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003219 goto must_grow;
3220 }
Christoph Lametere498be72005-09-09 13:03:32 -07003221
Joonsoo Kim8456a642013-10-24 10:07:49 +09003222 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003223 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003224
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003225 STATS_INC_NODEALLOCS(cachep);
3226 STATS_INC_ACTIVE(cachep);
3227 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003228
Joonsoo Kim8456a642013-10-24 10:07:49 +09003229 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003230
Joonsoo Kim8456a642013-10-24 10:07:49 +09003231 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003232 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003233 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003234 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003235
Joonsoo Kim8456a642013-10-24 10:07:49 +09003236 if (page->active == cachep->num)
3237 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003238 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003239 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003240
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003241 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003242 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003243
Andrew Mortona737b3e2006-03-22 00:08:11 -08003244must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003245 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003246 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003247 if (x)
3248 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003249
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003250 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003251
Andrew Mortona737b3e2006-03-22 00:08:11 -08003252done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003253 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003254}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003255
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003256static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003257slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003258 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003259{
3260 unsigned long save_flags;
3261 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003262 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003263
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003264 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003265
Nick Piggincf40bd12009-01-21 08:12:39 +01003266 lockdep_trace_alloc(flags);
3267
Akinobu Mita773ff602008-12-23 19:37:01 +09003268 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003269 return NULL;
3270
Glauber Costad79923f2012-12-18 14:22:48 -08003271 cachep = memcg_kmem_get_cache(cachep, flags);
3272
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003273 cache_alloc_debugcheck_before(cachep, flags);
3274 local_irq_save(save_flags);
3275
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003276 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003277 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003278
Christoph Lameter6a673682013-01-10 19:14:19 +00003279 if (unlikely(!cachep->node[nodeid])) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003280 /* Node not bootstrapped yet */
3281 ptr = fallback_alloc(cachep, flags);
3282 goto out;
3283 }
3284
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003285 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003286 /*
3287 * Use the locally cached objects if possible.
3288 * However ____cache_alloc does not allow fallback
3289 * to other nodes. It may fail while we still have
3290 * objects on other nodes available.
3291 */
3292 ptr = ____cache_alloc(cachep, flags);
3293 if (ptr)
3294 goto out;
3295 }
3296 /* ___cache_alloc_node can fall back to other nodes */
3297 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3298 out:
3299 local_irq_restore(save_flags);
3300 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003301 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003302 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003303
Pekka Enbergc175eea2008-05-09 20:35:53 +02003304 if (likely(ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003305 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003306
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003307 if (unlikely((flags & __GFP_ZERO) && ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003308 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003309
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003310 return ptr;
3311}
3312
3313static __always_inline void *
3314__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3315{
3316 void *objp;
3317
3318 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3319 objp = alternate_node_alloc(cache, flags);
3320 if (objp)
3321 goto out;
3322 }
3323 objp = ____cache_alloc(cache, flags);
3324
3325 /*
3326 * We may just have run out of memory on the local node.
3327 * ____cache_alloc_node() knows how to locate memory on other nodes
3328 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003329 if (!objp)
3330 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003331
3332 out:
3333 return objp;
3334}
3335#else
3336
3337static __always_inline void *
3338__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3339{
3340 return ____cache_alloc(cachep, flags);
3341}
3342
3343#endif /* CONFIG_NUMA */
3344
3345static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003346slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003347{
3348 unsigned long save_flags;
3349 void *objp;
3350
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003351 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003352
Nick Piggincf40bd12009-01-21 08:12:39 +01003353 lockdep_trace_alloc(flags);
3354
Akinobu Mita773ff602008-12-23 19:37:01 +09003355 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003356 return NULL;
3357
Glauber Costad79923f2012-12-18 14:22:48 -08003358 cachep = memcg_kmem_get_cache(cachep, flags);
3359
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003360 cache_alloc_debugcheck_before(cachep, flags);
3361 local_irq_save(save_flags);
3362 objp = __do_cache_alloc(cachep, flags);
3363 local_irq_restore(save_flags);
3364 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003365 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003366 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003367 prefetchw(objp);
3368
Pekka Enbergc175eea2008-05-09 20:35:53 +02003369 if (likely(objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003370 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003371
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003372 if (unlikely((flags & __GFP_ZERO) && objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003373 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003374
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003375 return objp;
3376}
Christoph Lametere498be72005-09-09 13:03:32 -07003377
3378/*
3379 * Caller needs to acquire correct kmem_list's list_lock
3380 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003381static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003382 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383{
3384 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003385 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386
3387 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003388 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003389 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390
Mel Gorman072bb0a2012-07-31 16:43:58 -07003391 clear_obj_pfmemalloc(&objpp[i]);
3392 objp = objpp[i];
3393
Joonsoo Kim8456a642013-10-24 10:07:49 +09003394 page = virt_to_head_page(objp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003395 n = cachep->node[node];
Joonsoo Kim8456a642013-10-24 10:07:49 +09003396 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003397 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003398 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003400 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401
3402 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003403 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003404 if (n->free_objects > n->free_limit) {
3405 n->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003406 /* No need to drop any previously held
3407 * lock here, even if we have a off-slab slab
3408 * descriptor it is guaranteed to come from
3409 * a different cache, refer to comments before
3410 * alloc_slabmgmt.
