blob: 266449d604bd170c4c1656318c1ded64fcc5404a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
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>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700109#include <linux/rtmutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111#include <asm/uaccess.h>
112#include <asm/cacheflush.h>
113#include <asm/tlbflush.h>
114#include <asm/page.h>
115
116/*
117 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
118 * SLAB_RED_ZONE & SLAB_POISON.
119 * 0 for faster, smaller code (especially in the critical paths).
120 *
121 * STATS - 1 to collect stats for /proc/slabinfo.
122 * 0 for faster, smaller code (especially in the critical paths).
123 *
124 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 */
126
127#ifdef CONFIG_DEBUG_SLAB
128#define DEBUG 1
129#define STATS 1
130#define FORCED_DEBUG 1
131#else
132#define DEBUG 0
133#define STATS 0
134#define FORCED_DEBUG 0
135#endif
136
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137/* Shouldn't this be in a header file somewhere? */
138#define BYTES_PER_WORD sizeof(void *)
139
140#ifndef cache_line_size
141#define cache_line_size() L1_CACHE_BYTES
142#endif
143
144#ifndef ARCH_KMALLOC_MINALIGN
145/*
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
150 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
151 * Note that this flag disables some debug features.
152 */
153#define ARCH_KMALLOC_MINALIGN 0
154#endif
155
156#ifndef ARCH_SLAB_MINALIGN
157/*
158 * Enforce a minimum alignment for all caches.
159 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
160 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
161 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
162 * some debug features.
163 */
164#define ARCH_SLAB_MINALIGN 0
165#endif
166
167#ifndef ARCH_KMALLOC_FLAGS
168#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169#endif
170
171/* Legal flag mask for kmem_cache_create(). */
172#if DEBUG
173# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800175 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800180# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
182 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800183 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184#endif
185
186/*
187 * kmem_bufctl_t:
188 *
189 * Bufctl's are used for linking objs within a slab
190 * linked offsets.
191 *
192 * This implementation relies on "struct page" for locating the cache &
193 * slab an object belongs to.
194 * This allows the bufctl structure to be small (one int), but limits
195 * the number of objects a slab (not a cache) can contain when off-slab
196 * bufctls are used. The limit is the size of the largest general cache
197 * that does not use off-slab slabs.
198 * For 32bit archs with 4 kB pages, is this 56.
199 * This is not serious, as it is only for large objects, when it is unwise
200 * to have too many per slab.
201 * Note: This limit can be raised by introducing a general cache whose size
202 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 */
204
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700205typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
207#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800208#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
209#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211/*
212 * struct slab
213 *
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
217 */
218struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800219 struct list_head list;
220 unsigned long colouroff;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
223 kmem_bufctl_t free;
224 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225};
226
227/*
228 * struct slab_rcu
229 *
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
237 *
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
240 *
241 * We assume struct slab_rcu can overlay struct slab when destroying.
242 */
243struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800244 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800245 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247};
248
249/*
250 * struct array_cache
251 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 * Purpose:
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
256 *
257 * The limit is stored in the per-cpu structure to reduce the data cache
258 * footprint.
259 *
260 */
261struct array_cache {
262 unsigned int avail;
263 unsigned int limit;
264 unsigned int batchcount;
265 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700266 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800267 void *entry[0]; /*
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
270 * the entries.
271 * [0] is for gcc 2.95. It should really be [].
272 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273};
274
Andrew Mortona737b3e2006-03-22 00:08:11 -0800275/*
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 */
279#define BOOT_CPUCACHE_ENTRIES 1
280struct arraycache_init {
281 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800282 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283};
284
285/*
Christoph Lametere498be72005-09-09 13:03:32 -0700286 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 */
288struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800293 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800294 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800298 unsigned long next_reap; /* updated without locking */
299 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300};
301
Christoph Lametere498be72005-09-09 13:03:32 -0700302/*
303 * Need this for bootstrapping a per node allocator.
304 */
305#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307#define CACHE_CACHE 0
308#define SIZE_AC 1
309#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Christoph Lametered11d9e2006-06-30 01:55:45 -0700311static int drain_freelist(struct kmem_cache *cache,
312 struct kmem_list3 *l3, int tofree);
313static void free_block(struct kmem_cache *cachep, void **objpp, int len,
314 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700315static int enable_cpucache(struct kmem_cache *cachep);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700316static void cache_reap(void *unused);
317
Christoph Lametere498be72005-09-09 13:03:32 -0700318/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800319 * This function must be completely optimized away if a constant is passed to
320 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700321 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700322static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700323{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800324 extern void __bad_size(void);
325
Christoph Lametere498be72005-09-09 13:03:32 -0700326 if (__builtin_constant_p(size)) {
327 int i = 0;
328
329#define CACHE(x) \
330 if (size <=x) \
331 return i; \
332 else \
333 i++;
334#include "linux/kmalloc_sizes.h"
335#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800336 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700337 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800338 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700339 return 0;
340}
341
Ingo Molnare0a42722006-06-23 02:03:46 -0700342static int slab_early_init = 1;
343
Christoph Lametere498be72005-09-09 13:03:32 -0700344#define INDEX_AC index_of(sizeof(struct arraycache_init))
345#define INDEX_L3 index_of(sizeof(struct kmem_list3))
346
Pekka Enberg5295a742006-02-01 03:05:48 -0800347static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700348{
349 INIT_LIST_HEAD(&parent->slabs_full);
350 INIT_LIST_HEAD(&parent->slabs_partial);
351 INIT_LIST_HEAD(&parent->slabs_free);
352 parent->shared = NULL;
353 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800354 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700355 spin_lock_init(&parent->list_lock);
356 parent->free_objects = 0;
357 parent->free_touched = 0;
358}
359
Andrew Mortona737b3e2006-03-22 00:08:11 -0800360#define MAKE_LIST(cachep, listp, slab, nodeid) \
361 do { \
362 INIT_LIST_HEAD(listp); \
363 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700364 } while (0)
365
Andrew Mortona737b3e2006-03-22 00:08:11 -0800366#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
367 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700368 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
369 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
371 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
373/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800374 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 *
376 * manages a cache.
377 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800378
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800379struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800381 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800382/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800383 unsigned int batchcount;
384 unsigned int limit;
385 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800386
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800387 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800388/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800389 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800390
Andrew Mortona737b3e2006-03-22 00:08:11 -0800391 unsigned int flags; /* constant flags */
392 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800394/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800396 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
Andrew Mortona737b3e2006-03-22 00:08:11 -0800401 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800402 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800403 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800404 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800405 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406
407 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800408 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800411 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800413/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 const char *name;
415 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800417/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
422 unsigned long grown;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700428 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800429 atomic_t allochit;
430 atomic_t allocmiss;
431 atomic_t freehit;
432 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433#endif
434#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800435 /*
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
440 */
441 int obj_offset;
442 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443#endif
444};
445
446#define CFLGS_OFF_SLAB (0x80000000UL)
447#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
448
449#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800450/*
451 * Optimization question: fewer reaps means less probability for unnessary
452 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100454 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 * which could lock up otherwise freeable slabs.
456 */
457#define REAPTIMEOUT_CPUC (2*HZ)
458#define REAPTIMEOUT_LIST3 (4*HZ)
459
460#if STATS
461#define STATS_INC_ACTIVE(x) ((x)->num_active++)
462#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
463#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
464#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700465#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800466#define STATS_SET_HIGH(x) \
467 do { \
468 if ((x)->num_active > (x)->high_mark) \
469 (x)->high_mark = (x)->num_active; \
470 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471#define STATS_INC_ERR(x) ((x)->errors++)
472#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700473#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700474#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800475#define STATS_SET_FREEABLE(x, i) \
476 do { \
477 if ((x)->max_freeable < i) \
478 (x)->max_freeable = i; \
479 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
481#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
482#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
483#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
484#else
485#define STATS_INC_ACTIVE(x) do { } while (0)
486#define STATS_DEC_ACTIVE(x) do { } while (0)
487#define STATS_INC_ALLOCED(x) do { } while (0)
488#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700489#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490#define STATS_SET_HIGH(x) do { } while (0)
491#define STATS_INC_ERR(x) do { } while (0)
492#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700493#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700494#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800495#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496#define STATS_INC_ALLOCHIT(x) do { } while (0)
497#define STATS_INC_ALLOCMISS(x) do { } while (0)
498#define STATS_INC_FREEHIT(x) do { } while (0)
499#define STATS_INC_FREEMISS(x) do { } while (0)
500#endif
501
502#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
Andrew Mortona737b3e2006-03-22 00:08:11 -0800504/*
505 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800507 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 * the end of an object is aligned with the end of the real
509 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800510 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800512 * cachep->obj_offset: The real object.
513 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800514 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
515 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800517static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800519 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520}
521
Pekka Enberg343e0d72006-02-01 03:05:50 -0800522static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800524 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525}
526
Pekka Enberg343e0d72006-02-01 03:05:50 -0800527static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
529 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531}
532
Pekka Enberg343e0d72006-02-01 03:05:50 -0800533static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
535 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
536 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800537 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800538 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800539 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540}
541
Pekka Enberg343e0d72006-02-01 03:05:50 -0800542static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543{
544 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800545 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546}
547
548#else
549
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800550#define obj_offset(x) 0
551#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
553#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
555
556#endif
557
558/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800559 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
560 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 */
562#if defined(CONFIG_LARGE_ALLOCS)
563#define MAX_OBJ_ORDER 13 /* up to 32Mb */
564#define MAX_GFP_ORDER 13 /* up to 32Mb */
565#elif defined(CONFIG_MMU)
566#define MAX_OBJ_ORDER 5 /* 32 pages */
567#define MAX_GFP_ORDER 5 /* 32 pages */
568#else
569#define MAX_OBJ_ORDER 8 /* up to 1Mb */
570#define MAX_GFP_ORDER 8 /* up to 1Mb */
571#endif
572
573/*
574 * Do not go above this order unless 0 objects fit into the slab.
575 */
576#define BREAK_GFP_ORDER_HI 1
577#define BREAK_GFP_ORDER_LO 0
578static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
579
Andrew Mortona737b3e2006-03-22 00:08:11 -0800580/*
581 * Functions for storing/retrieving the cachep and or slab from the page
582 * allocator. These are used to find the slab an obj belongs to. With kfree(),
583 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800585static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
586{
587 page->lru.next = (struct list_head *)cache;
588}
589
590static inline struct kmem_cache *page_get_cache(struct page *page)
591{
Nick Piggin84097512006-03-22 00:08:34 -0800592 if (unlikely(PageCompound(page)))
593 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700594 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800595 return (struct kmem_cache *)page->lru.next;
596}
597
598static inline void page_set_slab(struct page *page, struct slab *slab)
599{
600 page->lru.prev = (struct list_head *)slab;
601}
602
603static inline struct slab *page_get_slab(struct page *page)
604{
Nick Piggin84097512006-03-22 00:08:34 -0800605 if (unlikely(PageCompound(page)))
606 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700607 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800608 return (struct slab *)page->lru.prev;
609}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800611static inline struct kmem_cache *virt_to_cache(const void *obj)
612{
613 struct page *page = virt_to_page(obj);
614 return page_get_cache(page);
615}
616
617static inline struct slab *virt_to_slab(const void *obj)
618{
619 struct page *page = virt_to_page(obj);
620 return page_get_slab(page);
621}
622
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800623static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
624 unsigned int idx)
625{
626 return slab->s_mem + cache->buffer_size * idx;
627}
628
629static inline unsigned int obj_to_index(struct kmem_cache *cache,
630 struct slab *slab, void *obj)
631{
632 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
633}
634
Andrew Mortona737b3e2006-03-22 00:08:11 -0800635/*
636 * These are the default caches for kmalloc. Custom caches can have other sizes.
637 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638struct cache_sizes malloc_sizes[] = {
639#define CACHE(x) { .cs_size = (x) },
640#include <linux/kmalloc_sizes.h>
641 CACHE(ULONG_MAX)
642#undef CACHE
643};
644EXPORT_SYMBOL(malloc_sizes);
645
646/* Must match cache_sizes above. Out of line to keep cache footprint low. */
647struct cache_names {
648 char *name;
649 char *name_dma;
650};
651
652static struct cache_names __initdata cache_names[] = {
653#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
654#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800655 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656#undef CACHE
657};
658
659static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800660 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
664/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800665static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800666 .batchcount = 1,
667 .limit = BOOT_CPUCACHE_ENTRIES,
668 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800669 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800670 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800672 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673#endif
674};
675
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700676#define BAD_ALIEN_MAGIC 0x01020304ul
677
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200678#ifdef CONFIG_LOCKDEP
679
680/*
681 * Slab sometimes uses the kmalloc slabs to store the slab headers
682 * for other slabs "off slab".
683 * The locking for this is tricky in that it nests within the locks
684 * of all other slabs in a few places; to deal with this special
685 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700686 *
687 * We set lock class for alien array caches which are up during init.
