blob: 046607f05f3eb16414538201836c6e605135b0e8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
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>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700113#include <linux/debugobjects.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115#include <asm/cacheflush.h>
116#include <asm/tlbflush.h>
117#include <asm/page.h>
118
119/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700120 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 * 0 for faster, smaller code (especially in the critical paths).
122 *
123 * STATS - 1 to collect stats for /proc/slabinfo.
124 * 0 for faster, smaller code (especially in the critical paths).
125 *
126 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
127 */
128
129#ifdef CONFIG_DEBUG_SLAB
130#define DEBUG 1
131#define STATS 1
132#define FORCED_DEBUG 1
133#else
134#define DEBUG 0
135#define STATS 0
136#define FORCED_DEBUG 0
137#endif
138
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139/* Shouldn't this be in a header file somewhere? */
140#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400141#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143#ifndef ARCH_KMALLOC_MINALIGN
144/*
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700149 * alignment larger than the alignment of a 64-bit integer.
150 * ARCH_KMALLOC_MINALIGN allows that.
151 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700153#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154#endif
155
156#ifndef ARCH_SLAB_MINALIGN
157/*
158 * Enforce a minimum alignment for all caches.
159 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
160 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
161 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
162 * some debug features.
163 */
164#define ARCH_SLAB_MINALIGN 0
165#endif
166
167#ifndef ARCH_KMALLOC_FLAGS
168#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169#endif
170
171/* Legal flag mask for kmem_cache_create(). */
172#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700173# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800175 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700176 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
179 SLAB_DEBUG_OBJECTS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800181# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700182 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
185 SLAB_DEBUG_OBJECTS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#endif
187
188/*
189 * kmem_bufctl_t:
190 *
191 * Bufctl's are used for linking objs within a slab
192 * linked offsets.
193 *
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 */
206
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700207typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800210#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213/*
214 * struct slab
215 *
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 */
220struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227};
228
229/*
230 * struct slab_rcu
231 *
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
239 *
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
242 *
243 * We assume struct slab_rcu can overlay struct slab when destroying.
244 */
245struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800247 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
251/*
252 * struct array_cache
253 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * Purpose:
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
258 *
259 * The limit is stored in the per-cpu structure to reduce the data cache
260 * footprint.
261 *
262 */
263struct array_cache {
264 unsigned int avail;
265 unsigned int limit;
266 unsigned int batchcount;
267 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700268 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700269 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
272 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274};
275
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276/*
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 */
280#define BOOT_CPUCACHE_ENTRIES 1
281struct arraycache_init {
282 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800283 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284};
285
286/*
Christoph Lametere498be72005-09-09 13:03:32 -0700287 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 */
289struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800294 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800295 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301};
302
Christoph Lametere498be72005-09-09 13:03:32 -0700303/*
304 * Need this for bootstrapping a per node allocator.
305 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200306#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700307struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200309#define SIZE_AC MAX_NUMNODES
310#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
Christoph Lametered11d9e2006-06-30 01:55:45 -0700312static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314static void free_block(struct kmem_cache *cachep, void **objpp, int len,
315 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700316static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000317static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700318
Christoph Lametere498be72005-09-09 13:03:32 -0700319/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700322 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700323static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700324{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800325 extern void __bad_size(void);
326
Christoph Lametere498be72005-09-09 13:03:32 -0700327 if (__builtin_constant_p(size)) {
328 int i = 0;
329
330#define CACHE(x) \
331 if (size <=x) \
332 return i; \
333 else \
334 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800335#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700336#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800337 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700338 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700340 return 0;
341}
342
Ingo Molnare0a42722006-06-23 02:03:46 -0700343static int slab_early_init = 1;
344
Christoph Lametere498be72005-09-09 13:03:32 -0700345#define INDEX_AC index_of(sizeof(struct arraycache_init))
346#define INDEX_L3 index_of(sizeof(struct kmem_list3))
347
Pekka Enberg5295a742006-02-01 03:05:48 -0800348static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700349{
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800355 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
359}
360
Andrew Mortona737b3e2006-03-22 00:08:11 -0800361#define MAKE_LIST(cachep, listp, slab, nodeid) \
362 do { \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700365 } while (0)
366
Andrew Mortona737b3e2006-03-22 00:08:11 -0800367#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
368 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
372 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800375 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 *
377 * manages a cache.
378 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800379
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800380struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800382 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 unsigned int batchcount;
385 unsigned int limit;
386 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800387
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800388 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800389 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800390/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391
Andrew Mortona737b3e2006-03-22 00:08:11 -0800392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800395/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800400 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
Andrew Mortona737b3e2006-03-22 00:08:11 -0800402 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800403 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800404 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800406 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408 /* constructor func */
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -0700409 void (*ctor)(struct kmem_cache *, void *);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800411/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800412 const char *name;
413 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800415/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800417 unsigned long num_active;
418 unsigned long num_allocations;
419 unsigned long high_mark;
420 unsigned long grown;
421 unsigned long reaped;
422 unsigned long errors;
423 unsigned long max_freeable;
424 unsigned long node_allocs;
425 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700426 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800427 atomic_t allochit;
428 atomic_t allocmiss;
429 atomic_t freehit;
430 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431#endif
432#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800433 /*
434 * If debugging is enabled, then the allocator can add additional
435 * fields and/or padding to every object. buffer_size contains the total
436 * object size including these internal fields, the following two
437 * variables contain the offset to the user object and its size.
438 */
439 int obj_offset;
440 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700442 /*
443 * We put nodelists[] at the end of kmem_cache, because we want to size
444 * this array to nr_node_ids slots instead of MAX_NUMNODES
445 * (see kmem_cache_init())
446 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
447 * is statically defined, so we reserve the max number of nodes.
448 */
449 struct kmem_list3 *nodelists[MAX_NUMNODES];
450 /*
451 * Do not add fields after nodelists[]
452 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453};
454
455#define CFLGS_OFF_SLAB (0x80000000UL)
456#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
457
458#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800459/*
460 * Optimization question: fewer reaps means less probability for unnessary
461 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100463 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 * which could lock up otherwise freeable slabs.
465 */
466#define REAPTIMEOUT_CPUC (2*HZ)
467#define REAPTIMEOUT_LIST3 (4*HZ)
468
469#if STATS
470#define STATS_INC_ACTIVE(x) ((x)->num_active++)
471#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
472#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
473#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700474#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800475#define STATS_SET_HIGH(x) \
476 do { \
477 if ((x)->num_active > (x)->high_mark) \
478 (x)->high_mark = (x)->num_active; \
479 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480#define STATS_INC_ERR(x) ((x)->errors++)
481#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700482#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700483#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800484#define STATS_SET_FREEABLE(x, i) \
485 do { \
486 if ((x)->max_freeable < i) \
487 (x)->max_freeable = i; \
488 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
490#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
491#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
492#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
493#else
494#define STATS_INC_ACTIVE(x) do { } while (0)
495#define STATS_DEC_ACTIVE(x) do { } while (0)
496#define STATS_INC_ALLOCED(x) do { } while (0)
497#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700498#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499#define STATS_SET_HIGH(x) do { } while (0)
500#define STATS_INC_ERR(x) do { } while (0)
501#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700502#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700503#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800504#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505#define STATS_INC_ALLOCHIT(x) do { } while (0)
506#define STATS_INC_ALLOCMISS(x) do { } while (0)
507#define STATS_INC_FREEHIT(x) do { } while (0)
508#define STATS_INC_FREEMISS(x) do { } while (0)
509#endif
510
511#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512
Andrew Mortona737b3e2006-03-22 00:08:11 -0800513/*
514 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800516 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * the end of an object is aligned with the end of the real
518 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800519 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800521 * cachep->obj_offset: The real object.
522 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800523 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
524 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800526static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800528 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529}
530
Pekka Enberg343e0d72006-02-01 03:05:50 -0800531static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800533 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534}
535
David Woodhouseb46b8f12007-05-08 00:22:59 -0700536static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537{
538 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700539 return (unsigned long long*) (objp + obj_offset(cachep) -
540 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
David Woodhouseb46b8f12007-05-08 00:22:59 -0700543static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544{
545 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
546 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700547 return (unsigned long long *)(objp + cachep->buffer_size -
548 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400549 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700550 return (unsigned long long *) (objp + cachep->buffer_size -
551 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552}
553
Pekka Enberg343e0d72006-02-01 03:05:50 -0800554static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555{
556 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800557 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558}
559
560#else
561
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800562#define obj_offset(x) 0
563#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700564#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
565#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
567
568#endif
569
570/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 * Do not go above this order unless 0 objects fit into the slab.
572 */
573#define BREAK_GFP_ORDER_HI 1
574#define BREAK_GFP_ORDER_LO 0
575static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
576
Andrew Mortona737b3e2006-03-22 00:08:11 -0800577/*
578 * Functions for storing/retrieving the cachep and or slab from the page
579 * allocator. These are used to find the slab an obj belongs to. With kfree(),
580 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800582static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
583{
584 page->lru.next = (struct list_head *)cache;
585}
586
587static inline struct kmem_cache *page_get_cache(struct page *page)
588{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700589 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700590 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800591 return (struct kmem_cache *)page->lru.next;
592}
593
594static inline void page_set_slab(struct page *page, struct slab *slab)
595{
596 page->lru.prev = (struct list_head *)slab;
597}
598
599static inline struct slab *page_get_slab(struct page *page)
600{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700601 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800602 return (struct slab *)page->lru.prev;
603}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800605static inline struct kmem_cache *virt_to_cache(const void *obj)
606{
Christoph Lameterb49af682007-05-06 14:49:41 -0700607 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800608 return page_get_cache(page);
609}
610
611static inline struct slab *virt_to_slab(const void *obj)
612{
Christoph Lameterb49af682007-05-06 14:49:41 -0700613 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800614 return page_get_slab(page);
615}
616
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800617static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
618 unsigned int idx)
619{
620 return slab->s_mem + cache->buffer_size * idx;
621}
622
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800623/*
624 * We want to avoid an expensive divide : (offset / cache->buffer_size)
625 * Using the fact that buffer_size is a constant for a particular cache,
626 * we can replace (offset / cache->buffer_size) by
627 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
628 */
629static inline unsigned int obj_to_index(const struct kmem_cache *cache,
630 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800631{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800632 u32 offset = (obj - slab->s_mem);
633 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800634}
635
Andrew Mortona737b3e2006-03-22 00:08:11 -0800636/*
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639struct cache_sizes malloc_sizes[] = {
640#define CACHE(x) { .cs_size = (x) },
641#include <linux/kmalloc_sizes.h>
642 CACHE(ULONG_MAX)
643#undef CACHE
644};
645EXPORT_SYMBOL(malloc_sizes);
646
647/* Must match cache_sizes above. Out of line to keep cache footprint low. */
648struct cache_names {
649 char *name;
650 char *name_dma;
651};
652
653static struct cache_names __initdata cache_names[] = {
654#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800656 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657#undef CACHE
658};
659
660static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800666static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800667 .batchcount = 1,
668 .limit = BOOT_CPUCACHE_ENTRIES,
669 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800670 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800671 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672};
673
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700674#define BAD_ALIEN_MAGIC 0x01020304ul
675
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200676#ifdef CONFIG_LOCKDEP
677
678/*
679 * Slab sometimes uses the kmalloc slabs to store the slab headers
680 * for other slabs "off slab".
681 * The locking for this is tricky in that it nests within the locks
682 * of all other slabs in a few places; to deal with this special
683 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700684 *
685 * We set lock class for alien array caches which are up during init.
686 * The lock annotation will be lost if all cpus of a node goes down and
687 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200688 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700689static struct lock_class_key on_slab_l3_key;
690static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200691
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700692static inline void init_lock_keys(void)
693
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200694{
695 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700696 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200697
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700698 while (s->cs_size != ULONG_MAX) {
699 for_each_node(q) {
700 struct array_cache **alc;
701 int r;
702 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
703 if (!l3 || OFF_SLAB(s->cs_cachep))
704 continue;
705 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
706 alc = l3->alien;
707 /*
708 * FIXME: This check for BAD_ALIEN_MAGIC
709 * should go away when common slab code is taught to
710 * work even without alien caches.
711 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
712 * for alloc_alien_cache,
713 */
714 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
715 continue;
716 for_each_node(r) {
717 if (alc[r])
718 lockdep_set_class(&alc[r]->lock,
719 &on_slab_alc_key);
720 }
721 }
722 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200723 }
724}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200725#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700726static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200727{
728}
729#endif
730
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800731/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100732 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800733 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800734static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735static struct list_head cache_chain;
736
737/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 * chicken and egg problem: delay the per-cpu array allocation
739 * until the general caches are up.
