blob: b344e67071280c9df522ef77f56ffa62ed1b4d63 [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
29 * slabs and you must pass objects with the same intializations to
30 * 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
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800106#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800110#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700111#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800112#include <linux/reciprocal_div.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#include <asm/cacheflush.h>
115#include <asm/tlbflush.h>
116#include <asm/page.h>
117
118/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400140#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142#ifndef cache_line_size
143#define cache_line_size() L1_CACHE_BYTES
144#endif
145
146#ifndef ARCH_KMALLOC_MINALIGN
147/*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700156#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#endif
158
159#ifndef ARCH_SLAB_MINALIGN
160/*
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
166 */
167#define ARCH_SLAB_MINALIGN 0
168#endif
169
170#ifndef ARCH_KMALLOC_FLAGS
171#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172#endif
173
174/* Legal flag mask for kmem_cache_create(). */
175#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700176# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800178 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700179 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800183# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700184 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187#endif
188
189/*
190 * kmem_bufctl_t:
191 *
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
194 *
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 */
207
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700208typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800211#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228};
229
230/*
231 * struct slab_rcu
232 *
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
240 *
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
243 *
244 * We assume struct slab_rcu can overlay struct slab when destroying.
245 */
246struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800248 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250};
251
252/*
253 * struct array_cache
254 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
259 *
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
262 *
263 */
264struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700269 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800270 void *entry[0]; /*
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
274 * [0] is for gcc 2.95. It should really be [].
275 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276};
277
Andrew Mortona737b3e2006-03-22 00:08:11 -0800278/*
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
282#define BOOT_CPUCACHE_ENTRIES 1
283struct arraycache_init {
284 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800285 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286};
287
288/*
Christoph Lametere498be72005-09-09 13:03:32 -0700289 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 */
291struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800297 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303};
304
Christoph Lametere498be72005-09-09 13:03:32 -0700305/*
306 * Need this for bootstrapping a per node allocator.
307 */
308#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
309struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310#define CACHE_CACHE 0
311#define SIZE_AC 1
312#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Christoph Lametered11d9e2006-06-30 01:55:45 -0700314static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700318static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000319static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700320
Christoph Lametere498be72005-09-09 13:03:32 -0700321/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700324 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700325static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700326{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800327 extern void __bad_size(void);
328
Christoph Lametere498be72005-09-09 13:03:32 -0700329 if (__builtin_constant_p(size)) {
330 int i = 0;
331
332#define CACHE(x) \
333 if (size <=x) \
334 return i; \
335 else \
336 i++;
337#include "linux/kmalloc_sizes.h"
338#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700340 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800341 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700342 return 0;
343}
344
Ingo Molnare0a42722006-06-23 02:03:46 -0700345static int slab_early_init = 1;
346
Christoph Lametere498be72005-09-09 13:03:32 -0700347#define INDEX_AC index_of(sizeof(struct arraycache_init))
348#define INDEX_L3 index_of(sizeof(struct kmem_list3))
349
Pekka Enberg5295a742006-02-01 03:05:48 -0800350static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700351{
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800357 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
361}
362
Andrew Mortona737b3e2006-03-22 00:08:11 -0800363#define MAKE_LIST(cachep, listp, slab, nodeid) \
364 do { \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700367 } while (0)
368
Andrew Mortona737b3e2006-03-22 00:08:11 -0800369#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
376/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800377 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 *
379 * manages a cache.
380 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800381
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800382struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800385/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800386 unsigned int batchcount;
387 unsigned int limit;
388 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800390 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800391 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800393
Andrew Mortona737b3e2006-03-22 00:08:11 -0800394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800397/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800402 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Andrew Mortona737b3e2006-03-22 00:08:11 -0800404 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800407 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800408 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800411 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800413/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 const char *name;
415 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800417/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
422 unsigned long grown;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700428 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800429 atomic_t allochit;
430 atomic_t allocmiss;
431 atomic_t freehit;
432 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433#endif
434#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800435 /*
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
440 */
441 int obj_offset;
442 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700444 /*
445 * We put nodelists[] at the end of kmem_cache, because we want to size
446 * this array to nr_node_ids slots instead of MAX_NUMNODES
447 * (see kmem_cache_init())
448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
449 * is statically defined, so we reserve the max number of nodes.
450 */
451 struct kmem_list3 *nodelists[MAX_NUMNODES];
452 /*
453 * Do not add fields after nodelists[]
454 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455};
456
457#define CFLGS_OFF_SLAB (0x80000000UL)
458#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459
460#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800461/*
462 * Optimization question: fewer reaps means less probability for unnessary
463 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100465 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 * which could lock up otherwise freeable slabs.
467 */
468#define REAPTIMEOUT_CPUC (2*HZ)
469#define REAPTIMEOUT_LIST3 (4*HZ)
470
471#if STATS
472#define STATS_INC_ACTIVE(x) ((x)->num_active++)
473#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
474#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
475#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700476#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800477#define STATS_SET_HIGH(x) \
478 do { \
479 if ((x)->num_active > (x)->high_mark) \
480 (x)->high_mark = (x)->num_active; \
481 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482#define STATS_INC_ERR(x) ((x)->errors++)
483#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700484#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700485#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800486#define STATS_SET_FREEABLE(x, i) \
487 do { \
488 if ((x)->max_freeable < i) \
489 (x)->max_freeable = i; \
490 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
492#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
493#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
494#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
495#else
496#define STATS_INC_ACTIVE(x) do { } while (0)
497#define STATS_DEC_ACTIVE(x) do { } while (0)
498#define STATS_INC_ALLOCED(x) do { } while (0)
499#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700500#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501#define STATS_SET_HIGH(x) do { } while (0)
502#define STATS_INC_ERR(x) do { } while (0)
503#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700504#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700505#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800506#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507#define STATS_INC_ALLOCHIT(x) do { } while (0)
508#define STATS_INC_ALLOCMISS(x) do { } while (0)
509#define STATS_INC_FREEHIT(x) do { } while (0)
510#define STATS_INC_FREEMISS(x) do { } while (0)
511#endif
512
513#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514
Andrew Mortona737b3e2006-03-22 00:08:11 -0800515/*
516 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 * the end of an object is aligned with the end of the real
520 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800523 * cachep->obj_offset: The real object.
524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
526 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531}
532
Pekka Enberg343e0d72006-02-01 03:05:50 -0800533static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800535 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536}
537
David Woodhouseb46b8f12007-05-08 00:22:59 -0700538static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539{
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700541 return (unsigned long long*) (objp + obj_offset(cachep) -
542 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
David Woodhouseb46b8f12007-05-08 00:22:59 -0700545static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700549 return (unsigned long long *)(objp + cachep->buffer_size -
550 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400551 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700552 return (unsigned long long *) (objp + cachep->buffer_size -
553 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554}
555
Pekka Enberg343e0d72006-02-01 03:05:50 -0800556static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
558 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
562#else
563
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800564#define obj_offset(x) 0
565#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700566#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
569
570#endif
571
572/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 * Do not go above this order unless 0 objects fit into the slab.
574 */
575#define BREAK_GFP_ORDER_HI 1
576#define BREAK_GFP_ORDER_LO 0
577static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
578
Andrew Mortona737b3e2006-03-22 00:08:11 -0800579/*
580 * Functions for storing/retrieving the cachep and or slab from the page
581 * allocator. These are used to find the slab an obj belongs to. With kfree(),
582 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800584static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
585{
586 page->lru.next = (struct list_head *)cache;
587}
588
589static inline struct kmem_cache *page_get_cache(struct page *page)
590{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700591 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700592 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800593 return (struct kmem_cache *)page->lru.next;
594}
595
596static inline void page_set_slab(struct page *page, struct slab *slab)
597{
598 page->lru.prev = (struct list_head *)slab;
599}
600
601static inline struct slab *page_get_slab(struct page *page)
602{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700603 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800604 return (struct slab *)page->lru.prev;
605}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800607static inline struct kmem_cache *virt_to_cache(const void *obj)
608{
Christoph Lameterb49af682007-05-06 14:49:41 -0700609 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800610 return page_get_cache(page);
611}
612
613static inline struct slab *virt_to_slab(const void *obj)
614{
Christoph Lameterb49af682007-05-06 14:49:41 -0700615 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800616 return page_get_slab(page);
617}
618
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800619static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
620 unsigned int idx)
621{
622 return slab->s_mem + cache->buffer_size * idx;
623}
624
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800625/*
626 * We want to avoid an expensive divide : (offset / cache->buffer_size)
627 * Using the fact that buffer_size is a constant for a particular cache,
628 * we can replace (offset / cache->buffer_size) by
629 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
630 */
631static inline unsigned int obj_to_index(const struct kmem_cache *cache,
632 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800633{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800634 u32 offset = (obj - slab->s_mem);
635 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800636}
637
Andrew Mortona737b3e2006-03-22 00:08:11 -0800638/*
639 * These are the default caches for kmalloc. Custom caches can have other sizes.
640 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641struct cache_sizes malloc_sizes[] = {
642#define CACHE(x) { .cs_size = (x) },
643#include <linux/kmalloc_sizes.h>
644 CACHE(ULONG_MAX)
645#undef CACHE
646};
647EXPORT_SYMBOL(malloc_sizes);
648
649/* Must match cache_sizes above. Out of line to keep cache footprint low. */
650struct cache_names {
651 char *name;
652 char *name_dma;
653};
654
655static struct cache_names __initdata cache_names[] = {
656#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
657#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800658 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659#undef CACHE
660};
661
662static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800665 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666
667/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800668static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800669 .batchcount = 1,
670 .limit = BOOT_CPUCACHE_ENTRIES,
671 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800672 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800673 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674};
675
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700676#define BAD_ALIEN_MAGIC 0x01020304ul
677
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200678#ifdef CONFIG_LOCKDEP
679
680/*
681 * Slab sometimes uses the kmalloc slabs to store the slab headers
682 * for other slabs "off slab".
683 * The locking for this is tricky in that it nests within the locks
684 * of all other slabs in a few places; to deal with this special
685 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700686 *
687 * We set lock class for alien array caches which are up during init.
688 * The lock annotation will be lost if all cpus of a node goes down and
689 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200690 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700691static struct lock_class_key on_slab_l3_key;
692static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200693
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700694static inline void init_lock_keys(void)
695
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200696{
697 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700698 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200699
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700700 while (s->cs_size != ULONG_MAX) {
701 for_each_node(q) {
702 struct array_cache **alc;
703 int r;
704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
705 if (!l3 || OFF_SLAB(s->cs_cachep))
706 continue;
707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
708 alc = l3->alien;
709 /*
710 * FIXME: This check for BAD_ALIEN_MAGIC
711 * should go away when common slab code is taught to
712 * work even without alien caches.
713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
714 * for alloc_alien_cache,
715 */
716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
717 continue;
718 for_each_node(r) {
719 if (alc[r])
720 lockdep_set_class(&alc[r]->lock,
721 &on_slab_alc_key);
722 }
723 }
724 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200725 }
726}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200727#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700728static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200729{
730}
731#endif
732
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800733/*
734 * 1. Guard access to the cache-chain.
735 * 2. Protect sanity of cpu_online_map against cpu hotplug events
736 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800737static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738static struct list_head cache_chain;
739
740/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 * chicken and egg problem: delay the per-cpu array allocation
742 * until the general caches are up.
743 */
744static enum {
745 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700746 PARTIAL_AC,
747 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 FULL
749} g_cpucache_up;
750
Mike Kravetz39d24e62006-05-15 09:44:13 -0700751/*
752 * used by boot code to determine if it can use slab based allocator
753 */
754int slab_is_available(void)
755{
756 return g_cpucache_up == FULL;
757}
758
David Howells52bad642006-11-22 14:54:01 +0000759static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
Pekka Enberg343e0d72006-02-01 03:05:50 -0800761static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762{
763 return cachep->array[smp_processor_id()];
764}
765
Andrew Mortona737b3e2006-03-22 00:08:11 -0800766static inline struct kmem_cache *__find_general_cachep(size_t size,
767 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
769 struct cache_sizes *csizep = malloc_sizes;
770
771#if DEBUG
772 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800773 * kmem_cache_create(), or __kmalloc(), before
774 * the generic caches are initialized.
775 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700776 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777#endif
778 while (size > csizep->cs_size)
779 csizep++;
780
781 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700782 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 * has cs_{dma,}cachep==NULL. Thus no special case
784 * for large kmalloc calls required.
785 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800786#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 if (unlikely(gfpflags & GFP_DMA))
788 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800789#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 return csizep->cs_cachep;
791}
792
Adrian Bunkb2213852006-09-25 23:31:02 -0700793static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700794{
795 return __find_general_cachep(size, gfpflags);
796}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700797
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800798static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800800 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
801}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Andrew Mortona737b3e2006-03-22 00:08:11 -0800803/*
804 * Calculate the number of objects and left-over bytes for a given buffer size.
805 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800806static void cache_estimate(unsigned long gfporder, size_t buffer_size,
807 size_t align, int flags, size_t *left_over,
808 unsigned int *num)
809{
810 int nr_objs;
811 size_t mgmt_size;
812 size_t slab_size = PAGE_SIZE << gfporder;
813
814 /*
815 * The slab management structure can be either off the slab or
816 * on it. For the latter case, the memory allocated for a
817 * slab is used for:
818 *
819 * - The struct slab
820 * - One kmem_bufctl_t for each object
821 * - Padding to respect alignment of @align
822 * - @buffer_size bytes for each object
823 *
824 * If the slab management structure is off the slab, then the
825 * alignment will already be calculated into the size. Because
826 * the slabs are all pages aligned, the objects will be at the
827 * correct alignment when allocated.
