blob: d1d55279202e0759f1d60c5b19ba483606e1973b [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
89#include <linux/config.h>
90#include <linux/slab.h>
91#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070092#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070093#include <linux/swap.h>
94#include <linux/cache.h>
95#include <linux/interrupt.h>
96#include <linux/init.h>
97#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080098#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
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>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111#include <asm/uaccess.h>
112#include <asm/cacheflush.h>
113#include <asm/tlbflush.h>
114#include <asm/page.h>
115
116/*
117 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
118 * SLAB_RED_ZONE & SLAB_POISON.
119 * 0 for faster, smaller code (especially in the critical paths).
120 *
121 * STATS - 1 to collect stats for /proc/slabinfo.
122 * 0 for faster, smaller code (especially in the critical paths).
123 *
124 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 */
126
127#ifdef CONFIG_DEBUG_SLAB
128#define DEBUG 1
129#define STATS 1
130#define FORCED_DEBUG 1
131#else
132#define DEBUG 0
133#define STATS 0
134#define FORCED_DEBUG 0
135#endif
136
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137/* Shouldn't this be in a header file somewhere? */
138#define BYTES_PER_WORD sizeof(void *)
139
140#ifndef cache_line_size
141#define cache_line_size() L1_CACHE_BYTES
142#endif
143
144#ifndef ARCH_KMALLOC_MINALIGN
145/*
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
150 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
151 * Note that this flag disables some debug features.
152 */
153#define ARCH_KMALLOC_MINALIGN 0
154#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
173# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800175 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800180# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
182 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800183 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184#endif
185
186/*
187 * kmem_bufctl_t:
188 *
189 * Bufctl's are used for linking objs within a slab
190 * linked offsets.
191 *
192 * This implementation relies on "struct page" for locating the cache &
193 * slab an object belongs to.
194 * This allows the bufctl structure to be small (one int), but limits
195 * the number of objects a slab (not a cache) can contain when off-slab
196 * bufctls are used. The limit is the size of the largest general cache
197 * that does not use off-slab slabs.
198 * For 32bit archs with 4 kB pages, is this 56.
199 * This is not serious, as it is only for large objects, when it is unwise
200 * to have too many per slab.
201 * Note: This limit can be raised by introducing a general cache whose size
202 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 */
204
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700205typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
207#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800208#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
209#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211/*
212 * struct slab
213 *
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
217 */
218struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800219 struct list_head list;
220 unsigned long colouroff;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
223 kmem_bufctl_t free;
224 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225};
226
227/*
228 * struct slab_rcu
229 *
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
237 *
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
240 *
241 * We assume struct slab_rcu can overlay struct slab when destroying.
242 */
243struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800244 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800245 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247};
248
249/*
250 * struct array_cache
251 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 * Purpose:
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
256 *
257 * The limit is stored in the per-cpu structure to reduce the data cache
258 * footprint.
259 *
260 */
261struct array_cache {
262 unsigned int avail;
263 unsigned int limit;
264 unsigned int batchcount;
265 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700266 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800267 void *entry[0]; /*
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
270 * the entries.
271 * [0] is for gcc 2.95. It should really be [].
272 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273};
274
Andrew Mortona737b3e2006-03-22 00:08:11 -0800275/*
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 */
279#define BOOT_CPUCACHE_ENTRIES 1
280struct arraycache_init {
281 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800282 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283};
284
285/*
Christoph Lametere498be72005-09-09 13:03:32 -0700286 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 */
288struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800293 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800294 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800298 unsigned long next_reap; /* updated without locking */
299 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300};
301
Christoph Lametere498be72005-09-09 13:03:32 -0700302/*
303 * Need this for bootstrapping a per node allocator.
304 */
305#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307#define CACHE_CACHE 0
308#define SIZE_AC 1
309#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Christoph Lametere498be72005-09-09 13:03:32 -0700311/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800312 * This function must be completely optimized away if a constant is passed to
313 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700314 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700315static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700316{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800317 extern void __bad_size(void);
318
Christoph Lametere498be72005-09-09 13:03:32 -0700319 if (__builtin_constant_p(size)) {
320 int i = 0;
321
322#define CACHE(x) \
323 if (size <=x) \
324 return i; \
325 else \
326 i++;
327#include "linux/kmalloc_sizes.h"
328#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800329 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700330 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800331 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700332 return 0;
333}
334
Ingo Molnare0a42722006-06-23 02:03:46 -0700335static int slab_early_init = 1;
336
Christoph Lametere498be72005-09-09 13:03:32 -0700337#define INDEX_AC index_of(sizeof(struct arraycache_init))
338#define INDEX_L3 index_of(sizeof(struct kmem_list3))
339
Pekka Enberg5295a742006-02-01 03:05:48 -0800340static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700341{
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800347 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700348 spin_lock_init(&parent->list_lock);
349 parent->free_objects = 0;
350 parent->free_touched = 0;
351}
352
Andrew Mortona737b3e2006-03-22 00:08:11 -0800353#define MAKE_LIST(cachep, listp, slab, nodeid) \
354 do { \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700357 } while (0)
358
Andrew Mortona737b3e2006-03-22 00:08:11 -0800359#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
360 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
364 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
366/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800367 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 *
369 * manages a cache.
370 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800371
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800372struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800374 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800375/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800376 unsigned int batchcount;
377 unsigned int limit;
378 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800379
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800380 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800381/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800382 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383
Andrew Mortona737b3e2006-03-22 00:08:11 -0800384 unsigned int flags; /* constant flags */
385 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800387/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800389 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
391 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800392 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Andrew Mortona737b3e2006-03-22 00:08:11 -0800394 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800395 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800396 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800398 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800401 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
403 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800404 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800406/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800407 const char *name;
408 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800410/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800412 unsigned long num_active;
413 unsigned long num_allocations;
414 unsigned long high_mark;
415 unsigned long grown;
416 unsigned long reaped;
417 unsigned long errors;
418 unsigned long max_freeable;
419 unsigned long node_allocs;
420 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700421 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800422 atomic_t allochit;
423 atomic_t allocmiss;
424 atomic_t freehit;
425 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426#endif
427#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800428 /*
429 * If debugging is enabled, then the allocator can add additional
430 * fields and/or padding to every object. buffer_size contains the total
431 * object size including these internal fields, the following two
432 * variables contain the offset to the user object and its size.
433 */
434 int obj_offset;
435 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436#endif
437};
438
439#define CFLGS_OFF_SLAB (0x80000000UL)
440#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
441
442#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800443/*
444 * Optimization question: fewer reaps means less probability for unnessary
445 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100447 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 * which could lock up otherwise freeable slabs.
449 */
450#define REAPTIMEOUT_CPUC (2*HZ)
451#define REAPTIMEOUT_LIST3 (4*HZ)
452
453#if STATS
454#define STATS_INC_ACTIVE(x) ((x)->num_active++)
455#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
456#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
457#define STATS_INC_GROWN(x) ((x)->grown++)
458#define STATS_INC_REAPED(x) ((x)->reaped++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800459#define STATS_SET_HIGH(x) \
460 do { \
461 if ((x)->num_active > (x)->high_mark) \
462 (x)->high_mark = (x)->num_active; \
463 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464#define STATS_INC_ERR(x) ((x)->errors++)
465#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700466#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700467#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800468#define STATS_SET_FREEABLE(x, i) \
469 do { \
470 if ((x)->max_freeable < i) \
471 (x)->max_freeable = i; \
472 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
474#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
475#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
476#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
477#else
478#define STATS_INC_ACTIVE(x) do { } while (0)
479#define STATS_DEC_ACTIVE(x) do { } while (0)
480#define STATS_INC_ALLOCED(x) do { } while (0)
481#define STATS_INC_GROWN(x) do { } while (0)
482#define STATS_INC_REAPED(x) do { } while (0)
483#define STATS_SET_HIGH(x) do { } while (0)
484#define STATS_INC_ERR(x) do { } while (0)
485#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700486#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700487#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800488#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489#define STATS_INC_ALLOCHIT(x) do { } while (0)
490#define STATS_INC_ALLOCMISS(x) do { } while (0)
491#define STATS_INC_FREEHIT(x) do { } while (0)
492#define STATS_INC_FREEMISS(x) do { } while (0)
493#endif
494
495#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496
Andrew Mortona737b3e2006-03-22 00:08:11 -0800497/*
498 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800500 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 * the end of an object is aligned with the end of the real
502 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800503 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800505 * cachep->obj_offset: The real object.
506 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800507 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
508 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800510static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800512 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513}
514
Pekka Enberg343e0d72006-02-01 03:05:50 -0800515static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800517 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518}
519
Pekka Enberg343e0d72006-02-01 03:05:50 -0800520static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
522 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800523 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524}
525
Pekka Enberg343e0d72006-02-01 03:05:50 -0800526static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527{
528 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
529 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800531 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800532 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533}
534
Pekka Enberg343e0d72006-02-01 03:05:50 -0800535static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
537 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800538 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539}
540
541#else
542
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800543#define obj_offset(x) 0
544#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
546#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
547#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
548
549#endif
550
551/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800552 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
553 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 */
555#if defined(CONFIG_LARGE_ALLOCS)
556#define MAX_OBJ_ORDER 13 /* up to 32Mb */
557#define MAX_GFP_ORDER 13 /* up to 32Mb */
558#elif defined(CONFIG_MMU)
559#define MAX_OBJ_ORDER 5 /* 32 pages */
560#define MAX_GFP_ORDER 5 /* 32 pages */
561#else
562#define MAX_OBJ_ORDER 8 /* up to 1Mb */
563#define MAX_GFP_ORDER 8 /* up to 1Mb */
564#endif
565
566/*
567 * Do not go above this order unless 0 objects fit into the slab.
568 */
569#define BREAK_GFP_ORDER_HI 1
570#define BREAK_GFP_ORDER_LO 0
571static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
572
Andrew Mortona737b3e2006-03-22 00:08:11 -0800573/*
574 * Functions for storing/retrieving the cachep and or slab from the page
575 * allocator. These are used to find the slab an obj belongs to. With kfree(),
576 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800578static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
579{
580 page->lru.next = (struct list_head *)cache;
581}
582
583static inline struct kmem_cache *page_get_cache(struct page *page)
584{
Nick Piggin84097512006-03-22 00:08:34 -0800585 if (unlikely(PageCompound(page)))
586 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700587 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800588 return (struct kmem_cache *)page->lru.next;
589}
590
591static inline void page_set_slab(struct page *page, struct slab *slab)
592{
593 page->lru.prev = (struct list_head *)slab;
594}
595
596static inline struct slab *page_get_slab(struct page *page)
597{
Nick Piggin84097512006-03-22 00:08:34 -0800598 if (unlikely(PageCompound(page)))
599 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700600 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800601 return (struct slab *)page->lru.prev;
602}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800604static inline struct kmem_cache *virt_to_cache(const void *obj)
605{
606 struct page *page = virt_to_page(obj);
607 return page_get_cache(page);
608}
609
610static inline struct slab *virt_to_slab(const void *obj)
611{
612 struct page *page = virt_to_page(obj);
613 return page_get_slab(page);
614}
615
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800616static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
617 unsigned int idx)
618{
619 return slab->s_mem + cache->buffer_size * idx;
620}
621
622static inline unsigned int obj_to_index(struct kmem_cache *cache,
623 struct slab *slab, void *obj)
624{
625 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
626}
627
Andrew Mortona737b3e2006-03-22 00:08:11 -0800628/*
629 * These are the default caches for kmalloc. Custom caches can have other sizes.
630 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631struct cache_sizes malloc_sizes[] = {
632#define CACHE(x) { .cs_size = (x) },
633#include <linux/kmalloc_sizes.h>
634 CACHE(ULONG_MAX)
635#undef CACHE
636};
637EXPORT_SYMBOL(malloc_sizes);
638
639/* Must match cache_sizes above. Out of line to keep cache footprint low. */
640struct cache_names {
641 char *name;
642 char *name_dma;
643};
644
645static struct cache_names __initdata cache_names[] = {
646#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
647#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800648 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649#undef CACHE
650};
651
652static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800653 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800655 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
657/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800658static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800659 .batchcount = 1,
660 .limit = BOOT_CPUCACHE_ENTRIES,
661 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800662 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800665 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666#endif
667};
668
669/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800670static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671static struct list_head cache_chain;
672
673/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800674 * vm_enough_memory() looks at this to determine how many slab-allocated pages
675 * are possibly freeable under pressure
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 *
677 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
678 */
679atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
681/*
682 * chicken and egg problem: delay the per-cpu array allocation
683 * until the general caches are up.
