blob: 2e71a328aa09540a75f83ab289a3da8931920914 [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>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800110#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700111#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800112#include <linux/reciprocal_div.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#include <asm/cacheflush.h>
115#include <asm/tlbflush.h>
116#include <asm/page.h>
117
118/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
140
141#ifndef cache_line_size
142#define cache_line_size() L1_CACHE_BYTES
143#endif
144
145#ifndef ARCH_KMALLOC_MINALIGN
146/*
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700155#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#endif
157
158#ifndef ARCH_SLAB_MINALIGN
159/*
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
165 */
166#define ARCH_SLAB_MINALIGN 0
167#endif
168
169#ifndef ARCH_KMALLOC_FLAGS
170#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171#endif
172
173/* Legal flag mask for kmem_cache_create(). */
174#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700175# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800177 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700178 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800182# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700183 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#endif
187
188/*
189 * kmem_bufctl_t:
190 *
191 * Bufctl's are used for linking objs within a slab
192 * linked offsets.
193 *
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 */
206
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700207typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800210#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213/*
214 * struct slab
215 *
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 */
220struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227};
228
229/*
230 * struct slab_rcu
231 *
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
239 *
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
242 *
243 * We assume struct slab_rcu can overlay struct slab when destroying.
244 */
245struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800247 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
251/*
252 * struct array_cache
253 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * Purpose:
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
258 *
259 * The limit is stored in the per-cpu structure to reduce the data cache
260 * footprint.
261 *
262 */
263struct array_cache {
264 unsigned int avail;
265 unsigned int limit;
266 unsigned int batchcount;
267 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700268 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800269 void *entry[0]; /*
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
272 * the entries.
273 * [0] is for gcc 2.95. It should really be [].
274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277/*
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 */
281#define BOOT_CPUCACHE_ENTRIES 1
282struct arraycache_init {
283 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800284 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
287/*
Christoph Lametere498be72005-09-09 13:03:32 -0700288 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
290struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800296 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Christoph Lametere498be72005-09-09 13:03:32 -0700304/*
305 * Need this for bootstrapping a per node allocator.
306 */
307#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
310#define SIZE_AC 1
311#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Christoph Lametered11d9e2006-06-30 01:55:45 -0700313static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700317static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000318static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700319
Christoph Lametere498be72005-09-09 13:03:32 -0700320/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700323 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700324static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700325{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800326 extern void __bad_size(void);
327
Christoph Lametere498be72005-09-09 13:03:32 -0700328 if (__builtin_constant_p(size)) {
329 int i = 0;
330
331#define CACHE(x) \
332 if (size <=x) \
333 return i; \
334 else \
335 i++;
336#include "linux/kmalloc_sizes.h"
337#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800338 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700339 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800340 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700341 return 0;
342}
343
Ingo Molnare0a42722006-06-23 02:03:46 -0700344static int slab_early_init = 1;
345
Christoph Lametere498be72005-09-09 13:03:32 -0700346#define INDEX_AC index_of(sizeof(struct arraycache_init))
347#define INDEX_L3 index_of(sizeof(struct kmem_list3))
348
Pekka Enberg5295a742006-02-01 03:05:48 -0800349static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700350{
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800356 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
360}
361
Andrew Mortona737b3e2006-03-22 00:08:11 -0800362#define MAKE_LIST(cachep, listp, slab, nodeid) \
363 do { \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700366 } while (0)
367
Andrew Mortona737b3e2006-03-22 00:08:11 -0800368#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
375/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800376 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 *
378 * manages a cache.
379 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800380
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800381struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800383 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800384/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800385 unsigned int batchcount;
386 unsigned int limit;
387 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800388
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800389 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800390 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392
Andrew Mortona737b3e2006-03-22 00:08:11 -0800393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800396/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800398 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800401 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
Andrew Mortona737b3e2006-03-22 00:08:11 -0800403 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800404 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800405 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800407 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800410 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800412/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800413 const char *name;
414 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800416/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800418 unsigned long num_active;
419 unsigned long num_allocations;
420 unsigned long high_mark;
421 unsigned long grown;
422 unsigned long reaped;
423 unsigned long errors;
424 unsigned long max_freeable;
425 unsigned long node_allocs;
426 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700427 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800428 atomic_t allochit;
429 atomic_t allocmiss;
430 atomic_t freehit;
431 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432#endif
433#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800434 /*
435 * If debugging is enabled, then the allocator can add additional
436 * fields and/or padding to every object. buffer_size contains the total
437 * object size including these internal fields, the following two
438 * variables contain the offset to the user object and its size.
439 */
440 int obj_offset;
441 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700443 /*
444 * We put nodelists[] at the end of kmem_cache, because we want to size
445 * this array to nr_node_ids slots instead of MAX_NUMNODES
446 * (see kmem_cache_init())
447 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
448 * is statically defined, so we reserve the max number of nodes.
449 */
450 struct kmem_list3 *nodelists[MAX_NUMNODES];
451 /*
452 * Do not add fields after nodelists[]
453 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454};
455
456#define CFLGS_OFF_SLAB (0x80000000UL)
457#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
458
459#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800460/*
461 * Optimization question: fewer reaps means less probability for unnessary
462 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100464 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 * which could lock up otherwise freeable slabs.
466 */
467#define REAPTIMEOUT_CPUC (2*HZ)
468#define REAPTIMEOUT_LIST3 (4*HZ)
469
470#if STATS
471#define STATS_INC_ACTIVE(x) ((x)->num_active++)
472#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
473#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
474#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700475#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800476#define STATS_SET_HIGH(x) \
477 do { \
478 if ((x)->num_active > (x)->high_mark) \
479 (x)->high_mark = (x)->num_active; \
480 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define STATS_INC_ERR(x) ((x)->errors++)
482#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700483#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700484#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800485#define STATS_SET_FREEABLE(x, i) \
486 do { \
487 if ((x)->max_freeable < i) \
488 (x)->max_freeable = i; \
489 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
491#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
492#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
493#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
494#else
495#define STATS_INC_ACTIVE(x) do { } while (0)
496#define STATS_DEC_ACTIVE(x) do { } while (0)
497#define STATS_INC_ALLOCED(x) do { } while (0)
498#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700499#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500#define STATS_SET_HIGH(x) do { } while (0)
501#define STATS_INC_ERR(x) do { } while (0)
502#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700503#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700504#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800505#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506#define STATS_INC_ALLOCHIT(x) do { } while (0)
507#define STATS_INC_ALLOCMISS(x) do { } while (0)
508#define STATS_INC_FREEHIT(x) do { } while (0)
509#define STATS_INC_FREEMISS(x) do { } while (0)
510#endif
511
512#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Andrew Mortona737b3e2006-03-22 00:08:11 -0800514/*
515 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800517 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 * the end of an object is aligned with the end of the real
519 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800522 * cachep->obj_offset: The real object.
523 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800524 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
525 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800527static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800529 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
531
Pekka Enberg343e0d72006-02-01 03:05:50 -0800532static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800534 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
David Woodhouseb46b8f12007-05-08 00:22:59 -0700537static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
539 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700540 return (unsigned long long*) (objp + obj_offset(cachep) -
541 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
David Woodhouseb46b8f12007-05-08 00:22:59 -0700544static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545{
546 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
547 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700548 return (unsigned long long *)(objp + cachep->buffer_size -
549 sizeof(unsigned long long) -
550 BYTES_PER_WORD);
551 return (unsigned long long *) (objp + cachep->buffer_size -
552 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553}
554
Pekka Enberg343e0d72006-02-01 03:05:50 -0800555static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556{
557 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800558 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
561#else
562
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800563#define obj_offset(x) 0
564#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700565#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
566#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
568
569#endif
570
571/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 * Do not go above this order unless 0 objects fit into the slab.
573 */
574#define BREAK_GFP_ORDER_HI 1
575#define BREAK_GFP_ORDER_LO 0
576static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
577
Andrew Mortona737b3e2006-03-22 00:08:11 -0800578/*
579 * Functions for storing/retrieving the cachep and or slab from the page
580 * allocator. These are used to find the slab an obj belongs to. With kfree(),
581 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800583static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
584{
585 page->lru.next = (struct list_head *)cache;
586}
587
588static inline struct kmem_cache *page_get_cache(struct page *page)
589{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700590 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700591 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800592 return (struct kmem_cache *)page->lru.next;
593}
594
595static inline void page_set_slab(struct page *page, struct slab *slab)
596{
597 page->lru.prev = (struct list_head *)slab;
598}
599
600static inline struct slab *page_get_slab(struct page *page)
601{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700602 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800603 return (struct slab *)page->lru.prev;
604}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800606static inline struct kmem_cache *virt_to_cache(const void *obj)
607{
Christoph Lameterb49af682007-05-06 14:49:41 -0700608 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800609 return page_get_cache(page);
610}
611
612static inline struct slab *virt_to_slab(const void *obj)
613{
Christoph Lameterb49af682007-05-06 14:49:41 -0700614 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800615 return page_get_slab(page);
616}
617
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800618static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
619 unsigned int idx)
620{
621 return slab->s_mem + cache->buffer_size * idx;
622}
623
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800624/*
625 * We want to avoid an expensive divide : (offset / cache->buffer_size)
626 * Using the fact that buffer_size is a constant for a particular cache,
627 * we can replace (offset / cache->buffer_size) by
628 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
629 */
630static inline unsigned int obj_to_index(const struct kmem_cache *cache,
631 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800632{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800633 u32 offset = (obj - slab->s_mem);
634 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800635}
636
Andrew Mortona737b3e2006-03-22 00:08:11 -0800637/*
638 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640struct cache_sizes malloc_sizes[] = {
641#define CACHE(x) { .cs_size = (x) },
642#include <linux/kmalloc_sizes.h>
643 CACHE(ULONG_MAX)
644#undef CACHE
645};
646EXPORT_SYMBOL(malloc_sizes);
647
648/* Must match cache_sizes above. Out of line to keep cache footprint low. */
649struct cache_names {
650 char *name;
651 char *name_dma;
652};
653
654static struct cache_names __initdata cache_names[] = {
655#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
656#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800657 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658#undef CACHE
659};
660
661static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665
666/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800667static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800668 .batchcount = 1,
669 .limit = BOOT_CPUCACHE_ENTRIES,
670 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800671 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800672 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673};
674
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700675#define BAD_ALIEN_MAGIC 0x01020304ul
676
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200677#ifdef CONFIG_LOCKDEP
678
679/*
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700685 *
686 * We set lock class for alien array caches which are up during init.
687 * The lock annotation will be lost if all cpus of a node goes down and
688 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200689 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700690static struct lock_class_key on_slab_l3_key;
691static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200692
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700693static inline void init_lock_keys(void)
694
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200695{
696 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700697 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200698
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700699 while (s->cs_size != ULONG_MAX) {
700 for_each_node(q) {
701 struct array_cache **alc;
702 int r;
703 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
704 if (!l3 || OFF_SLAB(s->cs_cachep))
705 continue;
706 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
707 alc = l3->alien;
708 /*
709 * FIXME: This check for BAD_ALIEN_MAGIC
710 * should go away when common slab code is taught to
711 * work even without alien caches.
712 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
713 * for alloc_alien_cache,
714 */
715 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
716 continue;
717 for_each_node(r) {
718 if (alc[r])
719 lockdep_set_class(&alc[r]->lock,
720 &on_slab_alc_key);
721 }
722 }
723 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200724 }
725}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200726#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700727static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200728{
729}
730#endif
731
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800732/*
733 * 1. Guard access to the cache-chain.
734 * 2. Protect sanity of cpu_online_map against cpu hotplug events
735 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800736static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737static struct list_head cache_chain;
738
739/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 * chicken and egg problem: delay the per-cpu array allocation
741 * until the general caches are up.
742 */
743static enum {
744 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700745 PARTIAL_AC,
746 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 FULL
748} g_cpucache_up;
749
Mike Kravetz39d24e62006-05-15 09:44:13 -0700750/*
751 * used by boot code to determine if it can use slab based allocator
752 */
753int slab_is_available(void)
754{
755 return g_cpucache_up == FULL;
756}
757
David Howells52bad642006-11-22 14:54:01 +0000758static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
Pekka Enberg343e0d72006-02-01 03:05:50 -0800760static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761{
762 return cachep->array[smp_processor_id()];
763}
764
Andrew Mortona737b3e2006-03-22 00:08:11 -0800765static inline struct kmem_cache *__find_general_cachep(size_t size,
766 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767{
768 struct cache_sizes *csizep = malloc_sizes;
769
770#if DEBUG
771 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800772 * kmem_cache_create(), or __kmalloc(), before
773 * the generic caches are initialized.
774 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700775 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776#endif
Christoph Lameter0b44f7a2007-05-16 22:10:53 -0700777 WARN_ON_ONCE(size == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 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 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800786#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 if (unlikely(gfpflags & GFP_DMA))
788 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800789#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 return csizep->cs_cachep;
791}
792
Adrian Bunkb2213852006-09-25 23:31:02 -0700793static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700794{
795 return __find_general_cachep(size, gfpflags);
796}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700797
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800798static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800800 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
801}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Andrew Mortona737b3e2006-03-22 00:08:11 -0800803/*
804 * Calculate the number of objects and left-over bytes for a given buffer size.