740 */
741static enum {
742 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700743 PARTIAL_AC,
744 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 FULL
746} g_cpucache_up;
747
Mike Kravetz39d24e62006-05-15 09:44:13 -0700748/*
749 * used by boot code to determine if it can use slab based allocator
750 */
751int slab_is_available(void)
752{
753 return g_cpucache_up == FULL;
754}
755
David Howells52bad642006-11-22 14:54:01 +0000756static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757
Pekka Enberg343e0d72006-02-01 03:05:50 -0800758static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759{
760 return cachep->array[smp_processor_id()];
761}
762
Andrew Mortona737b3e2006-03-22 00:08:11 -0800763static inline struct kmem_cache *__find_general_cachep(size_t size,
764 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765{
766 struct cache_sizes *csizep = malloc_sizes;
767
768#if DEBUG
769 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800770 * kmem_cache_create(), or __kmalloc(), before
771 * the generic caches are initialized.
772 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700773 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700775 if (!size)
776 return ZERO_SIZE_PTR;
777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 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
Harvey Harrisond40cee22008-04-30 00:55:07 -0700864#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865
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;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700883static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800884static int __init noaliencache_setup(char *s)
885{
886 use_alien_caches = 0;
887 return 1;
888}
889__setup("noaliencache", noaliencache_setup);
890
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800891#ifdef CONFIG_NUMA
892/*
893 * Special reaping functions for NUMA systems called from cache_reap().
894 * These take care of doing round robin flushing of alien caches (containing
895 * objects freed on different nodes from which they were allocated) and the
896 * flushing of remote pcps by calling drain_node_pages.
897 */
898static DEFINE_PER_CPU(unsigned long, reap_node);
899
900static void init_reap_node(int cpu)
901{
902 int node;
903
904 node = next_node(cpu_to_node(cpu), node_online_map);
905 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800906 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800907
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800908 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800909}
910
911static void next_reap_node(void)
912{
913 int node = __get_cpu_var(reap_node);
914
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800915 node = next_node(node, node_online_map);
916 if (unlikely(node >= MAX_NUMNODES))
917 node = first_node(node_online_map);
918 __get_cpu_var(reap_node) = node;
919}
920
921#else
922#define init_reap_node(cpu) do { } while (0)
923#define next_reap_node(void) do { } while (0)
924#endif
925
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926/*
927 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
928 * via the workqueue/eventd.
929 * Add the CPU number into the expiration time to minimize the possibility of
930 * the CPUs getting into lockstep and contending for the global cache chain
931 * lock.
932 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700933static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934{
David Howells52bad642006-11-22 14:54:01 +0000935 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
937 /*
938 * When this gets called from do_initcalls via cpucache_init(),
939 * init_workqueues() has already run, so keventd will be setup
940 * at that time.
941 */
David Howells52bad642006-11-22 14:54:01 +0000942 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800943 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000944 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800945 schedule_delayed_work_on(cpu, reap_work,
946 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 }
948}
949
Christoph Lametere498be72005-09-09 13:03:32 -0700950static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800951 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800953 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 struct array_cache *nc = NULL;
955
Christoph Lametere498be72005-09-09 13:03:32 -0700956 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 if (nc) {
958 nc->avail = 0;
959 nc->limit = entries;
960 nc->batchcount = batchcount;
961 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700962 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 }
964 return nc;
965}
966
Christoph Lameter3ded1752006-03-25 03:06:44 -0800967/*
968 * Transfer objects in one arraycache to another.
969 * Locking must be handled by the caller.
970 *
971 * Return the number of entries transferred.
972 */
973static int transfer_objects(struct array_cache *to,
974 struct array_cache *from, unsigned int max)
975{
976 /* Figure out how many entries to transfer */
977 int nr = min(min(from->avail, max), to->limit - to->avail);
978
979 if (!nr)
980 return 0;
981
982 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
983 sizeof(void *) *nr);
984
985 from->avail -= nr;
986 to->avail += nr;
987 to->touched = 1;
988 return nr;
989}
990
Christoph Lameter765c4502006-09-27 01:50:08 -0700991#ifndef CONFIG_NUMA
992
993#define drain_alien_cache(cachep, alien) do { } while (0)
994#define reap_alien(cachep, l3) do { } while (0)
995
996static inline struct array_cache **alloc_alien_cache(int node, int limit)
997{
998 return (struct array_cache **)BAD_ALIEN_MAGIC;
999}
1000
1001static inline void free_alien_cache(struct array_cache **ac_ptr)
1002{
1003}
1004
1005static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1006{
1007 return 0;
1008}
1009
1010static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1011 gfp_t flags)
1012{
1013 return NULL;
1014}
1015
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001016static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001017 gfp_t flags, int nodeid)
1018{
1019 return NULL;
1020}
1021
1022#else /* CONFIG_NUMA */
1023
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001024static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001025static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001026
Pekka Enberg5295a742006-02-01 03:05:48 -08001027static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001028{
1029 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001030 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001031 int i;
1032
1033 if (limit > 1)
1034 limit = 12;
1035 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1036 if (ac_ptr) {
1037 for_each_node(i) {
1038 if (i == node || !node_online(i)) {
1039 ac_ptr[i] = NULL;
1040 continue;
1041 }
1042 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1043 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001044 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001045 kfree(ac_ptr[i]);
1046 kfree(ac_ptr);
1047 return NULL;
1048 }
1049 }
1050 }
1051 return ac_ptr;
1052}
1053
Pekka Enberg5295a742006-02-01 03:05:48 -08001054static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001055{
1056 int i;
1057
1058 if (!ac_ptr)
1059 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001060 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001061 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001062 kfree(ac_ptr);
1063}
1064
Pekka Enberg343e0d72006-02-01 03:05:50 -08001065static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001066 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001067{
1068 struct kmem_list3 *rl3 = cachep->nodelists[node];
1069
1070 if (ac->avail) {
1071 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001072 /*
1073 * Stuff objects into the remote nodes shared array first.
1074 * That way we could avoid the overhead of putting the objects
1075 * into the free lists and getting them back later.
1076 */
shin, jacob693f7d32006-04-28 10:54:37 -05001077 if (rl3->shared)
1078 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001079
Christoph Lameterff694162005-09-22 21:44:02 -07001080 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001081 ac->avail = 0;
1082 spin_unlock(&rl3->list_lock);
1083 }
1084}
1085
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001086/*
1087 * Called from cache_reap() to regularly drain alien caches round robin.
1088 */
1089static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1090{
1091 int node = __get_cpu_var(reap_node);
1092
1093 if (l3->alien) {
1094 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001095
1096 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001097 __drain_alien_cache(cachep, ac, node);
1098 spin_unlock_irq(&ac->lock);
1099 }
1100 }
1101}
1102
Andrew Mortona737b3e2006-03-22 00:08:11 -08001103static void drain_alien_cache(struct kmem_cache *cachep,
1104 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001105{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001106 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001107 struct array_cache *ac;
1108 unsigned long flags;
1109
1110 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001111 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001112 if (ac) {
1113 spin_lock_irqsave(&ac->lock, flags);
1114 __drain_alien_cache(cachep, ac, i);
1115 spin_unlock_irqrestore(&ac->lock, flags);
1116 }
1117 }
1118}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001119
Ingo Molnar873623d2006-07-13 14:44:38 +02001120static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001121{
1122 struct slab *slabp = virt_to_slab(objp);
1123 int nodeid = slabp->nodeid;
1124 struct kmem_list3 *l3;
1125 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001126 int node;
1127
1128 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001129
1130 /*
1131 * Make sure we are not freeing a object from another node to the array
1132 * cache on this cpu.
1133 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001134 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001135 return 0;
1136
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001137 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001138 STATS_INC_NODEFREES(cachep);
1139 if (l3->alien && l3->alien[nodeid]) {
1140 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001141 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001142 if (unlikely(alien->avail == alien->limit)) {
1143 STATS_INC_ACOVERFLOW(cachep);
1144 __drain_alien_cache(cachep, alien, nodeid);
1145 }
1146 alien->entry[alien->avail++] = objp;
1147 spin_unlock(&alien->lock);
1148 } else {
1149 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1150 free_block(cachep, &objp, 1, nodeid);
1151 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1152 }
1153 return 1;
1154}
Christoph Lametere498be72005-09-09 13:03:32 -07001155#endif
1156
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001157static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001159 struct kmem_cache *cachep;
1160 struct kmem_list3 *l3 = NULL;
1161 int node = cpu_to_node(cpu);
Mike Travisc5f59f02008-04-04 18:11:10 -07001162 node_to_cpumask_ptr(mask, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001163
1164 list_for_each_entry(cachep, &cache_chain, next) {
1165 struct array_cache *nc;
1166 struct array_cache *shared;
1167 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001168
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001169 /* cpu is dead; no one can alloc from it. */
1170 nc = cachep->array[cpu];
1171 cachep->array[cpu] = NULL;
1172 l3 = cachep->nodelists[node];
1173
1174 if (!l3)
1175 goto free_array_cache;
1176
1177 spin_lock_irq(&l3->list_lock);
1178
1179 /* Free limit for this kmem_list3 */
1180 l3->free_limit -= cachep->batchcount;
1181 if (nc)
1182 free_block(cachep, nc->entry, nc->avail, node);
1183
Mike Travisc5f59f02008-04-04 18:11:10 -07001184 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 spin_unlock_irq(&l3->list_lock);
1186 goto free_array_cache;
1187 }
1188
1189 shared = l3->shared;
1190 if (shared) {
1191 free_block(cachep, shared->entry,
1192 shared->avail, node);
1193 l3->shared = NULL;
1194 }
1195
1196 alien = l3->alien;
1197 l3->alien = NULL;
1198
1199 spin_unlock_irq(&l3->list_lock);
1200
1201 kfree(shared);
1202 if (alien) {
1203 drain_alien_cache(cachep, alien);
1204 free_alien_cache(alien);
1205 }
1206free_array_cache:
1207 kfree(nc);
1208 }
1209 /*
1210 * In the previous loop, all the objects were freed to
1211 * the respective cache's slabs, now we can go ahead and
1212 * shrink each nodelist to its limit.
1213 */
1214 list_for_each_entry(cachep, &cache_chain, next) {
1215 l3 = cachep->nodelists[node];
1216 if (!l3)
1217 continue;
1218 drain_freelist(cachep, l3, l3->free_objects);
1219 }
1220}
1221
1222static int __cpuinit cpuup_prepare(long cpu)
1223{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001224 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001225 struct kmem_list3 *l3 = NULL;
1226 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001227 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001229 /*
1230 * We need to do this right in the beginning since
1231 * alloc_arraycache's are going to use this list.
1232 * kmalloc_node allows us to add the slab to the right
1233 * kmem_list3 and not this cpu's kmem_list3
1234 */
1235
1236 list_for_each_entry(cachep, &cache_chain, next) {
1237 /*
1238 * Set up the size64 kmemlist for cpu before we can
1239 * begin anything. Make sure some other cpu on this
1240 * node has not already allocated this
1241 */
1242 if (!cachep->nodelists[node]) {
1243 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1244 if (!l3)
1245 goto bad;
1246 kmem_list3_init(l3);
1247 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1248 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1249
1250 /*
1251 * The l3s don't come and go as CPUs come and
1252 * go. cache_chain_mutex is sufficient
1253 * protection here.
1254 */
1255 cachep->nodelists[node] = l3;
1256 }
1257
1258 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1259 cachep->nodelists[node]->free_limit =
1260 (1 + nr_cpus_node(node)) *
1261 cachep->batchcount + cachep->num;
1262 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1263 }
1264
1265 /*
1266 * Now we can go ahead with allocating the shared arrays and
1267 * array caches
1268 */
1269 list_for_each_entry(cachep, &cache_chain, next) {
1270 struct array_cache *nc;
1271 struct array_cache *shared = NULL;
1272 struct array_cache **alien = NULL;
1273
1274 nc = alloc_arraycache(node, cachep->limit,
1275 cachep->batchcount);
1276 if (!nc)
1277 goto bad;
1278 if (cachep->shared) {
1279 shared = alloc_arraycache(node,
1280 cachep->shared * cachep->batchcount,
1281 0xbaadf00d);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001282 if (!shared) {
1283 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001284 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001285 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001286 }
1287 if (use_alien_caches) {
1288 alien = alloc_alien_cache(node, cachep->limit);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001289 if (!alien) {
1290 kfree(shared);
1291 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001292 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001293 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001294 }
1295 cachep->array[cpu] = nc;
1296 l3 = cachep->nodelists[node];
1297 BUG_ON(!l3);
1298
1299 spin_lock_irq(&l3->list_lock);
1300 if (!l3->shared) {
1301 /*
1302 * We are serialised from CPU_DEAD or
1303 * CPU_UP_CANCELLED by the cpucontrol lock
1304 */
1305 l3->shared = shared;
1306 shared = NULL;
1307 }
1308#ifdef CONFIG_NUMA
1309 if (!l3->alien) {
1310 l3->alien = alien;
1311 alien = NULL;
1312 }
1313#endif
1314 spin_unlock_irq(&l3->list_lock);
1315 kfree(shared);
1316 free_alien_cache(alien);
1317 }
1318 return 0;
1319bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001320 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001321 return -ENOMEM;
1322}
1323
1324static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1325 unsigned long action, void *hcpu)
1326{
1327 long cpu = (long)hcpu;
1328 int err = 0;
1329
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001331 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001332 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001333 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001334 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001335 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 break;
1337 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001338 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 start_cpu_timer(cpu);
1340 break;
1341#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001342 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001343 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001344 /*
1345 * Shutdown cache reaper. Note that the cache_chain_mutex is
1346 * held so that if cache_reap() is invoked it cannot do
1347 * anything expensive but will only modify reap_work
1348 * and reschedule the timer.