688 * The lock annotation will be lost if all cpus of a node goes down and
689 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200690 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700691static struct lock_class_key on_slab_l3_key;
692static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200693
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700694static inline void init_lock_keys(void)
695
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200696{
697 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700698 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200699
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700700 while (s->cs_size != ULONG_MAX) {
701 for_each_node(q) {
702 struct array_cache **alc;
703 int r;
704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
705 if (!l3 || OFF_SLAB(s->cs_cachep))
706 continue;
707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
708 alc = l3->alien;
709 /*
710 * FIXME: This check for BAD_ALIEN_MAGIC
711 * should go away when common slab code is taught to
712 * work even without alien caches.
713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
714 * for alloc_alien_cache,
715 */
716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
717 continue;
718 for_each_node(r) {
719 if (alc[r])
720 lockdep_set_class(&alc[r]->lock,
721 &on_slab_alc_key);
722 }
723 }
724 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200725 }
726}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200727#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700728static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200729{
730}
731#endif
732
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800734static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735static struct list_head cache_chain;
736
737/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 * chicken and egg problem: delay the per-cpu array allocation
739 * until the general caches are up.
740 */
741static enum {
742 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700743 PARTIAL_AC,
744 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 FULL
746} g_cpucache_up;
747
Mike Kravetz39d24e62006-05-15 09:44:13 -0700748/*
749 * used by boot code to determine if it can use slab based allocator
750 */
751int slab_is_available(void)
752{
753 return g_cpucache_up == FULL;
754}
755
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756static DEFINE_PER_CPU(struct work_struct, reap_work);
757
Pekka Enberg343e0d72006-02-01 03:05:50 -0800758static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759{
760 return cachep->array[smp_processor_id()];
761}
762
Andrew Mortona737b3e2006-03-22 00:08:11 -0800763static inline struct kmem_cache *__find_general_cachep(size_t size,
764 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765{
766 struct cache_sizes *csizep = malloc_sizes;
767
768#if DEBUG
769 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800770 * kmem_cache_create(), or __kmalloc(), before
771 * the generic caches are initialized.
772 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700773 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774#endif
775 while (size > csizep->cs_size)
776 csizep++;
777
778 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700779 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 * has cs_{dma,}cachep==NULL. Thus no special case
781 * for large kmalloc calls required.
782 */
783 if (unlikely(gfpflags & GFP_DMA))
784 return csizep->cs_dmacachep;
785 return csizep->cs_cachep;
786}
787
Adrian Bunkb2213852006-09-25 23:31:02 -0700788static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700789{
790 return __find_general_cachep(size, gfpflags);
791}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700792
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800793static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800795 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
796}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797
Andrew Mortona737b3e2006-03-22 00:08:11 -0800798/*
799 * Calculate the number of objects and left-over bytes for a given buffer size.
800 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800801static void cache_estimate(unsigned long gfporder, size_t buffer_size,
802 size_t align, int flags, size_t *left_over,
803 unsigned int *num)
804{
805 int nr_objs;
806 size_t mgmt_size;
807 size_t slab_size = PAGE_SIZE << gfporder;
808
809 /*
810 * The slab management structure can be either off the slab or
811 * on it. For the latter case, the memory allocated for a
812 * slab is used for:
813 *
814 * - The struct slab
815 * - One kmem_bufctl_t for each object
816 * - Padding to respect alignment of @align
817 * - @buffer_size bytes for each object
818 *
819 * If the slab management structure is off the slab, then the
820 * alignment will already be calculated into the size. Because
821 * the slabs are all pages aligned, the objects will be at the
822 * correct alignment when allocated.
823 */
824 if (flags & CFLGS_OFF_SLAB) {
825 mgmt_size = 0;
826 nr_objs = slab_size / buffer_size;
827
828 if (nr_objs > SLAB_LIMIT)
829 nr_objs = SLAB_LIMIT;
830 } else {
831 /*
832 * Ignore padding for the initial guess. The padding
833 * is at most @align-1 bytes, and @buffer_size is at
834 * least @align. In the worst case, this result will
835 * be one greater than the number of objects that fit
836 * into the memory allocation when taking the padding
837 * into account.
838 */
839 nr_objs = (slab_size - sizeof(struct slab)) /
840 (buffer_size + sizeof(kmem_bufctl_t));
841
842 /*
843 * This calculated number will be either the right
844 * amount, or one greater than what we want.
845 */
846 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
847 > slab_size)
848 nr_objs--;
849
850 if (nr_objs > SLAB_LIMIT)
851 nr_objs = SLAB_LIMIT;
852
853 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800855 *num = nr_objs;
856 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857}
858
859#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
860
Andrew Mortona737b3e2006-03-22 00:08:11 -0800861static void __slab_error(const char *function, struct kmem_cache *cachep,
862 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863{
864 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800865 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 dump_stack();
867}
868
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800869#ifdef CONFIG_NUMA
870/*
871 * Special reaping functions for NUMA systems called from cache_reap().
872 * These take care of doing round robin flushing of alien caches (containing
873 * objects freed on different nodes from which they were allocated) and the
874 * flushing of remote pcps by calling drain_node_pages.
875 */
876static DEFINE_PER_CPU(unsigned long, reap_node);
877
878static void init_reap_node(int cpu)
879{
880 int node;
881
882 node = next_node(cpu_to_node(cpu), node_online_map);
883 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800884 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800885
886 __get_cpu_var(reap_node) = node;
887}
888
889static void next_reap_node(void)
890{
891 int node = __get_cpu_var(reap_node);
892
893 /*
894 * Also drain per cpu pages on remote zones
895 */
896 if (node != numa_node_id())
897 drain_node_pages(node);
898
899 node = next_node(node, node_online_map);
900 if (unlikely(node >= MAX_NUMNODES))
901 node = first_node(node_online_map);
902 __get_cpu_var(reap_node) = node;
903}
904
905#else
906#define init_reap_node(cpu) do { } while (0)
907#define next_reap_node(void) do { } while (0)
908#endif
909
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910/*
911 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
912 * via the workqueue/eventd.
913 * Add the CPU number into the expiration time to minimize the possibility of
914 * the CPUs getting into lockstep and contending for the global cache chain
915 * lock.
916 */
917static void __devinit start_cpu_timer(int cpu)
918{
919 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
920
921 /*
922 * When this gets called from do_initcalls via cpucache_init(),
923 * init_workqueues() has already run, so keventd will be setup
924 * at that time.
925 */
926 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800927 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 INIT_WORK(reap_work, cache_reap, NULL);
929 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
930 }
931}
932
Christoph Lametere498be72005-09-09 13:03:32 -0700933static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800934 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800936 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 struct array_cache *nc = NULL;
938
Christoph Lametere498be72005-09-09 13:03:32 -0700939 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 if (nc) {
941 nc->avail = 0;
942 nc->limit = entries;
943 nc->batchcount = batchcount;
944 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700945 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 }
947 return nc;
948}
949
Christoph Lameter3ded1752006-03-25 03:06:44 -0800950/*
951 * Transfer objects in one arraycache to another.
952 * Locking must be handled by the caller.
953 *
954 * Return the number of entries transferred.
955 */
956static int transfer_objects(struct array_cache *to,
957 struct array_cache *from, unsigned int max)
958{
959 /* Figure out how many entries to transfer */
960 int nr = min(min(from->avail, max), to->limit - to->avail);
961
962 if (!nr)
963 return 0;
964
965 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
966 sizeof(void *) *nr);
967
968 from->avail -= nr;
969 to->avail += nr;
970 to->touched = 1;
971 return nr;
972}
973
Christoph Lameter765c4502006-09-27 01:50:08 -0700974#ifndef CONFIG_NUMA
975
976#define drain_alien_cache(cachep, alien) do { } while (0)
977#define reap_alien(cachep, l3) do { } while (0)
978
979static inline struct array_cache **alloc_alien_cache(int node, int limit)
980{
981 return (struct array_cache **)BAD_ALIEN_MAGIC;
982}
983
984static inline void free_alien_cache(struct array_cache **ac_ptr)
985{
986}
987
988static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
989{
990 return 0;
991}
992
993static inline void *alternate_node_alloc(struct kmem_cache *cachep,
994 gfp_t flags)
995{
996 return NULL;
997}
998
999static inline void *__cache_alloc_node(struct kmem_cache *cachep,
1000 gfp_t flags, int nodeid)
1001{
1002 return NULL;
1003}
1004
1005#else /* CONFIG_NUMA */
1006
Pekka Enberg343e0d72006-02-01 03:05:50 -08001007static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001008static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001009
Pekka Enberg5295a742006-02-01 03:05:48 -08001010static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001011{
1012 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001013 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -07001014 int i;
1015
1016 if (limit > 1)
1017 limit = 12;
1018 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1019 if (ac_ptr) {
1020 for_each_node(i) {
1021 if (i == node || !node_online(i)) {
1022 ac_ptr[i] = NULL;
1023 continue;
1024 }
1025 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1026 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001027 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001028 kfree(ac_ptr[i]);
1029 kfree(ac_ptr);
1030 return NULL;
1031 }
1032 }
1033 }
1034 return ac_ptr;
1035}
1036
Pekka Enberg5295a742006-02-01 03:05:48 -08001037static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001038{
1039 int i;
1040
1041 if (!ac_ptr)
1042 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001043 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001044 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001045 kfree(ac_ptr);
1046}
1047
Pekka Enberg343e0d72006-02-01 03:05:50 -08001048static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001049 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001050{
1051 struct kmem_list3 *rl3 = cachep->nodelists[node];
1052
1053 if (ac->avail) {
1054 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001055 /*
1056 * Stuff objects into the remote nodes shared array first.
1057 * That way we could avoid the overhead of putting the objects
1058 * into the free lists and getting them back later.
1059 */
shin, jacob693f7d32006-04-28 10:54:37 -05001060 if (rl3->shared)
1061 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001062
Christoph Lameterff694162005-09-22 21:44:02 -07001063 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001064 ac->avail = 0;
1065 spin_unlock(&rl3->list_lock);
1066 }
1067}
1068
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001069/*
1070 * Called from cache_reap() to regularly drain alien caches round robin.
1071 */
1072static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1073{
1074 int node = __get_cpu_var(reap_node);
1075
1076 if (l3->alien) {
1077 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001078
1079 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001080 __drain_alien_cache(cachep, ac, node);
1081 spin_unlock_irq(&ac->lock);
1082 }
1083 }
1084}
1085
Andrew Mortona737b3e2006-03-22 00:08:11 -08001086static void drain_alien_cache(struct kmem_cache *cachep,
1087 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001088{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001089 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001090 struct array_cache *ac;
1091 unsigned long flags;
1092
1093 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001094 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001095 if (ac) {
1096 spin_lock_irqsave(&ac->lock, flags);
1097 __drain_alien_cache(cachep, ac, i);
1098 spin_unlock_irqrestore(&ac->lock, flags);
1099 }
1100 }
1101}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001102
Ingo Molnar873623d2006-07-13 14:44:38 +02001103static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001104{
1105 struct slab *slabp = virt_to_slab(objp);
1106 int nodeid = slabp->nodeid;
1107 struct kmem_list3 *l3;
1108 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001109 int node;
1110
1111 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001112
1113 /*
1114 * Make sure we are not freeing a object from another node to the array
1115 * cache on this cpu.
1116 */
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001117 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001118 return 0;
1119
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001120 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001121 STATS_INC_NODEFREES(cachep);
1122 if (l3->alien && l3->alien[nodeid]) {
1123 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001124 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001125 if (unlikely(alien->avail == alien->limit)) {
1126 STATS_INC_ACOVERFLOW(cachep);
1127 __drain_alien_cache(cachep, alien, nodeid);
1128 }
1129 alien->entry[alien->avail++] = objp;
1130 spin_unlock(&alien->lock);
1131 } else {
1132 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1133 free_block(cachep, &objp, 1, nodeid);
1134 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1135 }
1136 return 1;
1137}
Christoph Lametere498be72005-09-09 13:03:32 -07001138#endif
1139
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001140static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001141 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142{
1143 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001144 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001145 struct kmem_list3 *l3 = NULL;
1146 int node = cpu_to_node(cpu);
1147 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148
1149 switch (action) {
1150 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001151 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001152 /*
1153 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001154 * alloc_arraycache's are going to use this list.
1155 * kmalloc_node allows us to add the slab to the right
1156 * kmem_list3 and not this cpu's kmem_list3
1157 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
Christoph Lametere498be72005-09-09 13:03:32 -07001159 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001160 /*
1161 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001162 * begin anything. Make sure some other cpu on this
1163 * node has not already allocated this
1164 */
1165 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001166 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1167 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001168 goto bad;
1169 kmem_list3_init(l3);
1170 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001171 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001172
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001173 /*
1174 * The l3s don't come and go as CPUs come and
1175 * go. cache_chain_mutex is sufficient
1176 * protection here.
