blob: af5c5237e11a0df9d2f384383a2e8cd3b846a792 [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>
92#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700106#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800107#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800108#include <linux/mutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110#include <asm/uaccess.h>
111#include <asm/cacheflush.h>
112#include <asm/tlbflush.h>
113#include <asm/page.h>
114
115/*
116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
117 * SLAB_RED_ZONE & SLAB_POISON.
118 * 0 for faster, smaller code (especially in the critical paths).
119 *
120 * STATS - 1 to collect stats for /proc/slabinfo.
121 * 0 for faster, smaller code (especially in the critical paths).
122 *
123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
124 */
125
126#ifdef CONFIG_DEBUG_SLAB
127#define DEBUG 1
128#define STATS 1
129#define FORCED_DEBUG 1
130#else
131#define DEBUG 0
132#define STATS 0
133#define FORCED_DEBUG 0
134#endif
135
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136/* Shouldn't this be in a header file somewhere? */
137#define BYTES_PER_WORD sizeof(void *)
138
139#ifndef cache_line_size
140#define cache_line_size() L1_CACHE_BYTES
141#endif
142
143#ifndef ARCH_KMALLOC_MINALIGN
144/*
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
150 * Note that this flag disables some debug features.
151 */
152#define ARCH_KMALLOC_MINALIGN 0
153#endif
154
155#ifndef ARCH_SLAB_MINALIGN
156/*
157 * Enforce a minimum alignment for all caches.
158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
161 * some debug features.
162 */
163#define ARCH_SLAB_MINALIGN 0
164#endif
165
166#ifndef ARCH_KMALLOC_FLAGS
167#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
168#endif
169
170/* Legal flag mask for kmem_cache_create(). */
171#if DEBUG
172# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800174 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800179# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183#endif
184
185/*
186 * kmem_bufctl_t:
187 *
188 * Bufctl's are used for linking objs within a slab
189 * linked offsets.
190 *
191 * This implementation relies on "struct page" for locating the cache &
192 * slab an object belongs to.
193 * This allows the bufctl structure to be small (one int), but limits
194 * the number of objects a slab (not a cache) can contain when off-slab
195 * bufctls are used. The limit is the size of the largest general cache
196 * that does not use off-slab slabs.
197 * For 32bit archs with 4 kB pages, is this 56.
198 * This is not serious, as it is only for large objects, when it is unwise
199 * to have too many per slab.
200 * Note: This limit can be raised by introducing a general cache whose size
201 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
202 */
203
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700204typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
206#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800207#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
208#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209
210/* Max number of objs-per-slab for caches which use off-slab slabs.
211 * Needed to avoid a possible looping condition in cache_grow().
212 */
213static unsigned long offslab_limit;
214
215/*
216 * struct slab
217 *
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 */
222struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
227 kmem_bufctl_t free;
228 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229};
230
231/*
232 * struct slab_rcu
233 *
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
241 *
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
244 *
245 * We assume struct slab_rcu can overlay struct slab when destroying.
246 */
247struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800249 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800250 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251};
252
253/*
254 * struct array_cache
255 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 * Purpose:
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
260 *
261 * The limit is stored in the per-cpu structure to reduce the data cache
262 * footprint.
263 *
264 */
265struct array_cache {
266 unsigned int avail;
267 unsigned int limit;
268 unsigned int batchcount;
269 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700270 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800271 void *entry[0]; /*
272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
274 * the entries.
275 * [0] is for gcc 2.95. It should really be [].
276 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277};
278
Andrew Mortona737b3e2006-03-22 00:08:11 -0800279/*
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 */
283#define BOOT_CPUCACHE_ENTRIES 1
284struct arraycache_init {
285 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800286 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287};
288
289/*
Christoph Lametere498be72005-09-09 13:03:32 -0700290 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
292struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800298 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304};
305
Christoph Lametere498be72005-09-09 13:03:32 -0700306/*
307 * Need this for bootstrapping a per node allocator.
308 */
309#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
310struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
311#define CACHE_CACHE 0
312#define SIZE_AC 1
313#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314
Christoph Lametere498be72005-09-09 13:03:32 -0700315/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800316 * This function must be completely optimized away if a constant is passed to
317 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700318 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700319static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700320{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800321 extern void __bad_size(void);
322
Christoph Lametere498be72005-09-09 13:03:32 -0700323 if (__builtin_constant_p(size)) {
324 int i = 0;
325
326#define CACHE(x) \
327 if (size <=x) \
328 return i; \
329 else \
330 i++;
331#include "linux/kmalloc_sizes.h"
332#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800333 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700334 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800335 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700336 return 0;
337}
338
339#define INDEX_AC index_of(sizeof(struct arraycache_init))
340#define INDEX_L3 index_of(sizeof(struct kmem_list3))
341
Pekka Enberg5295a742006-02-01 03:05:48 -0800342static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700343{
344 INIT_LIST_HEAD(&parent->slabs_full);
345 INIT_LIST_HEAD(&parent->slabs_partial);
346 INIT_LIST_HEAD(&parent->slabs_free);
347 parent->shared = NULL;
348 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800349 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700350 spin_lock_init(&parent->list_lock);
351 parent->free_objects = 0;
352 parent->free_touched = 0;
353}
354
Andrew Mortona737b3e2006-03-22 00:08:11 -0800355#define MAKE_LIST(cachep, listp, slab, nodeid) \
356 do { \
357 INIT_LIST_HEAD(listp); \
358 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700359 } while (0)
360
Andrew Mortona737b3e2006-03-22 00:08:11 -0800361#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
362 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700363 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
365 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
366 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367
368/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800369 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 *
371 * manages a cache.
372 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800373
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800374struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800376 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800377/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800378 unsigned int batchcount;
379 unsigned int limit;
380 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800381
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800382 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800385
Andrew Mortona737b3e2006-03-22 00:08:11 -0800386 unsigned int flags; /* constant flags */
387 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800391 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392
393 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800394 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
Andrew Mortona737b3e2006-03-22 00:08:11 -0800396 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800398 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800400 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800403 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
405 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800408/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800409 const char *name;
410 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800412/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 unsigned long num_active;
415 unsigned long num_allocations;
416 unsigned long high_mark;
417 unsigned long grown;
418 unsigned long reaped;
419 unsigned long errors;
420 unsigned long max_freeable;
421 unsigned long node_allocs;
422 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700423 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800424 atomic_t allochit;
425 atomic_t allocmiss;
426 atomic_t freehit;
427 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428#endif
429#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800430 /*
431 * If debugging is enabled, then the allocator can add additional
432 * fields and/or padding to every object. buffer_size contains the total
433 * object size including these internal fields, the following two
434 * variables contain the offset to the user object and its size.
435 */
436 int obj_offset;
437 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438#endif
439};
440
441#define CFLGS_OFF_SLAB (0x80000000UL)
442#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
443
444#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800445/*
446 * Optimization question: fewer reaps means less probability for unnessary
447 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100449 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 * which could lock up otherwise freeable slabs.
451 */
452#define REAPTIMEOUT_CPUC (2*HZ)
453#define REAPTIMEOUT_LIST3 (4*HZ)
454
455#if STATS
456#define STATS_INC_ACTIVE(x) ((x)->num_active++)
457#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
458#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
459#define STATS_INC_GROWN(x) ((x)->grown++)
460#define STATS_INC_REAPED(x) ((x)->reaped++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800461#define STATS_SET_HIGH(x) \
462 do { \
463 if ((x)->num_active > (x)->high_mark) \
464 (x)->high_mark = (x)->num_active; \
465 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466#define STATS_INC_ERR(x) ((x)->errors++)
467#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700468#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700469#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800470#define STATS_SET_FREEABLE(x, i) \
471 do { \
472 if ((x)->max_freeable < i) \
473 (x)->max_freeable = i; \
474 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
476#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
477#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
478#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
479#else
480#define STATS_INC_ACTIVE(x) do { } while (0)
481#define STATS_DEC_ACTIVE(x) do { } while (0)
482#define STATS_INC_ALLOCED(x) do { } while (0)
483#define STATS_INC_GROWN(x) do { } while (0)
484#define STATS_INC_REAPED(x) do { } while (0)
485#define STATS_SET_HIGH(x) do { } while (0)
486#define STATS_INC_ERR(x) do { } while (0)
487#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700488#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700489#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800490#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_INC_ALLOCHIT(x) do { } while (0)
492#define STATS_INC_ALLOCMISS(x) do { } while (0)
493#define STATS_INC_FREEHIT(x) do { } while (0)
494#define STATS_INC_FREEMISS(x) do { } while (0)
495#endif
496
497#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -0800498/*
499 * Magic nums for obj red zoning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 * Placed in the first word before and the first word after an obj.
501 */
502#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
503#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
504
505/* ...and for poisoning */
506#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
507#define POISON_FREE 0x6b /* for use-after-free poisoning */
508#define POISON_END 0xa5 /* end-byte of poisoning */
509
Andrew Mortona737b3e2006-03-22 00:08:11 -0800510/*
511 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 * the end of an object is aligned with the end of the real
515 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800516 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800518 * cachep->obj_offset: The real object.
519 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800520 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
521 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800523static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531}
532
Pekka Enberg343e0d72006-02-01 03:05:50 -0800533static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
535 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800536 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537}
538
Pekka Enberg343e0d72006-02-01 03:05:50 -0800539static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540{
541 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
542 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800543 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800544 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800545 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546}
547
Pekka Enberg343e0d72006-02-01 03:05:50 -0800548static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549{
550 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800551 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552}
553
554#else
555
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800556#define obj_offset(x) 0
557#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
559#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
560#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
561
562#endif
563
564/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800565 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
566 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 */
568#if defined(CONFIG_LARGE_ALLOCS)
569#define MAX_OBJ_ORDER 13 /* up to 32Mb */
570#define MAX_GFP_ORDER 13 /* up to 32Mb */
571#elif defined(CONFIG_MMU)
572#define MAX_OBJ_ORDER 5 /* 32 pages */
573#define MAX_GFP_ORDER 5 /* 32 pages */
574#else
575#define MAX_OBJ_ORDER 8 /* up to 1Mb */
576#define MAX_GFP_ORDER 8 /* up to 1Mb */
577#endif
578
579/*
580 * Do not go above this order unless 0 objects fit into the slab.
581 */
582#define BREAK_GFP_ORDER_HI 1
583#define BREAK_GFP_ORDER_LO 0
584static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
585
Andrew Mortona737b3e2006-03-22 00:08:11 -0800586/*
587 * Functions for storing/retrieving the cachep and or slab from the page
588 * allocator. These are used to find the slab an obj belongs to. With kfree(),
589 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800591static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
592{
593 page->lru.next = (struct list_head *)cache;
594}
595
596static inline struct kmem_cache *page_get_cache(struct page *page)
597{
Nick Piggin84097512006-03-22 00:08:34 -0800598 if (unlikely(PageCompound(page)))
599 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800600 return (struct kmem_cache *)page->lru.next;
601}
602
603static inline void page_set_slab(struct page *page, struct slab *slab)
604{
605 page->lru.prev = (struct list_head *)slab;
606}
607
608static inline struct slab *page_get_slab(struct page *page)
609{
Nick Piggin84097512006-03-22 00:08:34 -0800610 if (unlikely(PageCompound(page)))
611 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800612 return (struct slab *)page->lru.prev;
613}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800615static inline struct kmem_cache *virt_to_cache(const void *obj)
616{
617 struct page *page = virt_to_page(obj);
618 return page_get_cache(page);
619}
620
621static inline struct slab *virt_to_slab(const void *obj)
622{
623 struct page *page = virt_to_page(obj);
624 return page_get_slab(page);
625}
626
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800627static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
628 unsigned int idx)
629{
630 return slab->s_mem + cache->buffer_size * idx;
631}
632
633static inline unsigned int obj_to_index(struct kmem_cache *cache,
634 struct slab *slab, void *obj)
635{
636 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
637}
638
Andrew Mortona737b3e2006-03-22 00:08:11 -0800639/*
640 * These are the default caches for kmalloc. Custom caches can have other sizes.