805 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800806static void cache_estimate(unsigned long gfporder, size_t buffer_size,
807 size_t align, int flags, size_t *left_over,
808 unsigned int *num)
809{
810 int nr_objs;
811 size_t mgmt_size;
812 size_t slab_size = PAGE_SIZE << gfporder;
813
814 /*
815 * The slab management structure can be either off the slab or
816 * on it. For the latter case, the memory allocated for a
817 * slab is used for:
818 *
819 * - The struct slab
820 * - One kmem_bufctl_t for each object
821 * - Padding to respect alignment of @align
822 * - @buffer_size bytes for each object
823 *
824 * If the slab management structure is off the slab, then the
825 * alignment will already be calculated into the size. Because
826 * the slabs are all pages aligned, the objects will be at the
827 * correct alignment when allocated.
828 */
829 if (flags & CFLGS_OFF_SLAB) {
830 mgmt_size = 0;
831 nr_objs = slab_size / buffer_size;
832
833 if (nr_objs > SLAB_LIMIT)
834 nr_objs = SLAB_LIMIT;
835 } else {
836 /*
837 * Ignore padding for the initial guess. The padding
838 * is at most @align-1 bytes, and @buffer_size is at
839 * least @align. In the worst case, this result will
840 * be one greater than the number of objects that fit
841 * into the memory allocation when taking the padding
842 * into account.
843 */
844 nr_objs = (slab_size - sizeof(struct slab)) /
845 (buffer_size + sizeof(kmem_bufctl_t));
846
847 /*
848 * This calculated number will be either the right
849 * amount, or one greater than what we want.
850 */
851 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
852 > slab_size)
853 nr_objs--;
854
855 if (nr_objs > SLAB_LIMIT)
856 nr_objs = SLAB_LIMIT;
857
858 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800860 *num = nr_objs;
861 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862}
863
864#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
865
Andrew Mortona737b3e2006-03-22 00:08:11 -0800866static void __slab_error(const char *function, struct kmem_cache *cachep,
867 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868{
869 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800870 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 dump_stack();
872}
873
Paul Menage3395ee02006-12-06 20:32:16 -0800874/*
875 * By default on NUMA we use alien caches to stage the freeing of
876 * objects allocated from other nodes. This causes massive memory
877 * inefficiencies when using fake NUMA setup to split memory into a
878 * large number of small nodes, so it can be disabled on the command
879 * line
880 */
881
882static int use_alien_caches __read_mostly = 1;
883static int __init noaliencache_setup(char *s)
884{
885 use_alien_caches = 0;
886 return 1;
887}
888__setup("noaliencache", noaliencache_setup);
889
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800890#ifdef CONFIG_NUMA
891/*
892 * Special reaping functions for NUMA systems called from cache_reap().
893 * These take care of doing round robin flushing of alien caches (containing
894 * objects freed on different nodes from which they were allocated) and the
895 * flushing of remote pcps by calling drain_node_pages.
896 */
897static DEFINE_PER_CPU(unsigned long, reap_node);
898
899static void init_reap_node(int cpu)
900{
901 int node;
902
903 node = next_node(cpu_to_node(cpu), node_online_map);
904 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800905 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800906
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800907 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800908}
909
910static void next_reap_node(void)
911{
912 int node = __get_cpu_var(reap_node);
913
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800914 node = next_node(node, node_online_map);
915 if (unlikely(node >= MAX_NUMNODES))
916 node = first_node(node_online_map);
917 __get_cpu_var(reap_node) = node;
918}
919
920#else
921#define init_reap_node(cpu) do { } while (0)
922#define next_reap_node(void) do { } while (0)
923#endif
924
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925/*
926 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
927 * via the workqueue/eventd.
928 * Add the CPU number into the expiration time to minimize the possibility of
929 * the CPUs getting into lockstep and contending for the global cache chain
930 * lock.
931 */
932static void __devinit start_cpu_timer(int cpu)
933{
David Howells52bad642006-11-22 14:54:01 +0000934 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
936 /*
937 * When this gets called from do_initcalls via cpucache_init(),
938 * init_workqueues() has already run, so keventd will be setup
939 * at that time.
940 */
David Howells52bad642006-11-22 14:54:01 +0000941 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800942 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000943 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800944 schedule_delayed_work_on(cpu, reap_work,
945 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 }
947}
948
Christoph Lametere498be72005-09-09 13:03:32 -0700949static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800950 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800952 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 struct array_cache *nc = NULL;
954
Christoph Lametere498be72005-09-09 13:03:32 -0700955 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 if (nc) {
957 nc->avail = 0;
958 nc->limit = entries;
959 nc->batchcount = batchcount;
960 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700961 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 }
963 return nc;
964}
965
Christoph Lameter3ded1752006-03-25 03:06:44 -0800966/*
967 * Transfer objects in one arraycache to another.
968 * Locking must be handled by the caller.
969 *
970 * Return the number of entries transferred.
971 */
972static int transfer_objects(struct array_cache *to,
973 struct array_cache *from, unsigned int max)
974{
975 /* Figure out how many entries to transfer */
976 int nr = min(min(from->avail, max), to->limit - to->avail);
977
978 if (!nr)
979 return 0;
980
981 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
982 sizeof(void *) *nr);
983
984 from->avail -= nr;
985 to->avail += nr;
986 to->touched = 1;
987 return nr;
988}
989
Christoph Lameter765c4502006-09-27 01:50:08 -0700990#ifndef CONFIG_NUMA
991
992#define drain_alien_cache(cachep, alien) do { } while (0)
993#define reap_alien(cachep, l3) do { } while (0)
994
995static inline struct array_cache **alloc_alien_cache(int node, int limit)
996{
997 return (struct array_cache **)BAD_ALIEN_MAGIC;
998}
999
1000static inline void free_alien_cache(struct array_cache **ac_ptr)
1001{
1002}
1003
1004static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1005{
1006 return 0;
1007}
1008
1009static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1010 gfp_t flags)
1011{
1012 return NULL;
1013}
1014
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001015static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001016 gfp_t flags, int nodeid)
1017{
1018 return NULL;
1019}
1020
1021#else /* CONFIG_NUMA */
1022
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001023static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001024static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001025
Pekka Enberg5295a742006-02-01 03:05:48 -08001026static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001027{
1028 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001029 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001030 int i;
1031
1032 if (limit > 1)
1033 limit = 12;
1034 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1035 if (ac_ptr) {
1036 for_each_node(i) {
1037 if (i == node || !node_online(i)) {
1038 ac_ptr[i] = NULL;
1039 continue;
1040 }
1041 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1042 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001043 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001044 kfree(ac_ptr[i]);
1045 kfree(ac_ptr);
1046 return NULL;
1047 }
1048 }
1049 }
1050 return ac_ptr;
1051}
1052
Pekka Enberg5295a742006-02-01 03:05:48 -08001053static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001054{
1055 int i;
1056
1057 if (!ac_ptr)
1058 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001059 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001060 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001061 kfree(ac_ptr);
1062}
1063
Pekka Enberg343e0d72006-02-01 03:05:50 -08001064static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001065 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001066{
1067 struct kmem_list3 *rl3 = cachep->nodelists[node];
1068
1069 if (ac->avail) {
1070 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001071 /*
1072 * Stuff objects into the remote nodes shared array first.
1073 * That way we could avoid the overhead of putting the objects
1074 * into the free lists and getting them back later.
1075 */
shin, jacob693f7d32006-04-28 10:54:37 -05001076 if (rl3->shared)
1077 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001078
Christoph Lameterff694162005-09-22 21:44:02 -07001079 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001080 ac->avail = 0;
1081 spin_unlock(&rl3->list_lock);
1082 }
1083}
1084
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001085/*
1086 * Called from cache_reap() to regularly drain alien caches round robin.
1087 */
1088static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1089{
1090 int node = __get_cpu_var(reap_node);
1091
1092 if (l3->alien) {
1093 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001094
1095 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001096 __drain_alien_cache(cachep, ac, node);
1097 spin_unlock_irq(&ac->lock);
1098 }
1099 }
1100}
1101
Andrew Mortona737b3e2006-03-22 00:08:11 -08001102static void drain_alien_cache(struct kmem_cache *cachep,
1103 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001104{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001105 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001106 struct array_cache *ac;
1107 unsigned long flags;
1108
1109 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001110 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001111 if (ac) {
1112 spin_lock_irqsave(&ac->lock, flags);
1113 __drain_alien_cache(cachep, ac, i);
1114 spin_unlock_irqrestore(&ac->lock, flags);
1115 }
1116 }
1117}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001118
Ingo Molnar873623d2006-07-13 14:44:38 +02001119static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001120{
1121 struct slab *slabp = virt_to_slab(objp);
1122 int nodeid = slabp->nodeid;
1123 struct kmem_list3 *l3;
1124 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001125 int node;
1126
1127 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001128
1129 /*
1130 * Make sure we are not freeing a object from another node to the array
1131 * cache on this cpu.
1132 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001133 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001134 return 0;
1135
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001136 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001137 STATS_INC_NODEFREES(cachep);
1138 if (l3->alien && l3->alien[nodeid]) {
1139 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001140 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001141 if (unlikely(alien->avail == alien->limit)) {
1142 STATS_INC_ACOVERFLOW(cachep);
1143 __drain_alien_cache(cachep, alien, nodeid);
1144 }
1145 alien->entry[alien->avail++] = objp;
1146 spin_unlock(&alien->lock);
1147 } else {
1148 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1149 free_block(cachep, &objp, 1, nodeid);
1150 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1151 }
1152 return 1;
1153}
Christoph Lametere498be72005-09-09 13:03:32 -07001154#endif
1155
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001156static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001157 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158{
1159 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001160 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001161 struct kmem_list3 *l3 = NULL;
1162 int node = cpu_to_node(cpu);
1163 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
1165 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001166 case CPU_LOCK_ACQUIRE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001167 mutex_lock(&cache_chain_mutex);
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001168 break;
1169 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001170 case CPU_UP_PREPARE_FROZEN:
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;
Eric Dumazet63109842007-05-06 14:49:28 -07001213 struct array_cache *shared = NULL;
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;
Eric Dumazet63109842007-05-06 14:49:28 -07001220 if (cachep->shared) {
1221 shared = alloc_arraycache(node,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001222 cachep->shared * cachep->batchcount,
1223 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07001224 if (!shared)
1225 goto bad;
1226 }
Paul Menage3395ee02006-12-06 20:32:16 -08001227 if (use_alien_caches) {
1228 alien = alloc_alien_cache(node, cachep->limit);
1229 if (!alien)
1230 goto bad;
1231 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001233 l3 = cachep->nodelists[node];
1234 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001235
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001236 spin_lock_irq(&l3->list_lock);
1237 if (!l3->shared) {
1238 /*
1239 * We are serialised from CPU_DEAD or
1240 * CPU_UP_CANCELLED by the cpucontrol lock
1241 */
1242 l3->shared = shared;
1243 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001244 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001245#ifdef CONFIG_NUMA
1246 if (!l3->alien) {
1247 l3->alien = alien;
1248 alien = NULL;
1249 }
1250#endif
1251 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001252 kfree(shared);
1253 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 break;
1256 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001257 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 start_cpu_timer(cpu);
1259 break;
1260#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001261 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001262 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001263 /*
1264 * Shutdown cache reaper. Note that the cache_chain_mutex is
1265 * held so that if cache_reap() is invoked it cannot do
1266 * anything expensive but will only modify reap_work
1267 * and reschedule the timer.
1268 */
1269 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1270 /* Now the cache_reaper is guaranteed to be not running. */
1271 per_cpu(reap_work, cpu).work.func = NULL;
1272 break;
1273 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001274 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001275 start_cpu_timer(cpu);
1276 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001278 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001279 /*
1280 * Even if all the cpus of a node are down, we don't free the
1281 * kmem_list3 of any cache. This to avoid a race between
1282 * cpu_down, and a kmalloc allocation from another cpu for
1283 * memory from the node of the cpu going down. The list3
1284 * structure is usually allocated from kmem_cache_create() and
1285 * gets destroyed at kmem_cache_destroy().
1286 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 /* fall thru */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001288#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001290 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 list_for_each_entry(cachep, &cache_chain, next) {
1292 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001293 struct array_cache *shared;
1294 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001295 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296
Christoph Lametere498be72005-09-09 13:03:32 -07001297 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 /* cpu is dead; no one can alloc from it. */
1299 nc = cachep->array[cpu];
1300 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001301 l3 = cachep->nodelists[node];
1302
1303 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001304 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001305
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001306 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001307
1308 /* Free limit for this kmem_list3 */
1309 l3->free_limit -= cachep->batchcount;
1310 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001311 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001312
1313 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001314 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001315 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001316 }
Christoph Lametere498be72005-09-09 13:03:32 -07001317
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001318 shared = l3->shared;
1319 if (shared) {
Eric Dumazet63109842007-05-06 14:49:28 -07001320 free_block(cachep, shared->entry,
1321 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001322 l3->shared = NULL;
1323 }
Christoph Lametere498be72005-09-09 13:03:32 -07001324
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001325 alien = l3->alien;
1326 l3->alien = NULL;
1327
1328 spin_unlock_irq(&l3->list_lock);
1329
1330 kfree(shared);
1331 if (alien) {
1332 drain_alien_cache(cachep, alien);
1333 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001334 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001335free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 kfree(nc);
1337 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001338 /*
1339 * In the previous loop, all the objects were freed to
1340 * the respective cache's slabs, now we can go ahead and
1341 * shrink each nodelist to its limit.
