blob: 068cb4503c15fcb50d6421b10f3b7d7b8f6ab771 [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>
Andrew Morton138ae662006-12-06 20:36:41 -0800106#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700110#include <linux/rtmutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112#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);
David Howells65f27f32006-11-22 14:55:48 +0000316static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700317
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
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800733/*
734 * 1. Guard access to the cache-chain.
735 * 2. Protect sanity of cpu_online_map against cpu hotplug events
736 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800737static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738static struct list_head cache_chain;
739
740/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 * chicken and egg problem: delay the per-cpu array allocation
742 * until the general caches are up.
743 */
744static enum {
745 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700746 PARTIAL_AC,
747 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 FULL
749} g_cpucache_up;
750
Mike Kravetz39d24e62006-05-15 09:44:13 -0700751/*
752 * used by boot code to determine if it can use slab based allocator
753 */
754int slab_is_available(void)
755{
756 return g_cpucache_up == FULL;
757}
758
David Howells52bad642006-11-22 14:54:01 +0000759static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
Pekka Enberg343e0d72006-02-01 03:05:50 -0800761static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762{
763 return cachep->array[smp_processor_id()];
764}
765
Andrew Mortona737b3e2006-03-22 00:08:11 -0800766static inline struct kmem_cache *__find_general_cachep(size_t size,
767 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
769 struct cache_sizes *csizep = malloc_sizes;
770
771#if DEBUG
772 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800773 * kmem_cache_create(), or __kmalloc(), before
774 * the generic caches are initialized.
775 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700776 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777#endif
778 while (size > csizep->cs_size)
779 csizep++;
780
781 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700782 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 * has cs_{dma,}cachep==NULL. Thus no special case
784 * for large kmalloc calls required.
785 */
786 if (unlikely(gfpflags & GFP_DMA))
787 return csizep->cs_dmacachep;
788 return csizep->cs_cachep;
789}
790
Adrian Bunkb2213852006-09-25 23:31:02 -0700791static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700792{
793 return __find_general_cachep(size, gfpflags);
794}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700795
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800796static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800798 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
799}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
Andrew Mortona737b3e2006-03-22 00:08:11 -0800801/*
802 * Calculate the number of objects and left-over bytes for a given buffer size.
803 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800804static void cache_estimate(unsigned long gfporder, size_t buffer_size,
805 size_t align, int flags, size_t *left_over,
806 unsigned int *num)
807{
808 int nr_objs;
809 size_t mgmt_size;
810 size_t slab_size = PAGE_SIZE << gfporder;
811
812 /*
813 * The slab management structure can be either off the slab or
814 * on it. For the latter case, the memory allocated for a
815 * slab is used for:
816 *
817 * - The struct slab
818 * - One kmem_bufctl_t for each object
819 * - Padding to respect alignment of @align
820 * - @buffer_size bytes for each object
821 *
822 * If the slab management structure is off the slab, then the
823 * alignment will already be calculated into the size. Because
824 * the slabs are all pages aligned, the objects will be at the
825 * correct alignment when allocated.
826 */
827 if (flags & CFLGS_OFF_SLAB) {
828 mgmt_size = 0;
829 nr_objs = slab_size / buffer_size;
830
831 if (nr_objs > SLAB_LIMIT)
832 nr_objs = SLAB_LIMIT;
833 } else {
834 /*
835 * Ignore padding for the initial guess. The padding
836 * is at most @align-1 bytes, and @buffer_size is at
837 * least @align. In the worst case, this result will
838 * be one greater than the number of objects that fit
839 * into the memory allocation when taking the padding
840 * into account.
841 */
842 nr_objs = (slab_size - sizeof(struct slab)) /
843 (buffer_size + sizeof(kmem_bufctl_t));
844
845 /*
846 * This calculated number will be either the right
847 * amount, or one greater than what we want.
848 */
849 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
850 > slab_size)
851 nr_objs--;
852
853 if (nr_objs > SLAB_LIMIT)
854 nr_objs = SLAB_LIMIT;
855
856 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800858 *num = nr_objs;
859 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860}
861
862#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
863
Andrew Mortona737b3e2006-03-22 00:08:11 -0800864static void __slab_error(const char *function, struct kmem_cache *cachep,
865 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866{
867 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800868 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 dump_stack();
870}
871
Paul Menage3395ee02006-12-06 20:32:16 -0800872/*
873 * By default on NUMA we use alien caches to stage the freeing of
874 * objects allocated from other nodes. This causes massive memory
875 * inefficiencies when using fake NUMA setup to split memory into a
876 * large number of small nodes, so it can be disabled on the command
877 * line
878 */
879
880static int use_alien_caches __read_mostly = 1;
881static int __init noaliencache_setup(char *s)
882{
883 use_alien_caches = 0;
884 return 1;
885}
886__setup("noaliencache", noaliencache_setup);
887
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800888#ifdef CONFIG_NUMA
889/*
890 * Special reaping functions for NUMA systems called from cache_reap().
891 * These take care of doing round robin flushing of alien caches (containing
892 * objects freed on different nodes from which they were allocated) and the
893 * flushing of remote pcps by calling drain_node_pages.
894 */
895static DEFINE_PER_CPU(unsigned long, reap_node);
896
897static void init_reap_node(int cpu)
898{
899 int node;
900
901 node = next_node(cpu_to_node(cpu), node_online_map);
902 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800903 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800904
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800905 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800906}
907
908static void next_reap_node(void)
909{
910 int node = __get_cpu_var(reap_node);
911
912 /*
913 * Also drain per cpu pages on remote zones
914 */
915 if (node != numa_node_id())
916 drain_node_pages(node);
917
918 node = next_node(node, node_online_map);
919 if (unlikely(node >= MAX_NUMNODES))
920 node = first_node(node_online_map);
921 __get_cpu_var(reap_node) = node;
922}
923
924#else
925#define init_reap_node(cpu) do { } while (0)
926#define next_reap_node(void) do { } while (0)
927#endif
928
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929/*
930 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
931 * via the workqueue/eventd.
932 * Add the CPU number into the expiration time to minimize the possibility of
933 * the CPUs getting into lockstep and contending for the global cache chain
934 * lock.
935 */
936static void __devinit start_cpu_timer(int cpu)
937{
David Howells52bad642006-11-22 14:54:01 +0000938 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939
940 /*
941 * When this gets called from do_initcalls via cpucache_init(),
942 * init_workqueues() has already run, so keventd will be setup
943 * at that time.
944 */
David Howells52bad642006-11-22 14:54:01 +0000945 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800946 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000947 INIT_DELAYED_WORK(reap_work, cache_reap);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
949 }
950}
951
Christoph Lametere498be72005-09-09 13:03:32 -0700952static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800953 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800955 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 struct array_cache *nc = NULL;
957
Christoph Lametere498be72005-09-09 13:03:32 -0700958 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 if (nc) {
960 nc->avail = 0;
961 nc->limit = entries;
962 nc->batchcount = batchcount;
963 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700964 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 }
966 return nc;
967}
968
Christoph Lameter3ded1752006-03-25 03:06:44 -0800969/*
970 * Transfer objects in one arraycache to another.
971 * Locking must be handled by the caller.
972 *
973 * Return the number of entries transferred.
974 */
975static int transfer_objects(struct array_cache *to,
976 struct array_cache *from, unsigned int max)
977{
978 /* Figure out how many entries to transfer */
979 int nr = min(min(from->avail, max), to->limit - to->avail);
980
981 if (!nr)
982 return 0;
983
984 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
985 sizeof(void *) *nr);
986
987 from->avail -= nr;
988 to->avail += nr;
989 to->touched = 1;
990 return nr;
991}
992
Christoph Lameter765c4502006-09-27 01:50:08 -0700993#ifndef CONFIG_NUMA
994
995#define drain_alien_cache(cachep, alien) do { } while (0)
996#define reap_alien(cachep, l3) do { } while (0)
997
998static inline struct array_cache **alloc_alien_cache(int node, int limit)
999{
1000 return (struct array_cache **)BAD_ALIEN_MAGIC;
1001}
1002
1003static inline void free_alien_cache(struct array_cache **ac_ptr)
1004{
1005}
1006
1007static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1008{
1009 return 0;
1010}
1011
1012static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1013 gfp_t flags)
1014{
1015 return NULL;
1016}
1017
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001018static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001019 gfp_t flags, int nodeid)
1020{
1021 return NULL;
1022}
1023
1024#else /* CONFIG_NUMA */
1025
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001026static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001027static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001028
Pekka Enberg5295a742006-02-01 03:05:48 -08001029static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001030{
1031 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001032 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -07001033 int i;
1034
1035 if (limit > 1)
1036 limit = 12;
1037 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1038 if (ac_ptr) {
1039 for_each_node(i) {
1040 if (i == node || !node_online(i)) {
1041 ac_ptr[i] = NULL;
1042 continue;
1043 }
1044 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1045 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001046 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001047 kfree(ac_ptr[i]);
1048 kfree(ac_ptr);
1049 return NULL;
1050 }
1051 }
1052 }
1053 return ac_ptr;
1054}
1055
Pekka Enberg5295a742006-02-01 03:05:48 -08001056static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001057{
1058 int i;
1059
1060 if (!ac_ptr)
1061 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001062 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001063 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001064 kfree(ac_ptr);
1065}
1066
Pekka Enberg343e0d72006-02-01 03:05:50 -08001067static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001068 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001069{
1070 struct kmem_list3 *rl3 = cachep->nodelists[node];
1071
1072 if (ac->avail) {
1073 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001074 /*
1075 * Stuff objects into the remote nodes shared array first.
1076 * That way we could avoid the overhead of putting the objects
1077 * into the free lists and getting them back later.
1078 */
shin, jacob693f7d32006-04-28 10:54:37 -05001079 if (rl3->shared)
1080 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001081
Christoph Lameterff694162005-09-22 21:44:02 -07001082 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001083 ac->avail = 0;
1084 spin_unlock(&rl3->list_lock);
1085 }
1086}
1087
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001088/*
1089 * Called from cache_reap() to regularly drain alien caches round robin.
1090 */
1091static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1092{
1093 int node = __get_cpu_var(reap_node);
1094
1095 if (l3->alien) {
1096 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001097
1098 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001099 __drain_alien_cache(cachep, ac, node);
1100 spin_unlock_irq(&ac->lock);
1101 }
1102 }
1103}
1104
Andrew Mortona737b3e2006-03-22 00:08:11 -08001105static void drain_alien_cache(struct kmem_cache *cachep,
1106 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001107{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001108 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001109 struct array_cache *ac;
1110 unsigned long flags;
1111
1112 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001113 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001114 if (ac) {
1115 spin_lock_irqsave(&ac->lock, flags);
1116 __drain_alien_cache(cachep, ac, i);
1117 spin_unlock_irqrestore(&ac->lock, flags);
1118 }
1119 }
1120}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001121
Ingo Molnar873623d2006-07-13 14:44:38 +02001122static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001123{
1124 struct slab *slabp = virt_to_slab(objp);
1125 int nodeid = slabp->nodeid;
1126 struct kmem_list3 *l3;
1127 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001128 int node;
1129
1130 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001131
1132 /*
1133 * Make sure we are not freeing a object from another node to the array
1134 * cache on this cpu.