641 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642struct cache_sizes malloc_sizes[] = {
643#define CACHE(x) { .cs_size = (x) },
644#include <linux/kmalloc_sizes.h>
645 CACHE(ULONG_MAX)
646#undef CACHE
647};
648EXPORT_SYMBOL(malloc_sizes);
649
650/* Must match cache_sizes above. Out of line to keep cache footprint low. */
651struct cache_names {
652 char *name;
653 char *name_dma;
654};
655
656static struct cache_names __initdata cache_names[] = {
657#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
658#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800659 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660#undef CACHE
661};
662
663static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800666 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
668/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800669static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800670 .batchcount = 1,
671 .limit = BOOT_CPUCACHE_ENTRIES,
672 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800673 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800674 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800676 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677#endif
678};
679
680/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800681static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682static struct list_head cache_chain;
683
684/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800685 * vm_enough_memory() looks at this to determine how many slab-allocated pages
686 * are possibly freeable under pressure
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 *
688 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
689 */
690atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
692/*
693 * chicken and egg problem: delay the per-cpu array allocation
694 * until the general caches are up.
695 */
696static enum {
697 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700698 PARTIAL_AC,
699 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 FULL
701} g_cpucache_up;
702
703static DEFINE_PER_CPU(struct work_struct, reap_work);
704
Andrew Mortona737b3e2006-03-22 00:08:11 -0800705static void free_block(struct kmem_cache *cachep, void **objpp, int len,
706 int node);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800707static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800708static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800709static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
Pekka Enberg343e0d72006-02-01 03:05:50 -0800711static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712{
713 return cachep->array[smp_processor_id()];
714}
715
Andrew Mortona737b3e2006-03-22 00:08:11 -0800716static inline struct kmem_cache *__find_general_cachep(size_t size,
717 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718{
719 struct cache_sizes *csizep = malloc_sizes;
720
721#if DEBUG
722 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800723 * kmem_cache_create(), or __kmalloc(), before
724 * the generic caches are initialized.
725 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700726 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727#endif
728 while (size > csizep->cs_size)
729 csizep++;
730
731 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700732 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 * has cs_{dma,}cachep==NULL. Thus no special case
734 * for large kmalloc calls required.
735 */
736 if (unlikely(gfpflags & GFP_DMA))
737 return csizep->cs_dmacachep;
738 return csizep->cs_cachep;
739}
740
Pekka Enberg343e0d72006-02-01 03:05:50 -0800741struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700742{
743 return __find_general_cachep(size, gfpflags);
744}
745EXPORT_SYMBOL(kmem_find_general_cachep);
746
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800747static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800749 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
750}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751
Andrew Mortona737b3e2006-03-22 00:08:11 -0800752/*
753 * Calculate the number of objects and left-over bytes for a given buffer size.
754 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800755static void cache_estimate(unsigned long gfporder, size_t buffer_size,
756 size_t align, int flags, size_t *left_over,
757 unsigned int *num)
758{
759 int nr_objs;
760 size_t mgmt_size;
761 size_t slab_size = PAGE_SIZE << gfporder;
762
763 /*
764 * The slab management structure can be either off the slab or
765 * on it. For the latter case, the memory allocated for a
766 * slab is used for:
767 *
768 * - The struct slab
769 * - One kmem_bufctl_t for each object
770 * - Padding to respect alignment of @align
771 * - @buffer_size bytes for each object
772 *
773 * If the slab management structure is off the slab, then the
774 * alignment will already be calculated into the size. Because
775 * the slabs are all pages aligned, the objects will be at the
776 * correct alignment when allocated.
777 */
778 if (flags & CFLGS_OFF_SLAB) {
779 mgmt_size = 0;
780 nr_objs = slab_size / buffer_size;
781
782 if (nr_objs > SLAB_LIMIT)
783 nr_objs = SLAB_LIMIT;
784 } else {
785 /*
786 * Ignore padding for the initial guess. The padding
787 * is at most @align-1 bytes, and @buffer_size is at
788 * least @align. In the worst case, this result will
789 * be one greater than the number of objects that fit
790 * into the memory allocation when taking the padding
791 * into account.
792 */
793 nr_objs = (slab_size - sizeof(struct slab)) /
794 (buffer_size + sizeof(kmem_bufctl_t));
795
796 /*
797 * This calculated number will be either the right
798 * amount, or one greater than what we want.
799 */
800 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
801 > slab_size)
802 nr_objs--;
803
804 if (nr_objs > SLAB_LIMIT)
805 nr_objs = SLAB_LIMIT;
806
807 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800809 *num = nr_objs;
810 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811}
812
813#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
814
Andrew Mortona737b3e2006-03-22 00:08:11 -0800815static void __slab_error(const char *function, struct kmem_cache *cachep,
816 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817{
818 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800819 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 dump_stack();
821}
822
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800823#ifdef CONFIG_NUMA
824/*
825 * Special reaping functions for NUMA systems called from cache_reap().
826 * These take care of doing round robin flushing of alien caches (containing
827 * objects freed on different nodes from which they were allocated) and the
828 * flushing of remote pcps by calling drain_node_pages.
829 */
830static DEFINE_PER_CPU(unsigned long, reap_node);
831
832static void init_reap_node(int cpu)
833{
834 int node;
835
836 node = next_node(cpu_to_node(cpu), node_online_map);
837 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800838 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800839
840 __get_cpu_var(reap_node) = node;
841}
842
843static void next_reap_node(void)
844{
845 int node = __get_cpu_var(reap_node);
846
847 /*
848 * Also drain per cpu pages on remote zones
849 */
850 if (node != numa_node_id())
851 drain_node_pages(node);
852
853 node = next_node(node, node_online_map);
854 if (unlikely(node >= MAX_NUMNODES))
855 node = first_node(node_online_map);
856 __get_cpu_var(reap_node) = node;
857}
858
859#else
860#define init_reap_node(cpu) do { } while (0)
861#define next_reap_node(void) do { } while (0)
862#endif
863
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864/*
865 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
866 * via the workqueue/eventd.
867 * Add the CPU number into the expiration time to minimize the possibility of
868 * the CPUs getting into lockstep and contending for the global cache chain
869 * lock.
870 */
871static void __devinit start_cpu_timer(int cpu)
872{
873 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
874
875 /*
876 * When this gets called from do_initcalls via cpucache_init(),
877 * init_workqueues() has already run, so keventd will be setup
878 * at that time.
879 */
880 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800881 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 INIT_WORK(reap_work, cache_reap, NULL);
883 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
884 }
885}
886
Christoph Lametere498be72005-09-09 13:03:32 -0700887static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800888 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800890 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 struct array_cache *nc = NULL;
892
Christoph Lametere498be72005-09-09 13:03:32 -0700893 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 if (nc) {
895 nc->avail = 0;
896 nc->limit = entries;
897 nc->batchcount = batchcount;
898 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700899 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 }
901 return nc;
902}
903
Christoph Lameter3ded1752006-03-25 03:06:44 -0800904/*
905 * Transfer objects in one arraycache to another.
906 * Locking must be handled by the caller.
907 *
908 * Return the number of entries transferred.
909 */
910static int transfer_objects(struct array_cache *to,
911 struct array_cache *from, unsigned int max)
912{
913 /* Figure out how many entries to transfer */
914 int nr = min(min(from->avail, max), to->limit - to->avail);
915
916 if (!nr)
917 return 0;
918
919 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
920 sizeof(void *) *nr);
921
922 from->avail -= nr;
923 to->avail += nr;
924 to->touched = 1;
925 return nr;
926}
927
Christoph Lametere498be72005-09-09 13:03:32 -0700928#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800929static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800930static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800931
Pekka Enberg5295a742006-02-01 03:05:48 -0800932static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700933{
934 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800935 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700936 int i;
937
938 if (limit > 1)
939 limit = 12;
940 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
941 if (ac_ptr) {
942 for_each_node(i) {
943 if (i == node || !node_online(i)) {
944 ac_ptr[i] = NULL;
945 continue;
946 }
947 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
948 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800949 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700950 kfree(ac_ptr[i]);
951 kfree(ac_ptr);
952 return NULL;
953 }
954 }
955 }
956 return ac_ptr;
957}
958
Pekka Enberg5295a742006-02-01 03:05:48 -0800959static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700960{
961 int i;
962
963 if (!ac_ptr)
964 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700965 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800966 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700967 kfree(ac_ptr);
968}
969
Pekka Enberg343e0d72006-02-01 03:05:50 -0800970static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800971 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700972{
973 struct kmem_list3 *rl3 = cachep->nodelists[node];
974
975 if (ac->avail) {
976 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800977 /*
978 * Stuff objects into the remote nodes shared array first.
979 * That way we could avoid the overhead of putting the objects
980 * into the free lists and getting them back later.
981 */
982 transfer_objects(rl3->shared, ac, ac->limit);
983
Christoph Lameterff694162005-09-22 21:44:02 -0700984 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700985 ac->avail = 0;
986 spin_unlock(&rl3->list_lock);
987 }
988}
989
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800990/*
991 * Called from cache_reap() to regularly drain alien caches round robin.
992 */
993static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
994{
995 int node = __get_cpu_var(reap_node);
996
997 if (l3->alien) {
998 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -0800999
1000 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001001 __drain_alien_cache(cachep, ac, node);
1002 spin_unlock_irq(&ac->lock);
1003 }
1004 }
1005}
1006
Andrew Mortona737b3e2006-03-22 00:08:11 -08001007static void drain_alien_cache(struct kmem_cache *cachep,
1008 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001009{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001010 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001011 struct array_cache *ac;
1012 unsigned long flags;
1013
1014 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001015 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001016 if (ac) {
1017 spin_lock_irqsave(&ac->lock, flags);
1018 __drain_alien_cache(cachep, ac, i);
1019 spin_unlock_irqrestore(&ac->lock, flags);
1020 }
1021 }
1022}
1023#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001024
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001025#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001026#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001027
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001028static inline struct array_cache **alloc_alien_cache(int node, int limit)
1029{
1030 return (struct array_cache **) 0x01020304ul;
1031}
1032
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001033static inline void free_alien_cache(struct array_cache **ac_ptr)
1034{
1035}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001036
Christoph Lametere498be72005-09-09 13:03:32 -07001037#endif
1038
Chandra Seetharaman83d722f2006-04-24 19:35:21 -07001039static int cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001040 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041{
1042 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001043 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001044 struct kmem_list3 *l3 = NULL;
1045 int node = cpu_to_node(cpu);
1046 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047
1048 switch (action) {
1049 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001050 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001051 /*
1052 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001053 * alloc_arraycache's are going to use this list.
1054 * kmalloc_node allows us to add the slab to the right
1055 * kmem_list3 and not this cpu's kmem_list3
1056 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057
Christoph Lametere498be72005-09-09 13:03:32 -07001058 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001059 /*
1060 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001061 * begin anything. Make sure some other cpu on this
1062 * node has not already allocated this
1063 */
1064 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001065 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1066 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001067 goto bad;
1068 kmem_list3_init(l3);
1069 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001070 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001071
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001072 /*
1073 * The l3s don't come and go as CPUs come and
1074 * go. cache_chain_mutex is sufficient
1075 * protection here.
1076 */
Christoph Lametere498be72005-09-09 13:03:32 -07001077 cachep->nodelists[node] = l3;
1078 }
1079
1080 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1081 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001082 (1 + nr_cpus_node(node)) *
1083 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001084 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1085 }
1086
Andrew Mortona737b3e2006-03-22 00:08:11 -08001087 /*
1088 * Now we can go ahead with allocating the shared arrays and
1089 * array caches
1090 */
Christoph Lametere498be72005-09-09 13:03:32 -07001091 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001092 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001093 struct array_cache *shared;
1094 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001095
Christoph Lametere498be72005-09-09 13:03:32 -07001096 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001097 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 if (!nc)
1099 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001100 shared = alloc_arraycache(node,
1101 cachep->shared * cachep->batchcount,
1102 0xbaadf00d);
1103 if (!shared)
1104 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001105
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001106 alien = alloc_alien_cache(node, cachep->limit);
1107 if (!alien)
1108 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001110 l3 = cachep->nodelists[node];
1111 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001112
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001113 spin_lock_irq(&l3->list_lock);
1114 if (!l3->shared) {
1115 /*
1116 * We are serialised from CPU_DEAD or
1117 * CPU_UP_CANCELLED by the cpucontrol lock
1118 */
1119 l3->shared = shared;
1120 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001121 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001122#ifdef CONFIG_NUMA
1123 if (!l3->alien) {
1124 l3->alien = alien;
1125 alien = NULL;
1126 }
1127#endif
1128 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001129 kfree(shared);
1130 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001132 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 break;
1134 case CPU_ONLINE:
1135 start_cpu_timer(cpu);
1136 break;
1137#ifdef CONFIG_HOTPLUG_CPU
1138 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001139 /*
1140 * Even if all the cpus of a node are down, we don't free the
1141 * kmem_list3 of any cache. This to avoid a race between
1142 * cpu_down, and a kmalloc allocation from another cpu for
1143 * memory from the node of the cpu going down. The list3
1144 * structure is usually allocated from kmem_cache_create() and
1145 * gets destroyed at kmem_cache_destroy().