1342 */
1343 list_for_each_entry(cachep, &cache_chain, next) {
1344 l3 = cachep->nodelists[node];
1345 if (!l3)
1346 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001347 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001348 }
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001349 break;
1350 case CPU_LOCK_RELEASE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001351 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 }
1354 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001355bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 return NOTIFY_BAD;
1357}
1358
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001359static struct notifier_block __cpuinitdata cpucache_notifier = {
1360 &cpuup_callback, NULL, 0
1361};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
Christoph Lametere498be72005-09-09 13:03:32 -07001363/*
1364 * swap the static kmem_list3 with kmalloced memory
1365 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001366static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1367 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001368{
1369 struct kmem_list3 *ptr;
1370
Christoph Lametere498be72005-09-09 13:03:32 -07001371 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1372 BUG_ON(!ptr);
1373
1374 local_irq_disable();
1375 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001376 /*
1377 * Do not assume that spinlocks can be initialized via memcpy:
1378 */
1379 spin_lock_init(&ptr->list_lock);
1380
Christoph Lametere498be72005-09-09 13:03:32 -07001381 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1382 cachep->nodelists[nodeid] = ptr;
1383 local_irq_enable();
1384}
1385
Andrew Mortona737b3e2006-03-22 00:08:11 -08001386/*
1387 * Initialisation. Called after the page allocator have been initialised and
1388 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 */
1390void __init kmem_cache_init(void)
1391{
1392 size_t left_over;
1393 struct cache_sizes *sizes;
1394 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001395 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001396 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001397 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001398
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001399 if (num_possible_nodes() == 1)
1400 use_alien_caches = 0;
1401
Christoph Lametere498be72005-09-09 13:03:32 -07001402 for (i = 0; i < NUM_INIT_LISTS; i++) {
1403 kmem_list3_init(&initkmem_list3[i]);
1404 if (i < MAX_NUMNODES)
1405 cache_cache.nodelists[i] = NULL;
1406 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
1408 /*
1409 * Fragmentation resistance on low memory - only use bigger
1410 * page orders on machines with more than 32MB of memory.
1411 */
1412 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1413 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1414
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 /* Bootstrap is tricky, because several objects are allocated
1416 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001417 * 1) initialize the cache_cache cache: it contains the struct
1418 * kmem_cache structures of all caches, except cache_cache itself:
1419 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001420 * Initially an __init data area is used for the head array and the
1421 * kmem_list3 structures, it's replaced with a kmalloc allocated
1422 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001424 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001425 * An __init data area is used for the head array.
1426 * 3) Create the remaining kmalloc caches, with minimally sized
1427 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 * 4) Replace the __init data head arrays for cache_cache and the first
1429 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001430 * 5) Replace the __init data for kmem_list3 for cache_cache and
1431 * the other cache's with kmalloc allocated memory.
1432 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 */
1434
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001435 node = numa_node_id();
1436
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 INIT_LIST_HEAD(&cache_chain);
1439 list_add(&cache_cache.next, &cache_chain);
1440 cache_cache.colour_off = cache_line_size();
1441 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001442 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443
Eric Dumazet8da34302007-05-06 14:49:29 -07001444 /*
1445 * struct kmem_cache size depends on nr_node_ids, which
1446 * can be less than MAX_NUMNODES.
1447 */
1448 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1449 nr_node_ids * sizeof(struct kmem_list3 *);
1450#if DEBUG
1451 cache_cache.obj_size = cache_cache.buffer_size;
1452#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001453 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1454 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001455 cache_cache.reciprocal_buffer_size =
1456 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457
Jack Steiner07ed76b2006-03-07 21:55:46 -08001458 for (order = 0; order < MAX_ORDER; order++) {
1459 cache_estimate(order, cache_cache.buffer_size,
1460 cache_line_size(), 0, &left_over, &cache_cache.num);
1461 if (cache_cache.num)
1462 break;
1463 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001464 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001465 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001466 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001467 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1468 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469
1470 /* 2+3) create the kmalloc caches */
1471 sizes = malloc_sizes;
1472 names = cache_names;
1473
Andrew Mortona737b3e2006-03-22 00:08:11 -08001474 /*
1475 * Initialize the caches that provide memory for the array cache and the
1476 * kmem_list3 structures first. Without this, further allocations will
1477 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001478 */
1479
1480 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001481 sizes[INDEX_AC].cs_size,
1482 ARCH_KMALLOC_MINALIGN,
1483 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1484 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001485
Andrew Mortona737b3e2006-03-22 00:08:11 -08001486 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001487 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001488 kmem_cache_create(names[INDEX_L3].name,
1489 sizes[INDEX_L3].cs_size,
1490 ARCH_KMALLOC_MINALIGN,
1491 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1492 NULL, NULL);
1493 }
Christoph Lametere498be72005-09-09 13:03:32 -07001494
Ingo Molnare0a42722006-06-23 02:03:46 -07001495 slab_early_init = 0;
1496
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001498 /*
1499 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 * This should be particularly beneficial on SMP boxes, as it
1501 * eliminates "false sharing".
1502 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001503 * allow tighter packing of the smaller caches.
1504 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001505 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001506 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001507 sizes->cs_size,
1508 ARCH_KMALLOC_MINALIGN,
1509 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1510 NULL, NULL);
1511 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001512#ifdef CONFIG_ZONE_DMA
1513 sizes->cs_dmacachep = kmem_cache_create(
1514 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001515 sizes->cs_size,
1516 ARCH_KMALLOC_MINALIGN,
1517 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1518 SLAB_PANIC,
1519 NULL, NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001520#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 sizes++;
1522 names++;
1523 }
1524 /* 4) Replace the bootstrap head arrays */
1525 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001526 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001527
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001529
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001531 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1532 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001533 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001534 /*
1535 * Do not assume that spinlocks can be initialized via memcpy:
1536 */
1537 spin_lock_init(&ptr->lock);
1538
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 cache_cache.array[smp_processor_id()] = ptr;
1540 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001541
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001543
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001545 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001546 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001547 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001548 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001549 /*
1550 * Do not assume that spinlocks can be initialized via memcpy:
1551 */
1552 spin_lock_init(&ptr->lock);
1553
Christoph Lametere498be72005-09-09 13:03:32 -07001554 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001555 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 local_irq_enable();
1557 }
Christoph Lametere498be72005-09-09 13:03:32 -07001558 /* 5) Replace the bootstrap kmem_list3's */
1559 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001560 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001562 /* Replace the static kmem_list3 structures for the boot cpu */
1563 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1564
1565 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001566 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001567 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001568
1569 if (INDEX_AC != INDEX_L3) {
1570 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001571 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001572 }
1573 }
1574 }
1575
1576 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001578 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001579 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001581 if (enable_cpucache(cachep))
1582 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001583 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 }
1585
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001586 /* Annotate slab for lockdep -- annotate the malloc caches */
1587 init_lock_keys();
1588
1589
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 /* Done! */
1591 g_cpucache_up = FULL;
1592
Andrew Mortona737b3e2006-03-22 00:08:11 -08001593 /*
1594 * Register a cpu startup notifier callback that initializes
1595 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 */
1597 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598
Andrew Mortona737b3e2006-03-22 00:08:11 -08001599 /*
1600 * The reap timers are started later, with a module init call: That part
1601 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 */
1603}
1604
1605static int __init cpucache_init(void)
1606{
1607 int cpu;
1608
Andrew Mortona737b3e2006-03-22 00:08:11 -08001609 /*
1610 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 */
Christoph Lametere498be72005-09-09 13:03:32 -07001612 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001613 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 return 0;
1615}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616__initcall(cpucache_init);
1617
1618/*
1619 * Interface to system's page allocator. No need to hold the cache-lock.
1620 *
1621 * If we requested dmaable memory, we will get it. Even if we
1622 * did not request dmaable memory, we might get it, but that
1623 * would be relatively rare and ignorable.
1624 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001625static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626{
1627 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001628 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 int i;
1630
Luke Yangd6fef9d2006-04-10 22:52:56 -07001631#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001632 /*
1633 * Nommu uses slab's for process anonymous memory allocations, and thus
1634 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001635 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001636 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001637#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001638
Christoph Lameter3c517a62006-12-06 20:33:29 -08001639 flags |= cachep->gfpflags;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001640
1641 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 if (!page)
1643 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001645 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001647 add_zone_page_state(page_zone(page),
1648 NR_SLAB_RECLAIMABLE, nr_pages);
1649 else
1650 add_zone_page_state(page_zone(page),
1651 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001652 for (i = 0; i < nr_pages; i++)
1653 __SetPageSlab(page + i);
1654 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655}
1656
1657/*
1658 * Interface to system's page release.
1659 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001660static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001662 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 struct page *page = virt_to_page(addr);
1664 const unsigned long nr_freed = i;
1665
Christoph Lameter972d1a72006-09-25 23:31:51 -07001666 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1667 sub_zone_page_state(page_zone(page),
1668 NR_SLAB_RECLAIMABLE, nr_freed);
1669 else
1670 sub_zone_page_state(page_zone(page),
1671 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001673 BUG_ON(!PageSlab(page));
1674 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 page++;
1676 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 if (current->reclaim_state)
1678 current->reclaim_state->reclaimed_slab += nr_freed;
1679 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680}
1681
1682static void kmem_rcu_free(struct rcu_head *head)
1683{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001684 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001685 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
1687 kmem_freepages(cachep, slab_rcu->addr);
1688 if (OFF_SLAB(cachep))
1689 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1690}
1691
1692#if DEBUG
1693
1694#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001695static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001696 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001698 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001700 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001702 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 return;
1704
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001705 *addr++ = 0x12345678;
1706 *addr++ = caller;
1707 *addr++ = smp_processor_id();
1708 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 {
1710 unsigned long *sptr = &caller;
1711 unsigned long svalue;
1712
1713 while (!kstack_end(sptr)) {
1714 svalue = *sptr++;
1715 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001716 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 size -= sizeof(unsigned long);
1718 if (size <= sizeof(unsigned long))
1719 break;
1720 }
1721 }
1722
1723 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001724 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725}
1726#endif
1727
Pekka Enberg343e0d72006-02-01 03:05:50 -08001728static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001730 int size = obj_size(cachep);
1731 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
1733 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001734 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735}
1736
1737static void dump_line(char *data, int offset, int limit)
1738{
1739 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001740 unsigned char error = 0;
1741 int bad_count = 0;
1742
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001744 for (i = 0; i < limit; i++) {
1745 if (data[offset + i] != POISON_FREE) {
1746 error = data[offset + i];
1747 bad_count++;
1748 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001749 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001752
1753 if (bad_count == 1) {
1754 error ^= POISON_FREE;
1755 if (!(error & (error - 1))) {
1756 printk(KERN_ERR "Single bit error detected. Probably "
1757 "bad RAM.\n");
1758#ifdef CONFIG_X86
1759 printk(KERN_ERR "Run memtest86+ or a similar memory "
1760 "test tool.\n");
1761#else
1762 printk(KERN_ERR "Run a memory test tool.\n");
1763#endif
1764 }
1765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766}
1767#endif
1768
1769#if DEBUG
1770
Pekka Enberg343e0d72006-02-01 03:05:50 -08001771static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772{
1773 int i, size;
1774 char *realobj;
1775
1776 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001777 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001778 *dbg_redzone1(cachep, objp),
1779 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 }
1781
1782 if (cachep->flags & SLAB_STORE_USER) {
1783 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001784 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001786 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 printk("\n");
1788 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001789 realobj = (char *)objp + obj_offset(cachep);
1790 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 int limit;
1793 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001794 if (i + limit > size)
1795 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 dump_line(realobj, i, limit);
1797 }
1798}
1799
Pekka Enberg343e0d72006-02-01 03:05:50 -08001800static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801{
1802 char *realobj;
1803 int size, i;
1804 int lines = 0;
1805
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001806 realobj = (char *)objp + obj_offset(cachep);
1807 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001809 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001811 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 exp = POISON_END;
1813 if (realobj[i] != exp) {
1814 int limit;
1815 /* Mismatch ! */
1816 /* Print header */
1817 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001818 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001819 "Slab corruption: %s start=%p, len=%d\n",
1820 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 print_objinfo(cachep, objp, 0);
1822 }
1823 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001824 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001826 if (i + limit > size)
1827 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 dump_line(realobj, i, limit);
1829 i += 16;
1830 lines++;
1831 /* Limit to 5 lines */
1832 if (lines > 5)
1833 break;
1834 }
1835 }
1836 if (lines != 0) {
1837 /* Print some data about the neighboring objects, if they
1838 * exist:
1839 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001840 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001841 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001843 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001845 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001846 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001848 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 print_objinfo(cachep, objp, 2);
1850 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001852 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001853 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001855 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 print_objinfo(cachep, objp, 2);
1857 }
1858 }
1859}
1860#endif
1861
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001863/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001864 * slab_destroy_objs - destroy a slab and its objects
1865 * @cachep: cache pointer being destroyed
1866 * @slabp: slab pointer being destroyed
1867 *
1868 * Call the registered destructor for each object in a slab that is being
1869 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001870 */
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 int i;
1874 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001875 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876
1877 if (cachep->flags & SLAB_POISON) {
1878#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001879 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1880 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001881 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001882 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 else
1884 check_poison_obj(cachep, objp);
1885#else
1886 check_poison_obj(cachep, objp);
1887#endif
1888 }
1889 if (cachep->flags & SLAB_RED_ZONE) {
1890 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1891 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001892 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1894 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001898}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001900static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001901{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001902}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903#endif
1904
Randy Dunlap911851e2006-03-22 00:08:14 -08001905/**
1906 * slab_destroy - destroy and release all objects in a slab
1907 * @cachep: cache pointer being destroyed
1908 * @slabp: slab pointer being destroyed
1909 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001910 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001911 * Before calling the slab must have been unlinked from the cache. The
1912 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001913 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001914static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001915{
1916 void *addr = slabp->s_mem - slabp->colouroff;
1917
1918 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1920 struct slab_rcu *slab_rcu;
1921
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 slab_rcu->cachep = cachep;
1924 slab_rcu->addr = addr;
1925 call_rcu(&slab_rcu->head, kmem_rcu_free);
1926 } else {
1927 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001928 if (OFF_SLAB(cachep))
1929 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 }
1931}
1932
Andrew Mortona737b3e2006-03-22 00:08:11 -08001933/*
1934 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1935 * size of kmem_list3.