1135 */
Paul Menage3395ee02006-12-06 20:32:16 -08001136 if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001137 return 0;
1138
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001139 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001140 STATS_INC_NODEFREES(cachep);
1141 if (l3->alien && l3->alien[nodeid]) {
1142 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001143 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001144 if (unlikely(alien->avail == alien->limit)) {
1145 STATS_INC_ACOVERFLOW(cachep);
1146 __drain_alien_cache(cachep, alien, nodeid);
1147 }
1148 alien->entry[alien->avail++] = objp;
1149 spin_unlock(&alien->lock);
1150 } else {
1151 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1152 free_block(cachep, &objp, 1, nodeid);
1153 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1154 }
1155 return 1;
1156}
Christoph Lametere498be72005-09-09 13:03:32 -07001157#endif
1158
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001159static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001160 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161{
1162 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001163 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001164 struct kmem_list3 *l3 = NULL;
1165 int node = cpu_to_node(cpu);
1166 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
1168 switch (action) {
1169 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001170 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001171 /*
1172 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001173 * alloc_arraycache's are going to use this list.
1174 * kmalloc_node allows us to add the slab to the right
1175 * kmem_list3 and not this cpu's kmem_list3
1176 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
Christoph Lametere498be72005-09-09 13:03:32 -07001178 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001179 /*
1180 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001181 * begin anything. Make sure some other cpu on this
1182 * node has not already allocated this
1183 */
1184 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001185 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1186 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001187 goto bad;
1188 kmem_list3_init(l3);
1189 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001190 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001191
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001192 /*
1193 * The l3s don't come and go as CPUs come and
1194 * go. cache_chain_mutex is sufficient
1195 * protection here.
1196 */
Christoph Lametere498be72005-09-09 13:03:32 -07001197 cachep->nodelists[node] = l3;
1198 }
1199
1200 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1201 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001202 (1 + nr_cpus_node(node)) *
1203 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001204 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1205 }
1206
Andrew Mortona737b3e2006-03-22 00:08:11 -08001207 /*
1208 * Now we can go ahead with allocating the shared arrays and
1209 * array caches
1210 */
Christoph Lametere498be72005-09-09 13:03:32 -07001211 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001212 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001213 struct array_cache *shared;
Paul Menage3395ee02006-12-06 20:32:16 -08001214 struct array_cache **alien = NULL;
Tobias Klausercd105df2006-01-08 01:00:59 -08001215
Christoph Lametere498be72005-09-09 13:03:32 -07001216 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001217 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 if (!nc)
1219 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001220 shared = alloc_arraycache(node,
1221 cachep->shared * cachep->batchcount,
1222 0xbaadf00d);
1223 if (!shared)
1224 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001225
Paul Menage3395ee02006-12-06 20:32:16 -08001226 if (use_alien_caches) {
1227 alien = alloc_alien_cache(node, cachep->limit);
1228 if (!alien)
1229 goto bad;
1230 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001232 l3 = cachep->nodelists[node];
1233 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001234
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001235 spin_lock_irq(&l3->list_lock);
1236 if (!l3->shared) {
1237 /*
1238 * We are serialised from CPU_DEAD or
1239 * CPU_UP_CANCELLED by the cpucontrol lock
1240 */
1241 l3->shared = shared;
1242 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001243 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001244#ifdef CONFIG_NUMA
1245 if (!l3->alien) {
1246 l3->alien = alien;
1247 alien = NULL;
1248 }
1249#endif
1250 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001251 kfree(shared);
1252 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 break;
1255 case CPU_ONLINE:
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001256 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 start_cpu_timer(cpu);
1258 break;
1259#ifdef CONFIG_HOTPLUG_CPU
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001260 case CPU_DOWN_PREPARE:
1261 mutex_lock(&cache_chain_mutex);
1262 break;
1263 case CPU_DOWN_FAILED:
1264 mutex_unlock(&cache_chain_mutex);
1265 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001267 /*
1268 * Even if all the cpus of a node are down, we don't free the
1269 * kmem_list3 of any cache. This to avoid a race between
1270 * cpu_down, and a kmalloc allocation from another cpu for
1271 * memory from the node of the cpu going down. The list3
1272 * structure is usually allocated from kmem_cache_create() and
1273 * gets destroyed at kmem_cache_destroy().
1274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 /* fall thru */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001276#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 case CPU_UP_CANCELED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 list_for_each_entry(cachep, &cache_chain, next) {
1279 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001280 struct array_cache *shared;
1281 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001282 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283
Christoph Lametere498be72005-09-09 13:03:32 -07001284 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 /* cpu is dead; no one can alloc from it. */
1286 nc = cachep->array[cpu];
1287 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001288 l3 = cachep->nodelists[node];
1289
1290 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001291 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001292
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001293 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001294
1295 /* Free limit for this kmem_list3 */
1296 l3->free_limit -= cachep->batchcount;
1297 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001298 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001299
1300 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001301 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001302 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001303 }
Christoph Lametere498be72005-09-09 13:03:32 -07001304
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001305 shared = l3->shared;
1306 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001307 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001308 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001309 l3->shared = NULL;
1310 }
Christoph Lametere498be72005-09-09 13:03:32 -07001311
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001312 alien = l3->alien;
1313 l3->alien = NULL;
1314
1315 spin_unlock_irq(&l3->list_lock);
1316
1317 kfree(shared);
1318 if (alien) {
1319 drain_alien_cache(cachep, alien);
1320 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001321 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001322free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 kfree(nc);
1324 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001325 /*
1326 * In the previous loop, all the objects were freed to
1327 * the respective cache's slabs, now we can go ahead and
1328 * shrink each nodelist to its limit.
1329 */
1330 list_for_each_entry(cachep, &cache_chain, next) {
1331 l3 = cachep->nodelists[node];
1332 if (!l3)
1333 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001334 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001335 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001336 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 }
1339 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001340bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 return NOTIFY_BAD;
1342}
1343
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001344static struct notifier_block __cpuinitdata cpucache_notifier = {
1345 &cpuup_callback, NULL, 0
1346};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
Christoph Lametere498be72005-09-09 13:03:32 -07001348/*
1349 * swap the static kmem_list3 with kmalloced memory
1350 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001351static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1352 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001353{
1354 struct kmem_list3 *ptr;
1355
Christoph Lametere498be72005-09-09 13:03:32 -07001356 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1357 BUG_ON(!ptr);
1358
1359 local_irq_disable();
1360 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001361 /*
1362 * Do not assume that spinlocks can be initialized via memcpy:
1363 */
1364 spin_lock_init(&ptr->list_lock);
1365
Christoph Lametere498be72005-09-09 13:03:32 -07001366 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1367 cachep->nodelists[nodeid] = ptr;
1368 local_irq_enable();
1369}
1370
Andrew Mortona737b3e2006-03-22 00:08:11 -08001371/*
1372 * Initialisation. Called after the page allocator have been initialised and
1373 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 */
1375void __init kmem_cache_init(void)
1376{
1377 size_t left_over;
1378 struct cache_sizes *sizes;
1379 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001380 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001381 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001382 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001383
1384 for (i = 0; i < NUM_INIT_LISTS; i++) {
1385 kmem_list3_init(&initkmem_list3[i]);
1386 if (i < MAX_NUMNODES)
1387 cache_cache.nodelists[i] = NULL;
1388 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389
1390 /*
1391 * Fragmentation resistance on low memory - only use bigger
1392 * page orders on machines with more than 32MB of memory.
1393 */
1394 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1395 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 /* Bootstrap is tricky, because several objects are allocated
1398 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001399 * 1) initialize the cache_cache cache: it contains the struct
1400 * kmem_cache structures of all caches, except cache_cache itself:
1401 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001402 * Initially an __init data area is used for the head array and the
1403 * kmem_list3 structures, it's replaced with a kmalloc allocated
1404 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001406 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001407 * An __init data area is used for the head array.
1408 * 3) Create the remaining kmalloc caches, with minimally sized
1409 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 * 4) Replace the __init data head arrays for cache_cache and the first
1411 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001412 * 5) Replace the __init data for kmem_list3 for cache_cache and
1413 * the other cache's with kmalloc allocated memory.
1414 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 */
1416
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001417 node = numa_node_id();
1418
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 INIT_LIST_HEAD(&cache_chain);
1421 list_add(&cache_cache.next, &cache_chain);
1422 cache_cache.colour_off = cache_line_size();
1423 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001424 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
Andrew Mortona737b3e2006-03-22 00:08:11 -08001426 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1427 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
Jack Steiner07ed76b2006-03-07 21:55:46 -08001429 for (order = 0; order < MAX_ORDER; order++) {
1430 cache_estimate(order, cache_cache.buffer_size,
1431 cache_line_size(), 0, &left_over, &cache_cache.num);
1432 if (cache_cache.num)
1433 break;
1434 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001435 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001436 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001437 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001438 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1439 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440
1441 /* 2+3) create the kmalloc caches */
1442 sizes = malloc_sizes;
1443 names = cache_names;
1444
Andrew Mortona737b3e2006-03-22 00:08:11 -08001445 /*
1446 * Initialize the caches that provide memory for the array cache and the
1447 * kmem_list3 structures first. Without this, further allocations will
1448 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001449 */
1450
1451 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001452 sizes[INDEX_AC].cs_size,
1453 ARCH_KMALLOC_MINALIGN,
1454 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1455 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001456
Andrew Mortona737b3e2006-03-22 00:08:11 -08001457 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001458 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001459 kmem_cache_create(names[INDEX_L3].name,
1460 sizes[INDEX_L3].cs_size,
1461 ARCH_KMALLOC_MINALIGN,
1462 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1463 NULL, NULL);
1464 }
Christoph Lametere498be72005-09-09 13:03:32 -07001465
Ingo Molnare0a42722006-06-23 02:03:46 -07001466 slab_early_init = 0;
1467
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001469 /*
1470 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 * This should be particularly beneficial on SMP boxes, as it
1472 * eliminates "false sharing".
1473 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001474 * allow tighter packing of the smaller caches.
1475 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001476 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001477 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001478 sizes->cs_size,
1479 ARCH_KMALLOC_MINALIGN,
1480 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1481 NULL, NULL);
1482 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001485 sizes->cs_size,
1486 ARCH_KMALLOC_MINALIGN,
1487 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1488 SLAB_PANIC,
1489 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 sizes++;
1491 names++;
1492 }
1493 /* 4) Replace the bootstrap head arrays */
1494 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001495 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001496
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001498
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001500 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1501 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001502 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001503 /*
1504 * Do not assume that spinlocks can be initialized via memcpy:
1505 */
1506 spin_lock_init(&ptr->lock);
1507
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 cache_cache.array[smp_processor_id()] = ptr;
1509 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001510
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001512
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001514 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001515 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001516 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001517 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001518 /*
1519 * Do not assume that spinlocks can be initialized via memcpy:
1520 */
1521 spin_lock_init(&ptr->lock);
1522
Christoph Lametere498be72005-09-09 13:03:32 -07001523 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001524 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 local_irq_enable();
1526 }
Christoph Lametere498be72005-09-09 13:03:32 -07001527 /* 5) Replace the bootstrap kmem_list3's */
1528 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001529 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001531 /* Replace the static kmem_list3 structures for the boot cpu */
1532 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1533
1534 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001535 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001536 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001537
1538 if (INDEX_AC != INDEX_L3) {
1539 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001540 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001541 }
1542 }
1543 }
1544
1545 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001547 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001548 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001550 if (enable_cpucache(cachep))
1551 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001552 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 }
1554
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001555 /* Annotate slab for lockdep -- annotate the malloc caches */
1556 init_lock_keys();
1557
1558
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 /* Done! */
1560 g_cpucache_up = FULL;
1561
Andrew Mortona737b3e2006-03-22 00:08:11 -08001562 /*
1563 * Register a cpu startup notifier callback that initializes
1564 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 */
1566 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567
Andrew Mortona737b3e2006-03-22 00:08:11 -08001568 /*
1569 * The reap timers are started later, with a module init call: That part
1570 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 */
1572}
1573
1574static int __init cpucache_init(void)
1575{
1576 int cpu;
1577
Andrew Mortona737b3e2006-03-22 00:08:11 -08001578 /*
1579 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 */
Christoph Lametere498be72005-09-09 13:03:32 -07001581 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001582 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 return 0;
1584}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585__initcall(cpucache_init);
1586
1587/*
1588 * Interface to system's page allocator. No need to hold the cache-lock.