3411 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003412 slab_destroy(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003414 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415 }
3416 } else {
3417 /* Unconditionally move a slab to the end of the
3418 * partial list on free - maximum time for the
3419 * other objects to be freed, too.
3420 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003421 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 }
3423 }
3424}
3425
Pekka Enberg343e0d72006-02-01 03:05:50 -08003426static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427{
3428 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003429 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003430 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431
3432 batchcount = ac->batchcount;
3433#if DEBUG
3434 BUG_ON(!batchcount || batchcount > ac->avail);
3435#endif
3436 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003437 n = cachep->node[node];
3438 spin_lock(&n->list_lock);
3439 if (n->shared) {
3440 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003441 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 if (max) {
3443 if (batchcount > max)
3444 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003445 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003446 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 shared_array->avail += batchcount;
3448 goto free_done;
3449 }
3450 }
3451
Christoph Lameterff694162005-09-22 21:44:02 -07003452 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003453free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454#if STATS
3455 {
3456 int i = 0;
3457 struct list_head *p;
3458
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003459 p = n->slabs_free.next;
3460 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003461 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462
Joonsoo Kim8456a642013-10-24 10:07:49 +09003463 page = list_entry(p, struct page, lru);
3464 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
3466 i++;
3467 p = p->next;
3468 }
3469 STATS_SET_FREEABLE(cachep, i);
3470 }
3471#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003472 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003474 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475}
3476
3477/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003478 * Release an obj back to its cache. If the obj has a constructed state, it must
3479 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003481static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003482 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003484 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485
3486 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003487 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003488 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003490 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003491
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003492 /*
3493 * Skip calling cache_free_alien() when the platform is not numa.
3494 * This will avoid cache misses that happen while accessing slabp (which
3495 * is per page memory reference) to get nodeid. Instead use a global
3496 * variable to skip the call, which is mostly likely to be present in
3497 * the cache.
3498 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003499 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003500 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003501
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 if (likely(ac->avail < ac->limit)) {
3503 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504 } else {
3505 STATS_INC_FREEMISS(cachep);
3506 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003508
Mel Gorman072bb0a2012-07-31 16:43:58 -07003509 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510}
3511
3512/**
3513 * kmem_cache_alloc - Allocate an object
3514 * @cachep: The cache to allocate from.
3515 * @flags: See kmalloc().
3516 *
3517 * Allocate an object from this cache. The flags are only relevant
3518 * if the cache has no available objects.
3519 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003520void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003522 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003523
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003524 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003525 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003526
3527 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528}
3529EXPORT_SYMBOL(kmem_cache_alloc);
3530
Li Zefan0f24f122009-12-11 15:45:30 +08003531#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003532void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003533kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003534{
Steven Rostedt85beb582010-11-24 16:23:34 -05003535 void *ret;
3536
Ezequiel Garcia48356302012-09-08 17:47:57 -03003537 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003538
3539 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003540 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003541 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003542}
Steven Rostedt85beb582010-11-24 16:23:34 -05003543EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003544#endif
3545
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003547/**
3548 * kmem_cache_alloc_node - Allocate an object on the specified node
3549 * @cachep: The cache to allocate from.