684 */
685static enum {
686 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700687 PARTIAL_AC,
688 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 FULL
690} g_cpucache_up;
691
Mike Kravetz39d24e62006-05-15 09:44:13 -0700692/*
693 * used by boot code to determine if it can use slab based allocator
694 */
695int slab_is_available(void)
696{
697 return g_cpucache_up == FULL;
698}
699
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700static DEFINE_PER_CPU(struct work_struct, reap_work);
701
Andrew Mortona737b3e2006-03-22 00:08:11 -0800702static void free_block(struct kmem_cache *cachep, void **objpp, int len,
703 int node);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800704static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800705static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800706static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707
Pekka Enberg343e0d72006-02-01 03:05:50 -0800708static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
710 return cachep->array[smp_processor_id()];
711}
712
Andrew Mortona737b3e2006-03-22 00:08:11 -0800713static inline struct kmem_cache *__find_general_cachep(size_t size,
714 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715{
716 struct cache_sizes *csizep = malloc_sizes;
717
718#if DEBUG
719 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800720 * kmem_cache_create(), or __kmalloc(), before
721 * the generic caches are initialized.
722 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700723 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724#endif
725 while (size > csizep->cs_size)
726 csizep++;
727
728 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700729 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 * has cs_{dma,}cachep==NULL. Thus no special case
731 * for large kmalloc calls required.
732 */
733 if (unlikely(gfpflags & GFP_DMA))
734 return csizep->cs_dmacachep;
735 return csizep->cs_cachep;
736}
737
Pekka Enberg343e0d72006-02-01 03:05:50 -0800738struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700739{
740 return __find_general_cachep(size, gfpflags);
741}
742EXPORT_SYMBOL(kmem_find_general_cachep);
743
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800744static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800746 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
747}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Andrew Mortona737b3e2006-03-22 00:08:11 -0800749/*
750 * Calculate the number of objects and left-over bytes for a given buffer size.
751 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800752static void cache_estimate(unsigned long gfporder, size_t buffer_size,
753 size_t align, int flags, size_t *left_over,
754 unsigned int *num)
755{
756 int nr_objs;
757 size_t mgmt_size;
758 size_t slab_size = PAGE_SIZE << gfporder;
759
760 /*
761 * The slab management structure can be either off the slab or
762 * on it. For the latter case, the memory allocated for a
763 * slab is used for:
764 *
765 * - The struct slab
766 * - One kmem_bufctl_t for each object
767 * - Padding to respect alignment of @align
768 * - @buffer_size bytes for each object
769 *
770 * If the slab management structure is off the slab, then the
771 * alignment will already be calculated into the size. Because
772 * the slabs are all pages aligned, the objects will be at the
773 * correct alignment when allocated.
774 */
775 if (flags & CFLGS_OFF_SLAB) {
776 mgmt_size = 0;
777 nr_objs = slab_size / buffer_size;
778
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
781 } else {
782 /*
783 * Ignore padding for the initial guess. The padding
784 * is at most @align-1 bytes, and @buffer_size is at
785 * least @align. In the worst case, this result will
786 * be one greater than the number of objects that fit
787 * into the memory allocation when taking the padding
788 * into account.
789 */
790 nr_objs = (slab_size - sizeof(struct slab)) /
791 (buffer_size + sizeof(kmem_bufctl_t));
792
793 /*
794 * This calculated number will be either the right
795 * amount, or one greater than what we want.
796 */
797 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
798 > slab_size)
799 nr_objs--;
800
801 if (nr_objs > SLAB_LIMIT)
802 nr_objs = SLAB_LIMIT;
803
804 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800806 *num = nr_objs;
807 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808}
809
810#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
811
Andrew Mortona737b3e2006-03-22 00:08:11 -0800812static void __slab_error(const char *function, struct kmem_cache *cachep,
813 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
815 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800816 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 dump_stack();
818}
819
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800820#ifdef CONFIG_NUMA
821/*
822 * Special reaping functions for NUMA systems called from cache_reap().
823 * These take care of doing round robin flushing of alien caches (containing
824 * objects freed on different nodes from which they were allocated) and the
825 * flushing of remote pcps by calling drain_node_pages.
826 */
827static DEFINE_PER_CPU(unsigned long, reap_node);
828
829static void init_reap_node(int cpu)
830{
831 int node;
832
833 node = next_node(cpu_to_node(cpu), node_online_map);
834 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800835 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800836
837 __get_cpu_var(reap_node) = node;
838}
839
840static void next_reap_node(void)
841{
842 int node = __get_cpu_var(reap_node);
843
844 /*
845 * Also drain per cpu pages on remote zones
846 */
847 if (node != numa_node_id())
848 drain_node_pages(node);
849
850 node = next_node(node, node_online_map);
851 if (unlikely(node >= MAX_NUMNODES))
852 node = first_node(node_online_map);
853 __get_cpu_var(reap_node) = node;
854}
855
856#else
857#define init_reap_node(cpu) do { } while (0)
858#define next_reap_node(void) do { } while (0)
859#endif
860
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861/*
862 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
863 * via the workqueue/eventd.
864 * Add the CPU number into the expiration time to minimize the possibility of
865 * the CPUs getting into lockstep and contending for the global cache chain
866 * lock.
867 */
868static void __devinit start_cpu_timer(int cpu)
869{
870 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
871
872 /*
873 * When this gets called from do_initcalls via cpucache_init(),
874 * init_workqueues() has already run, so keventd will be setup
875 * at that time.
876 */
877 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800878 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 INIT_WORK(reap_work, cache_reap, NULL);
880 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
881 }
882}
883
Christoph Lametere498be72005-09-09 13:03:32 -0700884static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800885 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800887 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 struct array_cache *nc = NULL;
889
Christoph Lametere498be72005-09-09 13:03:32 -0700890 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 if (nc) {
892 nc->avail = 0;
893 nc->limit = entries;
894 nc->batchcount = batchcount;
895 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700896 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 }
898 return nc;
899}
900
Christoph Lameter3ded1752006-03-25 03:06:44 -0800901/*
902 * Transfer objects in one arraycache to another.
903 * Locking must be handled by the caller.
904 *
905 * Return the number of entries transferred.
906 */
907static int transfer_objects(struct array_cache *to,
908 struct array_cache *from, unsigned int max)
909{
910 /* Figure out how many entries to transfer */
911 int nr = min(min(from->avail, max), to->limit - to->avail);
912
913 if (!nr)
914 return 0;
915
916 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
917 sizeof(void *) *nr);
918
919 from->avail -= nr;
920 to->avail += nr;
921 to->touched = 1;
922 return nr;
923}
924
Christoph Lametere498be72005-09-09 13:03:32 -0700925#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800926static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800927static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800928
Pekka Enberg5295a742006-02-01 03:05:48 -0800929static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700930{
931 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800932 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700933 int i;
934
935 if (limit > 1)
936 limit = 12;
937 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
938 if (ac_ptr) {
939 for_each_node(i) {
940 if (i == node || !node_online(i)) {
941 ac_ptr[i] = NULL;
942 continue;
943 }
944 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
945 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800946 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700947 kfree(ac_ptr[i]);
948 kfree(ac_ptr);
949 return NULL;
950 }
951 }
952 }
953 return ac_ptr;
954}
955
Pekka Enberg5295a742006-02-01 03:05:48 -0800956static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700957{
958 int i;
959
960 if (!ac_ptr)
961 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700962 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800963 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700964 kfree(ac_ptr);
965}
966
Pekka Enberg343e0d72006-02-01 03:05:50 -0800967static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800968 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700969{
970 struct kmem_list3 *rl3 = cachep->nodelists[node];
971
972 if (ac->avail) {
973 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800974 /*
975 * Stuff objects into the remote nodes shared array first.
976 * That way we could avoid the overhead of putting the objects
977 * into the free lists and getting them back later.
978 */
shin, jacob693f7d32006-04-28 10:54:37 -0500979 if (rl3->shared)
980 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800981
Christoph Lameterff694162005-09-22 21:44:02 -0700982 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700983 ac->avail = 0;
984 spin_unlock(&rl3->list_lock);
985 }
986}
987
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800988/*
989 * Called from cache_reap() to regularly drain alien caches round robin.
990 */
991static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
992{
993 int node = __get_cpu_var(reap_node);
994
995 if (l3->alien) {
996 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -0800997
998 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800999 __drain_alien_cache(cachep, ac, node);
1000 spin_unlock_irq(&ac->lock);
1001 }
1002 }
1003}
1004
Andrew Mortona737b3e2006-03-22 00:08:11 -08001005static void drain_alien_cache(struct kmem_cache *cachep,
1006 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001007{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001008 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001009 struct array_cache *ac;
1010 unsigned long flags;
1011
1012 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001013 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001014 if (ac) {
1015 spin_lock_irqsave(&ac->lock, flags);
1016 __drain_alien_cache(cachep, ac, i);
1017 spin_unlock_irqrestore(&ac->lock, flags);
1018 }
1019 }
1020}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001021
1022static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1023{
1024 struct slab *slabp = virt_to_slab(objp);
1025 int nodeid = slabp->nodeid;
1026 struct kmem_list3 *l3;
1027 struct array_cache *alien = NULL;
1028
1029 /*
1030 * Make sure we are not freeing a object from another node to the array
1031 * cache on this cpu.
1032 */
1033 if (likely(slabp->nodeid == numa_node_id()))
1034 return 0;
1035
1036 l3 = cachep->nodelists[numa_node_id()];
1037 STATS_INC_NODEFREES(cachep);
1038 if (l3->alien && l3->alien[nodeid]) {
1039 alien = l3->alien[nodeid];
1040 spin_lock(&alien->lock);
1041 if (unlikely(alien->avail == alien->limit)) {
1042 STATS_INC_ACOVERFLOW(cachep);
1043 __drain_alien_cache(cachep, alien, nodeid);
1044 }
1045 alien->entry[alien->avail++] = objp;
1046 spin_unlock(&alien->lock);
1047 } else {
1048 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1049 free_block(cachep, &objp, 1, nodeid);
1050 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1051 }
1052 return 1;
1053}
1054
Christoph Lametere498be72005-09-09 13:03:32 -07001055#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001056
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001057#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001058#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001059
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001060static inline struct array_cache **alloc_alien_cache(int node, int limit)
1061{
1062 return (struct array_cache **) 0x01020304ul;
1063}
1064
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001065static inline void free_alien_cache(struct array_cache **ac_ptr)
1066{
1067}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001068
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001069static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1070{
1071 return 0;
1072}
1073
Christoph Lametere498be72005-09-09 13:03:32 -07001074#endif
1075
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07001076static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001077 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078{
1079 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001080 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001081 struct kmem_list3 *l3 = NULL;
1082 int node = cpu_to_node(cpu);
1083 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
1085 switch (action) {
1086 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001087 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001088 /*
1089 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001090 * alloc_arraycache's are going to use this list.
1091 * kmalloc_node allows us to add the slab to the right
1092 * kmem_list3 and not this cpu's kmem_list3
1093 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094
Christoph Lametere498be72005-09-09 13:03:32 -07001095 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001096 /*
1097 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001098 * begin anything. Make sure some other cpu on this
1099 * node has not already allocated this
1100 */
1101 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001102 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1103 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001104 goto bad;
1105 kmem_list3_init(l3);
1106 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001107 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001108
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001109 /*
1110 * The l3s don't come and go as CPUs come and
1111 * go. cache_chain_mutex is sufficient
1112 * protection here.