1589 *
1590 * If we requested dmaable memory, we will get it. Even if we
1591 * did not request dmaable memory, we might get it, but that
1592 * would be relatively rare and ignorable.
1593 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001594static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595{
1596 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001597 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 int i;
1599
Luke Yangd6fef9d2006-04-10 22:52:56 -07001600#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001601 /*
1602 * Nommu uses slab's for process anonymous memory allocations, and thus
1603 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001604 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001605 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001606#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001607
Christoph Lameter3c517a62006-12-06 20:33:29 -08001608 flags |= cachep->gfpflags;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001609
1610 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 if (!page)
1612 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001614 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001616 add_zone_page_state(page_zone(page),
1617 NR_SLAB_RECLAIMABLE, nr_pages);
1618 else
1619 add_zone_page_state(page_zone(page),
1620 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001621 for (i = 0; i < nr_pages; i++)
1622 __SetPageSlab(page + i);
1623 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624}
1625
1626/*
1627 * Interface to system's page release.
1628 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001629static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001631 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 struct page *page = virt_to_page(addr);
1633 const unsigned long nr_freed = i;
1634
Christoph Lameter972d1a72006-09-25 23:31:51 -07001635 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1636 sub_zone_page_state(page_zone(page),
1637 NR_SLAB_RECLAIMABLE, nr_freed);
1638 else
1639 sub_zone_page_state(page_zone(page),
1640 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001642 BUG_ON(!PageSlab(page));
1643 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 page++;
1645 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 if (current->reclaim_state)
1647 current->reclaim_state->reclaimed_slab += nr_freed;
1648 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649}
1650
1651static void kmem_rcu_free(struct rcu_head *head)
1652{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001653 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001654 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655
1656 kmem_freepages(cachep, slab_rcu->addr);
1657 if (OFF_SLAB(cachep))
1658 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1659}
1660
1661#if DEBUG
1662
1663#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001664static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001665 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001667 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001669 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001671 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 return;
1673
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001674 *addr++ = 0x12345678;
1675 *addr++ = caller;
1676 *addr++ = smp_processor_id();
1677 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 {
1679 unsigned long *sptr = &caller;
1680 unsigned long svalue;
1681
1682 while (!kstack_end(sptr)) {
1683 svalue = *sptr++;
1684 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001685 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 size -= sizeof(unsigned long);
1687 if (size <= sizeof(unsigned long))
1688 break;
1689 }
1690 }
1691
1692 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001693 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694}
1695#endif
1696
Pekka Enberg343e0d72006-02-01 03:05:50 -08001697static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001699 int size = obj_size(cachep);
1700 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
1702 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001703 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704}
1705
1706static void dump_line(char *data, int offset, int limit)
1707{
1708 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001709 unsigned char error = 0;
1710 int bad_count = 0;
1711
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001713 for (i = 0; i < limit; i++) {
1714 if (data[offset + i] != POISON_FREE) {
1715 error = data[offset + i];
1716 bad_count++;
1717 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001718 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001719 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001721
1722 if (bad_count == 1) {
1723 error ^= POISON_FREE;
1724 if (!(error & (error - 1))) {
1725 printk(KERN_ERR "Single bit error detected. Probably "
1726 "bad RAM.\n");
1727#ifdef CONFIG_X86
1728 printk(KERN_ERR "Run memtest86+ or a similar memory "
1729 "test tool.\n");
1730#else
1731 printk(KERN_ERR "Run a memory test tool.\n");
1732#endif
1733 }
1734 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735}
1736#endif
1737
1738#if DEBUG
1739
Pekka Enberg343e0d72006-02-01 03:05:50 -08001740static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741{
1742 int i, size;
1743 char *realobj;
1744
1745 if (cachep->flags & SLAB_RED_ZONE) {
1746 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001747 *dbg_redzone1(cachep, objp),
1748 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 }
1750
1751 if (cachep->flags & SLAB_STORE_USER) {
1752 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001753 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001755 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 printk("\n");
1757 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001758 realobj = (char *)objp + obj_offset(cachep);
1759 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001760 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 int limit;
1762 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001763 if (i + limit > size)
1764 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 dump_line(realobj, i, limit);
1766 }
1767}
1768
Pekka Enberg343e0d72006-02-01 03:05:50 -08001769static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770{
1771 char *realobj;
1772 int size, i;
1773 int lines = 0;
1774
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001775 realobj = (char *)objp + obj_offset(cachep);
1776 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001778 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001780 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 exp = POISON_END;
1782 if (realobj[i] != exp) {
1783 int limit;
1784 /* Mismatch ! */
1785 /* Print header */
1786 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001787 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001788 "Slab corruption: start=%p, len=%d\n",
1789 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 print_objinfo(cachep, objp, 0);
1791 }
1792 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001793 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001795 if (i + limit > size)
1796 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 dump_line(realobj, i, limit);
1798 i += 16;
1799 lines++;
1800 /* Limit to 5 lines */
1801 if (lines > 5)
1802 break;
1803 }
1804 }
1805 if (lines != 0) {
1806 /* Print some data about the neighboring objects, if they
1807 * exist:
1808 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001809 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001810 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001812 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001814 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001815 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001817 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 print_objinfo(cachep, objp, 2);
1819 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001820 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001821 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001822 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001824 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 print_objinfo(cachep, objp, 2);
1826 }
1827 }
1828}
1829#endif
1830
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001832/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001833 * slab_destroy_objs - destroy a slab and its objects
1834 * @cachep: cache pointer being destroyed
1835 * @slabp: slab pointer being destroyed
1836 *
1837 * Call the registered destructor for each object in a slab that is being
1838 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001839 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001840static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001841{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 int i;
1843 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001844 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845
1846 if (cachep->flags & SLAB_POISON) {
1847#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001848 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1849 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001850 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001851 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 else
1853 check_poison_obj(cachep, objp);
1854#else
1855 check_poison_obj(cachep, objp);
1856#endif
1857 }
1858 if (cachep->flags & SLAB_RED_ZONE) {
1859 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1860 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1863 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001864 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 }
1866 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001867 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001869}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001871static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001872{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 if (cachep->dtor) {
1874 int i;
1875 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001876 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 }
1879 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001880}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881#endif
1882
Randy Dunlap911851e2006-03-22 00:08:14 -08001883/**
1884 * slab_destroy - destroy and release all objects in a slab
1885 * @cachep: cache pointer being destroyed
1886 * @slabp: slab pointer being destroyed
1887 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001888 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001889 * Before calling the slab must have been unlinked from the cache. The
1890 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001891 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001892static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001893{
1894 void *addr = slabp->s_mem - slabp->colouroff;
1895
1896 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1898 struct slab_rcu *slab_rcu;
1899
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001900 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 slab_rcu->cachep = cachep;
1902 slab_rcu->addr = addr;
1903 call_rcu(&slab_rcu->head, kmem_rcu_free);
1904 } else {
1905 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001906 if (OFF_SLAB(cachep))
1907 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 }
1909}
1910
Andrew Mortona737b3e2006-03-22 00:08:11 -08001911/*
1912 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1913 * size of kmem_list3.
1914 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001915static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001916{
1917 int node;
1918
1919 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001920 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001921 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 REAPTIMEOUT_LIST3 +
1923 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001924 }
1925}
1926
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001927static void __kmem_cache_destroy(struct kmem_cache *cachep)
1928{
1929 int i;
1930 struct kmem_list3 *l3;
1931
1932 for_each_online_cpu(i)
1933 kfree(cachep->array[i]);
1934
1935 /* NUMA: free the list3 structures */
1936 for_each_online_node(i) {
1937 l3 = cachep->nodelists[i];
1938 if (l3) {
1939 kfree(l3->shared);
1940 free_alien_cache(l3->alien);
1941 kfree(l3);
1942 }
1943 }
1944 kmem_cache_free(&cache_cache, cachep);
1945}
1946
1947
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001949 * calculate_slab_order - calculate size (page order) of slabs
1950 * @cachep: pointer to the cache that is being created
1951 * @size: size of objects to be created in this cache.
1952 * @align: required alignment for the objects.
1953 * @flags: slab allocation flags
1954 *
1955 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001956 *
1957 * This could be made much more intelligent. For now, try to avoid using
1958 * high order pages for slabs. When the gfp() functions are more friendly
1959 * towards high-order requests, this should be changed.
1960 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001961static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001962 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001963{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001964 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001965 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001966 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001967
Andrew Mortona737b3e2006-03-22 00:08:11 -08001968 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001969 unsigned int num;
1970 size_t remainder;
1971
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001972 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001973 if (!num)
1974 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001975
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001976 if (flags & CFLGS_OFF_SLAB) {
1977 /*
1978 * Max number of objs-per-slab for caches which
1979 * use off-slab slabs. Needed to avoid a possible
1980 * looping condition in cache_grow().
1981 */
1982 offslab_limit = size - sizeof(struct slab);
1983 offslab_limit /= sizeof(kmem_bufctl_t);
1984
1985 if (num > offslab_limit)
1986 break;
1987 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001988
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001989 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001990 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001991 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001992 left_over = remainder;
1993
1994 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001995 * A VFS-reclaimable slab tends to have most allocations
1996 * as GFP_NOFS and we really don't want to have to be allocating
1997 * higher-order pages when we are unable to shrink dcache.
1998 */
1999 if (flags & SLAB_RECLAIM_ACCOUNT)
2000 break;
2001
2002 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002003 * Large number of objects is good, but very large slabs are
2004 * currently bad for the gfp()s.
2005 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002006 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002007 break;
2008
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002009 /*
2010 * Acceptable internal fragmentation?
2011 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002012 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002013 break;
2014 }
2015 return left_over;
2016}
2017
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002018static int setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002019{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002020 if (g_cpucache_up == FULL)
2021 return enable_cpucache(cachep);
2022
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002023 if (g_cpucache_up == NONE) {
2024 /*
2025 * Note: the first kmem_cache_create must create the cache
2026 * that's used by kmalloc(24), otherwise the creation of
2027 * further caches will BUG().
2028 */
2029 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2030
2031 /*
2032 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2033 * the first cache, then we need to set up all its list3s,
2034 * otherwise the creation of further caches will BUG().
2035 */
2036 set_up_list3s(cachep, SIZE_AC);
2037 if (INDEX_AC == INDEX_L3)
2038 g_cpucache_up = PARTIAL_L3;
2039 else
2040 g_cpucache_up = PARTIAL_AC;
2041 } else {
2042 cachep->array[smp_processor_id()] =
2043 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2044
2045 if (g_cpucache_up == PARTIAL_AC) {
2046 set_up_list3s(cachep, SIZE_L3);
2047 g_cpucache_up = PARTIAL_L3;
2048 } else {
2049 int node;
2050 for_each_online_node(node) {
2051 cachep->nodelists[node] =
2052 kmalloc_node(sizeof(struct kmem_list3),
2053 GFP_KERNEL, node);
2054 BUG_ON(!cachep->nodelists[node]);
2055 kmem_list3_init(cachep->nodelists[node]);
2056 }
2057 }
2058 }
2059 cachep->nodelists[numa_node_id()]->next_reap =
2060 jiffies + REAPTIMEOUT_LIST3 +
2061 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2062
2063 cpu_cache_get(cachep)->avail = 0;
2064 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2065 cpu_cache_get(cachep)->batchcount = 1;
2066 cpu_cache_get(cachep)->touched = 0;
2067 cachep->batchcount = 1;
2068 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002069 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002070}
2071
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002072/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 * kmem_cache_create - Create a cache.
2074 * @name: A string which is used in /proc/slabinfo to identify this cache.