1177 */
Christoph Lametere498be72005-09-09 13:03:32 -07001178 cachep->nodelists[node] = l3;
1179 }
1180
1181 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1182 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001183 (1 + nr_cpus_node(node)) *
1184 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001185 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1186 }
1187
Andrew Mortona737b3e2006-03-22 00:08:11 -08001188 /*
1189 * Now we can go ahead with allocating the shared arrays and
1190 * array caches
1191 */
Christoph Lametere498be72005-09-09 13:03:32 -07001192 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001193 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001194 struct array_cache *shared;
1195 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001196
Christoph Lametere498be72005-09-09 13:03:32 -07001197 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001198 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 if (!nc)
1200 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001201 shared = alloc_arraycache(node,
1202 cachep->shared * cachep->batchcount,
1203 0xbaadf00d);
1204 if (!shared)
1205 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001206
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001207 alien = alloc_alien_cache(node, cachep->limit);
1208 if (!alien)
1209 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001211 l3 = cachep->nodelists[node];
1212 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001213
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001214 spin_lock_irq(&l3->list_lock);
1215 if (!l3->shared) {
1216 /*
1217 * We are serialised from CPU_DEAD or
1218 * CPU_UP_CANCELLED by the cpucontrol lock
1219 */
1220 l3->shared = shared;
1221 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001222 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001223#ifdef CONFIG_NUMA
1224 if (!l3->alien) {
1225 l3->alien = alien;
1226 alien = NULL;
1227 }
1228#endif
1229 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001230 kfree(shared);
1231 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001233 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 break;
1235 case CPU_ONLINE:
1236 start_cpu_timer(cpu);
1237 break;
1238#ifdef CONFIG_HOTPLUG_CPU
1239 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001240 /*
1241 * Even if all the cpus of a node are down, we don't free the
1242 * kmem_list3 of any cache. This to avoid a race between
1243 * cpu_down, and a kmalloc allocation from another cpu for
1244 * memory from the node of the cpu going down. The list3
1245 * structure is usually allocated from kmem_cache_create() and
1246 * gets destroyed at kmem_cache_destroy().
1247 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 /* fall thru */
1249 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001250 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 list_for_each_entry(cachep, &cache_chain, next) {
1252 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001253 struct array_cache *shared;
1254 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001255 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
Christoph Lametere498be72005-09-09 13:03:32 -07001257 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 /* cpu is dead; no one can alloc from it. */
1259 nc = cachep->array[cpu];
1260 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001261 l3 = cachep->nodelists[node];
1262
1263 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001264 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001265
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001266 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001267
1268 /* Free limit for this kmem_list3 */
1269 l3->free_limit -= cachep->batchcount;
1270 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001271 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001272
1273 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001274 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001275 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001276 }
Christoph Lametere498be72005-09-09 13:03:32 -07001277
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001278 shared = l3->shared;
1279 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001280 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001281 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001282 l3->shared = NULL;
1283 }
Christoph Lametere498be72005-09-09 13:03:32 -07001284
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001285 alien = l3->alien;
1286 l3->alien = NULL;
1287
1288 spin_unlock_irq(&l3->list_lock);
1289
1290 kfree(shared);
1291 if (alien) {
1292 drain_alien_cache(cachep, alien);
1293 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001294 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001295free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 kfree(nc);
1297 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001298 /*
1299 * In the previous loop, all the objects were freed to
1300 * the respective cache's slabs, now we can go ahead and
1301 * shrink each nodelist to its limit.
1302 */
1303 list_for_each_entry(cachep, &cache_chain, next) {
1304 l3 = cachep->nodelists[node];
1305 if (!l3)
1306 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001307 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001308 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001309 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 break;
1311#endif
1312 }
1313 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001314bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001315 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 return NOTIFY_BAD;
1317}
1318
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001319static struct notifier_block __cpuinitdata cpucache_notifier = {
1320 &cpuup_callback, NULL, 0
1321};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Christoph Lametere498be72005-09-09 13:03:32 -07001323/*
1324 * swap the static kmem_list3 with kmalloced memory
1325 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001326static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1327 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001328{
1329 struct kmem_list3 *ptr;
1330
Christoph Lametere498be72005-09-09 13:03:32 -07001331 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1332 BUG_ON(!ptr);
1333
1334 local_irq_disable();
1335 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001336 /*
1337 * Do not assume that spinlocks can be initialized via memcpy:
1338 */
1339 spin_lock_init(&ptr->list_lock);
1340
Christoph Lametere498be72005-09-09 13:03:32 -07001341 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1342 cachep->nodelists[nodeid] = ptr;
1343 local_irq_enable();
1344}
1345
Andrew Mortona737b3e2006-03-22 00:08:11 -08001346/*
1347 * Initialisation. Called after the page allocator have been initialised and
1348 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 */
1350void __init kmem_cache_init(void)
1351{
1352 size_t left_over;
1353 struct cache_sizes *sizes;
1354 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001355 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001356 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001357 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001358
1359 for (i = 0; i < NUM_INIT_LISTS; i++) {
1360 kmem_list3_init(&initkmem_list3[i]);
1361 if (i < MAX_NUMNODES)
1362 cache_cache.nodelists[i] = NULL;
1363 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364
1365 /*
1366 * Fragmentation resistance on low memory - only use bigger
1367 * page orders on machines with more than 32MB of memory.
1368 */
1369 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1370 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1371
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 /* Bootstrap is tricky, because several objects are allocated
1373 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001374 * 1) initialize the cache_cache cache: it contains the struct
1375 * kmem_cache structures of all caches, except cache_cache itself:
1376 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001377 * Initially an __init data area is used for the head array and the
1378 * kmem_list3 structures, it's replaced with a kmalloc allocated
1379 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001381 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001382 * An __init data area is used for the head array.
1383 * 3) Create the remaining kmalloc caches, with minimally sized
1384 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 * 4) Replace the __init data head arrays for cache_cache and the first
1386 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001387 * 5) Replace the __init data for kmem_list3 for cache_cache and
1388 * the other cache's with kmalloc allocated memory.
1389 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 */
1391
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001392 node = numa_node_id();
1393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 INIT_LIST_HEAD(&cache_chain);
1396 list_add(&cache_cache.next, &cache_chain);
1397 cache_cache.colour_off = cache_line_size();
1398 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001399 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400
Andrew Mortona737b3e2006-03-22 00:08:11 -08001401 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1402 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
Jack Steiner07ed76b2006-03-07 21:55:46 -08001404 for (order = 0; order < MAX_ORDER; order++) {
1405 cache_estimate(order, cache_cache.buffer_size,
1406 cache_line_size(), 0, &left_over, &cache_cache.num);
1407 if (cache_cache.num)
1408 break;
1409 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001410 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001411 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001412 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001413 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1414 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415
1416 /* 2+3) create the kmalloc caches */
1417 sizes = malloc_sizes;
1418 names = cache_names;
1419
Andrew Mortona737b3e2006-03-22 00:08:11 -08001420 /*
1421 * Initialize the caches that provide memory for the array cache and the
1422 * kmem_list3 structures first. Without this, further allocations will
1423 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001424 */
1425
1426 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001427 sizes[INDEX_AC].cs_size,
1428 ARCH_KMALLOC_MINALIGN,
1429 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1430 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001431
Andrew Mortona737b3e2006-03-22 00:08:11 -08001432 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001433 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001434 kmem_cache_create(names[INDEX_L3].name,
1435 sizes[INDEX_L3].cs_size,
1436 ARCH_KMALLOC_MINALIGN,
1437 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1438 NULL, NULL);
1439 }
Christoph Lametere498be72005-09-09 13:03:32 -07001440
Ingo Molnare0a42722006-06-23 02:03:46 -07001441 slab_early_init = 0;
1442
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001444 /*
1445 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446 * This should be particularly beneficial on SMP boxes, as it
1447 * eliminates "false sharing".
1448 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001449 * allow tighter packing of the smaller caches.
1450 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001451 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001452 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001453 sizes->cs_size,
1454 ARCH_KMALLOC_MINALIGN,
1455 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1456 NULL, NULL);
1457 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001460 sizes->cs_size,
1461 ARCH_KMALLOC_MINALIGN,
1462 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1463 SLAB_PANIC,
1464 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 sizes++;
1466 names++;
1467 }
1468 /* 4) Replace the bootstrap head arrays */
1469 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001470 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001475 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1476 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001477 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001478 /*
1479 * Do not assume that spinlocks can be initialized via memcpy:
1480 */
1481 spin_lock_init(&ptr->lock);
1482
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 cache_cache.array[smp_processor_id()] = ptr;
1484 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001485
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001487
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001489 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001490 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001491 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001492 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001493 /*
1494 * Do not assume that spinlocks can be initialized via memcpy:
1495 */
1496 spin_lock_init(&ptr->lock);
1497
Christoph Lametere498be72005-09-09 13:03:32 -07001498 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001499 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 local_irq_enable();
1501 }
Christoph Lametere498be72005-09-09 13:03:32 -07001502 /* 5) Replace the bootstrap kmem_list3's */
1503 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001504 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001506 /* Replace the static kmem_list3 structures for the boot cpu */
1507 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1508
1509 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001510 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001511 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001512
1513 if (INDEX_AC != INDEX_L3) {
1514 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001515 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001516 }
1517 }
1518 }
1519
1520 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001522 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001523 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001525 if (enable_cpucache(cachep))
1526 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001527 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 }
1529
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001530 /* Annotate slab for lockdep -- annotate the malloc caches */
1531 init_lock_keys();
1532
1533
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 /* Done! */
1535 g_cpucache_up = FULL;
1536
Andrew Mortona737b3e2006-03-22 00:08:11 -08001537 /*
1538 * Register a cpu startup notifier callback that initializes
1539 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 */
1541 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542
Andrew Mortona737b3e2006-03-22 00:08:11 -08001543 /*
1544 * The reap timers are started later, with a module init call: That part
1545 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 */
1547}
1548
1549static int __init cpucache_init(void)
1550{
1551 int cpu;
1552
Andrew Mortona737b3e2006-03-22 00:08:11 -08001553 /*
1554 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 */
Christoph Lametere498be72005-09-09 13:03:32 -07001556 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 return 0;
1559}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560__initcall(cpucache_init);
1561
1562/*
1563 * Interface to system's page allocator. No need to hold the cache-lock.
1564 *
1565 * If we requested dmaable memory, we will get it. Even if we
1566 * did not request dmaable memory, we might get it, but that
1567 * would be relatively rare and ignorable.
1568 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001569static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570{
1571 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001572 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 int i;
1574
Luke Yangd6fef9d2006-04-10 22:52:56 -07001575#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001576 /*
1577 * Nommu uses slab's for process anonymous memory allocations, and thus
1578 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001579 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001580 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001581#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001582
1583 /*
1584 * Under NUMA we want memory on the indicated node. We will handle
1585 * the needed fallback ourselves since we want to serve from our
1586 * per node object lists first for other nodes.
1587 */
1588 flags |= cachep->gfpflags | GFP_THISNODE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001589
1590 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 if (!page)
1592 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001594 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001596 add_zone_page_state(page_zone(page),
1597 NR_SLAB_RECLAIMABLE, nr_pages);
1598 else
1599 add_zone_page_state(page_zone(page),
1600 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001601 for (i = 0; i < nr_pages; i++)
1602 __SetPageSlab(page + i);
1603 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604}
1605
1606/*
1607 * Interface to system's page release.