1349 */
1350 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1351 /* Now the cache_reaper is guaranteed to be not running. */
1352 per_cpu(reap_work, cpu).work.func = NULL;
1353 break;
1354 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001355 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001356 start_cpu_timer(cpu);
1357 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001359 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001360 /*
1361 * Even if all the cpus of a node are down, we don't free the
1362 * kmem_list3 of any cache. This to avoid a race between
1363 * cpu_down, and a kmalloc allocation from another cpu for
1364 * memory from the node of the cpu going down. The list3
1365 * structure is usually allocated from kmem_cache_create() and
1366 * gets destroyed at kmem_cache_destroy().
1367 */
Simon Arlott183ff222007-10-20 01:27:18 +02001368 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001369#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001371 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001372 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001373 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001374 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001377 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378}
1379
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001380static struct notifier_block __cpuinitdata cpucache_notifier = {
1381 &cpuup_callback, NULL, 0
1382};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383
Christoph Lametere498be72005-09-09 13:03:32 -07001384/*
1385 * swap the static kmem_list3 with kmalloced memory
1386 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001387static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1388 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001389{
1390 struct kmem_list3 *ptr;
1391
Christoph Lametere498be72005-09-09 13:03:32 -07001392 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1393 BUG_ON(!ptr);
1394
1395 local_irq_disable();
1396 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001397 /*
1398 * Do not assume that spinlocks can be initialized via memcpy:
1399 */
1400 spin_lock_init(&ptr->list_lock);
1401
Christoph Lametere498be72005-09-09 13:03:32 -07001402 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1403 cachep->nodelists[nodeid] = ptr;
1404 local_irq_enable();
1405}
1406
Andrew Mortona737b3e2006-03-22 00:08:11 -08001407/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001408 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1409 * size of kmem_list3.
1410 */
1411static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1412{
1413 int node;
1414
1415 for_each_online_node(node) {
1416 cachep->nodelists[node] = &initkmem_list3[index + node];
1417 cachep->nodelists[node]->next_reap = jiffies +
1418 REAPTIMEOUT_LIST3 +
1419 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1420 }
1421}
1422
1423/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001424 * Initialisation. Called after the page allocator have been initialised and
1425 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 */
1427void __init kmem_cache_init(void)
1428{
1429 size_t left_over;
1430 struct cache_sizes *sizes;
1431 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001432 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001433 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001434 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001435
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001436 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001437 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001438 numa_platform = 0;
1439 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001440
Christoph Lametere498be72005-09-09 13:03:32 -07001441 for (i = 0; i < NUM_INIT_LISTS; i++) {
1442 kmem_list3_init(&initkmem_list3[i]);
1443 if (i < MAX_NUMNODES)
1444 cache_cache.nodelists[i] = NULL;
1445 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001446 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447
1448 /*
1449 * Fragmentation resistance on low memory - only use bigger
1450 * page orders on machines with more than 32MB of memory.
1451 */
1452 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1453 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1454
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 /* Bootstrap is tricky, because several objects are allocated
1456 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001457 * 1) initialize the cache_cache cache: it contains the struct
1458 * kmem_cache structures of all caches, except cache_cache itself:
1459 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001460 * Initially an __init data area is used for the head array and the
1461 * kmem_list3 structures, it's replaced with a kmalloc allocated
1462 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001464 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001465 * An __init data area is used for the head array.
1466 * 3) Create the remaining kmalloc caches, with minimally sized
1467 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 * 4) Replace the __init data head arrays for cache_cache and the first
1469 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001470 * 5) Replace the __init data for kmem_list3 for cache_cache and
1471 * the other cache's with kmalloc allocated memory.
1472 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 */
1474
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001475 node = numa_node_id();
1476
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 INIT_LIST_HEAD(&cache_chain);
1479 list_add(&cache_cache.next, &cache_chain);
1480 cache_cache.colour_off = cache_line_size();
1481 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001482 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
Eric Dumazet8da34302007-05-06 14:49:29 -07001484 /*
1485 * struct kmem_cache size depends on nr_node_ids, which
1486 * can be less than MAX_NUMNODES.
1487 */
1488 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1489 nr_node_ids * sizeof(struct kmem_list3 *);
1490#if DEBUG
1491 cache_cache.obj_size = cache_cache.buffer_size;
1492#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001493 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1494 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001495 cache_cache.reciprocal_buffer_size =
1496 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497
Jack Steiner07ed76b2006-03-07 21:55:46 -08001498 for (order = 0; order < MAX_ORDER; order++) {
1499 cache_estimate(order, cache_cache.buffer_size,
1500 cache_line_size(), 0, &left_over, &cache_cache.num);
1501 if (cache_cache.num)
1502 break;
1503 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001504 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001505 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001506 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001507 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1508 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
1510 /* 2+3) create the kmalloc caches */
1511 sizes = malloc_sizes;
1512 names = cache_names;
1513
Andrew Mortona737b3e2006-03-22 00:08:11 -08001514 /*
1515 * Initialize the caches that provide memory for the array cache and the
1516 * kmem_list3 structures first. Without this, further allocations will
1517 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001518 */
1519
1520 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001521 sizes[INDEX_AC].cs_size,
1522 ARCH_KMALLOC_MINALIGN,
1523 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001524 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001525
Andrew Mortona737b3e2006-03-22 00:08:11 -08001526 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001527 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001528 kmem_cache_create(names[INDEX_L3].name,
1529 sizes[INDEX_L3].cs_size,
1530 ARCH_KMALLOC_MINALIGN,
1531 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001532 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001533 }
Christoph Lametere498be72005-09-09 13:03:32 -07001534
Ingo Molnare0a42722006-06-23 02:03:46 -07001535 slab_early_init = 0;
1536
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001538 /*
1539 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 * This should be particularly beneficial on SMP boxes, as it
1541 * eliminates "false sharing".
1542 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001543 * allow tighter packing of the smaller caches.
1544 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001545 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001546 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001547 sizes->cs_size,
1548 ARCH_KMALLOC_MINALIGN,
1549 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001550 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001551 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001552#ifdef CONFIG_ZONE_DMA
1553 sizes->cs_dmacachep = kmem_cache_create(
1554 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001555 sizes->cs_size,
1556 ARCH_KMALLOC_MINALIGN,
1557 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1558 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001559 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001560#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 sizes++;
1562 names++;
1563 }
1564 /* 4) Replace the bootstrap head arrays */
1565 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001566 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001567
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001569
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001571 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1572 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001573 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001574 /*
1575 * Do not assume that spinlocks can be initialized via memcpy:
1576 */
1577 spin_lock_init(&ptr->lock);
1578
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 cache_cache.array[smp_processor_id()] = ptr;
1580 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001581
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001583
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001585 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001586 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001587 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001588 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001589 /*
1590 * Do not assume that spinlocks can be initialized via memcpy:
1591 */
1592 spin_lock_init(&ptr->lock);
1593
Christoph Lametere498be72005-09-09 13:03:32 -07001594 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001595 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 local_irq_enable();
1597 }
Christoph Lametere498be72005-09-09 13:03:32 -07001598 /* 5) Replace the bootstrap kmem_list3's */
1599 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001600 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Mel Gorman9c09a952008-01-24 05:49:54 -08001602 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001603 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001604
Christoph Lametere498be72005-09-09 13:03:32 -07001605 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001606 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001607
1608 if (INDEX_AC != INDEX_L3) {
1609 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001610 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001611 }
1612 }
1613 }
1614
1615 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001617 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001618 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001620 if (enable_cpucache(cachep))
1621 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001622 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 }
1624
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001625 /* Annotate slab for lockdep -- annotate the malloc caches */
1626 init_lock_keys();
1627
1628
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 /* Done! */
1630 g_cpucache_up = FULL;
1631
Andrew Mortona737b3e2006-03-22 00:08:11 -08001632 /*
1633 * Register a cpu startup notifier callback that initializes
1634 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 */
1636 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637
Andrew Mortona737b3e2006-03-22 00:08:11 -08001638 /*
1639 * The reap timers are started later, with a module init call: That part
1640 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 */
1642}
1643
1644static int __init cpucache_init(void)
1645{
1646 int cpu;
1647
Andrew Mortona737b3e2006-03-22 00:08:11 -08001648 /*
1649 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 */
Christoph Lametere498be72005-09-09 13:03:32 -07001651 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001652 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 return 0;
1654}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655__initcall(cpucache_init);
1656
1657/*
1658 * Interface to system's page allocator. No need to hold the cache-lock.
1659 *
1660 * If we requested dmaable memory, we will get it. Even if we
1661 * did not request dmaable memory, we might get it, but that
1662 * would be relatively rare and ignorable.
1663 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001664static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665{
1666 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001667 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 int i;
1669
Luke Yangd6fef9d2006-04-10 22:52:56 -07001670#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001671 /*
1672 * Nommu uses slab's for process anonymous memory allocations, and thus
1673 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001674 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001675 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001676#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001677
Christoph Lameter3c517a62006-12-06 20:33:29 -08001678 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001679 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1680 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001681
1682 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 if (!page)
1684 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001686 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001688 add_zone_page_state(page_zone(page),
1689 NR_SLAB_RECLAIMABLE, nr_pages);
1690 else
1691 add_zone_page_state(page_zone(page),
1692 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001693 for (i = 0; i < nr_pages; i++)
1694 __SetPageSlab(page + i);
1695 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696}
1697
1698/*
1699 * Interface to system's page release.