1113 */
Christoph Lametere498be72005-09-09 13:03:32 -07001114 cachep->nodelists[node] = l3;
1115 }
1116
1117 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1118 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001119 (1 + nr_cpus_node(node)) *
1120 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001121 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1122 }
1123
Andrew Mortona737b3e2006-03-22 00:08:11 -08001124 /*
1125 * Now we can go ahead with allocating the shared arrays and
1126 * array caches
1127 */
Christoph Lametere498be72005-09-09 13:03:32 -07001128 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001129 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001130 struct array_cache *shared;
1131 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001132
Christoph Lametere498be72005-09-09 13:03:32 -07001133 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001134 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 if (!nc)
1136 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001137 shared = alloc_arraycache(node,
1138 cachep->shared * cachep->batchcount,
1139 0xbaadf00d);
1140 if (!shared)
1141 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001142
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001143 alien = alloc_alien_cache(node, cachep->limit);
1144 if (!alien)
1145 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001147 l3 = cachep->nodelists[node];
1148 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001149
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001150 spin_lock_irq(&l3->list_lock);
1151 if (!l3->shared) {
1152 /*
1153 * We are serialised from CPU_DEAD or
1154 * CPU_UP_CANCELLED by the cpucontrol lock
1155 */
1156 l3->shared = shared;
1157 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001158 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001159#ifdef CONFIG_NUMA
1160 if (!l3->alien) {
1161 l3->alien = alien;
1162 alien = NULL;
1163 }
1164#endif
1165 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001166 kfree(shared);
1167 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001169 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 break;
1171 case CPU_ONLINE:
1172 start_cpu_timer(cpu);
1173 break;
1174#ifdef CONFIG_HOTPLUG_CPU
1175 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001176 /*
1177 * Even if all the cpus of a node are down, we don't free the
1178 * kmem_list3 of any cache. This to avoid a race between
1179 * cpu_down, and a kmalloc allocation from another cpu for
1180 * memory from the node of the cpu going down. The list3
1181 * structure is usually allocated from kmem_cache_create() and
1182 * gets destroyed at kmem_cache_destroy().
1183 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 /* fall thru */
1185 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001186 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 list_for_each_entry(cachep, &cache_chain, next) {
1188 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001189 struct array_cache *shared;
1190 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001191 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192
Christoph Lametere498be72005-09-09 13:03:32 -07001193 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 /* cpu is dead; no one can alloc from it. */
1195 nc = cachep->array[cpu];
1196 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001197 l3 = cachep->nodelists[node];
1198
1199 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001200 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001201
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001202 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001203
1204 /* Free limit for this kmem_list3 */
1205 l3->free_limit -= cachep->batchcount;
1206 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001207 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001208
1209 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001210 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001211 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001212 }
Christoph Lametere498be72005-09-09 13:03:32 -07001213
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001214 shared = l3->shared;
1215 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001216 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001217 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001218 l3->shared = NULL;
1219 }
Christoph Lametere498be72005-09-09 13:03:32 -07001220
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001221 alien = l3->alien;
1222 l3->alien = NULL;
1223
1224 spin_unlock_irq(&l3->list_lock);
1225
1226 kfree(shared);
1227 if (alien) {
1228 drain_alien_cache(cachep, alien);
1229 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001230 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001231free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 kfree(nc);
1233 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001234 /*
1235 * In the previous loop, all the objects were freed to
1236 * the respective cache's slabs, now we can go ahead and
1237 * shrink each nodelist to its limit.
1238 */
1239 list_for_each_entry(cachep, &cache_chain, next) {
1240 l3 = cachep->nodelists[node];
1241 if (!l3)
1242 continue;
1243 spin_lock_irq(&l3->list_lock);
1244 /* free slabs belonging to this node */
1245 __node_shrink(cachep, node);
1246 spin_unlock_irq(&l3->list_lock);
1247 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001248 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 break;
1250#endif
1251 }
1252 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001253bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001254 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 return NOTIFY_BAD;
1256}
1257
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001258static struct notifier_block __cpuinitdata cpucache_notifier = {
1259 &cpuup_callback, NULL, 0
1260};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261
Christoph Lametere498be72005-09-09 13:03:32 -07001262/*
1263 * swap the static kmem_list3 with kmalloced memory
1264 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001265static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1266 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001267{
1268 struct kmem_list3 *ptr;
1269
1270 BUG_ON(cachep->nodelists[nodeid] != list);
1271 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1272 BUG_ON(!ptr);
1273
1274 local_irq_disable();
1275 memcpy(ptr, list, sizeof(struct kmem_list3));
1276 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1277 cachep->nodelists[nodeid] = ptr;
1278 local_irq_enable();
1279}
1280
Andrew Mortona737b3e2006-03-22 00:08:11 -08001281/*
1282 * Initialisation. Called after the page allocator have been initialised and
1283 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 */
1285void __init kmem_cache_init(void)
1286{
1287 size_t left_over;
1288 struct cache_sizes *sizes;
1289 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001290 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001291 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001292
1293 for (i = 0; i < NUM_INIT_LISTS; i++) {
1294 kmem_list3_init(&initkmem_list3[i]);
1295 if (i < MAX_NUMNODES)
1296 cache_cache.nodelists[i] = NULL;
1297 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
1299 /*
1300 * Fragmentation resistance on low memory - only use bigger
1301 * page orders on machines with more than 32MB of memory.
1302 */
1303 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1304 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1305
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 /* Bootstrap is tricky, because several objects are allocated
1307 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001308 * 1) initialize the cache_cache cache: it contains the struct
1309 * kmem_cache structures of all caches, except cache_cache itself:
1310 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001311 * Initially an __init data area is used for the head array and the
1312 * kmem_list3 structures, it's replaced with a kmalloc allocated
1313 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001315 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001316 * An __init data area is used for the head array.
1317 * 3) Create the remaining kmalloc caches, with minimally sized
1318 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 * 4) Replace the __init data head arrays for cache_cache and the first
1320 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001321 * 5) Replace the __init data for kmem_list3 for cache_cache and
1322 * the other cache's with kmalloc allocated memory.
1323 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 */
1325
1326 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 INIT_LIST_HEAD(&cache_chain);
1328 list_add(&cache_cache.next, &cache_chain);
1329 cache_cache.colour_off = cache_line_size();
1330 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001331 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
Andrew Mortona737b3e2006-03-22 00:08:11 -08001333 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1334 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335
Jack Steiner07ed76b2006-03-07 21:55:46 -08001336 for (order = 0; order < MAX_ORDER; order++) {
1337 cache_estimate(order, cache_cache.buffer_size,
1338 cache_line_size(), 0, &left_over, &cache_cache.num);
1339 if (cache_cache.num)
1340 break;
1341 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001342 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001343 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001344 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001345 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1346 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 /* 2+3) create the kmalloc caches */
1349 sizes = malloc_sizes;
1350 names = cache_names;
1351
Andrew Mortona737b3e2006-03-22 00:08:11 -08001352 /*
1353 * Initialize the caches that provide memory for the array cache and the
1354 * kmem_list3 structures first. Without this, further allocations will
1355 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001356 */
1357
1358 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001359 sizes[INDEX_AC].cs_size,
1360 ARCH_KMALLOC_MINALIGN,
1361 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1362 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001363
Andrew Mortona737b3e2006-03-22 00:08:11 -08001364 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001365 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001366 kmem_cache_create(names[INDEX_L3].name,
1367 sizes[INDEX_L3].cs_size,
1368 ARCH_KMALLOC_MINALIGN,
1369 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1370 NULL, NULL);
1371 }
Christoph Lametere498be72005-09-09 13:03:32 -07001372
Ingo Molnare0a42722006-06-23 02:03:46 -07001373 slab_early_init = 0;
1374
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001376 /*
1377 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 * This should be particularly beneficial on SMP boxes, as it
1379 * eliminates "false sharing".
1380 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001381 * allow tighter packing of the smaller caches.
1382 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001383 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001384 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001385 sizes->cs_size,
1386 ARCH_KMALLOC_MINALIGN,
1387 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1388 NULL, NULL);
1389 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001392 sizes->cs_size,
1393 ARCH_KMALLOC_MINALIGN,
1394 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1395 SLAB_PANIC,
1396 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 sizes++;
1398 names++;
1399 }
1400 /* 4) Replace the bootstrap head arrays */
1401 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001402 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001405
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001407 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1408 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001409 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 cache_cache.array[smp_processor_id()] = ptr;
1411 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001412
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001414
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001416 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001417 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001418 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001419 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001420 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001421 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 local_irq_enable();
1423 }
Christoph Lametere498be72005-09-09 13:03:32 -07001424 /* 5) Replace the bootstrap kmem_list3's */
1425 {
1426 int node;
1427 /* Replace the static kmem_list3 structures for the boot cpu */
1428 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001429 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430
Christoph Lametere498be72005-09-09 13:03:32 -07001431 for_each_online_node(node) {
1432 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001433 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001434
1435 if (INDEX_AC != INDEX_L3) {
1436 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001437 &initkmem_list3[SIZE_L3 + node],
1438 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001439 }
1440 }
1441 }
1442
1443 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001445 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001446 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001448 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001449 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 }
1451
1452 /* Done! */
1453 g_cpucache_up = FULL;
1454
Andrew Mortona737b3e2006-03-22 00:08:11 -08001455 /*
1456 * Register a cpu startup notifier callback that initializes
1457 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458 */
1459 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460
Andrew Mortona737b3e2006-03-22 00:08:11 -08001461 /*
1462 * The reap timers are started later, with a module init call: That part
1463 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 */
1465}
1466
1467static int __init cpucache_init(void)
1468{
1469 int cpu;
1470
Andrew Mortona737b3e2006-03-22 00:08:11 -08001471 /*
1472 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 */
Christoph Lametere498be72005-09-09 13:03:32 -07001474 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001475 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 return 0;
1477}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478__initcall(cpucache_init);
1479
1480/*
1481 * Interface to system's page allocator. No need to hold the cache-lock.
1482 *
1483 * If we requested dmaable memory, we will get it. Even if we
1484 * did not request dmaable memory, we might get it, but that
1485 * would be relatively rare and ignorable.
1486 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001487static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488{
1489 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001490 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 int i;
1492
Luke Yangd6fef9d2006-04-10 22:52:56 -07001493#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001494 /*
1495 * Nommu uses slab's for process anonymous memory allocations, and thus
1496 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001497 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001498 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001499#endif
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001500 flags |= cachep->gfpflags;
1501
1502 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 if (!page)
1504 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001506 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001508 atomic_add(nr_pages, &slab_reclaim_pages);
1509 add_page_state(nr_slab, nr_pages);
1510 for (i = 0; i < nr_pages; i++)
1511 __SetPageSlab(page + i);
1512 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513}
1514
1515/*
1516 * Interface to system's page release.