828 */
829 if (flags & CFLGS_OFF_SLAB) {
830 mgmt_size = 0;
831 nr_objs = slab_size / buffer_size;
832
833 if (nr_objs > SLAB_LIMIT)
834 nr_objs = SLAB_LIMIT;
835 } else {
836 /*
837 * Ignore padding for the initial guess. The padding
838 * is at most @align-1 bytes, and @buffer_size is at
839 * least @align. In the worst case, this result will
840 * be one greater than the number of objects that fit
841 * into the memory allocation when taking the padding
842 * into account.
843 */
844 nr_objs = (slab_size - sizeof(struct slab)) /
845 (buffer_size + sizeof(kmem_bufctl_t));
846
847 /*
848 * This calculated number will be either the right
849 * amount, or one greater than what we want.
850 */
851 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
852 > slab_size)
853 nr_objs--;
854
855 if (nr_objs > SLAB_LIMIT)
856 nr_objs = SLAB_LIMIT;
857
858 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800860 *num = nr_objs;
861 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862}
863
864#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
865
Andrew Mortona737b3e2006-03-22 00:08:11 -0800866static void __slab_error(const char *function, struct kmem_cache *cachep,
867 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868{
869 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800870 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 dump_stack();
872}
873
Paul Menage3395ee02006-12-06 20:32:16 -0800874/*
875 * By default on NUMA we use alien caches to stage the freeing of
876 * objects allocated from other nodes. This causes massive memory
877 * inefficiencies when using fake NUMA setup to split memory into a
878 * large number of small nodes, so it can be disabled on the command
879 * line
880 */
881
882static int use_alien_caches __read_mostly = 1;
883static int __init noaliencache_setup(char *s)
884{
885 use_alien_caches = 0;
886 return 1;
887}
888__setup("noaliencache", noaliencache_setup);
889
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800890#ifdef CONFIG_NUMA
891/*
892 * Special reaping functions for NUMA systems called from cache_reap().
893 * These take care of doing round robin flushing of alien caches (containing
894 * objects freed on different nodes from which they were allocated) and the
895 * flushing of remote pcps by calling drain_node_pages.
896 */
897static DEFINE_PER_CPU(unsigned long, reap_node);
898
899static void init_reap_node(int cpu)
900{
901 int node;
902
903 node = next_node(cpu_to_node(cpu), node_online_map);
904 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800905 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800906
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800907 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800908}
909
910static void next_reap_node(void)
911{
912 int node = __get_cpu_var(reap_node);
913
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800914 node = next_node(node, node_online_map);
915 if (unlikely(node >= MAX_NUMNODES))
916 node = first_node(node_online_map);
917 __get_cpu_var(reap_node) = node;
918}
919
920#else
921#define init_reap_node(cpu) do { } while (0)
922#define next_reap_node(void) do { } while (0)
923#endif
924
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925/*
926 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
927 * via the workqueue/eventd.
928 * Add the CPU number into the expiration time to minimize the possibility of
929 * the CPUs getting into lockstep and contending for the global cache chain
930 * lock.
931 */
932static void __devinit start_cpu_timer(int cpu)
933{
David Howells52bad642006-11-22 14:54:01 +0000934 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
936 /*
937 * When this gets called from do_initcalls via cpucache_init(),
938 * init_workqueues() has already run, so keventd will be setup
939 * at that time.
940 */
David Howells52bad642006-11-22 14:54:01 +0000941 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800942 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000943 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800944 schedule_delayed_work_on(cpu, reap_work,
945 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 }
947}
948
Christoph Lametere498be72005-09-09 13:03:32 -0700949static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800950 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800952 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 struct array_cache *nc = NULL;
954
Christoph Lametere498be72005-09-09 13:03:32 -0700955 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 if (nc) {
957 nc->avail = 0;
958 nc->limit = entries;
959 nc->batchcount = batchcount;
960 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700961 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 }
963 return nc;
964}
965
Christoph Lameter3ded1752006-03-25 03:06:44 -0800966/*
967 * Transfer objects in one arraycache to another.
968 * Locking must be handled by the caller.
969 *
970 * Return the number of entries transferred.
971 */
972static int transfer_objects(struct array_cache *to,
973 struct array_cache *from, unsigned int max)
974{
975 /* Figure out how many entries to transfer */
976 int nr = min(min(from->avail, max), to->limit - to->avail);
977
978 if (!nr)
979 return 0;
980
981 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
982 sizeof(void *) *nr);
983
984 from->avail -= nr;
985 to->avail += nr;
986 to->touched = 1;
987 return nr;
988}
989
Christoph Lameter765c4502006-09-27 01:50:08 -0700990#ifndef CONFIG_NUMA
991
992#define drain_alien_cache(cachep, alien) do { } while (0)
993#define reap_alien(cachep, l3) do { } while (0)
994
995static inline struct array_cache **alloc_alien_cache(int node, int limit)
996{
997 return (struct array_cache **)BAD_ALIEN_MAGIC;
998}
999
1000static inline void free_alien_cache(struct array_cache **ac_ptr)
1001{
1002}
1003
1004static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1005{
1006 return 0;
1007}
1008
1009static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1010 gfp_t flags)
1011{
1012 return NULL;
1013}
1014
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001015static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001016 gfp_t flags, int nodeid)
1017{
1018 return NULL;
1019}
1020
1021#else /* CONFIG_NUMA */
1022
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001023static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001024static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001025
Pekka Enberg5295a742006-02-01 03:05:48 -08001026static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001027{
1028 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001029 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001030 int i;
1031
1032 if (limit > 1)
1033 limit = 12;
1034 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1035 if (ac_ptr) {
1036 for_each_node(i) {
1037 if (i == node || !node_online(i)) {
1038 ac_ptr[i] = NULL;
1039 continue;
1040 }
1041 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1042 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001043 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001044 kfree(ac_ptr[i]);
1045 kfree(ac_ptr);
1046 return NULL;
1047 }
1048 }
1049 }
1050 return ac_ptr;
1051}
1052
Pekka Enberg5295a742006-02-01 03:05:48 -08001053static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001054{
1055 int i;
1056
1057 if (!ac_ptr)
1058 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001059 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001060 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001061 kfree(ac_ptr);
1062}
1063
Pekka Enberg343e0d72006-02-01 03:05:50 -08001064static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001065 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001066{
1067 struct kmem_list3 *rl3 = cachep->nodelists[node];
1068
1069 if (ac->avail) {
1070 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001071 /*
1072 * Stuff objects into the remote nodes shared array first.
1073 * That way we could avoid the overhead of putting the objects
1074 * into the free lists and getting them back later.
1075 */
shin, jacob693f7d32006-04-28 10:54:37 -05001076 if (rl3->shared)
1077 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001078
Christoph Lameterff694162005-09-22 21:44:02 -07001079 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001080 ac->avail = 0;
1081 spin_unlock(&rl3->list_lock);
1082 }
1083}
1084
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001085/*
1086 * Called from cache_reap() to regularly drain alien caches round robin.
1087 */
1088static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1089{
1090 int node = __get_cpu_var(reap_node);
1091
1092 if (l3->alien) {
1093 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001094
1095 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001096 __drain_alien_cache(cachep, ac, node);
1097 spin_unlock_irq(&ac->lock);
1098 }
1099 }
1100}
1101
Andrew Mortona737b3e2006-03-22 00:08:11 -08001102static void drain_alien_cache(struct kmem_cache *cachep,
1103 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001104{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001105 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001106 struct array_cache *ac;
1107 unsigned long flags;
1108
1109 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001110 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001111 if (ac) {
1112 spin_lock_irqsave(&ac->lock, flags);
1113 __drain_alien_cache(cachep, ac, i);
1114 spin_unlock_irqrestore(&ac->lock, flags);
1115 }
1116 }
1117}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001118
Ingo Molnar873623d2006-07-13 14:44:38 +02001119static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001120{
1121 struct slab *slabp = virt_to_slab(objp);
1122 int nodeid = slabp->nodeid;
1123 struct kmem_list3 *l3;
1124 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001125 int node;
1126
1127 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001128
1129 /*
1130 * Make sure we are not freeing a object from another node to the array
1131 * cache on this cpu.
1132 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001133 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001134 return 0;
1135
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001136 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001137 STATS_INC_NODEFREES(cachep);
1138 if (l3->alien && l3->alien[nodeid]) {
1139 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001140 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001141 if (unlikely(alien->avail == alien->limit)) {
1142 STATS_INC_ACOVERFLOW(cachep);
1143 __drain_alien_cache(cachep, alien, nodeid);
1144 }
1145 alien->entry[alien->avail++] = objp;
1146 spin_unlock(&alien->lock);
1147 } else {
1148 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1149 free_block(cachep, &objp, 1, nodeid);
1150 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1151 }
1152 return 1;
1153}
Christoph Lametere498be72005-09-09 13:03:32 -07001154#endif
1155
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001156static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001157 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158{
1159 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001160 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001161 struct kmem_list3 *l3 = NULL;
1162 int node = cpu_to_node(cpu);
1163 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
1165 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001166 case CPU_LOCK_ACQUIRE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001167 mutex_lock(&cache_chain_mutex);
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001168 break;
1169 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001170 case CPU_UP_PREPARE_FROZEN:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001171 /*
1172 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001173 * alloc_arraycache's are going to use this list.
1174 * kmalloc_node allows us to add the slab to the right
1175 * kmem_list3 and not this cpu's kmem_list3
1176 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
Christoph Lametere498be72005-09-09 13:03:32 -07001178 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001179 /*
1180 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001181 * begin anything. Make sure some other cpu on this
1182 * node has not already allocated this
1183 */
1184 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001185 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1186 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001187 goto bad;
1188 kmem_list3_init(l3);
1189 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001190 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001191
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001192 /*
1193 * The l3s don't come and go as CPUs come and
1194 * go. cache_chain_mutex is sufficient
1195 * protection here.
1196 */
Christoph Lametere498be72005-09-09 13:03:32 -07001197 cachep->nodelists[node] = l3;
1198 }
1199
1200 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1201 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001202 (1 + nr_cpus_node(node)) *
1203 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001204 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1205 }
1206
Andrew Mortona737b3e2006-03-22 00:08:11 -08001207 /*
1208 * Now we can go ahead with allocating the shared arrays and
1209 * array caches
1210 */
Christoph Lametere498be72005-09-09 13:03:32 -07001211 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001212 struct array_cache *nc;
Eric Dumazet63109842007-05-06 14:49:28 -07001213 struct array_cache *shared = NULL;
Paul Menage3395ee02006-12-06 20:32:16 -08001214 struct array_cache **alien = NULL;
Tobias Klausercd105df2006-01-08 01:00:59 -08001215
Christoph Lametere498be72005-09-09 13:03:32 -07001216 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001217 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 if (!nc)
1219 goto bad;
Eric Dumazet63109842007-05-06 14:49:28 -07001220 if (cachep->shared) {
1221 shared = alloc_arraycache(node,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001222 cachep->shared * cachep->batchcount,
1223 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07001224 if (!shared)
1225 goto bad;
1226 }
Paul Menage3395ee02006-12-06 20:32:16 -08001227 if (use_alien_caches) {
1228 alien = alloc_alien_cache(node, cachep->limit);
1229 if (!alien)
1230 goto bad;
1231 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001233 l3 = cachep->nodelists[node];
1234 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001235
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001236 spin_lock_irq(&l3->list_lock);
1237 if (!l3->shared) {
1238 /*
1239 * We are serialised from CPU_DEAD or
1240 * CPU_UP_CANCELLED by the cpucontrol lock
1241 */
1242 l3->shared = shared;
1243 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001244 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001245#ifdef CONFIG_NUMA
1246 if (!l3->alien) {
1247 l3->alien = alien;
1248 alien = NULL;
1249 }
1250#endif
1251 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001252 kfree(shared);
1253 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 break;
1256 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001257 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 start_cpu_timer(cpu);
1259 break;
1260#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001261 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001262 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001263 /*
1264 * Shutdown cache reaper. Note that the cache_chain_mutex is
1265 * held so that if cache_reap() is invoked it cannot do
1266 * anything expensive but will only modify reap_work
1267 * and reschedule the timer.
1268 */
1269 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1270 /* Now the cache_reaper is guaranteed to be not running. */
1271 per_cpu(reap_work, cpu).work.func = NULL;
1272 break;
1273 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001274 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001275 start_cpu_timer(cpu);
1276 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001278 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001279 /*
1280 * Even if all the cpus of a node are down, we don't free the
1281 * kmem_list3 of any cache. This to avoid a race between
1282 * cpu_down, and a kmalloc allocation from another cpu for
1283 * memory from the node of the cpu going down. The list3
1284 * structure is usually allocated from kmem_cache_create() and
1285 * gets destroyed at kmem_cache_destroy().
