blob: c32af7e7581ef0ff11ef112f94b6a84da50cafc4 [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 */
shin, jacob693f7d32006-04-28 10:54:37 -0500982 if (rl3->shared)
983 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800984
Christoph Lameterff694162005-09-22 21:44:02 -0700985 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700986 ac->avail = 0;
987 spin_unlock(&rl3->list_lock);
988 }
989}
990
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800991/*
992 * Called from cache_reap() to regularly drain alien caches round robin.
993 */
994static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
995{
996 int node = __get_cpu_var(reap_node);
997
998 if (l3->alien) {
999 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001000
1001 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001002 __drain_alien_cache(cachep, ac, node);
1003 spin_unlock_irq(&ac->lock);
1004 }
1005 }
1006}
1007
Andrew Mortona737b3e2006-03-22 00:08:11 -08001008static void drain_alien_cache(struct kmem_cache *cachep,
1009 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001010{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001011 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001012 struct array_cache *ac;
1013 unsigned long flags;
1014
1015 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001016 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001017 if (ac) {
1018 spin_lock_irqsave(&ac->lock, flags);
1019 __drain_alien_cache(cachep, ac, i);
1020 spin_unlock_irqrestore(&ac->lock, flags);
1021 }
1022 }
1023}
1024#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001025
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001026#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001027#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001028
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001029static inline struct array_cache **alloc_alien_cache(int node, int limit)
1030{
1031 return (struct array_cache **) 0x01020304ul;
1032}
1033
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001034static inline void free_alien_cache(struct array_cache **ac_ptr)
1035{
1036}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001037
Christoph Lametere498be72005-09-09 13:03:32 -07001038#endif
1039
Chandra Seetharaman83d722f2006-04-24 19:35:21 -07001040static int cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001041 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042{
1043 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001044 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001045 struct kmem_list3 *l3 = NULL;
1046 int node = cpu_to_node(cpu);
1047 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048
1049 switch (action) {
1050 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001051 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001052 /*
1053 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001054 * alloc_arraycache's are going to use this list.
1055 * kmalloc_node allows us to add the slab to the right
1056 * kmem_list3 and not this cpu's kmem_list3
1057 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058
Christoph Lametere498be72005-09-09 13:03:32 -07001059 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001060 /*
1061 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001062 * begin anything. Make sure some other cpu on this
1063 * node has not already allocated this
1064 */
1065 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001066 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1067 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001068 goto bad;
1069 kmem_list3_init(l3);
1070 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001071 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001072
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001073 /*
1074 * The l3s don't come and go as CPUs come and
1075 * go. cache_chain_mutex is sufficient
1076 * protection here.
1077 */
Christoph Lametere498be72005-09-09 13:03:32 -07001078 cachep->nodelists[node] = l3;
1079 }
1080
1081 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1082 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001083 (1 + nr_cpus_node(node)) *
1084 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001085 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1086 }
1087
Andrew Mortona737b3e2006-03-22 00:08:11 -08001088 /*
1089 * Now we can go ahead with allocating the shared arrays and
1090 * array caches
1091 */
Christoph Lametere498be72005-09-09 13:03:32 -07001092 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001093 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001094 struct array_cache *shared;
1095 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001096
Christoph Lametere498be72005-09-09 13:03:32 -07001097 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001098 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 if (!nc)
1100 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001101 shared = alloc_arraycache(node,
1102 cachep->shared * cachep->batchcount,
1103 0xbaadf00d);
1104 if (!shared)
1105 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001106
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001107 alien = alloc_alien_cache(node, cachep->limit);
1108 if (!alien)
1109 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001111 l3 = cachep->nodelists[node];
1112 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001113
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001114 spin_lock_irq(&l3->list_lock);
1115 if (!l3->shared) {
1116 /*
1117 * We are serialised from CPU_DEAD or
1118 * CPU_UP_CANCELLED by the cpucontrol lock
1119 */
1120 l3->shared = shared;
1121 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001122 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001123#ifdef CONFIG_NUMA
1124 if (!l3->alien) {
1125 l3->alien = alien;
1126 alien = NULL;
1127 }
1128#endif
1129 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001130 kfree(shared);
1131 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001133 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 break;
1135 case CPU_ONLINE:
1136 start_cpu_timer(cpu);
1137 break;
1138#ifdef CONFIG_HOTPLUG_CPU
1139 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001140 /*
1141 * Even if all the cpus of a node are down, we don't free the
1142 * kmem_list3 of any cache. This to avoid a race between
1143 * cpu_down, and a kmalloc allocation from another cpu for
1144 * memory from the node of the cpu going down. The list3
1145 * structure is usually allocated from kmem_cache_create() and
1146 * gets destroyed at kmem_cache_destroy().
1147 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 /* fall thru */
1149 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001150 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 list_for_each_entry(cachep, &cache_chain, next) {
1152 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001153 struct array_cache *shared;
1154 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001155 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156
Christoph Lametere498be72005-09-09 13:03:32 -07001157 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 /* cpu is dead; no one can alloc from it. */
1159 nc = cachep->array[cpu];
1160 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001161 l3 = cachep->nodelists[node];
1162
1163 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001164 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001165
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001166 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001167
1168 /* Free limit for this kmem_list3 */
1169 l3->free_limit -= cachep->batchcount;
1170 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001171 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001172
1173 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001174 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001175 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001176 }
Christoph Lametere498be72005-09-09 13:03:32 -07001177
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001178 shared = l3->shared;
1179 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001180 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001181 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001182 l3->shared = NULL;
1183 }
Christoph Lametere498be72005-09-09 13:03:32 -07001184
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001185 alien = l3->alien;
1186 l3->alien = NULL;
1187
1188 spin_unlock_irq(&l3->list_lock);
1189
1190 kfree(shared);
1191 if (alien) {
1192 drain_alien_cache(cachep, alien);
1193 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001194 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001195free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 kfree(nc);
1197 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001198 /*
1199 * In the previous loop, all the objects were freed to
1200 * the respective cache's slabs, now we can go ahead and
1201 * shrink each nodelist to its limit.
1202 */
1203 list_for_each_entry(cachep, &cache_chain, next) {
1204 l3 = cachep->nodelists[node];
1205 if (!l3)
1206 continue;
1207 spin_lock_irq(&l3->list_lock);
1208 /* free slabs belonging to this node */
1209 __node_shrink(cachep, node);
1210 spin_unlock_irq(&l3->list_lock);
1211 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001212 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 break;
1214#endif
1215 }
1216 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001217bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001218 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 return NOTIFY_BAD;
1220}
1221
1222static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1223
Christoph Lametere498be72005-09-09 13:03:32 -07001224/*
1225 * swap the static kmem_list3 with kmalloced memory
1226 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001227static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1228 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001229{
1230 struct kmem_list3 *ptr;
1231
1232 BUG_ON(cachep->nodelists[nodeid] != list);
1233 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1234 BUG_ON(!ptr);
1235
1236 local_irq_disable();
1237 memcpy(ptr, list, sizeof(struct kmem_list3));
1238 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1239 cachep->nodelists[nodeid] = ptr;
1240 local_irq_enable();
1241}
1242
Andrew Mortona737b3e2006-03-22 00:08:11 -08001243/*
1244 * Initialisation. Called after the page allocator have been initialised and
1245 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 */
1247void __init kmem_cache_init(void)
1248{
1249 size_t left_over;
1250 struct cache_sizes *sizes;
1251 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001252 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001253 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001254
1255 for (i = 0; i < NUM_INIT_LISTS; i++) {
1256 kmem_list3_init(&initkmem_list3[i]);
1257 if (i < MAX_NUMNODES)
1258 cache_cache.nodelists[i] = NULL;
1259 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
1261 /*
1262 * Fragmentation resistance on low memory - only use bigger
1263 * page orders on machines with more than 32MB of memory.
1264 */
1265 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1266 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1267
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 /* Bootstrap is tricky, because several objects are allocated
1269 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001270 * 1) initialize the cache_cache cache: it contains the struct
1271 * kmem_cache structures of all caches, except cache_cache itself:
1272 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001273 * Initially an __init data area is used for the head array and the
1274 * kmem_list3 structures, it's replaced with a kmalloc allocated
1275 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001277 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001278 * An __init data area is used for the head array.
1279 * 3) Create the remaining kmalloc caches, with minimally sized
1280 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 * 4) Replace the __init data head arrays for cache_cache and the first
1282 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001283 * 5) Replace the __init data for kmem_list3 for cache_cache and
1284 * the other cache's with kmalloc allocated memory.
1285 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 */
1287
1288 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 INIT_LIST_HEAD(&cache_chain);
1290 list_add(&cache_cache.next, &cache_chain);
1291 cache_cache.colour_off = cache_line_size();
1292 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001293 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294
Andrew Mortona737b3e2006-03-22 00:08:11 -08001295 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1296 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297
Jack Steiner07ed76b2006-03-07 21:55:46 -08001298 for (order = 0; order < MAX_ORDER; order++) {
1299 cache_estimate(order, cache_cache.buffer_size,
1300 cache_line_size(), 0, &left_over, &cache_cache.num);
1301 if (cache_cache.num)
1302 break;
1303 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001304 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001305 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001306 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001307 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1308 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309
1310 /* 2+3) create the kmalloc caches */
1311 sizes = malloc_sizes;
1312 names = cache_names;
1313
Andrew Mortona737b3e2006-03-22 00:08:11 -08001314 /*
1315 * Initialize the caches that provide memory for the array cache and the
1316 * kmem_list3 structures first. Without this, further allocations will
1317 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001318 */
1319
1320 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001321 sizes[INDEX_AC].cs_size,
1322 ARCH_KMALLOC_MINALIGN,
1323 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1324 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001325
Andrew Mortona737b3e2006-03-22 00:08:11 -08001326 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001327 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001328 kmem_cache_create(names[INDEX_L3].name,
1329 sizes[INDEX_L3].cs_size,
1330 ARCH_KMALLOC_MINALIGN,
1331 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1332 NULL, NULL);
1333 }
Christoph Lametere498be72005-09-09 13:03:32 -07001334
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001336 /*
1337 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 * This should be particularly beneficial on SMP boxes, as it
1339 * eliminates "false sharing".