2075 * @size: The size of objects to be created in this cache.
2076 * @align: The required alignment for the objects.
2077 * @flags: SLAB flags
2078 * @ctor: A constructor for the objects.
2079 * @dtor: A destructor for the objects.
2080 *
2081 * Returns a ptr to the cache on success, NULL on failure.
2082 * Cannot be called within a int, but can be interrupted.
2083 * The @ctor is run when new pages are allocated by the cache
2084 * and the @dtor is run before the pages are handed back.
2085 *
2086 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002087 * the module calling this has to destroy the cache before getting unloaded.
2088 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 * The flags are
2090 *
2091 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2092 * to catch references to uninitialised memory.
2093 *
2094 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2095 * for buffer overruns.
2096 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2098 * cacheline. This can be beneficial if you're counting cycles as closely
2099 * as davem.
2100 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002101struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002103 unsigned long flags,
2104 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002105 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106{
2107 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002108 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109
2110 /*
2111 * Sanity checks... these are all serious usage bugs.
2112 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002113 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002114 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002115 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2116 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002117 BUG();
2118 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002120 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002121 * We use cache_chain_mutex to ensure a consistent view of
2122 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002123 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002124 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002125
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002126 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002127 char tmp;
2128 int res;
2129
2130 /*
2131 * This happens when the module gets unloaded and doesn't
2132 * destroy its slab cache and no-one else reuses the vmalloc
2133 * area of the module. Print a warning.
2134 */
Andrew Morton138ae662006-12-06 20:36:41 -08002135 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002136 if (res) {
2137 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002138 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002139 continue;
2140 }
2141
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002142 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002143 printk("kmem_cache_create: duplicate cache %s\n", name);
2144 dump_stack();
2145 goto oops;
2146 }
2147 }
2148
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149#if DEBUG
2150 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2151 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2152 /* No constructor, but inital state check requested */
2153 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002154 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 flags &= ~SLAB_DEBUG_INITIAL;
2156 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157#if FORCED_DEBUG
2158 /*
2159 * Enable redzoning and last user accounting, except for caches with
2160 * large objects, if the increased size would increase the object size
2161 * above the next power of two: caches with object sizes just above a
2162 * power of two have a significant amount of internal fragmentation.
2163 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002164 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002165 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 if (!(flags & SLAB_DESTROY_BY_RCU))
2167 flags |= SLAB_POISON;
2168#endif
2169 if (flags & SLAB_DESTROY_BY_RCU)
2170 BUG_ON(flags & SLAB_POISON);
2171#endif
2172 if (flags & SLAB_DESTROY_BY_RCU)
2173 BUG_ON(dtor);
2174
2175 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002176 * Always checks flags, a caller might be expecting debug support which
2177 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002179 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180
Andrew Mortona737b3e2006-03-22 00:08:11 -08002181 /*
2182 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 * unaligned accesses for some archs when redzoning is used, and makes
2184 * sure any on-slab bufctl's are also correctly aligned.
2185 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002186 if (size & (BYTES_PER_WORD - 1)) {
2187 size += (BYTES_PER_WORD - 1);
2188 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 }
2190
Andrew Mortona737b3e2006-03-22 00:08:11 -08002191 /* calculate the final buffer alignment: */
2192
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 /* 1) arch recommendation: can be overridden for debug */
2194 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002195 /*
2196 * Default alignment: as specified by the arch code. Except if
2197 * an object is really small, then squeeze multiple objects into
2198 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 */
2200 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 ralign /= 2;
2203 } else {
2204 ralign = BYTES_PER_WORD;
2205 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002206
2207 /*
2208 * Redzoning and user store require word alignment. Note this will be
2209 * overridden by architecture or caller mandated alignment if either
2210 * is greater than BYTES_PER_WORD.
2211 */
2212 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2213 ralign = BYTES_PER_WORD;
2214
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002215 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 if (ralign < ARCH_SLAB_MINALIGN) {
2217 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002219 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 if (ralign < align) {
2221 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002223 /* disable debug if necessary */
2224 if (ralign > BYTES_PER_WORD)
2225 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002226 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002227 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 */
2229 align = ralign;
2230
2231 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002232 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002234 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235
2236#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002237 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238
Pekka Enbergca5f9702006-09-25 23:31:25 -07002239 /*
2240 * Both debugging options require word-alignment which is calculated
2241 * into align above.
2242 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002245 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002246 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 }
2248 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002249 /* user store requires one word storage behind the end of
2250 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 size += BYTES_PER_WORD;
2253 }
2254#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002255 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002256 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2257 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 size = PAGE_SIZE;
2259 }
2260#endif
2261#endif
2262
Ingo Molnare0a42722006-06-23 02:03:46 -07002263 /*
2264 * Determine if the slab management is 'on' or 'off' slab.
2265 * (bootstrapping cannot cope with offslab caches so don't do
2266 * it too early on.)
2267 */
2268 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 /*
2270 * Size is large, assume best to place the slab management obj
2271 * off-slab (should allow better packing of objs).
2272 */
2273 flags |= CFLGS_OFF_SLAB;
2274
2275 size = ALIGN(size, align);
2276
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002277 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278
2279 if (!cachep->num) {
2280 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2281 kmem_cache_free(&cache_cache, cachep);
2282 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002283 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002285 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2286 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
2288 /*
2289 * If the slab has been placed off-slab, and we have enough space then
2290 * move it on-slab. This is at the expense of any extra colouring.
2291 */
2292 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2293 flags &= ~CFLGS_OFF_SLAB;
2294 left_over -= slab_size;
2295 }
2296
2297 if (flags & CFLGS_OFF_SLAB) {
2298 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002299 slab_size =
2300 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 }
2302
2303 cachep->colour_off = cache_line_size();
2304 /* Offset must be a multiple of the alignment. */
2305 if (cachep->colour_off < align)
2306 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002307 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 cachep->slab_size = slab_size;
2309 cachep->flags = flags;
2310 cachep->gfpflags = 0;
2311 if (flags & SLAB_CACHE_DMA)
2312 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002313 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002315 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002316 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002317 /*
2318 * This is a possibility for one of the malloc_sizes caches.
2319 * But since we go off slab only for object size greater than
2320 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2321 * this should not happen at all.
2322 * But leave a BUG_ON for some lucky dude.
2323 */
2324 BUG_ON(!cachep->slabp_cache);
2325 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 cachep->ctor = ctor;
2327 cachep->dtor = dtor;
2328 cachep->name = name;
2329
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002330 if (setup_cpu_cache(cachep)) {
2331 __kmem_cache_destroy(cachep);
2332 cachep = NULL;
2333 goto oops;
2334 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 /* cache setup completed, link it into the list */
2337 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002338oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 if (!cachep && (flags & SLAB_PANIC))
2340 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002341 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002342 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 return cachep;
2344}
2345EXPORT_SYMBOL(kmem_cache_create);
2346
2347#if DEBUG
2348static void check_irq_off(void)
2349{
2350 BUG_ON(!irqs_disabled());
2351}
2352
2353static void check_irq_on(void)
2354{
2355 BUG_ON(irqs_disabled());
2356}
2357
Pekka Enberg343e0d72006-02-01 03:05:50 -08002358static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359{
2360#ifdef CONFIG_SMP
2361 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002362 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363#endif
2364}
Christoph Lametere498be72005-09-09 13:03:32 -07002365
Pekka Enberg343e0d72006-02-01 03:05:50 -08002366static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002367{
2368#ifdef CONFIG_SMP
2369 check_irq_off();
2370 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2371#endif
2372}
2373
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374#else
2375#define check_irq_off() do { } while(0)
2376#define check_irq_on() do { } while(0)
2377#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002378#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379#endif
2380
Christoph Lameteraab22072006-03-22 00:09:06 -08002381static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2382 struct array_cache *ac,
2383 int force, int node);
2384
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385static void do_drain(void *arg)
2386{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002387 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002389 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390
2391 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002392 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002393 spin_lock(&cachep->nodelists[node]->list_lock);
2394 free_block(cachep, ac->entry, ac->avail, node);
2395 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 ac->avail = 0;
2397}
2398
Pekka Enberg343e0d72006-02-01 03:05:50 -08002399static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400{
Christoph Lametere498be72005-09-09 13:03:32 -07002401 struct kmem_list3 *l3;
2402 int node;
2403
Andrew Mortona07fa392006-03-22 00:08:17 -08002404 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002406 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002407 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002408 if (l3 && l3->alien)
2409 drain_alien_cache(cachep, l3->alien);
2410 }
2411
2412 for_each_online_node(node) {
2413 l3 = cachep->nodelists[node];
2414 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002415 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002416 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417}
2418
Christoph Lametered11d9e2006-06-30 01:55:45 -07002419/*
2420 * Remove slabs from the list of free slabs.
2421 * Specify the number of slabs to drain in tofree.
2422 *
2423 * Returns the actual number of slabs released.
2424 */
2425static int drain_freelist(struct kmem_cache *cache,
2426 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002428 struct list_head *p;
2429 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431
Christoph Lametered11d9e2006-06-30 01:55:45 -07002432 nr_freed = 0;
2433 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434
Christoph Lametered11d9e2006-06-30 01:55:45 -07002435 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002436 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002437 if (p == &l3->slabs_free) {
2438 spin_unlock_irq(&l3->list_lock);
2439 goto out;
2440 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441
Christoph Lametered11d9e2006-06-30 01:55:45 -07002442 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002444 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445#endif
2446 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002447 /*
2448 * Safe to drop the lock. The slab is no longer linked
2449 * to the cache.
2450 */
2451 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002452 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002453 slab_destroy(cache, slabp);
2454 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002456out:
2457 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458}
2459
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002460/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002461static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002462{
2463 int ret = 0, i = 0;
2464 struct kmem_list3 *l3;
2465
2466 drain_cpu_caches(cachep);
2467
2468 check_irq_on();
2469 for_each_online_node(i) {
2470 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002471 if (!l3)
2472 continue;
2473
2474 drain_freelist(cachep, l3, l3->free_objects);
2475
2476 ret += !list_empty(&l3->slabs_full) ||
2477 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002478 }
2479 return (ret ? 1 : 0);
2480}
2481
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482/**
2483 * kmem_cache_shrink - Shrink a cache.
2484 * @cachep: The cache to shrink.
2485 *
2486 * Releases as many slabs as possible for a cache.
2487 * To help debugging, a zero exit status indicates all slabs were released.
2488 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002489int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002491 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002492 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002494 mutex_lock(&cache_chain_mutex);
2495 ret = __cache_shrink(cachep);
2496 mutex_unlock(&cache_chain_mutex);
2497 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498}
2499EXPORT_SYMBOL(kmem_cache_shrink);
2500
2501/**
2502 * kmem_cache_destroy - delete a cache
2503 * @cachep: the cache to destroy
2504 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002505 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 *
2507 * It is expected this function will be called by a module when it is
2508 * unloaded. This will remove the cache completely, and avoid a duplicate
2509 * cache being allocated each time a module is loaded and unloaded, if the
2510 * module doesn't have persistent in-kernel storage across loads and unloads.
2511 *
2512 * The cache must be empty before calling this function.
2513 *
2514 * The caller must guarantee that noone will allocate memory from the cache
2515 * during the kmem_cache_destroy().
2516 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002517void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002519 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002522 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 /*
2524 * the chain is never empty, cache_cache is never destroyed
2525 */
2526 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 if (__cache_shrink(cachep)) {
2528 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002529 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002530 mutex_unlock(&cache_chain_mutex);
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002531 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 }
2533
2534 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002535 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002537 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002538 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539}
2540EXPORT_SYMBOL(kmem_cache_destroy);
2541
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002542/*
2543 * Get the memory for a slab management obj.