1286 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 /* fall thru */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001288#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001290 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 list_for_each_entry(cachep, &cache_chain, next) {
1292 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001293 struct array_cache *shared;
1294 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001295 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296
Christoph Lametere498be72005-09-09 13:03:32 -07001297 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 /* cpu is dead; no one can alloc from it. */
1299 nc = cachep->array[cpu];
1300 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001301 l3 = cachep->nodelists[node];
1302
1303 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001304 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001305
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001306 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001307
1308 /* Free limit for this kmem_list3 */
1309 l3->free_limit -= cachep->batchcount;
1310 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001311 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001312
1313 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001314 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001315 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001316 }
Christoph Lametere498be72005-09-09 13:03:32 -07001317
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001318 shared = l3->shared;
1319 if (shared) {
Eric Dumazet63109842007-05-06 14:49:28 -07001320 free_block(cachep, shared->entry,
1321 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001322 l3->shared = NULL;
1323 }
Christoph Lametere498be72005-09-09 13:03:32 -07001324
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001325 alien = l3->alien;
1326 l3->alien = NULL;
1327
1328 spin_unlock_irq(&l3->list_lock);
1329
1330 kfree(shared);
1331 if (alien) {
1332 drain_alien_cache(cachep, alien);
1333 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001334 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001335free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 kfree(nc);
1337 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001338 /*
1339 * In the previous loop, all the objects were freed to
1340 * the respective cache's slabs, now we can go ahead and
1341 * shrink each nodelist to its limit.
1342 */
1343 list_for_each_entry(cachep, &cache_chain, next) {
1344 l3 = cachep->nodelists[node];
1345 if (!l3)
1346 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001347 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001348 }
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001349 break;
1350 case CPU_LOCK_RELEASE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001351 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 }
1354 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001355bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 return NOTIFY_BAD;
1357}
1358
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001359static struct notifier_block __cpuinitdata cpucache_notifier = {
1360 &cpuup_callback, NULL, 0
1361};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
Christoph Lametere498be72005-09-09 13:03:32 -07001363/*
1364 * swap the static kmem_list3 with kmalloced memory
1365 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001366static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1367 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001368{
1369 struct kmem_list3 *ptr;
1370
Christoph Lametere498be72005-09-09 13:03:32 -07001371 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1372 BUG_ON(!ptr);
1373
1374 local_irq_disable();
1375 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001376 /*
1377 * Do not assume that spinlocks can be initialized via memcpy:
1378 */
1379 spin_lock_init(&ptr->list_lock);
1380
Christoph Lametere498be72005-09-09 13:03:32 -07001381 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1382 cachep->nodelists[nodeid] = ptr;
1383 local_irq_enable();
1384}
1385
Andrew Mortona737b3e2006-03-22 00:08:11 -08001386/*
1387 * Initialisation. Called after the page allocator have been initialised and
1388 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 */
1390void __init kmem_cache_init(void)
1391{
1392 size_t left_over;
1393 struct cache_sizes *sizes;
1394 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001395 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001396 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001397 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001398
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001399 if (num_possible_nodes() == 1)
1400 use_alien_caches = 0;
1401
Christoph Lametere498be72005-09-09 13:03:32 -07001402 for (i = 0; i < NUM_INIT_LISTS; i++) {
1403 kmem_list3_init(&initkmem_list3[i]);
1404 if (i < MAX_NUMNODES)
1405 cache_cache.nodelists[i] = NULL;
1406 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
1408 /*
1409 * Fragmentation resistance on low memory - only use bigger
1410 * page orders on machines with more than 32MB of memory.
1411 */
1412 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1413 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1414
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 /* Bootstrap is tricky, because several objects are allocated
1416 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001417 * 1) initialize the cache_cache cache: it contains the struct
1418 * kmem_cache structures of all caches, except cache_cache itself:
1419 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001420 * Initially an __init data area is used for the head array and the
1421 * kmem_list3 structures, it's replaced with a kmalloc allocated
1422 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001424 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001425 * An __init data area is used for the head array.
1426 * 3) Create the remaining kmalloc caches, with minimally sized
1427 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 * 4) Replace the __init data head arrays for cache_cache and the first
1429 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001430 * 5) Replace the __init data for kmem_list3 for cache_cache and
1431 * the other cache's with kmalloc allocated memory.
1432 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 */
1434
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001435 node = numa_node_id();
1436
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 INIT_LIST_HEAD(&cache_chain);
1439 list_add(&cache_cache.next, &cache_chain);
1440 cache_cache.colour_off = cache_line_size();
1441 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001442 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443
Eric Dumazet8da34302007-05-06 14:49:29 -07001444 /*
1445 * struct kmem_cache size depends on nr_node_ids, which
1446 * can be less than MAX_NUMNODES.
1447 */
1448 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1449 nr_node_ids * sizeof(struct kmem_list3 *);
1450#if DEBUG
1451 cache_cache.obj_size = cache_cache.buffer_size;
1452#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001453 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1454 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001455 cache_cache.reciprocal_buffer_size =
1456 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457
Jack Steiner07ed76b2006-03-07 21:55:46 -08001458 for (order = 0; order < MAX_ORDER; order++) {
1459 cache_estimate(order, cache_cache.buffer_size,
1460 cache_line_size(), 0, &left_over, &cache_cache.num);
1461 if (cache_cache.num)
1462 break;
1463 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001464 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001465 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001466 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001467 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1468 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469
1470 /* 2+3) create the kmalloc caches */
1471 sizes = malloc_sizes;
1472 names = cache_names;
1473
Andrew Mortona737b3e2006-03-22 00:08:11 -08001474 /*
1475 * Initialize the caches that provide memory for the array cache and the
1476 * kmem_list3 structures first. Without this, further allocations will
1477 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001478 */
1479
1480 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001481 sizes[INDEX_AC].cs_size,
1482 ARCH_KMALLOC_MINALIGN,
1483 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1484 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001485
Andrew Mortona737b3e2006-03-22 00:08:11 -08001486 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001487 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001488 kmem_cache_create(names[INDEX_L3].name,
1489 sizes[INDEX_L3].cs_size,
1490 ARCH_KMALLOC_MINALIGN,
1491 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1492 NULL, NULL);
1493 }
Christoph Lametere498be72005-09-09 13:03:32 -07001494
Ingo Molnare0a42722006-06-23 02:03:46 -07001495 slab_early_init = 0;
1496
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001498 /*
1499 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 * This should be particularly beneficial on SMP boxes, as it
1501 * eliminates "false sharing".
1502 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001503 * allow tighter packing of the smaller caches.
1504 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001505 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001506 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001507 sizes->cs_size,
1508 ARCH_KMALLOC_MINALIGN,
1509 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1510 NULL, NULL);
1511 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001512#ifdef CONFIG_ZONE_DMA
1513 sizes->cs_dmacachep = kmem_cache_create(
1514 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001515 sizes->cs_size,
1516 ARCH_KMALLOC_MINALIGN,
1517 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1518 SLAB_PANIC,
1519 NULL, NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001520#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 sizes++;
1522 names++;
1523 }
1524 /* 4) Replace the bootstrap head arrays */
1525 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001526 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001527
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001529
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001531 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1532 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001533 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001534 /*
1535 * Do not assume that spinlocks can be initialized via memcpy:
1536 */
1537 spin_lock_init(&ptr->lock);
1538
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 cache_cache.array[smp_processor_id()] = ptr;
1540 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001541
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001543
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001545 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001546 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001547 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001548 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001549 /*
1550 * Do not assume that spinlocks can be initialized via memcpy:
1551 */
1552 spin_lock_init(&ptr->lock);
1553
Christoph Lametere498be72005-09-09 13:03:32 -07001554 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001555 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 local_irq_enable();
1557 }
Christoph Lametere498be72005-09-09 13:03:32 -07001558 /* 5) Replace the bootstrap kmem_list3's */
1559 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001560 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001562 /* Replace the static kmem_list3 structures for the boot cpu */
1563 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1564
1565 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001566 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001567 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001568
1569 if (INDEX_AC != INDEX_L3) {
1570 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001571 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001572 }
1573 }
1574 }
1575
1576 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001578 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001579 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001581 if (enable_cpucache(cachep))
1582 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001583 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 }
1585
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001586 /* Annotate slab for lockdep -- annotate the malloc caches */
1587 init_lock_keys();
1588
1589
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 /* Done! */
1591 g_cpucache_up = FULL;
1592
Andrew Mortona737b3e2006-03-22 00:08:11 -08001593 /*
1594 * Register a cpu startup notifier callback that initializes
1595 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 */
1597 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598
Andrew Mortona737b3e2006-03-22 00:08:11 -08001599 /*
1600 * The reap timers are started later, with a module init call: That part
1601 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 */
1603}
1604
1605static int __init cpucache_init(void)
1606{
1607 int cpu;
1608
Andrew Mortona737b3e2006-03-22 00:08:11 -08001609 /*
1610 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 */
Christoph Lametere498be72005-09-09 13:03:32 -07001612 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001613 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 return 0;
1615}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616__initcall(cpucache_init);
1617
1618/*
1619 * Interface to system's page allocator. No need to hold the cache-lock.
1620 *
1621 * If we requested dmaable memory, we will get it. Even if we
1622 * did not request dmaable memory, we might get it, but that
1623 * would be relatively rare and ignorable.
1624 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001625static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626{
1627 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001628 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 int i;
1630
Luke Yangd6fef9d2006-04-10 22:52:56 -07001631#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001632 /*
1633 * Nommu uses slab's for process anonymous memory allocations, and thus
1634 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001635 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001636 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001637#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001638
Christoph Lameter3c517a62006-12-06 20:33:29 -08001639 flags |= cachep->gfpflags;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001640
1641 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 if (!page)
1643 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001645 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001647 add_zone_page_state(page_zone(page),
1648 NR_SLAB_RECLAIMABLE, nr_pages);
1649 else
1650 add_zone_page_state(page_zone(page),
1651 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001652 for (i = 0; i < nr_pages; i++)
1653 __SetPageSlab(page + i);
1654 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655}
1656
1657/*
1658 * Interface to system's page release.
1659 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001660static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001662 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 struct page *page = virt_to_page(addr);
1664 const unsigned long nr_freed = i;
1665
Christoph Lameter972d1a72006-09-25 23:31:51 -07001666 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1667 sub_zone_page_state(page_zone(page),
1668 NR_SLAB_RECLAIMABLE, nr_freed);
1669 else
1670 sub_zone_page_state(page_zone(page),
1671 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001673 BUG_ON(!PageSlab(page));
1674 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 page++;
1676 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 if (current->reclaim_state)
1678 current->reclaim_state->reclaimed_slab += nr_freed;
1679 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680}
1681
1682static void kmem_rcu_free(struct rcu_head *head)
1683{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001684 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001685 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
1687 kmem_freepages(cachep, slab_rcu->addr);
1688 if (OFF_SLAB(cachep))
1689 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1690}
1691
1692#if DEBUG
1693
1694#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001695static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001696 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001698 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001700 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001702 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 return;
1704
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001705 *addr++ = 0x12345678;
1706 *addr++ = caller;
1707 *addr++ = smp_processor_id();
1708 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 {
1710 unsigned long *sptr = &caller;
1711 unsigned long svalue;
1712
1713 while (!kstack_end(sptr)) {
1714 svalue = *sptr++;
1715 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001716 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 size -= sizeof(unsigned long);
1718 if (size <= sizeof(unsigned long))
1719 break;
1720 }
1721 }
1722
1723 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001724 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725}
1726#endif
1727
Pekka Enberg343e0d72006-02-01 03:05:50 -08001728static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001730 int size = obj_size(cachep);
1731 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
1733 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001734 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735}
1736
1737static void dump_line(char *data, int offset, int limit)
1738{
1739 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001740 unsigned char error = 0;
1741 int bad_count = 0;
1742
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001744 for (i = 0; i < limit; i++) {
1745 if (data[offset + i] != POISON_FREE) {
1746 error = data[offset + i];
1747 bad_count++;
1748 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001749 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001752
1753 if (bad_count == 1) {
1754 error ^= POISON_FREE;
1755 if (!(error & (error - 1))) {
1756 printk(KERN_ERR "Single bit error detected. Probably "
1757 "bad RAM.\n");
1758#ifdef CONFIG_X86
1759 printk(KERN_ERR "Run memtest86+ or a similar memory "
1760 "test tool.\n");
1761#else
1762 printk(KERN_ERR "Run a memory test tool.\n");
1763#endif
1764 }
1765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766}
1767#endif
1768
1769#if DEBUG
1770
Pekka Enberg343e0d72006-02-01 03:05:50 -08001771static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772{
1773 int i, size;
1774 char *realobj;
1775
1776 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001777 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001778 *dbg_redzone1(cachep, objp),
1779 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 }
1781
1782 if (cachep->flags & SLAB_STORE_USER) {
1783 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001784 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001786 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 printk("\n");
1788 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001789 realobj = (char *)objp + obj_offset(cachep);
1790 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 int limit;
1793 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001794 if (i + limit > size)
1795 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 dump_line(realobj, i, limit);
1797 }
1798}
1799
Pekka Enberg343e0d72006-02-01 03:05:50 -08001800static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801{
1802 char *realobj;
1803 int size, i;
1804 int lines = 0;
1805
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001806 realobj = (char *)objp + obj_offset(cachep);
1807 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001809 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001811 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 exp = POISON_END;
1813 if (realobj[i] != exp) {
1814 int limit;
1815 /* Mismatch ! */
1816 /* Print header */
1817 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001818 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001819 "Slab corruption: %s start=%p, len=%d\n",
1820 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 print_objinfo(cachep, objp, 0);
1822 }
1823 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001824 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001826 if (i + limit > size)
1827 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 dump_line(realobj, i, limit);
1829 i += 16;
1830 lines++;
1831 /* Limit to 5 lines */
1832 if (lines > 5)
1833 break;
1834 }
1835 }
1836 if (lines != 0) {
1837 /* Print some data about the neighboring objects, if they
1838 * exist:
1839 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001840 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001841 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001843 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001845 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001846 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001848 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 print_objinfo(cachep, objp, 2);
1850 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001852 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001853 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001855 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 print_objinfo(cachep, objp, 2);
1857 }
1858 }
1859}
1860#endif
1861
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001863/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001864 * slab_destroy_objs - destroy a slab and its objects
1865 * @cachep: cache pointer being destroyed
1866 * @slabp: slab pointer being destroyed
1867 *
1868 * Call the registered destructor for each object in a slab that is being
1869 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001870 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001871static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001872{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 int i;
1874 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001875 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876
1877 if (cachep->flags & SLAB_POISON) {
1878#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001879 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1880 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001881 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001882 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 else
1884 check_poison_obj(cachep, objp);
1885#else
1886 check_poison_obj(cachep, objp);
1887#endif
1888 }
1889 if (cachep->flags & SLAB_RED_ZONE) {
1890 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1891 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001892 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1894 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001898}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001900static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001901{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001902}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903#endif
1904
Randy Dunlap911851e2006-03-22 00:08:14 -08001905/**
1906 * slab_destroy - destroy and release all objects in a slab
1907 * @cachep: cache pointer being destroyed
1908 * @slabp: slab pointer being destroyed
1909 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001910 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001911 * Before calling the slab must have been unlinked from the cache. The
1912 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001913 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001914static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001915{
1916 void *addr = slabp->s_mem - slabp->colouroff;
1917
1918 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1920 struct slab_rcu *slab_rcu;
1921
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 slab_rcu->cachep = cachep;
1924 slab_rcu->addr = addr;
1925 call_rcu(&slab_rcu->head, kmem_rcu_free);
1926 } else {
1927 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001928 if (OFF_SLAB(cachep))
1929 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 }
1931}
1932
Andrew Mortona737b3e2006-03-22 00:08:11 -08001933/*
1934 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1935 * size of kmem_list3.