1340 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001341 * allow tighter packing of the smaller caches.
1342 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001343 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001344 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001345 sizes->cs_size,
1346 ARCH_KMALLOC_MINALIGN,
1347 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1348 NULL, NULL);
1349 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350
1351 /* Inc off-slab bufctl limit until the ceiling is hit. */
1352 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001353 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 offslab_limit /= sizeof(kmem_bufctl_t);
1355 }
1356
1357 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001358 sizes->cs_size,
1359 ARCH_KMALLOC_MINALIGN,
1360 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1361 SLAB_PANIC,
1362 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 sizes++;
1364 names++;
1365 }
1366 /* 4) Replace the bootstrap head arrays */
1367 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001368 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001369
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001371
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001373 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1374 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001375 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 cache_cache.array[smp_processor_id()] = ptr;
1377 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001378
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001380
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001382 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001383 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001384 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001385 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001386 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001387 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 local_irq_enable();
1389 }
Christoph Lametere498be72005-09-09 13:03:32 -07001390 /* 5) Replace the bootstrap kmem_list3's */
1391 {
1392 int node;
1393 /* Replace the static kmem_list3 structures for the boot cpu */
1394 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001395 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
Christoph Lametere498be72005-09-09 13:03:32 -07001397 for_each_online_node(node) {
1398 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001399 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001400
1401 if (INDEX_AC != INDEX_L3) {
1402 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001403 &initkmem_list3[SIZE_L3 + node],
1404 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001405 }
1406 }
1407 }
1408
1409 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001411 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001412 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001414 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001415 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 }
1417
1418 /* Done! */
1419 g_cpucache_up = FULL;
1420
Andrew Mortona737b3e2006-03-22 00:08:11 -08001421 /*
1422 * Register a cpu startup notifier callback that initializes
1423 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 */
1425 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426
Andrew Mortona737b3e2006-03-22 00:08:11 -08001427 /*
1428 * The reap timers are started later, with a module init call: That part
1429 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 */
1431}
1432
1433static int __init cpucache_init(void)
1434{
1435 int cpu;
1436
Andrew Mortona737b3e2006-03-22 00:08:11 -08001437 /*
1438 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 */
Christoph Lametere498be72005-09-09 13:03:32 -07001440 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001441 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 return 0;
1443}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444__initcall(cpucache_init);
1445
1446/*
1447 * Interface to system's page allocator. No need to hold the cache-lock.
1448 *
1449 * If we requested dmaable memory, we will get it. Even if we
1450 * did not request dmaable memory, we might get it, but that
1451 * would be relatively rare and ignorable.
1452 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001453static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454{
1455 struct page *page;
1456 void *addr;
1457 int i;
1458
1459 flags |= cachep->gfpflags;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001460#ifndef CONFIG_MMU
1461 /* nommu uses slab's for process anonymous memory allocations, so
1462 * requires __GFP_COMP to properly refcount higher order allocations"
1463 */
1464 page = alloc_pages_node(nodeid, (flags | __GFP_COMP), cachep->gfporder);
1465#else
Christoph Lameter50c85a12005-11-13 16:06:47 -08001466 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Luke Yangd6fef9d2006-04-10 22:52:56 -07001467#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 if (!page)
1469 return NULL;
1470 addr = page_address(page);
1471
1472 i = (1 << cachep->gfporder);
1473 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1474 atomic_add(i, &slab_reclaim_pages);
1475 add_page_state(nr_slab, i);
1476 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001477 __SetPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 page++;
1479 }
1480 return addr;
1481}
1482
1483/*
1484 * Interface to system's page release.
1485 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001486static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001488 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 struct page *page = virt_to_page(addr);
1490 const unsigned long nr_freed = i;
1491
1492 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001493 BUG_ON(!PageSlab(page));
1494 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 page++;
1496 }
1497 sub_page_state(nr_slab, nr_freed);
1498 if (current->reclaim_state)
1499 current->reclaim_state->reclaimed_slab += nr_freed;
1500 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001501 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1502 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503}
1504
1505static void kmem_rcu_free(struct rcu_head *head)
1506{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001507 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001508 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
1510 kmem_freepages(cachep, slab_rcu->addr);
1511 if (OFF_SLAB(cachep))
1512 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1513}
1514
1515#if DEBUG
1516
1517#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001518static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001519 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001521 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001523 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001525 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 return;
1527
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001528 *addr++ = 0x12345678;
1529 *addr++ = caller;
1530 *addr++ = smp_processor_id();
1531 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 {
1533 unsigned long *sptr = &caller;
1534 unsigned long svalue;
1535
1536 while (!kstack_end(sptr)) {
1537 svalue = *sptr++;
1538 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001539 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 size -= sizeof(unsigned long);
1541 if (size <= sizeof(unsigned long))
1542 break;
1543 }
1544 }
1545
1546 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001547 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548}
1549#endif
1550
Pekka Enberg343e0d72006-02-01 03:05:50 -08001551static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001553 int size = obj_size(cachep);
1554 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555
1556 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001557 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558}
1559
1560static void dump_line(char *data, int offset, int limit)
1561{
1562 int i;
1563 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001564 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001565 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 printk("\n");
1567}
1568#endif
1569
1570#if DEBUG
1571
Pekka Enberg343e0d72006-02-01 03:05:50 -08001572static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573{
1574 int i, size;
1575 char *realobj;
1576
1577 if (cachep->flags & SLAB_RED_ZONE) {
1578 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001579 *dbg_redzone1(cachep, objp),
1580 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581 }
1582
1583 if (cachep->flags & SLAB_STORE_USER) {
1584 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001585 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001587 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 printk("\n");
1589 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001590 realobj = (char *)objp + obj_offset(cachep);
1591 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001592 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 int limit;
1594 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001595 if (i + limit > size)
1596 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 dump_line(realobj, i, limit);
1598 }
1599}
1600
Pekka Enberg343e0d72006-02-01 03:05:50 -08001601static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602{
1603 char *realobj;
1604 int size, i;
1605 int lines = 0;
1606
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001607 realobj = (char *)objp + obj_offset(cachep);
1608 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001610 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001612 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 exp = POISON_END;
1614 if (realobj[i] != exp) {
1615 int limit;
1616 /* Mismatch ! */
1617 /* Print header */
1618 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001619 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001620 "Slab corruption: start=%p, len=%d\n",
1621 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 print_objinfo(cachep, objp, 0);
1623 }
1624 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001625 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001627 if (i + limit > size)
1628 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 dump_line(realobj, i, limit);
1630 i += 16;
1631 lines++;
1632 /* Limit to 5 lines */
1633 if (lines > 5)
1634 break;
1635 }
1636 }
1637 if (lines != 0) {
1638 /* Print some data about the neighboring objects, if they
1639 * exist:
1640 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001641 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001642 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001644 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001646 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001647 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001649 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 print_objinfo(cachep, objp, 2);
1651 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001652 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001653 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001654 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001656 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 print_objinfo(cachep, objp, 2);
1658 }
1659 }
1660}
1661#endif
1662
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001664/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001665 * slab_destroy_objs - destroy a slab and its objects
1666 * @cachep: cache pointer being destroyed
1667 * @slabp: slab pointer being destroyed
1668 *
1669 * Call the registered destructor for each object in a slab that is being
1670 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001671 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001672static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001673{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 int i;
1675 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001676 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677
1678 if (cachep->flags & SLAB_POISON) {
1679#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001680 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1681 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001682 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001683 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 else
1685 check_poison_obj(cachep, objp);
1686#else
1687 check_poison_obj(cachep, objp);
1688#endif
1689 }
1690 if (cachep->flags & SLAB_RED_ZONE) {
1691 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1692 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001693 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1695 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001696 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 }
1698 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001699 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001701}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001703static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001704{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 if (cachep->dtor) {
1706 int i;
1707 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001708 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001709 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 }
1711 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001712}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713#endif
1714
Randy Dunlap911851e2006-03-22 00:08:14 -08001715/**
1716 * slab_destroy - destroy and release all objects in a slab
1717 * @cachep: cache pointer being destroyed
1718 * @slabp: slab pointer being destroyed
1719 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001720 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001721 * Before calling the slab must have been unlinked from the cache. The
1722 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001723 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001724static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001725{
1726 void *addr = slabp->s_mem - slabp->colouroff;
1727
1728 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1730 struct slab_rcu *slab_rcu;
1731
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001732 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 slab_rcu->cachep = cachep;
1734 slab_rcu->addr = addr;
1735 call_rcu(&slab_rcu->head, kmem_rcu_free);
1736 } else {
1737 kmem_freepages(cachep, addr);
1738 if (OFF_SLAB(cachep))
1739 kmem_cache_free(cachep->slabp_cache, slabp);
1740 }
1741}
1742
Andrew Mortona737b3e2006-03-22 00:08:11 -08001743/*
1744 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1745 * size of kmem_list3.
1746 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001747static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001748{
1749 int node;
1750
1751 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001752 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001753 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001754 REAPTIMEOUT_LIST3 +
1755 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001756 }
1757}
1758
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001760 * calculate_slab_order - calculate size (page order) of slabs
1761 * @cachep: pointer to the cache that is being created
1762 * @size: size of objects to be created in this cache.
1763 * @align: required alignment for the objects.
1764 * @flags: slab allocation flags
1765 *
1766 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001767 *
1768 * This could be made much more intelligent. For now, try to avoid using
1769 * high order pages for slabs. When the gfp() functions are more friendly
1770 * towards high-order requests, this should be changed.
1771 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001772static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001773 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001774{
1775 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001776 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001777
Andrew Mortona737b3e2006-03-22 00:08:11 -08001778 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001779 unsigned int num;
1780 size_t remainder;
1781
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001782 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001783 if (!num)
1784 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001785
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001786 /* More than offslab_limit objects will cause problems */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001787 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001788 break;
1789
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001790 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001791 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001792 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001793 left_over = remainder;
1794
1795 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001796 * A VFS-reclaimable slab tends to have most allocations
1797 * as GFP_NOFS and we really don't want to have to be allocating
1798 * higher-order pages when we are unable to shrink dcache.