1146 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 /* fall thru */
1148 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001149 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 list_for_each_entry(cachep, &cache_chain, next) {
1151 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001152 struct array_cache *shared;
1153 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001154 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Christoph Lametere498be72005-09-09 13:03:32 -07001156 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 /* cpu is dead; no one can alloc from it. */
1158 nc = cachep->array[cpu];
1159 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001160 l3 = cachep->nodelists[node];
1161
1162 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001163 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001164
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001165 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001166
1167 /* Free limit for this kmem_list3 */
1168 l3->free_limit -= cachep->batchcount;
1169 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001170 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001171
1172 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001173 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001174 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001175 }
Christoph Lametere498be72005-09-09 13:03:32 -07001176
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001177 shared = l3->shared;
1178 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001179 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001180 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001181 l3->shared = NULL;
1182 }
Christoph Lametere498be72005-09-09 13:03:32 -07001183
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001184 alien = l3->alien;
1185 l3->alien = NULL;
1186
1187 spin_unlock_irq(&l3->list_lock);
1188
1189 kfree(shared);
1190 if (alien) {
1191 drain_alien_cache(cachep, alien);
1192 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001193 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001194free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 kfree(nc);
1196 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001197 /*
1198 * In the previous loop, all the objects were freed to
1199 * the respective cache's slabs, now we can go ahead and
1200 * shrink each nodelist to its limit.
1201 */
1202 list_for_each_entry(cachep, &cache_chain, next) {
1203 l3 = cachep->nodelists[node];
1204 if (!l3)
1205 continue;
1206 spin_lock_irq(&l3->list_lock);
1207 /* free slabs belonging to this node */
1208 __node_shrink(cachep, node);
1209 spin_unlock_irq(&l3->list_lock);
1210 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001211 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 break;
1213#endif
1214 }
1215 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001216bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001217 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 return NOTIFY_BAD;
1219}
1220
1221static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1222
Christoph Lametere498be72005-09-09 13:03:32 -07001223/*
1224 * swap the static kmem_list3 with kmalloced memory
1225 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001226static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1227 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001228{
1229 struct kmem_list3 *ptr;
1230
1231 BUG_ON(cachep->nodelists[nodeid] != list);
1232 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1233 BUG_ON(!ptr);
1234
1235 local_irq_disable();
1236 memcpy(ptr, list, sizeof(struct kmem_list3));
1237 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1238 cachep->nodelists[nodeid] = ptr;
1239 local_irq_enable();
1240}
1241
Andrew Mortona737b3e2006-03-22 00:08:11 -08001242/*
1243 * Initialisation. Called after the page allocator have been initialised and
1244 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 */
1246void __init kmem_cache_init(void)
1247{
1248 size_t left_over;
1249 struct cache_sizes *sizes;
1250 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001251 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001252 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001253
1254 for (i = 0; i < NUM_INIT_LISTS; i++) {
1255 kmem_list3_init(&initkmem_list3[i]);
1256 if (i < MAX_NUMNODES)
1257 cache_cache.nodelists[i] = NULL;
1258 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
1260 /*
1261 * Fragmentation resistance on low memory - only use bigger
1262 * page orders on machines with more than 32MB of memory.
1263 */
1264 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1265 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1266
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 /* Bootstrap is tricky, because several objects are allocated
1268 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001269 * 1) initialize the cache_cache cache: it contains the struct
1270 * kmem_cache structures of all caches, except cache_cache itself:
1271 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001272 * Initially an __init data area is used for the head array and the
1273 * kmem_list3 structures, it's replaced with a kmalloc allocated
1274 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001276 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001277 * An __init data area is used for the head array.
1278 * 3) Create the remaining kmalloc caches, with minimally sized
1279 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 * 4) Replace the __init data head arrays for cache_cache and the first
1281 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001282 * 5) Replace the __init data for kmem_list3 for cache_cache and
1283 * the other cache's with kmalloc allocated memory.
1284 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 */
1286
1287 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 INIT_LIST_HEAD(&cache_chain);
1289 list_add(&cache_cache.next, &cache_chain);
1290 cache_cache.colour_off = cache_line_size();
1291 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001292 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293
Andrew Mortona737b3e2006-03-22 00:08:11 -08001294 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1295 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296
Jack Steiner07ed76b2006-03-07 21:55:46 -08001297 for (order = 0; order < MAX_ORDER; order++) {
1298 cache_estimate(order, cache_cache.buffer_size,
1299 cache_line_size(), 0, &left_over, &cache_cache.num);
1300 if (cache_cache.num)
1301 break;
1302 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001303 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001304 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001305 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001306 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1307 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308
1309 /* 2+3) create the kmalloc caches */
1310 sizes = malloc_sizes;
1311 names = cache_names;
1312
Andrew Mortona737b3e2006-03-22 00:08:11 -08001313 /*
1314 * Initialize the caches that provide memory for the array cache and the
1315 * kmem_list3 structures first. Without this, further allocations will
1316 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001317 */
1318
1319 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001320 sizes[INDEX_AC].cs_size,
1321 ARCH_KMALLOC_MINALIGN,
1322 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1323 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001324
Andrew Mortona737b3e2006-03-22 00:08:11 -08001325 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001326 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001327 kmem_cache_create(names[INDEX_L3].name,
1328 sizes[INDEX_L3].cs_size,
1329 ARCH_KMALLOC_MINALIGN,
1330 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1331 NULL, NULL);
1332 }
Christoph Lametere498be72005-09-09 13:03:32 -07001333
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001335 /*
1336 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 * This should be particularly beneficial on SMP boxes, as it
1338 * eliminates "false sharing".
1339 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001340 * allow tighter packing of the smaller caches.
1341 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001342 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001343 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001344 sizes->cs_size,
1345 ARCH_KMALLOC_MINALIGN,
1346 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1347 NULL, NULL);
1348 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349
1350 /* Inc off-slab bufctl limit until the ceiling is hit. */
1351 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001352 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 offslab_limit /= sizeof(kmem_bufctl_t);
1354 }
1355
1356 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001357 sizes->cs_size,
1358 ARCH_KMALLOC_MINALIGN,
1359 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1360 SLAB_PANIC,
1361 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 sizes++;
1363 names++;
1364 }
1365 /* 4) Replace the bootstrap head arrays */
1366 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001367 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001370
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001372 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1373 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001374 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 cache_cache.array[smp_processor_id()] = ptr;
1376 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001377
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001379
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001381 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001382 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001383 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001384 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001385 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001386 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 local_irq_enable();
1388 }
Christoph Lametere498be72005-09-09 13:03:32 -07001389 /* 5) Replace the bootstrap kmem_list3's */
1390 {
1391 int node;
1392 /* Replace the static kmem_list3 structures for the boot cpu */
1393 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001394 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395
Christoph Lametere498be72005-09-09 13:03:32 -07001396 for_each_online_node(node) {
1397 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001398 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001399
1400 if (INDEX_AC != INDEX_L3) {
1401 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001402 &initkmem_list3[SIZE_L3 + node],
1403 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001404 }
1405 }
1406 }
1407
1408 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001410 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001411 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001413 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001414 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 }
1416
1417 /* Done! */
1418 g_cpucache_up = FULL;
1419
Andrew Mortona737b3e2006-03-22 00:08:11 -08001420 /*
1421 * Register a cpu startup notifier callback that initializes
1422 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 */
1424 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
Andrew Mortona737b3e2006-03-22 00:08:11 -08001426 /*
1427 * The reap timers are started later, with a module init call: That part
1428 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429 */
1430}
1431
1432static int __init cpucache_init(void)
1433{
1434 int cpu;
1435
Andrew Mortona737b3e2006-03-22 00:08:11 -08001436 /*
1437 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 */
Christoph Lametere498be72005-09-09 13:03:32 -07001439 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001440 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 return 0;
1442}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443__initcall(cpucache_init);
1444
1445/*
1446 * Interface to system's page allocator. No need to hold the cache-lock.
1447 *
1448 * If we requested dmaable memory, we will get it. Even if we
1449 * did not request dmaable memory, we might get it, but that
1450 * would be relatively rare and ignorable.
1451 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001452static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453{
1454 struct page *page;
1455 void *addr;
1456 int i;
1457
1458 flags |= cachep->gfpflags;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001459#ifndef CONFIG_MMU
1460 /* nommu uses slab's for process anonymous memory allocations, so
1461 * requires __GFP_COMP to properly refcount higher order allocations"
1462 */
1463 page = alloc_pages_node(nodeid, (flags | __GFP_COMP), cachep->gfporder);
1464#else
Christoph Lameter50c85a12005-11-13 16:06:47 -08001465 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Luke Yangd6fef9d2006-04-10 22:52:56 -07001466#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 if (!page)
1468 return NULL;
1469 addr = page_address(page);
1470
1471 i = (1 << cachep->gfporder);
1472 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1473 atomic_add(i, &slab_reclaim_pages);
1474 add_page_state(nr_slab, i);
1475 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001476 __SetPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 page++;
1478 }
1479 return addr;
1480}
1481
1482/*
1483 * Interface to system's page release.
1484 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001485static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001487 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 struct page *page = virt_to_page(addr);
1489 const unsigned long nr_freed = i;
1490
1491 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001492 BUG_ON(!PageSlab(page));
1493 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 page++;
1495 }
1496 sub_page_state(nr_slab, nr_freed);
1497 if (current->reclaim_state)
1498 current->reclaim_state->reclaimed_slab += nr_freed;
1499 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001500 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1501 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502}
1503
1504static void kmem_rcu_free(struct rcu_head *head)
1505{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001506 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001507 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
1509 kmem_freepages(cachep, slab_rcu->addr);
1510 if (OFF_SLAB(cachep))
1511 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1512}
1513
1514#if DEBUG
1515
1516#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001517static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001518 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001520 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001522 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001524 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 return;
1526
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001527 *addr++ = 0x12345678;
1528 *addr++ = caller;
1529 *addr++ = smp_processor_id();
1530 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 {
1532 unsigned long *sptr = &caller;
1533 unsigned long svalue;
1534
1535 while (!kstack_end(sptr)) {
1536 svalue = *sptr++;
1537 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001538 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 size -= sizeof(unsigned long);
1540 if (size <= sizeof(unsigned long))
1541 break;
1542 }
1543 }
1544
1545 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001546 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547}
1548#endif
1549
Pekka Enberg343e0d72006-02-01 03:05:50 -08001550static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001552 int size = obj_size(cachep);
1553 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554
1555 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001556 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557}
1558
1559static void dump_line(char *data, int offset, int limit)
1560{
1561 int i;
1562 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001563 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001564 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 printk("\n");
1566}
1567#endif
1568
1569#if DEBUG
1570
Pekka Enberg343e0d72006-02-01 03:05:50 -08001571static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572{
1573 int i, size;
1574 char *realobj;
1575
1576 if (cachep->flags & SLAB_RED_ZONE) {
1577 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001578 *dbg_redzone1(cachep, objp),
1579 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 }
1581
1582 if (cachep->flags & SLAB_STORE_USER) {
1583 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001584 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001586 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 printk("\n");
1588 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001589 realobj = (char *)objp + obj_offset(cachep);
1590 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001591 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 int limit;
1593 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001594 if (i + limit > size)
1595 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 dump_line(realobj, i, limit);
1597 }
1598}
1599
Pekka Enberg343e0d72006-02-01 03:05:50 -08001600static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601{
1602 char *realobj;
1603 int size, i;
1604 int lines = 0;
1605
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001606 realobj = (char *)objp + obj_offset(cachep);
1607 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001609 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001611 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 exp = POISON_END;
1613 if (realobj[i] != exp) {
1614 int limit;
1615 /* Mismatch ! */
1616 /* Print header */
1617 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001618 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001619 "Slab corruption: start=%p, len=%d\n",
1620 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 print_objinfo(cachep, objp, 0);
1622 }
1623 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001624 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001626 if (i + limit > size)
1627 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 dump_line(realobj, i, limit);
1629 i += 16;
1630 lines++;
1631 /* Limit to 5 lines */
1632 if (lines > 5)
1633 break;
1634 }
1635 }
1636 if (lines != 0) {
1637 /* Print some data about the neighboring objects, if they
1638 * exist:
1639 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001640 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001641 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001643 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001645 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001646 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001648 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 print_objinfo(cachep, objp, 2);
1650 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001651 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001652 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001653 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001655 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 print_objinfo(cachep, objp, 2);
1657 }
1658 }
1659}
1660#endif
1661
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001663/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001664 * slab_destroy_objs - destroy a slab and its objects
1665 * @cachep: cache pointer being destroyed
1666 * @slabp: slab pointer being destroyed
1667 *
1668 * Call the registered destructor for each object in a slab that is being
1669 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001670 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001671static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001672{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 int i;
1674 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001675 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676
1677 if (cachep->flags & SLAB_POISON) {
1678#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001679 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1680 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001681 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001682 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 else
1684 check_poison_obj(cachep, objp);
1685#else
1686 check_poison_obj(cachep, objp);
1687#endif
1688 }
1689 if (cachep->flags & SLAB_RED_ZONE) {
1690 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1691 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001692 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1694 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001695 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 }
1697 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001698 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001700}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001702static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001703{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 if (cachep->dtor) {
1705 int i;
1706 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001707 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001708 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 }
1710 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001711}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712#endif
1713
Randy Dunlap911851e2006-03-22 00:08:14 -08001714/**
1715 * slab_destroy - destroy and release all objects in a slab
1716 * @cachep: cache pointer being destroyed
1717 * @slabp: slab pointer being destroyed
1718 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001719 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001720 * Before calling the slab must have been unlinked from the cache. The
1721 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001722 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001723static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001724{
1725 void *addr = slabp->s_mem - slabp->colouroff;
1726
1727 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1729 struct slab_rcu *slab_rcu;
1730
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001731 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 slab_rcu->cachep = cachep;
1733 slab_rcu->addr = addr;
1734 call_rcu(&slab_rcu->head, kmem_rcu_free);
1735 } else {
1736 kmem_freepages(cachep, addr);
1737 if (OFF_SLAB(cachep))
1738 kmem_cache_free(cachep->slabp_cache, slabp);
1739 }
1740}
1741
Andrew Mortona737b3e2006-03-22 00:08:11 -08001742/*
1743 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1744 * size of kmem_list3.