1936 */
Andrew Mortona3a02be2007-05-06 14:49:31 -07001937static void __init set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001938{
1939 int node;
1940
1941 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001942 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001943 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001944 REAPTIMEOUT_LIST3 +
1945 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001946 }
1947}
1948
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001949static void __kmem_cache_destroy(struct kmem_cache *cachep)
1950{
1951 int i;
1952 struct kmem_list3 *l3;
1953
1954 for_each_online_cpu(i)
1955 kfree(cachep->array[i]);
1956
1957 /* NUMA: free the list3 structures */
1958 for_each_online_node(i) {
1959 l3 = cachep->nodelists[i];
1960 if (l3) {
1961 kfree(l3->shared);
1962 free_alien_cache(l3->alien);
1963 kfree(l3);
1964 }
1965 }
1966 kmem_cache_free(&cache_cache, cachep);
1967}
1968
1969
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001971 * calculate_slab_order - calculate size (page order) of slabs
1972 * @cachep: pointer to the cache that is being created
1973 * @size: size of objects to be created in this cache.
1974 * @align: required alignment for the objects.
1975 * @flags: slab allocation flags
1976 *
1977 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001978 *
1979 * This could be made much more intelligent. For now, try to avoid using
1980 * high order pages for slabs. When the gfp() functions are more friendly
1981 * towards high-order requests, this should be changed.
1982 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001983static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001984 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001985{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001986 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001987 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001988 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001989
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001990 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001991 unsigned int num;
1992 size_t remainder;
1993
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001994 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001995 if (!num)
1996 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001997
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001998 if (flags & CFLGS_OFF_SLAB) {
1999 /*
2000 * Max number of objs-per-slab for caches which
2001 * use off-slab slabs. Needed to avoid a possible
2002 * looping condition in cache_grow().
2003 */
2004 offslab_limit = size - sizeof(struct slab);
2005 offslab_limit /= sizeof(kmem_bufctl_t);
2006
2007 if (num > offslab_limit)
2008 break;
2009 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002011 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002013 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002014 left_over = remainder;
2015
2016 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002017 * A VFS-reclaimable slab tends to have most allocations
2018 * as GFP_NOFS and we really don't want to have to be allocating
2019 * higher-order pages when we are unable to shrink dcache.
2020 */
2021 if (flags & SLAB_RECLAIM_ACCOUNT)
2022 break;
2023
2024 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002025 * Large number of objects is good, but very large slabs are
2026 * currently bad for the gfp()s.
2027 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002028 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002029 break;
2030
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002031 /*
2032 * Acceptable internal fragmentation?
2033 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002034 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002035 break;
2036 }
2037 return left_over;
2038}
2039
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002040static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002041{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002042 if (g_cpucache_up == FULL)
2043 return enable_cpucache(cachep);
2044
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002045 if (g_cpucache_up == NONE) {
2046 /*
2047 * Note: the first kmem_cache_create must create the cache
2048 * that's used by kmalloc(24), otherwise the creation of
2049 * further caches will BUG().
2050 */
2051 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2052
2053 /*
2054 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2055 * the first cache, then we need to set up all its list3s,
2056 * otherwise the creation of further caches will BUG().
2057 */
2058 set_up_list3s(cachep, SIZE_AC);
2059 if (INDEX_AC == INDEX_L3)
2060 g_cpucache_up = PARTIAL_L3;
2061 else
2062 g_cpucache_up = PARTIAL_AC;
2063 } else {
2064 cachep->array[smp_processor_id()] =
2065 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2066
2067 if (g_cpucache_up == PARTIAL_AC) {
2068 set_up_list3s(cachep, SIZE_L3);
2069 g_cpucache_up = PARTIAL_L3;
2070 } else {
2071 int node;
2072 for_each_online_node(node) {
2073 cachep->nodelists[node] =
2074 kmalloc_node(sizeof(struct kmem_list3),
2075 GFP_KERNEL, node);
2076 BUG_ON(!cachep->nodelists[node]);
2077 kmem_list3_init(cachep->nodelists[node]);
2078 }
2079 }
2080 }
2081 cachep->nodelists[numa_node_id()]->next_reap =
2082 jiffies + REAPTIMEOUT_LIST3 +
2083 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2084
2085 cpu_cache_get(cachep)->avail = 0;
2086 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2087 cpu_cache_get(cachep)->batchcount = 1;
2088 cpu_cache_get(cachep)->touched = 0;
2089 cachep->batchcount = 1;
2090 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002091 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002092}
2093
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002094/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 * kmem_cache_create - Create a cache.
2096 * @name: A string which is used in /proc/slabinfo to identify this cache.
2097 * @size: The size of objects to be created in this cache.
2098 * @align: The required alignment for the objects.
2099 * @flags: SLAB flags
2100 * @ctor: A constructor for the objects.
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002101 * @dtor: A destructor for the objects (not implemented anymore).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 *
2103 * Returns a ptr to the cache on success, NULL on failure.
2104 * Cannot be called within a int, but can be interrupted.
2105 * The @ctor is run when new pages are allocated by the cache
2106 * and the @dtor is run before the pages are handed back.
2107 *
2108 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002109 * the module calling this has to destroy the cache before getting unloaded.
2110 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 * The flags are
2112 *
2113 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2114 * to catch references to uninitialised memory.
2115 *
2116 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2117 * for buffer overruns.
2118 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2120 * cacheline. This can be beneficial if you're counting cycles as closely
2121 * as davem.
2122 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002123struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002125 unsigned long flags,
2126 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002127 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128{
2129 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002130 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
2132 /*
2133 * Sanity checks... these are all serious usage bugs.
2134 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002135 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002136 size > KMALLOC_MAX_SIZE || dtor) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002137 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2138 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002139 BUG();
2140 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002142 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002143 * We use cache_chain_mutex to ensure a consistent view of
2144 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002145 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002146 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002147
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002148 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002149 char tmp;
2150 int res;
2151
2152 /*
2153 * This happens when the module gets unloaded and doesn't
2154 * destroy its slab cache and no-one else reuses the vmalloc
2155 * area of the module. Print a warning.
2156 */
Andrew Morton138ae662006-12-06 20:36:41 -08002157 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002158 if (res) {
matzeb4169522007-05-06 14:49:52 -07002159 printk(KERN_ERR
2160 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002161 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002162 continue;
2163 }
2164
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002165 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002166 printk(KERN_ERR
2167 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002168 dump_stack();
2169 goto oops;
2170 }
2171 }
2172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173#if DEBUG
2174 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175#if FORCED_DEBUG
2176 /*
2177 * Enable redzoning and last user accounting, except for caches with
2178 * large objects, if the increased size would increase the object size
2179 * above the next power of two: caches with object sizes just above a
2180 * power of two have a significant amount of internal fragmentation.
2181 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002182 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2183 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002184 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 if (!(flags & SLAB_DESTROY_BY_RCU))
2186 flags |= SLAB_POISON;
2187#endif
2188 if (flags & SLAB_DESTROY_BY_RCU)
2189 BUG_ON(flags & SLAB_POISON);
2190#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002192 * Always checks flags, a caller might be expecting debug support which
2193 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002195 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
Andrew Mortona737b3e2006-03-22 00:08:11 -08002197 /*
2198 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 * unaligned accesses for some archs when redzoning is used, and makes
2200 * sure any on-slab bufctl's are also correctly aligned.
2201 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002202 if (size & (BYTES_PER_WORD - 1)) {
2203 size += (BYTES_PER_WORD - 1);
2204 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 }
2206
Andrew Mortona737b3e2006-03-22 00:08:11 -08002207 /* calculate the final buffer alignment: */
2208
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 /* 1) arch recommendation: can be overridden for debug */
2210 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002211 /*
2212 * Default alignment: as specified by the arch code. Except if
2213 * an object is really small, then squeeze multiple objects into
2214 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 */
2216 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002217 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 ralign /= 2;
2219 } else {
2220 ralign = BYTES_PER_WORD;
2221 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002222
2223 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002224 * Redzoning and user store require word alignment or possibly larger.
2225 * Note this will be overridden by architecture or caller mandated
2226 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002227 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002228 if (flags & SLAB_STORE_USER)
2229 ralign = BYTES_PER_WORD;
2230
2231 if (flags & SLAB_RED_ZONE) {
2232 ralign = REDZONE_ALIGN;
2233 /* If redzoning, ensure that the second redzone is suitably
2234 * aligned, by adjusting the object size accordingly. */
2235 size += REDZONE_ALIGN - 1;
2236 size &= ~(REDZONE_ALIGN - 1);
2237 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002238
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002239 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 if (ralign < ARCH_SLAB_MINALIGN) {
2241 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002243 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 if (ralign < align) {
2245 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002247 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002248 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002249 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002250 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002251 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 */
2253 align = ralign;
2254
2255 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002256 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002258 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
2260#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002261 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262
Pekka Enbergca5f9702006-09-25 23:31:25 -07002263 /*
2264 * Both debugging options require word-alignment which is calculated
2265 * into align above.
2266 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002269 cachep->obj_offset += sizeof(unsigned long long);
2270 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 }
2272 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002273 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002274 * the real object. But if the second red zone needs to be
2275 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002277 if (flags & SLAB_RED_ZONE)
2278 size += REDZONE_ALIGN;
2279 else
2280 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 }
2282#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002283 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002284 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2285 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 size = PAGE_SIZE;
2287 }
2288#endif
2289#endif
2290
Ingo Molnare0a42722006-06-23 02:03:46 -07002291 /*
2292 * Determine if the slab management is 'on' or 'off' slab.
2293 * (bootstrapping cannot cope with offslab caches so don't do
2294 * it too early on.)
2295 */
2296 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 /*
2298 * Size is large, assume best to place the slab management obj
2299 * off-slab (should allow better packing of objs).
2300 */
2301 flags |= CFLGS_OFF_SLAB;
2302
2303 size = ALIGN(size, align);
2304
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002305 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306
2307 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002308 printk(KERN_ERR
2309 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 kmem_cache_free(&cache_cache, cachep);
2311 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002312 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002314 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2315 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316
2317 /*
2318 * If the slab has been placed off-slab, and we have enough space then
2319 * move it on-slab. This is at the expense of any extra colouring.
2320 */
2321 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2322 flags &= ~CFLGS_OFF_SLAB;
2323 left_over -= slab_size;
2324 }
2325
2326 if (flags & CFLGS_OFF_SLAB) {
2327 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002328 slab_size =
2329 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 }
2331
2332 cachep->colour_off = cache_line_size();
2333 /* Offset must be a multiple of the alignment. */
2334 if (cachep->colour_off < align)
2335 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002336 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 cachep->slab_size = slab_size;
2338 cachep->flags = flags;
2339 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002340 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002342 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002343 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002345 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002346 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002347 /*
2348 * This is a possibility for one of the malloc_sizes caches.
2349 * But since we go off slab only for object size greater than
2350 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2351 * this should not happen at all.
2352 * But leave a BUG_ON for some lucky dude.