1799 */
1800 if (flags & SLAB_RECLAIM_ACCOUNT)
1801 break;
1802
1803 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001804 * Large number of objects is good, but very large slabs are
1805 * currently bad for the gfp()s.
1806 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001807 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001808 break;
1809
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001810 /*
1811 * Acceptable internal fragmentation?
1812 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001813 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001814 break;
1815 }
1816 return left_over;
1817}
1818
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001819static void setup_cpu_cache(struct kmem_cache *cachep)
1820{
1821 if (g_cpucache_up == FULL) {
1822 enable_cpucache(cachep);
1823 return;
1824 }
1825 if (g_cpucache_up == NONE) {
1826 /*
1827 * Note: the first kmem_cache_create must create the cache
1828 * that's used by kmalloc(24), otherwise the creation of
1829 * further caches will BUG().
1830 */
1831 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1832
1833 /*
1834 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1835 * the first cache, then we need to set up all its list3s,
1836 * otherwise the creation of further caches will BUG().
1837 */
1838 set_up_list3s(cachep, SIZE_AC);
1839 if (INDEX_AC == INDEX_L3)
1840 g_cpucache_up = PARTIAL_L3;
1841 else
1842 g_cpucache_up = PARTIAL_AC;
1843 } else {
1844 cachep->array[smp_processor_id()] =
1845 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1846
1847 if (g_cpucache_up == PARTIAL_AC) {
1848 set_up_list3s(cachep, SIZE_L3);
1849 g_cpucache_up = PARTIAL_L3;
1850 } else {
1851 int node;
1852 for_each_online_node(node) {
1853 cachep->nodelists[node] =
1854 kmalloc_node(sizeof(struct kmem_list3),
1855 GFP_KERNEL, node);
1856 BUG_ON(!cachep->nodelists[node]);
1857 kmem_list3_init(cachep->nodelists[node]);
1858 }
1859 }
1860 }
1861 cachep->nodelists[numa_node_id()]->next_reap =
1862 jiffies + REAPTIMEOUT_LIST3 +
1863 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1864
1865 cpu_cache_get(cachep)->avail = 0;
1866 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1867 cpu_cache_get(cachep)->batchcount = 1;
1868 cpu_cache_get(cachep)->touched = 0;
1869 cachep->batchcount = 1;
1870 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1871}
1872
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001873/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 * kmem_cache_create - Create a cache.
1875 * @name: A string which is used in /proc/slabinfo to identify this cache.
1876 * @size: The size of objects to be created in this cache.
1877 * @align: The required alignment for the objects.
1878 * @flags: SLAB flags
1879 * @ctor: A constructor for the objects.
1880 * @dtor: A destructor for the objects.
1881 *
1882 * Returns a ptr to the cache on success, NULL on failure.
1883 * Cannot be called within a int, but can be interrupted.
1884 * The @ctor is run when new pages are allocated by the cache
1885 * and the @dtor is run before the pages are handed back.
1886 *
1887 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001888 * the module calling this has to destroy the cache before getting unloaded.
1889 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 * The flags are
1891 *
1892 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1893 * to catch references to uninitialised memory.
1894 *
1895 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1896 * for buffer overruns.
1897 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1899 * cacheline. This can be beneficial if you're counting cycles as closely
1900 * as davem.
1901 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001902struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001904 unsigned long flags,
1905 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001906 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907{
1908 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001909 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001910 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
1912 /*
1913 * Sanity checks... these are all serious usage bugs.
1914 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001915 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001916 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001917 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1918 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001919 BUG();
1920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001922 /*
1923 * Prevent CPUs from coming and going.
1924 * lock_cpu_hotplug() nests outside cache_chain_mutex
1925 */
1926 lock_cpu_hotplug();
1927
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001928 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001929
1930 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001931 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001932 mm_segment_t old_fs = get_fs();
1933 char tmp;
1934 int res;
1935
1936 /*
1937 * This happens when the module gets unloaded and doesn't
1938 * destroy its slab cache and no-one else reuses the vmalloc
1939 * area of the module. Print a warning.
1940 */
1941 set_fs(KERNEL_DS);
1942 res = __get_user(tmp, pc->name);
1943 set_fs(old_fs);
1944 if (res) {
1945 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001946 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001947 continue;
1948 }
1949
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001950 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001951 printk("kmem_cache_create: duplicate cache %s\n", name);
1952 dump_stack();
1953 goto oops;
1954 }
1955 }
1956
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957#if DEBUG
1958 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1959 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1960 /* No constructor, but inital state check requested */
1961 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001962 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 flags &= ~SLAB_DEBUG_INITIAL;
1964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965#if FORCED_DEBUG
1966 /*
1967 * Enable redzoning and last user accounting, except for caches with
1968 * large objects, if the increased size would increase the object size
1969 * above the next power of two: caches with object sizes just above a
1970 * power of two have a significant amount of internal fragmentation.
1971 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001972 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001973 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 if (!(flags & SLAB_DESTROY_BY_RCU))
1975 flags |= SLAB_POISON;
1976#endif
1977 if (flags & SLAB_DESTROY_BY_RCU)
1978 BUG_ON(flags & SLAB_POISON);
1979#endif
1980 if (flags & SLAB_DESTROY_BY_RCU)
1981 BUG_ON(dtor);
1982
1983 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001984 * Always checks flags, a caller might be expecting debug support which
1985 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001987 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988
Andrew Mortona737b3e2006-03-22 00:08:11 -08001989 /*
1990 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 * unaligned accesses for some archs when redzoning is used, and makes
1992 * sure any on-slab bufctl's are also correctly aligned.
1993 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001994 if (size & (BYTES_PER_WORD - 1)) {
1995 size += (BYTES_PER_WORD - 1);
1996 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 }
1998
Andrew Mortona737b3e2006-03-22 00:08:11 -08001999 /* calculate the final buffer alignment: */
2000
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 /* 1) arch recommendation: can be overridden for debug */
2002 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002003 /*
2004 * Default alignment: as specified by the arch code. Except if
2005 * an object is really small, then squeeze multiple objects into
2006 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007 */
2008 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002009 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 ralign /= 2;
2011 } else {
2012 ralign = BYTES_PER_WORD;
2013 }
2014 /* 2) arch mandated alignment: disables debug if necessary */
2015 if (ralign < ARCH_SLAB_MINALIGN) {
2016 ralign = ARCH_SLAB_MINALIGN;
2017 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002018 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 }
2020 /* 3) caller mandated alignment: disables debug if necessary */
2021 if (ralign < align) {
2022 ralign = align;
2023 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002024 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002026 /*
2027 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 * the alignment to BYTES_PER_WORD.
2029 */
2030 align = ralign;
2031
2032 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002033 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002035 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036
2037#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002038 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039
2040 if (flags & SLAB_RED_ZONE) {
2041 /* redzoning only works with word aligned caches */
2042 align = BYTES_PER_WORD;
2043
2044 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002045 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002046 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 }
2048 if (flags & SLAB_STORE_USER) {
2049 /* user store requires word alignment and
2050 * one word storage behind the end of the real
2051 * object.
2052 */
2053 align = BYTES_PER_WORD;
2054 size += BYTES_PER_WORD;
2055 }
2056#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002057 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002058 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2059 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 size = PAGE_SIZE;
2061 }
2062#endif
2063#endif
2064
2065 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002066 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 /*
2068 * Size is large, assume best to place the slab management obj
2069 * off-slab (should allow better packing of objs).
2070 */
2071 flags |= CFLGS_OFF_SLAB;
2072
2073 size = ALIGN(size, align);
2074
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002075 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076
2077 if (!cachep->num) {
2078 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2079 kmem_cache_free(&cache_cache, cachep);
2080 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002081 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002083 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2084 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085
2086 /*
2087 * If the slab has been placed off-slab, and we have enough space then
2088 * move it on-slab. This is at the expense of any extra colouring.