1745 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001746static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001747{
1748 int node;
1749
1750 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001751 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001752 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001753 REAPTIMEOUT_LIST3 +
1754 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001755 }
1756}
1757
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001759 * calculate_slab_order - calculate size (page order) of slabs
1760 * @cachep: pointer to the cache that is being created
1761 * @size: size of objects to be created in this cache.
1762 * @align: required alignment for the objects.
1763 * @flags: slab allocation flags
1764 *
1765 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001766 *
1767 * This could be made much more intelligent. For now, try to avoid using
1768 * high order pages for slabs. When the gfp() functions are more friendly
1769 * towards high-order requests, this should be changed.
1770 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001771static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001772 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001773{
1774 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001775 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001776
Andrew Mortona737b3e2006-03-22 00:08:11 -08001777 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001778 unsigned int num;
1779 size_t remainder;
1780
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001781 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001782 if (!num)
1783 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001784
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001785 /* More than offslab_limit objects will cause problems */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001786 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001787 break;
1788
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001789 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001790 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001791 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001792 left_over = remainder;
1793
1794 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001795 * A VFS-reclaimable slab tends to have most allocations
1796 * as GFP_NOFS and we really don't want to have to be allocating
1797 * higher-order pages when we are unable to shrink dcache.
1798 */
1799 if (flags & SLAB_RECLAIM_ACCOUNT)
1800 break;
1801
1802 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001803 * Large number of objects is good, but very large slabs are
1804 * currently bad for the gfp()s.
1805 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001806 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001807 break;
1808
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001809 /*
1810 * Acceptable internal fragmentation?
1811 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001812 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001813 break;
1814 }
1815 return left_over;
1816}
1817
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001818static void setup_cpu_cache(struct kmem_cache *cachep)
1819{
1820 if (g_cpucache_up == FULL) {
1821 enable_cpucache(cachep);
1822 return;
1823 }
1824 if (g_cpucache_up == NONE) {
1825 /*
1826 * Note: the first kmem_cache_create must create the cache
1827 * that's used by kmalloc(24), otherwise the creation of
1828 * further caches will BUG().
1829 */
1830 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1831
1832 /*
1833 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1834 * the first cache, then we need to set up all its list3s,
1835 * otherwise the creation of further caches will BUG().
1836 */
1837 set_up_list3s(cachep, SIZE_AC);
1838 if (INDEX_AC == INDEX_L3)
1839 g_cpucache_up = PARTIAL_L3;
1840 else
1841 g_cpucache_up = PARTIAL_AC;
1842 } else {
1843 cachep->array[smp_processor_id()] =
1844 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1845
1846 if (g_cpucache_up == PARTIAL_AC) {
1847 set_up_list3s(cachep, SIZE_L3);
1848 g_cpucache_up = PARTIAL_L3;
1849 } else {
1850 int node;
1851 for_each_online_node(node) {
1852 cachep->nodelists[node] =
1853 kmalloc_node(sizeof(struct kmem_list3),
1854 GFP_KERNEL, node);
1855 BUG_ON(!cachep->nodelists[node]);
1856 kmem_list3_init(cachep->nodelists[node]);
1857 }
1858 }
1859 }
1860 cachep->nodelists[numa_node_id()]->next_reap =
1861 jiffies + REAPTIMEOUT_LIST3 +
1862 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1863
1864 cpu_cache_get(cachep)->avail = 0;
1865 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1866 cpu_cache_get(cachep)->batchcount = 1;
1867 cpu_cache_get(cachep)->touched = 0;
1868 cachep->batchcount = 1;
1869 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1870}
1871
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001872/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 * kmem_cache_create - Create a cache.
1874 * @name: A string which is used in /proc/slabinfo to identify this cache.
1875 * @size: The size of objects to be created in this cache.
1876 * @align: The required alignment for the objects.
1877 * @flags: SLAB flags
1878 * @ctor: A constructor for the objects.
1879 * @dtor: A destructor for the objects.
1880 *
1881 * Returns a ptr to the cache on success, NULL on failure.
1882 * Cannot be called within a int, but can be interrupted.
1883 * The @ctor is run when new pages are allocated by the cache
1884 * and the @dtor is run before the pages are handed back.
1885 *
1886 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001887 * the module calling this has to destroy the cache before getting unloaded.
1888 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 * The flags are
1890 *
1891 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1892 * to catch references to uninitialised memory.
1893 *
1894 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1895 * for buffer overruns.
1896 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1898 * cacheline. This can be beneficial if you're counting cycles as closely
1899 * as davem.
1900 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001901struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001903 unsigned long flags,
1904 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001905 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906{
1907 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001908 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001909 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910
1911 /*
1912 * Sanity checks... these are all serious usage bugs.
1913 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001914 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001915 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001916 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1917 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001918 BUG();
1919 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001921 /*
1922 * Prevent CPUs from coming and going.
1923 * lock_cpu_hotplug() nests outside cache_chain_mutex
1924 */
1925 lock_cpu_hotplug();
1926
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001927 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001928
1929 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001930 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001931 mm_segment_t old_fs = get_fs();
1932 char tmp;
1933 int res;
1934
1935 /*
1936 * This happens when the module gets unloaded and doesn't
1937 * destroy its slab cache and no-one else reuses the vmalloc
1938 * area of the module. Print a warning.
1939 */
1940 set_fs(KERNEL_DS);
1941 res = __get_user(tmp, pc->name);
1942 set_fs(old_fs);
1943 if (res) {
1944 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001945 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001946 continue;
1947 }
1948
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001949 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001950 printk("kmem_cache_create: duplicate cache %s\n", name);
1951 dump_stack();
1952 goto oops;
1953 }
1954 }
1955
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956#if DEBUG
1957 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1958 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1959 /* No constructor, but inital state check requested */
1960 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001961 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 flags &= ~SLAB_DEBUG_INITIAL;
1963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964#if FORCED_DEBUG
1965 /*
1966 * Enable redzoning and last user accounting, except for caches with
1967 * large objects, if the increased size would increase the object size
1968 * above the next power of two: caches with object sizes just above a
1969 * power of two have a significant amount of internal fragmentation.
1970 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001971 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001972 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 if (!(flags & SLAB_DESTROY_BY_RCU))
1974 flags |= SLAB_POISON;
1975#endif
1976 if (flags & SLAB_DESTROY_BY_RCU)
1977 BUG_ON(flags & SLAB_POISON);
1978#endif
1979 if (flags & SLAB_DESTROY_BY_RCU)
1980 BUG_ON(dtor);
1981
1982 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001983 * Always checks flags, a caller might be expecting debug support which
1984 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001986 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987
Andrew Mortona737b3e2006-03-22 00:08:11 -08001988 /*
1989 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 * unaligned accesses for some archs when redzoning is used, and makes
1991 * sure any on-slab bufctl's are also correctly aligned.
1992 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001993 if (size & (BYTES_PER_WORD - 1)) {
1994 size += (BYTES_PER_WORD - 1);
1995 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 }
1997
Andrew Mortona737b3e2006-03-22 00:08:11 -08001998 /* calculate the final buffer alignment: */
1999
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000 /* 1) arch recommendation: can be overridden for debug */
2001 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002002 /*
2003 * Default alignment: as specified by the arch code. Except if
2004 * an object is really small, then squeeze multiple objects into
2005 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 */
2007 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002008 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 ralign /= 2;
2010 } else {
2011 ralign = BYTES_PER_WORD;
2012 }
2013 /* 2) arch mandated alignment: disables debug if necessary */
2014 if (ralign < ARCH_SLAB_MINALIGN) {
2015 ralign = ARCH_SLAB_MINALIGN;
2016 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002017 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 }
2019 /* 3) caller mandated alignment: disables debug if necessary */
2020 if (ralign < align) {
2021 ralign = align;
2022 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002023 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002025 /*
2026 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 * the alignment to BYTES_PER_WORD.
2028 */
2029 align = ralign;
2030
2031 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002032 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002034 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035
2036#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002037 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
2039 if (flags & SLAB_RED_ZONE) {
2040 /* redzoning only works with word aligned caches */
2041 align = BYTES_PER_WORD;
2042
2043 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002044 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002045 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 }
2047 if (flags & SLAB_STORE_USER) {
2048 /* user store requires word alignment and
2049 * one word storage behind the end of the real
2050 * object.
2051 */
2052 align = BYTES_PER_WORD;
2053 size += BYTES_PER_WORD;
2054 }
2055#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002056 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002057 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2058 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 size = PAGE_SIZE;
2060 }
2061#endif
2062#endif
2063
2064 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002065 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002066 /*
2067 * Size is large, assume best to place the slab management obj
2068 * off-slab (should allow better packing of objs).
2069 */
2070 flags |= CFLGS_OFF_SLAB;
2071
2072 size = ALIGN(size, align);
2073
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002074 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075
2076 if (!cachep->num) {
2077 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2078 kmem_cache_free(&cache_cache, cachep);
2079 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002080 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002082 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2083 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084
2085 /*
2086 * If the slab has been placed off-slab, and we have enough space then
2087 * move it on-slab. This is at the expense of any extra colouring.