1700 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001701static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001703 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 struct page *page = virt_to_page(addr);
1705 const unsigned long nr_freed = i;
1706
Christoph Lameter972d1a72006-09-25 23:31:51 -07001707 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1708 sub_zone_page_state(page_zone(page),
1709 NR_SLAB_RECLAIMABLE, nr_freed);
1710 else
1711 sub_zone_page_state(page_zone(page),
1712 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001714 BUG_ON(!PageSlab(page));
1715 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 page++;
1717 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 if (current->reclaim_state)
1719 current->reclaim_state->reclaimed_slab += nr_freed;
1720 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721}
1722
1723static void kmem_rcu_free(struct rcu_head *head)
1724{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001725 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001726 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727
1728 kmem_freepages(cachep, slab_rcu->addr);
1729 if (OFF_SLAB(cachep))
1730 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1731}
1732
1733#if DEBUG
1734
1735#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001736static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001737 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001739 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001741 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001743 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 return;
1745
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001746 *addr++ = 0x12345678;
1747 *addr++ = caller;
1748 *addr++ = smp_processor_id();
1749 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 {
1751 unsigned long *sptr = &caller;
1752 unsigned long svalue;
1753
1754 while (!kstack_end(sptr)) {
1755 svalue = *sptr++;
1756 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001757 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 size -= sizeof(unsigned long);
1759 if (size <= sizeof(unsigned long))
1760 break;
1761 }
1762 }
1763
1764 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001765 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766}
1767#endif
1768
Pekka Enberg343e0d72006-02-01 03:05:50 -08001769static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001771 int size = obj_size(cachep);
1772 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773
1774 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001775 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776}
1777
1778static void dump_line(char *data, int offset, int limit)
1779{
1780 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001781 unsigned char error = 0;
1782 int bad_count = 0;
1783
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001785 for (i = 0; i < limit; i++) {
1786 if (data[offset + i] != POISON_FREE) {
1787 error = data[offset + i];
1788 bad_count++;
1789 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001790 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001791 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001793
1794 if (bad_count == 1) {
1795 error ^= POISON_FREE;
1796 if (!(error & (error - 1))) {
1797 printk(KERN_ERR "Single bit error detected. Probably "
1798 "bad RAM.\n");
1799#ifdef CONFIG_X86
1800 printk(KERN_ERR "Run memtest86+ or a similar memory "
1801 "test tool.\n");
1802#else
1803 printk(KERN_ERR "Run a memory test tool.\n");
1804#endif
1805 }
1806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807}
1808#endif
1809
1810#if DEBUG
1811
Pekka Enberg343e0d72006-02-01 03:05:50 -08001812static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813{
1814 int i, size;
1815 char *realobj;
1816
1817 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001818 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001819 *dbg_redzone1(cachep, objp),
1820 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 }
1822
1823 if (cachep->flags & SLAB_STORE_USER) {
1824 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001825 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001827 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 printk("\n");
1829 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001830 realobj = (char *)objp + obj_offset(cachep);
1831 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001832 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 int limit;
1834 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001835 if (i + limit > size)
1836 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837 dump_line(realobj, i, limit);
1838 }
1839}
1840
Pekka Enberg343e0d72006-02-01 03:05:50 -08001841static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842{
1843 char *realobj;
1844 int size, i;
1845 int lines = 0;
1846
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001847 realobj = (char *)objp + obj_offset(cachep);
1848 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001850 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001852 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 exp = POISON_END;
1854 if (realobj[i] != exp) {
1855 int limit;
1856 /* Mismatch ! */
1857 /* Print header */
1858 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001859 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001860 "Slab corruption: %s start=%p, len=%d\n",
1861 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 print_objinfo(cachep, objp, 0);
1863 }
1864 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001865 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001867 if (i + limit > size)
1868 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 dump_line(realobj, i, limit);
1870 i += 16;
1871 lines++;
1872 /* Limit to 5 lines */
1873 if (lines > 5)
1874 break;
1875 }
1876 }
1877 if (lines != 0) {
1878 /* Print some data about the neighboring objects, if they
1879 * exist:
1880 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001881 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001882 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001884 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001886 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001887 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001889 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 print_objinfo(cachep, objp, 2);
1891 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001892 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001893 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001894 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001896 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 print_objinfo(cachep, objp, 2);
1898 }
1899 }
1900}
1901#endif
1902
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001904/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001905 * slab_destroy_objs - destroy a slab and its objects
1906 * @cachep: cache pointer being destroyed
1907 * @slabp: slab pointer being destroyed
1908 *
1909 * Call the registered destructor for each object in a slab that is being
1910 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001911 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001912static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001913{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 int i;
1915 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001916 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
1918 if (cachep->flags & SLAB_POISON) {
1919#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001920 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1921 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001923 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 else
1925 check_poison_obj(cachep, objp);
1926#else
1927 check_poison_obj(cachep, objp);
1928#endif
1929 }
1930 if (cachep->flags & SLAB_RED_ZONE) {
1931 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1932 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001933 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1935 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001936 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001939}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001941static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001942{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001943}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944#endif
1945
Randy Dunlap911851e2006-03-22 00:08:14 -08001946/**
1947 * slab_destroy - destroy and release all objects in a slab
1948 * @cachep: cache pointer being destroyed
1949 * @slabp: slab pointer being destroyed
1950 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001951 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001952 * Before calling the slab must have been unlinked from the cache. The
1953 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001954 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001955static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001956{
1957 void *addr = slabp->s_mem - slabp->colouroff;
1958
1959 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1961 struct slab_rcu *slab_rcu;
1962
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001963 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 slab_rcu->cachep = cachep;
1965 slab_rcu->addr = addr;
1966 call_rcu(&slab_rcu->head, kmem_rcu_free);
1967 } else {
1968 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001969 if (OFF_SLAB(cachep))
1970 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 }
1972}
1973
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001974static void __kmem_cache_destroy(struct kmem_cache *cachep)
1975{
1976 int i;
1977 struct kmem_list3 *l3;
1978
1979 for_each_online_cpu(i)
1980 kfree(cachep->array[i]);
1981
1982 /* NUMA: free the list3 structures */
1983 for_each_online_node(i) {
1984 l3 = cachep->nodelists[i];
1985 if (l3) {
1986 kfree(l3->shared);
1987 free_alien_cache(l3->alien);
1988 kfree(l3);
1989 }
1990 }
1991 kmem_cache_free(&cache_cache, cachep);
1992}
1993
1994
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001996 * calculate_slab_order - calculate size (page order) of slabs
1997 * @cachep: pointer to the cache that is being created
1998 * @size: size of objects to be created in this cache.
1999 * @align: required alignment for the objects.
2000 * @flags: slab allocation flags
2001 *
2002 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002003 *
2004 * This could be made much more intelligent. For now, try to avoid using
2005 * high order pages for slabs. When the gfp() functions are more friendly
2006 * towards high-order requests, this should be changed.
2007 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002008static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002009 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002011 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002013 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002014
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002015 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002016 unsigned int num;
2017 size_t remainder;
2018
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002019 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002020 if (!num)
2021 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002022
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002023 if (flags & CFLGS_OFF_SLAB) {
2024 /*
2025 * Max number of objs-per-slab for caches which
2026 * use off-slab slabs. Needed to avoid a possible
2027 * looping condition in cache_grow().
2028 */
2029 offslab_limit = size - sizeof(struct slab);
2030 offslab_limit /= sizeof(kmem_bufctl_t);
2031
2032 if (num > offslab_limit)
2033 break;
2034 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002035
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002036 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002037 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002038 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002039 left_over = remainder;
2040
2041 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002042 * A VFS-reclaimable slab tends to have most allocations
2043 * as GFP_NOFS and we really don't want to have to be allocating
2044 * higher-order pages when we are unable to shrink dcache.
2045 */
2046 if (flags & SLAB_RECLAIM_ACCOUNT)
2047 break;
2048
2049 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002050 * Large number of objects is good, but very large slabs are
2051 * currently bad for the gfp()s.
2052 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002053 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002054 break;
2055
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002056 /*
2057 * Acceptable internal fragmentation?
2058 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002059 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002060 break;
2061 }
2062 return left_over;
2063}
2064
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002065static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002066{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002067 if (g_cpucache_up == FULL)
2068 return enable_cpucache(cachep);
2069
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002070 if (g_cpucache_up == NONE) {
2071 /*
2072 * Note: the first kmem_cache_create must create the cache
2073 * that's used by kmalloc(24), otherwise the creation of
2074 * further caches will BUG().
2075 */
2076 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2077
2078 /*
2079 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2080 * the first cache, then we need to set up all its list3s,
2081 * otherwise the creation of further caches will BUG().
2082 */
2083 set_up_list3s(cachep, SIZE_AC);
2084 if (INDEX_AC == INDEX_L3)
2085 g_cpucache_up = PARTIAL_L3;
2086 else
2087 g_cpucache_up = PARTIAL_AC;
2088 } else {
2089 cachep->array[smp_processor_id()] =
2090 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2091
2092 if (g_cpucache_up == PARTIAL_AC) {
2093 set_up_list3s(cachep, SIZE_L3);
2094 g_cpucache_up = PARTIAL_L3;
2095 } else {
2096 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002097 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002098 cachep->nodelists[node] =
2099 kmalloc_node(sizeof(struct kmem_list3),
2100 GFP_KERNEL, node);
2101 BUG_ON(!cachep->nodelists[node]);
2102 kmem_list3_init(cachep->nodelists[node]);
2103 }
2104 }
2105 }
2106 cachep->nodelists[numa_node_id()]->next_reap =
2107 jiffies + REAPTIMEOUT_LIST3 +
2108 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2109
2110 cpu_cache_get(cachep)->avail = 0;
2111 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2112 cpu_cache_get(cachep)->batchcount = 1;
2113 cpu_cache_get(cachep)->touched = 0;
2114 cachep->batchcount = 1;
2115 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002116 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002117}
2118
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002119/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 * kmem_cache_create - Create a cache.
2121 * @name: A string which is used in /proc/slabinfo to identify this cache.
2122 * @size: The size of objects to be created in this cache.
2123 * @align: The required alignment for the objects.
2124 * @flags: SLAB flags
2125 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 *
2127 * Returns a ptr to the cache on success, NULL on failure.
2128 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002129 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 *
2131 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002132 * the module calling this has to destroy the cache before getting unloaded.
2133 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 * The flags are
2135 *
2136 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2137 * to catch references to uninitialised memory.
2138 *
2139 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2140 * for buffer overruns.
2141 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2143 * cacheline. This can be beneficial if you're counting cycles as closely
2144 * as davem.
2145 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002146struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002148 unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002149 void (*ctor)(struct kmem_cache *, void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150{
2151 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002152 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153
2154 /*
2155 * Sanity checks... these are all serious usage bugs.
2156 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002157 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002158 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002159 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002160 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002161 BUG();
2162 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002164 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002165 * We use cache_chain_mutex to ensure a consistent view of
2166 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002167 */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002168 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002169 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002170
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002171 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002172 char tmp;
2173 int res;
2174
2175 /*
2176 * This happens when the module gets unloaded and doesn't
2177 * destroy its slab cache and no-one else reuses the vmalloc
2178 * area of the module. Print a warning.
2179 */
Andrew Morton138ae662006-12-06 20:36:41 -08002180 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002181 if (res) {
matzeb4169522007-05-06 14:49:52 -07002182 printk(KERN_ERR
2183 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002184 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002185 continue;
2186 }
2187
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002188 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002189 printk(KERN_ERR
2190 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002191 dump_stack();
2192 goto oops;
2193 }
2194 }
2195
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196#if DEBUG
2197 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198#if FORCED_DEBUG
2199 /*
2200 * Enable redzoning and last user accounting, except for caches with
2201 * large objects, if the increased size would increase the object size
2202 * above the next power of two: caches with object sizes just above a
2203 * power of two have a significant amount of internal fragmentation.
2204 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002205 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2206 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002207 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 if (!(flags & SLAB_DESTROY_BY_RCU))
2209 flags |= SLAB_POISON;
2210#endif
2211 if (flags & SLAB_DESTROY_BY_RCU)
2212 BUG_ON(flags & SLAB_POISON);
2213#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002215 * Always checks flags, a caller might be expecting debug support which
2216 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002218 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219
Andrew Mortona737b3e2006-03-22 00:08:11 -08002220 /*
2221 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 * unaligned accesses for some archs when redzoning is used, and makes
2223 * sure any on-slab bufctl's are also correctly aligned.
2224 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002225 if (size & (BYTES_PER_WORD - 1)) {
2226 size += (BYTES_PER_WORD - 1);
2227 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 }
2229
Andrew Mortona737b3e2006-03-22 00:08:11 -08002230 /* calculate the final buffer alignment: */
2231
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 /* 1) arch recommendation: can be overridden for debug */
2233 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002234 /*
2235 * Default alignment: as specified by the arch code. Except if
2236 * an object is really small, then squeeze multiple objects into
2237 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 */
2239 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002240 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 ralign /= 2;
2242 } else {
2243 ralign = BYTES_PER_WORD;
2244 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002245
2246 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002247 * Redzoning and user store require word alignment or possibly larger.
2248 * Note this will be overridden by architecture or caller mandated
2249 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002250 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002251 if (flags & SLAB_STORE_USER)
2252 ralign = BYTES_PER_WORD;
2253
2254 if (flags & SLAB_RED_ZONE) {
2255 ralign = REDZONE_ALIGN;
2256 /* If redzoning, ensure that the second redzone is suitably
2257 * aligned, by adjusting the object size accordingly. */
2258 size += REDZONE_ALIGN - 1;
2259 size &= ~(REDZONE_ALIGN - 1);
2260 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002261
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002262 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 if (ralign < ARCH_SLAB_MINALIGN) {
2264 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002266 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 if (ralign < align) {
2268 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002270 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002271 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002272 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002273 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002274 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 */
2276 align = ralign;
2277
2278 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002279 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002281 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282
2283#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002284 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
Pekka Enbergca5f9702006-09-25 23:31:25 -07002286 /*
2287 * Both debugging options require word-alignment which is calculated
2288 * into align above.
2289 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002292 cachep->obj_offset += sizeof(unsigned long long);
2293 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 }
2295 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002296 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002297 * the real object. But if the second red zone needs to be
2298 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002300 if (flags & SLAB_RED_ZONE)
2301 size += REDZONE_ALIGN;
2302 else
2303 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 }
2305#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002306 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002307 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2308 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 size = PAGE_SIZE;
2310 }
2311#endif
2312#endif
2313
Ingo Molnare0a42722006-06-23 02:03:46 -07002314 /*
2315 * Determine if the slab management is 'on' or 'off' slab.
2316 * (bootstrapping cannot cope with offslab caches so don't do
2317 * it too early on.)
2318 */
2319 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 /*
2321 * Size is large, assume best to place the slab management obj
2322 * off-slab (should allow better packing of objs).
2323 */
2324 flags |= CFLGS_OFF_SLAB;
2325
2326 size = ALIGN(size, align);
2327
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002328 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329
2330 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002331 printk(KERN_ERR
2332 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 kmem_cache_free(&cache_cache, cachep);
2334 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002335 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002337 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2338 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339
2340 /*
2341 * If the slab has been placed off-slab, and we have enough space then
2342 * move it on-slab. This is at the expense of any extra colouring.
2343 */
2344 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2345 flags &= ~CFLGS_OFF_SLAB;
2346 left_over -= slab_size;
2347 }
2348
2349 if (flags & CFLGS_OFF_SLAB) {
2350 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002351 slab_size =
2352 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 }
2354
2355 cachep->colour_off = cache_line_size();
2356 /* Offset must be a multiple of the alignment. */
2357 if (cachep->colour_off < align)
2358 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002359 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 cachep->slab_size = slab_size;
2361 cachep->flags = flags;
2362 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002363 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002365 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002366 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002368 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002369 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002370 /*
2371 * This is a possibility for one of the malloc_sizes caches.
2372 * But since we go off slab only for object size greater than
2373 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2374 * this should not happen at all.