1517 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001518static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001520 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 struct page *page = virt_to_page(addr);
1522 const unsigned long nr_freed = i;
1523
1524 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001525 BUG_ON(!PageSlab(page));
1526 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 page++;
1528 }
1529 sub_page_state(nr_slab, nr_freed);
1530 if (current->reclaim_state)
1531 current->reclaim_state->reclaimed_slab += nr_freed;
1532 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001533 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1534 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535}
1536
1537static void kmem_rcu_free(struct rcu_head *head)
1538{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001539 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001540 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541
1542 kmem_freepages(cachep, slab_rcu->addr);
1543 if (OFF_SLAB(cachep))
1544 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1545}
1546
1547#if DEBUG
1548
1549#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001550static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001551 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001553 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001555 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001557 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 return;
1559
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001560 *addr++ = 0x12345678;
1561 *addr++ = caller;
1562 *addr++ = smp_processor_id();
1563 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 {
1565 unsigned long *sptr = &caller;
1566 unsigned long svalue;
1567
1568 while (!kstack_end(sptr)) {
1569 svalue = *sptr++;
1570 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001571 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 size -= sizeof(unsigned long);
1573 if (size <= sizeof(unsigned long))
1574 break;
1575 }
1576 }
1577
1578 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001579 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580}
1581#endif
1582
Pekka Enberg343e0d72006-02-01 03:05:50 -08001583static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001585 int size = obj_size(cachep);
1586 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587
1588 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590}
1591
1592static void dump_line(char *data, int offset, int limit)
1593{
1594 int i;
1595 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001597 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 printk("\n");
1599}
1600#endif
1601
1602#if DEBUG
1603
Pekka Enberg343e0d72006-02-01 03:05:50 -08001604static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605{
1606 int i, size;
1607 char *realobj;
1608
1609 if (cachep->flags & SLAB_RED_ZONE) {
1610 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001611 *dbg_redzone1(cachep, objp),
1612 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 }
1614
1615 if (cachep->flags & SLAB_STORE_USER) {
1616 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001617 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001619 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 printk("\n");
1621 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001622 realobj = (char *)objp + obj_offset(cachep);
1623 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001624 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 int limit;
1626 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001627 if (i + limit > size)
1628 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 dump_line(realobj, i, limit);
1630 }
1631}
1632
Pekka Enberg343e0d72006-02-01 03:05:50 -08001633static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634{
1635 char *realobj;
1636 int size, i;
1637 int lines = 0;
1638
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001639 realobj = (char *)objp + obj_offset(cachep);
1640 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001642 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001644 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 exp = POISON_END;
1646 if (realobj[i] != exp) {
1647 int limit;
1648 /* Mismatch ! */
1649 /* Print header */
1650 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001651 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001652 "Slab corruption: start=%p, len=%d\n",
1653 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 print_objinfo(cachep, objp, 0);
1655 }
1656 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001657 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001659 if (i + limit > size)
1660 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 dump_line(realobj, i, limit);
1662 i += 16;
1663 lines++;
1664 /* Limit to 5 lines */
1665 if (lines > 5)
1666 break;
1667 }
1668 }
1669 if (lines != 0) {
1670 /* Print some data about the neighboring objects, if they
1671 * exist:
1672 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001673 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001674 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001676 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001678 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001679 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001681 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 print_objinfo(cachep, objp, 2);
1683 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001684 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001685 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001686 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001688 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 print_objinfo(cachep, objp, 2);
1690 }
1691 }
1692}
1693#endif
1694
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001696/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001697 * slab_destroy_objs - destroy a slab and its objects
1698 * @cachep: cache pointer being destroyed
1699 * @slabp: slab pointer being destroyed
1700 *
1701 * Call the registered destructor for each object in a slab that is being
1702 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001703 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001704static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001705{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 int i;
1707 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001708 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
1710 if (cachep->flags & SLAB_POISON) {
1711#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001712 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1713 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001714 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001715 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 else
1717 check_poison_obj(cachep, objp);
1718#else
1719 check_poison_obj(cachep, objp);
1720#endif
1721 }
1722 if (cachep->flags & SLAB_RED_ZONE) {
1723 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1724 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001725 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1727 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001728 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
1730 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001731 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001733}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001735static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001736{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 if (cachep->dtor) {
1738 int i;
1739 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001740 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001741 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 }
1743 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001744}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745#endif
1746
Randy Dunlap911851e2006-03-22 00:08:14 -08001747/**
1748 * slab_destroy - destroy and release all objects in a slab
1749 * @cachep: cache pointer being destroyed
1750 * @slabp: slab pointer being destroyed
1751 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001752 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001753 * Before calling the slab must have been unlinked from the cache. The
1754 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001755 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001756static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001757{
1758 void *addr = slabp->s_mem - slabp->colouroff;
1759
1760 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1762 struct slab_rcu *slab_rcu;
1763
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001764 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 slab_rcu->cachep = cachep;
1766 slab_rcu->addr = addr;
1767 call_rcu(&slab_rcu->head, kmem_rcu_free);
1768 } else {
1769 kmem_freepages(cachep, addr);
1770 if (OFF_SLAB(cachep))
1771 kmem_cache_free(cachep->slabp_cache, slabp);
1772 }
1773}
1774
Andrew Mortona737b3e2006-03-22 00:08:11 -08001775/*
1776 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1777 * size of kmem_list3.
1778 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001779static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001780{
1781 int node;
1782
1783 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001784 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001785 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001786 REAPTIMEOUT_LIST3 +
1787 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001788 }
1789}
1790
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001792 * calculate_slab_order - calculate size (page order) of slabs
1793 * @cachep: pointer to the cache that is being created
1794 * @size: size of objects to be created in this cache.
1795 * @align: required alignment for the objects.
1796 * @flags: slab allocation flags
1797 *
1798 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001799 *
1800 * This could be made much more intelligent. For now, try to avoid using
1801 * high order pages for slabs. When the gfp() functions are more friendly
1802 * towards high-order requests, this should be changed.
1803 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001804static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001805 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001806{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001807 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001808 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001809 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001810
Andrew Mortona737b3e2006-03-22 00:08:11 -08001811 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001812 unsigned int num;
1813 size_t remainder;
1814
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001815 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001816 if (!num)
1817 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001818
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001819 if (flags & CFLGS_OFF_SLAB) {
1820 /*
1821 * Max number of objs-per-slab for caches which
1822 * use off-slab slabs. Needed to avoid a possible
1823 * looping condition in cache_grow().
1824 */
1825 offslab_limit = size - sizeof(struct slab);
1826 offslab_limit /= sizeof(kmem_bufctl_t);
1827
1828 if (num > offslab_limit)
1829 break;
1830 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001831
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001832 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001833 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001834 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001835 left_over = remainder;
1836
1837 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001838 * A VFS-reclaimable slab tends to have most allocations
1839 * as GFP_NOFS and we really don't want to have to be allocating
1840 * higher-order pages when we are unable to shrink dcache.
1841 */
1842 if (flags & SLAB_RECLAIM_ACCOUNT)
1843 break;
1844
1845 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001846 * Large number of objects is good, but very large slabs are
1847 * currently bad for the gfp()s.
1848 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001849 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001850 break;
1851
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001852 /*
1853 * Acceptable internal fragmentation?
1854 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001855 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001856 break;
1857 }
1858 return left_over;
1859}
1860
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001861static void setup_cpu_cache(struct kmem_cache *cachep)
1862{
1863 if (g_cpucache_up == FULL) {
1864 enable_cpucache(cachep);
1865 return;
1866 }
1867 if (g_cpucache_up == NONE) {
1868 /*
1869 * Note: the first kmem_cache_create must create the cache
1870 * that's used by kmalloc(24), otherwise the creation of
1871 * further caches will BUG().
1872 */
1873 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1874
1875 /*
1876 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1877 * the first cache, then we need to set up all its list3s,
1878 * otherwise the creation of further caches will BUG().
1879 */
1880 set_up_list3s(cachep, SIZE_AC);
1881 if (INDEX_AC == INDEX_L3)
1882 g_cpucache_up = PARTIAL_L3;
1883 else
1884 g_cpucache_up = PARTIAL_AC;
1885 } else {
1886 cachep->array[smp_processor_id()] =
1887 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1888
1889 if (g_cpucache_up == PARTIAL_AC) {
1890 set_up_list3s(cachep, SIZE_L3);
1891 g_cpucache_up = PARTIAL_L3;
1892 } else {
1893 int node;
1894 for_each_online_node(node) {
1895 cachep->nodelists[node] =
1896 kmalloc_node(sizeof(struct kmem_list3),
1897 GFP_KERNEL, node);
1898 BUG_ON(!cachep->nodelists[node]);
1899 kmem_list3_init(cachep->nodelists[node]);
1900 }
1901 }
1902 }
1903 cachep->nodelists[numa_node_id()]->next_reap =
1904 jiffies + REAPTIMEOUT_LIST3 +
1905 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1906
1907 cpu_cache_get(cachep)->avail = 0;
1908 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1909 cpu_cache_get(cachep)->batchcount = 1;
1910 cpu_cache_get(cachep)->touched = 0;
1911 cachep->batchcount = 1;
1912 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1913}
1914
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001915/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 * kmem_cache_create - Create a cache.
1917 * @name: A string which is used in /proc/slabinfo to identify this cache.
1918 * @size: The size of objects to be created in this cache.
1919 * @align: The required alignment for the objects.
1920 * @flags: SLAB flags
1921 * @ctor: A constructor for the objects.
1922 * @dtor: A destructor for the objects.
1923 *
1924 * Returns a ptr to the cache on success, NULL on failure.
1925 * Cannot be called within a int, but can be interrupted.
1926 * The @ctor is run when new pages are allocated by the cache
1927 * and the @dtor is run before the pages are handed back.
1928 *
1929 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001930 * the module calling this has to destroy the cache before getting unloaded.
1931 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 * The flags are
1933 *
1934 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1935 * to catch references to uninitialised memory.
1936 *
1937 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1938 * for buffer overruns.
1939 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1941 * cacheline. This can be beneficial if you're counting cycles as closely
1942 * as davem.
1943 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001944struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001946 unsigned long flags,
1947 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001948 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949{
1950 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07001951 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952
1953 /*
1954 * Sanity checks... these are all serious usage bugs.
1955 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001956 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001957 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001958 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1959 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001960 BUG();
1961 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001963 /*
1964 * Prevent CPUs from coming and going.
1965 * lock_cpu_hotplug() nests outside cache_chain_mutex
1966 */
1967 lock_cpu_hotplug();
1968
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001969 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001970
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07001971 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001972 mm_segment_t old_fs = get_fs();
1973 char tmp;
1974 int res;
1975
1976 /*
1977 * This happens when the module gets unloaded and doesn't
1978 * destroy its slab cache and no-one else reuses the vmalloc
1979 * area of the module. Print a warning.
1980 */
1981 set_fs(KERNEL_DS);
1982 res = __get_user(tmp, pc->name);
1983 set_fs(old_fs);
1984 if (res) {
1985 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001986 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001987 continue;
1988 }
1989
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001990 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001991 printk("kmem_cache_create: duplicate cache %s\n", name);
1992 dump_stack();
1993 goto oops;
1994 }
1995 }
1996
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997#if DEBUG
1998 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1999 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2000 /* No constructor, but inital state check requested */
2001 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002002 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 flags &= ~SLAB_DEBUG_INITIAL;
2004 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005#if FORCED_DEBUG
2006 /*
2007 * Enable redzoning and last user accounting, except for caches with
2008 * large objects, if the increased size would increase the object size
2009 * above the next power of two: caches with object sizes just above a
2010 * power of two have a significant amount of internal fragmentation.
2011 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002012 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002013 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 if (!(flags & SLAB_DESTROY_BY_RCU))
2015 flags |= SLAB_POISON;
2016#endif
2017 if (flags & SLAB_DESTROY_BY_RCU)
2018 BUG_ON(flags & SLAB_POISON);
2019#endif
2020 if (flags & SLAB_DESTROY_BY_RCU)
2021 BUG_ON(dtor);
2022
2023 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002024 * Always checks flags, a caller might be expecting debug support which
2025 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002027 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
Andrew Mortona737b3e2006-03-22 00:08:11 -08002029 /*
2030 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 * unaligned accesses for some archs when redzoning is used, and makes
2032 * sure any on-slab bufctl's are also correctly aligned.
2033 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002034 if (size & (BYTES_PER_WORD - 1)) {
2035 size += (BYTES_PER_WORD - 1);
2036 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 }
2038
Andrew Mortona737b3e2006-03-22 00:08:11 -08002039 /* calculate the final buffer alignment: */
2040
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 /* 1) arch recommendation: can be overridden for debug */
2042 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002043 /*
2044 * Default alignment: as specified by the arch code. Except if
2045 * an object is really small, then squeeze multiple objects into
2046 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 */
2048 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002049 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 ralign /= 2;
2051 } else {
2052 ralign = BYTES_PER_WORD;
2053 }
2054 /* 2) arch mandated alignment: disables debug if necessary */
2055 if (ralign < ARCH_SLAB_MINALIGN) {
2056 ralign = ARCH_SLAB_MINALIGN;
2057 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002058 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 }
2060 /* 3) caller mandated alignment: disables debug if necessary */
2061 if (ralign < align) {
2062 ralign = align;
2063 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002064 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002066 /*
2067 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 * the alignment to BYTES_PER_WORD.
2069 */
2070 align = ralign;
2071
2072 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002073 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002075 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076
2077#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002078 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079
2080 if (flags & SLAB_RED_ZONE) {
2081 /* redzoning only works with word aligned caches */
2082 align = BYTES_PER_WORD;
2083
2084 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002085 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002086 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 }
2088 if (flags & SLAB_STORE_USER) {
2089 /* user store requires word alignment and
2090 * one word storage behind the end of the real
2091 * object.
2092 */
2093 align = BYTES_PER_WORD;
2094 size += BYTES_PER_WORD;
2095 }
2096#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002097 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002098 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2099 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 size = PAGE_SIZE;
2101 }
2102#endif
2103#endif
2104
Ingo Molnare0a42722006-06-23 02:03:46 -07002105 /*
2106 * Determine if the slab management is 'on' or 'off' slab.
2107 * (bootstrapping cannot cope with offslab caches so don't do
2108 * it too early on.)
2109 */
2110 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 /*
2112 * Size is large, assume best to place the slab management obj
2113 * off-slab (should allow better packing of objs).
2114 */
2115 flags |= CFLGS_OFF_SLAB;
2116
2117 size = ALIGN(size, align);
2118
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002119 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
2121 if (!cachep->num) {
2122 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2123 kmem_cache_free(&cache_cache, cachep);
2124 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002125 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002127 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2128 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
2130 /*
2131 * If the slab has been placed off-slab, and we have enough space then
2132 * move it on-slab. This is at the expense of any extra colouring.