2088 */
2089 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2090 flags &= ~CFLGS_OFF_SLAB;
2091 left_over -= slab_size;
2092 }
2093
2094 if (flags & CFLGS_OFF_SLAB) {
2095 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002096 slab_size =
2097 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 }
2099
2100 cachep->colour_off = cache_line_size();
2101 /* Offset must be a multiple of the alignment. */
2102 if (cachep->colour_off < align)
2103 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002104 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 cachep->slab_size = slab_size;
2106 cachep->flags = flags;
2107 cachep->gfpflags = 0;
2108 if (flags & SLAB_CACHE_DMA)
2109 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002110 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111
2112 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002113 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 cachep->ctor = ctor;
2115 cachep->dtor = dtor;
2116 cachep->name = name;
2117
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002119 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 /* cache setup completed, link it into the list */
2122 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002123oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 if (!cachep && (flags & SLAB_PANIC))
2125 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002126 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002127 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002128 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 return cachep;
2130}
2131EXPORT_SYMBOL(kmem_cache_create);
2132
2133#if DEBUG
2134static void check_irq_off(void)
2135{
2136 BUG_ON(!irqs_disabled());
2137}
2138
2139static void check_irq_on(void)
2140{
2141 BUG_ON(irqs_disabled());
2142}
2143
Pekka Enberg343e0d72006-02-01 03:05:50 -08002144static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145{
2146#ifdef CONFIG_SMP
2147 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002148 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149#endif
2150}
Christoph Lametere498be72005-09-09 13:03:32 -07002151
Pekka Enberg343e0d72006-02-01 03:05:50 -08002152static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002153{
2154#ifdef CONFIG_SMP
2155 check_irq_off();
2156 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2157#endif
2158}
2159
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160#else
2161#define check_irq_off() do { } while(0)
2162#define check_irq_on() do { } while(0)
2163#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002164#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165#endif
2166
Christoph Lameteraab22072006-03-22 00:09:06 -08002167static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2168 struct array_cache *ac,
2169 int force, int node);
2170
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171static void do_drain(void *arg)
2172{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002173 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002175 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176
2177 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002178 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002179 spin_lock(&cachep->nodelists[node]->list_lock);
2180 free_block(cachep, ac->entry, ac->avail, node);
2181 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 ac->avail = 0;
2183}
2184
Pekka Enberg343e0d72006-02-01 03:05:50 -08002185static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186{
Christoph Lametere498be72005-09-09 13:03:32 -07002187 struct kmem_list3 *l3;
2188 int node;
2189
Andrew Mortona07fa392006-03-22 00:08:17 -08002190 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002192 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002193 l3 = cachep->nodelists[node];
2194 if (l3) {
Christoph Lameteraab22072006-03-22 00:09:06 -08002195 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002196 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002197 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002198 }
2199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200}
2201
Pekka Enberg343e0d72006-02-01 03:05:50 -08002202static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203{
2204 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002205 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 int ret;
2207
Christoph Lametere498be72005-09-09 13:03:32 -07002208 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 struct list_head *p;
2210
Christoph Lametere498be72005-09-09 13:03:32 -07002211 p = l3->slabs_free.prev;
2212 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 break;
2214
Christoph Lametere498be72005-09-09 13:03:32 -07002215 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002217 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218#endif
2219 list_del(&slabp->list);
2220
Christoph Lametere498be72005-09-09 13:03:32 -07002221 l3->free_objects -= cachep->num;
2222 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002224 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002226 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 return ret;
2228}
2229
Pekka Enberg343e0d72006-02-01 03:05:50 -08002230static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002231{
2232 int ret = 0, i = 0;
2233 struct kmem_list3 *l3;
2234
2235 drain_cpu_caches(cachep);
2236
2237 check_irq_on();
2238 for_each_online_node(i) {
2239 l3 = cachep->nodelists[i];
2240 if (l3) {
2241 spin_lock_irq(&l3->list_lock);
2242 ret += __node_shrink(cachep, i);
2243 spin_unlock_irq(&l3->list_lock);
2244 }
2245 }
2246 return (ret ? 1 : 0);
2247}
2248
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249/**
2250 * kmem_cache_shrink - Shrink a cache.
2251 * @cachep: The cache to shrink.
2252 *
2253 * Releases as many slabs as possible for a cache.
2254 * To help debugging, a zero exit status indicates all slabs were released.
2255 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002256int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002258 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
2260 return __cache_shrink(cachep);
2261}
2262EXPORT_SYMBOL(kmem_cache_shrink);
2263
2264/**
2265 * kmem_cache_destroy - delete a cache
2266 * @cachep: the cache to destroy
2267 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002268 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 * Returns 0 on success.
2270 *
2271 * It is expected this function will be called by a module when it is
2272 * unloaded. This will remove the cache completely, and avoid a duplicate
2273 * cache being allocated each time a module is loaded and unloaded, if the
2274 * module doesn't have persistent in-kernel storage across loads and unloads.
2275 *
2276 * The cache must be empty before calling this function.
2277 *
2278 * The caller must guarantee that noone will allocate memory from the cache
2279 * during the kmem_cache_destroy().
2280 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002281int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282{
2283 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002284 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002286 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
2288 /* Don't let CPUs to come and go */
2289 lock_cpu_hotplug();
2290
2291 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002292 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 /*
2294 * the chain is never empty, cache_cache is never destroyed
2295 */
2296 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002297 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298
2299 if (__cache_shrink(cachep)) {
2300 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002301 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002302 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002303 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 unlock_cpu_hotplug();
2305 return 1;
2306 }
2307
2308 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002309 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310
Christoph Lametere498be72005-09-09 13:03:32 -07002311 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002312 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313
2314 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002315 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002316 l3 = cachep->nodelists[i];
2317 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002318 kfree(l3->shared);
2319 free_alien_cache(l3->alien);
2320 kfree(l3);
2321 }
2322 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 return 0;
2326}
2327EXPORT_SYMBOL(kmem_cache_destroy);
2328
2329/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002330static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002331 int colour_off, gfp_t local_flags,
2332 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333{
2334 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002335
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 if (OFF_SLAB(cachep)) {
2337 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002338 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2339 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 if (!slabp)
2341 return NULL;
2342 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002343 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 colour_off += cachep->slab_size;
2345 }
2346 slabp->inuse = 0;
2347 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002348 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002349 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 return slabp;
2351}
2352
2353static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2354{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002355 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356}
2357
Pekka Enberg343e0d72006-02-01 03:05:50 -08002358static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002359 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360{
2361 int i;
2362
2363 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002364 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365#if DEBUG
2366 /* need to poison the objs? */
2367 if (cachep->flags & SLAB_POISON)
2368 poison_obj(cachep, objp, POISON_FREE);
2369 if (cachep->flags & SLAB_STORE_USER)
2370 *dbg_userword(cachep, objp) = NULL;
2371
2372 if (cachep->flags & SLAB_RED_ZONE) {
2373 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2374 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2375 }
2376 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002377 * Constructors are not allowed to allocate memory from the same
2378 * cache which they are a constructor for. Otherwise, deadlock.
2379 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 */
2381 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002382 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002383 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384
2385 if (cachep->flags & SLAB_RED_ZONE) {
2386 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2387 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002388 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2390 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002391 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002393 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2394 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002395 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002396 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397#else
2398 if (cachep->ctor)
2399 cachep->ctor(objp, cachep, ctor_flags);
2400#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002401 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002403 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404 slabp->free = 0;
2405}
2406
Pekka Enberg343e0d72006-02-01 03:05:50 -08002407static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002409 if (flags & SLAB_DMA)
2410 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2411 else
2412 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413}
2414
Andrew Mortona737b3e2006-03-22 00:08:11 -08002415static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2416 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002417{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002418 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002419 kmem_bufctl_t next;
2420
2421 slabp->inuse++;
2422 next = slab_bufctl(slabp)[slabp->free];
2423#if DEBUG
2424 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2425 WARN_ON(slabp->nodeid != nodeid);
2426#endif
2427 slabp->free = next;
2428
2429 return objp;
2430}
2431
Andrew Mortona737b3e2006-03-22 00:08:11 -08002432static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2433 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002434{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002435 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002436
2437#if DEBUG
2438 /* Verify that the slab belongs to the intended node */
2439 WARN_ON(slabp->nodeid != nodeid);
2440
Al Viro871751e2006-03-25 03:06:39 -08002441 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002442 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002443 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002444 BUG();
2445 }
2446#endif
2447 slab_bufctl(slabp)[objnr] = slabp->free;
2448 slabp->free = objnr;
2449 slabp->inuse--;
2450}
2451
Andrew Mortona737b3e2006-03-22 00:08:11 -08002452static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2453 void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454{
2455 int i;
2456 struct page *page;
2457
2458 /* Nasty!!!!!! I hope this is OK. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 page = virt_to_page(objp);
Nick Piggin84097512006-03-22 00:08:34 -08002460
2461 i = 1;
2462 if (likely(!PageCompound(page)))
2463 i <<= cachep->gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002465 page_set_cache(page, cachep);
2466 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 page++;
2468 } while (--i);
2469}
2470
2471/*
2472 * Grow (by 1) the number of slabs within a cache. This is called by
2473 * kmem_cache_alloc() when there are no active objs left in a cache.
2474 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002475static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002477 struct slab *slabp;
2478 void *objp;
2479 size_t offset;
2480 gfp_t local_flags;
2481 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002482 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483
Andrew Mortona737b3e2006-03-22 00:08:11 -08002484 /*
2485 * Be lazy and only check for valid flags here, keeping it out of the
2486 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002488 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 if (flags & SLAB_NO_GROW)
2490 return 0;
2491
2492 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2493 local_flags = (flags & SLAB_LEVEL_MASK);
2494 if (!(local_flags & __GFP_WAIT))
2495 /*
2496 * Not allowed to sleep. Need to tell a constructor about
2497 * this - it might need to know...
2498 */
2499 ctor_flags |= SLAB_CTOR_ATOMIC;
2500
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002501 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002503 l3 = cachep->nodelists[nodeid];
2504 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505
2506 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002507 offset = l3->colour_next;
2508 l3->colour_next++;
2509 if (l3->colour_next >= cachep->colour)
2510 l3->colour_next = 0;
2511 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002513 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514
2515 if (local_flags & __GFP_WAIT)
2516 local_irq_enable();
2517
2518 /*
2519 * The test for missing atomic flag is performed here, rather than
2520 * the more obvious place, simply to reduce the critical path length
2521 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2522 * will eventually be caught here (where it matters).
2523 */
2524 kmem_flagcheck(cachep, flags);
2525
Andrew Mortona737b3e2006-03-22 00:08:11 -08002526 /*
2527 * Get mem for the objs. Attempt to allocate a physical page from
2528 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002529 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002530 objp = kmem_getpages(cachep, flags, nodeid);
2531 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 goto failed;
2533
2534 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002535 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002536 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 goto opps1;
2538
Christoph Lametere498be72005-09-09 13:03:32 -07002539 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 set_slab_attr(cachep, slabp, objp);
2541
2542 cache_init_objs(cachep, slabp, ctor_flags);
2543
2544 if (local_flags & __GFP_WAIT)
2545 local_irq_disable();
2546 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002547 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548
2549 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002550 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002552 l3->free_objects += cachep->num;
2553 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002555opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002557failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 if (local_flags & __GFP_WAIT)
2559 local_irq_disable();
2560 return 0;
2561}
2562
2563#if DEBUG
2564
2565/*
2566 * Perform extra freeing checks:
2567 * - detect bad pointers.