2089 */
2090 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2091 flags &= ~CFLGS_OFF_SLAB;
2092 left_over -= slab_size;
2093 }
2094
2095 if (flags & CFLGS_OFF_SLAB) {
2096 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002097 slab_size =
2098 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 }
2100
2101 cachep->colour_off = cache_line_size();
2102 /* Offset must be a multiple of the alignment. */
2103 if (cachep->colour_off < align)
2104 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002105 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 cachep->slab_size = slab_size;
2107 cachep->flags = flags;
2108 cachep->gfpflags = 0;
2109 if (flags & SLAB_CACHE_DMA)
2110 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002111 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112
2113 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002114 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 cachep->ctor = ctor;
2116 cachep->dtor = dtor;
2117 cachep->name = name;
2118
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002120 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 /* cache setup completed, link it into the list */
2123 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002124oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 if (!cachep && (flags & SLAB_PANIC))
2126 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002127 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002128 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002129 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 return cachep;
2131}
2132EXPORT_SYMBOL(kmem_cache_create);
2133
2134#if DEBUG
2135static void check_irq_off(void)
2136{
2137 BUG_ON(!irqs_disabled());
2138}
2139
2140static void check_irq_on(void)
2141{
2142 BUG_ON(irqs_disabled());
2143}
2144
Pekka Enberg343e0d72006-02-01 03:05:50 -08002145static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146{
2147#ifdef CONFIG_SMP
2148 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002149 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150#endif
2151}
Christoph Lametere498be72005-09-09 13:03:32 -07002152
Pekka Enberg343e0d72006-02-01 03:05:50 -08002153static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002154{
2155#ifdef CONFIG_SMP
2156 check_irq_off();
2157 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2158#endif
2159}
2160
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161#else
2162#define check_irq_off() do { } while(0)
2163#define check_irq_on() do { } while(0)
2164#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002165#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166#endif
2167
Christoph Lameteraab22072006-03-22 00:09:06 -08002168static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2169 struct array_cache *ac,
2170 int force, int node);
2171
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172static void do_drain(void *arg)
2173{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002174 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002176 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177
2178 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002179 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002180 spin_lock(&cachep->nodelists[node]->list_lock);
2181 free_block(cachep, ac->entry, ac->avail, node);
2182 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 ac->avail = 0;
2184}
2185
Pekka Enberg343e0d72006-02-01 03:05:50 -08002186static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187{
Christoph Lametere498be72005-09-09 13:03:32 -07002188 struct kmem_list3 *l3;
2189 int node;
2190
Andrew Mortona07fa392006-03-22 00:08:17 -08002191 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002193 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002194 l3 = cachep->nodelists[node];
2195 if (l3) {
Christoph Lameteraab22072006-03-22 00:09:06 -08002196 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002197 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002198 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002199 }
2200 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201}
2202
Pekka Enberg343e0d72006-02-01 03:05:50 -08002203static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204{
2205 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002206 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 int ret;
2208
Christoph Lametere498be72005-09-09 13:03:32 -07002209 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 struct list_head *p;
2211
Christoph Lametere498be72005-09-09 13:03:32 -07002212 p = l3->slabs_free.prev;
2213 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 break;
2215
Christoph Lametere498be72005-09-09 13:03:32 -07002216 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002218 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219#endif
2220 list_del(&slabp->list);
2221
Christoph Lametere498be72005-09-09 13:03:32 -07002222 l3->free_objects -= cachep->num;
2223 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002225 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002227 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 return ret;
2229}
2230
Pekka Enberg343e0d72006-02-01 03:05:50 -08002231static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002232{
2233 int ret = 0, i = 0;
2234 struct kmem_list3 *l3;
2235
2236 drain_cpu_caches(cachep);
2237
2238 check_irq_on();
2239 for_each_online_node(i) {
2240 l3 = cachep->nodelists[i];
2241 if (l3) {
2242 spin_lock_irq(&l3->list_lock);
2243 ret += __node_shrink(cachep, i);
2244 spin_unlock_irq(&l3->list_lock);
2245 }
2246 }
2247 return (ret ? 1 : 0);
2248}
2249
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250/**
2251 * kmem_cache_shrink - Shrink a cache.
2252 * @cachep: The cache to shrink.
2253 *
2254 * Releases as many slabs as possible for a cache.
2255 * To help debugging, a zero exit status indicates all slabs were released.
2256 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002257int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002259 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260
2261 return __cache_shrink(cachep);
2262}
2263EXPORT_SYMBOL(kmem_cache_shrink);
2264
2265/**
2266 * kmem_cache_destroy - delete a cache
2267 * @cachep: the cache to destroy
2268 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002269 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 * Returns 0 on success.
2271 *
2272 * It is expected this function will be called by a module when it is
2273 * unloaded. This will remove the cache completely, and avoid a duplicate
2274 * cache being allocated each time a module is loaded and unloaded, if the
2275 * module doesn't have persistent in-kernel storage across loads and unloads.
2276 *
2277 * The cache must be empty before calling this function.
2278 *
2279 * The caller must guarantee that noone will allocate memory from the cache
2280 * during the kmem_cache_destroy().
2281 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002282int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283{
2284 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002285 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002287 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288
2289 /* Don't let CPUs to come and go */
2290 lock_cpu_hotplug();
2291
2292 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002293 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 /*
2295 * the chain is never empty, cache_cache is never destroyed
2296 */
2297 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002298 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299
2300 if (__cache_shrink(cachep)) {
2301 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002302 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002303 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002304 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 unlock_cpu_hotplug();
2306 return 1;
2307 }
2308
2309 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002310 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311
Christoph Lametere498be72005-09-09 13:03:32 -07002312 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002313 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314
2315 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002316 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002317 l3 = cachep->nodelists[i];
2318 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002319 kfree(l3->shared);
2320 free_alien_cache(l3->alien);
2321 kfree(l3);
2322 }
2323 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 return 0;
2327}
2328EXPORT_SYMBOL(kmem_cache_destroy);
2329
2330/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002331static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002332 int colour_off, gfp_t local_flags,
2333 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334{
2335 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002336
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 if (OFF_SLAB(cachep)) {
2338 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002339 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2340 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 if (!slabp)
2342 return NULL;
2343 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002344 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 colour_off += cachep->slab_size;
2346 }
2347 slabp->inuse = 0;
2348 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002349 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002350 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 return slabp;
2352}
2353
2354static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2355{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002356 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357}
2358
Pekka Enberg343e0d72006-02-01 03:05:50 -08002359static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002360 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361{
2362 int i;
2363
2364 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002365 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366#if DEBUG
2367 /* need to poison the objs? */
2368 if (cachep->flags & SLAB_POISON)
2369 poison_obj(cachep, objp, POISON_FREE);
2370 if (cachep->flags & SLAB_STORE_USER)
2371 *dbg_userword(cachep, objp) = NULL;
2372
2373 if (cachep->flags & SLAB_RED_ZONE) {
2374 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2375 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2376 }
2377 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002378 * Constructors are not allowed to allocate memory from the same
2379 * cache which they are a constructor for. Otherwise, deadlock.
2380 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 */
2382 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002383 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002384 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385
2386 if (cachep->flags & SLAB_RED_ZONE) {
2387 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2388 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002389 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2391 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002392 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002394 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2395 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002396 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002397 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398#else
2399 if (cachep->ctor)
2400 cachep->ctor(objp, cachep, ctor_flags);
2401#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002402 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002404 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 slabp->free = 0;
2406}
2407
Pekka Enberg343e0d72006-02-01 03:05:50 -08002408static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002410 if (flags & SLAB_DMA)
2411 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2412 else
2413 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414}
2415
Andrew Mortona737b3e2006-03-22 00:08:11 -08002416static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2417 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002418{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002419 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002420 kmem_bufctl_t next;
2421
2422 slabp->inuse++;
2423 next = slab_bufctl(slabp)[slabp->free];
2424#if DEBUG
2425 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2426 WARN_ON(slabp->nodeid != nodeid);
2427#endif
2428 slabp->free = next;
2429
2430 return objp;
2431}
2432
Andrew Mortona737b3e2006-03-22 00:08:11 -08002433static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2434 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002435{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002436 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002437
2438#if DEBUG
2439 /* Verify that the slab belongs to the intended node */
2440 WARN_ON(slabp->nodeid != nodeid);
2441
Al Viro871751e2006-03-25 03:06:39 -08002442 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002443 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002444 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002445 BUG();
2446 }
2447#endif
2448 slab_bufctl(slabp)[objnr] = slabp->free;
2449 slabp->free = objnr;
2450 slabp->inuse--;
2451}
2452
Andrew Mortona737b3e2006-03-22 00:08:11 -08002453static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2454 void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455{
2456 int i;
2457 struct page *page;
2458
2459 /* Nasty!!!!!! I hope this is OK. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 page = virt_to_page(objp);
Nick Piggin84097512006-03-22 00:08:34 -08002461
2462 i = 1;
2463 if (likely(!PageCompound(page)))
2464 i <<= cachep->gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002466 page_set_cache(page, cachep);
2467 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 page++;
2469 } while (--i);
2470}
2471
2472/*
2473 * Grow (by 1) the number of slabs within a cache. This is called by
2474 * kmem_cache_alloc() when there are no active objs left in a cache.
2475 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002476static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002478 struct slab *slabp;
2479 void *objp;
2480 size_t offset;
2481 gfp_t local_flags;
2482 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002483 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484
Andrew Mortona737b3e2006-03-22 00:08:11 -08002485 /*
2486 * Be lazy and only check for valid flags here, keeping it out of the
2487 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002489 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 if (flags & SLAB_NO_GROW)
2491 return 0;
2492
2493 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2494 local_flags = (flags & SLAB_LEVEL_MASK);
2495 if (!(local_flags & __GFP_WAIT))
2496 /*
2497 * Not allowed to sleep. Need to tell a constructor about
2498 * this - it might need to know...
2499 */
2500 ctor_flags |= SLAB_CTOR_ATOMIC;
2501
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002502 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002504 l3 = cachep->nodelists[nodeid];
2505 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
2507 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002508 offset = l3->colour_next;
2509 l3->colour_next++;
2510 if (l3->colour_next >= cachep->colour)
2511 l3->colour_next = 0;
2512 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002514 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515
2516 if (local_flags & __GFP_WAIT)
2517 local_irq_enable();
2518
2519 /*
2520 * The test for missing atomic flag is performed here, rather than
2521 * the more obvious place, simply to reduce the critical path length
2522 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2523 * will eventually be caught here (where it matters).
2524 */
2525 kmem_flagcheck(cachep, flags);
2526
Andrew Mortona737b3e2006-03-22 00:08:11 -08002527 /*
2528 * Get mem for the objs. Attempt to allocate a physical page from
2529 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002530 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002531 objp = kmem_getpages(cachep, flags, nodeid);
2532 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 goto failed;
2534
2535 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002536 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002537 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 goto opps1;
2539
Christoph Lametere498be72005-09-09 13:03:32 -07002540 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 set_slab_attr(cachep, slabp, objp);
2542
2543 cache_init_objs(cachep, slabp, ctor_flags);
2544
2545 if (local_flags & __GFP_WAIT)
2546 local_irq_disable();
2547 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002548 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549
2550 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002551 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002553 l3->free_objects += cachep->num;
2554 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002556opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002558failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 if (local_flags & __GFP_WAIT)
2560 local_irq_disable();
2561 return 0;
2562}
2563
2564#if DEBUG
2565
2566/*
2567 * Perform extra freeing checks:
2568 * - detect bad pointers.