2133 */
2134 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2135 flags &= ~CFLGS_OFF_SLAB;
2136 left_over -= slab_size;
2137 }
2138
2139 if (flags & CFLGS_OFF_SLAB) {
2140 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002141 slab_size =
2142 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 }
2144
2145 cachep->colour_off = cache_line_size();
2146 /* Offset must be a multiple of the alignment. */
2147 if (cachep->colour_off < align)
2148 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002149 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 cachep->slab_size = slab_size;
2151 cachep->flags = flags;
2152 cachep->gfpflags = 0;
2153 if (flags & SLAB_CACHE_DMA)
2154 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002155 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
2157 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002158 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 cachep->ctor = ctor;
2160 cachep->dtor = dtor;
2161 cachep->name = name;
2162
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002164 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 /* cache setup completed, link it into the list */
2167 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002168oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 if (!cachep && (flags & SLAB_PANIC))
2170 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002171 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002172 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002173 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 return cachep;
2175}
2176EXPORT_SYMBOL(kmem_cache_create);
2177
2178#if DEBUG
2179static void check_irq_off(void)
2180{
2181 BUG_ON(!irqs_disabled());
2182}
2183
2184static void check_irq_on(void)
2185{
2186 BUG_ON(irqs_disabled());
2187}
2188
Pekka Enberg343e0d72006-02-01 03:05:50 -08002189static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190{
2191#ifdef CONFIG_SMP
2192 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002193 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194#endif
2195}
Christoph Lametere498be72005-09-09 13:03:32 -07002196
Pekka Enberg343e0d72006-02-01 03:05:50 -08002197static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002198{
2199#ifdef CONFIG_SMP
2200 check_irq_off();
2201 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2202#endif
2203}
2204
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205#else
2206#define check_irq_off() do { } while(0)
2207#define check_irq_on() do { } while(0)
2208#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002209#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210#endif
2211
Christoph Lameteraab22072006-03-22 00:09:06 -08002212static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2213 struct array_cache *ac,
2214 int force, int node);
2215
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216static void do_drain(void *arg)
2217{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002218 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002220 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
2222 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002223 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002224 spin_lock(&cachep->nodelists[node]->list_lock);
2225 free_block(cachep, ac->entry, ac->avail, node);
2226 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 ac->avail = 0;
2228}
2229
Pekka Enberg343e0d72006-02-01 03:05:50 -08002230static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231{
Christoph Lametere498be72005-09-09 13:03:32 -07002232 struct kmem_list3 *l3;
2233 int node;
2234
Andrew Mortona07fa392006-03-22 00:08:17 -08002235 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002237 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002238 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002239 if (l3 && l3->alien)
2240 drain_alien_cache(cachep, l3->alien);
2241 }
2242
2243 for_each_online_node(node) {
2244 l3 = cachep->nodelists[node];
2245 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002246 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248}
2249
Pekka Enberg343e0d72006-02-01 03:05:50 -08002250static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251{
2252 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002253 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 int ret;
2255
Christoph Lametere498be72005-09-09 13:03:32 -07002256 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 struct list_head *p;
2258
Christoph Lametere498be72005-09-09 13:03:32 -07002259 p = l3->slabs_free.prev;
2260 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 break;
2262
Christoph Lametere498be72005-09-09 13:03:32 -07002263 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002265 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266#endif
2267 list_del(&slabp->list);
2268
Christoph Lametere498be72005-09-09 13:03:32 -07002269 l3->free_objects -= cachep->num;
2270 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002272 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002274 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 return ret;
2276}
2277
Pekka Enberg343e0d72006-02-01 03:05:50 -08002278static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002279{
2280 int ret = 0, i = 0;
2281 struct kmem_list3 *l3;
2282
2283 drain_cpu_caches(cachep);
2284
2285 check_irq_on();
2286 for_each_online_node(i) {
2287 l3 = cachep->nodelists[i];
2288 if (l3) {
2289 spin_lock_irq(&l3->list_lock);
2290 ret += __node_shrink(cachep, i);
2291 spin_unlock_irq(&l3->list_lock);
2292 }
2293 }
2294 return (ret ? 1 : 0);
2295}
2296
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297/**
2298 * kmem_cache_shrink - Shrink a cache.
2299 * @cachep: The cache to shrink.
2300 *
2301 * Releases as many slabs as possible for a cache.
2302 * To help debugging, a zero exit status indicates all slabs were released.
2303 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002304int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002306 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
2308 return __cache_shrink(cachep);
2309}
2310EXPORT_SYMBOL(kmem_cache_shrink);
2311
2312/**
2313 * kmem_cache_destroy - delete a cache
2314 * @cachep: the cache to destroy
2315 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002316 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 * Returns 0 on success.
2318 *
2319 * It is expected this function will be called by a module when it is
2320 * unloaded. This will remove the cache completely, and avoid a duplicate
2321 * cache being allocated each time a module is loaded and unloaded, if the
2322 * module doesn't have persistent in-kernel storage across loads and unloads.
2323 *
2324 * The cache must be empty before calling this function.
2325 *
2326 * The caller must guarantee that noone will allocate memory from the cache
2327 * during the kmem_cache_destroy().
2328 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002329int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330{
2331 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002332 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002334 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335
2336 /* Don't let CPUs to come and go */
2337 lock_cpu_hotplug();
2338
2339 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002340 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 /*
2342 * the chain is never empty, cache_cache is never destroyed
2343 */
2344 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002345 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
2347 if (__cache_shrink(cachep)) {
2348 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002349 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002350 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002351 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 unlock_cpu_hotplug();
2353 return 1;
2354 }
2355
2356 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002357 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
Christoph Lametere498be72005-09-09 13:03:32 -07002359 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002360 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
2362 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002363 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002364 l3 = cachep->nodelists[i];
2365 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002366 kfree(l3->shared);
2367 free_alien_cache(l3->alien);
2368 kfree(l3);
2369 }
2370 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 return 0;
2374}
2375EXPORT_SYMBOL(kmem_cache_destroy);
2376
2377/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002378static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002379 int colour_off, gfp_t local_flags,
2380 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381{
2382 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002383
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 if (OFF_SLAB(cachep)) {
2385 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002386 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2387 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 if (!slabp)
2389 return NULL;
2390 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002391 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 colour_off += cachep->slab_size;
2393 }
2394 slabp->inuse = 0;
2395 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002396 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002397 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 return slabp;
2399}
2400
2401static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2402{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002403 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404}
2405
Pekka Enberg343e0d72006-02-01 03:05:50 -08002406static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002407 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408{
2409 int i;
2410
2411 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002412 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413#if DEBUG
2414 /* need to poison the objs? */
2415 if (cachep->flags & SLAB_POISON)
2416 poison_obj(cachep, objp, POISON_FREE);
2417 if (cachep->flags & SLAB_STORE_USER)
2418 *dbg_userword(cachep, objp) = NULL;
2419
2420 if (cachep->flags & SLAB_RED_ZONE) {
2421 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2422 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2423 }
2424 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002425 * Constructors are not allowed to allocate memory from the same
2426 * cache which they are a constructor for. Otherwise, deadlock.
2427 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 */
2429 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002430 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002431 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432
2433 if (cachep->flags & SLAB_RED_ZONE) {
2434 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2435 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002436 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2438 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002439 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002441 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2442 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002443 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002444 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445#else
2446 if (cachep->ctor)
2447 cachep->ctor(objp, cachep, ctor_flags);
2448#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002449 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002451 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 slabp->free = 0;
2453}
2454
Pekka Enberg343e0d72006-02-01 03:05:50 -08002455static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002457 if (flags & SLAB_DMA)
2458 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2459 else
2460 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461}
2462
Andrew Mortona737b3e2006-03-22 00:08:11 -08002463static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2464 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002465{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002466 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002467 kmem_bufctl_t next;
2468
2469 slabp->inuse++;
2470 next = slab_bufctl(slabp)[slabp->free];
2471#if DEBUG
2472 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2473 WARN_ON(slabp->nodeid != nodeid);
2474#endif
2475 slabp->free = next;
2476
2477 return objp;
2478}
2479
Andrew Mortona737b3e2006-03-22 00:08:11 -08002480static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2481 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002482{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002483 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002484
2485#if DEBUG
2486 /* Verify that the slab belongs to the intended node */
2487 WARN_ON(slabp->nodeid != nodeid);
2488
Al Viro871751e2006-03-25 03:06:39 -08002489 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002490 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002491 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002492 BUG();
2493 }
2494#endif
2495 slab_bufctl(slabp)[objnr] = slabp->free;
2496 slabp->free = objnr;
2497 slabp->inuse--;
2498}
2499
Pekka Enberg47768742006-06-23 02:03:07 -07002500/*
2501 * Map pages beginning at addr to the given cache and slab. This is required
2502 * for the slab allocator to be able to lookup the cache and slab of a
2503 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2504 */
2505static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2506 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507{
Pekka Enberg47768742006-06-23 02:03:07 -07002508 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 struct page *page;
2510
Pekka Enberg47768742006-06-23 02:03:07 -07002511 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002512
Pekka Enberg47768742006-06-23 02:03:07 -07002513 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002514 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002515 nr_pages <<= cache->gfporder;
2516
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002518 page_set_cache(page, cache);
2519 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002521 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522}
2523
2524/*
2525 * Grow (by 1) the number of slabs within a cache. This is called by
2526 * kmem_cache_alloc() when there are no active objs left in a cache.
2527 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002528static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002530 struct slab *slabp;
2531 void *objp;
2532 size_t offset;
2533 gfp_t local_flags;
2534 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002535 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536
Andrew Mortona737b3e2006-03-22 00:08:11 -08002537 /*
2538 * Be lazy and only check for valid flags here, keeping it out of the
2539 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002541 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 if (flags & SLAB_NO_GROW)
2543 return 0;
2544
2545 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2546 local_flags = (flags & SLAB_LEVEL_MASK);
2547 if (!(local_flags & __GFP_WAIT))
2548 /*
2549 * Not allowed to sleep. Need to tell a constructor about
2550 * this - it might need to know...
2551 */
2552 ctor_flags |= SLAB_CTOR_ATOMIC;
2553
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002554 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002556 l3 = cachep->nodelists[nodeid];
2557 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558
2559 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002560 offset = l3->colour_next;
2561 l3->colour_next++;
2562 if (l3->colour_next >= cachep->colour)
2563 l3->colour_next = 0;
2564 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002566 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567
2568 if (local_flags & __GFP_WAIT)
2569 local_irq_enable();
2570
2571 /*
2572 * The test for missing atomic flag is performed here, rather than
2573 * the more obvious place, simply to reduce the critical path length
2574 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2575 * will eventually be caught here (where it matters).
2576 */
2577 kmem_flagcheck(cachep, flags);
2578
Andrew Mortona737b3e2006-03-22 00:08:11 -08002579 /*
2580 * Get mem for the objs. Attempt to allocate a physical page from
2581 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002582 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002583 objp = kmem_getpages(cachep, flags, nodeid);
2584 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585 goto failed;
2586
2587 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002588 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002589 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 goto opps1;
2591
Christoph Lametere498be72005-09-09 13:03:32 -07002592 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002593 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594
2595 cache_init_objs(cachep, slabp, ctor_flags);
2596
2597 if (local_flags & __GFP_WAIT)
2598 local_irq_disable();
2599 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002600 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601
2602 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002603 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002605 l3->free_objects += cachep->num;
2606 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002608opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002610failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 if (local_flags & __GFP_WAIT)
2612 local_irq_disable();
2613 return 0;
2614}
2615
2616#if DEBUG
2617
2618/*
2619 * Perform extra freeing checks:
2620 * - detect bad pointers.
2621 * - POISON/RED_ZONE checking
2622 * - destructor calls, for caches with POISON+dtor
2623 */
2624static void kfree_debugcheck(const void *objp)
2625{
2626 struct page *page;
2627
2628 if (!virt_addr_valid(objp)) {
2629 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002630 (unsigned long)objp);
2631 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 }
2633 page = virt_to_page(objp);
2634 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002635 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2636 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 BUG();
2638 }
2639}
2640
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002641static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2642{
2643 unsigned long redzone1, redzone2;
2644
2645 redzone1 = *dbg_redzone1(cache, obj);
2646 redzone2 = *dbg_redzone2(cache, obj);
2647
2648 /*
2649 * Redzone is ok.