2568 * - POISON/RED_ZONE checking
2569 * - destructor calls, for caches with POISON+dtor
2570 */
2571static void kfree_debugcheck(const void *objp)
2572{
2573 struct page *page;
2574
2575 if (!virt_addr_valid(objp)) {
2576 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002577 (unsigned long)objp);
2578 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 }
2580 page = virt_to_page(objp);
2581 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002582 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2583 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 BUG();
2585 }
2586}
2587
Pekka Enberg343e0d72006-02-01 03:05:50 -08002588static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002589 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590{
2591 struct page *page;
2592 unsigned int objnr;
2593 struct slab *slabp;
2594
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002595 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 kfree_debugcheck(objp);
2597 page = virt_to_page(objp);
2598
Pekka Enberg065d41c2005-11-13 16:06:46 -08002599 if (page_get_cache(page) != cachep) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002600 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2601 "cache %p, got %p\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002602 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002604 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2605 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 WARN_ON(1);
2607 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002608 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609
2610 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002611 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2612 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2613 slab_error(cachep, "double free, or memory outside"
2614 " object was overwritten");
2615 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2616 "redzone 2:0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002617 objp, *dbg_redzone1(cachep, objp),
2618 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 }
2620 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2621 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2622 }
2623 if (cachep->flags & SLAB_STORE_USER)
2624 *dbg_userword(cachep, objp) = caller;
2625
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002626 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627
2628 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002629 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630
2631 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002632 /*
2633 * Need to call the slab's constructor so the caller can
2634 * perform a verify of its state (debugging). Called without
2635 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002637 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639 }
2640 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2641 /* we want to cache poison the object,
2642 * call the destruction callback
2643 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002644 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 }
Al Viro871751e2006-03-25 03:06:39 -08002646#ifdef CONFIG_DEBUG_SLAB_LEAK
2647 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2648#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 if (cachep->flags & SLAB_POISON) {
2650#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002651 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002653 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002654 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 } else {
2656 poison_obj(cachep, objp, POISON_FREE);
2657 }
2658#else
2659 poison_obj(cachep, objp, POISON_FREE);
2660#endif
2661 }
2662 return objp;
2663}
2664
Pekka Enberg343e0d72006-02-01 03:05:50 -08002665static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666{
2667 kmem_bufctl_t i;
2668 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002669
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 /* Check slab's freelist to see if this obj is there. */
2671 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2672 entries++;
2673 if (entries > cachep->num || i >= cachep->num)
2674 goto bad;
2675 }
2676 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002677bad:
2678 printk(KERN_ERR "slab: Internal list corruption detected in "
2679 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2680 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002681 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002682 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002683 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002684 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002686 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 }
2688 printk("\n");
2689 BUG();
2690 }
2691}
2692#else
2693#define kfree_debugcheck(x) do { } while(0)
2694#define cache_free_debugcheck(x,objp,z) (objp)
2695#define check_slabp(x,y) do { } while(0)
2696#endif
2697
Pekka Enberg343e0d72006-02-01 03:05:50 -08002698static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699{
2700 int batchcount;
2701 struct kmem_list3 *l3;
2702 struct array_cache *ac;
2703
2704 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002705 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002706retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 batchcount = ac->batchcount;
2708 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002709 /*
2710 * If there was little recent activity on this cache, then
2711 * perform only a partial refill. Otherwise we could generate
2712 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 */
2714 batchcount = BATCHREFILL_LIMIT;
2715 }
Christoph Lametere498be72005-09-09 13:03:32 -07002716 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717
Christoph Lametere498be72005-09-09 13:03:32 -07002718 BUG_ON(ac->avail > 0 || !l3);
2719 spin_lock(&l3->list_lock);
2720
Christoph Lameter3ded1752006-03-25 03:06:44 -08002721 /* See if we can refill from the shared array */
2722 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2723 goto alloc_done;
2724
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 while (batchcount > 0) {
2726 struct list_head *entry;
2727 struct slab *slabp;
2728 /* Get slab alloc is to come from. */
2729 entry = l3->slabs_partial.next;
2730 if (entry == &l3->slabs_partial) {
2731 l3->free_touched = 1;
2732 entry = l3->slabs_free.next;
2733 if (entry == &l3->slabs_free)
2734 goto must_grow;
2735 }
2736
2737 slabp = list_entry(entry, struct slab, list);
2738 check_slabp(cachep, slabp);
2739 check_spinlock_acquired(cachep);
2740 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741 STATS_INC_ALLOCED(cachep);
2742 STATS_INC_ACTIVE(cachep);
2743 STATS_SET_HIGH(cachep);
2744
Matthew Dobson78d382d2006-02-01 03:05:47 -08002745 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2746 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 }
2748 check_slabp(cachep, slabp);
2749
2750 /* move slabp to correct slabp list: */
2751 list_del(&slabp->list);
2752 if (slabp->free == BUFCTL_END)
2753 list_add(&slabp->list, &l3->slabs_full);
2754 else
2755 list_add(&slabp->list, &l3->slabs_partial);
2756 }
2757
Andrew Mortona737b3e2006-03-22 00:08:11 -08002758must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002760alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002761 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762
2763 if (unlikely(!ac->avail)) {
2764 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002765 x = cache_grow(cachep, flags, numa_node_id());
2766
Andrew Mortona737b3e2006-03-22 00:08:11 -08002767 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002768 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002769 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 return NULL;
2771
Andrew Mortona737b3e2006-03-22 00:08:11 -08002772 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 goto retry;
2774 }
2775 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002776 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777}
2778
Andrew Mortona737b3e2006-03-22 00:08:11 -08002779static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2780 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781{
2782 might_sleep_if(flags & __GFP_WAIT);
2783#if DEBUG
2784 kmem_flagcheck(cachep, flags);
2785#endif
2786}
2787
2788#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002789static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2790 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002792 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002794 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002796 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002797 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002798 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 else
2800 check_poison_obj(cachep, objp);
2801#else
2802 check_poison_obj(cachep, objp);
2803#endif
2804 poison_obj(cachep, objp, POISON_INUSE);
2805 }
2806 if (cachep->flags & SLAB_STORE_USER)
2807 *dbg_userword(cachep, objp) = caller;
2808
2809 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002810 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2811 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2812 slab_error(cachep, "double free, or memory outside"
2813 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002814 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002815 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2816 objp, *dbg_redzone1(cachep, objp),
2817 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 }
2819 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2820 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2821 }
Al Viro871751e2006-03-25 03:06:39 -08002822#ifdef CONFIG_DEBUG_SLAB_LEAK
2823 {
2824 struct slab *slabp;
2825 unsigned objnr;
2826
2827 slabp = page_get_slab(virt_to_page(objp));
2828 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2829 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2830 }
2831#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002832 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002834 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835
2836 if (!(flags & __GFP_WAIT))
2837 ctor_flags |= SLAB_CTOR_ATOMIC;
2838
2839 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002840 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 return objp;
2842}
2843#else
2844#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2845#endif
2846
Pekka Enberg343e0d72006-02-01 03:05:50 -08002847static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002849 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 struct array_cache *ac;
2851
Christoph Lameterdc85da12006-01-18 17:42:36 -08002852#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002853 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002854 objp = alternate_node_alloc(cachep, flags);
2855 if (objp != NULL)
2856 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08002857 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08002858#endif
2859
Alok N Kataria5c382302005-09-27 21:45:46 -07002860 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002861 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 if (likely(ac->avail)) {
2863 STATS_INC_ALLOCHIT(cachep);
2864 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002865 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 } else {
2867 STATS_INC_ALLOCMISS(cachep);
2868 objp = cache_alloc_refill(cachep, flags);
2869 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002870 return objp;
2871}
2872
Andrew Mortona737b3e2006-03-22 00:08:11 -08002873static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2874 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002875{
2876 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002877 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002878
2879 cache_alloc_debugcheck_before(cachep, flags);
2880
2881 local_irq_save(save_flags);
2882 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002884 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002885 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002886 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 return objp;
2888}
2889
Christoph Lametere498be72005-09-09 13:03:32 -07002890#ifdef CONFIG_NUMA
2891/*
Paul Jacksonb2455392006-03-24 03:16:12 -08002892 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002893 *
2894 * If we are in_interrupt, then process context, including cpusets and
2895 * mempolicy, may not apply and should not be used for allocation policy.
2896 */
2897static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2898{
2899 int nid_alloc, nid_here;
2900
2901 if (in_interrupt())
2902 return NULL;
2903 nid_alloc = nid_here = numa_node_id();
2904 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2905 nid_alloc = cpuset_mem_spread_node();
2906 else if (current->mempolicy)
2907 nid_alloc = slab_node(current->mempolicy);
2908 if (nid_alloc != nid_here)
2909 return __cache_alloc_node(cachep, flags, nid_alloc);
2910 return NULL;
2911}
2912
2913/*
Christoph Lametere498be72005-09-09 13:03:32 -07002914 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002916static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2917 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002918{
2919 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002920 struct slab *slabp;
2921 struct kmem_list3 *l3;
2922 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002923 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002925 l3 = cachep->nodelists[nodeid];
2926 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002927
Andrew Mortona737b3e2006-03-22 00:08:11 -08002928retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002929 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002930 spin_lock(&l3->list_lock);
2931 entry = l3->slabs_partial.next;
2932 if (entry == &l3->slabs_partial) {
2933 l3->free_touched = 1;
2934 entry = l3->slabs_free.next;
2935 if (entry == &l3->slabs_free)
2936 goto must_grow;
2937 }
Christoph Lametere498be72005-09-09 13:03:32 -07002938
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002939 slabp = list_entry(entry, struct slab, list);
2940 check_spinlock_acquired_node(cachep, nodeid);
2941 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002942
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002943 STATS_INC_NODEALLOCS(cachep);
2944 STATS_INC_ACTIVE(cachep);
2945 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002946
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002947 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002948
Matthew Dobson78d382d2006-02-01 03:05:47 -08002949 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002950 check_slabp(cachep, slabp);
2951 l3->free_objects--;
2952 /* move slabp to correct slabp list: */
2953 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002954
Andrew Mortona737b3e2006-03-22 00:08:11 -08002955 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002956 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002957 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002958 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002959
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002960 spin_unlock(&l3->list_lock);
2961 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002962
Andrew Mortona737b3e2006-03-22 00:08:11 -08002963must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002964 spin_unlock(&l3->list_lock);
2965 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002966
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002967 if (!x)
2968 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002969
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002970 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002971done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002972 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002973}
2974#endif
2975
2976/*
2977 * Caller needs to acquire correct kmem_list's list_lock
2978 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002979static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002980 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981{
2982 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002983 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984
2985 for (i = 0; i < nr_objects; i++) {
2986 void *objp = objpp[i];
2987 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002989 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002990 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002992 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002994 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002996 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997 check_slabp(cachep, slabp);
2998
2999 /* fixup slab chains */
3000 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003001 if (l3->free_objects > l3->free_limit) {
3002 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 slab_destroy(cachep, slabp);
3004 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003005 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 }
3007 } else {
3008 /* Unconditionally move a slab to the end of the
3009 * partial list on free - maximum time for the
3010 * other objects to be freed, too.
3011 */
Christoph Lametere498be72005-09-09 13:03:32 -07003012 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 }
3014 }
3015}
3016
Pekka Enberg343e0d72006-02-01 03:05:50 -08003017static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018{
3019 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003020 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003021 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022
3023 batchcount = ac->batchcount;
3024#if DEBUG
3025 BUG_ON(!batchcount || batchcount > ac->avail);
3026#endif
3027 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003028 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07003029 spin_lock(&l3->list_lock);
3030 if (l3->shared) {
3031 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003032 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 if (max) {
3034 if (batchcount > max)
3035 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003036 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003037 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 shared_array->avail += batchcount;
3039 goto free_done;
3040 }
3041 }
3042
Christoph Lameterff694162005-09-22 21:44:02 -07003043 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003044free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045#if STATS
3046 {
3047 int i = 0;
3048 struct list_head *p;
3049
Christoph Lametere498be72005-09-09 13:03:32 -07003050 p = l3->slabs_free.next;
3051 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052 struct slab *slabp;
3053
3054 slabp = list_entry(p, struct slab, list);
3055 BUG_ON(slabp->inuse);
3056
3057 i++;
3058 p = p->next;
3059 }
3060 STATS_SET_FREEABLE(cachep, i);
3061 }
3062#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003063 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003065 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066}
3067
3068/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003069 * Release an obj back to its cache. If the obj has a constructed state, it must
3070 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003072static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003074 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075
3076 check_irq_off();
3077 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3078
Christoph Lametere498be72005-09-09 13:03:32 -07003079 /* Make sure we are not freeing a object from another
3080 * node to the array cache on this cpu.
3081 */
3082#ifdef CONFIG_NUMA
3083 {
3084 struct slab *slabp;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003085 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07003086 if (unlikely(slabp->nodeid != numa_node_id())) {
3087 struct array_cache *alien = NULL;
3088 int nodeid = slabp->nodeid;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003089 struct kmem_list3 *l3;
Christoph Lametere498be72005-09-09 13:03:32 -07003090
Andrew Mortona737b3e2006-03-22 00:08:11 -08003091 l3 = cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07003092 STATS_INC_NODEFREES(cachep);
3093 if (l3->alien && l3->alien[nodeid]) {
3094 alien = l3->alien[nodeid];
3095 spin_lock(&alien->lock);
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003096 if (unlikely(alien->avail == alien->limit)) {
3097 STATS_INC_ACOVERFLOW(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003098 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003099 alien, nodeid);
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003100 }
Christoph Lametere498be72005-09-09 13:03:32 -07003101 alien->entry[alien->avail++] = objp;
3102 spin_unlock(&alien->lock);
3103 } else {
3104 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003105 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003106 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003107 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003108 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003109 }
3110 return;
3111 }
3112 }
3113#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114 if (likely(ac->avail < ac->limit)) {
3115 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003116 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 return;
3118 } else {
3119 STATS_INC_FREEMISS(cachep);
3120 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003121 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 }
3123}
3124
3125/**
3126 * kmem_cache_alloc - Allocate an object
3127 * @cachep: The cache to allocate from.
3128 * @flags: See kmalloc().
3129 *
3130 * Allocate an object from this cache. The flags are only relevant
3131 * if the cache has no available objects.
3132 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003133void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003135 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136}
3137EXPORT_SYMBOL(kmem_cache_alloc);
3138
3139/**
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003140 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3141 * @cache: The cache to allocate from.
3142 * @flags: See kmalloc().
3143 *
3144 * Allocate an object from this cache and set the allocated memory to zero.
3145 * The flags are only relevant if the cache has no available objects.
3146 */
3147void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3148{
3149 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3150 if (ret)
3151 memset(ret, 0, obj_size(cache));
3152 return ret;
3153}
3154EXPORT_SYMBOL(kmem_cache_zalloc);
3155
3156/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 * kmem_ptr_validate - check if an untrusted pointer might
3158 * be a slab entry.
3159 * @cachep: the cache we're checking against
3160 * @ptr: pointer to validate
3161 *
3162 * This verifies that the untrusted pointer looks sane:
3163 * it is _not_ a guarantee that the pointer is actually
3164 * part of the slab cache in question, but it at least
3165 * validates that the pointer can be dereferenced and
3166 * looks half-way sane.