3550 * @flags: See kmalloc().
3551 * @nodeid: node number of the target node.
3552 *
3553 * Identical to kmem_cache_alloc but it will allocate memory on the given
3554 * node, which can improve the performance for cpu bound structures.
3555 *
3556 * Fallback to other node is possible if __GFP_THISNODE is not set.
3557 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003558void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3559{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003560 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003561
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003562 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003563 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003564 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003565
3566 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003567}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568EXPORT_SYMBOL(kmem_cache_alloc_node);
3569
Li Zefan0f24f122009-12-11 15:45:30 +08003570#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003571void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003572 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003573 int nodeid,
3574 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003575{
Steven Rostedt85beb582010-11-24 16:23:34 -05003576 void *ret;
3577
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003578 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003579
Steven Rostedt85beb582010-11-24 16:23:34 -05003580 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003581 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003582 flags, nodeid);
3583 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003584}
Steven Rostedt85beb582010-11-24 16:23:34 -05003585EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003586#endif
3587
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003588static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003589__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003590{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003591 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003592
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003593 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003594 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3595 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003596 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003597}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003598
Li Zefan0bb38a52009-12-11 15:45:50 +08003599#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003600void *__kmalloc_node(size_t size, gfp_t flags, int node)
3601{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003602 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003603}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003604EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003605
3606void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003607 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003608{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003609 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003610}
3611EXPORT_SYMBOL(__kmalloc_node_track_caller);
3612#else
3613void *__kmalloc_node(size_t size, gfp_t flags, int node)
3614{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003615 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003616}
3617EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003618#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003619#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003620
3621/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003622 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003624 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003625 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003627static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003628 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003630 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003631 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003633 /* If you want to save a few bytes .text space: replace
3634 * __ with kmem_.
3635 * Then kmalloc uses the uninlined functions instead of the inline
3636 * functions.
3637 */
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003638 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003639 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3640 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003641 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003642
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003643 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003644 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003645
3646 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003647}
3648
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003649
Li Zefan0bb38a52009-12-11 15:45:50 +08003650#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003651void *__kmalloc(size_t size, gfp_t flags)
3652{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003653 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654}
3655EXPORT_SYMBOL(__kmalloc);
3656
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003657void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003658{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003659 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003660}
3661EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003662
3663#else
3664void *__kmalloc(size_t size, gfp_t flags)
3665{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003666 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003667}
3668EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003669#endif
3670
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671/**
3672 * kmem_cache_free - Deallocate an object
3673 * @cachep: The cache the allocation was from.
3674 * @objp: The previously allocated object.
3675 *
3676 * Free an object which was previously allocated from this
3677 * cache.
3678 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003679void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680{
3681 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003682 cachep = cache_from_obj(cachep, objp);
3683 if (!cachep)
3684 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685
3686 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003687 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003688 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003689 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003690 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003692
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003693 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694}
3695EXPORT_SYMBOL(kmem_cache_free);
3696
3697/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 * kfree - free previously allocated memory
3699 * @objp: pointer returned by kmalloc.
3700 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003701 * If @objp is NULL, no operation is performed.
3702 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 * Don't free memory not originally allocated by kmalloc()
3704 * or you will run into trouble.