2353 */
2354 BUG_ON(!cachep->slabp_cache);
2355 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 cachep->name = name;
2358
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002359 if (setup_cpu_cache(cachep)) {
2360 __kmem_cache_destroy(cachep);
2361 cachep = NULL;
2362 goto oops;
2363 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 /* cache setup completed, link it into the list */
2366 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002367oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 if (!cachep && (flags & SLAB_PANIC))
2369 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002370 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002371 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 return cachep;
2373}
2374EXPORT_SYMBOL(kmem_cache_create);
2375
2376#if DEBUG
2377static void check_irq_off(void)
2378{
2379 BUG_ON(!irqs_disabled());
2380}
2381
2382static void check_irq_on(void)
2383{
2384 BUG_ON(irqs_disabled());
2385}
2386
Pekka Enberg343e0d72006-02-01 03:05:50 -08002387static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388{
2389#ifdef CONFIG_SMP
2390 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002391 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392#endif
2393}
Christoph Lametere498be72005-09-09 13:03:32 -07002394
Pekka Enberg343e0d72006-02-01 03:05:50 -08002395static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002396{
2397#ifdef CONFIG_SMP
2398 check_irq_off();
2399 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2400#endif
2401}
2402
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403#else
2404#define check_irq_off() do { } while(0)
2405#define check_irq_on() do { } while(0)
2406#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002407#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408#endif
2409
Christoph Lameteraab22072006-03-22 00:09:06 -08002410static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2411 struct array_cache *ac,
2412 int force, int node);
2413
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414static void do_drain(void *arg)
2415{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002416 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002418 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
2420 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002421 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002422 spin_lock(&cachep->nodelists[node]->list_lock);
2423 free_block(cachep, ac->entry, ac->avail, node);
2424 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 ac->avail = 0;
2426}
2427
Pekka Enberg343e0d72006-02-01 03:05:50 -08002428static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429{
Christoph Lametere498be72005-09-09 13:03:32 -07002430 struct kmem_list3 *l3;
2431 int node;
2432
Andrew Mortona07fa392006-03-22 00:08:17 -08002433 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002435 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002436 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002437 if (l3 && l3->alien)
2438 drain_alien_cache(cachep, l3->alien);
2439 }
2440
2441 for_each_online_node(node) {
2442 l3 = cachep->nodelists[node];
2443 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002444 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446}
2447
Christoph Lametered11d9e2006-06-30 01:55:45 -07002448/*
2449 * Remove slabs from the list of free slabs.
2450 * Specify the number of slabs to drain in tofree.
2451 *
2452 * Returns the actual number of slabs released.
2453 */
2454static int drain_freelist(struct kmem_cache *cache,
2455 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002457 struct list_head *p;
2458 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
Christoph Lametered11d9e2006-06-30 01:55:45 -07002461 nr_freed = 0;
2462 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463
Christoph Lametered11d9e2006-06-30 01:55:45 -07002464 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002465 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002466 if (p == &l3->slabs_free) {
2467 spin_unlock_irq(&l3->list_lock);
2468 goto out;
2469 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470
Christoph Lametered11d9e2006-06-30 01:55:45 -07002471 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002473 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474#endif
2475 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002476 /*
2477 * Safe to drop the lock. The slab is no longer linked
2478 * to the cache.
2479 */
2480 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002481 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002482 slab_destroy(cache, slabp);
2483 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002485out:
2486 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487}
2488
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002489/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002490static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002491{
2492 int ret = 0, i = 0;
2493 struct kmem_list3 *l3;
2494
2495 drain_cpu_caches(cachep);
2496
2497 check_irq_on();
2498 for_each_online_node(i) {
2499 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002500 if (!l3)
2501 continue;
2502
2503 drain_freelist(cachep, l3, l3->free_objects);
2504
2505 ret += !list_empty(&l3->slabs_full) ||
2506 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002507 }
2508 return (ret ? 1 : 0);
2509}
2510
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511/**
2512 * kmem_cache_shrink - Shrink a cache.
2513 * @cachep: The cache to shrink.
2514 *
2515 * Releases as many slabs as possible for a cache.
2516 * To help debugging, a zero exit status indicates all slabs were released.
2517 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002518int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002520 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002521 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002523 mutex_lock(&cache_chain_mutex);
2524 ret = __cache_shrink(cachep);
2525 mutex_unlock(&cache_chain_mutex);
2526 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527}
2528EXPORT_SYMBOL(kmem_cache_shrink);
2529
2530/**
2531 * kmem_cache_destroy - delete a cache
2532 * @cachep: the cache to destroy
2533 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002534 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 *
2536 * It is expected this function will be called by a module when it is
2537 * unloaded. This will remove the cache completely, and avoid a duplicate
2538 * cache being allocated each time a module is loaded and unloaded, if the
2539 * module doesn't have persistent in-kernel storage across loads and unloads.
2540 *
2541 * The cache must be empty before calling this function.
2542 *
2543 * The caller must guarantee that noone will allocate memory from the cache
2544 * during the kmem_cache_destroy().
2545 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002546void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002548 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002551 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 /*
2553 * the chain is never empty, cache_cache is never destroyed
2554 */
2555 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 if (__cache_shrink(cachep)) {
2557 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002558 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002559 mutex_unlock(&cache_chain_mutex);
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002560 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 }
2562
2563 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002564 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002566 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002567 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568}
2569EXPORT_SYMBOL(kmem_cache_destroy);
2570
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002571/*
2572 * Get the memory for a slab management obj.
2573 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2574 * always come from malloc_sizes caches. The slab descriptor cannot
2575 * come from the same cache which is getting created because,
2576 * when we are searching for an appropriate cache for these
2577 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2578 * If we are creating a malloc_sizes cache here it would not be visible to
2579 * kmem_find_general_cachep till the initialization is complete.
2580 * Hence we cannot have slabp_cache same as the original cache.
2581 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002582static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002583 int colour_off, gfp_t local_flags,
2584 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585{
2586 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002587
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 if (OFF_SLAB(cachep)) {
2589 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002590 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002591 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 if (!slabp)
2593 return NULL;
2594 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002595 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 colour_off += cachep->slab_size;
2597 }
2598 slabp->inuse = 0;
2599 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002600 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002601 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 return slabp;
2603}
2604
2605static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2606{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002607 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608}
2609
Pekka Enberg343e0d72006-02-01 03:05:50 -08002610static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002611 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612{
2613 int i;
2614
2615 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002616 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617#if DEBUG
2618 /* need to poison the objs? */
2619 if (cachep->flags & SLAB_POISON)
2620 poison_obj(cachep, objp, POISON_FREE);
2621 if (cachep->flags & SLAB_STORE_USER)
2622 *dbg_userword(cachep, objp) = NULL;
2623
2624 if (cachep->flags & SLAB_RED_ZONE) {
2625 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2626 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2627 }
2628 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002629 * Constructors are not allowed to allocate memory from the same
2630 * cache which they are a constructor for. Otherwise, deadlock.
2631 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 */
2633 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002634 cachep->ctor(objp + obj_offset(cachep), cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002635 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
2637 if (cachep->flags & SLAB_RED_ZONE) {
2638 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2639 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2642 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002643 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002645 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2646 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002647 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002648 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649#else
2650 if (cachep->ctor)
Christoph Lametera35afb82007-05-16 22:10:57 -07002651 cachep->ctor(objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002653 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002655 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 slabp->free = 0;
2657}
2658
Pekka Enberg343e0d72006-02-01 03:05:50 -08002659static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002661 if (CONFIG_ZONE_DMA_FLAG) {
2662 if (flags & GFP_DMA)
2663 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2664 else
2665 BUG_ON(cachep->gfpflags & GFP_DMA);
2666 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667}
2668
Andrew Mortona737b3e2006-03-22 00:08:11 -08002669static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2670 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002671{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002672 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002673 kmem_bufctl_t next;
2674
2675 slabp->inuse++;
2676 next = slab_bufctl(slabp)[slabp->free];
2677#if DEBUG
2678 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2679 WARN_ON(slabp->nodeid != nodeid);
2680#endif
2681 slabp->free = next;
2682
2683 return objp;
2684}
2685
Andrew Mortona737b3e2006-03-22 00:08:11 -08002686static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2687 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002688{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002689 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002690
2691#if DEBUG
2692 /* Verify that the slab belongs to the intended node */
2693 WARN_ON(slabp->nodeid != nodeid);
2694
Al Viro871751e2006-03-25 03:06:39 -08002695 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002696 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002697 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002698 BUG();
2699 }
2700#endif
2701 slab_bufctl(slabp)[objnr] = slabp->free;
2702 slabp->free = objnr;
2703 slabp->inuse--;
2704}
2705
Pekka Enberg47768742006-06-23 02:03:07 -07002706/*
2707 * Map pages beginning at addr to the given cache and slab. This is required
2708 * for the slab allocator to be able to lookup the cache and slab of a
2709 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2710 */
2711static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2712 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713{
Pekka Enberg47768742006-06-23 02:03:07 -07002714 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 struct page *page;
2716
Pekka Enberg47768742006-06-23 02:03:07 -07002717 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002718
Pekka Enberg47768742006-06-23 02:03:07 -07002719 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002720 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002721 nr_pages <<= cache->gfporder;
2722
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002724 page_set_cache(page, cache);
2725 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002727 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728}
2729
2730/*
2731 * Grow (by 1) the number of slabs within a cache. This is called by
2732 * kmem_cache_alloc() when there are no active objs left in a cache.
2733 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002734static int cache_grow(struct kmem_cache *cachep,
2735 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002737 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002738 size_t offset;
2739 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002740 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
Andrew Mortona737b3e2006-03-22 00:08:11 -08002742 /*
2743 * Be lazy and only check for valid flags here, keeping it out of the
2744 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 */
Christoph Lametercfce6602007-05-06 14:50:17 -07002746 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
Christoph Lametera06d72c2006-12-06 20:33:12 -08002748 local_flags = (flags & GFP_LEVEL_MASK);
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002749 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002751 l3 = cachep->nodelists[nodeid];
2752 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753
2754 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002755 offset = l3->colour_next;
2756 l3->colour_next++;
2757 if (l3->colour_next >= cachep->colour)
2758 l3->colour_next = 0;
2759 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002761 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762
2763 if (local_flags & __GFP_WAIT)
2764 local_irq_enable();
2765
2766 /*
2767 * The test for missing atomic flag is performed here, rather than
2768 * the more obvious place, simply to reduce the critical path length
2769 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2770 * will eventually be caught here (where it matters).
2771 */
2772 kmem_flagcheck(cachep, flags);
2773
Andrew Mortona737b3e2006-03-22 00:08:11 -08002774 /*
2775 * Get mem for the objs. Attempt to allocate a physical page from
2776 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002777 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002778 if (!objp)
2779 objp = kmem_getpages(cachep, flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002780 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 goto failed;
2782
2783 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002784 slabp = alloc_slabmgmt(cachep, objp, offset,
2785 local_flags & ~GFP_THISNODE, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002786 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 goto opps1;
2788
Christoph Lametere498be72005-09-09 13:03:32 -07002789 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002790 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
Christoph Lametera35afb82007-05-16 22:10:57 -07002792 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
2794 if (local_flags & __GFP_WAIT)
2795 local_irq_disable();
2796 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002797 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798
2799 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002800 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002802 l3->free_objects += cachep->num;
2803 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002805opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002807failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 if (local_flags & __GFP_WAIT)
2809 local_irq_disable();
2810 return 0;
2811}
2812
2813#if DEBUG
2814
2815/*
2816 * Perform extra freeing checks:
2817 * - detect bad pointers.
2818 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 */
2820static void kfree_debugcheck(const void *objp)
2821{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 if (!virt_addr_valid(objp)) {
2823 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002824 (unsigned long)objp);
2825 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827}
2828
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002829static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2830{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002831 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002832
2833 redzone1 = *dbg_redzone1(cache, obj);
2834 redzone2 = *dbg_redzone2(cache, obj);
2835
2836 /*
2837 * Redzone is ok.
2838 */
2839 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2840 return;
2841
2842 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2843 slab_error(cache, "double free detected");
2844 else
2845 slab_error(cache, "memory outside object was overwritten");
2846
David Woodhouseb46b8f12007-05-08 00:22:59 -07002847 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002848 obj, redzone1, redzone2);
2849}
2850
Pekka Enberg343e0d72006-02-01 03:05:50 -08002851static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002852 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853{
2854 struct page *page;
2855 unsigned int objnr;
2856 struct slab *slabp;
2857
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002858 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002860 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861
Pekka Enberg065d41c2005-11-13 16:06:46 -08002862 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863
2864 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002865 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2867 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2868 }
2869 if (cachep->flags & SLAB_STORE_USER)
2870 *dbg_userword(cachep, objp) = caller;
2871
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002872 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
2874 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002875 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876
Al Viro871751e2006-03-25 03:06:39 -08002877#ifdef CONFIG_DEBUG_SLAB_LEAK
2878 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2879#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 if (cachep->flags & SLAB_POISON) {
2881#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002882 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002884 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002885 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 } else {
2887 poison_obj(cachep, objp, POISON_FREE);
2888 }
2889#else
2890 poison_obj(cachep, objp, POISON_FREE);
2891#endif
2892 }
2893 return objp;
2894}
2895
Pekka Enberg343e0d72006-02-01 03:05:50 -08002896static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897{
2898 kmem_bufctl_t i;
2899 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002900
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 /* Check slab's freelist to see if this obj is there. */
2902 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2903 entries++;
2904 if (entries > cachep->num || i >= cachep->num)
2905 goto bad;
2906 }
2907 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002908bad:
2909 printk(KERN_ERR "slab: Internal list corruption detected in "
2910 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2911 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002912 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002913 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002914 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002915 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002917 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 }
2919 printk("\n");
2920 BUG();
2921 }
2922}
2923#else
2924#define kfree_debugcheck(x) do { } while(0)
2925#define cache_free_debugcheck(x,objp,z) (objp)
2926#define check_slabp(x,y) do { } while(0)
2927#endif
2928
Pekka Enberg343e0d72006-02-01 03:05:50 -08002929static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930{
2931 int batchcount;
2932 struct kmem_list3 *l3;
2933 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002934 int node;
2935
2936 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937
2938 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002939 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002940retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 batchcount = ac->batchcount;
2942 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002943 /*
2944 * If there was little recent activity on this cache, then
2945 * perform only a partial refill. Otherwise we could generate
2946 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 */
2948 batchcount = BATCHREFILL_LIMIT;
2949 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002950 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951
Christoph Lametere498be72005-09-09 13:03:32 -07002952 BUG_ON(ac->avail > 0 || !l3);
2953 spin_lock(&l3->list_lock);
2954
Christoph Lameter3ded1752006-03-25 03:06:44 -08002955 /* See if we can refill from the shared array */
2956 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2957 goto alloc_done;
2958
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 while (batchcount > 0) {
2960 struct list_head *entry;
2961 struct slab *slabp;
2962 /* Get slab alloc is to come from. */
2963 entry = l3->slabs_partial.next;
2964 if (entry == &l3->slabs_partial) {
2965 l3->free_touched = 1;
2966 entry = l3->slabs_free.next;
2967 if (entry == &l3->slabs_free)
2968 goto must_grow;
2969 }
2970
2971 slabp = list_entry(entry, struct slab, list);
2972 check_slabp(cachep, slabp);
2973 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07002974
2975 /*
2976 * The slab was either on partial or free list so
2977 * there must be at least one object available for
2978 * allocation.