2544 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2545 * always come from malloc_sizes caches. The slab descriptor cannot
2546 * come from the same cache which is getting created because,
2547 * when we are searching for an appropriate cache for these
2548 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2549 * If we are creating a malloc_sizes cache here it would not be visible to
2550 * kmem_find_general_cachep till the initialization is complete.
2551 * Hence we cannot have slabp_cache same as the original cache.
2552 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002553static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002554 int colour_off, gfp_t local_flags,
2555 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556{
2557 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002558
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 if (OFF_SLAB(cachep)) {
2560 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002561 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002562 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 if (!slabp)
2564 return NULL;
2565 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002566 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 colour_off += cachep->slab_size;
2568 }
2569 slabp->inuse = 0;
2570 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002571 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002572 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 return slabp;
2574}
2575
2576static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2577{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002578 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579}
2580
Pekka Enberg343e0d72006-02-01 03:05:50 -08002581static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002582 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583{
2584 int i;
2585
2586 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002587 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588#if DEBUG
2589 /* need to poison the objs? */
2590 if (cachep->flags & SLAB_POISON)
2591 poison_obj(cachep, objp, POISON_FREE);
2592 if (cachep->flags & SLAB_STORE_USER)
2593 *dbg_userword(cachep, objp) = NULL;
2594
2595 if (cachep->flags & SLAB_RED_ZONE) {
2596 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2597 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2598 }
2599 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002600 * Constructors are not allowed to allocate memory from the same
2601 * cache which they are a constructor for. Otherwise, deadlock.
2602 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 */
2604 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002605 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002606 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607
2608 if (cachep->flags & SLAB_RED_ZONE) {
2609 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2610 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002611 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2613 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002614 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002616 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2617 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002618 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002619 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620#else
2621 if (cachep->ctor)
2622 cachep->ctor(objp, cachep, ctor_flags);
2623#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002624 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002626 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 slabp->free = 0;
2628}
2629
Pekka Enberg343e0d72006-02-01 03:05:50 -08002630static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631{
Christoph Lameter441e1432006-12-06 20:33:19 -08002632 if (flags & GFP_DMA)
Andrew Mortona737b3e2006-03-22 00:08:11 -08002633 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2634 else
2635 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636}
2637
Andrew Mortona737b3e2006-03-22 00:08:11 -08002638static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2639 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002640{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002641 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002642 kmem_bufctl_t next;
2643
2644 slabp->inuse++;
2645 next = slab_bufctl(slabp)[slabp->free];
2646#if DEBUG
2647 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2648 WARN_ON(slabp->nodeid != nodeid);
2649#endif
2650 slabp->free = next;
2651
2652 return objp;
2653}
2654
Andrew Mortona737b3e2006-03-22 00:08:11 -08002655static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2656 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002657{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002658 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002659
2660#if DEBUG
2661 /* Verify that the slab belongs to the intended node */
2662 WARN_ON(slabp->nodeid != nodeid);
2663
Al Viro871751e2006-03-25 03:06:39 -08002664 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002665 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002666 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002667 BUG();
2668 }
2669#endif
2670 slab_bufctl(slabp)[objnr] = slabp->free;
2671 slabp->free = objnr;
2672 slabp->inuse--;
2673}
2674
Pekka Enberg47768742006-06-23 02:03:07 -07002675/*
2676 * Map pages beginning at addr to the given cache and slab. This is required
2677 * for the slab allocator to be able to lookup the cache and slab of a
2678 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2679 */
2680static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2681 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682{
Pekka Enberg47768742006-06-23 02:03:07 -07002683 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 struct page *page;
2685
Pekka Enberg47768742006-06-23 02:03:07 -07002686 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002687
Pekka Enberg47768742006-06-23 02:03:07 -07002688 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002689 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002690 nr_pages <<= cache->gfporder;
2691
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002693 page_set_cache(page, cache);
2694 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002696 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697}
2698
2699/*
2700 * Grow (by 1) the number of slabs within a cache. This is called by
2701 * kmem_cache_alloc() when there are no active objs left in a cache.
2702 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002703static int cache_grow(struct kmem_cache *cachep,
2704 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002706 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002707 size_t offset;
2708 gfp_t local_flags;
2709 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002710 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
Andrew Mortona737b3e2006-03-22 00:08:11 -08002712 /*
2713 * Be lazy and only check for valid flags here, keeping it out of the
2714 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 */
Christoph Lameter441e1432006-12-06 20:33:19 -08002716 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW));
Christoph Lameter6e0eaa42006-12-06 20:33:10 -08002717 if (flags & __GFP_NO_GROW)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 return 0;
2719
2720 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Christoph Lametera06d72c2006-12-06 20:33:12 -08002721 local_flags = (flags & GFP_LEVEL_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 if (!(local_flags & __GFP_WAIT))
2723 /*
2724 * Not allowed to sleep. Need to tell a constructor about
2725 * this - it might need to know...
2726 */
2727 ctor_flags |= SLAB_CTOR_ATOMIC;
2728
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002729 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002731 l3 = cachep->nodelists[nodeid];
2732 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
2734 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002735 offset = l3->colour_next;
2736 l3->colour_next++;
2737 if (l3->colour_next >= cachep->colour)
2738 l3->colour_next = 0;
2739 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002741 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742
2743 if (local_flags & __GFP_WAIT)
2744 local_irq_enable();
2745
2746 /*
2747 * The test for missing atomic flag is performed here, rather than
2748 * the more obvious place, simply to reduce the critical path length
2749 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2750 * will eventually be caught here (where it matters).
2751 */
2752 kmem_flagcheck(cachep, flags);
2753
Andrew Mortona737b3e2006-03-22 00:08:11 -08002754 /*
2755 * Get mem for the objs. Attempt to allocate a physical page from
2756 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002757 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002758 if (!objp)
2759 objp = kmem_getpages(cachep, flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002760 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 goto failed;
2762
2763 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002764 slabp = alloc_slabmgmt(cachep, objp, offset,
2765 local_flags & ~GFP_THISNODE, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002766 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 goto opps1;
2768
Christoph Lametere498be72005-09-09 13:03:32 -07002769 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002770 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
2772 cache_init_objs(cachep, slabp, ctor_flags);
2773
2774 if (local_flags & __GFP_WAIT)
2775 local_irq_disable();
2776 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002777 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
2779 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002780 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002782 l3->free_objects += cachep->num;
2783 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002785opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002787failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 if (local_flags & __GFP_WAIT)
2789 local_irq_disable();
2790 return 0;
2791}
2792
2793#if DEBUG
2794
2795/*
2796 * Perform extra freeing checks:
2797 * - detect bad pointers.
2798 * - POISON/RED_ZONE checking
2799 * - destructor calls, for caches with POISON+dtor
2800 */
2801static void kfree_debugcheck(const void *objp)
2802{
2803 struct page *page;
2804
2805 if (!virt_addr_valid(objp)) {
2806 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002807 (unsigned long)objp);
2808 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 }
2810 page = virt_to_page(objp);
2811 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002812 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2813 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 BUG();
2815 }
2816}
2817
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002818static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2819{
2820 unsigned long redzone1, redzone2;
2821
2822 redzone1 = *dbg_redzone1(cache, obj);
2823 redzone2 = *dbg_redzone2(cache, obj);
2824
2825 /*
2826 * Redzone is ok.
2827 */
2828 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2829 return;
2830
2831 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2832 slab_error(cache, "double free detected");
2833 else
2834 slab_error(cache, "memory outside object was overwritten");
2835
2836 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2837 obj, redzone1, redzone2);
2838}
2839
Pekka Enberg343e0d72006-02-01 03:05:50 -08002840static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002841 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842{
2843 struct page *page;
2844 unsigned int objnr;
2845 struct slab *slabp;
2846
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002847 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 kfree_debugcheck(objp);
2849 page = virt_to_page(objp);
2850
Pekka Enberg065d41c2005-11-13 16:06:46 -08002851 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852
2853 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002854 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2856 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2857 }
2858 if (cachep->flags & SLAB_STORE_USER)
2859 *dbg_userword(cachep, objp) = caller;
2860
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002861 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862
2863 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002864 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865
2866 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002867 /*
2868 * Need to call the slab's constructor so the caller can
2869 * perform a verify of its state (debugging). Called without
2870 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002872 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002873 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 }
2875 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2876 /* we want to cache poison the object,
2877 * call the destruction callback
2878 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002879 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 }
Al Viro871751e2006-03-25 03:06:39 -08002881#ifdef CONFIG_DEBUG_SLAB_LEAK
2882 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2883#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 if (cachep->flags & SLAB_POISON) {
2885#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002886 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002888 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002889 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 } else {
2891 poison_obj(cachep, objp, POISON_FREE);
2892 }
2893#else
2894 poison_obj(cachep, objp, POISON_FREE);
2895#endif
2896 }
2897 return objp;
2898}
2899
Pekka Enberg343e0d72006-02-01 03:05:50 -08002900static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901{
2902 kmem_bufctl_t i;
2903 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002904
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 /* Check slab's freelist to see if this obj is there. */
2906 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2907 entries++;
2908 if (entries > cachep->num || i >= cachep->num)
2909 goto bad;
2910 }
2911 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002912bad:
2913 printk(KERN_ERR "slab: Internal list corruption detected in "
2914 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2915 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002916 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002917 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002918 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002919 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002921 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922 }
2923 printk("\n");
2924 BUG();
2925 }
2926}
2927#else
2928#define kfree_debugcheck(x) do { } while(0)
2929#define cache_free_debugcheck(x,objp,z) (objp)
2930#define check_slabp(x,y) do { } while(0)
2931#endif
2932
Pekka Enberg343e0d72006-02-01 03:05:50 -08002933static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934{
2935 int batchcount;
2936 struct kmem_list3 *l3;
2937 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002938 int node;
2939
2940 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
2942 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002943 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002944retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 batchcount = ac->batchcount;
2946 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002947 /*
2948 * If there was little recent activity on this cache, then
2949 * perform only a partial refill. Otherwise we could generate
2950 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 */
2952 batchcount = BATCHREFILL_LIMIT;
2953 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002954 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955
Christoph Lametere498be72005-09-09 13:03:32 -07002956 BUG_ON(ac->avail > 0 || !l3);
2957 spin_lock(&l3->list_lock);
2958
Christoph Lameter3ded1752006-03-25 03:06:44 -08002959 /* See if we can refill from the shared array */
2960 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2961 goto alloc_done;
2962
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 while (batchcount > 0) {
2964 struct list_head *entry;
2965 struct slab *slabp;
2966 /* Get slab alloc is to come from. */
2967 entry = l3->slabs_partial.next;
2968 if (entry == &l3->slabs_partial) {
2969 l3->free_touched = 1;
2970 entry = l3->slabs_free.next;
2971 if (entry == &l3->slabs_free)
2972 goto must_grow;
2973 }
2974
2975 slabp = list_entry(entry, struct slab, list);
2976 check_slabp(cachep, slabp);
2977 check_spinlock_acquired(cachep);
2978 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 STATS_INC_ALLOCED(cachep);
2980 STATS_INC_ACTIVE(cachep);
2981 STATS_SET_HIGH(cachep);
2982
Matthew Dobson78d382d2006-02-01 03:05:47 -08002983 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002984 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 }
2986 check_slabp(cachep, slabp);
2987
2988 /* move slabp to correct slabp list: */
2989 list_del(&slabp->list);
2990 if (slabp->free == BUFCTL_END)
2991 list_add(&slabp->list, &l3->slabs_full);
2992 else
2993 list_add(&slabp->list, &l3->slabs_partial);
2994 }
2995
Andrew Mortona737b3e2006-03-22 00:08:11 -08002996must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002998alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002999 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000
3001 if (unlikely(!ac->avail)) {
3002 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003003 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003004
Andrew Mortona737b3e2006-03-22 00:08:11 -08003005 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003006 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003007 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008 return NULL;
3009
Andrew Mortona737b3e2006-03-22 00:08:11 -08003010 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 goto retry;
3012 }
3013 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003014 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015}
3016
Andrew Mortona737b3e2006-03-22 00:08:11 -08003017static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3018 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019{
3020 might_sleep_if(flags & __GFP_WAIT);
3021#if DEBUG
3022 kmem_flagcheck(cachep, flags);
3023#endif
3024}
3025
3026#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003027static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3028 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003030 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003032 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003034 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003035 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003036 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 else
3038 check_poison_obj(cachep, objp);
3039#else
3040 check_poison_obj(cachep, objp);
3041#endif
3042 poison_obj(cachep, objp, POISON_INUSE);
3043 }
3044 if (cachep->flags & SLAB_STORE_USER)
3045 *dbg_userword(cachep, objp) = caller;
3046
3047 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003048 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3049 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3050 slab_error(cachep, "double free, or memory outside"
3051 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003052 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08003053 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3054 objp, *dbg_redzone1(cachep, objp),
3055 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 }
3057 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3058 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3059 }
Al Viro871751e2006-03-25 03:06:39 -08003060#ifdef CONFIG_DEBUG_SLAB_LEAK
3061 {
3062 struct slab *slabp;
3063 unsigned objnr;
3064
3065 slabp = page_get_slab(virt_to_page(objp));
3066 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3067 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3068 }
3069#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003070 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003072 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073
3074 if (!(flags & __GFP_WAIT))
3075 ctor_flags |= SLAB_CTOR_ATOMIC;
3076
3077 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003078 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003079#if ARCH_SLAB_MINALIGN
3080 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3081 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3082 objp, ARCH_SLAB_MINALIGN);
3083 }
3084#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 return objp;
3086}
3087#else
3088#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3089#endif
3090
Pekka Enberg343e0d72006-02-01 03:05:50 -08003091static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003093 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 struct array_cache *ac;
3095
Alok N Kataria5c382302005-09-27 21:45:46 -07003096 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003097 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 if (likely(ac->avail)) {
3099 STATS_INC_ALLOCHIT(cachep);
3100 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003101 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102 } else {
3103 STATS_INC_ALLOCMISS(cachep);
3104 objp = cache_alloc_refill(cachep, flags);
3105 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003106 return objp;
3107}
3108
Andrew Mortona737b3e2006-03-22 00:08:11 -08003109static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3110 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07003111{
3112 unsigned long save_flags;
Christoph Lameterde3083e2006-09-27 01:50:03 -07003113 void *objp = NULL;
Alok N Kataria5c382302005-09-27 21:45:46 -07003114
3115 cache_alloc_debugcheck_before(cachep, flags);
3116
3117 local_irq_save(save_flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003118
Christoph Lameter765c4502006-09-27 01:50:08 -07003119 if (unlikely(NUMA_BUILD &&
3120 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
Christoph Lameterde3083e2006-09-27 01:50:03 -07003121 objp = alternate_node_alloc(cachep, flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003122
3123 if (!objp)
3124 objp = ____cache_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003125 /*
3126 * We may just have run out of memory on the local node.