2375 * But leave a BUG_ON for some lucky dude.
2376 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002377 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002378 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 cachep->name = name;
2381
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002382 if (setup_cpu_cache(cachep)) {
2383 __kmem_cache_destroy(cachep);
2384 cachep = NULL;
2385 goto oops;
2386 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 /* cache setup completed, link it into the list */
2389 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002390oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 if (!cachep && (flags & SLAB_PANIC))
2392 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002393 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002394 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002395 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 return cachep;
2397}
2398EXPORT_SYMBOL(kmem_cache_create);
2399
2400#if DEBUG
2401static void check_irq_off(void)
2402{
2403 BUG_ON(!irqs_disabled());
2404}
2405
2406static void check_irq_on(void)
2407{
2408 BUG_ON(irqs_disabled());
2409}
2410
Pekka Enberg343e0d72006-02-01 03:05:50 -08002411static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412{
2413#ifdef CONFIG_SMP
2414 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002415 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416#endif
2417}
Christoph Lametere498be72005-09-09 13:03:32 -07002418
Pekka Enberg343e0d72006-02-01 03:05:50 -08002419static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002420{
2421#ifdef CONFIG_SMP
2422 check_irq_off();
2423 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2424#endif
2425}
2426
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427#else
2428#define check_irq_off() do { } while(0)
2429#define check_irq_on() do { } while(0)
2430#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002431#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432#endif
2433
Christoph Lameteraab22072006-03-22 00:09:06 -08002434static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2435 struct array_cache *ac,
2436 int force, int node);
2437
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438static void do_drain(void *arg)
2439{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002440 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002442 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443
2444 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002445 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002446 spin_lock(&cachep->nodelists[node]->list_lock);
2447 free_block(cachep, ac->entry, ac->avail, node);
2448 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 ac->avail = 0;
2450}
2451
Pekka Enberg343e0d72006-02-01 03:05:50 -08002452static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453{
Christoph Lametere498be72005-09-09 13:03:32 -07002454 struct kmem_list3 *l3;
2455 int node;
2456
Andrew Mortona07fa392006-03-22 00:08:17 -08002457 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002459 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002460 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002461 if (l3 && l3->alien)
2462 drain_alien_cache(cachep, l3->alien);
2463 }
2464
2465 for_each_online_node(node) {
2466 l3 = cachep->nodelists[node];
2467 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002468 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002469 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470}
2471
Christoph Lametered11d9e2006-06-30 01:55:45 -07002472/*
2473 * Remove slabs from the list of free slabs.
2474 * Specify the number of slabs to drain in tofree.
2475 *
2476 * Returns the actual number of slabs released.
2477 */
2478static int drain_freelist(struct kmem_cache *cache,
2479 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002481 struct list_head *p;
2482 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484
Christoph Lametered11d9e2006-06-30 01:55:45 -07002485 nr_freed = 0;
2486 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487
Christoph Lametered11d9e2006-06-30 01:55:45 -07002488 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002489 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002490 if (p == &l3->slabs_free) {
2491 spin_unlock_irq(&l3->list_lock);
2492 goto out;
2493 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494
Christoph Lametered11d9e2006-06-30 01:55:45 -07002495 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002497 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498#endif
2499 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002500 /*
2501 * Safe to drop the lock. The slab is no longer linked
2502 * to the cache.
2503 */
2504 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002505 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002506 slab_destroy(cache, slabp);
2507 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002509out:
2510 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511}
2512
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002513/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002514static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002515{
2516 int ret = 0, i = 0;
2517 struct kmem_list3 *l3;
2518
2519 drain_cpu_caches(cachep);
2520
2521 check_irq_on();
2522 for_each_online_node(i) {
2523 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002524 if (!l3)
2525 continue;
2526
2527 drain_freelist(cachep, l3, l3->free_objects);
2528
2529 ret += !list_empty(&l3->slabs_full) ||
2530 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002531 }
2532 return (ret ? 1 : 0);
2533}
2534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535/**
2536 * kmem_cache_shrink - Shrink a cache.
2537 * @cachep: The cache to shrink.
2538 *
2539 * Releases as many slabs as possible for a cache.
2540 * To help debugging, a zero exit status indicates all slabs were released.
2541 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002542int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002544 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002545 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002547 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002548 mutex_lock(&cache_chain_mutex);
2549 ret = __cache_shrink(cachep);
2550 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002551 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002552 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553}
2554EXPORT_SYMBOL(kmem_cache_shrink);
2555
2556/**
2557 * kmem_cache_destroy - delete a cache
2558 * @cachep: the cache to destroy
2559 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002560 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 *
2562 * It is expected this function will be called by a module when it is
2563 * unloaded. This will remove the cache completely, and avoid a duplicate
2564 * cache being allocated each time a module is loaded and unloaded, if the
2565 * module doesn't have persistent in-kernel storage across loads and unloads.
2566 *
2567 * The cache must be empty before calling this function.
2568 *
2569 * The caller must guarantee that noone will allocate memory from the cache
2570 * during the kmem_cache_destroy().
2571 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002572void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002574 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002577 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002578 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 /*
2580 * the chain is never empty, cache_cache is never destroyed
2581 */
2582 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 if (__cache_shrink(cachep)) {
2584 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002585 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002586 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002587 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002588 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 }
2590
2591 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002592 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002594 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002595 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002596 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597}
2598EXPORT_SYMBOL(kmem_cache_destroy);
2599
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002600/*
2601 * Get the memory for a slab management obj.
2602 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2603 * always come from malloc_sizes caches. The slab descriptor cannot
2604 * come from the same cache which is getting created because,
2605 * when we are searching for an appropriate cache for these
2606 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2607 * If we are creating a malloc_sizes cache here it would not be visible to
2608 * kmem_find_general_cachep till the initialization is complete.
2609 * Hence we cannot have slabp_cache same as the original cache.
2610 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002611static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002612 int colour_off, gfp_t local_flags,
2613 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614{
2615 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002616
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 if (OFF_SLAB(cachep)) {
2618 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002619 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002620 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 if (!slabp)
2622 return NULL;
2623 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002624 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 colour_off += cachep->slab_size;
2626 }
2627 slabp->inuse = 0;
2628 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002629 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002630 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002631 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 return slabp;
2633}
2634
2635static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2636{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002637 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638}
2639
Pekka Enberg343e0d72006-02-01 03:05:50 -08002640static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002641 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642{
2643 int i;
2644
2645 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002646 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647#if DEBUG
2648 /* need to poison the objs? */
2649 if (cachep->flags & SLAB_POISON)
2650 poison_obj(cachep, objp, POISON_FREE);
2651 if (cachep->flags & SLAB_STORE_USER)
2652 *dbg_userword(cachep, objp) = NULL;
2653
2654 if (cachep->flags & SLAB_RED_ZONE) {
2655 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2656 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2657 }
2658 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002659 * Constructors are not allowed to allocate memory from the same
2660 * cache which they are a constructor for. Otherwise, deadlock.
2661 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 */
2663 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002664 cachep->ctor(cachep, objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665
2666 if (cachep->flags & SLAB_RED_ZONE) {
2667 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2668 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002669 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2671 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002672 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002674 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2675 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002677 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678#else
2679 if (cachep->ctor)
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002680 cachep->ctor(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002682 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002684 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685}
2686
Pekka Enberg343e0d72006-02-01 03:05:50 -08002687static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002689 if (CONFIG_ZONE_DMA_FLAG) {
2690 if (flags & GFP_DMA)
2691 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2692 else
2693 BUG_ON(cachep->gfpflags & GFP_DMA);
2694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695}
2696
Andrew Mortona737b3e2006-03-22 00:08:11 -08002697static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2698 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002699{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002700 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002701 kmem_bufctl_t next;
2702
2703 slabp->inuse++;
2704 next = slab_bufctl(slabp)[slabp->free];
2705#if DEBUG
2706 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2707 WARN_ON(slabp->nodeid != nodeid);
2708#endif
2709 slabp->free = next;
2710
2711 return objp;
2712}
2713
Andrew Mortona737b3e2006-03-22 00:08:11 -08002714static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2715 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002716{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002717 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002718
2719#if DEBUG
2720 /* Verify that the slab belongs to the intended node */
2721 WARN_ON(slabp->nodeid != nodeid);
2722
Al Viro871751e2006-03-25 03:06:39 -08002723 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002724 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002725 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002726 BUG();
2727 }
2728#endif
2729 slab_bufctl(slabp)[objnr] = slabp->free;
2730 slabp->free = objnr;
2731 slabp->inuse--;
2732}
2733
Pekka Enberg47768742006-06-23 02:03:07 -07002734/*
2735 * Map pages beginning at addr to the given cache and slab. This is required
2736 * for the slab allocator to be able to lookup the cache and slab of a
2737 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2738 */
2739static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2740 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741{
Pekka Enberg47768742006-06-23 02:03:07 -07002742 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 struct page *page;
2744
Pekka Enberg47768742006-06-23 02:03:07 -07002745 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002746
Pekka Enberg47768742006-06-23 02:03:07 -07002747 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002748 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002749 nr_pages <<= cache->gfporder;
2750
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002752 page_set_cache(page, cache);
2753 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002755 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756}
2757
2758/*
2759 * Grow (by 1) the number of slabs within a cache. This is called by
2760 * kmem_cache_alloc() when there are no active objs left in a cache.
2761 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002762static int cache_grow(struct kmem_cache *cachep,
2763 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002765 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002766 size_t offset;
2767 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002768 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769
Andrew Mortona737b3e2006-03-22 00:08:11 -08002770 /*
2771 * Be lazy and only check for valid flags here, keeping it out of the
2772 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002774 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2775 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002777 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002779 l3 = cachep->nodelists[nodeid];
2780 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781
2782 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002783 offset = l3->colour_next;
2784 l3->colour_next++;
2785 if (l3->colour_next >= cachep->colour)
2786 l3->colour_next = 0;
2787 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002789 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790
2791 if (local_flags & __GFP_WAIT)
2792 local_irq_enable();
2793
2794 /*
2795 * The test for missing atomic flag is performed here, rather than
2796 * the more obvious place, simply to reduce the critical path length
2797 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2798 * will eventually be caught here (where it matters).
2799 */
2800 kmem_flagcheck(cachep, flags);
2801
Andrew Mortona737b3e2006-03-22 00:08:11 -08002802 /*
2803 * Get mem for the objs. Attempt to allocate a physical page from
2804 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002805 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002806 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002807 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002808 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 goto failed;
2810
2811 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002812 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002813 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002814 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 goto opps1;
2816
Pekka Enberg47768742006-06-23 02:03:07 -07002817 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818
Christoph Lametera35afb82007-05-16 22:10:57 -07002819 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
2821 if (local_flags & __GFP_WAIT)
2822 local_irq_disable();
2823 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002824 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825
2826 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002827 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002829 l3->free_objects += cachep->num;
2830 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002832opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002834failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 if (local_flags & __GFP_WAIT)
2836 local_irq_disable();
2837 return 0;
2838}
2839
2840#if DEBUG
2841
2842/*
2843 * Perform extra freeing checks:
2844 * - detect bad pointers.
2845 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 */
2847static void kfree_debugcheck(const void *objp)
2848{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 if (!virt_addr_valid(objp)) {
2850 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002851 (unsigned long)objp);
2852 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854}
2855
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002856static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2857{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002858 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002859
2860 redzone1 = *dbg_redzone1(cache, obj);
2861 redzone2 = *dbg_redzone2(cache, obj);
2862
2863 /*
2864 * Redzone is ok.