2979 */
2980 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2981
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 STATS_INC_ALLOCED(cachep);
2984 STATS_INC_ACTIVE(cachep);
2985 STATS_SET_HIGH(cachep);
2986
Matthew Dobson78d382d2006-02-01 03:05:47 -08002987 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002988 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 }
2990 check_slabp(cachep, slabp);
2991
2992 /* move slabp to correct slabp list: */
2993 list_del(&slabp->list);
2994 if (slabp->free == BUFCTL_END)
2995 list_add(&slabp->list, &l3->slabs_full);
2996 else
2997 list_add(&slabp->list, &l3->slabs_partial);
2998 }
2999
Andrew Mortona737b3e2006-03-22 00:08:11 -08003000must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003002alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003003 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004
3005 if (unlikely(!ac->avail)) {
3006 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003007 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003008
Andrew Mortona737b3e2006-03-22 00:08:11 -08003009 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003010 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003011 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 return NULL;
3013
Andrew Mortona737b3e2006-03-22 00:08:11 -08003014 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 goto retry;
3016 }
3017 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003018 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019}
3020
Andrew Mortona737b3e2006-03-22 00:08:11 -08003021static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3022 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023{
3024 might_sleep_if(flags & __GFP_WAIT);
3025#if DEBUG
3026 kmem_flagcheck(cachep, flags);
3027#endif
3028}
3029
3030#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3032 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003034 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003036 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003038 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003039 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003040 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 else
3042 check_poison_obj(cachep, objp);
3043#else
3044 check_poison_obj(cachep, objp);
3045#endif
3046 poison_obj(cachep, objp, POISON_INUSE);
3047 }
3048 if (cachep->flags & SLAB_STORE_USER)
3049 *dbg_userword(cachep, objp) = caller;
3050
3051 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003052 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3053 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3054 slab_error(cachep, "double free, or memory outside"
3055 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003056 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003057 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003058 objp, *dbg_redzone1(cachep, objp),
3059 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 }
3061 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3062 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3063 }
Al Viro871751e2006-03-25 03:06:39 -08003064#ifdef CONFIG_DEBUG_SLAB_LEAK
3065 {
3066 struct slab *slabp;
3067 unsigned objnr;
3068
Christoph Lameterb49af682007-05-06 14:49:41 -07003069 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003070 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3071 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3072 }
3073#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003074 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003075 if (cachep->ctor && cachep->flags & SLAB_POISON)
Christoph Lametera35afb82007-05-16 22:10:57 -07003076 cachep->ctor(objp, cachep, 0);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003077#if ARCH_SLAB_MINALIGN
3078 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3079 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3080 objp, ARCH_SLAB_MINALIGN);
3081 }
3082#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 return objp;
3084}
3085#else
3086#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3087#endif
3088
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003089#ifdef CONFIG_FAILSLAB
3090
3091static struct failslab_attr {
3092
3093 struct fault_attr attr;
3094
3095 u32 ignore_gfp_wait;
3096#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3097 struct dentry *ignore_gfp_wait_file;
3098#endif
3099
3100} failslab = {
3101 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003102 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003103};
3104
3105static int __init setup_failslab(char *str)
3106{
3107 return setup_fault_attr(&failslab.attr, str);
3108}
3109__setup("failslab=", setup_failslab);
3110
3111static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3112{
3113 if (cachep == &cache_cache)
3114 return 0;
3115 if (flags & __GFP_NOFAIL)
3116 return 0;
3117 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3118 return 0;
3119
3120 return should_fail(&failslab.attr, obj_size(cachep));
3121}
3122
3123#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3124
3125static int __init failslab_debugfs(void)
3126{
3127 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3128 struct dentry *dir;
3129 int err;
3130
Akinobu Mita824ebef2007-05-06 14:49:58 -07003131 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003132 if (err)
3133 return err;
3134 dir = failslab.attr.dentries.dir;
3135
3136 failslab.ignore_gfp_wait_file =
3137 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3138 &failslab.ignore_gfp_wait);
3139
3140 if (!failslab.ignore_gfp_wait_file) {
3141 err = -ENOMEM;
3142 debugfs_remove(failslab.ignore_gfp_wait_file);
3143 cleanup_fault_attr_dentries(&failslab.attr);
3144 }
3145
3146 return err;
3147}
3148
3149late_initcall(failslab_debugfs);
3150
3151#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3152
3153#else /* CONFIG_FAILSLAB */
3154
3155static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3156{
3157 return 0;
3158}
3159
3160#endif /* CONFIG_FAILSLAB */
3161
Pekka Enberg343e0d72006-02-01 03:05:50 -08003162static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003164 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 struct array_cache *ac;
3166
Alok N Kataria5c382302005-09-27 21:45:46 -07003167 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003168
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003169 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170 if (likely(ac->avail)) {
3171 STATS_INC_ALLOCHIT(cachep);
3172 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003173 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 } else {
3175 STATS_INC_ALLOCMISS(cachep);
3176 objp = cache_alloc_refill(cachep, flags);
3177 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003178 return objp;
3179}
3180
Christoph Lametere498be72005-09-09 13:03:32 -07003181#ifdef CONFIG_NUMA
3182/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003183 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003184 *
3185 * If we are in_interrupt, then process context, including cpusets and
3186 * mempolicy, may not apply and should not be used for allocation policy.
3187 */
3188static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3189{
3190 int nid_alloc, nid_here;
3191
Christoph Lameter765c4502006-09-27 01:50:08 -07003192 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003193 return NULL;
3194 nid_alloc = nid_here = numa_node_id();
3195 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3196 nid_alloc = cpuset_mem_spread_node();
3197 else if (current->mempolicy)
3198 nid_alloc = slab_node(current->mempolicy);
3199 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003200 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003201 return NULL;
3202}
3203
3204/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003205 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003206 * certain node and fall back is permitted. First we scan all the
3207 * available nodelists for available objects. If that fails then we
3208 * perform an allocation without specifying a node. This allows the page
3209 * allocator to do its reclaim / fallback magic. We then insert the
3210 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003211 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003212static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003213{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003214 struct zonelist *zonelist;
3215 gfp_t local_flags;
Christoph Lameter765c4502006-09-27 01:50:08 -07003216 struct zone **z;
3217 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003218 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003219
3220 if (flags & __GFP_THISNODE)
3221 return NULL;
3222
3223 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3224 ->node_zonelists[gfp_zone(flags)];
3225 local_flags = (flags & GFP_LEVEL_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003226
Christoph Lameter3c517a62006-12-06 20:33:29 -08003227retry:
3228 /*
3229 * Look through allowed nodes for objects available
3230 * from existing per node queues.
3231 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003232 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003233 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003234
Paul Jackson02a0e532006-12-13 00:34:25 -08003235 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003236 cache->nodelists[nid] &&
3237 cache->nodelists[nid]->free_objects)
3238 obj = ____cache_alloc_node(cache,
3239 flags | GFP_THISNODE, nid);
3240 }
3241
Christoph Lametercfce6602007-05-06 14:50:17 -07003242 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003243 /*
3244 * This allocation will be performed within the constraints
3245 * of the current cpuset / memory policy requirements.
3246 * We may trigger various forms of reclaim on the allowed
3247 * set and go into memory reserves if necessary.
3248 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003249 if (local_flags & __GFP_WAIT)
3250 local_irq_enable();
3251 kmem_flagcheck(cache, flags);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003252 obj = kmem_getpages(cache, flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003253 if (local_flags & __GFP_WAIT)
3254 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003255 if (obj) {
3256 /*
3257 * Insert into the appropriate per node queues
3258 */
3259 nid = page_to_nid(virt_to_page(obj));
3260 if (cache_grow(cache, flags, nid, obj)) {
3261 obj = ____cache_alloc_node(cache,
3262 flags | GFP_THISNODE, nid);
3263 if (!obj)
3264 /*
3265 * Another processor may allocate the
3266 * objects in the slab since we are
3267 * not holding any locks.
3268 */
3269 goto retry;
3270 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003271 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003272 obj = NULL;
3273 }
3274 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003275 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003276 return obj;
3277}
3278
3279/*
Christoph Lametere498be72005-09-09 13:03:32 -07003280 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003282static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003283 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003284{
3285 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003286 struct slab *slabp;
3287 struct kmem_list3 *l3;
3288 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003289 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 l3 = cachep->nodelists[nodeid];
3292 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003293
Andrew Mortona737b3e2006-03-22 00:08:11 -08003294retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003295 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003296 spin_lock(&l3->list_lock);
3297 entry = l3->slabs_partial.next;
3298 if (entry == &l3->slabs_partial) {
3299 l3->free_touched = 1;
3300 entry = l3->slabs_free.next;
3301 if (entry == &l3->slabs_free)
3302 goto must_grow;
3303 }
Christoph Lametere498be72005-09-09 13:03:32 -07003304
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003305 slabp = list_entry(entry, struct slab, list);
3306 check_spinlock_acquired_node(cachep, nodeid);
3307 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003308
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003309 STATS_INC_NODEALLOCS(cachep);
3310 STATS_INC_ACTIVE(cachep);
3311 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003312
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003313 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003314
Matthew Dobson78d382d2006-02-01 03:05:47 -08003315 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003316 check_slabp(cachep, slabp);
3317 l3->free_objects--;
3318 /* move slabp to correct slabp list: */
3319 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003320
Andrew Mortona737b3e2006-03-22 00:08:11 -08003321 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003322 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003323 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003324 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003325
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003326 spin_unlock(&l3->list_lock);
3327 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003328
Andrew Mortona737b3e2006-03-22 00:08:11 -08003329must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003330 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003331 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003332 if (x)
3333 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003334
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003335 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003336
Andrew Mortona737b3e2006-03-22 00:08:11 -08003337done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003339}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003340
3341/**
3342 * kmem_cache_alloc_node - Allocate an object on the specified node
3343 * @cachep: The cache to allocate from.
3344 * @flags: See kmalloc().
3345 * @nodeid: node number of the target node.
3346 * @caller: return address of caller, used for debug information
3347 *
3348 * Identical to kmem_cache_alloc but it will allocate memory on the given
3349 * node, which can improve the performance for cpu bound structures.
3350 *
3351 * Fallback to other node is possible if __GFP_THISNODE is not set.
3352 */
3353static __always_inline void *
3354__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3355 void *caller)
3356{
3357 unsigned long save_flags;
3358 void *ptr;
3359
Akinobu Mita824ebef2007-05-06 14:49:58 -07003360 if (should_failslab(cachep, flags))
3361 return NULL;
3362
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003363 cache_alloc_debugcheck_before(cachep, flags);
3364 local_irq_save(save_flags);
3365
3366 if (unlikely(nodeid == -1))
3367 nodeid = numa_node_id();
3368
3369 if (unlikely(!cachep->nodelists[nodeid])) {
3370 /* Node not bootstrapped yet */
3371 ptr = fallback_alloc(cachep, flags);
3372 goto out;
3373 }
3374
3375 if (nodeid == numa_node_id()) {
3376 /*
3377 * Use the locally cached objects if possible.
3378 * However ____cache_alloc does not allow fallback
3379 * to other nodes. It may fail while we still have
3380 * objects on other nodes available.
3381 */
3382 ptr = ____cache_alloc(cachep, flags);
3383 if (ptr)
3384 goto out;
3385 }
3386 /* ___cache_alloc_node can fall back to other nodes */
3387 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3388 out:
3389 local_irq_restore(save_flags);
3390 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3391
3392 return ptr;
3393}
3394
3395static __always_inline void *
3396__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3397{
3398 void *objp;
3399
3400 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3401 objp = alternate_node_alloc(cache, flags);
3402 if (objp)
3403 goto out;
3404 }
3405 objp = ____cache_alloc(cache, flags);
3406
3407 /*
3408 * We may just have run out of memory on the local node.