3705 */
3706void kfree(const void *objp)
3707{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003708 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709 unsigned long flags;
3710
Pekka Enberg2121db72009-03-25 11:05:57 +02003711 trace_kfree(_RET_IP_, objp);
3712
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003713 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714 return;
3715 local_irq_save(flags);
3716 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003717 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003718 debug_check_no_locks_freed(objp, c->object_size);
3719
3720 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003721 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722 local_irq_restore(flags);
3723}
3724EXPORT_SYMBOL(kfree);
3725
Christoph Lametere498be72005-09-09 13:03:32 -07003726/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003727 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003728 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003729static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003730{
3731 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003732 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003733 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003734 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003735
Mel Gorman9c09a952008-01-24 05:49:54 -08003736 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003737
Paul Menage3395ee02006-12-06 20:32:16 -08003738 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003739 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003740 if (!new_alien)
3741 goto fail;
3742 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003743
Eric Dumazet63109842007-05-06 14:49:28 -07003744 new_shared = NULL;
3745 if (cachep->shared) {
3746 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003747 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003748 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003749 if (!new_shared) {
3750 free_alien_cache(new_alien);
3751 goto fail;
3752 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003753 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003754
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003755 n = cachep->node[node];
3756 if (n) {
3757 struct array_cache *shared = n->shared;
Christoph Lametercafeb022006-03-25 03:06:46 -08003758
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003759 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003760
Christoph Lametercafeb022006-03-25 03:06:46 -08003761 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003762 free_block(cachep, shared->entry,
3763 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003764
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003765 n->shared = new_shared;
3766 if (!n->alien) {
3767 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003768 new_alien = NULL;
3769 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003770 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003771 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003772 spin_unlock_irq(&n->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003773 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003774 free_alien_cache(new_alien);
3775 continue;
3776 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003777 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3778 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003779 free_alien_cache(new_alien);
3780 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003781 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003782 }
Christoph Lametere498be72005-09-09 13:03:32 -07003783
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003784 kmem_cache_node_init(n);
3785 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003786 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003787 n->shared = new_shared;
3788 n->alien = new_alien;
3789 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003790 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003791 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003792 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003793 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003794
Andrew Mortona737b3e2006-03-22 00:08:11 -08003795fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003796 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003797 /* Cache is not active yet. Roll back what we did */
3798 node--;
3799 while (node >= 0) {
Christoph Lameter6a673682013-01-10 19:14:19 +00003800 if (cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003801 n = cachep->node[node];
Christoph Lameter0718dc22006-03-25 03:06:47 -08003802
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003803 kfree(n->shared);
3804 free_alien_cache(n->alien);
3805 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003806 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003807 }
3808 node--;
3809 }
3810 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003811 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003812}
3813
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003815 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003816 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817};
3818
3819static void do_ccupdate_local(void *info)
3820{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003821 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003822 struct array_cache *old;
3823
3824 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003825 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003826
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3828 new->new[smp_processor_id()] = old;
3829}
3830
Christoph Lameter18004c52012-07-06 15:25:12 -05003831/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003832static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003833 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003834{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003835 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003836 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003838 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3839 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003840 if (!new)
3841 return -ENOMEM;
3842
Christoph Lametere498be72005-09-09 13:03:32 -07003843 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003844 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003845 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003846 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003847 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003848 kfree(new->new[i]);
3849 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003850 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851 }
3852 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003853 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003855 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003856
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 cachep->batchcount = batchcount;
3859 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003860 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861
Christoph Lametere498be72005-09-09 13:03:32 -07003862 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003863 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 if (!ccold)
3865 continue;
Christoph Lameter6a673682013-01-10 19:14:19 +00003866 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003867 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
Christoph Lameter6a673682013-01-10 19:14:19 +00003868 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003869 kfree(ccold);
3870 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003871 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003872 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873}
3874
Glauber Costa943a4512012-12-18 14:23:03 -08003875static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3876 int batchcount, int shared, gfp_t gfp)
3877{
3878 int ret;
3879 struct kmem_cache *c = NULL;
3880 int i = 0;
3881
3882 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3883
3884 if (slab_state < FULL)
3885 return ret;
3886
3887 if ((ret < 0) || !is_root_cache(cachep))
3888 return ret;
3889
Glauber Costaebe945c2012-12-18 14:23:10 -08003890 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003891 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003892 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003893 if (c)
3894 /* return value determined by the parent cache only */
3895 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3896 }
3897
3898 return ret;
3899}
3900
Christoph Lameter18004c52012-07-06 15:25:12 -05003901/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003902static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903{
3904 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003905 int limit = 0;
3906 int shared = 0;
3907 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908
Glauber Costa943a4512012-12-18 14:23:03 -08003909 if (!is_root_cache(cachep)) {
3910 struct kmem_cache *root = memcg_root_cache(cachep);
3911 limit = root->limit;
3912 shared = root->shared;
3913 batchcount = root->batchcount;
3914 }
3915
3916 if (limit && shared && batchcount)
3917 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003918 /*
3919 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 * - create a LIFO ordering, i.e. return objects that are cache-warm
3921 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003922 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923 * bufctl chains: array operations are cheaper.