1936 */
Andrew Mortona3a02be2007-05-06 14:49:31 -07001937static void __init set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001938{
1939 int node;
1940
1941 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001942 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001943 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001944 REAPTIMEOUT_LIST3 +
1945 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001946 }
1947}
1948
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001949static void __kmem_cache_destroy(struct kmem_cache *cachep)
1950{
1951 int i;
1952 struct kmem_list3 *l3;
1953
1954 for_each_online_cpu(i)
1955 kfree(cachep->array[i]);
1956
1957 /* NUMA: free the list3 structures */
1958 for_each_online_node(i) {
1959 l3 = cachep->nodelists[i];
1960 if (l3) {
1961 kfree(l3->shared);
1962 free_alien_cache(l3->alien);
1963 kfree(l3);
1964 }
1965 }
1966 kmem_cache_free(&cache_cache, cachep);
1967}
1968
1969
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001971 * calculate_slab_order - calculate size (page order) of slabs
1972 * @cachep: pointer to the cache that is being created
1973 * @size: size of objects to be created in this cache.
1974 * @align: required alignment for the objects.
1975 * @flags: slab allocation flags
1976 *
1977 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001978 *
1979 * This could be made much more intelligent. For now, try to avoid using
1980 * high order pages for slabs. When the gfp() functions are more friendly
1981 * towards high-order requests, this should be changed.
1982 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001983static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001984 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001985{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001986 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001987 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001988 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001989
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001990 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001991 unsigned int num;
1992 size_t remainder;
1993
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001994 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001995 if (!num)
1996 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001997
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001998 if (flags & CFLGS_OFF_SLAB) {
1999 /*
2000 * Max number of objs-per-slab for caches which
2001 * use off-slab slabs. Needed to avoid a possible
2002 * looping condition in cache_grow().
2003 */
2004 offslab_limit = size - sizeof(struct slab);
2005 offslab_limit /= sizeof(kmem_bufctl_t);
2006
2007 if (num > offslab_limit)
2008 break;
2009 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002011 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002013 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002014 left_over = remainder;
2015
2016 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002017 * A VFS-reclaimable slab tends to have most allocations
2018 * as GFP_NOFS and we really don't want to have to be allocating
2019 * higher-order pages when we are unable to shrink dcache.
2020 */
2021 if (flags & SLAB_RECLAIM_ACCOUNT)
2022 break;
2023
2024 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002025 * Large number of objects is good, but very large slabs are
2026 * currently bad for the gfp()s.
2027 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002028 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002029 break;
2030
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002031 /*
2032 * Acceptable internal fragmentation?
2033 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002034 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002035 break;
2036 }
2037 return left_over;
2038}
2039
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002040static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002041{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002042 if (g_cpucache_up == FULL)
2043 return enable_cpucache(cachep);
2044
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002045 if (g_cpucache_up == NONE) {
2046 /*
2047 * Note: the first kmem_cache_create must create the cache
2048 * that's used by kmalloc(24), otherwise the creation of
2049 * further caches will BUG().
2050 */
2051 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2052
2053 /*
2054 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2055 * the first cache, then we need to set up all its list3s,
2056 * otherwise the creation of further caches will BUG().
2057 */
2058 set_up_list3s(cachep, SIZE_AC);
2059 if (INDEX_AC == INDEX_L3)
2060 g_cpucache_up = PARTIAL_L3;
2061 else
2062 g_cpucache_up = PARTIAL_AC;
2063 } else {
2064 cachep->array[smp_processor_id()] =
2065 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2066
2067 if (g_cpucache_up == PARTIAL_AC) {
2068 set_up_list3s(cachep, SIZE_L3);
2069 g_cpucache_up = PARTIAL_L3;
2070 } else {
2071 int node;
2072 for_each_online_node(node) {
2073 cachep->nodelists[node] =
2074 kmalloc_node(sizeof(struct kmem_list3),
2075 GFP_KERNEL, node);
2076 BUG_ON(!cachep->nodelists[node]);
2077 kmem_list3_init(cachep->nodelists[node]);
2078 }
2079 }
2080 }
2081 cachep->nodelists[numa_node_id()]->next_reap =
2082 jiffies + REAPTIMEOUT_LIST3 +
2083 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2084
2085 cpu_cache_get(cachep)->avail = 0;
2086 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2087 cpu_cache_get(cachep)->batchcount = 1;
2088 cpu_cache_get(cachep)->touched = 0;
2089 cachep->batchcount = 1;
2090 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002091 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002092}
2093
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002094/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 * kmem_cache_create - Create a cache.
2096 * @name: A string which is used in /proc/slabinfo to identify this cache.
2097 * @size: The size of objects to be created in this cache.
2098 * @align: The required alignment for the objects.
2099 * @flags: SLAB flags
2100 * @ctor: A constructor for the objects.
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002101 * @dtor: A destructor for the objects (not implemented anymore).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 *
2103 * Returns a ptr to the cache on success, NULL on failure.
2104 * Cannot be called within a int, but can be interrupted.
2105 * The @ctor is run when new pages are allocated by the cache
2106 * and the @dtor is run before the pages are handed back.
2107 *
2108 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002109 * the module calling this has to destroy the cache before getting unloaded.
2110 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 * The flags are
2112 *
2113 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2114 * to catch references to uninitialised memory.
2115 *
2116 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2117 * for buffer overruns.
2118 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2120 * cacheline. This can be beneficial if you're counting cycles as closely
2121 * as davem.
2122 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002123struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002125 unsigned long flags,
2126 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002127 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128{
2129 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002130 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
2132 /*
2133 * Sanity checks... these are all serious usage bugs.
2134 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002135 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002136 size > KMALLOC_MAX_SIZE || dtor) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002137 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2138 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002139 BUG();
2140 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002142 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002143 * We use cache_chain_mutex to ensure a consistent view of
2144 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002145 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002146 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002147
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002148 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002149 char tmp;
2150 int res;
2151
2152 /*
2153 * This happens when the module gets unloaded and doesn't
2154 * destroy its slab cache and no-one else reuses the vmalloc
2155 * area of the module. Print a warning.
2156 */
Andrew Morton138ae662006-12-06 20:36:41 -08002157 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002158 if (res) {
matzeb4169522007-05-06 14:49:52 -07002159 printk(KERN_ERR
2160 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002161 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002162 continue;
2163 }
2164
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002165 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002166 printk(KERN_ERR
2167 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002168 dump_stack();
2169 goto oops;
2170 }
2171 }
2172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173#if DEBUG
2174 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175#if FORCED_DEBUG
2176 /*
2177 * Enable redzoning and last user accounting, except for caches with
2178 * large objects, if the increased size would increase the object size
2179 * above the next power of two: caches with object sizes just above a
2180 * power of two have a significant amount of internal fragmentation.
2181 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002182 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002183 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 if (!(flags & SLAB_DESTROY_BY_RCU))
2185 flags |= SLAB_POISON;
2186#endif
2187 if (flags & SLAB_DESTROY_BY_RCU)
2188 BUG_ON(flags & SLAB_POISON);
2189#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002191 * Always checks flags, a caller might be expecting debug support which
2192 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002194 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195
Andrew Mortona737b3e2006-03-22 00:08:11 -08002196 /*
2197 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 * unaligned accesses for some archs when redzoning is used, and makes
2199 * sure any on-slab bufctl's are also correctly aligned.
2200 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 if (size & (BYTES_PER_WORD - 1)) {
2202 size += (BYTES_PER_WORD - 1);
2203 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 }
2205
Andrew Mortona737b3e2006-03-22 00:08:11 -08002206 /* calculate the final buffer alignment: */
2207
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 /* 1) arch recommendation: can be overridden for debug */
2209 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002210 /*
2211 * Default alignment: as specified by the arch code. Except if
2212 * an object is really small, then squeeze multiple objects into
2213 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 */
2215 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002216 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 ralign /= 2;
2218 } else {
2219 ralign = BYTES_PER_WORD;
2220 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002221
2222 /*
2223 * Redzoning and user store require word alignment. Note this will be
2224 * overridden by architecture or caller mandated alignment if either
2225 * is greater than BYTES_PER_WORD.
2226 */
2227 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -07002228 ralign = __alignof__(unsigned long long);
Pekka Enbergca5f9702006-09-25 23:31:25 -07002229
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002230 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 if (ralign < ARCH_SLAB_MINALIGN) {
2232 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002234 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 if (ralign < align) {
2236 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002238 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002239 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002240 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002241 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002242 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 */
2244 align = ralign;
2245
2246 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002247 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002249 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250
2251#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002252 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253
Pekka Enbergca5f9702006-09-25 23:31:25 -07002254 /*
2255 * Both debugging options require word-alignment which is calculated
2256 * into align above.
2257 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002260 cachep->obj_offset += sizeof(unsigned long long);
2261 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 }
2263 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002264 /* user store requires one word storage behind the end of
2265 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 size += BYTES_PER_WORD;
2268 }
2269#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002270 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002271 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2272 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 size = PAGE_SIZE;
2274 }
2275#endif
2276#endif
2277
Ingo Molnare0a42722006-06-23 02:03:46 -07002278 /*
2279 * Determine if the slab management is 'on' or 'off' slab.
2280 * (bootstrapping cannot cope with offslab caches so don't do
2281 * it too early on.)
2282 */
2283 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 /*
2285 * Size is large, assume best to place the slab management obj
2286 * off-slab (should allow better packing of objs).
2287 */
2288 flags |= CFLGS_OFF_SLAB;
2289
2290 size = ALIGN(size, align);
2291
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002292 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
2294 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002295 printk(KERN_ERR
2296 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 kmem_cache_free(&cache_cache, cachep);
2298 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002299 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002301 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2302 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303
2304 /*
2305 * If the slab has been placed off-slab, and we have enough space then
2306 * move it on-slab. This is at the expense of any extra colouring.
2307 */
2308 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2309 flags &= ~CFLGS_OFF_SLAB;
2310 left_over -= slab_size;
2311 }
2312
2313 if (flags & CFLGS_OFF_SLAB) {
2314 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002315 slab_size =
2316 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 }
2318
2319 cachep->colour_off = cache_line_size();
2320 /* Offset must be a multiple of the alignment. */
2321 if (cachep->colour_off < align)
2322 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002323 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 cachep->slab_size = slab_size;
2325 cachep->flags = flags;
2326 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002327 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002329 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002330 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002332 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002333 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002334 /*
2335 * This is a possibility for one of the malloc_sizes caches.
2336 * But since we go off slab only for object size greater than
2337 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2338 * this should not happen at all.
2339 * But leave a BUG_ON for some lucky dude.
2340 */
2341 BUG_ON(!cachep->slabp_cache);
2342 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 cachep->name = name;
2345
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002346 if (setup_cpu_cache(cachep)) {
2347 __kmem_cache_destroy(cachep);
2348 cachep = NULL;
2349 goto oops;
2350 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 /* cache setup completed, link it into the list */
2353 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002354oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 if (!cachep && (flags & SLAB_PANIC))
2356 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002357 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002358 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 return cachep;
2360}
2361EXPORT_SYMBOL(kmem_cache_create);
2362
2363#if DEBUG
2364static void check_irq_off(void)
2365{
2366 BUG_ON(!irqs_disabled());
2367}
2368
2369static void check_irq_on(void)
2370{
2371 BUG_ON(irqs_disabled());
2372}
2373
Pekka Enberg343e0d72006-02-01 03:05:50 -08002374static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375{
2376#ifdef CONFIG_SMP
2377 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002378 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379#endif
2380}
Christoph Lametere498be72005-09-09 13:03:32 -07002381
Pekka Enberg343e0d72006-02-01 03:05:50 -08002382static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002383{
2384#ifdef CONFIG_SMP
2385 check_irq_off();
2386 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2387#endif
2388}
2389
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390#else
2391#define check_irq_off() do { } while(0)
2392#define check_irq_on() do { } while(0)
2393#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002394#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395#endif
2396
Christoph Lameteraab22072006-03-22 00:09:06 -08002397static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2398 struct array_cache *ac,
2399 int force, int node);
2400
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401static void do_drain(void *arg)
2402{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002403 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002405 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406
2407 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002408 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002409 spin_lock(&cachep->nodelists[node]->list_lock);
2410 free_block(cachep, ac->entry, ac->avail, node);
2411 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 ac->avail = 0;
2413}
2414
Pekka Enberg343e0d72006-02-01 03:05:50 -08002415static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416{
Christoph Lametere498be72005-09-09 13:03:32 -07002417 struct kmem_list3 *l3;
2418 int node;
2419
Andrew Mortona07fa392006-03-22 00:08:17 -08002420 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002422 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002423 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002424 if (l3 && l3->alien)
2425 drain_alien_cache(cachep, l3->alien);
2426 }
2427
2428 for_each_online_node(node) {
2429 l3 = cachep->nodelists[node];
2430 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002431 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002432 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433}
2434
Christoph Lametered11d9e2006-06-30 01:55:45 -07002435/*
2436 * Remove slabs from the list of free slabs.