3409 * ____cache_alloc_node() knows how to locate memory on other nodes
3410 */
3411 if (!objp)
3412 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3413
3414 out:
3415 return objp;
3416}
3417#else
3418
3419static __always_inline void *
3420__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3421{
3422 return ____cache_alloc(cachep, flags);
3423}
3424
3425#endif /* CONFIG_NUMA */
3426
3427static __always_inline void *
3428__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3429{
3430 unsigned long save_flags;
3431 void *objp;
3432
Akinobu Mita824ebef2007-05-06 14:49:58 -07003433 if (should_failslab(cachep, flags))
3434 return NULL;
3435
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003436 cache_alloc_debugcheck_before(cachep, flags);
3437 local_irq_save(save_flags);
3438 objp = __do_cache_alloc(cachep, flags);
3439 local_irq_restore(save_flags);
3440 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3441 prefetchw(objp);
3442
3443 return objp;
3444}
Christoph Lametere498be72005-09-09 13:03:32 -07003445
3446/*
3447 * Caller needs to acquire correct kmem_list's list_lock
3448 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003449static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003450 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451{
3452 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003453 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454
3455 for (i = 0; i < nr_objects; i++) {
3456 void *objp = objpp[i];
3457 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003459 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003460 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003462 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003464 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003466 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 check_slabp(cachep, slabp);
3468
3469 /* fixup slab chains */
3470 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003471 if (l3->free_objects > l3->free_limit) {
3472 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003473 /* No need to drop any previously held
3474 * lock here, even if we have a off-slab slab
3475 * descriptor it is guaranteed to come from
3476 * a different cache, refer to comments before
3477 * alloc_slabmgmt.
3478 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 slab_destroy(cachep, slabp);
3480 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003481 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 }
3483 } else {
3484 /* Unconditionally move a slab to the end of the
3485 * partial list on free - maximum time for the
3486 * other objects to be freed, too.
3487 */
Christoph Lametere498be72005-09-09 13:03:32 -07003488 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 }
3490 }
3491}
3492
Pekka Enberg343e0d72006-02-01 03:05:50 -08003493static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494{
3495 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003496 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003497 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498
3499 batchcount = ac->batchcount;
3500#if DEBUG
3501 BUG_ON(!batchcount || batchcount > ac->avail);
3502#endif
3503 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003504 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003505 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003506 if (l3->shared) {
3507 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003508 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509 if (max) {
3510 if (batchcount > max)
3511 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003512 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003513 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 shared_array->avail += batchcount;
3515 goto free_done;
3516 }
3517 }
3518
Christoph Lameterff694162005-09-22 21:44:02 -07003519 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003520free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521#if STATS
3522 {
3523 int i = 0;
3524 struct list_head *p;
3525
Christoph Lametere498be72005-09-09 13:03:32 -07003526 p = l3->slabs_free.next;
3527 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528 struct slab *slabp;
3529
3530 slabp = list_entry(p, struct slab, list);
3531 BUG_ON(slabp->inuse);
3532
3533 i++;
3534 p = p->next;
3535 }
3536 STATS_SET_FREEABLE(cachep, i);
3537 }
3538#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003539 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003541 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542}
3543
3544/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003545 * Release an obj back to its cache. If the obj has a constructed state, it must
3546 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003548static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003550 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551
3552 check_irq_off();
3553 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3554
Christoph Lameter3cdc0ed2007-06-08 13:46:46 -07003555 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003556 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003557
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558 if (likely(ac->avail < ac->limit)) {
3559 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003560 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561 return;
3562 } else {
3563 STATS_INC_FREEMISS(cachep);
3564 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003565 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 }
3567}
3568
3569/**
3570 * kmem_cache_alloc - Allocate an object
3571 * @cachep: The cache to allocate from.
3572 * @flags: See kmalloc().
3573 *
3574 * Allocate an object from this cache. The flags are only relevant
3575 * if the cache has no available objects.
3576 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003577void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003579 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580}
3581EXPORT_SYMBOL(kmem_cache_alloc);
3582
3583/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003584 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003585 * @cache: The cache to allocate from.
3586 * @flags: See kmalloc().
3587 *
3588 * Allocate an object from this cache and set the allocated memory to zero.
3589 * The flags are only relevant if the cache has no available objects.
3590 */
3591void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3592{
3593 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3594 if (ret)
3595 memset(ret, 0, obj_size(cache));
3596 return ret;
3597}
3598EXPORT_SYMBOL(kmem_cache_zalloc);
3599
3600/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601 * kmem_ptr_validate - check if an untrusted pointer might
3602 * be a slab entry.
3603 * @cachep: the cache we're checking against
3604 * @ptr: pointer to validate
3605 *
3606 * This verifies that the untrusted pointer looks sane:
3607 * it is _not_ a guarantee that the pointer is actually
3608 * part of the slab cache in question, but it at least
3609 * validates that the pointer can be dereferenced and
3610 * looks half-way sane.
3611 *
3612 * Currently only used for dentry validation.
3613 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003614int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003616 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003618 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003619 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003620 struct page *page;
3621
3622 if (unlikely(addr < min_addr))
3623 goto out;
3624 if (unlikely(addr > (unsigned long)high_memory - size))
3625 goto out;
3626 if (unlikely(addr & align_mask))
3627 goto out;
3628 if (unlikely(!kern_addr_valid(addr)))
3629 goto out;
3630 if (unlikely(!kern_addr_valid(addr + size - 1)))
3631 goto out;
3632 page = virt_to_page(ptr);
3633 if (unlikely(!PageSlab(page)))
3634 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003635 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 goto out;
3637 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003638out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 return 0;
3640}
3641
3642#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003643void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3644{
3645 return __cache_alloc_node(cachep, flags, nodeid,
3646 __builtin_return_address(0));
3647}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648EXPORT_SYMBOL(kmem_cache_alloc_node);
3649
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003650static __always_inline void *
3651__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003652{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003653 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003654
3655 cachep = kmem_find_general_cachep(size, flags);
3656 if (unlikely(cachep == NULL))
3657 return NULL;
3658 return kmem_cache_alloc_node(cachep, flags, node);
3659}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003660
3661#ifdef CONFIG_DEBUG_SLAB
3662void *__kmalloc_node(size_t size, gfp_t flags, int node)
3663{
3664 return __do_kmalloc_node(size, flags, node,
3665 __builtin_return_address(0));
3666}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003667EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003668
3669void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3670 int node, void *caller)
3671{
3672 return __do_kmalloc_node(size, flags, node, caller);
3673}
3674EXPORT_SYMBOL(__kmalloc_node_track_caller);
3675#else
3676void *__kmalloc_node(size_t size, gfp_t flags, int node)
3677{
3678 return __do_kmalloc_node(size, flags, node, NULL);
3679}
3680EXPORT_SYMBOL(__kmalloc_node);
3681#endif /* CONFIG_DEBUG_SLAB */
3682#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683
3684/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003685 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003687 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003688 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003690static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3691 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003693 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003695 /* If you want to save a few bytes .text space: replace
3696 * __ with kmem_.
3697 * Then kmalloc uses the uninlined functions instead of the inline
3698 * functions.
3699 */
3700 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003701 if (unlikely(cachep == NULL))
3702 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003703 return __cache_alloc(cachep, flags, caller);
3704}
3705
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003706
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003707#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003708void *__kmalloc(size_t size, gfp_t flags)
3709{
Al Viro871751e2006-03-25 03:06:39 -08003710 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711}
3712EXPORT_SYMBOL(__kmalloc);
3713
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003714void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3715{
3716 return __do_kmalloc(size, flags, caller);
3717}
3718EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003719
3720#else
3721void *__kmalloc(size_t size, gfp_t flags)
3722{
3723 return __do_kmalloc(size, flags, NULL);
3724}
3725EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003726#endif
3727
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728/**
Pekka Enbergfd76bab2007-05-06 14:48:40 -07003729 * krealloc - reallocate memory. The contents will remain unchanged.
Pekka Enbergfd76bab2007-05-06 14:48:40 -07003730 * @p: object to reallocate memory for.
3731 * @new_size: how many bytes of memory are required.
3732 * @flags: the type of memory to allocate.
3733 *
3734 * The contents of the object pointed to are preserved up to the
3735 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3736 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3737 * %NULL pointer, the object pointed to is freed.
3738 */
3739void *krealloc(const void *p, size_t new_size, gfp_t flags)
3740{
3741 struct kmem_cache *cache, *new_cache;
3742 void *ret;
3743
3744 if (unlikely(!p))
3745 return kmalloc_track_caller(new_size, flags);
3746
3747 if (unlikely(!new_size)) {
3748 kfree(p);
3749 return NULL;
3750 }
3751
3752 cache = virt_to_cache(p);
3753 new_cache = __find_general_cachep(new_size, flags);
3754
3755 /*
3756 * If new size fits in the current cache, bail out.
3757 */
3758 if (likely(cache == new_cache))
3759 return (void *)p;
3760
3761 /*
3762 * We are on the slow-path here so do not use __cache_alloc
3763 * because it bloats kernel text.
3764 */
3765 ret = kmalloc_track_caller(new_size, flags);
3766 if (ret) {
3767 memcpy(ret, p, min(new_size, ksize(p)));
3768 kfree(p);
3769 }
3770 return ret;
3771}
3772EXPORT_SYMBOL(krealloc);
3773
3774/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 * kmem_cache_free - Deallocate an object
3776 * @cachep: The cache the allocation was from.
3777 * @objp: The previously allocated object.
3778 *
3779 * Free an object which was previously allocated from this
3780 * cache.
3781 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003782void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783{
3784 unsigned long flags;
3785
Pekka Enbergddc2e812006-06-23 02:03:40 -07003786 BUG_ON(virt_to_cache(objp) != cachep);
3787
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003789 debug_check_no_locks_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003790 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 local_irq_restore(flags);
3792}
3793EXPORT_SYMBOL(kmem_cache_free);
3794
3795/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 * kfree - free previously allocated memory
3797 * @objp: pointer returned by kmalloc.
3798 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003799 * If @objp is NULL, no operation is performed.
3800 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801 * Don't free memory not originally allocated by kmalloc()
3802 * or you will run into trouble.
3803 */
3804void kfree(const void *objp)
3805{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003806 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 unsigned long flags;
3808
3809 if (unlikely(!objp))
3810 return;
3811 local_irq_save(flags);
3812 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003813 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003814 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003815 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816 local_irq_restore(flags);
3817}
3818EXPORT_SYMBOL(kfree);
3819
Pekka Enberg343e0d72006-02-01 03:05:50 -08003820unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003822 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823}
3824EXPORT_SYMBOL(kmem_cache_size);
3825
Pekka Enberg343e0d72006-02-01 03:05:50 -08003826const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003827{
3828 return cachep->name;
3829}
3830EXPORT_SYMBOL_GPL(kmem_cache_name);
3831
Christoph Lametere498be72005-09-09 13:03:32 -07003832/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003833 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003834 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003835static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003836{
3837 int node;
3838 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003839 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003840 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003841
3842 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003843
Paul Menage3395ee02006-12-06 20:32:16 -08003844 if (use_alien_caches) {
3845 new_alien = alloc_alien_cache(node, cachep->limit);
3846 if (!new_alien)
3847 goto fail;
3848 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003849
Eric Dumazet63109842007-05-06 14:49:28 -07003850 new_shared = NULL;
3851 if (cachep->shared) {
3852 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003853 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003854 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003855 if (!new_shared) {
3856 free_alien_cache(new_alien);
3857 goto fail;
3858 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003859 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003860
Andrew Mortona737b3e2006-03-22 00:08:11 -08003861 l3 = cachep->nodelists[node];
3862 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003863 struct array_cache *shared = l3->shared;
3864
Christoph Lametere498be72005-09-09 13:03:32 -07003865 spin_lock_irq(&l3->list_lock);
3866
Christoph Lametercafeb022006-03-25 03:06:46 -08003867 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003868 free_block(cachep, shared->entry,
3869 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003870
Christoph Lametercafeb022006-03-25 03:06:46 -08003871 l3->shared = new_shared;
3872 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003873 l3->alien = new_alien;
3874 new_alien = NULL;
3875 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003876 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003877 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003878 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003879 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003880 free_alien_cache(new_alien);
3881 continue;
3882 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003883 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003884 if (!l3) {
3885 free_alien_cache(new_alien);
3886 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003887 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003888 }
Christoph Lametere498be72005-09-09 13:03:32 -07003889
3890 kmem_list3_init(l3);
3891 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003892 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003893 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003894 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003895 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003896 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003897 cachep->nodelists[node] = l3;
3898 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003899 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003900
Andrew Mortona737b3e2006-03-22 00:08:11 -08003901fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003902 if (!cachep->next.next) {
3903 /* Cache is not active yet. Roll back what we did */
3904 node--;
3905 while (node >= 0) {
3906 if (cachep->nodelists[node]) {
3907 l3 = cachep->nodelists[node];
3908
3909 kfree(l3->shared);
3910 free_alien_cache(l3->alien);
3911 kfree(l3);
3912 cachep->nodelists[node] = NULL;
3913 }
3914 node--;
3915 }
3916 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003917 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003918}
3919
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003921 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922 struct array_cache *new[NR_CPUS];
3923};
3924
3925static void do_ccupdate_local(void *info)
3926{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003927 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928 struct array_cache *old;
3929
3930 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003931 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003932
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3934 new->new[smp_processor_id()] = old;
3935}
3936
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003937/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003938static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3939 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003941 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003942 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003944 new = kzalloc(sizeof(*new), GFP_KERNEL);
3945 if (!new)
3946 return -ENOMEM;
3947
Christoph Lametere498be72005-09-09 13:03:32 -07003948 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003949 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003950 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003951 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003952 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003953 kfree(new->new[i]);
3954 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003955 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 }
3957 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003958 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003960 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003961
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963 cachep->batchcount = batchcount;
3964 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003965 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966
Christoph Lametere498be72005-09-09 13:03:32 -07003967 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003968 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 if (!ccold)
3970 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003971 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003972 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003973 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 kfree(ccold);
3975 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003976 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003977 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978}
3979
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003980/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003981static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982{
3983 int err;
3984 int limit, shared;
3985
Andrew Mortona737b3e2006-03-22 00:08:11 -08003986 /*
3987 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 * - create a LIFO ordering, i.e. return objects that are cache-warm
3989 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003990 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 * bufctl chains: array operations are cheaper.