2650 */
2651 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2652 return;
2653
2654 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2655 slab_error(cache, "double free detected");
2656 else
2657 slab_error(cache, "memory outside object was overwritten");
2658
2659 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2660 obj, redzone1, redzone2);
2661}
2662
Pekka Enberg343e0d72006-02-01 03:05:50 -08002663static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002664 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665{
2666 struct page *page;
2667 unsigned int objnr;
2668 struct slab *slabp;
2669
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002670 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 kfree_debugcheck(objp);
2672 page = virt_to_page(objp);
2673
Pekka Enberg065d41c2005-11-13 16:06:46 -08002674 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675
2676 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002677 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2679 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2680 }
2681 if (cachep->flags & SLAB_STORE_USER)
2682 *dbg_userword(cachep, objp) = caller;
2683
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002684 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685
2686 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002687 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688
2689 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002690 /*
2691 * Need to call the slab's constructor so the caller can
2692 * perform a verify of its state (debugging). Called without
2693 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002695 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002696 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 }
2698 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2699 /* we want to cache poison the object,
2700 * call the destruction callback
2701 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002702 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 }
Al Viro871751e2006-03-25 03:06:39 -08002704#ifdef CONFIG_DEBUG_SLAB_LEAK
2705 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2706#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 if (cachep->flags & SLAB_POISON) {
2708#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002709 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002711 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002712 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 } else {
2714 poison_obj(cachep, objp, POISON_FREE);
2715 }
2716#else
2717 poison_obj(cachep, objp, POISON_FREE);
2718#endif
2719 }
2720 return objp;
2721}
2722
Pekka Enberg343e0d72006-02-01 03:05:50 -08002723static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724{
2725 kmem_bufctl_t i;
2726 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002727
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 /* Check slab's freelist to see if this obj is there. */
2729 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2730 entries++;
2731 if (entries > cachep->num || i >= cachep->num)
2732 goto bad;
2733 }
2734 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002735bad:
2736 printk(KERN_ERR "slab: Internal list corruption detected in "
2737 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2738 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002739 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002740 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002741 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002742 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002744 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 }
2746 printk("\n");
2747 BUG();
2748 }
2749}
2750#else
2751#define kfree_debugcheck(x) do { } while(0)
2752#define cache_free_debugcheck(x,objp,z) (objp)
2753#define check_slabp(x,y) do { } while(0)
2754#endif
2755
Pekka Enberg343e0d72006-02-01 03:05:50 -08002756static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757{
2758 int batchcount;
2759 struct kmem_list3 *l3;
2760 struct array_cache *ac;
2761
2762 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002763 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002764retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 batchcount = ac->batchcount;
2766 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002767 /*
2768 * If there was little recent activity on this cache, then
2769 * perform only a partial refill. Otherwise we could generate
2770 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 */
2772 batchcount = BATCHREFILL_LIMIT;
2773 }
Christoph Lametere498be72005-09-09 13:03:32 -07002774 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775
Christoph Lametere498be72005-09-09 13:03:32 -07002776 BUG_ON(ac->avail > 0 || !l3);
2777 spin_lock(&l3->list_lock);
2778
Christoph Lameter3ded1752006-03-25 03:06:44 -08002779 /* See if we can refill from the shared array */
2780 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2781 goto alloc_done;
2782
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 while (batchcount > 0) {
2784 struct list_head *entry;
2785 struct slab *slabp;
2786 /* Get slab alloc is to come from. */
2787 entry = l3->slabs_partial.next;
2788 if (entry == &l3->slabs_partial) {
2789 l3->free_touched = 1;
2790 entry = l3->slabs_free.next;
2791 if (entry == &l3->slabs_free)
2792 goto must_grow;
2793 }
2794
2795 slabp = list_entry(entry, struct slab, list);
2796 check_slabp(cachep, slabp);
2797 check_spinlock_acquired(cachep);
2798 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 STATS_INC_ALLOCED(cachep);
2800 STATS_INC_ACTIVE(cachep);
2801 STATS_SET_HIGH(cachep);
2802
Matthew Dobson78d382d2006-02-01 03:05:47 -08002803 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2804 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 }
2806 check_slabp(cachep, slabp);
2807
2808 /* move slabp to correct slabp list: */
2809 list_del(&slabp->list);
2810 if (slabp->free == BUFCTL_END)
2811 list_add(&slabp->list, &l3->slabs_full);
2812 else
2813 list_add(&slabp->list, &l3->slabs_partial);
2814 }
2815
Andrew Mortona737b3e2006-03-22 00:08:11 -08002816must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002818alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002819 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
2821 if (unlikely(!ac->avail)) {
2822 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002823 x = cache_grow(cachep, flags, numa_node_id());
2824
Andrew Mortona737b3e2006-03-22 00:08:11 -08002825 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002826 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002827 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 return NULL;
2829
Andrew Mortona737b3e2006-03-22 00:08:11 -08002830 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 goto retry;
2832 }
2833 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002834 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835}
2836
Andrew Mortona737b3e2006-03-22 00:08:11 -08002837static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2838 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839{
2840 might_sleep_if(flags & __GFP_WAIT);
2841#if DEBUG
2842 kmem_flagcheck(cachep, flags);
2843#endif
2844}
2845
2846#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002847static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2848 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002850 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002852 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002854 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002855 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002856 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 else
2858 check_poison_obj(cachep, objp);
2859#else
2860 check_poison_obj(cachep, objp);
2861#endif
2862 poison_obj(cachep, objp, POISON_INUSE);
2863 }
2864 if (cachep->flags & SLAB_STORE_USER)
2865 *dbg_userword(cachep, objp) = caller;
2866
2867 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002868 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2869 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2870 slab_error(cachep, "double free, or memory outside"
2871 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002872 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002873 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2874 objp, *dbg_redzone1(cachep, objp),
2875 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 }
2877 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2878 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2879 }
Al Viro871751e2006-03-25 03:06:39 -08002880#ifdef CONFIG_DEBUG_SLAB_LEAK
2881 {
2882 struct slab *slabp;
2883 unsigned objnr;
2884
2885 slabp = page_get_slab(virt_to_page(objp));
2886 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2887 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2888 }
2889#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002890 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002892 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893
2894 if (!(flags & __GFP_WAIT))
2895 ctor_flags |= SLAB_CTOR_ATOMIC;
2896
2897 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002898 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 return objp;
2900}
2901#else
2902#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2903#endif
2904
Pekka Enberg343e0d72006-02-01 03:05:50 -08002905static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002907 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 struct array_cache *ac;
2909
Christoph Lameterdc85da12006-01-18 17:42:36 -08002910#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002911 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002912 objp = alternate_node_alloc(cachep, flags);
2913 if (objp != NULL)
2914 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08002915 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08002916#endif
2917
Alok N Kataria5c382302005-09-27 21:45:46 -07002918 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002919 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 if (likely(ac->avail)) {
2921 STATS_INC_ALLOCHIT(cachep);
2922 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002923 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 } else {
2925 STATS_INC_ALLOCMISS(cachep);
2926 objp = cache_alloc_refill(cachep, flags);
2927 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002928 return objp;
2929}
2930
Andrew Mortona737b3e2006-03-22 00:08:11 -08002931static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2932 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002933{
2934 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002935 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002936
2937 cache_alloc_debugcheck_before(cachep, flags);
2938
2939 local_irq_save(save_flags);
2940 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002942 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002943 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002944 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 return objp;
2946}
2947
Christoph Lametere498be72005-09-09 13:03:32 -07002948#ifdef CONFIG_NUMA
2949/*
Paul Jacksonb2455392006-03-24 03:16:12 -08002950 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002951 *
2952 * If we are in_interrupt, then process context, including cpusets and
2953 * mempolicy, may not apply and should not be used for allocation policy.
2954 */
2955static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2956{
2957 int nid_alloc, nid_here;
2958
2959 if (in_interrupt())
2960 return NULL;
2961 nid_alloc = nid_here = numa_node_id();
2962 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2963 nid_alloc = cpuset_mem_spread_node();
2964 else if (current->mempolicy)
2965 nid_alloc = slab_node(current->mempolicy);
2966 if (nid_alloc != nid_here)
2967 return __cache_alloc_node(cachep, flags, nid_alloc);
2968 return NULL;
2969}
2970
2971/*
Christoph Lametere498be72005-09-09 13:03:32 -07002972 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002974static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2975 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002976{
2977 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002978 struct slab *slabp;
2979 struct kmem_list3 *l3;
2980 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002981 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002983 l3 = cachep->nodelists[nodeid];
2984 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002985
Andrew Mortona737b3e2006-03-22 00:08:11 -08002986retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002987 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002988 spin_lock(&l3->list_lock);
2989 entry = l3->slabs_partial.next;
2990 if (entry == &l3->slabs_partial) {
2991 l3->free_touched = 1;
2992 entry = l3->slabs_free.next;
2993 if (entry == &l3->slabs_free)
2994 goto must_grow;
2995 }
Christoph Lametere498be72005-09-09 13:03:32 -07002996
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002997 slabp = list_entry(entry, struct slab, list);
2998 check_spinlock_acquired_node(cachep, nodeid);
2999 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003000
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003001 STATS_INC_NODEALLOCS(cachep);
3002 STATS_INC_ACTIVE(cachep);
3003 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003004
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003005 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003006
Matthew Dobson78d382d2006-02-01 03:05:47 -08003007 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003008 check_slabp(cachep, slabp);
3009 l3->free_objects--;
3010 /* move slabp to correct slabp list: */
3011 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003012
Andrew Mortona737b3e2006-03-22 00:08:11 -08003013 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003014 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003015 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003016 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003017
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003018 spin_unlock(&l3->list_lock);
3019 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003020
Andrew Mortona737b3e2006-03-22 00:08:11 -08003021must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003022 spin_unlock(&l3->list_lock);
3023 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003024
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003025 if (!x)
3026 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003027
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003028 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003029done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003030 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003031}
3032#endif
3033
3034/*
3035 * Caller needs to acquire correct kmem_list's list_lock
3036 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003037static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003038 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039{
3040 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003041 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042
3043 for (i = 0; i < nr_objects; i++) {
3044 void *objp = objpp[i];
3045 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003047 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003048 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003050 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003052 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003054 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 check_slabp(cachep, slabp);
3056
3057 /* fixup slab chains */
3058 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003059 if (l3->free_objects > l3->free_limit) {
3060 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 slab_destroy(cachep, slabp);
3062 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003063 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 }
3065 } else {
3066 /* Unconditionally move a slab to the end of the
3067 * partial list on free - maximum time for the
3068 * other objects to be freed, too.
3069 */
Christoph Lametere498be72005-09-09 13:03:32 -07003070 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 }
3072 }
3073}
3074
Pekka Enberg343e0d72006-02-01 03:05:50 -08003075static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076{
3077 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003078 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003079 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080
3081 batchcount = ac->batchcount;
3082#if DEBUG
3083 BUG_ON(!batchcount || batchcount > ac->avail);
3084#endif
3085 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003086 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07003087 spin_lock(&l3->list_lock);
3088 if (l3->shared) {
3089 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003090 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 if (max) {
3092 if (batchcount > max)
3093 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003094 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003095 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 shared_array->avail += batchcount;
3097 goto free_done;
3098 }
3099 }
3100
Christoph Lameterff694162005-09-22 21:44:02 -07003101 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003102free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103#if STATS
3104 {
3105 int i = 0;
3106 struct list_head *p;
3107
Christoph Lametere498be72005-09-09 13:03:32 -07003108 p = l3->slabs_free.next;
3109 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 struct slab *slabp;
3111
3112 slabp = list_entry(p, struct slab, list);
3113 BUG_ON(slabp->inuse);
3114
3115 i++;
3116 p = p->next;
3117 }
3118 STATS_SET_FREEABLE(cachep, i);
3119 }
3120#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003121 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003123 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124}
3125
3126/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003127 * Release an obj back to its cache. If the obj has a constructed state, it must
3128 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003130static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003132 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133
3134 check_irq_off();
3135 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3136
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003137 if (cache_free_alien(cachep, objp))
3138 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003139
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 if (likely(ac->avail < ac->limit)) {
3141 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003142 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 return;
3144 } else {
3145 STATS_INC_FREEMISS(cachep);
3146 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003147 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148 }
3149}
3150
3151/**
3152 * kmem_cache_alloc - Allocate an object
3153 * @cachep: The cache to allocate from.
3154 * @flags: See kmalloc().
3155 *
3156 * Allocate an object from this cache. The flags are only relevant
3157 * if the cache has no available objects.