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003127 * ____cache_alloc_node() knows how to locate memory on other nodes
Christoph Lameter765c4502006-09-27 01:50:08 -07003128 */
3129 if (NUMA_BUILD && !objp)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003130 objp = ____cache_alloc_node(cachep, flags, numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07003132 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003133 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07003134 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135 return objp;
3136}
3137
Christoph Lametere498be72005-09-09 13:03:32 -07003138#ifdef CONFIG_NUMA
3139/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003140 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003141 *
3142 * If we are in_interrupt, then process context, including cpusets and
3143 * mempolicy, may not apply and should not be used for allocation policy.
3144 */
3145static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3146{
3147 int nid_alloc, nid_here;
3148
Christoph Lameter765c4502006-09-27 01:50:08 -07003149 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003150 return NULL;
3151 nid_alloc = nid_here = numa_node_id();
3152 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3153 nid_alloc = cpuset_mem_spread_node();
3154 else if (current->mempolicy)
3155 nid_alloc = slab_node(current->mempolicy);
3156 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003157 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003158 return NULL;
3159}
3160
3161/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003162 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003163 * certain node and fall back is permitted. First we scan all the
3164 * available nodelists for available objects. If that fails then we
3165 * perform an allocation without specifying a node. This allows the page
3166 * allocator to do its reclaim / fallback magic. We then insert the
3167 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003168 */
3169void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3170{
3171 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3172 ->node_zonelists[gfp_zone(flags)];
3173 struct zone **z;
3174 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003175 int nid;
Christoph Lameter765c4502006-09-27 01:50:08 -07003176
Christoph Lameter3c517a62006-12-06 20:33:29 -08003177retry:
3178 /*
3179 * Look through allowed nodes for objects available
3180 * from existing per node queues.
3181 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003182 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003183 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003184
Christoph Lameter3c517a62006-12-06 20:33:29 -08003185 if (cpuset_zone_allowed(*z, flags) &&
3186 cache->nodelists[nid] &&
3187 cache->nodelists[nid]->free_objects)
3188 obj = ____cache_alloc_node(cache,
3189 flags | GFP_THISNODE, nid);
3190 }
3191
3192 if (!obj) {
3193 /*
3194 * This allocation will be performed within the constraints
3195 * of the current cpuset / memory policy requirements.
3196 * We may trigger various forms of reclaim on the allowed
3197 * set and go into memory reserves if necessary.
3198 */
3199 obj = kmem_getpages(cache, flags, -1);
3200 if (obj) {
3201 /*
3202 * Insert into the appropriate per node queues
3203 */
3204 nid = page_to_nid(virt_to_page(obj));
3205 if (cache_grow(cache, flags, nid, obj)) {
3206 obj = ____cache_alloc_node(cache,
3207 flags | GFP_THISNODE, nid);
3208 if (!obj)
3209 /*
3210 * Another processor may allocate the
3211 * objects in the slab since we are
3212 * not holding any locks.
3213 */
3214 goto retry;
3215 } else {
3216 kmem_freepages(cache, obj);
3217 obj = NULL;
3218 }
3219 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003220 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003221 return obj;
3222}
3223
3224/*
Christoph Lametere498be72005-09-09 13:03:32 -07003225 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003227static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003228 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003229{
3230 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003231 struct slab *slabp;
3232 struct kmem_list3 *l3;
3233 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003234 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003236 l3 = cachep->nodelists[nodeid];
3237 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003238
Andrew Mortona737b3e2006-03-22 00:08:11 -08003239retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003240 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003241 spin_lock(&l3->list_lock);
3242 entry = l3->slabs_partial.next;
3243 if (entry == &l3->slabs_partial) {
3244 l3->free_touched = 1;
3245 entry = l3->slabs_free.next;
3246 if (entry == &l3->slabs_free)
3247 goto must_grow;
3248 }
Christoph Lametere498be72005-09-09 13:03:32 -07003249
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003250 slabp = list_entry(entry, struct slab, list);
3251 check_spinlock_acquired_node(cachep, nodeid);
3252 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003253
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003254 STATS_INC_NODEALLOCS(cachep);
3255 STATS_INC_ACTIVE(cachep);
3256 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003257
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003258 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003259
Matthew Dobson78d382d2006-02-01 03:05:47 -08003260 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003261 check_slabp(cachep, slabp);
3262 l3->free_objects--;
3263 /* move slabp to correct slabp list: */
3264 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003265
Andrew Mortona737b3e2006-03-22 00:08:11 -08003266 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003267 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003268 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003269 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003270
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003271 spin_unlock(&l3->list_lock);
3272 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003273
Andrew Mortona737b3e2006-03-22 00:08:11 -08003274must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003275 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003276 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003277 if (x)
3278 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003279
Christoph Lameter765c4502006-09-27 01:50:08 -07003280 if (!(flags & __GFP_THISNODE))
3281 /* Unable to grow the cache. Fall back to other nodes. */
3282 return fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003283
Christoph Lameter765c4502006-09-27 01:50:08 -07003284 return NULL;
3285
Andrew Mortona737b3e2006-03-22 00:08:11 -08003286done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003287 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003288}
3289#endif
3290
3291/*
3292 * Caller needs to acquire correct kmem_list's list_lock
3293 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003294static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003295 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296{
3297 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003298 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
3300 for (i = 0; i < nr_objects; i++) {
3301 void *objp = objpp[i];
3302 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003304 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003305 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003307 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003309 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003311 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 check_slabp(cachep, slabp);
3313
3314 /* fixup slab chains */
3315 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003316 if (l3->free_objects > l3->free_limit) {
3317 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003318 /* No need to drop any previously held
3319 * lock here, even if we have a off-slab slab
3320 * descriptor it is guaranteed to come from
3321 * a different cache, refer to comments before
3322 * alloc_slabmgmt.
3323 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324 slab_destroy(cachep, slabp);
3325 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003326 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 }
3328 } else {
3329 /* Unconditionally move a slab to the end of the
3330 * partial list on free - maximum time for the
3331 * other objects to be freed, too.
3332 */
Christoph Lametere498be72005-09-09 13:03:32 -07003333 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334 }
3335 }
3336}
3337
Pekka Enberg343e0d72006-02-01 03:05:50 -08003338static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339{
3340 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003341 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003342 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343
3344 batchcount = ac->batchcount;
3345#if DEBUG
3346 BUG_ON(!batchcount || batchcount > ac->avail);
3347#endif
3348 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003349 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003350 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003351 if (l3->shared) {
3352 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003353 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354 if (max) {
3355 if (batchcount > max)
3356 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003357 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003358 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 shared_array->avail += batchcount;
3360 goto free_done;
3361 }
3362 }
3363
Christoph Lameterff694162005-09-22 21:44:02 -07003364 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003365free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366#if STATS
3367 {
3368 int i = 0;
3369 struct list_head *p;
3370
Christoph Lametere498be72005-09-09 13:03:32 -07003371 p = l3->slabs_free.next;
3372 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 struct slab *slabp;
3374
3375 slabp = list_entry(p, struct slab, list);
3376 BUG_ON(slabp->inuse);
3377
3378 i++;
3379 p = p->next;
3380 }
3381 STATS_SET_FREEABLE(cachep, i);
3382 }
3383#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003384 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003386 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387}
3388
3389/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003390 * Release an obj back to its cache. If the obj has a constructed state, it must
3391 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003393static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003395 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
3397 check_irq_off();
3398 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3399
Ingo Molnar873623d2006-07-13 14:44:38 +02003400 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003401 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003402
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 if (likely(ac->avail < ac->limit)) {
3404 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003405 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406 return;
3407 } else {
3408 STATS_INC_FREEMISS(cachep);
3409 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003410 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411 }
3412}
3413
3414/**
3415 * kmem_cache_alloc - Allocate an object
3416 * @cachep: The cache to allocate from.
3417 * @flags: See kmalloc().
3418 *
3419 * Allocate an object from this cache. The flags are only relevant
3420 * if the cache has no available objects.
3421 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003422void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003424 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425}
3426EXPORT_SYMBOL(kmem_cache_alloc);
3427
3428/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003429 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003430 * @cache: The cache to allocate from.
3431 * @flags: See kmalloc().
3432 *
3433 * Allocate an object from this cache and set the allocated memory to zero.
3434 * The flags are only relevant if the cache has no available objects.
3435 */
3436void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3437{
3438 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3439 if (ret)
3440 memset(ret, 0, obj_size(cache));
3441 return ret;
3442}
3443EXPORT_SYMBOL(kmem_cache_zalloc);
3444
3445/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 * kmem_ptr_validate - check if an untrusted pointer might
3447 * be a slab entry.