3167 *
3168 * Currently only used for dentry validation.
3169 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003170int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003172 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003174 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003175 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 struct page *page;
3177
3178 if (unlikely(addr < min_addr))
3179 goto out;
3180 if (unlikely(addr > (unsigned long)high_memory - size))
3181 goto out;
3182 if (unlikely(addr & align_mask))
3183 goto out;
3184 if (unlikely(!kern_addr_valid(addr)))
3185 goto out;
3186 if (unlikely(!kern_addr_valid(addr + size - 1)))
3187 goto out;
3188 page = virt_to_page(ptr);
3189 if (unlikely(!PageSlab(page)))
3190 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003191 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192 goto out;
3193 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003194out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195 return 0;
3196}
3197
3198#ifdef CONFIG_NUMA
3199/**
3200 * kmem_cache_alloc_node - Allocate an object on the specified node
3201 * @cachep: The cache to allocate from.
3202 * @flags: See kmalloc().
3203 * @nodeid: node number of the target node.
3204 *
3205 * Identical to kmem_cache_alloc, except that this function is slow
3206 * and can sleep. And it will allocate memory on the given node, which
3207 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003208 * New and improved: it will now make sure that the object gets
3209 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003211void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212{
Christoph Lametere498be72005-09-09 13:03:32 -07003213 unsigned long save_flags;
3214 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215
Christoph Lametere498be72005-09-09 13:03:32 -07003216 cache_alloc_debugcheck_before(cachep, flags);
3217 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003218
3219 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003220 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003221 ptr = ____cache_alloc(cachep, flags);
3222 else
3223 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003224 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003225
3226 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3227 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228
Christoph Lametere498be72005-09-09 13:03:32 -07003229 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230}
3231EXPORT_SYMBOL(kmem_cache_alloc_node);
3232
Al Virodd0fc662005-10-07 07:46:04 +01003233void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003234{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003235 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003236
3237 cachep = kmem_find_general_cachep(size, flags);
3238 if (unlikely(cachep == NULL))
3239 return NULL;
3240 return kmem_cache_alloc_node(cachep, flags, node);
3241}
3242EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243#endif
3244
3245/**
3246 * kmalloc - allocate memory
3247 * @size: how many bytes of memory are required.
3248 * @flags: the type of memory to allocate.
Randy Dunlap911851e2006-03-22 00:08:14 -08003249 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250 *
3251 * kmalloc is the normal method of allocating memory
3252 * in the kernel.
3253 *
3254 * The @flags argument may be one of:
3255 *
3256 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3257 *
3258 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3259 *
3260 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3261 *
3262 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3263 * must be suitable for DMA. This can mean different things on different
3264 * platforms. For example, on i386, it means that the memory must come
3265 * from the first 16MB.
3266 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003267static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3268 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003270 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003272 /* If you want to save a few bytes .text space: replace
3273 * __ with kmem_.
3274 * Then kmalloc uses the uninlined functions instead of the inline
3275 * functions.
3276 */
3277 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003278 if (unlikely(cachep == NULL))
3279 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003280 return __cache_alloc(cachep, flags, caller);
3281}
3282
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003283
3284void *__kmalloc(size_t size, gfp_t flags)
3285{
Al Viro871751e2006-03-25 03:06:39 -08003286#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003287 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003288#else
3289 return __do_kmalloc(size, flags, __builtin_return_address(0));
3290#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291}
3292EXPORT_SYMBOL(__kmalloc);
3293
Al Viro871751e2006-03-25 03:06:39 -08003294#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003295void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3296{
3297 return __do_kmalloc(size, flags, caller);
3298}
3299EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003300#endif
3301
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302#ifdef CONFIG_SMP
3303/**
3304 * __alloc_percpu - allocate one copy of the object for every present
3305 * cpu in the system, zeroing them.
3306 * Objects should be dereferenced using the per_cpu_ptr macro only.
3307 *
3308 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003309 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003310void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311{
3312 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003313 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314
3315 if (!pdata)
3316 return NULL;
3317
Christoph Lametere498be72005-09-09 13:03:32 -07003318 /*
3319 * Cannot use for_each_online_cpu since a cpu may come online
3320 * and we have no way of figuring out how to fix the array
3321 * that we have allocated then....
3322 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003323 for_each_possible_cpu(i) {
Christoph Lametere498be72005-09-09 13:03:32 -07003324 int node = cpu_to_node(i);
3325
3326 if (node_online(node))
3327 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3328 else
3329 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330
3331 if (!pdata->ptrs[i])
3332 goto unwind_oom;
3333 memset(pdata->ptrs[i], 0, size);
3334 }
3335
3336 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003337 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338
Andrew Mortona737b3e2006-03-22 00:08:11 -08003339unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340 while (--i >= 0) {
3341 if (!cpu_possible(i))
3342 continue;
3343 kfree(pdata->ptrs[i]);
3344 }
3345 kfree(pdata);
3346 return NULL;
3347}
3348EXPORT_SYMBOL(__alloc_percpu);
3349#endif
3350
3351/**
3352 * kmem_cache_free - Deallocate an object
3353 * @cachep: The cache the allocation was from.
3354 * @objp: The previously allocated object.
3355 *
3356 * Free an object which was previously allocated from this
3357 * cache.
3358 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003359void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360{
3361 unsigned long flags;
3362
3363 local_irq_save(flags);
3364 __cache_free(cachep, objp);
3365 local_irq_restore(flags);
3366}
3367EXPORT_SYMBOL(kmem_cache_free);
3368
3369/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370 * kfree - free previously allocated memory
3371 * @objp: pointer returned by kmalloc.
3372 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003373 * If @objp is NULL, no operation is performed.
3374 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 * Don't free memory not originally allocated by kmalloc()
3376 * or you will run into trouble.
3377 */
3378void kfree(const void *objp)
3379{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003380 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 unsigned long flags;
3382
3383 if (unlikely(!objp))
3384 return;
3385 local_irq_save(flags);
3386 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003387 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003388 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003389 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 local_irq_restore(flags);
3391}
3392EXPORT_SYMBOL(kfree);
3393
3394#ifdef CONFIG_SMP
3395/**
3396 * free_percpu - free previously allocated percpu memory
3397 * @objp: pointer returned by alloc_percpu.
3398 *
3399 * Don't free memory not originally allocated by alloc_percpu()
3400 * The complemented objp is to check for that.
3401 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003402void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403{
3404 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003405 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406
Christoph Lametere498be72005-09-09 13:03:32 -07003407 /*
3408 * We allocate for all cpus so we cannot use for online cpu here.
3409 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003410 for_each_possible_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003411 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412 kfree(p);
3413}
3414EXPORT_SYMBOL(free_percpu);
3415#endif
3416
Pekka Enberg343e0d72006-02-01 03:05:50 -08003417unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003419 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420}
3421EXPORT_SYMBOL(kmem_cache_size);
3422
Pekka Enberg343e0d72006-02-01 03:05:50 -08003423const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003424{
3425 return cachep->name;
3426}
3427EXPORT_SYMBOL_GPL(kmem_cache_name);
3428
Christoph Lametere498be72005-09-09 13:03:32 -07003429/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003430 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003431 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003432static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003433{
3434 int node;
3435 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003436 struct array_cache *new_shared;
3437 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003438
3439 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003440
Andrew Mortona737b3e2006-03-22 00:08:11 -08003441 new_alien = alloc_alien_cache(node, cachep->limit);
3442 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003443 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003444
Christoph Lameter0718dc22006-03-25 03:06:47 -08003445 new_shared = alloc_arraycache(node,
3446 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003447 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003448 if (!new_shared) {
3449 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003450 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003451 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003452
Andrew Mortona737b3e2006-03-22 00:08:11 -08003453 l3 = cachep->nodelists[node];
3454 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003455 struct array_cache *shared = l3->shared;
3456
Christoph Lametere498be72005-09-09 13:03:32 -07003457 spin_lock_irq(&l3->list_lock);
3458
Christoph Lametercafeb022006-03-25 03:06:46 -08003459 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003460 free_block(cachep, shared->entry,
3461 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003462
Christoph Lametercafeb022006-03-25 03:06:46 -08003463 l3->shared = new_shared;
3464 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003465 l3->alien = new_alien;
3466 new_alien = NULL;
3467 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003468 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003469 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003470 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003471 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003472 free_alien_cache(new_alien);
3473 continue;
3474 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003475 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003476 if (!l3) {
3477 free_alien_cache(new_alien);
3478 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003479 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003480 }
Christoph Lametere498be72005-09-09 13:03:32 -07003481
3482 kmem_list3_init(l3);
3483 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003484 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003485 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003486 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003487 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003488 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003489 cachep->nodelists[node] = l3;
3490 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003491 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003492
Andrew Mortona737b3e2006-03-22 00:08:11 -08003493fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003494 if (!cachep->next.next) {
3495 /* Cache is not active yet. Roll back what we did */
3496 node--;
3497 while (node >= 0) {
3498 if (cachep->nodelists[node]) {
3499 l3 = cachep->nodelists[node];
3500
3501 kfree(l3->shared);
3502 free_alien_cache(l3->alien);
3503 kfree(l3);
3504 cachep->nodelists[node] = NULL;
3505 }
3506 node--;
3507 }
3508 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003509 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003510}
3511
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003513 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 struct array_cache *new[NR_CPUS];
3515};
3516
3517static void do_ccupdate_local(void *info)
3518{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003519 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 struct array_cache *old;
3521
3522 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003523 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003524
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3526 new->new[smp_processor_id()] = old;
3527}
3528
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003529/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003530static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3531 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532{
3533 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003534 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003536 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003537 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003538 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3539 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003540 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003541 for (i--; i >= 0; i--)
3542 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003543 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 }
3545 }
3546 new.cachep = cachep;
3547
Andrew Mortona07fa392006-03-22 00:08:17 -08003548 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003549
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 cachep->batchcount = batchcount;
3552 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003553 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554
Christoph Lametere498be72005-09-09 13:03:32 -07003555 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556 struct array_cache *ccold = new.new[i];
3557 if (!ccold)
3558 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003559 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003560 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003561 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 kfree(ccold);
3563 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564
Christoph Lametere498be72005-09-09 13:03:32 -07003565 err = alloc_kmemlist(cachep);
3566 if (err) {
3567 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003568 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003569 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 return 0;
3572}
3573
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003574/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003575static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576{
3577 int err;
3578 int limit, shared;
3579
Andrew Mortona737b3e2006-03-22 00:08:11 -08003580 /*
3581 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582 * - create a LIFO ordering, i.e. return objects that are cache-warm
3583 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003584 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 * bufctl chains: array operations are cheaper.
3586 * The numbers are guessed, we should auto-tune as described by
3587 * Bonwick.
3588 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003589 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003591 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003593 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003595 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596 limit = 54;
3597 else
3598 limit = 120;
3599
Andrew Mortona737b3e2006-03-22 00:08:11 -08003600 /*
3601 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602 * allocation behaviour: Most allocs on one cpu, most free operations
3603 * on another cpu. For these cases, an efficient object passing between
3604 * cpus is necessary. This is provided by a shared array. The array
3605 * replaces Bonwick's magazine layer.
3606 * On uniprocessor, it's functionally equivalent (but less efficient)
3607 * to a larger limit. Thus disabled by default.
3608 */
3609 shared = 0;
3610#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003611 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612 shared = 8;
3613#endif
3614
3615#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003616 /*
3617 * With debugging enabled, large batchcount lead to excessively long
3618 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 */
3620 if (limit > 32)
3621 limit = 32;
3622#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003623 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 if (err)
3625 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003626 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627}
3628
Christoph Lameter1b552532006-03-22 00:09:07 -08003629/*
3630 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003631 * necessary. Note that the l3 listlock also protects the array_cache
3632 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003633 */
3634void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3635 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636{
3637 int tofree;
3638
Christoph Lameter1b552532006-03-22 00:09:07 -08003639 if (!ac || !ac->avail)
3640 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641 if (ac->touched && !force) {
3642 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003643 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003644 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003645 if (ac->avail) {
3646 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3647 if (tofree > ac->avail)
3648 tofree = (ac->avail + 1) / 2;
3649 free_block(cachep, ac->entry, tofree, node);
3650 ac->avail -= tofree;
3651 memmove(ac->entry, &(ac->entry[tofree]),
3652 sizeof(void *) * ac->avail);
3653 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003654 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655 }
3656}
3657
3658/**
3659 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003660 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003661 *
3662 * Called from workqueue/eventd every few seconds.
3663 * Purpose:
3664 * - clear the per-cpu caches for this CPU.