2569 * - POISON/RED_ZONE checking
2570 * - destructor calls, for caches with POISON+dtor
2571 */
2572static void kfree_debugcheck(const void *objp)
2573{
2574 struct page *page;
2575
2576 if (!virt_addr_valid(objp)) {
2577 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002578 (unsigned long)objp);
2579 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 }
2581 page = virt_to_page(objp);
2582 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002583 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2584 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585 BUG();
2586 }
2587}
2588
Pekka Enberg343e0d72006-02-01 03:05:50 -08002589static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002590 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591{
2592 struct page *page;
2593 unsigned int objnr;
2594 struct slab *slabp;
2595
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002596 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 kfree_debugcheck(objp);
2598 page = virt_to_page(objp);
2599
Pekka Enberg065d41c2005-11-13 16:06:46 -08002600 if (page_get_cache(page) != cachep) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002601 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2602 "cache %p, got %p\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002603 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002605 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2606 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 WARN_ON(1);
2608 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002609 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610
2611 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002612 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2613 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2614 slab_error(cachep, "double free, or memory outside"
2615 " object was overwritten");
2616 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2617 "redzone 2:0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002618 objp, *dbg_redzone1(cachep, objp),
2619 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 }
2621 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2622 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2623 }
2624 if (cachep->flags & SLAB_STORE_USER)
2625 *dbg_userword(cachep, objp) = caller;
2626
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002627 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628
2629 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002630 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631
2632 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002633 /*
2634 * Need to call the slab's constructor so the caller can
2635 * perform a verify of its state (debugging). Called without
2636 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002638 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002639 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 }
2641 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2642 /* we want to cache poison the object,
2643 * call the destruction callback
2644 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002645 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646 }
Al Viro871751e2006-03-25 03:06:39 -08002647#ifdef CONFIG_DEBUG_SLAB_LEAK
2648 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2649#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 if (cachep->flags & SLAB_POISON) {
2651#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002652 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002654 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002655 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 } else {
2657 poison_obj(cachep, objp, POISON_FREE);
2658 }
2659#else
2660 poison_obj(cachep, objp, POISON_FREE);
2661#endif
2662 }
2663 return objp;
2664}
2665
Pekka Enberg343e0d72006-02-01 03:05:50 -08002666static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667{
2668 kmem_bufctl_t i;
2669 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002670
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 /* Check slab's freelist to see if this obj is there. */
2672 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2673 entries++;
2674 if (entries > cachep->num || i >= cachep->num)
2675 goto bad;
2676 }
2677 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002678bad:
2679 printk(KERN_ERR "slab: Internal list corruption detected in "
2680 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2681 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002682 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002683 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002684 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002685 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002687 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688 }
2689 printk("\n");
2690 BUG();
2691 }
2692}
2693#else
2694#define kfree_debugcheck(x) do { } while(0)
2695#define cache_free_debugcheck(x,objp,z) (objp)
2696#define check_slabp(x,y) do { } while(0)
2697#endif
2698
Pekka Enberg343e0d72006-02-01 03:05:50 -08002699static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700{
2701 int batchcount;
2702 struct kmem_list3 *l3;
2703 struct array_cache *ac;
2704
2705 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002706 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002707retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 batchcount = ac->batchcount;
2709 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002710 /*
2711 * If there was little recent activity on this cache, then
2712 * perform only a partial refill. Otherwise we could generate
2713 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 */
2715 batchcount = BATCHREFILL_LIMIT;
2716 }
Christoph Lametere498be72005-09-09 13:03:32 -07002717 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718
Christoph Lametere498be72005-09-09 13:03:32 -07002719 BUG_ON(ac->avail > 0 || !l3);
2720 spin_lock(&l3->list_lock);
2721
Christoph Lameter3ded1752006-03-25 03:06:44 -08002722 /* See if we can refill from the shared array */
2723 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2724 goto alloc_done;
2725
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 while (batchcount > 0) {
2727 struct list_head *entry;
2728 struct slab *slabp;
2729 /* Get slab alloc is to come from. */
2730 entry = l3->slabs_partial.next;
2731 if (entry == &l3->slabs_partial) {
2732 l3->free_touched = 1;
2733 entry = l3->slabs_free.next;
2734 if (entry == &l3->slabs_free)
2735 goto must_grow;
2736 }
2737
2738 slabp = list_entry(entry, struct slab, list);
2739 check_slabp(cachep, slabp);
2740 check_spinlock_acquired(cachep);
2741 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 STATS_INC_ALLOCED(cachep);
2743 STATS_INC_ACTIVE(cachep);
2744 STATS_SET_HIGH(cachep);
2745
Matthew Dobson78d382d2006-02-01 03:05:47 -08002746 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2747 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 }
2749 check_slabp(cachep, slabp);
2750
2751 /* move slabp to correct slabp list: */
2752 list_del(&slabp->list);
2753 if (slabp->free == BUFCTL_END)
2754 list_add(&slabp->list, &l3->slabs_full);
2755 else
2756 list_add(&slabp->list, &l3->slabs_partial);
2757 }
2758
Andrew Mortona737b3e2006-03-22 00:08:11 -08002759must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002761alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002762 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
2764 if (unlikely(!ac->avail)) {
2765 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002766 x = cache_grow(cachep, flags, numa_node_id());
2767
Andrew Mortona737b3e2006-03-22 00:08:11 -08002768 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002769 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002770 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 return NULL;
2772
Andrew Mortona737b3e2006-03-22 00:08:11 -08002773 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 goto retry;
2775 }
2776 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002777 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778}
2779
Andrew Mortona737b3e2006-03-22 00:08:11 -08002780static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2781 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782{
2783 might_sleep_if(flags & __GFP_WAIT);
2784#if DEBUG
2785 kmem_flagcheck(cachep, flags);
2786#endif
2787}
2788
2789#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002790static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2791 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002793 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002795 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002797 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002798 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002799 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 else
2801 check_poison_obj(cachep, objp);
2802#else
2803 check_poison_obj(cachep, objp);
2804#endif
2805 poison_obj(cachep, objp, POISON_INUSE);
2806 }
2807 if (cachep->flags & SLAB_STORE_USER)
2808 *dbg_userword(cachep, objp) = caller;
2809
2810 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002811 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2812 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2813 slab_error(cachep, "double free, or memory outside"
2814 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002815 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002816 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2817 objp, *dbg_redzone1(cachep, objp),
2818 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 }
2820 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2821 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2822 }
Al Viro871751e2006-03-25 03:06:39 -08002823#ifdef CONFIG_DEBUG_SLAB_LEAK
2824 {
2825 struct slab *slabp;
2826 unsigned objnr;
2827
2828 slabp = page_get_slab(virt_to_page(objp));
2829 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2830 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2831 }
2832#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002833 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002835 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836
2837 if (!(flags & __GFP_WAIT))
2838 ctor_flags |= SLAB_CTOR_ATOMIC;
2839
2840 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002841 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 return objp;
2843}
2844#else
2845#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2846#endif
2847
Pekka Enberg343e0d72006-02-01 03:05:50 -08002848static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002850 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 struct array_cache *ac;
2852
Christoph Lameterdc85da12006-01-18 17:42:36 -08002853#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002854 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002855 objp = alternate_node_alloc(cachep, flags);
2856 if (objp != NULL)
2857 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08002858 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08002859#endif
2860
Alok N Kataria5c382302005-09-27 21:45:46 -07002861 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002862 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 if (likely(ac->avail)) {
2864 STATS_INC_ALLOCHIT(cachep);
2865 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002866 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 } else {
2868 STATS_INC_ALLOCMISS(cachep);
2869 objp = cache_alloc_refill(cachep, flags);
2870 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002871 return objp;
2872}
2873
Andrew Mortona737b3e2006-03-22 00:08:11 -08002874static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2875 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002876{
2877 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002878 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002879
2880 cache_alloc_debugcheck_before(cachep, flags);
2881
2882 local_irq_save(save_flags);
2883 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002885 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002886 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002887 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 return objp;
2889}
2890
Christoph Lametere498be72005-09-09 13:03:32 -07002891#ifdef CONFIG_NUMA
2892/*
Paul Jacksonb2455392006-03-24 03:16:12 -08002893 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002894 *
2895 * If we are in_interrupt, then process context, including cpusets and
2896 * mempolicy, may not apply and should not be used for allocation policy.
2897 */
2898static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2899{
2900 int nid_alloc, nid_here;
2901
2902 if (in_interrupt())
2903 return NULL;
2904 nid_alloc = nid_here = numa_node_id();
2905 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2906 nid_alloc = cpuset_mem_spread_node();
2907 else if (current->mempolicy)
2908 nid_alloc = slab_node(current->mempolicy);
2909 if (nid_alloc != nid_here)
2910 return __cache_alloc_node(cachep, flags, nid_alloc);
2911 return NULL;
2912}
2913
2914/*
Christoph Lametere498be72005-09-09 13:03:32 -07002915 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002917static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2918 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002919{
2920 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002921 struct slab *slabp;
2922 struct kmem_list3 *l3;
2923 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002924 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002926 l3 = cachep->nodelists[nodeid];
2927 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002928
Andrew Mortona737b3e2006-03-22 00:08:11 -08002929retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002930 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002931 spin_lock(&l3->list_lock);
2932 entry = l3->slabs_partial.next;
2933 if (entry == &l3->slabs_partial) {
2934 l3->free_touched = 1;
2935 entry = l3->slabs_free.next;
2936 if (entry == &l3->slabs_free)
2937 goto must_grow;
2938 }
Christoph Lametere498be72005-09-09 13:03:32 -07002939
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002940 slabp = list_entry(entry, struct slab, list);
2941 check_spinlock_acquired_node(cachep, nodeid);
2942 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002943
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002944 STATS_INC_NODEALLOCS(cachep);
2945 STATS_INC_ACTIVE(cachep);
2946 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002947
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002948 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002949
Matthew Dobson78d382d2006-02-01 03:05:47 -08002950 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002951 check_slabp(cachep, slabp);
2952 l3->free_objects--;
2953 /* move slabp to correct slabp list: */
2954 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002955
Andrew Mortona737b3e2006-03-22 00:08:11 -08002956 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002957 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002958 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002959 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002960
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002961 spin_unlock(&l3->list_lock);
2962 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002963
Andrew Mortona737b3e2006-03-22 00:08:11 -08002964must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002965 spin_unlock(&l3->list_lock);
2966 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002967
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002968 if (!x)
2969 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002970
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002971 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002972done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002973 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002974}
2975#endif
2976
2977/*
2978 * Caller needs to acquire correct kmem_list's list_lock
2979 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002980static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002981 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982{
2983 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002984 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985
2986 for (i = 0; i < nr_objects; i++) {
2987 void *objp = objpp[i];
2988 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002990 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002991 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002993 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002995 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002997 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 check_slabp(cachep, slabp);
2999
3000 /* fixup slab chains */
3001 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003002 if (l3->free_objects > l3->free_limit) {
3003 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004 slab_destroy(cachep, slabp);
3005 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003006 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 }
3008 } else {
3009 /* Unconditionally move a slab to the end of the
3010 * partial list on free - maximum time for the
3011 * other objects to be freed, too.