1608 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001609static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001611 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 struct page *page = virt_to_page(addr);
1613 const unsigned long nr_freed = i;
1614
Christoph Lameter972d1a72006-09-25 23:31:51 -07001615 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1616 sub_zone_page_state(page_zone(page),
1617 NR_SLAB_RECLAIMABLE, nr_freed);
1618 else
1619 sub_zone_page_state(page_zone(page),
1620 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001622 BUG_ON(!PageSlab(page));
1623 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 page++;
1625 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 if (current->reclaim_state)
1627 current->reclaim_state->reclaimed_slab += nr_freed;
1628 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629}
1630
1631static void kmem_rcu_free(struct rcu_head *head)
1632{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001633 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001634 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635
1636 kmem_freepages(cachep, slab_rcu->addr);
1637 if (OFF_SLAB(cachep))
1638 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1639}
1640
1641#if DEBUG
1642
1643#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001644static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001645 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001647 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001649 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001651 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 return;
1653
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001654 *addr++ = 0x12345678;
1655 *addr++ = caller;
1656 *addr++ = smp_processor_id();
1657 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 {
1659 unsigned long *sptr = &caller;
1660 unsigned long svalue;
1661
1662 while (!kstack_end(sptr)) {
1663 svalue = *sptr++;
1664 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001665 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 size -= sizeof(unsigned long);
1667 if (size <= sizeof(unsigned long))
1668 break;
1669 }
1670 }
1671
1672 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001673 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674}
1675#endif
1676
Pekka Enberg343e0d72006-02-01 03:05:50 -08001677static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001679 int size = obj_size(cachep);
1680 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
1682 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001683 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684}
1685
1686static void dump_line(char *data, int offset, int limit)
1687{
1688 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001689 unsigned char error = 0;
1690 int bad_count = 0;
1691
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001693 for (i = 0; i < limit; i++) {
1694 if (data[offset + i] != POISON_FREE) {
1695 error = data[offset + i];
1696 bad_count++;
1697 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001698 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001699 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001701
1702 if (bad_count == 1) {
1703 error ^= POISON_FREE;
1704 if (!(error & (error - 1))) {
1705 printk(KERN_ERR "Single bit error detected. Probably "
1706 "bad RAM.\n");
1707#ifdef CONFIG_X86
1708 printk(KERN_ERR "Run memtest86+ or a similar memory "
1709 "test tool.\n");
1710#else
1711 printk(KERN_ERR "Run a memory test tool.\n");
1712#endif
1713 }
1714 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715}
1716#endif
1717
1718#if DEBUG
1719
Pekka Enberg343e0d72006-02-01 03:05:50 -08001720static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721{
1722 int i, size;
1723 char *realobj;
1724
1725 if (cachep->flags & SLAB_RED_ZONE) {
1726 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001727 *dbg_redzone1(cachep, objp),
1728 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
1730
1731 if (cachep->flags & SLAB_STORE_USER) {
1732 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001733 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001735 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 printk("\n");
1737 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001738 realobj = (char *)objp + obj_offset(cachep);
1739 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001740 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 int limit;
1742 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001743 if (i + limit > size)
1744 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 dump_line(realobj, i, limit);
1746 }
1747}
1748
Pekka Enberg343e0d72006-02-01 03:05:50 -08001749static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750{
1751 char *realobj;
1752 int size, i;
1753 int lines = 0;
1754
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001755 realobj = (char *)objp + obj_offset(cachep);
1756 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001758 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001760 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 exp = POISON_END;
1762 if (realobj[i] != exp) {
1763 int limit;
1764 /* Mismatch ! */
1765 /* Print header */
1766 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001767 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001768 "Slab corruption: start=%p, len=%d\n",
1769 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 print_objinfo(cachep, objp, 0);
1771 }
1772 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001773 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001775 if (i + limit > size)
1776 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 dump_line(realobj, i, limit);
1778 i += 16;
1779 lines++;
1780 /* Limit to 5 lines */
1781 if (lines > 5)
1782 break;
1783 }
1784 }
1785 if (lines != 0) {
1786 /* Print some data about the neighboring objects, if they
1787 * exist:
1788 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001789 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001790 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001792 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001794 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001795 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001797 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798 print_objinfo(cachep, objp, 2);
1799 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001801 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001802 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001804 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 print_objinfo(cachep, objp, 2);
1806 }
1807 }
1808}
1809#endif
1810
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001812/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001813 * slab_destroy_objs - destroy a slab and its objects
1814 * @cachep: cache pointer being destroyed
1815 * @slabp: slab pointer being destroyed
1816 *
1817 * Call the registered destructor for each object in a slab that is being
1818 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001819 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001820static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001821{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 int i;
1823 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001824 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
1826 if (cachep->flags & SLAB_POISON) {
1827#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001828 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1829 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001830 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001831 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 else
1833 check_poison_obj(cachep, objp);
1834#else
1835 check_poison_obj(cachep, objp);
1836#endif
1837 }
1838 if (cachep->flags & SLAB_RED_ZONE) {
1839 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1840 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001841 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1843 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001844 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 }
1846 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001847 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001849}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001851static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001852{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 if (cachep->dtor) {
1854 int i;
1855 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001856 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001857 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 }
1859 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001860}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861#endif
1862
Randy Dunlap911851e2006-03-22 00:08:14 -08001863/**
1864 * slab_destroy - destroy and release all objects in a slab
1865 * @cachep: cache pointer being destroyed
1866 * @slabp: slab pointer being destroyed
1867 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001868 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001869 * Before calling the slab must have been unlinked from the cache. The
1870 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001871 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001872static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001873{
1874 void *addr = slabp->s_mem - slabp->colouroff;
1875
1876 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1878 struct slab_rcu *slab_rcu;
1879
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001880 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 slab_rcu->cachep = cachep;
1882 slab_rcu->addr = addr;
1883 call_rcu(&slab_rcu->head, kmem_rcu_free);
1884 } else {
1885 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001886 if (OFF_SLAB(cachep))
1887 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 }
1889}
1890
Andrew Mortona737b3e2006-03-22 00:08:11 -08001891/*
1892 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1893 * size of kmem_list3.
1894 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001895static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001896{
1897 int node;
1898
1899 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001900 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001901 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001902 REAPTIMEOUT_LIST3 +
1903 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001904 }
1905}
1906
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001907static void __kmem_cache_destroy(struct kmem_cache *cachep)
1908{
1909 int i;
1910 struct kmem_list3 *l3;
1911
1912 for_each_online_cpu(i)
1913 kfree(cachep->array[i]);
1914
1915 /* NUMA: free the list3 structures */
1916 for_each_online_node(i) {
1917 l3 = cachep->nodelists[i];
1918 if (l3) {
1919 kfree(l3->shared);
1920 free_alien_cache(l3->alien);
1921 kfree(l3);
1922 }
1923 }
1924 kmem_cache_free(&cache_cache, cachep);
1925}
1926
1927
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001929 * calculate_slab_order - calculate size (page order) of slabs
1930 * @cachep: pointer to the cache that is being created
1931 * @size: size of objects to be created in this cache.
1932 * @align: required alignment for the objects.
1933 * @flags: slab allocation flags
1934 *
1935 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001936 *
1937 * This could be made much more intelligent. For now, try to avoid using
1938 * high order pages for slabs. When the gfp() functions are more friendly
1939 * towards high-order requests, this should be changed.
1940 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001941static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001942 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001943{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001944 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001945 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001946 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001947
Andrew Mortona737b3e2006-03-22 00:08:11 -08001948 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001949 unsigned int num;
1950 size_t remainder;
1951
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001952 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001953 if (!num)
1954 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001955
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001956 if (flags & CFLGS_OFF_SLAB) {
1957 /*
1958 * Max number of objs-per-slab for caches which
1959 * use off-slab slabs. Needed to avoid a possible
1960 * looping condition in cache_grow().
1961 */
1962 offslab_limit = size - sizeof(struct slab);
1963 offslab_limit /= sizeof(kmem_bufctl_t);
1964
1965 if (num > offslab_limit)
1966 break;
1967 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001968
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001969 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001970 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001971 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001972 left_over = remainder;
1973
1974 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001975 * A VFS-reclaimable slab tends to have most allocations
1976 * as GFP_NOFS and we really don't want to have to be allocating
1977 * higher-order pages when we are unable to shrink dcache.
1978 */
1979 if (flags & SLAB_RECLAIM_ACCOUNT)
1980 break;
1981
1982 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001983 * Large number of objects is good, but very large slabs are
1984 * currently bad for the gfp()s.
1985 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001986 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001987 break;
1988
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001989 /*
1990 * Acceptable internal fragmentation?
1991 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001992 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001993 break;
1994 }
1995 return left_over;
1996}
1997
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001998static int setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001999{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002000 if (g_cpucache_up == FULL)
2001 return enable_cpucache(cachep);
2002
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002003 if (g_cpucache_up == NONE) {
2004 /*
2005 * Note: the first kmem_cache_create must create the cache
2006 * that's used by kmalloc(24), otherwise the creation of
2007 * further caches will BUG().
2008 */
2009 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2010
2011 /*
2012 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2013 * the first cache, then we need to set up all its list3s,
2014 * otherwise the creation of further caches will BUG().
2015 */
2016 set_up_list3s(cachep, SIZE_AC);
2017 if (INDEX_AC == INDEX_L3)
2018 g_cpucache_up = PARTIAL_L3;
2019 else
2020 g_cpucache_up = PARTIAL_AC;
2021 } else {
2022 cachep->array[smp_processor_id()] =
2023 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2024
2025 if (g_cpucache_up == PARTIAL_AC) {
2026 set_up_list3s(cachep, SIZE_L3);
2027 g_cpucache_up = PARTIAL_L3;
2028 } else {
2029 int node;
2030 for_each_online_node(node) {
2031 cachep->nodelists[node] =
2032 kmalloc_node(sizeof(struct kmem_list3),
2033 GFP_KERNEL, node);
2034 BUG_ON(!cachep->nodelists[node]);
2035 kmem_list3_init(cachep->nodelists[node]);
2036 }
2037 }
2038 }
2039 cachep->nodelists[numa_node_id()]->next_reap =
2040 jiffies + REAPTIMEOUT_LIST3 +
2041 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2042
2043 cpu_cache_get(cachep)->avail = 0;
2044 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2045 cpu_cache_get(cachep)->batchcount = 1;
2046 cpu_cache_get(cachep)->touched = 0;
2047 cachep->batchcount = 1;
2048 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002049 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002050}
2051
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002052/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 * kmem_cache_create - Create a cache.
2054 * @name: A string which is used in /proc/slabinfo to identify this cache.
2055 * @size: The size of objects to be created in this cache.
2056 * @align: The required alignment for the objects.
2057 * @flags: SLAB flags
2058 * @ctor: A constructor for the objects.
2059 * @dtor: A destructor for the objects.
2060 *
2061 * Returns a ptr to the cache on success, NULL on failure.
2062 * Cannot be called within a int, but can be interrupted.
2063 * The @ctor is run when new pages are allocated by the cache
2064 * and the @dtor is run before the pages are handed back.
2065 *
2066 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002067 * the module calling this has to destroy the cache before getting unloaded.
2068 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 * The flags are
2070 *
2071 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2072 * to catch references to uninitialised memory.
2073 *
2074 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2075 * for buffer overruns.
2076 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2078 * cacheline. This can be beneficial if you're counting cycles as closely
2079 * as davem.
2080 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002081struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002083 unsigned long flags,
2084 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002085 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086{
2087 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002088 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089
2090 /*
2091 * Sanity checks... these are all serious usage bugs.
2092 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002093 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002094 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002095 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2096 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002097 BUG();
2098 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002100 /*
2101 * Prevent CPUs from coming and going.
2102 * lock_cpu_hotplug() nests outside cache_chain_mutex
2103 */
2104 lock_cpu_hotplug();
2105
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002106 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002107
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002108 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002109 mm_segment_t old_fs = get_fs();
2110 char tmp;
2111 int res;
2112
2113 /*
2114 * This happens when the module gets unloaded and doesn't
2115 * destroy its slab cache and no-one else reuses the vmalloc
2116 * area of the module. Print a warning.
2117 */
2118 set_fs(KERNEL_DS);
2119 res = __get_user(tmp, pc->name);
2120 set_fs(old_fs);
2121 if (res) {
2122 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002123 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002124 continue;
2125 }
2126
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002127 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002128 printk("kmem_cache_create: duplicate cache %s\n", name);
2129 dump_stack();
2130 goto oops;
2131 }
2132 }
2133
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134#if DEBUG
2135 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2136 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2137 /* No constructor, but inital state check requested */
2138 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002139 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 flags &= ~SLAB_DEBUG_INITIAL;
2141 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142#if FORCED_DEBUG
2143 /*
2144 * Enable redzoning and last user accounting, except for caches with
2145 * large objects, if the increased size would increase the object size
2146 * above the next power of two: caches with object sizes just above a
2147 * power of two have a significant amount of internal fragmentation.
2148 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002149 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002150 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 if (!(flags & SLAB_DESTROY_BY_RCU))
2152 flags |= SLAB_POISON;
2153#endif
2154 if (flags & SLAB_DESTROY_BY_RCU)
2155 BUG_ON(flags & SLAB_POISON);
2156#endif
2157 if (flags & SLAB_DESTROY_BY_RCU)
2158 BUG_ON(dtor);
2159
2160 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002161 * Always checks flags, a caller might be expecting debug support which
2162 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002164 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
Andrew Mortona737b3e2006-03-22 00:08:11 -08002166 /*
2167 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 * unaligned accesses for some archs when redzoning is used, and makes
2169 * sure any on-slab bufctl's are also correctly aligned.
2170 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002171 if (size & (BYTES_PER_WORD - 1)) {
2172 size += (BYTES_PER_WORD - 1);
2173 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 }
2175
Andrew Mortona737b3e2006-03-22 00:08:11 -08002176 /* calculate the final buffer alignment: */
2177
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 /* 1) arch recommendation: can be overridden for debug */
2179 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002180 /*
2181 * Default alignment: as specified by the arch code. Except if
2182 * an object is really small, then squeeze multiple objects into
2183 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 */
2185 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002186 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 ralign /= 2;
2188 } else {
2189 ralign = BYTES_PER_WORD;
2190 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002191
2192 /*
2193 * Redzoning and user store require word alignment. Note this will be
2194 * overridden by architecture or caller mandated alignment if either
2195 * is greater than BYTES_PER_WORD.