3158 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003159void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003161 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162}
3163EXPORT_SYMBOL(kmem_cache_alloc);
3164
3165/**
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003166 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3167 * @cache: The cache to allocate from.
3168 * @flags: See kmalloc().
3169 *
3170 * Allocate an object from this cache and set the allocated memory to zero.
3171 * The flags are only relevant if the cache has no available objects.
3172 */
3173void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3174{
3175 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3176 if (ret)
3177 memset(ret, 0, obj_size(cache));
3178 return ret;
3179}
3180EXPORT_SYMBOL(kmem_cache_zalloc);
3181
3182/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 * kmem_ptr_validate - check if an untrusted pointer might
3184 * be a slab entry.
3185 * @cachep: the cache we're checking against
3186 * @ptr: pointer to validate
3187 *
3188 * This verifies that the untrusted pointer looks sane:
3189 * it is _not_ a guarantee that the pointer is actually
3190 * part of the slab cache in question, but it at least
3191 * validates that the pointer can be dereferenced and
3192 * looks half-way sane.
3193 *
3194 * Currently only used for dentry validation.
3195 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003196int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003198 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003200 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003201 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 struct page *page;
3203
3204 if (unlikely(addr < min_addr))
3205 goto out;
3206 if (unlikely(addr > (unsigned long)high_memory - size))
3207 goto out;
3208 if (unlikely(addr & align_mask))
3209 goto out;
3210 if (unlikely(!kern_addr_valid(addr)))
3211 goto out;
3212 if (unlikely(!kern_addr_valid(addr + size - 1)))
3213 goto out;
3214 page = virt_to_page(ptr);
3215 if (unlikely(!PageSlab(page)))
3216 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003217 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 goto out;
3219 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003220out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 return 0;
3222}
3223
3224#ifdef CONFIG_NUMA
3225/**
3226 * kmem_cache_alloc_node - Allocate an object on the specified node
3227 * @cachep: The cache to allocate from.
3228 * @flags: See kmalloc().
3229 * @nodeid: node number of the target node.
3230 *
3231 * Identical to kmem_cache_alloc, except that this function is slow
3232 * and can sleep. And it will allocate memory on the given node, which
3233 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003234 * New and improved: it will now make sure that the object gets
3235 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003237void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238{
Christoph Lametere498be72005-09-09 13:03:32 -07003239 unsigned long save_flags;
3240 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241
Christoph Lametere498be72005-09-09 13:03:32 -07003242 cache_alloc_debugcheck_before(cachep, flags);
3243 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003244
3245 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003246 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003247 ptr = ____cache_alloc(cachep, flags);
3248 else
3249 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003250 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003251
3252 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3253 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254
Christoph Lametere498be72005-09-09 13:03:32 -07003255 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256}
3257EXPORT_SYMBOL(kmem_cache_alloc_node);
3258
Al Virodd0fc662005-10-07 07:46:04 +01003259void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003260{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003261 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003262
3263 cachep = kmem_find_general_cachep(size, flags);
3264 if (unlikely(cachep == NULL))
3265 return NULL;
3266 return kmem_cache_alloc_node(cachep, flags, node);
3267}
3268EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269#endif
3270
3271/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003272 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003274 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003275 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003277static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3278 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003280 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003282 /* If you want to save a few bytes .text space: replace
3283 * __ with kmem_.
3284 * Then kmalloc uses the uninlined functions instead of the inline
3285 * functions.
3286 */
3287 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003288 if (unlikely(cachep == NULL))
3289 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003290 return __cache_alloc(cachep, flags, caller);
3291}
3292
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003293
3294void *__kmalloc(size_t size, gfp_t flags)
3295{
Al Viro871751e2006-03-25 03:06:39 -08003296#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003297 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003298#else
3299 return __do_kmalloc(size, flags, __builtin_return_address(0));
3300#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301}
3302EXPORT_SYMBOL(__kmalloc);
3303
Al Viro871751e2006-03-25 03:06:39 -08003304#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003305void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3306{
3307 return __do_kmalloc(size, flags, caller);
3308}
3309EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003310#endif
3311
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312#ifdef CONFIG_SMP
3313/**
3314 * __alloc_percpu - allocate one copy of the object for every present
3315 * cpu in the system, zeroing them.
3316 * Objects should be dereferenced using the per_cpu_ptr macro only.
3317 *
3318 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003320void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321{
3322 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003323 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324
3325 if (!pdata)
3326 return NULL;
3327
Christoph Lametere498be72005-09-09 13:03:32 -07003328 /*
3329 * Cannot use for_each_online_cpu since a cpu may come online
3330 * and we have no way of figuring out how to fix the array
3331 * that we have allocated then....
3332 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003333 for_each_possible_cpu(i) {
Christoph Lametere498be72005-09-09 13:03:32 -07003334 int node = cpu_to_node(i);
3335
3336 if (node_online(node))
3337 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3338 else
3339 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340
3341 if (!pdata->ptrs[i])
3342 goto unwind_oom;
3343 memset(pdata->ptrs[i], 0, size);
3344 }
3345
3346 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003347 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348
Andrew Mortona737b3e2006-03-22 00:08:11 -08003349unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350 while (--i >= 0) {
3351 if (!cpu_possible(i))
3352 continue;
3353 kfree(pdata->ptrs[i]);
3354 }
3355 kfree(pdata);
3356 return NULL;
3357}
3358EXPORT_SYMBOL(__alloc_percpu);
3359#endif
3360
3361/**
3362 * kmem_cache_free - Deallocate an object
3363 * @cachep: The cache the allocation was from.
3364 * @objp: The previously allocated object.
3365 *
3366 * Free an object which was previously allocated from this
3367 * cache.
3368 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003369void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370{
3371 unsigned long flags;
3372
Pekka Enbergddc2e812006-06-23 02:03:40 -07003373 BUG_ON(virt_to_cache(objp) != cachep);
3374
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 local_irq_save(flags);
3376 __cache_free(cachep, objp);
3377 local_irq_restore(flags);
3378}
3379EXPORT_SYMBOL(kmem_cache_free);
3380
3381/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 * kfree - free previously allocated memory
3383 * @objp: pointer returned by kmalloc.
3384 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003385 * If @objp is NULL, no operation is performed.
3386 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387 * Don't free memory not originally allocated by kmalloc()
3388 * or you will run into trouble.
3389 */
3390void kfree(const void *objp)
3391{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003392 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 unsigned long flags;
3394
3395 if (unlikely(!objp))
3396 return;
3397 local_irq_save(flags);
3398 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003399 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003400 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003401 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 local_irq_restore(flags);
3403}
3404EXPORT_SYMBOL(kfree);
3405
3406#ifdef CONFIG_SMP
3407/**
3408 * free_percpu - free previously allocated percpu memory
3409 * @objp: pointer returned by alloc_percpu.
3410 *
3411 * Don't free memory not originally allocated by alloc_percpu()
3412 * The complemented objp is to check for that.
3413 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003414void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415{
3416 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003417 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418
Christoph Lametere498be72005-09-09 13:03:32 -07003419 /*
3420 * We allocate for all cpus so we cannot use for online cpu here.
3421 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003422 for_each_possible_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003423 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 kfree(p);
3425}
3426EXPORT_SYMBOL(free_percpu);
3427#endif
3428
Pekka Enberg343e0d72006-02-01 03:05:50 -08003429unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003431 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432}
3433EXPORT_SYMBOL(kmem_cache_size);
3434
Pekka Enberg343e0d72006-02-01 03:05:50 -08003435const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003436{
3437 return cachep->name;
3438}
3439EXPORT_SYMBOL_GPL(kmem_cache_name);
3440
Christoph Lametere498be72005-09-09 13:03:32 -07003441/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003442 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003443 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003444static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003445{
3446 int node;
3447 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003448 struct array_cache *new_shared;
3449 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003450
3451 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003452
Andrew Mortona737b3e2006-03-22 00:08:11 -08003453 new_alien = alloc_alien_cache(node, cachep->limit);
3454 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003455 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003456
Christoph Lameter0718dc22006-03-25 03:06:47 -08003457 new_shared = alloc_arraycache(node,
3458 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003459 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003460 if (!new_shared) {
3461 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003462 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003463 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003464
Andrew Mortona737b3e2006-03-22 00:08:11 -08003465 l3 = cachep->nodelists[node];
3466 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003467 struct array_cache *shared = l3->shared;
3468
Christoph Lametere498be72005-09-09 13:03:32 -07003469 spin_lock_irq(&l3->list_lock);
3470
Christoph Lametercafeb022006-03-25 03:06:46 -08003471 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003472 free_block(cachep, shared->entry,
3473 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003474
Christoph Lametercafeb022006-03-25 03:06:46 -08003475 l3->shared = new_shared;
3476 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003477 l3->alien = new_alien;
3478 new_alien = NULL;
3479 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003480 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003481 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003482 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003483 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003484 free_alien_cache(new_alien);
3485 continue;
3486 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003487 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003488 if (!l3) {
3489 free_alien_cache(new_alien);
3490 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003491 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003492 }
Christoph Lametere498be72005-09-09 13:03:32 -07003493
3494 kmem_list3_init(l3);
3495 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003496 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003497 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003498 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003499 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003500 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003501 cachep->nodelists[node] = l3;
3502 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003503 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003504
Andrew Mortona737b3e2006-03-22 00:08:11 -08003505fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003506 if (!cachep->next.next) {
3507 /* Cache is not active yet. Roll back what we did */
3508 node--;
3509 while (node >= 0) {
3510 if (cachep->nodelists[node]) {
3511 l3 = cachep->nodelists[node];
3512
3513 kfree(l3->shared);
3514 free_alien_cache(l3->alien);
3515 kfree(l3);
3516 cachep->nodelists[node] = NULL;
3517 }
3518 node--;
3519 }
3520 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003521 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003522}
3523
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003525 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 struct array_cache *new[NR_CPUS];
3527};
3528
3529static void do_ccupdate_local(void *info)
3530{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003531 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 struct array_cache *old;
3533
3534 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003535 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003536
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3538 new->new[smp_processor_id()] = old;
3539}
3540
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003541/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003542static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3543 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544{
3545 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003546 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003548 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003549 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003550 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3551 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003552 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003553 for (i--; i >= 0; i--)
3554 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003555 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556 }
3557 }
3558 new.cachep = cachep;
3559
Andrew Mortona07fa392006-03-22 00:08:17 -08003560 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003561
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 cachep->batchcount = batchcount;
3564 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003565 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566
Christoph Lametere498be72005-09-09 13:03:32 -07003567 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568 struct array_cache *ccold = new.new[i];
3569 if (!ccold)
3570 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003571 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003572 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003573 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574 kfree(ccold);
3575 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576
Christoph Lametere498be72005-09-09 13:03:32 -07003577 err = alloc_kmemlist(cachep);
3578 if (err) {
3579 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003580 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003581 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 return 0;
3584}
3585
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003586/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003587static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588{
3589 int err;
3590 int limit, shared;
3591
Andrew Mortona737b3e2006-03-22 00:08:11 -08003592 /*
3593 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 * - create a LIFO ordering, i.e. return objects that are cache-warm
3595 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003596 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 * bufctl chains: array operations are cheaper.
3598 * The numbers are guessed, we should auto-tune as described by
3599 * Bonwick.
3600 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003601 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003603 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003605 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003607 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 limit = 54;
3609 else
3610 limit = 120;
3611
Andrew Mortona737b3e2006-03-22 00:08:11 -08003612 /*
3613 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 * allocation behaviour: Most allocs on one cpu, most free operations
3615 * on another cpu. For these cases, an efficient object passing between
3616 * cpus is necessary. This is provided by a shared array. The array
3617 * replaces Bonwick's magazine layer.
3618 * On uniprocessor, it's functionally equivalent (but less efficient)
3619 * to a larger limit. Thus disabled by default.
3620 */
3621 shared = 0;
3622#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003623 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 shared = 8;
3625#endif
3626
3627#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003628 /*
3629 * With debugging enabled, large batchcount lead to excessively long
3630 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 */
3632 if (limit > 32)
3633 limit = 32;
3634#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003635 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 if (err)
3637 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003638 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639}
3640
Christoph Lameter1b552532006-03-22 00:09:07 -08003641/*
3642 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003643 * necessary. Note that the l3 listlock also protects the array_cache
3644 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003645 */
3646void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3647 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648{
3649 int tofree;
3650
Christoph Lameter1b552532006-03-22 00:09:07 -08003651 if (!ac || !ac->avail)
3652 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 if (ac->touched && !force) {
3654 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003655 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003656 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003657 if (ac->avail) {
3658 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3659 if (tofree > ac->avail)
3660 tofree = (ac->avail + 1) / 2;
3661 free_block(cachep, ac->entry, tofree, node);
3662 ac->avail -= tofree;
3663 memmove(ac->entry, &(ac->entry[tofree]),
3664 sizeof(void *) * ac->avail);
3665 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003666 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 }
3668}
3669
3670/**
3671 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003672 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 *
3674 * Called from workqueue/eventd every few seconds.