3012 */
Christoph Lametere498be72005-09-09 13:03:32 -07003013 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 }
3015 }
3016}
3017
Pekka Enberg343e0d72006-02-01 03:05:50 -08003018static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019{
3020 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003021 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003022 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023
3024 batchcount = ac->batchcount;
3025#if DEBUG
3026 BUG_ON(!batchcount || batchcount > ac->avail);
3027#endif
3028 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003029 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07003030 spin_lock(&l3->list_lock);
3031 if (l3->shared) {
3032 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003033 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 if (max) {
3035 if (batchcount > max)
3036 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003037 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003038 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 shared_array->avail += batchcount;
3040 goto free_done;
3041 }
3042 }
3043
Christoph Lameterff694162005-09-22 21:44:02 -07003044 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003045free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046#if STATS
3047 {
3048 int i = 0;
3049 struct list_head *p;
3050
Christoph Lametere498be72005-09-09 13:03:32 -07003051 p = l3->slabs_free.next;
3052 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 struct slab *slabp;
3054
3055 slabp = list_entry(p, struct slab, list);
3056 BUG_ON(slabp->inuse);
3057
3058 i++;
3059 p = p->next;
3060 }
3061 STATS_SET_FREEABLE(cachep, i);
3062 }
3063#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003064 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003066 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067}
3068
3069/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003070 * Release an obj back to its cache. If the obj has a constructed state, it must
3071 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003073static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003075 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076
3077 check_irq_off();
3078 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3079
Christoph Lametere498be72005-09-09 13:03:32 -07003080 /* Make sure we are not freeing a object from another
3081 * node to the array cache on this cpu.
3082 */
3083#ifdef CONFIG_NUMA
3084 {
3085 struct slab *slabp;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003086 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07003087 if (unlikely(slabp->nodeid != numa_node_id())) {
3088 struct array_cache *alien = NULL;
3089 int nodeid = slabp->nodeid;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003090 struct kmem_list3 *l3;
Christoph Lametere498be72005-09-09 13:03:32 -07003091
Andrew Mortona737b3e2006-03-22 00:08:11 -08003092 l3 = cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07003093 STATS_INC_NODEFREES(cachep);
3094 if (l3->alien && l3->alien[nodeid]) {
3095 alien = l3->alien[nodeid];
3096 spin_lock(&alien->lock);
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003097 if (unlikely(alien->avail == alien->limit)) {
3098 STATS_INC_ACOVERFLOW(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003099 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003100 alien, nodeid);
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003101 }
Christoph Lametere498be72005-09-09 13:03:32 -07003102 alien->entry[alien->avail++] = objp;
3103 spin_unlock(&alien->lock);
3104 } else {
3105 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003106 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003107 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003108 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003109 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003110 }
3111 return;
3112 }
3113 }
3114#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 if (likely(ac->avail < ac->limit)) {
3116 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003117 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 return;
3119 } else {
3120 STATS_INC_FREEMISS(cachep);
3121 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003122 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123 }
3124}
3125
3126/**
3127 * kmem_cache_alloc - Allocate an object
3128 * @cachep: The cache to allocate from.
3129 * @flags: See kmalloc().
3130 *
3131 * Allocate an object from this cache. The flags are only relevant
3132 * if the cache has no available objects.
3133 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003134void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003136 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137}
3138EXPORT_SYMBOL(kmem_cache_alloc);
3139
3140/**
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003141 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3142 * @cache: The cache to allocate from.
3143 * @flags: See kmalloc().
3144 *
3145 * Allocate an object from this cache and set the allocated memory to zero.
3146 * The flags are only relevant if the cache has no available objects.
3147 */
3148void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3149{
3150 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3151 if (ret)
3152 memset(ret, 0, obj_size(cache));
3153 return ret;
3154}
3155EXPORT_SYMBOL(kmem_cache_zalloc);
3156
3157/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 * kmem_ptr_validate - check if an untrusted pointer might
3159 * be a slab entry.
3160 * @cachep: the cache we're checking against
3161 * @ptr: pointer to validate
3162 *
3163 * This verifies that the untrusted pointer looks sane:
3164 * it is _not_ a guarantee that the pointer is actually
3165 * part of the slab cache in question, but it at least
3166 * validates that the pointer can be dereferenced and
3167 * looks half-way sane.
3168 *
3169 * Currently only used for dentry validation.
3170 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003171int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003173 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003175 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003176 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 struct page *page;
3178
3179 if (unlikely(addr < min_addr))
3180 goto out;
3181 if (unlikely(addr > (unsigned long)high_memory - size))
3182 goto out;
3183 if (unlikely(addr & align_mask))
3184 goto out;
3185 if (unlikely(!kern_addr_valid(addr)))
3186 goto out;
3187 if (unlikely(!kern_addr_valid(addr + size - 1)))
3188 goto out;
3189 page = virt_to_page(ptr);
3190 if (unlikely(!PageSlab(page)))
3191 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003192 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 goto out;
3194 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003195out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 return 0;
3197}
3198
3199#ifdef CONFIG_NUMA
3200/**
3201 * kmem_cache_alloc_node - Allocate an object on the specified node
3202 * @cachep: The cache to allocate from.
3203 * @flags: See kmalloc().
3204 * @nodeid: node number of the target node.
3205 *
3206 * Identical to kmem_cache_alloc, except that this function is slow
3207 * and can sleep. And it will allocate memory on the given node, which
3208 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003209 * New and improved: it will now make sure that the object gets
3210 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003212void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213{
Christoph Lametere498be72005-09-09 13:03:32 -07003214 unsigned long save_flags;
3215 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216
Christoph Lametere498be72005-09-09 13:03:32 -07003217 cache_alloc_debugcheck_before(cachep, flags);
3218 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003219
3220 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003221 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003222 ptr = ____cache_alloc(cachep, flags);
3223 else
3224 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003225 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003226
3227 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3228 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229
Christoph Lametere498be72005-09-09 13:03:32 -07003230 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231}
3232EXPORT_SYMBOL(kmem_cache_alloc_node);
3233
Al Virodd0fc662005-10-07 07:46:04 +01003234void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003235{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003236 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003237
3238 cachep = kmem_find_general_cachep(size, flags);
3239 if (unlikely(cachep == NULL))
3240 return NULL;
3241 return kmem_cache_alloc_node(cachep, flags, node);
3242}
3243EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003244#endif
3245
3246/**
3247 * kmalloc - allocate memory
3248 * @size: how many bytes of memory are required.
3249 * @flags: the type of memory to allocate.
Randy Dunlap911851e2006-03-22 00:08:14 -08003250 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 *
3252 * kmalloc is the normal method of allocating memory
3253 * in the kernel.
3254 *
3255 * The @flags argument may be one of:
3256 *
3257 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3258 *
3259 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3260 *
3261 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3262 *
3263 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3264 * must be suitable for DMA. This can mean different things on different
3265 * platforms. For example, on i386, it means that the memory must come
3266 * from the first 16MB.
3267 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003268static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3269 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003271 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003273 /* If you want to save a few bytes .text space: replace
3274 * __ with kmem_.
3275 * Then kmalloc uses the uninlined functions instead of the inline
3276 * functions.
3277 */
3278 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003279 if (unlikely(cachep == NULL))
3280 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003281 return __cache_alloc(cachep, flags, caller);
3282}
3283
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003284
3285void *__kmalloc(size_t size, gfp_t flags)
3286{
Al Viro871751e2006-03-25 03:06:39 -08003287#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003288 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003289#else
3290 return __do_kmalloc(size, flags, __builtin_return_address(0));
3291#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292}
3293EXPORT_SYMBOL(__kmalloc);
3294
Al Viro871751e2006-03-25 03:06:39 -08003295#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003296void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3297{
3298 return __do_kmalloc(size, flags, caller);
3299}
3300EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003301#endif
3302
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303#ifdef CONFIG_SMP
3304/**
3305 * __alloc_percpu - allocate one copy of the object for every present
3306 * cpu in the system, zeroing them.
3307 * Objects should be dereferenced using the per_cpu_ptr macro only.
3308 *
3309 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003311void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312{
3313 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003314 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315
3316 if (!pdata)
3317 return NULL;
3318
Christoph Lametere498be72005-09-09 13:03:32 -07003319 /*
3320 * Cannot use for_each_online_cpu since a cpu may come online
3321 * and we have no way of figuring out how to fix the array
3322 * that we have allocated then....
3323 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003324 for_each_possible_cpu(i) {
Christoph Lametere498be72005-09-09 13:03:32 -07003325 int node = cpu_to_node(i);
3326
3327 if (node_online(node))
3328 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3329 else
3330 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331
3332 if (!pdata->ptrs[i])
3333 goto unwind_oom;
3334 memset(pdata->ptrs[i], 0, size);
3335 }
3336
3337 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339
Andrew Mortona737b3e2006-03-22 00:08:11 -08003340unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341 while (--i >= 0) {
3342 if (!cpu_possible(i))
3343 continue;
3344 kfree(pdata->ptrs[i]);
3345 }
3346 kfree(pdata);
3347 return NULL;
3348}
3349EXPORT_SYMBOL(__alloc_percpu);
3350#endif
3351
3352/**
3353 * kmem_cache_free - Deallocate an object
3354 * @cachep: The cache the allocation was from.
3355 * @objp: The previously allocated object.
3356 *
3357 * Free an object which was previously allocated from this
3358 * cache.