2196 */
2197 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2198 ralign = BYTES_PER_WORD;
2199
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 /* 2) arch mandated alignment: disables debug if necessary */
2201 if (ralign < ARCH_SLAB_MINALIGN) {
2202 ralign = ARCH_SLAB_MINALIGN;
2203 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002204 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 }
2206 /* 3) caller mandated alignment: disables debug if necessary */
2207 if (ralign < align) {
2208 ralign = align;
2209 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002210 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002212 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002213 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 */
2215 align = ralign;
2216
2217 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002218 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002220 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
2222#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002223 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224
Pekka Enbergca5f9702006-09-25 23:31:25 -07002225 /*
2226 * Both debugging options require word-alignment which is calculated
2227 * into align above.
2228 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002231 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002232 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 }
2234 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002235 /* user store requires one word storage behind the end of
2236 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 size += BYTES_PER_WORD;
2239 }
2240#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002241 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002242 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2243 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 size = PAGE_SIZE;
2245 }
2246#endif
2247#endif
2248
Ingo Molnare0a42722006-06-23 02:03:46 -07002249 /*
2250 * Determine if the slab management is 'on' or 'off' slab.
2251 * (bootstrapping cannot cope with offslab caches so don't do
2252 * it too early on.)
2253 */
2254 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 /*
2256 * Size is large, assume best to place the slab management obj
2257 * off-slab (should allow better packing of objs).
2258 */
2259 flags |= CFLGS_OFF_SLAB;
2260
2261 size = ALIGN(size, align);
2262
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002263 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264
2265 if (!cachep->num) {
2266 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2267 kmem_cache_free(&cache_cache, cachep);
2268 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002269 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002271 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2272 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273
2274 /*
2275 * If the slab has been placed off-slab, and we have enough space then
2276 * move it on-slab. This is at the expense of any extra colouring.
2277 */
2278 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2279 flags &= ~CFLGS_OFF_SLAB;
2280 left_over -= slab_size;
2281 }
2282
2283 if (flags & CFLGS_OFF_SLAB) {
2284 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002285 slab_size =
2286 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 }
2288
2289 cachep->colour_off = cache_line_size();
2290 /* Offset must be a multiple of the alignment. */
2291 if (cachep->colour_off < align)
2292 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002293 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 cachep->slab_size = slab_size;
2295 cachep->flags = flags;
2296 cachep->gfpflags = 0;
2297 if (flags & SLAB_CACHE_DMA)
2298 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002299 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002301 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002302 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002303 /*
2304 * This is a possibility for one of the malloc_sizes caches.
2305 * But since we go off slab only for object size greater than
2306 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2307 * this should not happen at all.
2308 * But leave a BUG_ON for some lucky dude.
2309 */
2310 BUG_ON(!cachep->slabp_cache);
2311 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 cachep->ctor = ctor;
2313 cachep->dtor = dtor;
2314 cachep->name = name;
2315
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002316 if (setup_cpu_cache(cachep)) {
2317 __kmem_cache_destroy(cachep);
2318 cachep = NULL;
2319 goto oops;
2320 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 /* cache setup completed, link it into the list */
2323 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002324oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 if (!cachep && (flags & SLAB_PANIC))
2326 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002327 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002328 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002329 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 return cachep;
2331}
2332EXPORT_SYMBOL(kmem_cache_create);
2333
2334#if DEBUG
2335static void check_irq_off(void)
2336{
2337 BUG_ON(!irqs_disabled());
2338}
2339
2340static void check_irq_on(void)
2341{
2342 BUG_ON(irqs_disabled());
2343}
2344
Pekka Enberg343e0d72006-02-01 03:05:50 -08002345static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346{
2347#ifdef CONFIG_SMP
2348 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002349 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350#endif
2351}
Christoph Lametere498be72005-09-09 13:03:32 -07002352
Pekka Enberg343e0d72006-02-01 03:05:50 -08002353static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002354{
2355#ifdef CONFIG_SMP
2356 check_irq_off();
2357 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2358#endif
2359}
2360
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361#else
2362#define check_irq_off() do { } while(0)
2363#define check_irq_on() do { } while(0)
2364#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002365#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366#endif
2367
Christoph Lameteraab22072006-03-22 00:09:06 -08002368static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2369 struct array_cache *ac,
2370 int force, int node);
2371
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372static void do_drain(void *arg)
2373{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002374 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002376 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377
2378 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002379 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002380 spin_lock(&cachep->nodelists[node]->list_lock);
2381 free_block(cachep, ac->entry, ac->avail, node);
2382 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 ac->avail = 0;
2384}
2385
Pekka Enberg343e0d72006-02-01 03:05:50 -08002386static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387{
Christoph Lametere498be72005-09-09 13:03:32 -07002388 struct kmem_list3 *l3;
2389 int node;
2390
Andrew Mortona07fa392006-03-22 00:08:17 -08002391 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002393 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002394 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002395 if (l3 && l3->alien)
2396 drain_alien_cache(cachep, l3->alien);
2397 }
2398
2399 for_each_online_node(node) {
2400 l3 = cachep->nodelists[node];
2401 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002402 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002403 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404}
2405
Christoph Lametered11d9e2006-06-30 01:55:45 -07002406/*
2407 * Remove slabs from the list of free slabs.
2408 * Specify the number of slabs to drain in tofree.
2409 *
2410 * Returns the actual number of slabs released.
2411 */
2412static int drain_freelist(struct kmem_cache *cache,
2413 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002415 struct list_head *p;
2416 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418
Christoph Lametered11d9e2006-06-30 01:55:45 -07002419 nr_freed = 0;
2420 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421
Christoph Lametered11d9e2006-06-30 01:55:45 -07002422 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002423 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002424 if (p == &l3->slabs_free) {
2425 spin_unlock_irq(&l3->list_lock);
2426 goto out;
2427 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428
Christoph Lametered11d9e2006-06-30 01:55:45 -07002429 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002431 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432#endif
2433 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002434 /*
2435 * Safe to drop the lock. The slab is no longer linked
2436 * to the cache.
2437 */
2438 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002439 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002440 slab_destroy(cache, slabp);
2441 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002443out:
2444 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445}
2446
Pekka Enberg343e0d72006-02-01 03:05:50 -08002447static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002448{
2449 int ret = 0, i = 0;
2450 struct kmem_list3 *l3;
2451
2452 drain_cpu_caches(cachep);
2453
2454 check_irq_on();
2455 for_each_online_node(i) {
2456 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002457 if (!l3)
2458 continue;
2459
2460 drain_freelist(cachep, l3, l3->free_objects);
2461
2462 ret += !list_empty(&l3->slabs_full) ||
2463 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002464 }
2465 return (ret ? 1 : 0);
2466}
2467
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468/**
2469 * kmem_cache_shrink - Shrink a cache.
2470 * @cachep: The cache to shrink.
2471 *
2472 * Releases as many slabs as possible for a cache.
2473 * To help debugging, a zero exit status indicates all slabs were released.
2474 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002475int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002477 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478
2479 return __cache_shrink(cachep);
2480}
2481EXPORT_SYMBOL(kmem_cache_shrink);
2482
2483/**
2484 * kmem_cache_destroy - delete a cache
2485 * @cachep: the cache to destroy
2486 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002487 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 *
2489 * It is expected this function will be called by a module when it is
2490 * unloaded. This will remove the cache completely, and avoid a duplicate
2491 * cache being allocated each time a module is loaded and unloaded, if the
2492 * module doesn't have persistent in-kernel storage across loads and unloads.
2493 *
2494 * The cache must be empty before calling this function.
2495 *
2496 * The caller must guarantee that noone will allocate memory from the cache
2497 * during the kmem_cache_destroy().
2498 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002499void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002501 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502
2503 /* Don't let CPUs to come and go */
2504 lock_cpu_hotplug();
2505
2506 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002507 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 /*
2509 * the chain is never empty, cache_cache is never destroyed
2510 */
2511 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002512 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513
2514 if (__cache_shrink(cachep)) {
2515 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002516 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002517 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002518 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 unlock_cpu_hotplug();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002520 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 }
2522
2523 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002524 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002526 __kmem_cache_destroy(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528}
2529EXPORT_SYMBOL(kmem_cache_destroy);
2530
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002531/*
2532 * Get the memory for a slab management obj.
2533 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2534 * always come from malloc_sizes caches. The slab descriptor cannot
2535 * come from the same cache which is getting created because,
2536 * when we are searching for an appropriate cache for these
2537 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2538 * If we are creating a malloc_sizes cache here it would not be visible to
2539 * kmem_find_general_cachep till the initialization is complete.
2540 * Hence we cannot have slabp_cache same as the original cache.
2541 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002542static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002543 int colour_off, gfp_t local_flags,
2544 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545{
2546 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002547
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 if (OFF_SLAB(cachep)) {
2549 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002550 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2551 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 if (!slabp)
2553 return NULL;
2554 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002555 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 colour_off += cachep->slab_size;
2557 }
2558 slabp->inuse = 0;
2559 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002560 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002561 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 return slabp;
2563}
2564
2565static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2566{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002567 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568}
2569
Pekka Enberg343e0d72006-02-01 03:05:50 -08002570static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002571 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572{
2573 int i;
2574
2575 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002576 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577#if DEBUG
2578 /* need to poison the objs? */
2579 if (cachep->flags & SLAB_POISON)
2580 poison_obj(cachep, objp, POISON_FREE);
2581 if (cachep->flags & SLAB_STORE_USER)
2582 *dbg_userword(cachep, objp) = NULL;
2583
2584 if (cachep->flags & SLAB_RED_ZONE) {
2585 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2586 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2587 }
2588 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002589 * Constructors are not allowed to allocate memory from the same
2590 * cache which they are a constructor for. Otherwise, deadlock.
2591 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 */
2593 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002594 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002595 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596
2597 if (cachep->flags & SLAB_RED_ZONE) {
2598 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2599 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002600 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2602 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002603 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002605 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2606 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002607 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002608 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609#else
2610 if (cachep->ctor)
2611 cachep->ctor(objp, cachep, ctor_flags);
2612#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002613 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002615 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 slabp->free = 0;
2617}
2618
Pekka Enberg343e0d72006-02-01 03:05:50 -08002619static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002621 if (flags & SLAB_DMA)
2622 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2623 else
2624 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625}
2626
Andrew Mortona737b3e2006-03-22 00:08:11 -08002627static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2628 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002629{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002630 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002631 kmem_bufctl_t next;
2632
2633 slabp->inuse++;
2634 next = slab_bufctl(slabp)[slabp->free];
2635#if DEBUG
2636 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2637 WARN_ON(slabp->nodeid != nodeid);
2638#endif
2639 slabp->free = next;
2640
2641 return objp;
2642}
2643
Andrew Mortona737b3e2006-03-22 00:08:11 -08002644static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2645 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002646{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002647 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002648
2649#if DEBUG
2650 /* Verify that the slab belongs to the intended node */
2651 WARN_ON(slabp->nodeid != nodeid);
2652
Al Viro871751e2006-03-25 03:06:39 -08002653 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002654 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002655 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002656 BUG();
2657 }
2658#endif
2659 slab_bufctl(slabp)[objnr] = slabp->free;
2660 slabp->free = objnr;
2661 slabp->inuse--;
2662}
2663
Pekka Enberg47768742006-06-23 02:03:07 -07002664/*
2665 * Map pages beginning at addr to the given cache and slab. This is required
2666 * for the slab allocator to be able to lookup the cache and slab of a
2667 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2668 */
2669static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2670 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671{
Pekka Enberg47768742006-06-23 02:03:07 -07002672 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 struct page *page;
2674
Pekka Enberg47768742006-06-23 02:03:07 -07002675 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002676
Pekka Enberg47768742006-06-23 02:03:07 -07002677 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002678 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002679 nr_pages <<= cache->gfporder;
2680
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002682 page_set_cache(page, cache);
2683 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002685 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686}
2687
2688/*
2689 * Grow (by 1) the number of slabs within a cache. This is called by
2690 * kmem_cache_alloc() when there are no active objs left in a cache.
2691 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002692static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002694 struct slab *slabp;
2695 void *objp;
2696 size_t offset;
2697 gfp_t local_flags;
2698 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002699 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700
Andrew Mortona737b3e2006-03-22 00:08:11 -08002701 /*
2702 * Be lazy and only check for valid flags here, keeping it out of the
2703 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002705 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 if (flags & SLAB_NO_GROW)
2707 return 0;
2708
2709 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2710 local_flags = (flags & SLAB_LEVEL_MASK);
2711 if (!(local_flags & __GFP_WAIT))
2712 /*
2713 * Not allowed to sleep. Need to tell a constructor about
2714 * this - it might need to know...