3924 * The numbers are guessed, we should auto-tune as described by
3925 * Bonwick.
3926 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003927 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003929 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003931 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003933 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 limit = 54;
3935 else
3936 limit = 120;
3937
Andrew Mortona737b3e2006-03-22 00:08:11 -08003938 /*
3939 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 * allocation behaviour: Most allocs on one cpu, most free operations
3941 * on another cpu. For these cases, an efficient object passing between
3942 * cpus is necessary. This is provided by a shared array. The array
3943 * replaces Bonwick's magazine layer.
3944 * On uniprocessor, it's functionally equivalent (but less efficient)
3945 * to a larger limit. Thus disabled by default.
3946 */
3947 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003948 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950
3951#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003952 /*
3953 * With debugging enabled, large batchcount lead to excessively long
3954 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955 */
3956 if (limit > 32)
3957 limit = 32;
3958#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003959 batchcount = (limit + 1) / 2;
3960skip_setup:
3961 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 if (err)
3963 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003964 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003965 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966}
3967
Christoph Lameter1b552532006-03-22 00:09:07 -08003968/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003969 * Drain an array if it contains any elements taking the node lock only if
3970 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003971 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003972 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003973static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003974 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975{
3976 int tofree;
3977
Christoph Lameter1b552532006-03-22 00:09:07 -08003978 if (!ac || !ac->avail)
3979 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 if (ac->touched && !force) {
3981 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003982 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003983 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003984 if (ac->avail) {
3985 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3986 if (tofree > ac->avail)
3987 tofree = (ac->avail + 1) / 2;
3988 free_block(cachep, ac->entry, tofree, node);
3989 ac->avail -= tofree;
3990 memmove(ac->entry, &(ac->entry[tofree]),
3991 sizeof(void *) * ac->avail);
3992 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003993 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 }
3995}
3996
3997/**
3998 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003999 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 *
4001 * Called from workqueue/eventd every few seconds.
4002 * Purpose:
4003 * - clear the per-cpu caches for this CPU.
4004 * - return freeable pages to the main free memory pool.
4005 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004006 * If we cannot acquire the cache chain mutex then just give up - we'll try
4007 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004009static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004011 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004012 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004013 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004014 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015
Christoph Lameter18004c52012-07-06 15:25:12 -05004016 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004018 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019
Christoph Lameter18004c52012-07-06 15:25:12 -05004020 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 check_irq_on();
4022
Christoph Lameter35386e32006-03-22 00:09:05 -08004023 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004024 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004025 * have established with reasonable certainty that
4026 * we can do some work if the lock was obtained.
4027 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004028 n = searchp->node[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004029
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004030 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004032 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033
Christoph Lameter35386e32006-03-22 00:09:05 -08004034 /*
4035 * These are racy checks but it does not matter
4036 * if we skip one check or scan twice.