2437 * Specify the number of slabs to drain in tofree.
2438 *
2439 * Returns the actual number of slabs released.
2440 */
2441static int drain_freelist(struct kmem_cache *cache,
2442 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002444 struct list_head *p;
2445 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447
Christoph Lametered11d9e2006-06-30 01:55:45 -07002448 nr_freed = 0;
2449 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450
Christoph Lametered11d9e2006-06-30 01:55:45 -07002451 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002452 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002453 if (p == &l3->slabs_free) {
2454 spin_unlock_irq(&l3->list_lock);
2455 goto out;
2456 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Christoph Lametered11d9e2006-06-30 01:55:45 -07002458 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002460 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461#endif
2462 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002463 /*
2464 * Safe to drop the lock. The slab is no longer linked
2465 * to the cache.
2466 */
2467 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002468 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002469 slab_destroy(cache, slabp);
2470 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002472out:
2473 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474}
2475
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002476/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002477static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002478{
2479 int ret = 0, i = 0;
2480 struct kmem_list3 *l3;
2481
2482 drain_cpu_caches(cachep);
2483
2484 check_irq_on();
2485 for_each_online_node(i) {
2486 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002487 if (!l3)
2488 continue;
2489
2490 drain_freelist(cachep, l3, l3->free_objects);
2491
2492 ret += !list_empty(&l3->slabs_full) ||
2493 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002494 }
2495 return (ret ? 1 : 0);
2496}
2497
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498/**
2499 * kmem_cache_shrink - Shrink a cache.
2500 * @cachep: The cache to shrink.
2501 *
2502 * Releases as many slabs as possible for a cache.
2503 * To help debugging, a zero exit status indicates all slabs were released.
2504 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002505int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002507 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002508 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002510 mutex_lock(&cache_chain_mutex);
2511 ret = __cache_shrink(cachep);
2512 mutex_unlock(&cache_chain_mutex);
2513 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514}
2515EXPORT_SYMBOL(kmem_cache_shrink);
2516
2517/**
2518 * kmem_cache_destroy - delete a cache
2519 * @cachep: the cache to destroy
2520 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002521 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 *
2523 * It is expected this function will be called by a module when it is
2524 * unloaded. This will remove the cache completely, and avoid a duplicate
2525 * cache being allocated each time a module is loaded and unloaded, if the
2526 * module doesn't have persistent in-kernel storage across loads and unloads.
2527 *
2528 * The cache must be empty before calling this function.
2529 *
2530 * The caller must guarantee that noone will allocate memory from the cache
2531 * during the kmem_cache_destroy().
2532 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002533void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002535 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002538 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 /*
2540 * the chain is never empty, cache_cache is never destroyed
2541 */
2542 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 if (__cache_shrink(cachep)) {
2544 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002545 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002546 mutex_unlock(&cache_chain_mutex);
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002547 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 }
2549
2550 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002551 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002553 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002554 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555}
2556EXPORT_SYMBOL(kmem_cache_destroy);
2557
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002558/*
2559 * Get the memory for a slab management obj.
2560 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2561 * always come from malloc_sizes caches. The slab descriptor cannot
2562 * come from the same cache which is getting created because,
2563 * when we are searching for an appropriate cache for these
2564 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2565 * If we are creating a malloc_sizes cache here it would not be visible to
2566 * kmem_find_general_cachep till the initialization is complete.
2567 * Hence we cannot have slabp_cache same as the original cache.
2568 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002569static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002570 int colour_off, gfp_t local_flags,
2571 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572{
2573 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002574
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575 if (OFF_SLAB(cachep)) {
2576 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002577 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002578 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 if (!slabp)
2580 return NULL;
2581 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002582 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 colour_off += cachep->slab_size;
2584 }
2585 slabp->inuse = 0;
2586 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002587 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002588 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 return slabp;
2590}
2591
2592static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2593{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002594 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595}
2596
Pekka Enberg343e0d72006-02-01 03:05:50 -08002597static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002598 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599{
2600 int i;
2601
2602 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002603 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604#if DEBUG
2605 /* need to poison the objs? */
2606 if (cachep->flags & SLAB_POISON)
2607 poison_obj(cachep, objp, POISON_FREE);
2608 if (cachep->flags & SLAB_STORE_USER)
2609 *dbg_userword(cachep, objp) = NULL;
2610
2611 if (cachep->flags & SLAB_RED_ZONE) {
2612 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2613 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2614 }
2615 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002616 * Constructors are not allowed to allocate memory from the same
2617 * cache which they are a constructor for. Otherwise, deadlock.
2618 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 */
2620 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002621 cachep->ctor(objp + obj_offset(cachep), cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002622 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623
2624 if (cachep->flags & SLAB_RED_ZONE) {
2625 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2626 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002627 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2629 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002630 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002632 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2633 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002634 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002635 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636#else
2637 if (cachep->ctor)
Christoph Lametera35afb82007-05-16 22:10:57 -07002638 cachep->ctor(objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002642 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 slabp->free = 0;
2644}
2645
Pekka Enberg343e0d72006-02-01 03:05:50 -08002646static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002648 if (CONFIG_ZONE_DMA_FLAG) {
2649 if (flags & GFP_DMA)
2650 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2651 else
2652 BUG_ON(cachep->gfpflags & GFP_DMA);
2653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654}
2655
Andrew Mortona737b3e2006-03-22 00:08:11 -08002656static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2657 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002658{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002659 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002660 kmem_bufctl_t next;
2661
2662 slabp->inuse++;
2663 next = slab_bufctl(slabp)[slabp->free];
2664#if DEBUG
2665 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2666 WARN_ON(slabp->nodeid != nodeid);
2667#endif
2668 slabp->free = next;
2669
2670 return objp;
2671}
2672
Andrew Mortona737b3e2006-03-22 00:08:11 -08002673static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2674 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002675{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002676 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002677
2678#if DEBUG
2679 /* Verify that the slab belongs to the intended node */
2680 WARN_ON(slabp->nodeid != nodeid);
2681
Al Viro871751e2006-03-25 03:06:39 -08002682 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002683 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002684 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002685 BUG();
2686 }
2687#endif
2688 slab_bufctl(slabp)[objnr] = slabp->free;
2689 slabp->free = objnr;
2690 slabp->inuse--;
2691}
2692
Pekka Enberg47768742006-06-23 02:03:07 -07002693/*
2694 * Map pages beginning at addr to the given cache and slab. This is required
2695 * for the slab allocator to be able to lookup the cache and slab of a
2696 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2697 */
2698static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2699 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700{
Pekka Enberg47768742006-06-23 02:03:07 -07002701 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 struct page *page;
2703
Pekka Enberg47768742006-06-23 02:03:07 -07002704 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002705
Pekka Enberg47768742006-06-23 02:03:07 -07002706 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002707 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002708 nr_pages <<= cache->gfporder;
2709
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002711 page_set_cache(page, cache);
2712 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002714 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715}
2716
2717/*
2718 * Grow (by 1) the number of slabs within a cache. This is called by
2719 * kmem_cache_alloc() when there are no active objs left in a cache.
2720 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002721static int cache_grow(struct kmem_cache *cachep,
2722 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002724 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002725 size_t offset;
2726 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002727 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728
Andrew Mortona737b3e2006-03-22 00:08:11 -08002729 /*
2730 * Be lazy and only check for valid flags here, keeping it out of the
2731 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 */
Christoph Lametercfce6602007-05-06 14:50:17 -07002733 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734
Christoph Lametera06d72c2006-12-06 20:33:12 -08002735 local_flags = (flags & GFP_LEVEL_MASK);
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002736 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002738 l3 = cachep->nodelists[nodeid];
2739 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740
2741 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002742 offset = l3->colour_next;
2743 l3->colour_next++;
2744 if (l3->colour_next >= cachep->colour)
2745 l3->colour_next = 0;
2746 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002748 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749
2750 if (local_flags & __GFP_WAIT)
2751 local_irq_enable();
2752
2753 /*
2754 * The test for missing atomic flag is performed here, rather than
2755 * the more obvious place, simply to reduce the critical path length
2756 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2757 * will eventually be caught here (where it matters).
2758 */
2759 kmem_flagcheck(cachep, flags);
2760
Andrew Mortona737b3e2006-03-22 00:08:11 -08002761 /*
2762 * Get mem for the objs. Attempt to allocate a physical page from
2763 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002764 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002765 if (!objp)
2766 objp = kmem_getpages(cachep, flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002767 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 goto failed;
2769
2770 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002771 slabp = alloc_slabmgmt(cachep, objp, offset,
2772 local_flags & ~GFP_THISNODE, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002773 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 goto opps1;
2775
Christoph Lametere498be72005-09-09 13:03:32 -07002776 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002777 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
Christoph Lametera35afb82007-05-16 22:10:57 -07002779 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780
2781 if (local_flags & __GFP_WAIT)
2782 local_irq_disable();
2783 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002784 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785
2786 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002787 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002789 l3->free_objects += cachep->num;
2790 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002792opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002794failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 if (local_flags & __GFP_WAIT)
2796 local_irq_disable();
2797 return 0;
2798}
2799
2800#if DEBUG
2801
2802/*
2803 * Perform extra freeing checks:
2804 * - detect bad pointers.
2805 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 */
2807static void kfree_debugcheck(const void *objp)
2808{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 if (!virt_addr_valid(objp)) {
2810 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002811 (unsigned long)objp);
2812 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814}
2815
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002816static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2817{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002818 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002819
2820 redzone1 = *dbg_redzone1(cache, obj);
2821 redzone2 = *dbg_redzone2(cache, obj);
2822
2823 /*
2824 * Redzone is ok.
2825 */
2826 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2827 return;
2828
2829 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2830 slab_error(cache, "double free detected");
2831 else
2832 slab_error(cache, "memory outside object was overwritten");
2833
David Woodhouseb46b8f12007-05-08 00:22:59 -07002834 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002835 obj, redzone1, redzone2);
2836}
2837
Pekka Enberg343e0d72006-02-01 03:05:50 -08002838static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002839 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840{
2841 struct page *page;
2842 unsigned int objnr;
2843 struct slab *slabp;
2844
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002845 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002847 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848
Pekka Enberg065d41c2005-11-13 16:06:46 -08002849 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850
2851 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002852 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2854 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2855 }
2856 if (cachep->flags & SLAB_STORE_USER)
2857 *dbg_userword(cachep, objp) = caller;
2858
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002859 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860
2861 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002862 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863
Al Viro871751e2006-03-25 03:06:39 -08002864#ifdef CONFIG_DEBUG_SLAB_LEAK
2865 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2866#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 if (cachep->flags & SLAB_POISON) {
2868#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002869 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002871 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002872 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 } else {
2874 poison_obj(cachep, objp, POISON_FREE);
2875 }
2876#else
2877 poison_obj(cachep, objp, POISON_FREE);
2878#endif
2879 }
2880 return objp;
2881}
2882
Pekka Enberg343e0d72006-02-01 03:05:50 -08002883static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884{
2885 kmem_bufctl_t i;
2886 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002887
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 /* Check slab's freelist to see if this obj is there. */
2889 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2890 entries++;
2891 if (entries > cachep->num || i >= cachep->num)
2892 goto bad;
2893 }
2894 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002895bad:
2896 printk(KERN_ERR "slab: Internal list corruption detected in "
2897 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2898 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002899 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002900 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002901 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002902 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002904 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 }
2906 printk("\n");
2907 BUG();
2908 }
2909}
2910#else
2911#define kfree_debugcheck(x) do { } while(0)
2912#define cache_free_debugcheck(x,objp,z) (objp)
2913#define check_slabp(x,y) do { } while(0)
2914#endif
2915
Pekka Enberg343e0d72006-02-01 03:05:50 -08002916static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
2918 int batchcount;
2919 struct kmem_list3 *l3;
2920 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002921 int node;
2922
2923 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
2925 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002926 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002927retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 batchcount = ac->batchcount;
2929 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002930 /*
2931 * If there was little recent activity on this cache, then
2932 * perform only a partial refill. Otherwise we could generate
2933 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 */
2935 batchcount = BATCHREFILL_LIMIT;
2936 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002937 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938
Christoph Lametere498be72005-09-09 13:03:32 -07002939 BUG_ON(ac->avail > 0 || !l3);
2940 spin_lock(&l3->list_lock);
2941
Christoph Lameter3ded1752006-03-25 03:06:44 -08002942 /* See if we can refill from the shared array */
2943 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2944 goto alloc_done;
2945
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 while (batchcount > 0) {
2947 struct list_head *entry;
2948 struct slab *slabp;
2949 /* Get slab alloc is to come from. */
2950 entry = l3->slabs_partial.next;
2951 if (entry == &l3->slabs_partial) {
2952 l3->free_touched = 1;
2953 entry = l3->slabs_free.next;
2954 if (entry == &l3->slabs_free)
2955 goto must_grow;
2956 }
2957
2958 slabp = list_entry(entry, struct slab, list);
2959 check_slabp(cachep, slabp);
2960 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07002961
2962 /*
2963 * The slab was either on partial or free list so
2964 * there must be at least one object available for
2965 * allocation.