2865 */
2866 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2867 return;
2868
2869 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2870 slab_error(cache, "double free detected");
2871 else
2872 slab_error(cache, "memory outside object was overwritten");
2873
David Woodhouseb46b8f12007-05-08 00:22:59 -07002874 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002875 obj, redzone1, redzone2);
2876}
2877
Pekka Enberg343e0d72006-02-01 03:05:50 -08002878static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002879 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880{
2881 struct page *page;
2882 unsigned int objnr;
2883 struct slab *slabp;
2884
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002885 BUG_ON(virt_to_cache(objp) != cachep);
2886
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002887 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002889 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890
Pekka Enberg065d41c2005-11-13 16:06:46 -08002891 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892
2893 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002894 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2896 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2897 }
2898 if (cachep->flags & SLAB_STORE_USER)
2899 *dbg_userword(cachep, objp) = caller;
2900
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002901 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902
2903 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002904 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
Al Viro871751e2006-03-25 03:06:39 -08002906#ifdef CONFIG_DEBUG_SLAB_LEAK
2907 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2908#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 if (cachep->flags & SLAB_POISON) {
2910#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002911 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002913 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002914 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 } else {
2916 poison_obj(cachep, objp, POISON_FREE);
2917 }
2918#else
2919 poison_obj(cachep, objp, POISON_FREE);
2920#endif
2921 }
2922 return objp;
2923}
2924
Pekka Enberg343e0d72006-02-01 03:05:50 -08002925static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926{
2927 kmem_bufctl_t i;
2928 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002929
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 /* Check slab's freelist to see if this obj is there. */
2931 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2932 entries++;
2933 if (entries > cachep->num || i >= cachep->num)
2934 goto bad;
2935 }
2936 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002937bad:
2938 printk(KERN_ERR "slab: Internal list corruption detected in "
2939 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2940 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002941 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002942 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002943 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002944 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002946 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 }
2948 printk("\n");
2949 BUG();
2950 }
2951}
2952#else
2953#define kfree_debugcheck(x) do { } while(0)
2954#define cache_free_debugcheck(x,objp,z) (objp)
2955#define check_slabp(x,y) do { } while(0)
2956#endif
2957
Pekka Enberg343e0d72006-02-01 03:05:50 -08002958static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959{
2960 int batchcount;
2961 struct kmem_list3 *l3;
2962 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002963 int node;
2964
Andrew Mortona737b3e2006-03-22 00:08:11 -08002965retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002966 check_irq_off();
2967 node = numa_node_id();
2968 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 batchcount = ac->batchcount;
2970 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002971 /*
2972 * If there was little recent activity on this cache, then
2973 * perform only a partial refill. Otherwise we could generate
2974 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 */
2976 batchcount = BATCHREFILL_LIMIT;
2977 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002978 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979
Christoph Lametere498be72005-09-09 13:03:32 -07002980 BUG_ON(ac->avail > 0 || !l3);
2981 spin_lock(&l3->list_lock);
2982
Christoph Lameter3ded1752006-03-25 03:06:44 -08002983 /* See if we can refill from the shared array */
2984 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2985 goto alloc_done;
2986
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 while (batchcount > 0) {
2988 struct list_head *entry;
2989 struct slab *slabp;
2990 /* Get slab alloc is to come from. */
2991 entry = l3->slabs_partial.next;
2992 if (entry == &l3->slabs_partial) {
2993 l3->free_touched = 1;
2994 entry = l3->slabs_free.next;
2995 if (entry == &l3->slabs_free)
2996 goto must_grow;
2997 }
2998
2999 slabp = list_entry(entry, struct slab, list);
3000 check_slabp(cachep, slabp);
3001 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003002
3003 /*
3004 * The slab was either on partial or free list so
3005 * there must be at least one object available for
3006 * allocation.
3007 */
3008 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3009
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 STATS_INC_ALLOCED(cachep);
3012 STATS_INC_ACTIVE(cachep);
3013 STATS_SET_HIGH(cachep);
3014
Matthew Dobson78d382d2006-02-01 03:05:47 -08003015 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003016 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 }
3018 check_slabp(cachep, slabp);
3019
3020 /* move slabp to correct slabp list: */
3021 list_del(&slabp->list);
3022 if (slabp->free == BUFCTL_END)
3023 list_add(&slabp->list, &l3->slabs_full);
3024 else
3025 list_add(&slabp->list, &l3->slabs_partial);
3026 }
3027
Andrew Mortona737b3e2006-03-22 00:08:11 -08003028must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003030alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003031 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
3033 if (unlikely(!ac->avail)) {
3034 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003035 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003036
Andrew Mortona737b3e2006-03-22 00:08:11 -08003037 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003038 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003039 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 return NULL;
3041
Andrew Mortona737b3e2006-03-22 00:08:11 -08003042 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 goto retry;
3044 }
3045 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003046 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047}
3048
Andrew Mortona737b3e2006-03-22 00:08:11 -08003049static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3050 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051{
3052 might_sleep_if(flags & __GFP_WAIT);
3053#if DEBUG
3054 kmem_flagcheck(cachep, flags);
3055#endif
3056}
3057
3058#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003059static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3060 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003062 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003064 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003066 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003067 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003068 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069 else
3070 check_poison_obj(cachep, objp);
3071#else
3072 check_poison_obj(cachep, objp);
3073#endif
3074 poison_obj(cachep, objp, POISON_INUSE);
3075 }
3076 if (cachep->flags & SLAB_STORE_USER)
3077 *dbg_userword(cachep, objp) = caller;
3078
3079 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003080 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3081 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3082 slab_error(cachep, "double free, or memory outside"
3083 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003084 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003085 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003086 objp, *dbg_redzone1(cachep, objp),
3087 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 }
3089 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3090 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3091 }
Al Viro871751e2006-03-25 03:06:39 -08003092#ifdef CONFIG_DEBUG_SLAB_LEAK
3093 {
3094 struct slab *slabp;
3095 unsigned objnr;
3096
Christoph Lameterb49af682007-05-06 14:49:41 -07003097 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003098 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3099 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3100 }
3101#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003102 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003103 if (cachep->ctor && cachep->flags & SLAB_POISON)
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003104 cachep->ctor(cachep, objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003105#if ARCH_SLAB_MINALIGN
3106 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3107 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3108 objp, ARCH_SLAB_MINALIGN);
3109 }
3110#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 return objp;
3112}
3113#else
3114#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3115#endif
3116
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003117#ifdef CONFIG_FAILSLAB
3118
3119static struct failslab_attr {
3120
3121 struct fault_attr attr;
3122
3123 u32 ignore_gfp_wait;
3124#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3125 struct dentry *ignore_gfp_wait_file;
3126#endif
3127
3128} failslab = {
3129 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003130 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003131};
3132
3133static int __init setup_failslab(char *str)
3134{
3135 return setup_fault_attr(&failslab.attr, str);
3136}
3137__setup("failslab=", setup_failslab);
3138
3139static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3140{
3141 if (cachep == &cache_cache)
3142 return 0;
3143 if (flags & __GFP_NOFAIL)
3144 return 0;
3145 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3146 return 0;
3147
3148 return should_fail(&failslab.attr, obj_size(cachep));
3149}
3150
3151#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3152
3153static int __init failslab_debugfs(void)
3154{
3155 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3156 struct dentry *dir;
3157 int err;
3158
Akinobu Mita824ebef2007-05-06 14:49:58 -07003159 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003160 if (err)
3161 return err;
3162 dir = failslab.attr.dentries.dir;
3163
3164 failslab.ignore_gfp_wait_file =
3165 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3166 &failslab.ignore_gfp_wait);
3167
3168 if (!failslab.ignore_gfp_wait_file) {
3169 err = -ENOMEM;
3170 debugfs_remove(failslab.ignore_gfp_wait_file);
3171 cleanup_fault_attr_dentries(&failslab.attr);
3172 }
3173
3174 return err;
3175}
3176
3177late_initcall(failslab_debugfs);
3178
3179#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3180
3181#else /* CONFIG_FAILSLAB */
3182
3183static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3184{
3185 return 0;
3186}
3187
3188#endif /* CONFIG_FAILSLAB */
3189
Pekka Enberg343e0d72006-02-01 03:05:50 -08003190static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003192 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 struct array_cache *ac;
3194
Alok N Kataria5c382302005-09-27 21:45:46 -07003195 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003196
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003197 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 if (likely(ac->avail)) {
3199 STATS_INC_ALLOCHIT(cachep);
3200 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003201 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 } else {
3203 STATS_INC_ALLOCMISS(cachep);
3204 objp = cache_alloc_refill(cachep, flags);
3205 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003206 return objp;
3207}
3208
Christoph Lametere498be72005-09-09 13:03:32 -07003209#ifdef CONFIG_NUMA
3210/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003211 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003212 *
3213 * If we are in_interrupt, then process context, including cpusets and
3214 * mempolicy, may not apply and should not be used for allocation policy.
3215 */
3216static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3217{
3218 int nid_alloc, nid_here;
3219
Christoph Lameter765c4502006-09-27 01:50:08 -07003220 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003221 return NULL;
3222 nid_alloc = nid_here = numa_node_id();
3223 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3224 nid_alloc = cpuset_mem_spread_node();
3225 else if (current->mempolicy)
3226 nid_alloc = slab_node(current->mempolicy);
3227 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003228 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003229 return NULL;
3230}
3231
3232/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003233 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003234 * certain node and fall back is permitted. First we scan all the
3235 * available nodelists for available objects. If that fails then we
3236 * perform an allocation without specifying a node. This allows the page
3237 * allocator to do its reclaim / fallback magic. We then insert the
3238 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003239 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003240static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003241{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003242 struct zonelist *zonelist;
3243 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003244 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003245 struct zone *zone;
3246 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003247 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003248 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003249
3250 if (flags & __GFP_THISNODE)
3251 return NULL;
3252
Mel Gorman0e884602008-04-28 02:12:14 -07003253 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003254 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003255
Christoph Lameter3c517a62006-12-06 20:33:29 -08003256retry:
3257 /*
3258 * Look through allowed nodes for objects available
3259 * from existing per node queues.
3260 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003261 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3262 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003263
Mel Gorman54a6eb52008-04-28 02:12:16 -07003264 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003265 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003266 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003267 obj = ____cache_alloc_node(cache,
3268 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003269 if (obj)
3270 break;
3271 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003272 }
3273
Christoph Lametercfce6602007-05-06 14:50:17 -07003274 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003275 /*
3276 * This allocation will be performed within the constraints
3277 * of the current cpuset / memory policy requirements.
3278 * We may trigger various forms of reclaim on the allowed
3279 * set and go into memory reserves if necessary.
3280 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003281 if (local_flags & __GFP_WAIT)
3282 local_irq_enable();
3283 kmem_flagcheck(cache, flags);
Christoph Lameter9ac33b22008-03-04 12:24:22 -08003284 obj = kmem_getpages(cache, local_flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003285 if (local_flags & __GFP_WAIT)
3286 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003287 if (obj) {
3288 /*
3289 * Insert into the appropriate per node queues
3290 */
3291 nid = page_to_nid(virt_to_page(obj));
3292 if (cache_grow(cache, flags, nid, obj)) {
3293 obj = ____cache_alloc_node(cache,
3294 flags | GFP_THISNODE, nid);
3295 if (!obj)
3296 /*
3297 * Another processor may allocate the
3298 * objects in the slab since we are
3299 * not holding any locks.
3300 */
3301 goto retry;
3302 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003303 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003304 obj = NULL;
3305 }
3306 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003307 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003308 return obj;
3309}
3310
3311/*
Christoph Lametere498be72005-09-09 13:03:32 -07003312 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003314static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003315 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003316{
3317 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318 struct slab *slabp;
3319 struct kmem_list3 *l3;
3320 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003323 l3 = cachep->nodelists[nodeid];
3324 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003325
Andrew Mortona737b3e2006-03-22 00:08:11 -08003326retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003327 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 spin_lock(&l3->list_lock);
3329 entry = l3->slabs_partial.next;
3330 if (entry == &l3->slabs_partial) {
3331 l3->free_touched = 1;
3332 entry = l3->slabs_free.next;
3333 if (entry == &l3->slabs_free)
3334 goto must_grow;
3335 }
Christoph Lametere498be72005-09-09 13:03:32 -07003336
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003337 slabp = list_entry(entry, struct slab, list);
3338 check_spinlock_acquired_node(cachep, nodeid);
3339 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003340
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003341 STATS_INC_NODEALLOCS(cachep);
3342 STATS_INC_ACTIVE(cachep);
3343 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003344
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003345 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Matthew Dobson78d382d2006-02-01 03:05:47 -08003347 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003348 check_slabp(cachep, slabp);
3349 l3->free_objects--;
3350 /* move slabp to correct slabp list: */
3351 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003352
Andrew Mortona737b3e2006-03-22 00:08:11 -08003353 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003354 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003355 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003356 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003357
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003358 spin_unlock(&l3->list_lock);
3359 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003360
Andrew Mortona737b3e2006-03-22 00:08:11 -08003361must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003362 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003363 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003364 if (x)
3365 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003366
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003367 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003368
Andrew Mortona737b3e2006-03-22 00:08:11 -08003369done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003370 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003371}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003372
3373/**
3374 * kmem_cache_alloc_node - Allocate an object on the specified node
3375 * @cachep: The cache to allocate from.
3376 * @flags: See kmalloc().
3377 * @nodeid: node number of the target node.
3378 * @caller: return address of caller, used for debug information
3379 *
3380 * Identical to kmem_cache_alloc but it will allocate memory on the given
3381 * node, which can improve the performance for cpu bound structures.
3382 *
3383 * Fallback to other node is possible if __GFP_THISNODE is not set.
3384 */
3385static __always_inline void *
3386__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3387 void *caller)
3388{
3389 unsigned long save_flags;
3390 void *ptr;
3391
Akinobu Mita824ebef2007-05-06 14:49:58 -07003392 if (should_failslab(cachep, flags))
3393 return NULL;
3394
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003395 cache_alloc_debugcheck_before(cachep, flags);
3396 local_irq_save(save_flags);
3397
3398 if (unlikely(nodeid == -1))
3399 nodeid = numa_node_id();
3400
3401 if (unlikely(!cachep->nodelists[nodeid])) {
3402 /* Node not bootstrapped yet */
3403 ptr = fallback_alloc(cachep, flags);
3404 goto out;
3405 }
3406
3407 if (nodeid == numa_node_id()) {
3408 /*
3409 * Use the locally cached objects if possible.