2715 */
2716 ctor_flags |= SLAB_CTOR_ATOMIC;
2717
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002718 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002720 l3 = cachep->nodelists[nodeid];
2721 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
2723 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002724 offset = l3->colour_next;
2725 l3->colour_next++;
2726 if (l3->colour_next >= cachep->colour)
2727 l3->colour_next = 0;
2728 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002730 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731
2732 if (local_flags & __GFP_WAIT)
2733 local_irq_enable();
2734
2735 /*
2736 * The test for missing atomic flag is performed here, rather than
2737 * the more obvious place, simply to reduce the critical path length
2738 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2739 * will eventually be caught here (where it matters).
2740 */
2741 kmem_flagcheck(cachep, flags);
2742
Andrew Mortona737b3e2006-03-22 00:08:11 -08002743 /*
2744 * Get mem for the objs. Attempt to allocate a physical page from
2745 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002746 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002747 objp = kmem_getpages(cachep, flags, nodeid);
2748 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 goto failed;
2750
2751 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002752 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002753 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 goto opps1;
2755
Christoph Lametere498be72005-09-09 13:03:32 -07002756 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002757 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758
2759 cache_init_objs(cachep, slabp, ctor_flags);
2760
2761 if (local_flags & __GFP_WAIT)
2762 local_irq_disable();
2763 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002764 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
2766 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002767 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002769 l3->free_objects += cachep->num;
2770 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002772opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002774failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 if (local_flags & __GFP_WAIT)
2776 local_irq_disable();
2777 return 0;
2778}
2779
2780#if DEBUG
2781
2782/*
2783 * Perform extra freeing checks:
2784 * - detect bad pointers.
2785 * - POISON/RED_ZONE checking
2786 * - destructor calls, for caches with POISON+dtor
2787 */
2788static void kfree_debugcheck(const void *objp)
2789{
2790 struct page *page;
2791
2792 if (!virt_addr_valid(objp)) {
2793 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002794 (unsigned long)objp);
2795 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 }
2797 page = virt_to_page(objp);
2798 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002799 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2800 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 BUG();
2802 }
2803}
2804
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002805static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2806{
2807 unsigned long redzone1, redzone2;
2808
2809 redzone1 = *dbg_redzone1(cache, obj);
2810 redzone2 = *dbg_redzone2(cache, obj);
2811
2812 /*
2813 * Redzone is ok.
2814 */
2815 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2816 return;
2817
2818 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2819 slab_error(cache, "double free detected");
2820 else
2821 slab_error(cache, "memory outside object was overwritten");
2822
2823 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2824 obj, redzone1, redzone2);
2825}
2826
Pekka Enberg343e0d72006-02-01 03:05:50 -08002827static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002828 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829{
2830 struct page *page;
2831 unsigned int objnr;
2832 struct slab *slabp;
2833
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002834 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 kfree_debugcheck(objp);
2836 page = virt_to_page(objp);
2837
Pekka Enberg065d41c2005-11-13 16:06:46 -08002838 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839
2840 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002841 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2843 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2844 }
2845 if (cachep->flags & SLAB_STORE_USER)
2846 *dbg_userword(cachep, objp) = caller;
2847
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002848 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849
2850 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002851 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852
2853 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002854 /*
2855 * Need to call the slab's constructor so the caller can
2856 * perform a verify of its state (debugging). Called without
2857 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002859 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002860 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 }
2862 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2863 /* we want to cache poison the object,
2864 * call the destruction callback
2865 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002866 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 }
Al Viro871751e2006-03-25 03:06:39 -08002868#ifdef CONFIG_DEBUG_SLAB_LEAK
2869 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2870#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 if (cachep->flags & SLAB_POISON) {
2872#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002873 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002875 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002876 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 } else {
2878 poison_obj(cachep, objp, POISON_FREE);
2879 }
2880#else
2881 poison_obj(cachep, objp, POISON_FREE);
2882#endif
2883 }
2884 return objp;
2885}
2886
Pekka Enberg343e0d72006-02-01 03:05:50 -08002887static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888{
2889 kmem_bufctl_t i;
2890 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002891
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 /* Check slab's freelist to see if this obj is there. */
2893 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2894 entries++;
2895 if (entries > cachep->num || i >= cachep->num)
2896 goto bad;
2897 }
2898 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002899bad:
2900 printk(KERN_ERR "slab: Internal list corruption detected in "
2901 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2902 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002903 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002904 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002905 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002906 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002908 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 }
2910 printk("\n");
2911 BUG();
2912 }
2913}
2914#else
2915#define kfree_debugcheck(x) do { } while(0)
2916#define cache_free_debugcheck(x,objp,z) (objp)
2917#define check_slabp(x,y) do { } while(0)
2918#endif
2919
Pekka Enberg343e0d72006-02-01 03:05:50 -08002920static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921{
2922 int batchcount;
2923 struct kmem_list3 *l3;
2924 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002925 int node;
2926
2927 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928
2929 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002930 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002931retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 batchcount = ac->batchcount;
2933 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002934 /*
2935 * If there was little recent activity on this cache, then
2936 * perform only a partial refill. Otherwise we could generate
2937 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 */
2939 batchcount = BATCHREFILL_LIMIT;
2940 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002941 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942
Christoph Lametere498be72005-09-09 13:03:32 -07002943 BUG_ON(ac->avail > 0 || !l3);
2944 spin_lock(&l3->list_lock);
2945
Christoph Lameter3ded1752006-03-25 03:06:44 -08002946 /* See if we can refill from the shared array */
2947 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2948 goto alloc_done;
2949
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 while (batchcount > 0) {
2951 struct list_head *entry;
2952 struct slab *slabp;
2953 /* Get slab alloc is to come from. */
2954 entry = l3->slabs_partial.next;
2955 if (entry == &l3->slabs_partial) {
2956 l3->free_touched = 1;
2957 entry = l3->slabs_free.next;
2958 if (entry == &l3->slabs_free)
2959 goto must_grow;
2960 }
2961
2962 slabp = list_entry(entry, struct slab, list);
2963 check_slabp(cachep, slabp);
2964 check_spinlock_acquired(cachep);
2965 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966 STATS_INC_ALLOCED(cachep);
2967 STATS_INC_ACTIVE(cachep);
2968 STATS_SET_HIGH(cachep);
2969
Matthew Dobson78d382d2006-02-01 03:05:47 -08002970 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002971 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 }
2973 check_slabp(cachep, slabp);
2974
2975 /* move slabp to correct slabp list: */
2976 list_del(&slabp->list);
2977 if (slabp->free == BUFCTL_END)
2978 list_add(&slabp->list, &l3->slabs_full);
2979 else
2980 list_add(&slabp->list, &l3->slabs_partial);
2981 }
2982
Andrew Mortona737b3e2006-03-22 00:08:11 -08002983must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002985alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002986 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987
2988 if (unlikely(!ac->avail)) {
2989 int x;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002990 x = cache_grow(cachep, flags, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002991
Andrew Mortona737b3e2006-03-22 00:08:11 -08002992 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002993 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002994 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 return NULL;
2996
Andrew Mortona737b3e2006-03-22 00:08:11 -08002997 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 goto retry;
2999 }
3000 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003001 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002}
3003
Andrew Mortona737b3e2006-03-22 00:08:11 -08003004static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3005 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006{
3007 might_sleep_if(flags & __GFP_WAIT);
3008#if DEBUG
3009 kmem_flagcheck(cachep, flags);
3010#endif
3011}
3012
3013#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003014static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3015 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003017 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003019 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003021 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003022 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003023 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 else
3025 check_poison_obj(cachep, objp);
3026#else
3027 check_poison_obj(cachep, objp);
3028#endif
3029 poison_obj(cachep, objp, POISON_INUSE);
3030 }
3031 if (cachep->flags & SLAB_STORE_USER)
3032 *dbg_userword(cachep, objp) = caller;
3033
3034 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003035 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3036 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3037 slab_error(cachep, "double free, or memory outside"
3038 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003039 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08003040 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3041 objp, *dbg_redzone1(cachep, objp),
3042 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 }
3044 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3045 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3046 }
Al Viro871751e2006-03-25 03:06:39 -08003047#ifdef CONFIG_DEBUG_SLAB_LEAK
3048 {
3049 struct slab *slabp;
3050 unsigned objnr;
3051
3052 slabp = page_get_slab(virt_to_page(objp));
3053 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3054 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3055 }
3056#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003057 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003059 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060
3061 if (!(flags & __GFP_WAIT))
3062 ctor_flags |= SLAB_CTOR_ATOMIC;
3063
3064 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003065 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 return objp;
3067}
3068#else
3069#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3070#endif
3071
Pekka Enberg343e0d72006-02-01 03:05:50 -08003072static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003074 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 struct array_cache *ac;
3076
Alok N Kataria5c382302005-09-27 21:45:46 -07003077 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003078 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 if (likely(ac->avail)) {
3080 STATS_INC_ALLOCHIT(cachep);
3081 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003082 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 } else {
3084 STATS_INC_ALLOCMISS(cachep);
3085 objp = cache_alloc_refill(cachep, flags);
3086 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003087 return objp;
3088}
3089
Andrew Mortona737b3e2006-03-22 00:08:11 -08003090static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3091 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07003092{
3093 unsigned long save_flags;
Christoph Lameterde3083e2006-09-27 01:50:03 -07003094 void *objp = NULL;
Alok N Kataria5c382302005-09-27 21:45:46 -07003095
3096 cache_alloc_debugcheck_before(cachep, flags);
3097
3098 local_irq_save(save_flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003099
Christoph Lameter765c4502006-09-27 01:50:08 -07003100 if (unlikely(NUMA_BUILD &&
3101 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
Christoph Lameterde3083e2006-09-27 01:50:03 -07003102 objp = alternate_node_alloc(cachep, flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003103
3104 if (!objp)
3105 objp = ____cache_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003106 /*
3107 * We may just have run out of memory on the local node.
3108 * __cache_alloc_node() knows how to locate memory on other nodes
3109 */
3110 if (NUMA_BUILD && !objp)
3111 objp = __cache_alloc_node(cachep, flags, numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07003113 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003114 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07003115 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 return objp;
3117}
3118
Christoph Lametere498be72005-09-09 13:03:32 -07003119#ifdef CONFIG_NUMA
3120/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003121 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003122 *
3123 * If we are in_interrupt, then process context, including cpusets and
3124 * mempolicy, may not apply and should not be used for allocation policy.
3125 */
3126static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3127{
3128 int nid_alloc, nid_here;
3129
Christoph Lameter765c4502006-09-27 01:50:08 -07003130 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003131 return NULL;
3132 nid_alloc = nid_here = numa_node_id();
3133 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3134 nid_alloc = cpuset_mem_spread_node();
3135 else if (current->mempolicy)
3136 nid_alloc = slab_node(current->mempolicy);
3137 if (nid_alloc != nid_here)
3138 return __cache_alloc_node(cachep, flags, nid_alloc);
3139 return NULL;
3140}
3141
3142/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003143 * Fallback function if there was no memory available and no objects on a
3144 * certain node and we are allowed to fall back. We mimick the behavior of
3145 * the page allocator. We fall back according to a zonelist determined by
3146 * the policy layer while obeying cpuset constraints.
3147 */
3148void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3149{
3150 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3151 ->node_zonelists[gfp_zone(flags)];
3152 struct zone **z;
3153 void *obj = NULL;
3154
3155 for (z = zonelist->zones; *z && !obj; z++)
3156 if (zone_idx(*z) <= ZONE_NORMAL &&
3157 cpuset_zone_allowed(*z, flags))
3158 obj = __cache_alloc_node(cache,
3159 flags | __GFP_THISNODE,
3160 zone_to_nid(*z));
3161 return obj;
3162}
3163
3164/*
Christoph Lametere498be72005-09-09 13:03:32 -07003165 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003167static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3168 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003169{
3170 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003171 struct slab *slabp;
3172 struct kmem_list3 *l3;
3173 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003174 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003176 l3 = cachep->nodelists[nodeid];
3177 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003178
Andrew Mortona737b3e2006-03-22 00:08:11 -08003179retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003180 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003181 spin_lock(&l3->list_lock);
3182 entry = l3->slabs_partial.next;
3183 if (entry == &l3->slabs_partial) {
3184 l3->free_touched = 1;
3185 entry = l3->slabs_free.next;
3186 if (entry == &l3->slabs_free)
3187 goto must_grow;
3188 }
Christoph Lametere498be72005-09-09 13:03:32 -07003189
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003190 slabp = list_entry(entry, struct slab, list);
3191 check_spinlock_acquired_node(cachep, nodeid);
3192 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003193
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003194 STATS_INC_NODEALLOCS(cachep);
3195 STATS_INC_ACTIVE(cachep);
3196 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003197
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003198 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003199
Matthew Dobson78d382d2006-02-01 03:05:47 -08003200 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003201 check_slabp(cachep, slabp);
3202 l3->free_objects--;
3203 /* move slabp to correct slabp list: */
3204 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003205
Andrew Mortona737b3e2006-03-22 00:08:11 -08003206 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003207 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003208 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003209 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003210
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003211 spin_unlock(&l3->list_lock);
3212 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003213
Andrew Mortona737b3e2006-03-22 00:08:11 -08003214must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003215 spin_unlock(&l3->list_lock);
3216 x = cache_grow(cachep, flags, nodeid);
Christoph Lameter765c4502006-09-27 01:50:08 -07003217 if (x)
3218 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003219
Christoph Lameter765c4502006-09-27 01:50:08 -07003220 if (!(flags & __GFP_THISNODE))
3221 /* Unable to grow the cache. Fall back to other nodes. */
3222 return fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003223
Christoph Lameter765c4502006-09-27 01:50:08 -07003224 return NULL;
3225
Andrew Mortona737b3e2006-03-22 00:08:11 -08003226done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003227 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003228}
3229#endif
3230
3231/*
3232 * Caller needs to acquire correct kmem_list's list_lock
3233 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003234static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003235 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236{
3237 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003238 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239
3240 for (i = 0; i < nr_objects; i++) {
3241 void *objp = objpp[i];
3242 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003244 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003245 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003247 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003249 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003251 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 check_slabp(cachep, slabp);
3253
3254 /* fixup slab chains */
3255 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003256 if (l3->free_objects > l3->free_limit) {
3257 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003258 /* No need to drop any previously held
3259 * lock here, even if we have a off-slab slab
3260 * descriptor it is guaranteed to come from
3261 * a different cache, refer to comments before
3262 * alloc_slabmgmt.