3448 * @cachep: the cache we're checking against
3449 * @ptr: pointer to validate
3450 *
3451 * This verifies that the untrusted pointer looks sane:
3452 * it is _not_ a guarantee that the pointer is actually
3453 * part of the slab cache in question, but it at least
3454 * validates that the pointer can be dereferenced and
3455 * looks half-way sane.
3456 *
3457 * Currently only used for dentry validation.
3458 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003459int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003461 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003463 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003464 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 struct page *page;
3466
3467 if (unlikely(addr < min_addr))
3468 goto out;
3469 if (unlikely(addr > (unsigned long)high_memory - size))
3470 goto out;
3471 if (unlikely(addr & align_mask))
3472 goto out;
3473 if (unlikely(!kern_addr_valid(addr)))
3474 goto out;
3475 if (unlikely(!kern_addr_valid(addr + size - 1)))
3476 goto out;
3477 page = virt_to_page(ptr);
3478 if (unlikely(!PageSlab(page)))
3479 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003480 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 goto out;
3482 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003483out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 return 0;
3485}
3486
3487#ifdef CONFIG_NUMA
3488/**
3489 * kmem_cache_alloc_node - Allocate an object on the specified node
3490 * @cachep: The cache to allocate from.
3491 * @flags: See kmalloc().
3492 * @nodeid: node number of the target node.
3493 *
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003494 * Identical to kmem_cache_alloc but it will allocate memory on the given
3495 * node, which can improve the performance for cpu bound structures.
3496 *
3497 * Fallback to other node is possible if __GFP_THISNODE is not set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003499static __always_inline void *
3500__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3501 int nodeid, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502{
Christoph Lametere498be72005-09-09 13:03:32 -07003503 unsigned long save_flags;
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003504 void *ptr = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505
Christoph Lametere498be72005-09-09 13:03:32 -07003506 cache_alloc_debugcheck_before(cachep, flags);
3507 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003508
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003509 if (unlikely(nodeid == -1))
3510 nodeid = numa_node_id();
Christoph Lameter18f820f2006-02-01 03:05:43 -08003511
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003512 if (likely(cachep->nodelists[nodeid])) {
3513 if (nodeid == numa_node_id()) {
3514 /*
3515 * Use the locally cached objects if possible.
3516 * However ____cache_alloc does not allow fallback
3517 * to other nodes. It may fail while we still have
3518 * objects on other nodes available.
3519 */
3520 ptr = ____cache_alloc(cachep, flags);
3521 }
3522 if (!ptr) {
3523 /* ___cache_alloc_node can fall back to other nodes */
3524 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3525 }
3526 } else {
3527 /* Node not bootstrapped yet */
3528 if (!(flags & __GFP_THISNODE))
3529 ptr = fallback_alloc(cachep, flags);
3530 }
3531
3532 local_irq_restore(save_flags);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003533 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534
Christoph Lametere498be72005-09-09 13:03:32 -07003535 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003537
3538void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3539{
3540 return __cache_alloc_node(cachep, flags, nodeid,
3541 __builtin_return_address(0));
3542}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543EXPORT_SYMBOL(kmem_cache_alloc_node);
3544
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003545static __always_inline void *
3546__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003547{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003548 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003549
3550 cachep = kmem_find_general_cachep(size, flags);
3551 if (unlikely(cachep == NULL))
3552 return NULL;
3553 return kmem_cache_alloc_node(cachep, flags, node);
3554}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003555
3556#ifdef CONFIG_DEBUG_SLAB
3557void *__kmalloc_node(size_t size, gfp_t flags, int node)
3558{
3559 return __do_kmalloc_node(size, flags, node,
3560 __builtin_return_address(0));
3561}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003562EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003563
3564void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3565 int node, void *caller)
3566{
3567 return __do_kmalloc_node(size, flags, node, caller);
3568}
3569EXPORT_SYMBOL(__kmalloc_node_track_caller);
3570#else
3571void *__kmalloc_node(size_t size, gfp_t flags, int node)
3572{
3573 return __do_kmalloc_node(size, flags, node, NULL);
3574}
3575EXPORT_SYMBOL(__kmalloc_node);
3576#endif /* CONFIG_DEBUG_SLAB */
3577#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578
3579/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003580 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003582 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003583 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003585static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3586 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003588 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003590 /* If you want to save a few bytes .text space: replace
3591 * __ with kmem_.
3592 * Then kmalloc uses the uninlined functions instead of the inline
3593 * functions.
3594 */
3595 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003596 if (unlikely(cachep == NULL))
3597 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003598 return __cache_alloc(cachep, flags, caller);
3599}
3600
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003601
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003602#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003603void *__kmalloc(size_t size, gfp_t flags)
3604{
Al Viro871751e2006-03-25 03:06:39 -08003605 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606}
3607EXPORT_SYMBOL(__kmalloc);
3608
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003609void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3610{
3611 return __do_kmalloc(size, flags, caller);
3612}
3613EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003614
3615#else
3616void *__kmalloc(size_t size, gfp_t flags)
3617{
3618 return __do_kmalloc(size, flags, NULL);
3619}
3620EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003621#endif
3622
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623/**
3624 * kmem_cache_free - Deallocate an object
3625 * @cachep: The cache the allocation was from.
3626 * @objp: The previously allocated object.
3627 *
3628 * Free an object which was previously allocated from this
3629 * cache.
3630 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003631void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632{
3633 unsigned long flags;
3634
Pekka Enbergddc2e812006-06-23 02:03:40 -07003635 BUG_ON(virt_to_cache(objp) != cachep);
3636
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637 local_irq_save(flags);
Ingo Molnar873623d2006-07-13 14:44:38 +02003638 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 local_irq_restore(flags);
3640}
3641EXPORT_SYMBOL(kmem_cache_free);
3642
3643/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 * kfree - free previously allocated memory
3645 * @objp: pointer returned by kmalloc.
3646 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003647 * If @objp is NULL, no operation is performed.
3648 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649 * Don't free memory not originally allocated by kmalloc()
3650 * or you will run into trouble.
3651 */
3652void kfree(const void *objp)
3653{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003654 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655 unsigned long flags;
3656
3657 if (unlikely(!objp))
3658 return;
3659 local_irq_save(flags);
3660 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003661 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003662 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003663 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 local_irq_restore(flags);
3665}
3666EXPORT_SYMBOL(kfree);
3667
Pekka Enberg343e0d72006-02-01 03:05:50 -08003668unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003670 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671}
3672EXPORT_SYMBOL(kmem_cache_size);
3673
Pekka Enberg343e0d72006-02-01 03:05:50 -08003674const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003675{
3676 return cachep->name;
3677}
3678EXPORT_SYMBOL_GPL(kmem_cache_name);
3679
Christoph Lametere498be72005-09-09 13:03:32 -07003680/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003681 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003682 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003683static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003684{
3685 int node;
3686 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003687 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003688 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003689
3690 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003691
Paul Menage3395ee02006-12-06 20:32:16 -08003692 if (use_alien_caches) {
3693 new_alien = alloc_alien_cache(node, cachep->limit);
3694 if (!new_alien)
3695 goto fail;
3696 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003697
Christoph Lameter0718dc22006-03-25 03:06:47 -08003698 new_shared = alloc_arraycache(node,
3699 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003700 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003701 if (!new_shared) {
3702 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003703 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003704 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003705
Andrew Mortona737b3e2006-03-22 00:08:11 -08003706 l3 = cachep->nodelists[node];
3707 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003708 struct array_cache *shared = l3->shared;
3709
Christoph Lametere498be72005-09-09 13:03:32 -07003710 spin_lock_irq(&l3->list_lock);
3711
Christoph Lametercafeb022006-03-25 03:06:46 -08003712 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003713 free_block(cachep, shared->entry,
3714 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003715
Christoph Lametercafeb022006-03-25 03:06:46 -08003716 l3->shared = new_shared;
3717 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003718 l3->alien = new_alien;
3719 new_alien = NULL;
3720 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003721 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003722 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003723 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003724 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003725 free_alien_cache(new_alien);
3726 continue;
3727 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003728 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003729 if (!l3) {
3730 free_alien_cache(new_alien);
3731 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003732 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003733 }
Christoph Lametere498be72005-09-09 13:03:32 -07003734
3735 kmem_list3_init(l3);
3736 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003737 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003738 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003739 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003740 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003741 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003742 cachep->nodelists[node] = l3;
3743 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003744 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003745
Andrew Mortona737b3e2006-03-22 00:08:11 -08003746fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003747 if (!cachep->next.next) {
3748 /* Cache is not active yet. Roll back what we did */
3749 node--;
3750 while (node >= 0) {
3751 if (cachep->nodelists[node]) {
3752 l3 = cachep->nodelists[node];
3753
3754 kfree(l3->shared);
3755 free_alien_cache(l3->alien);
3756 kfree(l3);
3757 cachep->nodelists[node] = NULL;
3758 }
3759 node--;
3760 }
3761 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003762 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003763}
3764
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003766 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767 struct array_cache *new[NR_CPUS];
3768};
3769
3770static void do_ccupdate_local(void *info)
3771{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003772 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773 struct array_cache *old;
3774
3775 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003776 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003777
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3779 new->new[smp_processor_id()] = old;
3780}
3781
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003782/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003783static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3784 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003786 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003787 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003789 new = kzalloc(sizeof(*new), GFP_KERNEL);
3790 if (!new)
3791 return -ENOMEM;
3792
Christoph Lametere498be72005-09-09 13:03:32 -07003793 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003794 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003795 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003796 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003797 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003798 kfree(new->new[i]);
3799 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003800 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801 }
3802 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003803 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003805 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003806
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 cachep->batchcount = batchcount;
3809 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003810 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811
Christoph Lametere498be72005-09-09 13:03:32 -07003812 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003813 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814 if (!ccold)
3815 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003816 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003817 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003818 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 kfree(ccold);
3820 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003821 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003822 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823}
3824
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003825/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003826static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827{
3828 int err;
3829 int limit, shared;
3830
Andrew Mortona737b3e2006-03-22 00:08:11 -08003831 /*
3832 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 * - create a LIFO ordering, i.e. return objects that are cache-warm
3834 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003835 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 * bufctl chains: array operations are cheaper.
3837 * The numbers are guessed, we should auto-tune as described by
3838 * Bonwick.
3839 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003840 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003842 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003844 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003846 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847 limit = 54;
3848 else
3849 limit = 120;
3850
Andrew Mortona737b3e2006-03-22 00:08:11 -08003851 /*
3852 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003853 * allocation behaviour: Most allocs on one cpu, most free operations
3854 * on another cpu. For these cases, an efficient object passing between
3855 * cpus is necessary. This is provided by a shared array. The array
3856 * replaces Bonwick's magazine layer.
3857 * On uniprocessor, it's functionally equivalent (but less efficient)
3858 * to a larger limit. Thus disabled by default.
3859 */
3860 shared = 0;
3861#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003862 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 shared = 8;
3864#endif
3865
3866#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003867 /*
3868 * With debugging enabled, large batchcount lead to excessively long
3869 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870 */
3871 if (limit > 32)
3872 limit = 32;
3873#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003874 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 if (err)
3876 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003877 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003878 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879}
3880
Christoph Lameter1b552532006-03-22 00:09:07 -08003881/*
3882 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003883 * necessary. Note that the l3 listlock also protects the array_cache
3884 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003885 */
3886void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3887 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888{
3889 int tofree;
3890
Christoph Lameter1b552532006-03-22 00:09:07 -08003891 if (!ac || !ac->avail)
3892 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893 if (ac->touched && !force) {
3894 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003895 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003896 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003897 if (ac->avail) {
3898 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3899 if (tofree > ac->avail)
3900 tofree = (ac->avail + 1) / 2;
3901 free_block(cachep, ac->entry, tofree, node);
3902 ac->avail -= tofree;
3903 memmove(ac->entry, &(ac->entry[tofree]),
3904 sizeof(void *) * ac->avail);
3905 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003906 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907 }
3908}
3909
3910/**
3911 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003912 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 *
3914 * Called from workqueue/eventd every few seconds.