3665 * - return freeable pages to the main free memory pool.
3666 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003667 * If we cannot acquire the cache chain mutex then just give up - we'll try
3668 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 */
3670static void cache_reap(void *unused)
3671{
3672 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003673 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003674 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003676 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003678 schedule_delayed_work(&__get_cpu_var(reap_work),
3679 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 return;
3681 }
3682
3683 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003684 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003685 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 int tofree;
3687 struct slab *slabp;
3688
Pekka Enberg343e0d72006-02-01 03:05:50 -08003689 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690 check_irq_on();
3691
Christoph Lameter35386e32006-03-22 00:09:05 -08003692 /*
3693 * We only take the l3 lock if absolutely necessary and we
3694 * have established with reasonable certainty that
3695 * we can do some work if the lock was obtained.
3696 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003697 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003698
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003699 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700
Christoph Lameteraab22072006-03-22 00:09:06 -08003701 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702
Christoph Lameter35386e32006-03-22 00:09:05 -08003703 /*
3704 * These are racy checks but it does not matter
3705 * if we skip one check or scan twice.
3706 */
Christoph Lametere498be72005-09-09 13:03:32 -07003707 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003708 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709
Christoph Lametere498be72005-09-09 13:03:32 -07003710 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
Christoph Lameteraab22072006-03-22 00:09:06 -08003712 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713
Christoph Lametere498be72005-09-09 13:03:32 -07003714 if (l3->free_touched) {
3715 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003716 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 }
3718
Andrew Mortona737b3e2006-03-22 00:08:11 -08003719 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3720 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003722 /*
3723 * Do not lock if there are no free blocks.
3724 */
3725 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 break;
3727
Christoph Lameter35386e32006-03-22 00:09:05 -08003728 spin_lock_irq(&l3->list_lock);
3729 p = l3->slabs_free.next;
3730 if (p == &(l3->slabs_free)) {
3731 spin_unlock_irq(&l3->list_lock);
3732 break;
3733 }
3734
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 slabp = list_entry(p, struct slab, list);
3736 BUG_ON(slabp->inuse);
3737 list_del(&slabp->list);
3738 STATS_INC_REAPED(searchp);
3739
Andrew Mortona737b3e2006-03-22 00:08:11 -08003740 /*
3741 * Safe to drop the lock. The slab is no longer linked
3742 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743 * cache_chain_lock
3744 */
Christoph Lametere498be72005-09-09 13:03:32 -07003745 l3->free_objects -= searchp->num;
3746 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003748 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003749next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750 cond_resched();
3751 }
3752 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003753 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003754 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003755 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003756 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757}
3758
3759#ifdef CONFIG_PROC_FS
3760
Pekka Enberg85289f92006-01-08 01:00:36 -08003761static void print_slabinfo_header(struct seq_file *m)
3762{
3763 /*
3764 * Output format version, so at least we can change it
3765 * without _too_ many complaints.
3766 */
3767#if STATS
3768 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3769#else
3770 seq_puts(m, "slabinfo - version: 2.1\n");
3771#endif
3772 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3773 "<objperslab> <pagesperslab>");
3774 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3775 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3776#if STATS
3777 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003778 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003779 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3780#endif
3781 seq_putc(m, '\n');
3782}
3783
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784static void *s_start(struct seq_file *m, loff_t *pos)
3785{
3786 loff_t n = *pos;
3787 struct list_head *p;
3788
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003789 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003790 if (!n)
3791 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 p = cache_chain.next;
3793 while (n--) {
3794 p = p->next;
3795 if (p == &cache_chain)
3796 return NULL;
3797 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003798 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799}
3800
3801static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3802{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003803 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003805 return cachep->next.next == &cache_chain ?
3806 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807}
3808
3809static void s_stop(struct seq_file *m, void *p)
3810{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003811 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003812}
3813
3814static int s_show(struct seq_file *m, void *p)
3815{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003816 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003818 struct slab *slabp;
3819 unsigned long active_objs;
3820 unsigned long num_objs;
3821 unsigned long active_slabs = 0;
3822 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003823 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003825 int node;
3826 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828 active_objs = 0;
3829 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003830 for_each_online_node(node) {
3831 l3 = cachep->nodelists[node];
3832 if (!l3)
3833 continue;
3834
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003835 check_irq_on();
3836 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003837
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003838 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003839 slabp = list_entry(q, struct slab, list);
3840 if (slabp->inuse != cachep->num && !error)
3841 error = "slabs_full accounting error";
3842 active_objs += cachep->num;
3843 active_slabs++;
3844 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003845 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003846 slabp = list_entry(q, struct slab, list);
3847 if (slabp->inuse == cachep->num && !error)
3848 error = "slabs_partial inuse accounting error";
3849 if (!slabp->inuse && !error)
3850 error = "slabs_partial/inuse accounting error";
3851 active_objs += slabp->inuse;
3852 active_slabs++;
3853 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003854 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003855 slabp = list_entry(q, struct slab, list);
3856 if (slabp->inuse && !error)
3857 error = "slabs_free/inuse accounting error";
3858 num_slabs++;
3859 }
3860 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003861 if (l3->shared)
3862 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003863
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003864 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003866 num_slabs += active_slabs;
3867 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003868 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003869 error = "free_objects accounting error";
3870
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003871 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872 if (error)
3873 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3874
3875 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003876 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003877 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003879 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003880 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003881 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003883 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 unsigned long high = cachep->high_mark;
3885 unsigned long allocs = cachep->num_allocations;
3886 unsigned long grown = cachep->grown;
3887 unsigned long reaped = cachep->reaped;
3888 unsigned long errors = cachep->errors;
3889 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003891 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003892 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893
Christoph Lametere498be72005-09-09 13:03:32 -07003894 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003895 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003896 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003897 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 }
3899 /* cpu stats */
3900 {
3901 unsigned long allochit = atomic_read(&cachep->allochit);
3902 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3903 unsigned long freehit = atomic_read(&cachep->freehit);
3904 unsigned long freemiss = atomic_read(&cachep->freemiss);
3905
3906 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003907 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 }
3909#endif
3910 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 return 0;
3912}
3913
3914/*
3915 * slabinfo_op - iterator that generates /proc/slabinfo
3916 *
3917 * Output layout:
3918 * cache-name
3919 * num-active-objs
3920 * total-objs
3921 * object size
3922 * num-active-slabs
3923 * total-slabs
3924 * num-pages-per-slab
3925 * + further values on SMP and with statistics enabled
3926 */
3927
3928struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003929 .start = s_start,
3930 .next = s_next,
3931 .stop = s_stop,
3932 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933};
3934
3935#define MAX_SLABINFO_WRITE 128
3936/**
3937 * slabinfo_write - Tuning for the slab allocator
3938 * @file: unused
3939 * @buffer: user buffer
3940 * @count: data length
3941 * @ppos: unused
3942 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003943ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3944 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003946 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 int limit, batchcount, shared, res;
3948 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003949
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 if (count > MAX_SLABINFO_WRITE)
3951 return -EINVAL;
3952 if (copy_from_user(&kbuf, buffer, count))
3953 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003954 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955
3956 tmp = strchr(kbuf, ' ');
3957 if (!tmp)
3958 return -EINVAL;
3959 *tmp = '\0';
3960 tmp++;
3961 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3962 return -EINVAL;
3963
3964 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003965 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003967 list_for_each(p, &cache_chain) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003968 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969
Andrew Mortona737b3e2006-03-22 00:08:11 -08003970 cachep = list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003972 if (limit < 1 || batchcount < 1 ||
3973 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003974 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003976 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003977 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 }
3979 break;
3980 }
3981 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003982 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 if (res >= 0)
3984 res = count;
3985 return res;
3986}
Al Viro871751e2006-03-25 03:06:39 -08003987
3988#ifdef CONFIG_DEBUG_SLAB_LEAK
3989
3990static void *leaks_start(struct seq_file *m, loff_t *pos)
3991{
3992 loff_t n = *pos;
3993 struct list_head *p;
3994
3995 mutex_lock(&cache_chain_mutex);
3996 p = cache_chain.next;
3997 while (n--) {
3998 p = p->next;
3999 if (p == &cache_chain)
4000 return NULL;
4001 }
4002 return list_entry(p, struct kmem_cache, next);
4003}
4004
4005static inline int add_caller(unsigned long *n, unsigned long v)
4006{
4007 unsigned long *p;
4008 int l;
4009 if (!v)
4010 return 1;
4011 l = n[1];
4012 p = n + 2;
4013 while (l) {
4014 int i = l/2;
4015 unsigned long *q = p + 2 * i;
4016 if (*q == v) {
4017 q[1]++;
4018 return 1;
4019 }
4020 if (*q > v) {
4021 l = i;
4022 } else {
4023 p = q + 2;
4024 l -= i + 1;
4025 }
4026 }
4027 if (++n[1] == n[0])
4028 return 0;
4029 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4030 p[0] = v;
4031 p[1] = 1;
4032 return 1;
4033}
4034
4035static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4036{
4037 void *p;
4038 int i;
4039 if (n[0] == n[1])
4040 return;
4041 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4042 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4043 continue;
4044 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4045 return;
4046 }
4047}
4048
4049static void show_symbol(struct seq_file *m, unsigned long address)
4050{
4051#ifdef CONFIG_KALLSYMS
4052 char *modname;
4053 const char *name;
4054 unsigned long offset, size;
4055 char namebuf[KSYM_NAME_LEN+1];
4056
4057 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4058
4059 if (name) {
4060 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4061 if (modname)
4062 seq_printf(m, " [%s]", modname);
4063 return;
4064 }
4065#endif
4066 seq_printf(m, "%p", (void *)address);
4067}
4068
4069static int leaks_show(struct seq_file *m, void *p)
4070{
4071 struct kmem_cache *cachep = p;
4072 struct list_head *q;
4073 struct slab *slabp;
4074 struct kmem_list3 *l3;
4075 const char *name;
4076 unsigned long *n = m->private;
4077 int node;
4078 int i;
4079
4080 if (!(cachep->flags & SLAB_STORE_USER))
4081 return 0;
4082 if (!(cachep->flags & SLAB_RED_ZONE))
4083 return 0;
4084
4085 /* OK, we can do it */
4086
4087 n[1] = 0;
4088
4089 for_each_online_node(node) {
4090 l3 = cachep->nodelists[node];
4091 if (!l3)
4092 continue;
4093
4094 check_irq_on();
4095 spin_lock_irq(&l3->list_lock);
4096
4097 list_for_each(q, &l3->slabs_full) {
4098 slabp = list_entry(q, struct slab, list);
4099 handle_slab(n, cachep, slabp);
4100 }
4101 list_for_each(q, &l3->slabs_partial) {
4102 slabp = list_entry(q, struct slab, list);
4103 handle_slab(n, cachep, slabp);
4104 }
4105 spin_unlock_irq(&l3->list_lock);
4106 }
4107 name = cachep->name;
4108 if (n[0] == n[1]) {
4109 /* Increase the buffer size */
4110 mutex_unlock(&cache_chain_mutex);
4111 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4112 if (!m->private) {
4113 /* Too bad, we are really out */
4114 m->private = n;
4115 mutex_lock(&cache_chain_mutex);
4116 return -ENOMEM;
4117 }
4118 *(unsigned long *)m->private = n[0] * 2;
4119 kfree(n);
4120 mutex_lock(&cache_chain_mutex);
4121 /* Now make sure this entry will be retried */
4122 m->count = m->size;
4123 return 0;
4124 }
4125 for (i = 0; i < n[1]; i++) {
4126 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4127 show_symbol(m, n[2*i+2]);
4128 seq_putc(m, '\n');
4129 }
4130 return 0;
4131}
4132
4133struct seq_operations slabstats_op = {
4134 .start = leaks_start,
4135 .next = s_next,
4136 .stop = s_stop,
4137 .show = leaks_show,
4138};
4139#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140#endif
4141
Manfred Spraul00e145b2005-09-03 15:55:07 -07004142/**
4143 * ksize - get the actual amount of memory allocated for a given object
4144 * @objp: Pointer to the object
4145 *
4146 * kmalloc may internally round up allocations and return more memory
4147 * than requested. ksize() can be used to determine the actual amount of
4148 * memory allocated. The caller may use this additional memory, even though
4149 * a smaller amount of memory was initially specified with the kmalloc call.
4150 * The caller must guarantee that objp points to a valid object previously
4151 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4152 * must not be freed during the duration of the call.
4153 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154unsigned int ksize(const void *objp)
4155{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004156 if (unlikely(objp == NULL))
4157 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004159 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160}