3263 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 slab_destroy(cachep, slabp);
3265 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003266 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267 }
3268 } else {
3269 /* Unconditionally move a slab to the end of the
3270 * partial list on free - maximum time for the
3271 * other objects to be freed, too.
3272 */
Christoph Lametere498be72005-09-09 13:03:32 -07003273 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274 }
3275 }
3276}
3277
Pekka Enberg343e0d72006-02-01 03:05:50 -08003278static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279{
3280 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003281 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003282 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283
3284 batchcount = ac->batchcount;
3285#if DEBUG
3286 BUG_ON(!batchcount || batchcount > ac->avail);
3287#endif
3288 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003289 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003290 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003291 if (l3->shared) {
3292 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003293 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 if (max) {
3295 if (batchcount > max)
3296 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003297 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003298 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 shared_array->avail += batchcount;
3300 goto free_done;
3301 }
3302 }
3303
Christoph Lameterff694162005-09-22 21:44:02 -07003304 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003305free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306#if STATS
3307 {
3308 int i = 0;
3309 struct list_head *p;
3310
Christoph Lametere498be72005-09-09 13:03:32 -07003311 p = l3->slabs_free.next;
3312 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 struct slab *slabp;
3314
3315 slabp = list_entry(p, struct slab, list);
3316 BUG_ON(slabp->inuse);
3317
3318 i++;
3319 p = p->next;
3320 }
3321 STATS_SET_FREEABLE(cachep, i);
3322 }
3323#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003324 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003326 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327}
3328
3329/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003330 * Release an obj back to its cache. If the obj has a constructed state, it must
3331 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003333static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003335 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336
3337 check_irq_off();
3338 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3339
Ingo Molnar873623d2006-07-13 14:44:38 +02003340 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003341 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003342
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 if (likely(ac->avail < ac->limit)) {
3344 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003345 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 return;
3347 } else {
3348 STATS_INC_FREEMISS(cachep);
3349 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003350 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351 }
3352}
3353
3354/**
3355 * kmem_cache_alloc - Allocate an object
3356 * @cachep: The cache to allocate from.
3357 * @flags: See kmalloc().
3358 *
3359 * Allocate an object from this cache. The flags are only relevant
3360 * if the cache has no available objects.
3361 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003362void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003364 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365}
3366EXPORT_SYMBOL(kmem_cache_alloc);
3367
3368/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003369 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003370 * @cache: The cache to allocate from.
3371 * @flags: See kmalloc().
3372 *
3373 * Allocate an object from this cache and set the allocated memory to zero.
3374 * The flags are only relevant if the cache has no available objects.
3375 */
3376void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3377{
3378 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3379 if (ret)
3380 memset(ret, 0, obj_size(cache));
3381 return ret;
3382}
3383EXPORT_SYMBOL(kmem_cache_zalloc);
3384
3385/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386 * kmem_ptr_validate - check if an untrusted pointer might
3387 * be a slab entry.
3388 * @cachep: the cache we're checking against
3389 * @ptr: pointer to validate
3390 *
3391 * This verifies that the untrusted pointer looks sane:
3392 * it is _not_ a guarantee that the pointer is actually
3393 * part of the slab cache in question, but it at least
3394 * validates that the pointer can be dereferenced and
3395 * looks half-way sane.
3396 *
3397 * Currently only used for dentry validation.
3398 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003399int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003401 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003403 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003404 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 struct page *page;
3406
3407 if (unlikely(addr < min_addr))
3408 goto out;
3409 if (unlikely(addr > (unsigned long)high_memory - size))
3410 goto out;
3411 if (unlikely(addr & align_mask))
3412 goto out;
3413 if (unlikely(!kern_addr_valid(addr)))
3414 goto out;
3415 if (unlikely(!kern_addr_valid(addr + size - 1)))
3416 goto out;
3417 page = virt_to_page(ptr);
3418 if (unlikely(!PageSlab(page)))
3419 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003420 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 goto out;
3422 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003423out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 return 0;
3425}
3426
3427#ifdef CONFIG_NUMA
3428/**
3429 * kmem_cache_alloc_node - Allocate an object on the specified node
3430 * @cachep: The cache to allocate from.
3431 * @flags: See kmalloc().
3432 * @nodeid: node number of the target node.
3433 *
3434 * Identical to kmem_cache_alloc, except that this function is slow
3435 * and can sleep. And it will allocate memory on the given node, which
3436 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003437 * New and improved: it will now make sure that the object gets
3438 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003440void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441{
Christoph Lametere498be72005-09-09 13:03:32 -07003442 unsigned long save_flags;
3443 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444
Christoph Lametere498be72005-09-09 13:03:32 -07003445 cache_alloc_debugcheck_before(cachep, flags);
3446 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003447
3448 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003449 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003450 ptr = ____cache_alloc(cachep, flags);
3451 else
3452 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003453 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003454
3455 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3456 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
Christoph Lametere498be72005-09-09 13:03:32 -07003458 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459}
3460EXPORT_SYMBOL(kmem_cache_alloc_node);
3461
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003462void *__kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003463{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003464 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003465
3466 cachep = kmem_find_general_cachep(size, flags);
3467 if (unlikely(cachep == NULL))
3468 return NULL;
3469 return kmem_cache_alloc_node(cachep, flags, node);
3470}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003471EXPORT_SYMBOL(__kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472#endif
3473
3474/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003475 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003477 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003478 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003480static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3481 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003483 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003485 /* If you want to save a few bytes .text space: replace
3486 * __ with kmem_.
3487 * Then kmalloc uses the uninlined functions instead of the inline
3488 * functions.
3489 */
3490 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003491 if (unlikely(cachep == NULL))
3492 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003493 return __cache_alloc(cachep, flags, caller);
3494}
3495
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003496
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003497#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003498void *__kmalloc(size_t size, gfp_t flags)
3499{
Al Viro871751e2006-03-25 03:06:39 -08003500 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501}
3502EXPORT_SYMBOL(__kmalloc);
3503
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003504void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3505{
3506 return __do_kmalloc(size, flags, caller);
3507}
3508EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003509
3510#else
3511void *__kmalloc(size_t size, gfp_t flags)
3512{
3513 return __do_kmalloc(size, flags, NULL);
3514}
3515EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003516#endif
3517
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518/**
3519 * kmem_cache_free - Deallocate an object
3520 * @cachep: The cache the allocation was from.
3521 * @objp: The previously allocated object.
3522 *
3523 * Free an object which was previously allocated from this
3524 * cache.
3525 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003526void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527{
3528 unsigned long flags;
3529
Pekka Enbergddc2e812006-06-23 02:03:40 -07003530 BUG_ON(virt_to_cache(objp) != cachep);
3531
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 local_irq_save(flags);
Ingo Molnar873623d2006-07-13 14:44:38 +02003533 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 local_irq_restore(flags);
3535}
3536EXPORT_SYMBOL(kmem_cache_free);
3537
3538/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 * kfree - free previously allocated memory
3540 * @objp: pointer returned by kmalloc.
3541 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003542 * If @objp is NULL, no operation is performed.
3543 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 * Don't free memory not originally allocated by kmalloc()
3545 * or you will run into trouble.
3546 */
3547void kfree(const void *objp)
3548{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003549 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550 unsigned long flags;
3551
3552 if (unlikely(!objp))
3553 return;
3554 local_irq_save(flags);
3555 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003556 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003557 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003558 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559 local_irq_restore(flags);
3560}
3561EXPORT_SYMBOL(kfree);
3562
Pekka Enberg343e0d72006-02-01 03:05:50 -08003563unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003565 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566}
3567EXPORT_SYMBOL(kmem_cache_size);
3568
Pekka Enberg343e0d72006-02-01 03:05:50 -08003569const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003570{
3571 return cachep->name;
3572}
3573EXPORT_SYMBOL_GPL(kmem_cache_name);
3574
Christoph Lametere498be72005-09-09 13:03:32 -07003575/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003576 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003577 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003578static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003579{
3580 int node;
3581 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003582 struct array_cache *new_shared;
3583 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003584
3585 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003586
Andrew Mortona737b3e2006-03-22 00:08:11 -08003587 new_alien = alloc_alien_cache(node, cachep->limit);
3588 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003589 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003590
Christoph Lameter0718dc22006-03-25 03:06:47 -08003591 new_shared = alloc_arraycache(node,
3592 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003593 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003594 if (!new_shared) {
3595 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003596 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003597 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003598
Andrew Mortona737b3e2006-03-22 00:08:11 -08003599 l3 = cachep->nodelists[node];
3600 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003601 struct array_cache *shared = l3->shared;
3602
Christoph Lametere498be72005-09-09 13:03:32 -07003603 spin_lock_irq(&l3->list_lock);
3604
Christoph Lametercafeb022006-03-25 03:06:46 -08003605 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003606 free_block(cachep, shared->entry,
3607 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003608
Christoph Lametercafeb022006-03-25 03:06:46 -08003609 l3->shared = new_shared;
3610 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003611 l3->alien = new_alien;
3612 new_alien = NULL;
3613 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003614 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003615 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003616 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003617 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003618 free_alien_cache(new_alien);
3619 continue;
3620 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003621 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003622 if (!l3) {
3623 free_alien_cache(new_alien);
3624 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003625 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003626 }
Christoph Lametere498be72005-09-09 13:03:32 -07003627
3628 kmem_list3_init(l3);
3629 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003630 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003631 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003632 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003633 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003634 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003635 cachep->nodelists[node] = l3;
3636 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003637 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003638
Andrew Mortona737b3e2006-03-22 00:08:11 -08003639fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003640 if (!cachep->next.next) {
3641 /* Cache is not active yet. Roll back what we did */
3642 node--;
3643 while (node >= 0) {
3644 if (cachep->nodelists[node]) {
3645 l3 = cachep->nodelists[node];
3646
3647 kfree(l3->shared);
3648 free_alien_cache(l3->alien);
3649 kfree(l3);
3650 cachep->nodelists[node] = NULL;
3651 }
3652 node--;
3653 }
3654 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003655 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003656}
3657
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003659 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 struct array_cache *new[NR_CPUS];
3661};
3662
3663static void do_ccupdate_local(void *info)
3664{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003665 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666 struct array_cache *old;
3667
3668 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003669 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003670
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3672 new->new[smp_processor_id()] = old;
3673}
3674
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003675/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003676static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3677 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003679 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003680 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003682 new = kzalloc(sizeof(*new), GFP_KERNEL);
3683 if (!new)
3684 return -ENOMEM;
3685
Christoph Lametere498be72005-09-09 13:03:32 -07003686 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003687 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003688 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003689 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003690 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003691 kfree(new->new[i]);
3692 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003693 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 }
3695 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003696 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003698 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003699
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701 cachep->batchcount = batchcount;
3702 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003703 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704
Christoph Lametere498be72005-09-09 13:03:32 -07003705 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003706 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707 if (!ccold)
3708 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003709 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003710 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003711 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712 kfree(ccold);
3713 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003714 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003715 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716}
3717
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003718/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003719static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720{
3721 int err;
3722 int limit, shared;
3723
Andrew Mortona737b3e2006-03-22 00:08:11 -08003724 /*
3725 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 * - create a LIFO ordering, i.e. return objects that are cache-warm
3727 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003728 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729 * bufctl chains: array operations are cheaper.
3730 * The numbers are guessed, we should auto-tune as described by
3731 * Bonwick.