3992 * The numbers are guessed, we should auto-tune as described by
3993 * Bonwick.
3994 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003995 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003997 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003999 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004001 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 limit = 54;
4003 else
4004 limit = 120;
4005
Andrew Mortona737b3e2006-03-22 00:08:11 -08004006 /*
4007 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 * allocation behaviour: Most allocs on one cpu, most free operations
4009 * on another cpu. For these cases, an efficient object passing between
4010 * cpus is necessary. This is provided by a shared array. The array
4011 * replaces Bonwick's magazine layer.
4012 * On uniprocessor, it's functionally equivalent (but less efficient)
4013 * to a larger limit. Thus disabled by default.
4014 */
4015 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004016 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018
4019#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004020 /*
4021 * With debugging enabled, large batchcount lead to excessively long
4022 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 */
4024 if (limit > 32)
4025 limit = 32;
4026#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004027 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028 if (err)
4029 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004030 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004031 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032}
4033
Christoph Lameter1b552532006-03-22 00:09:07 -08004034/*
4035 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004036 * necessary. Note that the l3 listlock also protects the array_cache
4037 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004038 */
4039void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4040 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041{
4042 int tofree;
4043
Christoph Lameter1b552532006-03-22 00:09:07 -08004044 if (!ac || !ac->avail)
4045 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046 if (ac->touched && !force) {
4047 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004048 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004049 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004050 if (ac->avail) {
4051 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4052 if (tofree > ac->avail)
4053 tofree = (ac->avail + 1) / 2;
4054 free_block(cachep, ac->entry, tofree, node);
4055 ac->avail -= tofree;
4056 memmove(ac->entry, &(ac->entry[tofree]),
4057 sizeof(void *) * ac->avail);
4058 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004059 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 }
4061}
4062
4063/**
4064 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004065 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066 *
4067 * Called from workqueue/eventd every few seconds.
4068 * Purpose:
4069 * - clear the per-cpu caches for this CPU.
4070 * - return freeable pages to the main free memory pool.
4071 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004072 * If we cannot acquire the cache chain mutex then just give up - we'll try
4073 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004075static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004077 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004078 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004079 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004080 struct delayed_work *work =
4081 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004083 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004085 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004087 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 check_irq_on();
4089
Christoph Lameter35386e32006-03-22 00:09:05 -08004090 /*
4091 * We only take the l3 lock if absolutely necessary and we
4092 * have established with reasonable certainty that
4093 * we can do some work if the lock was obtained.
4094 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004095 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004096
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004097 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098
Christoph Lameteraab22072006-03-22 00:09:06 -08004099 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100
Christoph Lameter35386e32006-03-22 00:09:05 -08004101 /*
4102 * These are racy checks but it does not matter
4103 * if we skip one check or scan twice.
4104 */
Christoph Lametere498be72005-09-09 13:03:32 -07004105 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004106 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107
Christoph Lametere498be72005-09-09 13:03:32 -07004108 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109
Christoph Lameteraab22072006-03-22 00:09:06 -08004110 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111
Christoph Lametered11d9e2006-06-30 01:55:45 -07004112 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004113 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004114 else {
4115 int freed;
4116
4117 freed = drain_freelist(searchp, l3, (l3->free_limit +
4118 5 * searchp->num - 1) / (5 * searchp->num));
4119 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004121next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122 cond_resched();
4123 }
4124 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004125 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004126 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004127out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004128 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004129 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130}
4131
4132#ifdef CONFIG_PROC_FS
4133
Pekka Enberg85289f92006-01-08 01:00:36 -08004134static void print_slabinfo_header(struct seq_file *m)
4135{
4136 /*
4137 * Output format version, so at least we can change it
4138 * without _too_ many complaints.
4139 */
4140#if STATS
4141 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4142#else
4143 seq_puts(m, "slabinfo - version: 2.1\n");
4144#endif
4145 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4146 "<objperslab> <pagesperslab>");
4147 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4148 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4149#if STATS
4150 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004151 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004152 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4153#endif
4154 seq_putc(m, '\n');
4155}
4156
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157static void *s_start(struct seq_file *m, loff_t *pos)
4158{
4159 loff_t n = *pos;
4160 struct list_head *p;
4161
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004162 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004163 if (!n)
4164 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 p = cache_chain.next;
4166 while (n--) {
4167 p = p->next;
4168 if (p == &cache_chain)
4169 return NULL;
4170 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08004171 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172}
4173
4174static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4175{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004176 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004178 return cachep->next.next == &cache_chain ?
4179 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180}
4181
4182static void s_stop(struct seq_file *m, void *p)
4183{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004184 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185}
4186
4187static int s_show(struct seq_file *m, void *p)
4188{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004189 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004190 struct slab *slabp;
4191 unsigned long active_objs;
4192 unsigned long num_objs;
4193 unsigned long active_slabs = 0;
4194 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004195 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004197 int node;
4198 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200 active_objs = 0;
4201 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004202 for_each_online_node(node) {
4203 l3 = cachep->nodelists[node];
4204 if (!l3)
4205 continue;
4206
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004207 check_irq_on();
4208 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004209
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004210 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004211 if (slabp->inuse != cachep->num && !error)
4212 error = "slabs_full accounting error";
4213 active_objs += cachep->num;
4214 active_slabs++;
4215 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004216 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004217 if (slabp->inuse == cachep->num && !error)
4218 error = "slabs_partial inuse accounting error";
4219 if (!slabp->inuse && !error)
4220 error = "slabs_partial/inuse accounting error";
4221 active_objs += slabp->inuse;
4222 active_slabs++;
4223 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004224 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004225 if (slabp->inuse && !error)
4226 error = "slabs_free/inuse accounting error";
4227 num_slabs++;
4228 }
4229 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004230 if (l3->shared)
4231 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004232
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004233 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004235 num_slabs += active_slabs;
4236 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004237 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 error = "free_objects accounting error";
4239
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004240 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004241 if (error)
4242 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4243
4244 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004245 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004246 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004248 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004249 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004250 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004252 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004253 unsigned long high = cachep->high_mark;
4254 unsigned long allocs = cachep->num_allocations;
4255 unsigned long grown = cachep->grown;
4256 unsigned long reaped = cachep->reaped;
4257 unsigned long errors = cachep->errors;
4258 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004259 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004260 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004261 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262
Christoph Lametere498be72005-09-09 13:03:32 -07004263 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004264 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004265 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004266 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267 }
4268 /* cpu stats */
4269 {
4270 unsigned long allochit = atomic_read(&cachep->allochit);
4271 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4272 unsigned long freehit = atomic_read(&cachep->freehit);
4273 unsigned long freemiss = atomic_read(&cachep->freemiss);
4274
4275 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004276 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277 }
4278#endif
4279 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280 return 0;
4281}
4282
4283/*
4284 * slabinfo_op - iterator that generates /proc/slabinfo
4285 *
4286 * Output layout:
4287 * cache-name
4288 * num-active-objs
4289 * total-objs
4290 * object size
4291 * num-active-slabs
4292 * total-slabs
4293 * num-pages-per-slab
4294 * + further values on SMP and with statistics enabled
4295 */
4296
Helge Deller15ad7cd2006-12-06 20:40:36 -08004297const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004298 .start = s_start,
4299 .next = s_next,
4300 .stop = s_stop,
4301 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302};
4303
4304#define MAX_SLABINFO_WRITE 128
4305/**
4306 * slabinfo_write - Tuning for the slab allocator
4307 * @file: unused
4308 * @buffer: user buffer
4309 * @count: data length
4310 * @ppos: unused
4311 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004312ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4313 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004315 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004317 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004318
Linus Torvalds1da177e2005-04-16 15:20:36 -07004319 if (count > MAX_SLABINFO_WRITE)
4320 return -EINVAL;
4321 if (copy_from_user(&kbuf, buffer, count))
4322 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004323 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324
4325 tmp = strchr(kbuf, ' ');
4326 if (!tmp)
4327 return -EINVAL;
4328 *tmp = '\0';
4329 tmp++;
4330 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4331 return -EINVAL;
4332
4333 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004334 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004335 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004336 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004338 if (limit < 1 || batchcount < 1 ||
4339 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004340 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004341 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004342 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004343 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004344 }
4345 break;
4346 }
4347 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004348 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004349 if (res >= 0)
4350 res = count;
4351 return res;
4352}
Al Viro871751e2006-03-25 03:06:39 -08004353
4354#ifdef CONFIG_DEBUG_SLAB_LEAK
4355
4356static void *leaks_start(struct seq_file *m, loff_t *pos)
4357{
4358 loff_t n = *pos;
4359 struct list_head *p;
4360
4361 mutex_lock(&cache_chain_mutex);
4362 p = cache_chain.next;
4363 while (n--) {
4364 p = p->next;
4365 if (p == &cache_chain)
4366 return NULL;
4367 }
4368 return list_entry(p, struct kmem_cache, next);
4369}
4370
4371static inline int add_caller(unsigned long *n, unsigned long v)
4372{
4373 unsigned long *p;
4374 int l;
4375 if (!v)
4376 return 1;
4377 l = n[1];
4378 p = n + 2;
4379 while (l) {
4380 int i = l/2;
4381 unsigned long *q = p + 2 * i;
4382 if (*q == v) {
4383 q[1]++;
4384 return 1;
4385 }
4386 if (*q > v) {
4387 l = i;
4388 } else {
4389 p = q + 2;
4390 l -= i + 1;
4391 }
4392 }
4393 if (++n[1] == n[0])
4394 return 0;
4395 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4396 p[0] = v;
4397 p[1] = 1;
4398 return 1;
4399}
4400
4401static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4402{
4403 void *p;
4404 int i;
4405 if (n[0] == n[1])
4406 return;
4407 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4408 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4409 continue;
4410 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4411 return;
4412 }
4413}
4414
4415static void show_symbol(struct seq_file *m, unsigned long address)
4416{
4417#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004418 unsigned long offset, size;
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004419 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1];
Al Viro871751e2006-03-25 03:06:39 -08004420
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004421 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004422 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004423 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004424 seq_printf(m, " [%s]", modname);
4425 return;
4426 }
4427#endif
4428 seq_printf(m, "%p", (void *)address);
4429}
4430
4431static int leaks_show(struct seq_file *m, void *p)
4432{
4433 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004434 struct slab *slabp;
4435 struct kmem_list3 *l3;
4436 const char *name;
4437 unsigned long *n = m->private;
4438 int node;
4439 int i;
4440
4441 if (!(cachep->flags & SLAB_STORE_USER))
4442 return 0;
4443 if (!(cachep->flags & SLAB_RED_ZONE))
4444 return 0;
4445
4446 /* OK, we can do it */
4447
4448 n[1] = 0;
4449
4450 for_each_online_node(node) {
4451 l3 = cachep->nodelists[node];
4452 if (!l3)
4453 continue;
4454
4455 check_irq_on();
4456 spin_lock_irq(&l3->list_lock);
4457
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004458 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004459 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004460 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004461 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004462 spin_unlock_irq(&l3->list_lock);
4463 }
4464 name = cachep->name;
4465 if (n[0] == n[1]) {
4466 /* Increase the buffer size */
4467 mutex_unlock(&cache_chain_mutex);
4468 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4469 if (!m->private) {
4470 /* Too bad, we are really out */
4471 m->private = n;
4472 mutex_lock(&cache_chain_mutex);
4473 return -ENOMEM;
4474 }
4475 *(unsigned long *)m->private = n[0] * 2;
4476 kfree(n);
4477 mutex_lock(&cache_chain_mutex);
4478 /* Now make sure this entry will be retried */
4479 m->count = m->size;
4480 return 0;
4481 }
4482 for (i = 0; i < n[1]; i++) {
4483 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4484 show_symbol(m, n[2*i+2]);
4485 seq_putc(m, '\n');
4486 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004487
Al Viro871751e2006-03-25 03:06:39 -08004488 return 0;
4489}
4490
Helge Deller15ad7cd2006-12-06 20:40:36 -08004491const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004492 .start = leaks_start,
4493 .next = s_next,
4494 .stop = s_stop,
4495 .show = leaks_show,
4496};
4497#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004498#endif
4499
Manfred Spraul00e145b2005-09-03 15:55:07 -07004500/**
4501 * ksize - get the actual amount of memory allocated for a given object
4502 * @objp: Pointer to the object
4503 *
4504 * kmalloc may internally round up allocations and return more memory
4505 * than requested. ksize() can be used to determine the actual amount of
4506 * memory allocated. The caller may use this additional memory, even though
4507 * a smaller amount of memory was initially specified with the kmalloc call.
4508 * The caller must guarantee that objp points to a valid object previously
4509 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4510 * must not be freed during the duration of the call.
4511 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004512size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004513{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004514 if (unlikely(objp == NULL))
4515 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004516
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004517 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004518}