2966 */
2967 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2968
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 STATS_INC_ALLOCED(cachep);
2971 STATS_INC_ACTIVE(cachep);
2972 STATS_SET_HIGH(cachep);
2973
Matthew Dobson78d382d2006-02-01 03:05:47 -08002974 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002975 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 }
2977 check_slabp(cachep, slabp);
2978
2979 /* move slabp to correct slabp list: */
2980 list_del(&slabp->list);
2981 if (slabp->free == BUFCTL_END)
2982 list_add(&slabp->list, &l3->slabs_full);
2983 else
2984 list_add(&slabp->list, &l3->slabs_partial);
2985 }
2986
Andrew Mortona737b3e2006-03-22 00:08:11 -08002987must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002989alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002990 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991
2992 if (unlikely(!ac->avail)) {
2993 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08002994 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002995
Andrew Mortona737b3e2006-03-22 00:08:11 -08002996 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002997 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002998 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 return NULL;
3000
Andrew Mortona737b3e2006-03-22 00:08:11 -08003001 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 goto retry;
3003 }
3004 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003005 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006}
3007
Andrew Mortona737b3e2006-03-22 00:08:11 -08003008static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3009 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010{
3011 might_sleep_if(flags & __GFP_WAIT);
3012#if DEBUG
3013 kmem_flagcheck(cachep, flags);
3014#endif
3015}
3016
3017#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003018static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3019 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003021 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003023 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003025 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003026 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003027 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 else
3029 check_poison_obj(cachep, objp);
3030#else
3031 check_poison_obj(cachep, objp);
3032#endif
3033 poison_obj(cachep, objp, POISON_INUSE);
3034 }
3035 if (cachep->flags & SLAB_STORE_USER)
3036 *dbg_userword(cachep, objp) = caller;
3037
3038 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003039 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3040 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3041 slab_error(cachep, "double free, or memory outside"
3042 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003043 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003044 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003045 objp, *dbg_redzone1(cachep, objp),
3046 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 }
3048 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3049 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3050 }
Al Viro871751e2006-03-25 03:06:39 -08003051#ifdef CONFIG_DEBUG_SLAB_LEAK
3052 {
3053 struct slab *slabp;
3054 unsigned objnr;
3055
Christoph Lameterb49af682007-05-06 14:49:41 -07003056 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003057 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3058 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3059 }
3060#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003061 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003062 if (cachep->ctor && cachep->flags & SLAB_POISON)
Christoph Lametera35afb82007-05-16 22:10:57 -07003063 cachep->ctor(objp, cachep, 0);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003064#if ARCH_SLAB_MINALIGN
3065 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3066 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3067 objp, ARCH_SLAB_MINALIGN);
3068 }
3069#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 return objp;
3071}
3072#else
3073#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3074#endif
3075
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003076#ifdef CONFIG_FAILSLAB
3077
3078static struct failslab_attr {
3079
3080 struct fault_attr attr;
3081
3082 u32 ignore_gfp_wait;
3083#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3084 struct dentry *ignore_gfp_wait_file;
3085#endif
3086
3087} failslab = {
3088 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003089 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003090};
3091
3092static int __init setup_failslab(char *str)
3093{
3094 return setup_fault_attr(&failslab.attr, str);
3095}
3096__setup("failslab=", setup_failslab);
3097
3098static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3099{
3100 if (cachep == &cache_cache)
3101 return 0;
3102 if (flags & __GFP_NOFAIL)
3103 return 0;
3104 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3105 return 0;
3106
3107 return should_fail(&failslab.attr, obj_size(cachep));
3108}
3109
3110#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3111
3112static int __init failslab_debugfs(void)
3113{
3114 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3115 struct dentry *dir;
3116 int err;
3117
Akinobu Mita824ebef2007-05-06 14:49:58 -07003118 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003119 if (err)
3120 return err;
3121 dir = failslab.attr.dentries.dir;
3122
3123 failslab.ignore_gfp_wait_file =
3124 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3125 &failslab.ignore_gfp_wait);
3126
3127 if (!failslab.ignore_gfp_wait_file) {
3128 err = -ENOMEM;
3129 debugfs_remove(failslab.ignore_gfp_wait_file);
3130 cleanup_fault_attr_dentries(&failslab.attr);
3131 }
3132
3133 return err;
3134}
3135
3136late_initcall(failslab_debugfs);
3137
3138#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3139
3140#else /* CONFIG_FAILSLAB */
3141
3142static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3143{
3144 return 0;
3145}
3146
3147#endif /* CONFIG_FAILSLAB */
3148
Pekka Enberg343e0d72006-02-01 03:05:50 -08003149static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003151 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 struct array_cache *ac;
3153
Alok N Kataria5c382302005-09-27 21:45:46 -07003154 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003155
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003156 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 if (likely(ac->avail)) {
3158 STATS_INC_ALLOCHIT(cachep);
3159 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003160 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 } else {
3162 STATS_INC_ALLOCMISS(cachep);
3163 objp = cache_alloc_refill(cachep, flags);
3164 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003165 return objp;
3166}
3167
Christoph Lametere498be72005-09-09 13:03:32 -07003168#ifdef CONFIG_NUMA
3169/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003170 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003171 *
3172 * If we are in_interrupt, then process context, including cpusets and
3173 * mempolicy, may not apply and should not be used for allocation policy.
3174 */
3175static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3176{
3177 int nid_alloc, nid_here;
3178
Christoph Lameter765c4502006-09-27 01:50:08 -07003179 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003180 return NULL;
3181 nid_alloc = nid_here = numa_node_id();
3182 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3183 nid_alloc = cpuset_mem_spread_node();
3184 else if (current->mempolicy)
3185 nid_alloc = slab_node(current->mempolicy);
3186 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003187 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003188 return NULL;
3189}
3190
3191/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003192 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003193 * certain node and fall back is permitted. First we scan all the
3194 * available nodelists for available objects. If that fails then we
3195 * perform an allocation without specifying a node. This allows the page
3196 * allocator to do its reclaim / fallback magic. We then insert the
3197 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003198 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003199static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003200{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003201 struct zonelist *zonelist;
3202 gfp_t local_flags;
Christoph Lameter765c4502006-09-27 01:50:08 -07003203 struct zone **z;
3204 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003205 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003206
3207 if (flags & __GFP_THISNODE)
3208 return NULL;
3209
3210 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3211 ->node_zonelists[gfp_zone(flags)];
3212 local_flags = (flags & GFP_LEVEL_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003213
Christoph Lameter3c517a62006-12-06 20:33:29 -08003214retry:
3215 /*
3216 * Look through allowed nodes for objects available
3217 * from existing per node queues.
3218 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003219 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003220 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003221
Paul Jackson02a0e532006-12-13 00:34:25 -08003222 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003223 cache->nodelists[nid] &&
3224 cache->nodelists[nid]->free_objects)
3225 obj = ____cache_alloc_node(cache,
3226 flags | GFP_THISNODE, nid);
3227 }
3228
Christoph Lametercfce6602007-05-06 14:50:17 -07003229 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003230 /*
3231 * This allocation will be performed within the constraints
3232 * of the current cpuset / memory policy requirements.
3233 * We may trigger various forms of reclaim on the allowed
3234 * set and go into memory reserves if necessary.
3235 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003236 if (local_flags & __GFP_WAIT)
3237 local_irq_enable();
3238 kmem_flagcheck(cache, flags);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003239 obj = kmem_getpages(cache, flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003240 if (local_flags & __GFP_WAIT)
3241 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003242 if (obj) {
3243 /*
3244 * Insert into the appropriate per node queues
3245 */
3246 nid = page_to_nid(virt_to_page(obj));
3247 if (cache_grow(cache, flags, nid, obj)) {
3248 obj = ____cache_alloc_node(cache,
3249 flags | GFP_THISNODE, nid);
3250 if (!obj)
3251 /*
3252 * Another processor may allocate the
3253 * objects in the slab since we are
3254 * not holding any locks.
3255 */
3256 goto retry;
3257 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003258 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003259 obj = NULL;
3260 }
3261 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003262 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003263 return obj;
3264}
3265
3266/*
Christoph Lametere498be72005-09-09 13:03:32 -07003267 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003269static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003270 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003271{
3272 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003273 struct slab *slabp;
3274 struct kmem_list3 *l3;
3275 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003276 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003278 l3 = cachep->nodelists[nodeid];
3279 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003280
Andrew Mortona737b3e2006-03-22 00:08:11 -08003281retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003282 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003283 spin_lock(&l3->list_lock);
3284 entry = l3->slabs_partial.next;
3285 if (entry == &l3->slabs_partial) {
3286 l3->free_touched = 1;
3287 entry = l3->slabs_free.next;
3288 if (entry == &l3->slabs_free)
3289 goto must_grow;
3290 }
Christoph Lametere498be72005-09-09 13:03:32 -07003291
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003292 slabp = list_entry(entry, struct slab, list);
3293 check_spinlock_acquired_node(cachep, nodeid);
3294 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003295
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003296 STATS_INC_NODEALLOCS(cachep);
3297 STATS_INC_ACTIVE(cachep);
3298 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003299
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003300 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003301
Matthew Dobson78d382d2006-02-01 03:05:47 -08003302 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003303 check_slabp(cachep, slabp);
3304 l3->free_objects--;
3305 /* move slabp to correct slabp list: */
3306 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003307
Andrew Mortona737b3e2006-03-22 00:08:11 -08003308 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003309 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003310 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003311 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003312
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003313 spin_unlock(&l3->list_lock);
3314 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003315
Andrew Mortona737b3e2006-03-22 00:08:11 -08003316must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003317 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003318 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003319 if (x)
3320 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003321
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003322 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003323
Andrew Mortona737b3e2006-03-22 00:08:11 -08003324done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003325 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003326}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003327
3328/**
3329 * kmem_cache_alloc_node - Allocate an object on the specified node
3330 * @cachep: The cache to allocate from.
3331 * @flags: See kmalloc().
3332 * @nodeid: node number of the target node.
3333 * @caller: return address of caller, used for debug information
3334 *
3335 * Identical to kmem_cache_alloc but it will allocate memory on the given
3336 * node, which can improve the performance for cpu bound structures.
3337 *
3338 * Fallback to other node is possible if __GFP_THISNODE is not set.
3339 */
3340static __always_inline void *
3341__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3342 void *caller)
3343{
3344 unsigned long save_flags;
3345 void *ptr;
3346
Akinobu Mita824ebef2007-05-06 14:49:58 -07003347 if (should_failslab(cachep, flags))
3348 return NULL;
3349
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003350 cache_alloc_debugcheck_before(cachep, flags);
3351 local_irq_save(save_flags);
3352
3353 if (unlikely(nodeid == -1))
3354 nodeid = numa_node_id();
3355
3356 if (unlikely(!cachep->nodelists[nodeid])) {
3357 /* Node not bootstrapped yet */
3358 ptr = fallback_alloc(cachep, flags);
3359 goto out;
3360 }
3361
3362 if (nodeid == numa_node_id()) {
3363 /*
3364 * Use the locally cached objects if possible.
3365 * However ____cache_alloc does not allow fallback
3366 * to other nodes. It may fail while we still have
3367 * objects on other nodes available.
3368 */
3369 ptr = ____cache_alloc(cachep, flags);
3370 if (ptr)
3371 goto out;
3372 }
3373 /* ___cache_alloc_node can fall back to other nodes */
3374 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3375 out:
3376 local_irq_restore(save_flags);
3377 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3378
3379 return ptr;
3380}
3381
3382static __always_inline void *
3383__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3384{
3385 void *objp;
3386
3387 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3388 objp = alternate_node_alloc(cache, flags);
3389 if (objp)
3390 goto out;
3391 }
3392 objp = ____cache_alloc(cache, flags);
3393
3394 /*
3395 * We may just have run out of memory on the local node.