3732 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003733 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003735 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003737 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003739 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740 limit = 54;
3741 else
3742 limit = 120;
3743
Andrew Mortona737b3e2006-03-22 00:08:11 -08003744 /*
3745 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746 * allocation behaviour: Most allocs on one cpu, most free operations
3747 * on another cpu. For these cases, an efficient object passing between
3748 * cpus is necessary. This is provided by a shared array. The array
3749 * replaces Bonwick's magazine layer.
3750 * On uniprocessor, it's functionally equivalent (but less efficient)
3751 * to a larger limit. Thus disabled by default.
3752 */
3753 shared = 0;
3754#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003755 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756 shared = 8;
3757#endif
3758
3759#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003760 /*
3761 * With debugging enabled, large batchcount lead to excessively long
3762 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763 */
3764 if (limit > 32)
3765 limit = 32;
3766#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003767 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003768 if (err)
3769 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003770 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003771 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772}
3773
Christoph Lameter1b552532006-03-22 00:09:07 -08003774/*
3775 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003776 * necessary. Note that the l3 listlock also protects the array_cache
3777 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003778 */
3779void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3780 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781{
3782 int tofree;
3783
Christoph Lameter1b552532006-03-22 00:09:07 -08003784 if (!ac || !ac->avail)
3785 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003786 if (ac->touched && !force) {
3787 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003788 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003789 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003790 if (ac->avail) {
3791 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3792 if (tofree > ac->avail)
3793 tofree = (ac->avail + 1) / 2;
3794 free_block(cachep, ac->entry, tofree, node);
3795 ac->avail -= tofree;
3796 memmove(ac->entry, &(ac->entry[tofree]),
3797 sizeof(void *) * ac->avail);
3798 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003799 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800 }
3801}
3802
3803/**
3804 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003805 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806 *
3807 * Called from workqueue/eventd every few seconds.
3808 * Purpose:
3809 * - clear the per-cpu caches for this CPU.
3810 * - return freeable pages to the main free memory pool.
3811 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003812 * If we cannot acquire the cache chain mutex then just give up - we'll try
3813 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814 */
3815static void cache_reap(void *unused)
3816{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003817 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07003818 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003819 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003820
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003821 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003822 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003823 schedule_delayed_work(&__get_cpu_var(reap_work),
3824 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 return;
3826 }
3827
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003828 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 check_irq_on();
3830
Christoph Lameter35386e32006-03-22 00:09:05 -08003831 /*
3832 * We only take the l3 lock if absolutely necessary and we
3833 * have established with reasonable certainty that
3834 * we can do some work if the lock was obtained.
3835 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003836 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003837
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003838 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839
Christoph Lameteraab22072006-03-22 00:09:06 -08003840 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841
Christoph Lameter35386e32006-03-22 00:09:05 -08003842 /*
3843 * These are racy checks but it does not matter
3844 * if we skip one check or scan twice.
3845 */
Christoph Lametere498be72005-09-09 13:03:32 -07003846 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003847 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848
Christoph Lametere498be72005-09-09 13:03:32 -07003849 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850
Christoph Lameteraab22072006-03-22 00:09:06 -08003851 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852
Christoph Lametered11d9e2006-06-30 01:55:45 -07003853 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07003854 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003855 else {
3856 int freed;
3857
3858 freed = drain_freelist(searchp, l3, (l3->free_limit +
3859 5 * searchp->num - 1) / (5 * searchp->num));
3860 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003862next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 cond_resched();
3864 }
3865 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003866 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003867 next_reap_node();
Christoph Lameter2244b952006-06-30 01:55:33 -07003868 refresh_cpu_vm_stats(smp_processor_id());
Andrew Mortona737b3e2006-03-22 00:08:11 -08003869 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003870 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871}
3872
3873#ifdef CONFIG_PROC_FS
3874
Pekka Enberg85289f92006-01-08 01:00:36 -08003875static void print_slabinfo_header(struct seq_file *m)
3876{
3877 /*
3878 * Output format version, so at least we can change it
3879 * without _too_ many complaints.
3880 */
3881#if STATS
3882 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3883#else
3884 seq_puts(m, "slabinfo - version: 2.1\n");
3885#endif
3886 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3887 "<objperslab> <pagesperslab>");
3888 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3889 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3890#if STATS
3891 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003892 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003893 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3894#endif
3895 seq_putc(m, '\n');
3896}
3897
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898static void *s_start(struct seq_file *m, loff_t *pos)
3899{
3900 loff_t n = *pos;
3901 struct list_head *p;
3902
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003903 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003904 if (!n)
3905 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906 p = cache_chain.next;
3907 while (n--) {
3908 p = p->next;
3909 if (p == &cache_chain)
3910 return NULL;
3911 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003912 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913}
3914
3915static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3916{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003917 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003919 return cachep->next.next == &cache_chain ?
3920 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921}
3922
3923static void s_stop(struct seq_file *m, void *p)
3924{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003925 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926}
3927
3928static int s_show(struct seq_file *m, void *p)
3929{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003930 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003931 struct slab *slabp;
3932 unsigned long active_objs;
3933 unsigned long num_objs;
3934 unsigned long active_slabs = 0;
3935 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003936 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003938 int node;
3939 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941 active_objs = 0;
3942 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003943 for_each_online_node(node) {
3944 l3 = cachep->nodelists[node];
3945 if (!l3)
3946 continue;
3947
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003948 check_irq_on();
3949 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003950
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003951 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003952 if (slabp->inuse != cachep->num && !error)
3953 error = "slabs_full accounting error";
3954 active_objs += cachep->num;
3955 active_slabs++;
3956 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003957 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003958 if (slabp->inuse == cachep->num && !error)
3959 error = "slabs_partial inuse accounting error";
3960 if (!slabp->inuse && !error)
3961 error = "slabs_partial/inuse accounting error";
3962 active_objs += slabp->inuse;
3963 active_slabs++;
3964 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003965 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003966 if (slabp->inuse && !error)
3967 error = "slabs_free/inuse accounting error";
3968 num_slabs++;
3969 }
3970 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003971 if (l3->shared)
3972 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003973
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003974 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003976 num_slabs += active_slabs;
3977 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003978 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 error = "free_objects accounting error";
3980
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003981 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 if (error)
3983 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3984
3985 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003986 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003987 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003989 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003990 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003991 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003993 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 unsigned long high = cachep->high_mark;
3995 unsigned long allocs = cachep->num_allocations;
3996 unsigned long grown = cachep->grown;
3997 unsigned long reaped = cachep->reaped;
3998 unsigned long errors = cachep->errors;
3999 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004001 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004002 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003
Christoph Lametere498be72005-09-09 13:03:32 -07004004 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004005 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004006 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004007 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 }
4009 /* cpu stats */
4010 {
4011 unsigned long allochit = atomic_read(&cachep->allochit);
4012 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4013 unsigned long freehit = atomic_read(&cachep->freehit);
4014 unsigned long freemiss = atomic_read(&cachep->freemiss);
4015
4016 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004017 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018 }
4019#endif
4020 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 return 0;
4022}
4023
4024/*
4025 * slabinfo_op - iterator that generates /proc/slabinfo
4026 *
4027 * Output layout:
4028 * cache-name
4029 * num-active-objs
4030 * total-objs
4031 * object size
4032 * num-active-slabs
4033 * total-slabs
4034 * num-pages-per-slab
4035 * + further values on SMP and with statistics enabled
4036 */
4037
4038struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004039 .start = s_start,
4040 .next = s_next,
4041 .stop = s_stop,
4042 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043};
4044
4045#define MAX_SLABINFO_WRITE 128
4046/**
4047 * slabinfo_write - Tuning for the slab allocator
4048 * @file: unused
4049 * @buffer: user buffer
4050 * @count: data length
4051 * @ppos: unused
4052 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004053ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4054 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004056 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004058 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004059
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 if (count > MAX_SLABINFO_WRITE)
4061 return -EINVAL;
4062 if (copy_from_user(&kbuf, buffer, count))
4063 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004064 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065
4066 tmp = strchr(kbuf, ' ');
4067 if (!tmp)
4068 return -EINVAL;
4069 *tmp = '\0';
4070 tmp++;
4071 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4072 return -EINVAL;
4073
4074 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004075 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004077 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004079 if (limit < 1 || batchcount < 1 ||
4080 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004081 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004083 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004084 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085 }
4086 break;
4087 }
4088 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004089 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090 if (res >= 0)
4091 res = count;
4092 return res;
4093}
Al Viro871751e2006-03-25 03:06:39 -08004094
4095#ifdef CONFIG_DEBUG_SLAB_LEAK
4096
4097static void *leaks_start(struct seq_file *m, loff_t *pos)
4098{
4099 loff_t n = *pos;
4100 struct list_head *p;
4101
4102 mutex_lock(&cache_chain_mutex);
4103 p = cache_chain.next;
4104 while (n--) {
4105 p = p->next;
4106 if (p == &cache_chain)
4107 return NULL;
4108 }
4109 return list_entry(p, struct kmem_cache, next);
4110}
4111
4112static inline int add_caller(unsigned long *n, unsigned long v)
4113{
4114 unsigned long *p;
4115 int l;
4116 if (!v)
4117 return 1;
4118 l = n[1];
4119 p = n + 2;
4120 while (l) {
4121 int i = l/2;
4122 unsigned long *q = p + 2 * i;
4123 if (*q == v) {
4124 q[1]++;
4125 return 1;
4126 }
4127 if (*q > v) {
4128 l = i;
4129 } else {
4130 p = q + 2;
4131 l -= i + 1;
4132 }
4133 }
4134 if (++n[1] == n[0])
4135 return 0;
4136 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4137 p[0] = v;
4138 p[1] = 1;
4139 return 1;
4140}
4141
4142static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4143{
4144 void *p;
4145 int i;
4146 if (n[0] == n[1])
4147 return;
4148 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4149 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4150 continue;
4151 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4152 return;
4153 }
4154}
4155
4156static void show_symbol(struct seq_file *m, unsigned long address)
4157{
4158#ifdef CONFIG_KALLSYMS
4159 char *modname;
4160 const char *name;
4161 unsigned long offset, size;
4162 char namebuf[KSYM_NAME_LEN+1];
4163
4164 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4165
4166 if (name) {
4167 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4168 if (modname)
4169 seq_printf(m, " [%s]", modname);
4170 return;
4171 }
4172#endif
4173 seq_printf(m, "%p", (void *)address);
4174}
4175
4176static int leaks_show(struct seq_file *m, void *p)
4177{
4178 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004179 struct slab *slabp;
4180 struct kmem_list3 *l3;
4181 const char *name;
4182 unsigned long *n = m->private;
4183 int node;
4184 int i;
4185
4186 if (!(cachep->flags & SLAB_STORE_USER))
4187 return 0;
4188 if (!(cachep->flags & SLAB_RED_ZONE))
4189 return 0;
4190
4191 /* OK, we can do it */
4192
4193 n[1] = 0;
4194
4195 for_each_online_node(node) {
4196 l3 = cachep->nodelists[node];
4197 if (!l3)
4198 continue;
4199
4200 check_irq_on();
4201 spin_lock_irq(&l3->list_lock);
4202
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004203 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004204 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004205 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004206 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004207 spin_unlock_irq(&l3->list_lock);
4208 }
4209 name = cachep->name;
4210 if (n[0] == n[1]) {
4211 /* Increase the buffer size */
4212 mutex_unlock(&cache_chain_mutex);
4213 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4214 if (!m->private) {
4215 /* Too bad, we are really out */
4216 m->private = n;
4217 mutex_lock(&cache_chain_mutex);
4218 return -ENOMEM;
4219 }
4220 *(unsigned long *)m->private = n[0] * 2;
4221 kfree(n);
4222 mutex_lock(&cache_chain_mutex);
4223 /* Now make sure this entry will be retried */
4224 m->count = m->size;
4225 return 0;
4226 }
4227 for (i = 0; i < n[1]; i++) {
4228 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4229 show_symbol(m, n[2*i+2]);
4230 seq_putc(m, '\n');
4231 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004232
Al Viro871751e2006-03-25 03:06:39 -08004233 return 0;
4234}
4235
4236struct seq_operations slabstats_op = {
4237 .start = leaks_start,
4238 .next = s_next,
4239 .stop = s_stop,
4240 .show = leaks_show,
4241};
4242#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243#endif
4244
Manfred Spraul00e145b2005-09-03 15:55:07 -07004245/**
4246 * ksize - get the actual amount of memory allocated for a given object
4247 * @objp: Pointer to the object
4248 *
4249 * kmalloc may internally round up allocations and return more memory
4250 * than requested. ksize() can be used to determine the actual amount of
4251 * memory allocated. The caller may use this additional memory, even though
4252 * a smaller amount of memory was initially specified with the kmalloc call.
4253 * The caller must guarantee that objp points to a valid object previously
4254 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4255 * must not be freed during the duration of the call.
4256 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004257unsigned int ksize(const void *objp)
4258{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004259 if (unlikely(objp == NULL))
4260 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004262 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004263}