3410 * However ____cache_alloc does not allow fallback
3411 * to other nodes. It may fail while we still have
3412 * objects on other nodes available.
3413 */
3414 ptr = ____cache_alloc(cachep, flags);
3415 if (ptr)
3416 goto out;
3417 }
3418 /* ___cache_alloc_node can fall back to other nodes */
3419 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3420 out:
3421 local_irq_restore(save_flags);
3422 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3423
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003424 if (unlikely((flags & __GFP_ZERO) && ptr))
3425 memset(ptr, 0, obj_size(cachep));
3426
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003427 return ptr;
3428}
3429
3430static __always_inline void *
3431__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3432{
3433 void *objp;
3434
3435 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3436 objp = alternate_node_alloc(cache, flags);
3437 if (objp)
3438 goto out;
3439 }
3440 objp = ____cache_alloc(cache, flags);
3441
3442 /*
3443 * We may just have run out of memory on the local node.
3444 * ____cache_alloc_node() knows how to locate memory on other nodes
3445 */
3446 if (!objp)
3447 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3448
3449 out:
3450 return objp;
3451}
3452#else
3453
3454static __always_inline void *
3455__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3456{
3457 return ____cache_alloc(cachep, flags);
3458}
3459
3460#endif /* CONFIG_NUMA */
3461
3462static __always_inline void *
3463__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3464{
3465 unsigned long save_flags;
3466 void *objp;
3467
Akinobu Mita824ebef2007-05-06 14:49:58 -07003468 if (should_failslab(cachep, flags))
3469 return NULL;
3470
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003471 cache_alloc_debugcheck_before(cachep, flags);
3472 local_irq_save(save_flags);
3473 objp = __do_cache_alloc(cachep, flags);
3474 local_irq_restore(save_flags);
3475 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3476 prefetchw(objp);
3477
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003478 if (unlikely((flags & __GFP_ZERO) && objp))
3479 memset(objp, 0, obj_size(cachep));
3480
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003481 return objp;
3482}
Christoph Lametere498be72005-09-09 13:03:32 -07003483
3484/*
3485 * Caller needs to acquire correct kmem_list's list_lock
3486 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003487static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003488 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489{
3490 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003491 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492
3493 for (i = 0; i < nr_objects; i++) {
3494 void *objp = objpp[i];
3495 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003497 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003498 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003500 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003502 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003504 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 check_slabp(cachep, slabp);
3506
3507 /* fixup slab chains */
3508 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003509 if (l3->free_objects > l3->free_limit) {
3510 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003511 /* No need to drop any previously held
3512 * lock here, even if we have a off-slab slab
3513 * descriptor it is guaranteed to come from
3514 * a different cache, refer to comments before
3515 * alloc_slabmgmt.
3516 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 slab_destroy(cachep, slabp);
3518 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003519 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 }
3521 } else {
3522 /* Unconditionally move a slab to the end of the
3523 * partial list on free - maximum time for the
3524 * other objects to be freed, too.
3525 */
Christoph Lametere498be72005-09-09 13:03:32 -07003526 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 }
3528 }
3529}
3530
Pekka Enberg343e0d72006-02-01 03:05:50 -08003531static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532{
3533 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003534 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003535 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
3537 batchcount = ac->batchcount;
3538#if DEBUG
3539 BUG_ON(!batchcount || batchcount > ac->avail);
3540#endif
3541 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003542 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003543 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003544 if (l3->shared) {
3545 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003546 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547 if (max) {
3548 if (batchcount > max)
3549 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003550 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003551 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 shared_array->avail += batchcount;
3553 goto free_done;
3554 }
3555 }
3556
Christoph Lameterff694162005-09-22 21:44:02 -07003557 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003558free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559#if STATS
3560 {
3561 int i = 0;
3562 struct list_head *p;
3563
Christoph Lametere498be72005-09-09 13:03:32 -07003564 p = l3->slabs_free.next;
3565 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 struct slab *slabp;
3567
3568 slabp = list_entry(p, struct slab, list);
3569 BUG_ON(slabp->inuse);
3570
3571 i++;
3572 p = p->next;
3573 }
3574 STATS_SET_FREEABLE(cachep, i);
3575 }
3576#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003577 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003579 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580}
3581
3582/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003583 * Release an obj back to its cache. If the obj has a constructed state, it must
3584 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003586static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003588 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
3590 check_irq_off();
3591 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3592
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003593 /*
3594 * Skip calling cache_free_alien() when the platform is not numa.
3595 * This will avoid cache misses that happen while accessing slabp (which
3596 * is per page memory reference) to get nodeid. Instead use a global
3597 * variable to skip the call, which is mostly likely to be present in
3598 * the cache.
3599 */
3600 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003601 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003602
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 if (likely(ac->avail < ac->limit)) {
3604 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003605 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 return;
3607 } else {
3608 STATS_INC_FREEMISS(cachep);
3609 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003610 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 }
3612}
3613
3614/**
3615 * kmem_cache_alloc - Allocate an object
3616 * @cachep: The cache to allocate from.
3617 * @flags: See kmalloc().
3618 *
3619 * Allocate an object from this cache. The flags are only relevant
3620 * if the cache has no available objects.
3621 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003622void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003624 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625}
3626EXPORT_SYMBOL(kmem_cache_alloc);
3627
3628/**
Randy Dunlap76824862008-03-19 17:00:40 -07003629 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630 * @cachep: the cache we're checking against
3631 * @ptr: pointer to validate
3632 *
Randy Dunlap76824862008-03-19 17:00:40 -07003633 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 * it is _not_ a guarantee that the pointer is actually
3635 * part of the slab cache in question, but it at least
3636 * validates that the pointer can be dereferenced and
3637 * looks half-way sane.
3638 *
3639 * Currently only used for dentry validation.
3640 */
Christoph Lameterb7f869a22006-12-22 01:06:44 -08003641int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003643 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003645 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003646 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647 struct page *page;
3648
3649 if (unlikely(addr < min_addr))
3650 goto out;
3651 if (unlikely(addr > (unsigned long)high_memory - size))
3652 goto out;
3653 if (unlikely(addr & align_mask))
3654 goto out;
3655 if (unlikely(!kern_addr_valid(addr)))
3656 goto out;
3657 if (unlikely(!kern_addr_valid(addr + size - 1)))
3658 goto out;
3659 page = virt_to_page(ptr);
3660 if (unlikely(!PageSlab(page)))
3661 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003662 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663 goto out;
3664 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003665out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666 return 0;
3667}
3668
3669#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003670void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3671{
3672 return __cache_alloc_node(cachep, flags, nodeid,
3673 __builtin_return_address(0));
3674}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675EXPORT_SYMBOL(kmem_cache_alloc_node);
3676
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003677static __always_inline void *
3678__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003679{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003680 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003681
3682 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003683 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3684 return cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003685 return kmem_cache_alloc_node(cachep, flags, node);
3686}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003687
3688#ifdef CONFIG_DEBUG_SLAB
3689void *__kmalloc_node(size_t size, gfp_t flags, int node)
3690{
3691 return __do_kmalloc_node(size, flags, node,
3692 __builtin_return_address(0));
3693}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003694EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003695
3696void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3697 int node, void *caller)
3698{
3699 return __do_kmalloc_node(size, flags, node, caller);
3700}
3701EXPORT_SYMBOL(__kmalloc_node_track_caller);
3702#else
3703void *__kmalloc_node(size_t size, gfp_t flags, int node)
3704{
3705 return __do_kmalloc_node(size, flags, node, NULL);
3706}
3707EXPORT_SYMBOL(__kmalloc_node);
3708#endif /* CONFIG_DEBUG_SLAB */
3709#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710
3711/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003712 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003714 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003715 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003717static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3718 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003720 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003722 /* If you want to save a few bytes .text space: replace
3723 * __ with kmem_.
3724 * Then kmalloc uses the uninlined functions instead of the inline
3725 * functions.
3726 */
3727 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003728 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3729 return cachep;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003730 return __cache_alloc(cachep, flags, caller);
3731}
3732
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003733
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003734#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003735void *__kmalloc(size_t size, gfp_t flags)
3736{
Al Viro871751e2006-03-25 03:06:39 -08003737 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738}
3739EXPORT_SYMBOL(__kmalloc);
3740
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003741void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3742{
3743 return __do_kmalloc(size, flags, caller);
3744}
3745EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003746
3747#else
3748void *__kmalloc(size_t size, gfp_t flags)
3749{
3750 return __do_kmalloc(size, flags, NULL);
3751}
3752EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003753#endif
3754
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755/**
3756 * kmem_cache_free - Deallocate an object
3757 * @cachep: The cache the allocation was from.
3758 * @objp: The previously allocated object.
3759 *
3760 * Free an object which was previously allocated from this
3761 * cache.
3762 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003763void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764{
3765 unsigned long flags;
3766
3767 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003768 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003769 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3770 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003771 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 local_irq_restore(flags);
3773}
3774EXPORT_SYMBOL(kmem_cache_free);
3775
3776/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 * kfree - free previously allocated memory
3778 * @objp: pointer returned by kmalloc.
3779 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003780 * If @objp is NULL, no operation is performed.
3781 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782 * Don't free memory not originally allocated by kmalloc()
3783 * or you will run into trouble.
3784 */
3785void kfree(const void *objp)
3786{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003787 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 unsigned long flags;
3789
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003790 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 return;
3792 local_irq_save(flags);
3793 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003794 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003795 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003796 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003797 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 local_irq_restore(flags);
3799}
3800EXPORT_SYMBOL(kfree);
3801
Pekka Enberg343e0d72006-02-01 03:05:50 -08003802unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003804 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805}
3806EXPORT_SYMBOL(kmem_cache_size);
3807
Pekka Enberg343e0d72006-02-01 03:05:50 -08003808const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003809{
3810 return cachep->name;
3811}
3812EXPORT_SYMBOL_GPL(kmem_cache_name);
3813
Christoph Lametere498be72005-09-09 13:03:32 -07003814/*
Simon Arlott183ff222007-10-20 01:27:18 +02003815 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003816 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003817static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003818{
3819 int node;
3820 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003821 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003822 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003823
Mel Gorman9c09a952008-01-24 05:49:54 -08003824 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003825
Paul Menage3395ee02006-12-06 20:32:16 -08003826 if (use_alien_caches) {
3827 new_alien = alloc_alien_cache(node, cachep->limit);
3828 if (!new_alien)
3829 goto fail;
3830 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003831
Eric Dumazet63109842007-05-06 14:49:28 -07003832 new_shared = NULL;
3833 if (cachep->shared) {
3834 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003835 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003836 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003837 if (!new_shared) {
3838 free_alien_cache(new_alien);
3839 goto fail;
3840 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003841 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003842
Andrew Mortona737b3e2006-03-22 00:08:11 -08003843 l3 = cachep->nodelists[node];
3844 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003845 struct array_cache *shared = l3->shared;
3846
Christoph Lametere498be72005-09-09 13:03:32 -07003847 spin_lock_irq(&l3->list_lock);
3848
Christoph Lametercafeb022006-03-25 03:06:46 -08003849 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003850 free_block(cachep, shared->entry,
3851 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003852
Christoph Lametercafeb022006-03-25 03:06:46 -08003853 l3->shared = new_shared;
3854 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003855 l3->alien = new_alien;
3856 new_alien = NULL;
3857 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003858 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003859 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003860 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003861 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003862 free_alien_cache(new_alien);
3863 continue;
3864 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003865 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003866 if (!l3) {
3867 free_alien_cache(new_alien);
3868 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003869 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003870 }
Christoph Lametere498be72005-09-09 13:03:32 -07003871
3872 kmem_list3_init(l3);
3873 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003874 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003875 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003876 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003877 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003878 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003879 cachep->nodelists[node] = l3;
3880 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003881 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003882
Andrew Mortona737b3e2006-03-22 00:08:11 -08003883fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003884 if (!cachep->next.next) {
3885 /* Cache is not active yet. Roll back what we did */
3886 node--;
3887 while (node >= 0) {
3888 if (cachep->nodelists[node]) {
3889 l3 = cachep->nodelists[node];
3890
3891 kfree(l3->shared);
3892 free_alien_cache(l3->alien);
3893 kfree(l3);
3894 cachep->nodelists[node] = NULL;
3895 }
3896 node--;
3897 }
3898 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003899 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003900}
3901
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003903 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 struct array_cache *new[NR_CPUS];
3905};
3906
3907static void do_ccupdate_local(void *info)
3908{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003909 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910 struct array_cache *old;
3911
3912 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003913 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003914
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3916 new->new[smp_processor_id()] = old;
3917}
3918
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003919/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003920static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3921 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003923 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003924 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003926 new = kzalloc(sizeof(*new), GFP_KERNEL);
3927 if (!new)
3928 return -ENOMEM;
3929
Christoph Lametere498be72005-09-09 13:03:32 -07003930 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003931 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003932 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003933 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003934 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003935 kfree(new->new[i]);
3936 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003937 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938 }
3939 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003940 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003942 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003943
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 cachep->batchcount = batchcount;
3946 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003947 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948
Christoph Lametere498be72005-09-09 13:03:32 -07003949 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003950 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 if (!ccold)
3952 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003953 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003954 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003955 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 kfree(ccold);
3957 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003958 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003959 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960}
3961
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003962/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003963static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964{
3965 int err;
3966 int limit, shared;
3967
Andrew Mortona737b3e2006-03-22 00:08:11 -08003968 /*
3969 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 * - create a LIFO ordering, i.e. return objects that are cache-warm
3971 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003972 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973 * bufctl chains: array operations are cheaper.