4037 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004038 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004039 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004041 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004043 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004045 if (n->free_touched)
4046 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004047 else {
4048 int freed;
4049
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004050 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004051 5 * searchp->num - 1) / (5 * searchp->num));
4052 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004054next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 cond_resched();
4056 }
4057 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004058 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004059 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004060out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004061 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004062 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063}
4064
Linus Torvalds158a9622008-01-02 13:04:48 -08004065#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004066void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004068 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004069 unsigned long active_objs;
4070 unsigned long num_objs;
4071 unsigned long active_slabs = 0;
4072 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004073 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004075 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004076 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 active_objs = 0;
4079 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004080 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004081 n = cachep->node[node];
4082 if (!n)
Christoph Lametere498be72005-09-09 13:03:32 -07004083 continue;
4084
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004085 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004086 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004087
Joonsoo Kim8456a642013-10-24 10:07:49 +09004088 list_for_each_entry(page, &n->slabs_full, lru) {
4089 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004090 error = "slabs_full accounting error";
4091 active_objs += cachep->num;
4092 active_slabs++;
4093 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004094 list_for_each_entry(page, &n->slabs_partial, lru) {
4095 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004096 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004097 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004098 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004099 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004100 active_slabs++;
4101 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004102 list_for_each_entry(page, &n->slabs_free, lru) {
4103 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004104 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004105 num_slabs++;
4106 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004107 free_objects += n->free_objects;
4108 if (n->shared)
4109 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004110
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004111 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004113 num_slabs += active_slabs;
4114 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004115 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116 error = "free_objects accounting error";
4117
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004118 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119 if (error)
4120 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4121
Glauber Costa0d7561c2012-10-19 18:20:27 +04004122 sinfo->active_objs = active_objs;
4123 sinfo->num_objs = num_objs;
4124 sinfo->active_slabs = active_slabs;
4125 sinfo->num_slabs = num_slabs;
4126 sinfo->shared_avail = shared_avail;
4127 sinfo->limit = cachep->limit;
4128 sinfo->batchcount = cachep->batchcount;
4129 sinfo->shared = cachep->shared;
4130 sinfo->objects_per_slab = cachep->num;
4131 sinfo->cache_order = cachep->gfporder;
4132}
4133
4134void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4135{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004137 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 unsigned long high = cachep->high_mark;
4139 unsigned long allocs = cachep->num_allocations;
4140 unsigned long grown = cachep->grown;
4141 unsigned long reaped = cachep->reaped;
4142 unsigned long errors = cachep->errors;
4143 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004145 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004146 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147
Joe Perchese92dd4f2010-03-26 19:27:58 -07004148 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4149 "%4lu %4lu %4lu %4lu %4lu",
4150 allocs, high, grown,
4151 reaped, errors, max_freeable, node_allocs,
4152 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 }
4154 /* cpu stats */
4155 {
4156 unsigned long allochit = atomic_read(&cachep->allochit);
4157 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4158 unsigned long freehit = atomic_read(&cachep->freehit);
4159 unsigned long freemiss = atomic_read(&cachep->freemiss);
4160
4161 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004162 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 }
4164#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165}
4166
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167#define MAX_SLABINFO_WRITE 128
4168/**
4169 * slabinfo_write - Tuning for the slab allocator
4170 * @file: unused
4171 * @buffer: user buffer
4172 * @count: data length
4173 * @ppos: unused
4174 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004175ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004176 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004178 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004180 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004181
Linus Torvalds1da177e2005-04-16 15:20:36 -07004182 if (count > MAX_SLABINFO_WRITE)
4183 return -EINVAL;
4184 if (copy_from_user(&kbuf, buffer, count))
4185 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004186 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187
4188 tmp = strchr(kbuf, ' ');
4189 if (!tmp)
4190 return -EINVAL;
4191 *tmp = '\0';
4192 tmp++;
4193 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4194 return -EINVAL;
4195
4196 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004197 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004199 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004201 if (limit < 1 || batchcount < 1 ||
4202 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004203 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004205 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004206 batchcount, shared,
4207 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 }
4209 break;
4210 }
4211 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004212 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 if (res >= 0)
4214 res = count;
4215 return res;
4216}
Al Viro871751e2006-03-25 03:06:39 -08004217