3359 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003360void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361{
3362 unsigned long flags;
3363
3364 local_irq_save(flags);
3365 __cache_free(cachep, objp);
3366 local_irq_restore(flags);
3367}
3368EXPORT_SYMBOL(kmem_cache_free);
3369
3370/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 * kfree - free previously allocated memory
3372 * @objp: pointer returned by kmalloc.
3373 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003374 * If @objp is NULL, no operation is performed.
3375 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 * Don't free memory not originally allocated by kmalloc()
3377 * or you will run into trouble.
3378 */
3379void kfree(const void *objp)
3380{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003381 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 unsigned long flags;
3383
3384 if (unlikely(!objp))
3385 return;
3386 local_irq_save(flags);
3387 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003388 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003389 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003390 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 local_irq_restore(flags);
3392}
3393EXPORT_SYMBOL(kfree);
3394
3395#ifdef CONFIG_SMP
3396/**
3397 * free_percpu - free previously allocated percpu memory
3398 * @objp: pointer returned by alloc_percpu.
3399 *
3400 * Don't free memory not originally allocated by alloc_percpu()
3401 * The complemented objp is to check for that.
3402 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003403void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404{
3405 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003406 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407
Christoph Lametere498be72005-09-09 13:03:32 -07003408 /*
3409 * We allocate for all cpus so we cannot use for online cpu here.
3410 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003411 for_each_possible_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003412 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 kfree(p);
3414}
3415EXPORT_SYMBOL(free_percpu);
3416#endif
3417
Pekka Enberg343e0d72006-02-01 03:05:50 -08003418unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003420 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421}
3422EXPORT_SYMBOL(kmem_cache_size);
3423
Pekka Enberg343e0d72006-02-01 03:05:50 -08003424const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003425{
3426 return cachep->name;
3427}
3428EXPORT_SYMBOL_GPL(kmem_cache_name);
3429
Christoph Lametere498be72005-09-09 13:03:32 -07003430/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003431 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003432 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003433static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003434{
3435 int node;
3436 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003437 struct array_cache *new_shared;
3438 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003439
3440 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003441
Andrew Mortona737b3e2006-03-22 00:08:11 -08003442 new_alien = alloc_alien_cache(node, cachep->limit);
3443 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003444 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003445
Christoph Lameter0718dc22006-03-25 03:06:47 -08003446 new_shared = alloc_arraycache(node,
3447 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003448 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003449 if (!new_shared) {
3450 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003451 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003452 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003453
Andrew Mortona737b3e2006-03-22 00:08:11 -08003454 l3 = cachep->nodelists[node];
3455 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003456 struct array_cache *shared = l3->shared;
3457
Christoph Lametere498be72005-09-09 13:03:32 -07003458 spin_lock_irq(&l3->list_lock);
3459
Christoph Lametercafeb022006-03-25 03:06:46 -08003460 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003461 free_block(cachep, shared->entry,
3462 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003463
Christoph Lametercafeb022006-03-25 03:06:46 -08003464 l3->shared = new_shared;
3465 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003466 l3->alien = new_alien;
3467 new_alien = NULL;
3468 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003469 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003470 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003471 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003472 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003473 free_alien_cache(new_alien);
3474 continue;
3475 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003476 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003477 if (!l3) {
3478 free_alien_cache(new_alien);
3479 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003480 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003481 }
Christoph Lametere498be72005-09-09 13:03:32 -07003482
3483 kmem_list3_init(l3);
3484 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003485 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003486 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003487 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003488 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003489 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003490 cachep->nodelists[node] = l3;
3491 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003492 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003493
Andrew Mortona737b3e2006-03-22 00:08:11 -08003494fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003495 if (!cachep->next.next) {
3496 /* Cache is not active yet. Roll back what we did */
3497 node--;
3498 while (node >= 0) {
3499 if (cachep->nodelists[node]) {
3500 l3 = cachep->nodelists[node];
3501
3502 kfree(l3->shared);
3503 free_alien_cache(l3->alien);
3504 kfree(l3);
3505 cachep->nodelists[node] = NULL;
3506 }
3507 node--;
3508 }
3509 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003510 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003511}
3512
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003514 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 struct array_cache *new[NR_CPUS];
3516};
3517
3518static void do_ccupdate_local(void *info)
3519{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003520 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 struct array_cache *old;
3522
3523 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003524 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003525
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3527 new->new[smp_processor_id()] = old;
3528}
3529
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003530/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003531static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3532 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533{
3534 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003535 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003537 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003538 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003539 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3540 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003541 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003542 for (i--; i >= 0; i--)
3543 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003544 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 }
3546 }
3547 new.cachep = cachep;
3548
Andrew Mortona07fa392006-03-22 00:08:17 -08003549 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003550
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 cachep->batchcount = batchcount;
3553 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003554 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555
Christoph Lametere498be72005-09-09 13:03:32 -07003556 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557 struct array_cache *ccold = new.new[i];
3558 if (!ccold)
3559 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003560 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003561 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003562 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 kfree(ccold);
3564 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565
Christoph Lametere498be72005-09-09 13:03:32 -07003566 err = alloc_kmemlist(cachep);
3567 if (err) {
3568 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003569 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003570 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 return 0;
3573}
3574
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003575/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003576static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577{
3578 int err;
3579 int limit, shared;
3580
Andrew Mortona737b3e2006-03-22 00:08:11 -08003581 /*
3582 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 * - create a LIFO ordering, i.e. return objects that are cache-warm
3584 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003585 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 * bufctl chains: array operations are cheaper.
3587 * The numbers are guessed, we should auto-tune as described by
3588 * Bonwick.
3589 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003590 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003592 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003594 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003596 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 limit = 54;
3598 else
3599 limit = 120;
3600
Andrew Mortona737b3e2006-03-22 00:08:11 -08003601 /*
3602 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 * allocation behaviour: Most allocs on one cpu, most free operations
3604 * on another cpu. For these cases, an efficient object passing between
3605 * cpus is necessary. This is provided by a shared array. The array
3606 * replaces Bonwick's magazine layer.
3607 * On uniprocessor, it's functionally equivalent (but less efficient)
3608 * to a larger limit. Thus disabled by default.
3609 */
3610 shared = 0;
3611#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003612 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613 shared = 8;
3614#endif
3615
3616#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003617 /*
3618 * With debugging enabled, large batchcount lead to excessively long
3619 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003620 */
3621 if (limit > 32)
3622 limit = 32;
3623#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003624 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625 if (err)
3626 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003627 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628}
3629
Christoph Lameter1b552532006-03-22 00:09:07 -08003630/*
3631 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003632 * necessary. Note that the l3 listlock also protects the array_cache
3633 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003634 */
3635void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3636 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637{
3638 int tofree;
3639
Christoph Lameter1b552532006-03-22 00:09:07 -08003640 if (!ac || !ac->avail)
3641 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 if (ac->touched && !force) {
3643 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003644 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003645 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003646 if (ac->avail) {
3647 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3648 if (tofree > ac->avail)
3649 tofree = (ac->avail + 1) / 2;
3650 free_block(cachep, ac->entry, tofree, node);
3651 ac->avail -= tofree;
3652 memmove(ac->entry, &(ac->entry[tofree]),
3653 sizeof(void *) * ac->avail);
3654 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003655 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656 }
3657}
3658
3659/**
3660 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003661 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662 *
3663 * Called from workqueue/eventd every few seconds.
3664 * Purpose:
3665 * - clear the per-cpu caches for this CPU.
3666 * - return freeable pages to the main free memory pool.
3667 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003668 * If we cannot acquire the cache chain mutex then just give up - we'll try
3669 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 */
3671static void cache_reap(void *unused)
3672{
3673 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003674 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003675 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003677 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003679 schedule_delayed_work(&__get_cpu_var(reap_work),
3680 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 return;
3682 }
3683
3684 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003685 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003686 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687 int tofree;
3688 struct slab *slabp;
3689
Pekka Enberg343e0d72006-02-01 03:05:50 -08003690 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691 check_irq_on();
3692
Christoph Lameter35386e32006-03-22 00:09:05 -08003693 /*
3694 * We only take the l3 lock if absolutely necessary and we
3695 * have established with reasonable certainty that
3696 * we can do some work if the lock was obtained.
3697 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003698 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003699
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003700 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701
Christoph Lameteraab22072006-03-22 00:09:06 -08003702 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703
Christoph Lameter35386e32006-03-22 00:09:05 -08003704 /*
3705 * These are racy checks but it does not matter
3706 * if we skip one check or scan twice.
3707 */
Christoph Lametere498be72005-09-09 13:03:32 -07003708 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003709 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710
Christoph Lametere498be72005-09-09 13:03:32 -07003711 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712
Christoph Lameteraab22072006-03-22 00:09:06 -08003713 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714
Christoph Lametere498be72005-09-09 13:03:32 -07003715 if (l3->free_touched) {
3716 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003717 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 }
3719
Andrew Mortona737b3e2006-03-22 00:08:11 -08003720 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3721 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003723 /*
3724 * Do not lock if there are no free blocks.
3725 */
3726 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 break;
3728
Christoph Lameter35386e32006-03-22 00:09:05 -08003729 spin_lock_irq(&l3->list_lock);
3730 p = l3->slabs_free.next;
3731 if (p == &(l3->slabs_free)) {
3732 spin_unlock_irq(&l3->list_lock);
3733 break;
3734 }
3735
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 slabp = list_entry(p, struct slab, list);
3737 BUG_ON(slabp->inuse);
3738 list_del(&slabp->list);
3739 STATS_INC_REAPED(searchp);
3740
Andrew Mortona737b3e2006-03-22 00:08:11 -08003741 /*
3742 * Safe to drop the lock. The slab is no longer linked
3743 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744 * cache_chain_lock
3745 */
Christoph Lametere498be72005-09-09 13:03:32 -07003746 l3->free_objects -= searchp->num;
3747 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003749 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003750next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751 cond_resched();
3752 }
3753 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003754 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003755 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003756 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003757 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758}
3759
3760#ifdef CONFIG_PROC_FS
3761
Pekka Enberg85289f92006-01-08 01:00:36 -08003762static void print_slabinfo_header(struct seq_file *m)
3763{
3764 /*
3765 * Output format version, so at least we can change it
3766 * without _too_ many complaints.