3396 * ____cache_alloc_node() knows how to locate memory on other nodes
3397 */
3398 if (!objp)
3399 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3400
3401 out:
3402 return objp;
3403}
3404#else
3405
3406static __always_inline void *
3407__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3408{
3409 return ____cache_alloc(cachep, flags);
3410}
3411
3412#endif /* CONFIG_NUMA */
3413
3414static __always_inline void *
3415__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3416{
3417 unsigned long save_flags;
3418 void *objp;
3419
Akinobu Mita824ebef2007-05-06 14:49:58 -07003420 if (should_failslab(cachep, flags))
3421 return NULL;
3422
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003423 cache_alloc_debugcheck_before(cachep, flags);
3424 local_irq_save(save_flags);
3425 objp = __do_cache_alloc(cachep, flags);
3426 local_irq_restore(save_flags);
3427 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3428 prefetchw(objp);
3429
3430 return objp;
3431}
Christoph Lametere498be72005-09-09 13:03:32 -07003432
3433/*
3434 * Caller needs to acquire correct kmem_list's list_lock
3435 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003436static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003437 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438{
3439 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003440 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441
3442 for (i = 0; i < nr_objects; i++) {
3443 void *objp = objpp[i];
3444 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003446 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003447 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003449 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003451 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003453 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454 check_slabp(cachep, slabp);
3455
3456 /* fixup slab chains */
3457 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003458 if (l3->free_objects > l3->free_limit) {
3459 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003460 /* No need to drop any previously held
3461 * lock here, even if we have a off-slab slab
3462 * descriptor it is guaranteed to come from
3463 * a different cache, refer to comments before
3464 * alloc_slabmgmt.
3465 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 slab_destroy(cachep, slabp);
3467 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003468 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 }
3470 } else {
3471 /* Unconditionally move a slab to the end of the
3472 * partial list on free - maximum time for the
3473 * other objects to be freed, too.
3474 */
Christoph Lametere498be72005-09-09 13:03:32 -07003475 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 }
3477 }
3478}
3479
Pekka Enberg343e0d72006-02-01 03:05:50 -08003480static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481{
3482 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003483 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003484 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485
3486 batchcount = ac->batchcount;
3487#if DEBUG
3488 BUG_ON(!batchcount || batchcount > ac->avail);
3489#endif
3490 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003491 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003492 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003493 if (l3->shared) {
3494 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003495 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 if (max) {
3497 if (batchcount > max)
3498 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003499 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003500 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 shared_array->avail += batchcount;
3502 goto free_done;
3503 }
3504 }
3505
Christoph Lameterff694162005-09-22 21:44:02 -07003506 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003507free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508#if STATS
3509 {
3510 int i = 0;
3511 struct list_head *p;
3512
Christoph Lametere498be72005-09-09 13:03:32 -07003513 p = l3->slabs_free.next;
3514 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 struct slab *slabp;
3516
3517 slabp = list_entry(p, struct slab, list);
3518 BUG_ON(slabp->inuse);
3519
3520 i++;
3521 p = p->next;
3522 }
3523 STATS_SET_FREEABLE(cachep, i);
3524 }
3525#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003526 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003528 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529}
3530
3531/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003532 * Release an obj back to its cache. If the obj has a constructed state, it must
3533 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003535static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003537 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538
3539 check_irq_off();
3540 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3541
Siddha, Suresh B62918a02007-05-02 19:27:18 +02003542 if (use_alien_caches && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003543 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003544
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 if (likely(ac->avail < ac->limit)) {
3546 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003547 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 return;
3549 } else {
3550 STATS_INC_FREEMISS(cachep);
3551 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003552 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 }
3554}
3555
3556/**
3557 * kmem_cache_alloc - Allocate an object
3558 * @cachep: The cache to allocate from.
3559 * @flags: See kmalloc().
3560 *
3561 * Allocate an object from this cache. The flags are only relevant
3562 * if the cache has no available objects.
3563 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003564void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003566 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567}
3568EXPORT_SYMBOL(kmem_cache_alloc);
3569
3570/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003571 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003572 * @cache: The cache to allocate from.
3573 * @flags: See kmalloc().
3574 *
3575 * Allocate an object from this cache and set the allocated memory to zero.
3576 * The flags are only relevant if the cache has no available objects.
3577 */
3578void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3579{
3580 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3581 if (ret)
3582 memset(ret, 0, obj_size(cache));
3583 return ret;
3584}
3585EXPORT_SYMBOL(kmem_cache_zalloc);
3586
3587/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 * kmem_ptr_validate - check if an untrusted pointer might
3589 * be a slab entry.
3590 * @cachep: the cache we're checking against
3591 * @ptr: pointer to validate
3592 *
3593 * This verifies that the untrusted pointer looks sane:
3594 * it is _not_ a guarantee that the pointer is actually
3595 * part of the slab cache in question, but it at least
3596 * validates that the pointer can be dereferenced and
3597 * looks half-way sane.
3598 *
3599 * Currently only used for dentry validation.
3600 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003601int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003603 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003605 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003606 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 struct page *page;
3608
3609 if (unlikely(addr < min_addr))
3610 goto out;
3611 if (unlikely(addr > (unsigned long)high_memory - size))
3612 goto out;
3613 if (unlikely(addr & align_mask))
3614 goto out;
3615 if (unlikely(!kern_addr_valid(addr)))
3616 goto out;
3617 if (unlikely(!kern_addr_valid(addr + size - 1)))
3618 goto out;
3619 page = virt_to_page(ptr);
3620 if (unlikely(!PageSlab(page)))
3621 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003622 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 goto out;
3624 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003625out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 return 0;
3627}
3628
3629#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003630void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3631{
3632 return __cache_alloc_node(cachep, flags, nodeid,
3633 __builtin_return_address(0));
3634}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635EXPORT_SYMBOL(kmem_cache_alloc_node);
3636
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003637static __always_inline void *
3638__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003639{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003640 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003641
3642 cachep = kmem_find_general_cachep(size, flags);
3643 if (unlikely(cachep == NULL))
3644 return NULL;
3645 return kmem_cache_alloc_node(cachep, flags, node);
3646}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003647
3648#ifdef CONFIG_DEBUG_SLAB
3649void *__kmalloc_node(size_t size, gfp_t flags, int node)
3650{
3651 return __do_kmalloc_node(size, flags, node,
3652 __builtin_return_address(0));
3653}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003654EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003655
3656void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3657 int node, void *caller)
3658{
3659 return __do_kmalloc_node(size, flags, node, caller);
3660}
3661EXPORT_SYMBOL(__kmalloc_node_track_caller);
3662#else
3663void *__kmalloc_node(size_t size, gfp_t flags, int node)
3664{
3665 return __do_kmalloc_node(size, flags, node, NULL);
3666}
3667EXPORT_SYMBOL(__kmalloc_node);
3668#endif /* CONFIG_DEBUG_SLAB */
3669#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670
3671/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003672 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003674 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003675 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003677static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3678 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003680 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003682 /* If you want to save a few bytes .text space: replace
3683 * __ with kmem_.
3684 * Then kmalloc uses the uninlined functions instead of the inline
3685 * functions.
3686 */
3687 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003688 if (unlikely(cachep == NULL))
3689 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003690 return __cache_alloc(cachep, flags, caller);
3691}
3692
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003693
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003694#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003695void *__kmalloc(size_t size, gfp_t flags)
3696{
Al Viro871751e2006-03-25 03:06:39 -08003697 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698}
3699EXPORT_SYMBOL(__kmalloc);
3700
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003701void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3702{
3703 return __do_kmalloc(size, flags, caller);
3704}
3705EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003706
3707#else
3708void *__kmalloc(size_t size, gfp_t flags)
3709{
3710 return __do_kmalloc(size, flags, NULL);
3711}
3712EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003713#endif
3714
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715/**
Pekka Enbergfd76bab2007-05-06 14:48:40 -07003716 * krealloc - reallocate memory. The contents will remain unchanged.
Pekka Enbergfd76bab2007-05-06 14:48:40 -07003717 * @p: object to reallocate memory for.
3718 * @new_size: how many bytes of memory are required.
3719 * @flags: the type of memory to allocate.
3720 *
3721 * The contents of the object pointed to are preserved up to the
3722 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3723 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3724 * %NULL pointer, the object pointed to is freed.
3725 */
3726void *krealloc(const void *p, size_t new_size, gfp_t flags)
3727{
3728 struct kmem_cache *cache, *new_cache;
3729 void *ret;
3730
3731 if (unlikely(!p))
3732 return kmalloc_track_caller(new_size, flags);
3733
3734 if (unlikely(!new_size)) {
3735 kfree(p);
3736 return NULL;
3737 }
3738
3739 cache = virt_to_cache(p);
3740 new_cache = __find_general_cachep(new_size, flags);
3741
3742 /*
3743 * If new size fits in the current cache, bail out.
3744 */
3745 if (likely(cache == new_cache))
3746 return (void *)p;
3747
3748 /*
3749 * We are on the slow-path here so do not use __cache_alloc
3750 * because it bloats kernel text.
3751 */
3752 ret = kmalloc_track_caller(new_size, flags);
3753 if (ret) {
3754 memcpy(ret, p, min(new_size, ksize(p)));
3755 kfree(p);
3756 }
3757 return ret;
3758}
3759EXPORT_SYMBOL(krealloc);
3760
3761/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 * kmem_cache_free - Deallocate an object
3763 * @cachep: The cache the allocation was from.
3764 * @objp: The previously allocated object.
3765 *
3766 * Free an object which was previously allocated from this
3767 * cache.
3768 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003769void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003770{
3771 unsigned long flags;
3772
Pekka Enbergddc2e812006-06-23 02:03:40 -07003773 BUG_ON(virt_to_cache(objp) != cachep);
3774
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003776 debug_check_no_locks_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003777 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 local_irq_restore(flags);
3779}
3780EXPORT_SYMBOL(kmem_cache_free);
3781
3782/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783 * kfree - free previously allocated memory
3784 * @objp: pointer returned by kmalloc.
3785 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003786 * If @objp is NULL, no operation is performed.
3787 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 * Don't free memory not originally allocated by kmalloc()
3789 * or you will run into trouble.
3790 */
3791void kfree(const void *objp)
3792{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003793 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 unsigned long flags;
3795
3796 if (unlikely(!objp))
3797 return;
3798 local_irq_save(flags);
3799 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003800 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003801 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003802 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 local_irq_restore(flags);
3804}
3805EXPORT_SYMBOL(kfree);
3806
Pekka Enberg343e0d72006-02-01 03:05:50 -08003807unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003809 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810}
3811EXPORT_SYMBOL(kmem_cache_size);
3812
Pekka Enberg343e0d72006-02-01 03:05:50 -08003813const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003814{
3815 return cachep->name;
3816}
3817EXPORT_SYMBOL_GPL(kmem_cache_name);
3818
Christoph Lametere498be72005-09-09 13:03:32 -07003819/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003820 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003821 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003822static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003823{
3824 int node;
3825 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003826 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003827 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003828
3829 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003830
Paul Menage3395ee02006-12-06 20:32:16 -08003831 if (use_alien_caches) {
3832 new_alien = alloc_alien_cache(node, cachep->limit);
3833 if (!new_alien)
3834 goto fail;
3835 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003836
Eric Dumazet63109842007-05-06 14:49:28 -07003837 new_shared = NULL;
3838 if (cachep->shared) {
3839 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003840 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003841 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003842 if (!new_shared) {
3843 free_alien_cache(new_alien);
3844 goto fail;
3845 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003846 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003847
Andrew Mortona737b3e2006-03-22 00:08:11 -08003848 l3 = cachep->nodelists[node];
3849 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003850 struct array_cache *shared = l3->shared;
3851
Christoph Lametere498be72005-09-09 13:03:32 -07003852 spin_lock_irq(&l3->list_lock);
3853
Christoph Lametercafeb022006-03-25 03:06:46 -08003854 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003855 free_block(cachep, shared->entry,
3856 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003857
Christoph Lametercafeb022006-03-25 03:06:46 -08003858 l3->shared = new_shared;
3859 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003860 l3->alien = new_alien;
3861 new_alien = NULL;
3862 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003863 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003864 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003865 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003866 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003867 free_alien_cache(new_alien);
3868 continue;
3869 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003870 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003871 if (!l3) {
3872 free_alien_cache(new_alien);
3873 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003874 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003875 }
Christoph Lametere498be72005-09-09 13:03:32 -07003876
3877 kmem_list3_init(l3);
3878 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003879 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003880 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003881 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003882 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003883 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003884 cachep->nodelists[node] = l3;
3885 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003886 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003887
Andrew Mortona737b3e2006-03-22 00:08:11 -08003888fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003889 if (!cachep->next.next) {
3890 /* Cache is not active yet. Roll back what we did */
3891 node--;
3892 while (node >= 0) {
3893 if (cachep->nodelists[node]) {
3894 l3 = cachep->nodelists[node];
3895
3896 kfree(l3->shared);
3897 free_alien_cache(l3->alien);
3898 kfree(l3);
3899 cachep->nodelists[node] = NULL;
3900 }
3901 node--;
3902 }
3903 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003904 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003905}
3906
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003908 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 struct array_cache *new[NR_CPUS];
3910};
3911
3912static void do_ccupdate_local(void *info)
3913{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003914 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915 struct array_cache *old;
3916
3917 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003918 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003919
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3921 new->new[smp_processor_id()] = old;
3922}
3923
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003924/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003925static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3926 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003928 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003929 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003931 new = kzalloc(sizeof(*new), GFP_KERNEL);
3932 if (!new)
3933 return -ENOMEM;
3934
Christoph Lametere498be72005-09-09 13:03:32 -07003935 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003936 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003937 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003938 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003939 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003940 kfree(new->new[i]);
3941 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003942 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 }
3944 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003945 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003947 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003948
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 cachep->batchcount = batchcount;
3951 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003952 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953
Christoph Lametere498be72005-09-09 13:03:32 -07003954 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003955 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 if (!ccold)
3957 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003958 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003959 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003960 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 kfree(ccold);
3962 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003963 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003964 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965}
3966
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003967/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003968static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969{
3970 int err;
3971 int limit, shared;
3972
Andrew Mortona737b3e2006-03-22 00:08:11 -08003973 /*
3974 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 * - create a LIFO ordering, i.e. return objects that are cache-warm
3976 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003977 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 * bufctl chains: array operations are cheaper.