3675 * Purpose:
3676 * - clear the per-cpu caches for this CPU.
3677 * - return freeable pages to the main free memory pool.
3678 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003679 * If we cannot acquire the cache chain mutex then just give up - we'll try
3680 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 */
3682static void cache_reap(void *unused)
3683{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003684 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07003685 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003686 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003688 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003690 schedule_delayed_work(&__get_cpu_var(reap_work),
3691 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692 return;
3693 }
3694
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003695 list_for_each_entry(searchp, &cache_chain, next) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003696 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697 int tofree;
3698 struct slab *slabp;
3699
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 check_irq_on();
3701
Christoph Lameter35386e32006-03-22 00:09:05 -08003702 /*
3703 * We only take the l3 lock if absolutely necessary and we
3704 * have established with reasonable certainty that
3705 * we can do some work if the lock was obtained.
3706 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003707 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003708
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003709 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710
Christoph Lameteraab22072006-03-22 00:09:06 -08003711 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712
Christoph Lameter35386e32006-03-22 00:09:05 -08003713 /*
3714 * These are racy checks but it does not matter
3715 * if we skip one check or scan twice.
3716 */
Christoph Lametere498be72005-09-09 13:03:32 -07003717 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003718 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719
Christoph Lametere498be72005-09-09 13:03:32 -07003720 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721
Christoph Lameteraab22072006-03-22 00:09:06 -08003722 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723
Christoph Lametere498be72005-09-09 13:03:32 -07003724 if (l3->free_touched) {
3725 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003726 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 }
3728
Andrew Mortona737b3e2006-03-22 00:08:11 -08003729 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3730 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003732 /*
3733 * Do not lock if there are no free blocks.
3734 */
3735 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 break;
3737
Christoph Lameter35386e32006-03-22 00:09:05 -08003738 spin_lock_irq(&l3->list_lock);
3739 p = l3->slabs_free.next;
3740 if (p == &(l3->slabs_free)) {
3741 spin_unlock_irq(&l3->list_lock);
3742 break;
3743 }
3744
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 slabp = list_entry(p, struct slab, list);
3746 BUG_ON(slabp->inuse);
3747 list_del(&slabp->list);
3748 STATS_INC_REAPED(searchp);
3749
Andrew Mortona737b3e2006-03-22 00:08:11 -08003750 /*
3751 * Safe to drop the lock. The slab is no longer linked
3752 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753 * cache_chain_lock
3754 */
Christoph Lametere498be72005-09-09 13:03:32 -07003755 l3->free_objects -= searchp->num;
3756 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003758 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003759next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 cond_resched();
3761 }
3762 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003763 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003764 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003765 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003766 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767}
3768
3769#ifdef CONFIG_PROC_FS
3770
Pekka Enberg85289f92006-01-08 01:00:36 -08003771static void print_slabinfo_header(struct seq_file *m)
3772{
3773 /*
3774 * Output format version, so at least we can change it
3775 * without _too_ many complaints.
3776 */
3777#if STATS
3778 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3779#else
3780 seq_puts(m, "slabinfo - version: 2.1\n");
3781#endif
3782 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3783 "<objperslab> <pagesperslab>");
3784 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3785 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3786#if STATS
3787 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003788 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003789 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3790#endif
3791 seq_putc(m, '\n');
3792}
3793
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794static void *s_start(struct seq_file *m, loff_t *pos)
3795{
3796 loff_t n = *pos;
3797 struct list_head *p;
3798
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003799 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003800 if (!n)
3801 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 p = cache_chain.next;
3803 while (n--) {
3804 p = p->next;
3805 if (p == &cache_chain)
3806 return NULL;
3807 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003808 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809}
3810
3811static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3812{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003813 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003815 return cachep->next.next == &cache_chain ?
3816 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817}
3818
3819static void s_stop(struct seq_file *m, void *p)
3820{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003821 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003822}
3823
3824static int s_show(struct seq_file *m, void *p)
3825{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003826 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003827 struct slab *slabp;
3828 unsigned long active_objs;
3829 unsigned long num_objs;
3830 unsigned long active_slabs = 0;
3831 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003832 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003834 int node;
3835 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837 active_objs = 0;
3838 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003839 for_each_online_node(node) {
3840 l3 = cachep->nodelists[node];
3841 if (!l3)
3842 continue;
3843
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003844 check_irq_on();
3845 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003846
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003847 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003848 if (slabp->inuse != cachep->num && !error)
3849 error = "slabs_full accounting error";
3850 active_objs += cachep->num;
3851 active_slabs++;
3852 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003853 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003854 if (slabp->inuse == cachep->num && !error)
3855 error = "slabs_partial inuse accounting error";
3856 if (!slabp->inuse && !error)
3857 error = "slabs_partial/inuse accounting error";
3858 active_objs += slabp->inuse;
3859 active_slabs++;
3860 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003861 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003862 if (slabp->inuse && !error)
3863 error = "slabs_free/inuse accounting error";
3864 num_slabs++;
3865 }
3866 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003867 if (l3->shared)
3868 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003869
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003870 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003872 num_slabs += active_slabs;
3873 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003874 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 error = "free_objects accounting error";
3876
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003877 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 if (error)
3879 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3880
3881 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003882 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003883 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003885 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003886 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003887 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003889 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890 unsigned long high = cachep->high_mark;
3891 unsigned long allocs = cachep->num_allocations;
3892 unsigned long grown = cachep->grown;
3893 unsigned long reaped = cachep->reaped;
3894 unsigned long errors = cachep->errors;
3895 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003897 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003898 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899
Christoph Lametere498be72005-09-09 13:03:32 -07003900 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003901 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003902 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003903 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 }
3905 /* cpu stats */
3906 {
3907 unsigned long allochit = atomic_read(&cachep->allochit);
3908 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3909 unsigned long freehit = atomic_read(&cachep->freehit);
3910 unsigned long freemiss = atomic_read(&cachep->freemiss);
3911
3912 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003913 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914 }
3915#endif
3916 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 return 0;
3918}
3919
3920/*
3921 * slabinfo_op - iterator that generates /proc/slabinfo
3922 *
3923 * Output layout:
3924 * cache-name
3925 * num-active-objs
3926 * total-objs
3927 * object size
3928 * num-active-slabs
3929 * total-slabs
3930 * num-pages-per-slab
3931 * + further values on SMP and with statistics enabled
3932 */
3933
3934struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003935 .start = s_start,
3936 .next = s_next,
3937 .stop = s_stop,
3938 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939};
3940
3941#define MAX_SLABINFO_WRITE 128
3942/**
3943 * slabinfo_write - Tuning for the slab allocator
3944 * @file: unused
3945 * @buffer: user buffer
3946 * @count: data length
3947 * @ppos: unused
3948 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003949ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3950 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003952 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003954 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003955
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 if (count > MAX_SLABINFO_WRITE)
3957 return -EINVAL;
3958 if (copy_from_user(&kbuf, buffer, count))
3959 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003960 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961
3962 tmp = strchr(kbuf, ' ');
3963 if (!tmp)
3964 return -EINVAL;
3965 *tmp = '\0';
3966 tmp++;
3967 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3968 return -EINVAL;
3969
3970 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003971 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003973 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003975 if (limit < 1 || batchcount < 1 ||
3976 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003977 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003979 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003980 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981 }
3982 break;
3983 }
3984 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003985 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 if (res >= 0)
3987 res = count;
3988 return res;
3989}
Al Viro871751e2006-03-25 03:06:39 -08003990
3991#ifdef CONFIG_DEBUG_SLAB_LEAK
3992
3993static void *leaks_start(struct seq_file *m, loff_t *pos)
3994{
3995 loff_t n = *pos;
3996 struct list_head *p;
3997
3998 mutex_lock(&cache_chain_mutex);
3999 p = cache_chain.next;
4000 while (n--) {
4001 p = p->next;
4002 if (p == &cache_chain)
4003 return NULL;
4004 }
4005 return list_entry(p, struct kmem_cache, next);
4006}
4007
4008static inline int add_caller(unsigned long *n, unsigned long v)
4009{
4010 unsigned long *p;
4011 int l;
4012 if (!v)
4013 return 1;
4014 l = n[1];
4015 p = n + 2;
4016 while (l) {
4017 int i = l/2;
4018 unsigned long *q = p + 2 * i;
4019 if (*q == v) {
4020 q[1]++;
4021 return 1;
4022 }
4023 if (*q > v) {
4024 l = i;
4025 } else {
4026 p = q + 2;
4027 l -= i + 1;
4028 }
4029 }
4030 if (++n[1] == n[0])
4031 return 0;
4032 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4033 p[0] = v;
4034 p[1] = 1;
4035 return 1;
4036}
4037
4038static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4039{
4040 void *p;
4041 int i;
4042 if (n[0] == n[1])
4043 return;
4044 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4045 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4046 continue;
4047 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4048 return;
4049 }
4050}
4051
4052static void show_symbol(struct seq_file *m, unsigned long address)
4053{
4054#ifdef CONFIG_KALLSYMS
4055 char *modname;
4056 const char *name;
4057 unsigned long offset, size;
4058 char namebuf[KSYM_NAME_LEN+1];
4059
4060 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4061
4062 if (name) {
4063 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4064 if (modname)
4065 seq_printf(m, " [%s]", modname);
4066 return;
4067 }
4068#endif
4069 seq_printf(m, "%p", (void *)address);
4070}
4071
4072static int leaks_show(struct seq_file *m, void *p)
4073{
4074 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004075 struct slab *slabp;
4076 struct kmem_list3 *l3;
4077 const char *name;
4078 unsigned long *n = m->private;
4079 int node;
4080 int i;
4081
4082 if (!(cachep->flags & SLAB_STORE_USER))
4083 return 0;
4084 if (!(cachep->flags & SLAB_RED_ZONE))
4085 return 0;
4086
4087 /* OK, we can do it */
4088
4089 n[1] = 0;
4090
4091 for_each_online_node(node) {
4092 l3 = cachep->nodelists[node];
4093 if (!l3)
4094 continue;
4095
4096 check_irq_on();
4097 spin_lock_irq(&l3->list_lock);
4098
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004099 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004100 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004101 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004102 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004103 spin_unlock_irq(&l3->list_lock);
4104 }
4105 name = cachep->name;
4106 if (n[0] == n[1]) {
4107 /* Increase the buffer size */
4108 mutex_unlock(&cache_chain_mutex);
4109 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4110 if (!m->private) {
4111 /* Too bad, we are really out */
4112 m->private = n;
4113 mutex_lock(&cache_chain_mutex);
4114 return -ENOMEM;
4115 }
4116 *(unsigned long *)m->private = n[0] * 2;
4117 kfree(n);
4118 mutex_lock(&cache_chain_mutex);
4119 /* Now make sure this entry will be retried */
4120 m->count = m->size;
4121 return 0;
4122 }
4123 for (i = 0; i < n[1]; i++) {
4124 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4125 show_symbol(m, n[2*i+2]);
4126 seq_putc(m, '\n');
4127 }
4128 return 0;
4129}
4130
4131struct seq_operations slabstats_op = {
4132 .start = leaks_start,
4133 .next = s_next,
4134 .stop = s_stop,
4135 .show = leaks_show,
4136};
4137#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138#endif
4139
Manfred Spraul00e145b2005-09-03 15:55:07 -07004140/**
4141 * ksize - get the actual amount of memory allocated for a given object
4142 * @objp: Pointer to the object
4143 *
4144 * kmalloc may internally round up allocations and return more memory
4145 * than requested. ksize() can be used to determine the actual amount of
4146 * memory allocated. The caller may use this additional memory, even though
4147 * a smaller amount of memory was initially specified with the kmalloc call.
4148 * The caller must guarantee that objp points to a valid object previously
4149 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4150 * must not be freed during the duration of the call.
4151 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152unsigned int ksize(const void *objp)
4153{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004154 if (unlikely(objp == NULL))
4155 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004157 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158}