3767 */
3768#if STATS
3769 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3770#else
3771 seq_puts(m, "slabinfo - version: 2.1\n");
3772#endif
3773 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3774 "<objperslab> <pagesperslab>");
3775 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3776 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3777#if STATS
3778 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003779 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003780 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3781#endif
3782 seq_putc(m, '\n');
3783}
3784
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785static void *s_start(struct seq_file *m, loff_t *pos)
3786{
3787 loff_t n = *pos;
3788 struct list_head *p;
3789
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003790 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003791 if (!n)
3792 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793 p = cache_chain.next;
3794 while (n--) {
3795 p = p->next;
3796 if (p == &cache_chain)
3797 return NULL;
3798 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003799 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800}
3801
3802static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3803{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003804 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003806 return cachep->next.next == &cache_chain ?
3807 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808}
3809
3810static void s_stop(struct seq_file *m, void *p)
3811{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003812 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813}
3814
3815static int s_show(struct seq_file *m, void *p)
3816{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003817 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003819 struct slab *slabp;
3820 unsigned long active_objs;
3821 unsigned long num_objs;
3822 unsigned long active_slabs = 0;
3823 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003824 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003826 int node;
3827 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 active_objs = 0;
3830 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003831 for_each_online_node(node) {
3832 l3 = cachep->nodelists[node];
3833 if (!l3)
3834 continue;
3835
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003836 check_irq_on();
3837 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003838
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003839 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003840 slabp = list_entry(q, struct slab, list);
3841 if (slabp->inuse != cachep->num && !error)
3842 error = "slabs_full accounting error";
3843 active_objs += cachep->num;
3844 active_slabs++;
3845 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003846 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003847 slabp = list_entry(q, struct slab, list);
3848 if (slabp->inuse == cachep->num && !error)
3849 error = "slabs_partial inuse accounting error";
3850 if (!slabp->inuse && !error)
3851 error = "slabs_partial/inuse accounting error";
3852 active_objs += slabp->inuse;
3853 active_slabs++;
3854 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003855 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003856 slabp = list_entry(q, struct slab, list);
3857 if (slabp->inuse && !error)
3858 error = "slabs_free/inuse accounting error";
3859 num_slabs++;
3860 }
3861 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003862 if (l3->shared)
3863 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003864
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003865 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003867 num_slabs += active_slabs;
3868 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003869 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870 error = "free_objects accounting error";
3871
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003872 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 if (error)
3874 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3875
3876 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003877 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003878 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003880 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003881 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003882 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003884 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885 unsigned long high = cachep->high_mark;
3886 unsigned long allocs = cachep->num_allocations;
3887 unsigned long grown = cachep->grown;
3888 unsigned long reaped = cachep->reaped;
3889 unsigned long errors = cachep->errors;
3890 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003891 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003892 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003893 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894
Christoph Lametere498be72005-09-09 13:03:32 -07003895 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003896 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003897 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003898 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899 }
3900 /* cpu stats */
3901 {
3902 unsigned long allochit = atomic_read(&cachep->allochit);
3903 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3904 unsigned long freehit = atomic_read(&cachep->freehit);
3905 unsigned long freemiss = atomic_read(&cachep->freemiss);
3906
3907 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003908 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 }
3910#endif
3911 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912 return 0;
3913}
3914
3915/*
3916 * slabinfo_op - iterator that generates /proc/slabinfo
3917 *
3918 * Output layout:
3919 * cache-name
3920 * num-active-objs
3921 * total-objs
3922 * object size
3923 * num-active-slabs
3924 * total-slabs
3925 * num-pages-per-slab
3926 * + further values on SMP and with statistics enabled
3927 */
3928
3929struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003930 .start = s_start,
3931 .next = s_next,
3932 .stop = s_stop,
3933 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934};
3935
3936#define MAX_SLABINFO_WRITE 128
3937/**
3938 * slabinfo_write - Tuning for the slab allocator
3939 * @file: unused
3940 * @buffer: user buffer
3941 * @count: data length
3942 * @ppos: unused
3943 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003944ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3945 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003947 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 int limit, batchcount, shared, res;
3949 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003950
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 if (count > MAX_SLABINFO_WRITE)
3952 return -EINVAL;
3953 if (copy_from_user(&kbuf, buffer, count))
3954 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003955 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956
3957 tmp = strchr(kbuf, ' ');
3958 if (!tmp)
3959 return -EINVAL;
3960 *tmp = '\0';
3961 tmp++;
3962 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3963 return -EINVAL;
3964
3965 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003966 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003968 list_for_each(p, &cache_chain) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003969 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970
Andrew Mortona737b3e2006-03-22 00:08:11 -08003971 cachep = list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003973 if (limit < 1 || batchcount < 1 ||
3974 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003975 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003977 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003978 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 }
3980 break;
3981 }
3982 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003983 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 if (res >= 0)
3985 res = count;
3986 return res;
3987}
Al Viro871751e2006-03-25 03:06:39 -08003988
3989#ifdef CONFIG_DEBUG_SLAB_LEAK
3990
3991static void *leaks_start(struct seq_file *m, loff_t *pos)
3992{
3993 loff_t n = *pos;
3994 struct list_head *p;
3995
3996 mutex_lock(&cache_chain_mutex);
3997 p = cache_chain.next;
3998 while (n--) {
3999 p = p->next;
4000 if (p == &cache_chain)
4001 return NULL;
4002 }
4003 return list_entry(p, struct kmem_cache, next);
4004}
4005
4006static inline int add_caller(unsigned long *n, unsigned long v)
4007{
4008 unsigned long *p;
4009 int l;
4010 if (!v)
4011 return 1;
4012 l = n[1];
4013 p = n + 2;
4014 while (l) {
4015 int i = l/2;
4016 unsigned long *q = p + 2 * i;
4017 if (*q == v) {
4018 q[1]++;
4019 return 1;
4020 }
4021 if (*q > v) {
4022 l = i;
4023 } else {
4024 p = q + 2;
4025 l -= i + 1;
4026 }
4027 }
4028 if (++n[1] == n[0])
4029 return 0;
4030 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4031 p[0] = v;
4032 p[1] = 1;
4033 return 1;
4034}
4035
4036static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4037{
4038 void *p;
4039 int i;
4040 if (n[0] == n[1])
4041 return;
4042 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4043 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4044 continue;
4045 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4046 return;
4047 }
4048}
4049
4050static void show_symbol(struct seq_file *m, unsigned long address)
4051{
4052#ifdef CONFIG_KALLSYMS
4053 char *modname;
4054 const char *name;
4055 unsigned long offset, size;
4056 char namebuf[KSYM_NAME_LEN+1];
4057
4058 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4059
4060 if (name) {
4061 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4062 if (modname)
4063 seq_printf(m, " [%s]", modname);
4064 return;
4065 }
4066#endif
4067 seq_printf(m, "%p", (void *)address);
4068}
4069
4070static int leaks_show(struct seq_file *m, void *p)
4071{
4072 struct kmem_cache *cachep = p;
4073 struct list_head *q;
4074 struct slab *slabp;
4075 struct kmem_list3 *l3;
4076 const char *name;
4077 unsigned long *n = m->private;
4078 int node;
4079 int i;
4080
4081 if (!(cachep->flags & SLAB_STORE_USER))
4082 return 0;
4083 if (!(cachep->flags & SLAB_RED_ZONE))
4084 return 0;
4085
4086 /* OK, we can do it */
4087
4088 n[1] = 0;
4089
4090 for_each_online_node(node) {
4091 l3 = cachep->nodelists[node];
4092 if (!l3)
4093 continue;
4094
4095 check_irq_on();
4096 spin_lock_irq(&l3->list_lock);
4097
4098 list_for_each(q, &l3->slabs_full) {
4099 slabp = list_entry(q, struct slab, list);
4100 handle_slab(n, cachep, slabp);
4101 }
4102 list_for_each(q, &l3->slabs_partial) {
4103 slabp = list_entry(q, struct slab, list);
4104 handle_slab(n, cachep, slabp);
4105 }
4106 spin_unlock_irq(&l3->list_lock);
4107 }
4108 name = cachep->name;
4109 if (n[0] == n[1]) {
4110 /* Increase the buffer size */
4111 mutex_unlock(&cache_chain_mutex);
4112 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4113 if (!m->private) {
4114 /* Too bad, we are really out */
4115 m->private = n;
4116 mutex_lock(&cache_chain_mutex);
4117 return -ENOMEM;
4118 }
4119 *(unsigned long *)m->private = n[0] * 2;
4120 kfree(n);
4121 mutex_lock(&cache_chain_mutex);
4122 /* Now make sure this entry will be retried */
4123 m->count = m->size;
4124 return 0;
4125 }
4126 for (i = 0; i < n[1]; i++) {
4127 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4128 show_symbol(m, n[2*i+2]);
4129 seq_putc(m, '\n');
4130 }
4131 return 0;
4132}
4133
4134struct seq_operations slabstats_op = {
4135 .start = leaks_start,
4136 .next = s_next,
4137 .stop = s_stop,
4138 .show = leaks_show,
4139};
4140#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004141#endif
4142
Manfred Spraul00e145b2005-09-03 15:55:07 -07004143/**
4144 * ksize - get the actual amount of memory allocated for a given object
4145 * @objp: Pointer to the object
4146 *
4147 * kmalloc may internally round up allocations and return more memory
4148 * than requested. ksize() can be used to determine the actual amount of
4149 * memory allocated. The caller may use this additional memory, even though
4150 * a smaller amount of memory was initially specified with the kmalloc call.
4151 * The caller must guarantee that objp points to a valid object previously
4152 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4153 * must not be freed during the duration of the call.
4154 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155unsigned int ksize(const void *objp)
4156{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004157 if (unlikely(objp == NULL))
4158 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004160 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161}