3974 * The numbers are guessed, we should auto-tune as described by
3975 * Bonwick.
3976 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003977 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003979 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003981 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003983 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 limit = 54;
3985 else
3986 limit = 120;
3987
Andrew Mortona737b3e2006-03-22 00:08:11 -08003988 /*
3989 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 * allocation behaviour: Most allocs on one cpu, most free operations
3991 * on another cpu. For these cases, an efficient object passing between
3992 * cpus is necessary. This is provided by a shared array. The array
3993 * replaces Bonwick's magazine layer.
3994 * On uniprocessor, it's functionally equivalent (but less efficient)
3995 * to a larger limit. Thus disabled by default.
3996 */
3997 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003998 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000
4001#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004002 /*
4003 * With debugging enabled, large batchcount lead to excessively long
4004 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 */
4006 if (limit > 32)
4007 limit = 32;
4008#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004009 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 if (err)
4011 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004012 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004013 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014}
4015
Christoph Lameter1b552532006-03-22 00:09:07 -08004016/*
4017 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004018 * necessary. Note that the l3 listlock also protects the array_cache
4019 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004020 */
4021void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4022 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023{
4024 int tofree;
4025
Christoph Lameter1b552532006-03-22 00:09:07 -08004026 if (!ac || !ac->avail)
4027 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028 if (ac->touched && !force) {
4029 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004030 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004031 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004032 if (ac->avail) {
4033 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4034 if (tofree > ac->avail)
4035 tofree = (ac->avail + 1) / 2;
4036 free_block(cachep, ac->entry, tofree, node);
4037 ac->avail -= tofree;
4038 memmove(ac->entry, &(ac->entry[tofree]),
4039 sizeof(void *) * ac->avail);
4040 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004041 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 }
4043}
4044
4045/**
4046 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004047 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 *
4049 * Called from workqueue/eventd every few seconds.
4050 * Purpose:
4051 * - clear the per-cpu caches for this CPU.
4052 * - return freeable pages to the main free memory pool.
4053 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004054 * If we cannot acquire the cache chain mutex then just give up - we'll try
4055 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004057static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004059 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004060 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004061 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004062 struct delayed_work *work =
4063 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004065 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004067 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004069 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004070 check_irq_on();
4071
Christoph Lameter35386e32006-03-22 00:09:05 -08004072 /*
4073 * We only take the l3 lock if absolutely necessary and we
4074 * have established with reasonable certainty that
4075 * we can do some work if the lock was obtained.
4076 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004077 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004078
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004079 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080
Christoph Lameteraab22072006-03-22 00:09:06 -08004081 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082
Christoph Lameter35386e32006-03-22 00:09:05 -08004083 /*
4084 * These are racy checks but it does not matter
4085 * if we skip one check or scan twice.
4086 */
Christoph Lametere498be72005-09-09 13:03:32 -07004087 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004088 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089
Christoph Lametere498be72005-09-09 13:03:32 -07004090 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091
Christoph Lameteraab22072006-03-22 00:09:06 -08004092 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004093
Christoph Lametered11d9e2006-06-30 01:55:45 -07004094 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004095 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004096 else {
4097 int freed;
4098
4099 freed = drain_freelist(searchp, l3, (l3->free_limit +
4100 5 * searchp->num - 1) / (5 * searchp->num));
4101 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004103next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 cond_resched();
4105 }
4106 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004107 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004108 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004109out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004110 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004111 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112}
4113
Linus Torvalds158a9622008-01-02 13:04:48 -08004114#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115
Pekka Enberg85289f92006-01-08 01:00:36 -08004116static void print_slabinfo_header(struct seq_file *m)
4117{
4118 /*
4119 * Output format version, so at least we can change it
4120 * without _too_ many complaints.
4121 */
4122#if STATS
4123 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4124#else
4125 seq_puts(m, "slabinfo - version: 2.1\n");
4126#endif
4127 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4128 "<objperslab> <pagesperslab>");
4129 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4130 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4131#if STATS
4132 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004133 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004134 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4135#endif
4136 seq_putc(m, '\n');
4137}
4138
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139static void *s_start(struct seq_file *m, loff_t *pos)
4140{
4141 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004143 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004144 if (!n)
4145 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004146
4147 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148}
4149
4150static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4151{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004152 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153}
4154
4155static void s_stop(struct seq_file *m, void *p)
4156{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004157 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158}
4159
4160static int s_show(struct seq_file *m, void *p)
4161{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004162 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004163 struct slab *slabp;
4164 unsigned long active_objs;
4165 unsigned long num_objs;
4166 unsigned long active_slabs = 0;
4167 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004168 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004170 int node;
4171 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 active_objs = 0;
4174 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004175 for_each_online_node(node) {
4176 l3 = cachep->nodelists[node];
4177 if (!l3)
4178 continue;
4179
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004180 check_irq_on();
4181 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004182
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004183 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004184 if (slabp->inuse != cachep->num && !error)
4185 error = "slabs_full accounting error";
4186 active_objs += cachep->num;
4187 active_slabs++;
4188 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004189 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004190 if (slabp->inuse == cachep->num && !error)
4191 error = "slabs_partial inuse accounting error";
4192 if (!slabp->inuse && !error)
4193 error = "slabs_partial/inuse accounting error";
4194 active_objs += slabp->inuse;
4195 active_slabs++;
4196 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004197 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004198 if (slabp->inuse && !error)
4199 error = "slabs_free/inuse accounting error";
4200 num_slabs++;
4201 }
4202 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004203 if (l3->shared)
4204 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004205
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004206 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004208 num_slabs += active_slabs;
4209 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004210 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211 error = "free_objects accounting error";
4212
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004213 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 if (error)
4215 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4216
4217 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004218 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004219 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004221 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004222 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004223 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004224#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004225 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 unsigned long high = cachep->high_mark;
4227 unsigned long allocs = cachep->num_allocations;
4228 unsigned long grown = cachep->grown;
4229 unsigned long reaped = cachep->reaped;
4230 unsigned long errors = cachep->errors;
4231 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004233 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004234 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004235
Christoph Lametere498be72005-09-09 13:03:32 -07004236 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004237 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004238 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004239 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 }
4241 /* cpu stats */
4242 {
4243 unsigned long allochit = atomic_read(&cachep->allochit);
4244 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4245 unsigned long freehit = atomic_read(&cachep->freehit);
4246 unsigned long freemiss = atomic_read(&cachep->freemiss);
4247
4248 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004249 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250 }
4251#endif
4252 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004253 return 0;
4254}
4255
4256/*
4257 * slabinfo_op - iterator that generates /proc/slabinfo
4258 *
4259 * Output layout:
4260 * cache-name
4261 * num-active-objs
4262 * total-objs
4263 * object size
4264 * num-active-slabs
4265 * total-slabs
4266 * num-pages-per-slab
4267 * + further values on SMP and with statistics enabled
4268 */
4269
Helge Deller15ad7cd2006-12-06 20:40:36 -08004270const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004271 .start = s_start,
4272 .next = s_next,
4273 .stop = s_stop,
4274 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275};
4276
4277#define MAX_SLABINFO_WRITE 128
4278/**
4279 * slabinfo_write - Tuning for the slab allocator
4280 * @file: unused
4281 * @buffer: user buffer
4282 * @count: data length
4283 * @ppos: unused
4284 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004285ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4286 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004287{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004288 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004290 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004291
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292 if (count > MAX_SLABINFO_WRITE)
4293 return -EINVAL;
4294 if (copy_from_user(&kbuf, buffer, count))
4295 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004296 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004297
4298 tmp = strchr(kbuf, ' ');
4299 if (!tmp)
4300 return -EINVAL;
4301 *tmp = '\0';
4302 tmp++;
4303 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4304 return -EINVAL;
4305
4306 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004307 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004309 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004311 if (limit < 1 || batchcount < 1 ||
4312 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004313 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004315 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004316 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004317 }
4318 break;
4319 }
4320 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004321 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004322 if (res >= 0)
4323 res = count;
4324 return res;
4325}
Al Viro871751e2006-03-25 03:06:39 -08004326
4327#ifdef CONFIG_DEBUG_SLAB_LEAK
4328
4329static void *leaks_start(struct seq_file *m, loff_t *pos)
4330{
Al Viro871751e2006-03-25 03:06:39 -08004331 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004332 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004333}
4334
4335static inline int add_caller(unsigned long *n, unsigned long v)
4336{
4337 unsigned long *p;
4338 int l;
4339 if (!v)
4340 return 1;
4341 l = n[1];
4342 p = n + 2;
4343 while (l) {
4344 int i = l/2;
4345 unsigned long *q = p + 2 * i;
4346 if (*q == v) {
4347 q[1]++;
4348 return 1;
4349 }
4350 if (*q > v) {
4351 l = i;
4352 } else {
4353 p = q + 2;
4354 l -= i + 1;
4355 }
4356 }
4357 if (++n[1] == n[0])
4358 return 0;
4359 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4360 p[0] = v;
4361 p[1] = 1;
4362 return 1;
4363}
4364
4365static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4366{
4367 void *p;
4368 int i;
4369 if (n[0] == n[1])
4370 return;
4371 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4372 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4373 continue;
4374 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4375 return;
4376 }
4377}
4378
4379static void show_symbol(struct seq_file *m, unsigned long address)
4380{
4381#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004382 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004383 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004384
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004385 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004386 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004387 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004388 seq_printf(m, " [%s]", modname);
4389 return;
4390 }
4391#endif
4392 seq_printf(m, "%p", (void *)address);
4393}
4394
4395static int leaks_show(struct seq_file *m, void *p)
4396{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004397 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004398 struct slab *slabp;
4399 struct kmem_list3 *l3;
4400 const char *name;
4401 unsigned long *n = m->private;
4402 int node;
4403 int i;
4404
4405 if (!(cachep->flags & SLAB_STORE_USER))
4406 return 0;
4407 if (!(cachep->flags & SLAB_RED_ZONE))
4408 return 0;
4409
4410 /* OK, we can do it */
4411
4412 n[1] = 0;
4413
4414 for_each_online_node(node) {
4415 l3 = cachep->nodelists[node];
4416 if (!l3)
4417 continue;
4418
4419 check_irq_on();
4420 spin_lock_irq(&l3->list_lock);
4421
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004422 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004423 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004424 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004425 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004426 spin_unlock_irq(&l3->list_lock);
4427 }
4428 name = cachep->name;
4429 if (n[0] == n[1]) {
4430 /* Increase the buffer size */
4431 mutex_unlock(&cache_chain_mutex);
4432 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4433 if (!m->private) {
4434 /* Too bad, we are really out */
4435 m->private = n;
4436 mutex_lock(&cache_chain_mutex);
4437 return -ENOMEM;
4438 }
4439 *(unsigned long *)m->private = n[0] * 2;
4440 kfree(n);
4441 mutex_lock(&cache_chain_mutex);
4442 /* Now make sure this entry will be retried */
4443 m->count = m->size;
4444 return 0;
4445 }
4446 for (i = 0; i < n[1]; i++) {
4447 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4448 show_symbol(m, n[2*i+2]);
4449 seq_putc(m, '\n');
4450 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004451
Al Viro871751e2006-03-25 03:06:39 -08004452 return 0;
4453}
4454
Helge Deller15ad7cd2006-12-06 20:40:36 -08004455const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004456 .start = leaks_start,
4457 .next = s_next,
4458 .stop = s_stop,
4459 .show = leaks_show,
4460};
4461#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462#endif
4463
Manfred Spraul00e145b2005-09-03 15:55:07 -07004464/**
4465 * ksize - get the actual amount of memory allocated for a given object
4466 * @objp: Pointer to the object
4467 *
4468 * kmalloc may internally round up allocations and return more memory
4469 * than requested. ksize() can be used to determine the actual amount of
4470 * memory allocated. The caller may use this additional memory, even though
4471 * a smaller amount of memory was initially specified with the kmalloc call.
4472 * The caller must guarantee that objp points to a valid object previously
4473 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4474 * must not be freed during the duration of the call.
4475 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004476size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004477{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004478 BUG_ON(!objp);
4479 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004480 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004482 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004483}
Tetsuo Handaf8fcc932007-12-04 23:45:08 -08004484EXPORT_SYMBOL(ksize);