3915 * Purpose:
3916 * - clear the per-cpu caches for this CPU.
3917 * - return freeable pages to the main free memory pool.
3918 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003919 * If we cannot acquire the cache chain mutex then just give up - we'll try
3920 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921 */
David Howells65f27f32006-11-22 14:55:48 +00003922static void cache_reap(struct work_struct *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003924 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07003925 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003926 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003928 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003930 schedule_delayed_work(&__get_cpu_var(reap_work),
3931 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932 return;
3933 }
3934
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003935 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936 check_irq_on();
3937
Christoph Lameter35386e32006-03-22 00:09:05 -08003938 /*
3939 * We only take the l3 lock if absolutely necessary and we
3940 * have established with reasonable certainty that
3941 * we can do some work if the lock was obtained.
3942 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003943 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003944
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003945 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946
Christoph Lameteraab22072006-03-22 00:09:06 -08003947 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948
Christoph Lameter35386e32006-03-22 00:09:05 -08003949 /*
3950 * These are racy checks but it does not matter
3951 * if we skip one check or scan twice.
3952 */
Christoph Lametere498be72005-09-09 13:03:32 -07003953 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003954 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955
Christoph Lametere498be72005-09-09 13:03:32 -07003956 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957
Christoph Lameteraab22072006-03-22 00:09:06 -08003958 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959
Christoph Lametered11d9e2006-06-30 01:55:45 -07003960 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07003961 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003962 else {
3963 int freed;
3964
3965 freed = drain_freelist(searchp, l3, (l3->free_limit +
3966 5 * searchp->num - 1) / (5 * searchp->num));
3967 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003969next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 cond_resched();
3971 }
3972 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003973 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003974 next_reap_node();
Christoph Lameter2244b952006-06-30 01:55:33 -07003975 refresh_cpu_vm_stats(smp_processor_id());
Andrew Mortona737b3e2006-03-22 00:08:11 -08003976 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003977 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978}
3979
3980#ifdef CONFIG_PROC_FS
3981
Pekka Enberg85289f92006-01-08 01:00:36 -08003982static void print_slabinfo_header(struct seq_file *m)
3983{
3984 /*
3985 * Output format version, so at least we can change it
3986 * without _too_ many complaints.
3987 */
3988#if STATS
3989 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3990#else
3991 seq_puts(m, "slabinfo - version: 2.1\n");
3992#endif
3993 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3994 "<objperslab> <pagesperslab>");
3995 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3996 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3997#if STATS
3998 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003999 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004000 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4001#endif
4002 seq_putc(m, '\n');
4003}
4004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005static void *s_start(struct seq_file *m, loff_t *pos)
4006{
4007 loff_t n = *pos;
4008 struct list_head *p;
4009
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004010 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004011 if (!n)
4012 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013 p = cache_chain.next;
4014 while (n--) {
4015 p = p->next;
4016 if (p == &cache_chain)
4017 return NULL;
4018 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08004019 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020}
4021
4022static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4023{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004024 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004026 return cachep->next.next == &cache_chain ?
4027 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028}
4029
4030static void s_stop(struct seq_file *m, void *p)
4031{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004032 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033}
4034
4035static int s_show(struct seq_file *m, void *p)
4036{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004037 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004038 struct slab *slabp;
4039 unsigned long active_objs;
4040 unsigned long num_objs;
4041 unsigned long active_slabs = 0;
4042 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004043 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004045 int node;
4046 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 active_objs = 0;
4049 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004050 for_each_online_node(node) {
4051 l3 = cachep->nodelists[node];
4052 if (!l3)
4053 continue;
4054
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004055 check_irq_on();
4056 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004057
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004058 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004059 if (slabp->inuse != cachep->num && !error)
4060 error = "slabs_full accounting error";
4061 active_objs += cachep->num;
4062 active_slabs++;
4063 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004064 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004065 if (slabp->inuse == cachep->num && !error)
4066 error = "slabs_partial inuse accounting error";
4067 if (!slabp->inuse && !error)
4068 error = "slabs_partial/inuse accounting error";
4069 active_objs += slabp->inuse;
4070 active_slabs++;
4071 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004072 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004073 if (slabp->inuse && !error)
4074 error = "slabs_free/inuse accounting error";
4075 num_slabs++;
4076 }
4077 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004078 if (l3->shared)
4079 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004080
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004081 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004083 num_slabs += active_slabs;
4084 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004085 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086 error = "free_objects accounting error";
4087
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004088 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 if (error)
4090 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4091
4092 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004093 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004094 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004096 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004097 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004098 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004100 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 unsigned long high = cachep->high_mark;
4102 unsigned long allocs = cachep->num_allocations;
4103 unsigned long grown = cachep->grown;
4104 unsigned long reaped = cachep->reaped;
4105 unsigned long errors = cachep->errors;
4106 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004108 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004109 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110
Christoph Lametere498be72005-09-09 13:03:32 -07004111 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004112 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004113 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004114 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115 }
4116 /* cpu stats */
4117 {
4118 unsigned long allochit = atomic_read(&cachep->allochit);
4119 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4120 unsigned long freehit = atomic_read(&cachep->freehit);
4121 unsigned long freemiss = atomic_read(&cachep->freemiss);
4122
4123 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004124 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125 }
4126#endif
4127 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 return 0;
4129}
4130
4131/*
4132 * slabinfo_op - iterator that generates /proc/slabinfo
4133 *
4134 * Output layout:
4135 * cache-name
4136 * num-active-objs
4137 * total-objs
4138 * object size
4139 * num-active-slabs
4140 * total-slabs
4141 * num-pages-per-slab
4142 * + further values on SMP and with statistics enabled
4143 */
4144
Helge Deller15ad7cd2006-12-06 20:40:36 -08004145const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004146 .start = s_start,
4147 .next = s_next,
4148 .stop = s_stop,
4149 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150};
4151
4152#define MAX_SLABINFO_WRITE 128
4153/**
4154 * slabinfo_write - Tuning for the slab allocator
4155 * @file: unused
4156 * @buffer: user buffer
4157 * @count: data length
4158 * @ppos: unused
4159 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004160ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4161 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004163 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004165 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004166
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 if (count > MAX_SLABINFO_WRITE)
4168 return -EINVAL;
4169 if (copy_from_user(&kbuf, buffer, count))
4170 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004171 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172
4173 tmp = strchr(kbuf, ' ');
4174 if (!tmp)
4175 return -EINVAL;
4176 *tmp = '\0';
4177 tmp++;
4178 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4179 return -EINVAL;
4180
4181 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004182 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004184 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004186 if (limit < 1 || batchcount < 1 ||
4187 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004188 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004190 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004191 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 }
4193 break;
4194 }
4195 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004196 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197 if (res >= 0)
4198 res = count;
4199 return res;
4200}
Al Viro871751e2006-03-25 03:06:39 -08004201
4202#ifdef CONFIG_DEBUG_SLAB_LEAK
4203
4204static void *leaks_start(struct seq_file *m, loff_t *pos)
4205{
4206 loff_t n = *pos;
4207 struct list_head *p;
4208
4209 mutex_lock(&cache_chain_mutex);
4210 p = cache_chain.next;
4211 while (n--) {
4212 p = p->next;
4213 if (p == &cache_chain)
4214 return NULL;
4215 }
4216 return list_entry(p, struct kmem_cache, next);
4217}
4218
4219static inline int add_caller(unsigned long *n, unsigned long v)
4220{
4221 unsigned long *p;
4222 int l;
4223 if (!v)
4224 return 1;
4225 l = n[1];
4226 p = n + 2;
4227 while (l) {
4228 int i = l/2;
4229 unsigned long *q = p + 2 * i;
4230 if (*q == v) {
4231 q[1]++;
4232 return 1;
4233 }
4234 if (*q > v) {
4235 l = i;
4236 } else {
4237 p = q + 2;
4238 l -= i + 1;
4239 }
4240 }
4241 if (++n[1] == n[0])
4242 return 0;
4243 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4244 p[0] = v;
4245 p[1] = 1;
4246 return 1;
4247}
4248
4249static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4250{
4251 void *p;
4252 int i;
4253 if (n[0] == n[1])
4254 return;
4255 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4256 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4257 continue;
4258 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4259 return;
4260 }
4261}
4262
4263static void show_symbol(struct seq_file *m, unsigned long address)
4264{
4265#ifdef CONFIG_KALLSYMS
4266 char *modname;
4267 const char *name;
4268 unsigned long offset, size;
4269 char namebuf[KSYM_NAME_LEN+1];
4270
4271 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4272
4273 if (name) {
4274 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4275 if (modname)
4276 seq_printf(m, " [%s]", modname);
4277 return;
4278 }
4279#endif
4280 seq_printf(m, "%p", (void *)address);
4281}
4282
4283static int leaks_show(struct seq_file *m, void *p)
4284{
4285 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004286 struct slab *slabp;
4287 struct kmem_list3 *l3;
4288 const char *name;
4289 unsigned long *n = m->private;
4290 int node;
4291 int i;
4292
4293 if (!(cachep->flags & SLAB_STORE_USER))
4294 return 0;
4295 if (!(cachep->flags & SLAB_RED_ZONE))
4296 return 0;
4297
4298 /* OK, we can do it */
4299
4300 n[1] = 0;
4301
4302 for_each_online_node(node) {
4303 l3 = cachep->nodelists[node];
4304 if (!l3)
4305 continue;
4306
4307 check_irq_on();
4308 spin_lock_irq(&l3->list_lock);
4309
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004310 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004311 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004312 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004313 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004314 spin_unlock_irq(&l3->list_lock);
4315 }
4316 name = cachep->name;
4317 if (n[0] == n[1]) {
4318 /* Increase the buffer size */
4319 mutex_unlock(&cache_chain_mutex);
4320 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4321 if (!m->private) {
4322 /* Too bad, we are really out */
4323 m->private = n;
4324 mutex_lock(&cache_chain_mutex);
4325 return -ENOMEM;
4326 }
4327 *(unsigned long *)m->private = n[0] * 2;
4328 kfree(n);
4329 mutex_lock(&cache_chain_mutex);
4330 /* Now make sure this entry will be retried */
4331 m->count = m->size;
4332 return 0;
4333 }
4334 for (i = 0; i < n[1]; i++) {
4335 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4336 show_symbol(m, n[2*i+2]);
4337 seq_putc(m, '\n');
4338 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004339
Al Viro871751e2006-03-25 03:06:39 -08004340 return 0;
4341}
4342
Helge Deller15ad7cd2006-12-06 20:40:36 -08004343const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004344 .start = leaks_start,
4345 .next = s_next,
4346 .stop = s_stop,
4347 .show = leaks_show,
4348};
4349#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350#endif
4351
Manfred Spraul00e145b2005-09-03 15:55:07 -07004352/**
4353 * ksize - get the actual amount of memory allocated for a given object
4354 * @objp: Pointer to the object
4355 *
4356 * kmalloc may internally round up allocations and return more memory
4357 * than requested. ksize() can be used to determine the actual amount of
4358 * memory allocated. The caller may use this additional memory, even though
4359 * a smaller amount of memory was initially specified with the kmalloc call.
4360 * The caller must guarantee that objp points to a valid object previously
4361 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4362 * must not be freed during the duration of the call.
4363 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004364unsigned int ksize(const void *objp)
4365{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004366 if (unlikely(objp == NULL))
4367 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004368
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004369 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004370}