3979 * The numbers are guessed, we should auto-tune as described by
3980 * Bonwick.
3981 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003982 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003984 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003986 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003988 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989 limit = 54;
3990 else
3991 limit = 120;
3992
Andrew Mortona737b3e2006-03-22 00:08:11 -08003993 /*
3994 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 * allocation behaviour: Most allocs on one cpu, most free operations
3996 * on another cpu. For these cases, an efficient object passing between
3997 * cpus is necessary. This is provided by a shared array. The array
3998 * replaces Bonwick's magazine layer.
3999 * On uniprocessor, it's functionally equivalent (but less efficient)
4000 * to a larger limit. Thus disabled by default.
4001 */
4002 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004003 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005
4006#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004007 /*
4008 * With debugging enabled, large batchcount lead to excessively long
4009 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 */
4011 if (limit > 32)
4012 limit = 32;
4013#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004014 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015 if (err)
4016 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004017 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004018 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019}
4020
Christoph Lameter1b552532006-03-22 00:09:07 -08004021/*
4022 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004023 * necessary. Note that the l3 listlock also protects the array_cache
4024 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004025 */
4026void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4027 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028{
4029 int tofree;
4030
Christoph Lameter1b552532006-03-22 00:09:07 -08004031 if (!ac || !ac->avail)
4032 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 if (ac->touched && !force) {
4034 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004035 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004036 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004037 if (ac->avail) {
4038 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4039 if (tofree > ac->avail)
4040 tofree = (ac->avail + 1) / 2;
4041 free_block(cachep, ac->entry, tofree, node);
4042 ac->avail -= tofree;
4043 memmove(ac->entry, &(ac->entry[tofree]),
4044 sizeof(void *) * ac->avail);
4045 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004046 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047 }
4048}
4049
4050/**
4051 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004052 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053 *
4054 * Called from workqueue/eventd every few seconds.
4055 * Purpose:
4056 * - clear the per-cpu caches for this CPU.
4057 * - return freeable pages to the main free memory pool.
4058 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004059 * If we cannot acquire the cache chain mutex then just give up - we'll try
4060 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004062static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004064 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004065 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004066 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004067 struct delayed_work *work =
4068 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004070 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004072 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004074 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 check_irq_on();
4076
Christoph Lameter35386e32006-03-22 00:09:05 -08004077 /*
4078 * We only take the l3 lock if absolutely necessary and we
4079 * have established with reasonable certainty that
4080 * we can do some work if the lock was obtained.
4081 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004082 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004083
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004084 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085
Christoph Lameteraab22072006-03-22 00:09:06 -08004086 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087
Christoph Lameter35386e32006-03-22 00:09:05 -08004088 /*
4089 * These are racy checks but it does not matter
4090 * if we skip one check or scan twice.
4091 */
Christoph Lametere498be72005-09-09 13:03:32 -07004092 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004093 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094
Christoph Lametere498be72005-09-09 13:03:32 -07004095 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096
Christoph Lameteraab22072006-03-22 00:09:06 -08004097 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098
Christoph Lametered11d9e2006-06-30 01:55:45 -07004099 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004100 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004101 else {
4102 int freed;
4103
4104 freed = drain_freelist(searchp, l3, (l3->free_limit +
4105 5 * searchp->num - 1) / (5 * searchp->num));
4106 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004108next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109 cond_resched();
4110 }
4111 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004112 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004113 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004114out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004115 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004116 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004117}
4118
4119#ifdef CONFIG_PROC_FS
4120
Pekka Enberg85289f92006-01-08 01:00:36 -08004121static void print_slabinfo_header(struct seq_file *m)
4122{
4123 /*
4124 * Output format version, so at least we can change it
4125 * without _too_ many complaints.
4126 */
4127#if STATS
4128 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4129#else
4130 seq_puts(m, "slabinfo - version: 2.1\n");
4131#endif
4132 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4133 "<objperslab> <pagesperslab>");
4134 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4135 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4136#if STATS
4137 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004138 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004139 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4140#endif
4141 seq_putc(m, '\n');
4142}
4143
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144static void *s_start(struct seq_file *m, loff_t *pos)
4145{
4146 loff_t n = *pos;
4147 struct list_head *p;
4148
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004149 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004150 if (!n)
4151 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 p = cache_chain.next;
4153 while (n--) {
4154 p = p->next;
4155 if (p == &cache_chain)
4156 return NULL;
4157 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08004158 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159}
4160
4161static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4162{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004163 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004165 return cachep->next.next == &cache_chain ?
4166 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167}
4168
4169static void s_stop(struct seq_file *m, void *p)
4170{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004171 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172}
4173
4174static int s_show(struct seq_file *m, void *p)
4175{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004176 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004177 struct slab *slabp;
4178 unsigned long active_objs;
4179 unsigned long num_objs;
4180 unsigned long active_slabs = 0;
4181 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004182 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004184 int node;
4185 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187 active_objs = 0;
4188 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004189 for_each_online_node(node) {
4190 l3 = cachep->nodelists[node];
4191 if (!l3)
4192 continue;
4193
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004194 check_irq_on();
4195 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004196
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004197 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004198 if (slabp->inuse != cachep->num && !error)
4199 error = "slabs_full accounting error";
4200 active_objs += cachep->num;
4201 active_slabs++;
4202 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004203 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004204 if (slabp->inuse == cachep->num && !error)
4205 error = "slabs_partial inuse accounting error";
4206 if (!slabp->inuse && !error)
4207 error = "slabs_partial/inuse accounting error";
4208 active_objs += slabp->inuse;
4209 active_slabs++;
4210 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004211 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004212 if (slabp->inuse && !error)
4213 error = "slabs_free/inuse accounting error";
4214 num_slabs++;
4215 }
4216 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004217 if (l3->shared)
4218 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004219
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004220 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004222 num_slabs += active_slabs;
4223 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004224 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225 error = "free_objects accounting error";
4226
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004227 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004228 if (error)
4229 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4230
4231 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004232 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004233 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004235 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004236 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004237 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004239 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 unsigned long high = cachep->high_mark;
4241 unsigned long allocs = cachep->num_allocations;
4242 unsigned long grown = cachep->grown;
4243 unsigned long reaped = cachep->reaped;
4244 unsigned long errors = cachep->errors;
4245 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004246 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004247 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004248 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249
Christoph Lametere498be72005-09-09 13:03:32 -07004250 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004251 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004252 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004253 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254 }
4255 /* cpu stats */
4256 {
4257 unsigned long allochit = atomic_read(&cachep->allochit);
4258 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4259 unsigned long freehit = atomic_read(&cachep->freehit);
4260 unsigned long freemiss = atomic_read(&cachep->freemiss);
4261
4262 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004263 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264 }
4265#endif
4266 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267 return 0;
4268}
4269
4270/*
4271 * slabinfo_op - iterator that generates /proc/slabinfo
4272 *
4273 * Output layout:
4274 * cache-name
4275 * num-active-objs
4276 * total-objs
4277 * object size
4278 * num-active-slabs
4279 * total-slabs
4280 * num-pages-per-slab
4281 * + further values on SMP and with statistics enabled
4282 */
4283
Helge Deller15ad7cd2006-12-06 20:40:36 -08004284const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004285 .start = s_start,
4286 .next = s_next,
4287 .stop = s_stop,
4288 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289};
4290
4291#define MAX_SLABINFO_WRITE 128
4292/**
4293 * slabinfo_write - Tuning for the slab allocator
4294 * @file: unused
4295 * @buffer: user buffer
4296 * @count: data length
4297 * @ppos: unused
4298 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004299ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4300 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004302 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004303 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004304 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004305
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306 if (count > MAX_SLABINFO_WRITE)
4307 return -EINVAL;
4308 if (copy_from_user(&kbuf, buffer, count))
4309 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004310 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311
4312 tmp = strchr(kbuf, ' ');
4313 if (!tmp)
4314 return -EINVAL;
4315 *tmp = '\0';
4316 tmp++;
4317 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4318 return -EINVAL;
4319
4320 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004321 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004322 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004323 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004325 if (limit < 1 || batchcount < 1 ||
4326 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004327 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004328 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004329 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004330 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331 }
4332 break;
4333 }
4334 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004335 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004336 if (res >= 0)
4337 res = count;
4338 return res;
4339}
Al Viro871751e2006-03-25 03:06:39 -08004340
4341#ifdef CONFIG_DEBUG_SLAB_LEAK
4342
4343static void *leaks_start(struct seq_file *m, loff_t *pos)
4344{
4345 loff_t n = *pos;
4346 struct list_head *p;
4347
4348 mutex_lock(&cache_chain_mutex);
4349 p = cache_chain.next;
4350 while (n--) {
4351 p = p->next;
4352 if (p == &cache_chain)
4353 return NULL;
4354 }
4355 return list_entry(p, struct kmem_cache, next);
4356}
4357
4358static inline int add_caller(unsigned long *n, unsigned long v)
4359{
4360 unsigned long *p;
4361 int l;
4362 if (!v)
4363 return 1;
4364 l = n[1];
4365 p = n + 2;
4366 while (l) {
4367 int i = l/2;
4368 unsigned long *q = p + 2 * i;
4369 if (*q == v) {
4370 q[1]++;
4371 return 1;
4372 }
4373 if (*q > v) {
4374 l = i;
4375 } else {
4376 p = q + 2;
4377 l -= i + 1;
4378 }
4379 }
4380 if (++n[1] == n[0])
4381 return 0;
4382 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4383 p[0] = v;
4384 p[1] = 1;
4385 return 1;
4386}
4387
4388static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4389{
4390 void *p;
4391 int i;
4392 if (n[0] == n[1])
4393 return;
4394 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4395 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4396 continue;
4397 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4398 return;
4399 }
4400}
4401
4402static void show_symbol(struct seq_file *m, unsigned long address)
4403{
4404#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004405 unsigned long offset, size;
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004406 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1];
Al Viro871751e2006-03-25 03:06:39 -08004407
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004408 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004409 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004410 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004411 seq_printf(m, " [%s]", modname);
4412 return;
4413 }
4414#endif
4415 seq_printf(m, "%p", (void *)address);
4416}
4417
4418static int leaks_show(struct seq_file *m, void *p)
4419{
4420 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004421 struct slab *slabp;
4422 struct kmem_list3 *l3;
4423 const char *name;
4424 unsigned long *n = m->private;
4425 int node;
4426 int i;
4427
4428 if (!(cachep->flags & SLAB_STORE_USER))
4429 return 0;
4430 if (!(cachep->flags & SLAB_RED_ZONE))
4431 return 0;
4432
4433 /* OK, we can do it */
4434
4435 n[1] = 0;
4436
4437 for_each_online_node(node) {
4438 l3 = cachep->nodelists[node];
4439 if (!l3)
4440 continue;
4441
4442 check_irq_on();
4443 spin_lock_irq(&l3->list_lock);
4444
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004445 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004446 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004447 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004448 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004449 spin_unlock_irq(&l3->list_lock);
4450 }
4451 name = cachep->name;
4452 if (n[0] == n[1]) {
4453 /* Increase the buffer size */
4454 mutex_unlock(&cache_chain_mutex);
4455 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4456 if (!m->private) {
4457 /* Too bad, we are really out */
4458 m->private = n;
4459 mutex_lock(&cache_chain_mutex);
4460 return -ENOMEM;
4461 }
4462 *(unsigned long *)m->private = n[0] * 2;
4463 kfree(n);
4464 mutex_lock(&cache_chain_mutex);
4465 /* Now make sure this entry will be retried */
4466 m->count = m->size;
4467 return 0;
4468 }
4469 for (i = 0; i < n[1]; i++) {
4470 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4471 show_symbol(m, n[2*i+2]);
4472 seq_putc(m, '\n');
4473 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004474
Al Viro871751e2006-03-25 03:06:39 -08004475 return 0;
4476}
4477
Helge Deller15ad7cd2006-12-06 20:40:36 -08004478const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004479 .start = leaks_start,
4480 .next = s_next,
4481 .stop = s_stop,
4482 .show = leaks_show,
4483};
4484#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004485#endif
4486
Manfred Spraul00e145b2005-09-03 15:55:07 -07004487/**
4488 * ksize - get the actual amount of memory allocated for a given object
4489 * @objp: Pointer to the object
4490 *
4491 * kmalloc may internally round up allocations and return more memory
4492 * than requested. ksize() can be used to determine the actual amount of
4493 * memory allocated. The caller may use this additional memory, even though
4494 * a smaller amount of memory was initially specified with the kmalloc call.
4495 * The caller must guarantee that objp points to a valid object previously
4496 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4497 * must not be freed during the duration of the call.
4498 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004499size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004500{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004501 if (unlikely(objp == NULL))
4502 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004503
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004504 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004505}