4218#ifdef CONFIG_DEBUG_SLAB_LEAK
4219
4220static void *leaks_start(struct seq_file *m, loff_t *pos)
4221{
Christoph Lameter18004c52012-07-06 15:25:12 -05004222 mutex_lock(&slab_mutex);
4223 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004224}
4225
4226static inline int add_caller(unsigned long *n, unsigned long v)
4227{
4228 unsigned long *p;
4229 int l;
4230 if (!v)
4231 return 1;
4232 l = n[1];
4233 p = n + 2;
4234 while (l) {
4235 int i = l/2;
4236 unsigned long *q = p + 2 * i;
4237 if (*q == v) {
4238 q[1]++;
4239 return 1;
4240 }
4241 if (*q > v) {
4242 l = i;
4243 } else {
4244 p = q + 2;
4245 l -= i + 1;
4246 }
4247 }
4248 if (++n[1] == n[0])
4249 return 0;
4250 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4251 p[0] = v;
4252 p[1] = 1;
4253 return 1;
4254}
4255
Joonsoo Kim8456a642013-10-24 10:07:49 +09004256static void handle_slab(unsigned long *n, struct kmem_cache *c,
4257 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004258{
4259 void *p;
Joonsoo Kimd35426e2014-06-23 13:22:06 -07004260 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004261
Al Viro871751e2006-03-25 03:06:39 -08004262 if (n[0] == n[1])
4263 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004264 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimd35426e2014-06-23 13:22:06 -07004265 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004266 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004267
Al Viro871751e2006-03-25 03:06:39 -08004268 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4269 return;
4270 }
4271}
4272
4273static void show_symbol(struct seq_file *m, unsigned long address)
4274{
4275#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004276 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004277 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004278
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004279 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004280 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004281 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004282 seq_printf(m, " [%s]", modname);
4283 return;
4284 }
4285#endif
4286 seq_printf(m, "%p", (void *)address);
4287}
4288
4289static int leaks_show(struct seq_file *m, void *p)
4290{
Thierry Reding0672aa72012-06-22 19:42:49 +02004291 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004292 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004293 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004294 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004295 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004296 int node;
4297 int i;
4298
4299 if (!(cachep->flags & SLAB_STORE_USER))
4300 return 0;
4301 if (!(cachep->flags & SLAB_RED_ZONE))
4302 return 0;
4303
4304 /* OK, we can do it */
4305
Christoph Lameterdb845062013-02-05 18:45:23 +00004306 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004307
4308 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004309 n = cachep->node[node];
4310 if (!n)
Al Viro871751e2006-03-25 03:06:39 -08004311 continue;
4312
4313 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004314 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004315
Joonsoo Kim8456a642013-10-24 10:07:49 +09004316 list_for_each_entry(page, &n->slabs_full, lru)
4317 handle_slab(x, cachep, page);
4318 list_for_each_entry(page, &n->slabs_partial, lru)
4319 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004320 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004321 }
4322 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004323 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004324 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004325 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004326 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004327 if (!m->private) {
4328 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004329 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004330 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004331 return -ENOMEM;
4332 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004333 *(unsigned long *)m->private = x[0] * 2;
4334 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004335 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004336 /* Now make sure this entry will be retried */
4337 m->count = m->size;
4338 return 0;
4339 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004340 for (i = 0; i < x[1]; i++) {
4341 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4342 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004343 seq_putc(m, '\n');
4344 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004345
Al Viro871751e2006-03-25 03:06:39 -08004346 return 0;
4347}
4348
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004349static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004350 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004351 .next = slab_next,
4352 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004353 .show = leaks_show,
4354};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004355
4356static int slabstats_open(struct inode *inode, struct file *file)
4357{
4358 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4359 int ret = -ENOMEM;
4360 if (n) {
4361 ret = seq_open(file, &slabstats_op);
4362 if (!ret) {
4363 struct seq_file *m = file->private_data;
4364 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4365 m->private = n;
4366 n = NULL;
4367 }
4368 kfree(n);
4369 }
4370 return ret;
4371}
4372
4373static const struct file_operations proc_slabstats_operations = {
4374 .open = slabstats_open,
4375 .read = seq_read,
4376 .llseek = seq_lseek,
4377 .release = seq_release_private,
4378};
Al Viro871751e2006-03-25 03:06:39 -08004379#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004380
4381static int __init slab_proc_init(void)
4382{
4383#ifdef CONFIG_DEBUG_SLAB_LEAK
4384 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4385#endif
4386 return 0;
4387}
4388module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004389#endif
4390
Manfred Spraul00e145b2005-09-03 15:55:07 -07004391/**
4392 * ksize - get the actual amount of memory allocated for a given object
4393 * @objp: Pointer to the object
4394 *
4395 * kmalloc may internally round up allocations and return more memory
4396 * than requested. ksize() can be used to determine the actual amount of
4397 * memory allocated. The caller may use this additional memory, even though
4398 * a smaller amount of memory was initially specified with the kmalloc call.
4399 * The caller must guarantee that objp points to a valid object previously
4400 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4401 * must not be freed during the duration of the call.
4402 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004403size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004405 BUG_ON(!objp);
4406 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004407 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004408
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004409 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004410}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004411EXPORT_SYMBOL(ksize);